text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
abstract: 'We study several properties of distillation protocols to purify multilevel qubit states (qudits) when applied to a certain family of initial mixed bipartite states. We find that it is possible to use qudits states to increase the stability region obtained with the flow equations to distill qubits. In particular, for qutrits we get the phase diagram of the distillation process with a rich structure of fixed points. We investigate the large-$D$ limit of qudits protocols and find an analytical solution in the continuum limit. The general solution of the distillation recursion relations is presented in an appendix. We stress the notion of weight amplification for distillation protocols as opposed to the quantum amplitude amplification that appears in the Grover algorithm. Likewise, we investigate the relations between quantum distillation and quantum renormalization processes.'
author:
- 'M.A. Martín-Delgado and M. Navascués'
title: |
Distillation Protocols for Mixed States of Multilevel Qubits and\
the Quantum Renormalization Group
---
Introduction {#sec1}
============
The experimental analysis of the intriguing properties of entanglement in quantum mechanics requires the availability of stable sources of entanglement. Despite the nice properties exhibited by entanglement, it has the odd behaviour of degrading by the unavoidable contact with the external environment. Thus, for the entanglement to be assessed as a precious mean, we must devise some method to pump it up to the entanglement source in order to sustain a prescribed degree of entanglement that we may need whether for quantum communication protocols (teleportation, cryptography, dense coding) or quantum computing (algorithmics) (for a review see [@review1], [@review2] and references therein).
Quantum distillation or purification protocols are precisely those methods, that have been devised to regenerate entanglement leakages of an entanglement source. Here we are interested in the purification of mixed states of bipartite type, having in mind the realization of a communication protocol by two parties, Alice and Bob. The seminal work of [@bennett1] has provided us with a standard distillation method that has been the focus to developing more protocols with the aim at improving its original performances. We shall refer to this distillation protocol as the BBPSSW protocol. There are feasible experimental proposals for this type of protocols using polarization beam splitters (PBS) [@zeilinger1]. Likewise, there also exist methods for the distillation of pure states [@gisin] that have been implemented experimentally [@gisin3].
In addition to the initial purpose for which the quantum distillation protocols were devised, they have found another very important application in connection to the problem of quantum error correction: quantum information needs to be protected from errors even more than classical information due to its tendency to become decoherent. To avoid these errors, one can resort to the ideas of quantum error correction codes [@shor], [@steane] and fault-tolerant quantum computation [@gottesman], [@preskill]. However, entanglement purification is another alternative to decoherence which gives a more powerful way of dealing with errors in quantum communication [@bennett2].
In a typical quantum communication experiment, Alice and Bob are two spatially separated parties sharing pairs of entangled qubits. The type of operations allowed on these qubits are denoted as LOCC (Local Operations and Classical Communication): they comprise local unitary operators $U_{\rm A}\otimes U_{\rm B}$ on each side, local quantum measurements and communication of the measurement results through a classical channel. These local quantum operations will suffer from imperfections producing local errors. Futhermore, Alice and Bob will also face transmission errors in their quantum channels due to dissipation and noise. To overcome these difficulties, they will have to set up an entanglement purification method. In short, a protocol like the BBPSSW creates a reduced set of maximally entangled pairs (within a certain accuracy) out of a larger set of imperfectly entangled pairs: entanglement is created at the expense of wasting extra pairs. The degree of purity of a mixed entangled pair is measured in terms of its fidelity with respect to a maximally entangled pure pair, which is the focus of the purification protocol. After the BBPSSW protocol, a new distillation protocol was introduced in [@qpa] by the name of Quantum Privacy Amplification(QPA) which converges much faster to the desired fidelity [@chiara1], [@chiara2]. Other protocols known as quantum repeaters [@repeaters-pur], [@repeaters-com] allow us to stablish quantum communication over long distances by avoiding absortion or depolarization errors that scale exponentially with the length of the quantum channel.
The advantages of dealing with $D$-dimensional or multilevel quantum states (qudits) instead of qubits are quite apparent: an increase in the information flux through the communication channels that could speed up quantum cryptography, etc. [@pasqui1], [@pasqui2], [@karlson]. Thus, it has been quite natural to propose extensions of the purification protocols for qudits. One of the proposals [@horodeckis] relies on an extension of the CNOT gate that is unitary, but not Hermitian. Recently, another very nice proposal has been introduced [@gisin1], [@gisin2] based on a generalization of the CNOT gate that is both unitary and Hermitian and gives a higher convergence. In this paper we make a study of the new purification protocols of [@gisin1], [@gisin2] when they are applied to mixed bipartite states of qudits that are not of the Werner form. In this way, we combine some of the tools employed by the QPA protocols [@qpa] with the advantages of the new methods.
This paper is organized as follows: in Sect. \[sec2\] we review simple distillation protocols for qubits not in Werner states and we generalize them for the purification of any of the Bell states. In Sect. \[sec3\] we extend the previous protocols to deal with multilevel qubits and obtain several results like an improvement in the size of the stability fidelity basin, analytical formulas for the distillation flows, phase diagrams, etc. In Sect. \[sec4\] we apply the distillation protocols for the purification of non-diagonal mixed states that are more easily realized experimentally. In Sect. \[sec5\] we study the large-$D$ limit of these protocols. In Sect. \[sec6\] we present a detailed investigation of the relationships between quantum distillation protocols and renormalization methods for quantum lattice Hamiltonians. Section \[sec7\] is devoted to conclusions. In appendix \[app\] we find the general solution for the distillation recursion relations used in the text in the general case of qudits.
Simple Distillation Protocols with Qubits {#sec2}
=========================================
Our starting point is the orthonormal basis of Bell states formed by the first qubit belonging to Alice and the second to Bob: $$\begin{split}
\ket{\Phi^{\pm}} &:= \frac{1}{\sqrt{2}}[\ket{00}\pm \ket{11}]\\
\ket{\Psi^{\pm}} &:= \frac{1}{\sqrt{2}}[\ket{01}\pm \ket{10}]
\end{split}
\label{s1}$$ We shall use the word “simple” applied to the distillation protocols to denote that the mixed state we shall be dealing with is made up of a combination of one state in the set $S:=\{ \ket{\Phi^+}, \ket{\Phi^-}\}$ of Bell states that have coincident bits in Alice’s and Bob’s qubits, with another state in the set $A:=\{ \ket{\Psi^+}, \ket{\Psi^-}\}$ of Bell states that do not have coincidences. Thus, we have 4 possible combinations to do this type of entanglement distillation.
To begin with, we shall choose the following mixed state in order to set up a simple distillation protocol $$\rho_{++} := F\ket{\Phi^{+}}\bra{\Phi^{+}} +
(1-F)\ket{\Psi^{+}}\bra{\Psi^{+}}
\label{s2}$$ Alice and Bob will also need to apply the CNOT gate defined as usual $$U_{\rm CNOT} \ket{i}\ket{j}:= \ket{i}\ket{i\oplus j}, \ i,j=0,1.
\label{s2b}$$
The distillation protocol can be arranged into a set of 5 instructions or steps [@bennett1],[@qpa],[@chiara2]:
1. Set up $\rho \longrightarrow \rho \otimes \rho$ with fidelity $F$.
2. Apply bilateral CNOT gate: $U_{\rm BCNOT}$.
3. Alice and Bob measure target qubits.
4. Classical communication of results: retain coincidences ($0_A0_B$ or $1_A1_B$).
5. Go to step 1) with $\rho'$ with fidelity $F'>F$.
The simplicity of this protocol also relies on the fact that we do not need any depolarization step, as it is the case when dealing with Werner states [@bennett1]. In Fig. \[fig1\] we show a schematic picture of a single aplication of the purification method. Let us comment on the outcomes corresponding to the most relevant steps in this protocol. After step 1/, the 4-quit mixed state $\rho \otimes \rho$ shared by Alice and Bob reads as follows $$\begin{split}
&\rho_{++} \otimes \rho_{++}
= F^2\ket{\Phi^{+}\Phi^{+}}\bra{\Phi^{+}\Phi^{+}}\\
&+ F(1-F)[\ket{\Phi^{+}\Psi^{+}}\bra{\Phi^{+}\Psi^{+}} +
\ket{\Psi^{+}\Phi^{+}}\bra{\Psi^{+}\Phi^{+}}] \\
&+ (1-F)^2\ket{\Psi^{+}\Psi^{+}}\bra{\Psi^{+}\Psi^{+}}
\end{split}
\label{s3}$$ In step 2/, Alice and Bob apply bilaterally the CNOT gate taking their first qubit as source and their second qubit as target, i.e., qubits first and third are source qubits while qubits second and fourth are target qubits. To obtain the transformed mixed state we must determine the action of the bilateral CNOT gate $U_{\rm BCNOT}$ [@bennett1] on the states of the form $\ket{\varphi_A\varphi_B}$. The results of this computation are shown in Table \[table1\]. With the help of this table we find $$\begin{split}
& U_{\rm BCNOT} \rho_{++} \otimes \rho_{++} U_{\rm BCNOT} =
F^2\ket{\Phi^{+}\Phi^{+}}\bra{\Phi^{+}\Phi^{+}}\\
&+ F(1-F)[\ket{\Phi^{+}\Psi^{+}}\bra{\Phi^{+}\Psi^{+}} +
\ket{\Psi^{+}\Psi^{+}}\bra{\Psi^{+}\Psi^{+}}] \\
&+ (1-F)^2\ket{\Psi^{+}\Phi^{+}}\bra{\Psi^{+}\Phi^{+}}
\end{split}
\label{s4}$$
\[Bc\]\[Bc\]\[1\]\[0\][Alice]{} \[Bc\]\[Bc\]\[1\]\[0\][Bob]{} \[Bc\]\[Bc\]\[1\]\[0\][$F$]{} \[Bc\]\[Bc\]\[1\]\[0\][$F$]{} \[Bc\]\[Bc\]\[1\]\[0\][$F$]{} \[Bc\]\[Bc\]\[1\]\[0\][$F$]{} \[Bc\]\[Bc\]\[1\]\[0\][CNOT]{} \[Bc\]\[Bc\]\[1\]\[0\][CNOT]{} \[Bc\]\[Bc\]\[1\]\[0\][Quantum]{} \[Bc\]\[Bc\]\[1\]\[0\][Distillation]{} \[Bc\]\[Bc\]\[1\]\[0\][$F'>F$]{} \[Bc\]\[Bc\]\[1\]\[0\][$F'>F$]{} ![Schematic representation of the distillation protocol by Alice and Bob. Originally, two pairs of shared entangled qubits represented by enclosed dots are transformed into a single pair of higher purity (doubly enclosed dots).[]{data-label="fig1"}](protocolqubits.eps "fig:"){width="6"}
After steps 3/ and 4/, Alice and Bob measure their target qubits and retain their source qubit whenever they find, via classical communication, the same results: either $0_A0_B$ or $1_A1_B$. This fact selects the state $\ket{\Phi^+}$ as the only admissible possibility for the target state. Thus, only the first and third terms in the RHS of (\[s4\]) survive to this process and the resulting 2-qubit state $\rho'_{++}$ is again of the same form as the original starting state (\[s2\]) in step 1/, but with a higher fidelity $F'>F$. In fact, we get $$\rho'_{++} := F'\ket{\Phi^{+}}\bra{\Phi^{+}} +
(1-F')\ket{\Psi^{+}}\bra{\Psi^{+}}
\label{s5}$$ with the new fidelity being $$F' = \frac{F^2}{F^2 + (1-F)^2}
\label{s6}$$ This relation defines a recursion scheme for entanglement purification: starting with say $N_P$ pairs of Bell states of fidelity $F$, after every application of the whole protocol we obtain $\frac{N_P}{2}$ pairs of higher fidelity $F'>F$. Thus, purification is achieved at the expense of halving the number of Bell pairs. The fixed points $F_c$ of the recursion relation (\[s6\]) are defined as $F'(F_c):=F_c$ and they are given by $F_c=0,\half,1$. The fixed points $F_c=0,1$ are stable, while $F_c=\half$ is unstable. The best way to recast these qualitative properties of the flow equation for the fidelities (\[s6\]) is to draw the corresponding flow diagram as shown in Fig. \[flowdiagram\].
$\ket{\varphi_A}\ket{\varphi_B}$ $U_{\rm BCNOT}\ket{\varphi_A}\ket{\varphi_B}$
---------------------------------- -----------------------------------------------
$\ket{\Phi^{+}}\ket{\Phi^{+}}$ $\ket{\Phi^{+}}\ket{\Phi^{+}}$
$\ket{\Phi^{+}}\ket{\Psi^{+}}$ $\ket{\Phi^{+}}\ket{\Psi^{+}}$
$\ket{\Psi^{+}}\ket{\Phi^{+}}$ $\ket{\Psi^{+}}\ket{\Psi^{+}}$
$\ket{\Psi^{+}}\ket{\Psi^{+}}$ $\ket{\Psi^{+}}\ket{\Phi^{+}}$
: This table shows the results of applying the bilateral CNOT gate to certain pairs of Bell states needed to distillation.[]{data-label="table1"}
\[Bc\]\[Bc\]\[1\]\[0\][0]{} \[Bc\]\[Bc\]\[1\]\[0\][1]{} \[Bc\]\[Bc\]\[1\]\[0\][$F_c$]{} \[Bc\]\[Bc\]\[1\]\[0\][$\half$]{} \[Bc\]\[Bc\]\[1\]\[0\][$F$]{} ![Flow diagram for the fidelity $F$ of the distillation protocol given by the recursion relation (\[s6\]).[]{data-label="flowdiagram"}](flowdiagram.eps "fig:"){width="6"}
Next, we may wonder whether it is possible to devise distillation protocols for the three possible combinations of Bell states, namely, $$\begin{split}
\rho_{+-} &:= F\ket{\Phi^{+}}\bra{\Phi^{+}} +
(1-F)\ket{\Psi^{-}}\bra{\Psi^{-}} \\
\rho_{-+} &:= F\ket{\Phi^{-}}\bra{\Phi^{-}} +
(1-F)\ket{\Psi^{+}}\bra{\Psi^{+}} \\
\rho_{--} &:= F\ket{\Phi^{-}}\bra{\Phi^{-}} +
(1-F)\ket{\Psi^{-}}\bra{\Psi^{-}}
\end{split}
\label{s7}$$ We can answer this question affirmatively by computing the action of the bilateral CNOT gate on the tensor product of these mixed states (\[s7\]). With a similar analysis which has led us to Table \[table1\] [@bennett1], we obtain $$\begin{split}
& U_{\rm BCNOT} \rho_{+-} \otimes \rho_{+-} U_{\rm BCNOT} =
F^2\ket{\Phi^{+}\Phi^{+}}\bra{\Phi^{+}\Phi^{+}}\\
&+ F(1-F)[\ket{\Phi^{-}\Psi^{-}}\bra{\Phi^{-}\Psi^{-}} +
\ket{\Psi^{-}\Psi^{+}}\bra{\Psi^{-}\Psi^{+}}] \\
&+ (1-F)^2\ket{\Psi^{+}\Phi^{-}}\bra{\Psi^{+}\Phi^{-}}\\
& U_{\rm BCNOT} \rho_{-+} \otimes \rho_{-+} U_{\rm BCNOT} =
F^2\ket{\Phi^{+}\Phi^{-}}\bra{\Phi^{+}\Phi^{-}}\\
&+ F(1-F)[\ket{\Phi^{-}\Psi^{+}}\bra{\Phi^{-}\Psi^{+}} +
\ket{\Psi^{-}\Psi^{-}}\bra{\Psi^{-}\Psi^{-}}] \\
&+ (1-F)^2\ket{\Psi^{+}\Phi^{+}}\bra{\Psi^{+}\Phi^{+}} \\
& U_{\rm BCNOT} \rho_{--} \otimes \rho_{--} U_{\rm BCNOT} =
F^2\ket{\Phi^{+}\Phi^{-}}\bra{\Phi^{+}\Phi^{-}}\\
&+ F(1-F)[\ket{\Phi^{+}\Psi^{-}}\bra{\Phi^{+}\Psi^{-}} +
\ket{\Psi^{+}\Psi^{-}}\bra{\Psi^{+}\Psi^{-}}] \\
&+ (1-F)^2\ket{\Psi^{+}\Phi^{-}}\bra{\Psi^{+}\Phi^{-}}
\end{split}
\label{s8}$$ We now realize that if we proceed to measure the target bits and classical communicate the results, we do not end up with the same type of mixed state as we had started with. That is, the protocol as it stands is not valid since it does not yield invariant mixed states. This problem has a solution provided we introduce an additional step prior to the measurement of the target qubits by Alice and Bob. This additional step corresponds to a local unitary operation $U_A\otimes U_B$ performed by Alice and Bob on their source qubits. The form of this local unitary depends on the mixed state we are distilling. We find the following results:
Step $2'$. Alice and Bob apply a local unitary transformation $U_A\otimes U_B$ to their source qubits:
For $\rho_{+-}$, $U_A = \half (1+\ii)(\sigma_x+\sigma_y)$, $U_B = \half (1-\ii)(\sigma_x-\sigma_y)$.
For $\rho_{-+}$, $U_A = U_B = \half (1+\ii)(\sigma_x+\sigma_y)$.
For $\rho_{--}$, $U_A = \sigma_z$, $U_B = 1$.
After this extra step, we can guarantee that the resulting 4-qubit mixed state has the appropriate Bell pairs at the source qubits so as to produce the same original state, once steps 3/ and 4/ are performed. Moreover, it is straightforward to prove that the new fidelity for these 3 protocols is also given by the same recursion relation (\[s6\]) as in the first protocol.
Finally, if the constraint that the state of fidelity $F$ must be a $\Phi^{\pm}$ state is relaxed, then there are two additional possible mixed states whose analysis can be carried out in a similar fashion.
Multilevel Extensions of Distillation Protocols {#sec3}
===============================================
In order to generalize the simple distillation protocol of the previous section to the case of qudits, we must notice that the two main ingredients in that distillation protocol are:
i\) the CNOT gate,
ii\) the Bell states (\[s1\]).
Regarding the CNOT gate, the extension of this gate to deal with qudits is not unique. As has been noted in [@gisin1],[@gisin2], the CNOT gate for qubits (\[s2b\]) has 3 properties that make it special, namely $$\begin{split}
U_{\rm CNOT}^{\dag} = U_{\rm CNOT}^{-1}, \\
U_{\rm CNOT}^{\dag} = U_{\rm CNOT}, \\
i\oplus j = 0 \Leftrightarrow i=j.
\end{split}
\label{g1}$$ The extension of the CNOT gate for qudits that satisfies these 3 properties (\[g1\]) is given by [@gisin1],[@gisin2] $$U_{\rm CNOT} \ket{i}\ket{j}:=\ket{i}\ket{i\ominus j}, \ i,j=0,\ldots, D-1
\label{g2}$$ where $i\ominus j:=i-j, {\rm mod} \ D$, denotes substraction modulus $D$. This is the definition that we shall adopt throughout this paper, unless otherwise stated.
As for the higher-dimensional extension of Bell states (\[s1\]), we shall also take the following generalization [@gisin1],[@gisin2] $$\ket{\Psi_{kj}}:=U_{\rm CNOT} \left[(U_{\rm F}\ket{k})\otimes \ket{j}\right], \ k,j=0,\ldots, D-1
\label{g3}$$ where $U_{\rm F}$ is the quantum Fourier transform (QFT) $$U_{\rm F}\ket{k}:=
\frac{1}{\sqrt{D}}\sum_{y=0}^{D-1}\ee^{\frac{2\pi \ii k y}{D}}\ket{y},
\label{g4}$$ which reduces to the Hadamard gate when dealing with qubits ($D=2$). As a matter of fact, we can readily check that for the special case of qubits $D=2$ we recover the standard Bell pairs (\[s1\]) in the following form $$\begin{split}
\ket{\Psi_{00}} = \ket{\Phi^+}, & \ket{\Psi_{01}} = \ket{\Psi^+},\\
\ket{\Psi_{10}} = \ket{\Phi^-}, & \ket{\Psi_{11}} = \ket{\Psi^-}
\end{split}
\label{g5}$$ Moreover, using the generalized CNOT gate (\[g2\]), the generalized Bell states are given by $$\ket{\Psi_{kj}} =
\frac{1}{\sqrt{D}}\sum_{y=0}^{D-1}\ee^{\frac{2\pi \ii k y}{D}}
\ket{y}\ket{y\ominus j}.
\label{g6}$$
With these extensions of the CNOT gate and the Bell states, we can set up a generalization of the simple distillation protocols of Sect. II for qudit states. These protocols have the same 5 steps as before.
[*Step 1/*]{}. We shall assume a general diagonal mixed state of the form $$\begin{split}
\rho &:= \sum_{k,j=0}^{D-1} q_{kj}\ket{\Psi_{ij}}\bra{\Psi_{ij}},\\
1 &=: \sum_{k,j=0}^{D-1} q_{kj},
\end{split}
\label{g7}$$ where $q_{kj}$ are normalized probabilities. For non-diagonal mixed states, we refer to Sect. IV. Then, Alice and Bob share pairs $\rho \times \rho$ of these states (\[g7\]).
[*Step 2/*]{}. Alice and Bob apply bilaterally the generalized CNOT gate (\[g2\]). To know the result of this operation on the state (\[g7\]) we need a previous result about the action of the gate $U_{\rm BCNOT}$ on pairs of generalized Bell states (\[g3\]). After some algebra, we arrive at the following expression $$U_{\rm BCNOT} \ket{\Psi_{kj}}\ket{\Psi_{k'j'}} =
\ket{\Psi_{k\oplus k',j}}\ket{\Psi_{D\ominus k',j\ominus j'}}
\label{g8}$$ This is a fundamental result for it means that [*the space of two-pairs of generalized Bell states is invariant under the action of the generalized bilateral CNOT gate.*]{} This is a very nice result that condenses in a single formula all the possibilities for the outcome of the action of the CNOT gates on Bell states, in particular, the whole table employed by Bennett et al. in [@bennett1] for the case of qubits is contained in equation (\[g8\]). This property is essential in order to have a closed distillation protocol. Actually, it would have been enough to have obtained only the source qubits as generalized Bell states.
Then, with the help of this property (\[g8\]) we obtain the action of $U_{\rm BCNOT}$ on pairs of $\rho$ states, as follows $$\begin{split}
& U_{\rm BCNOT} \rho \otimes \rho U_{\rm BCNOT} =
\sum_{k,j=0}^{D-1} \sum_{k',j'=0}^{D-1}
q_{k\ominus k',j} q_{k'j'} \\
& \ket{\Psi_{kj}\Psi_{D\ominus k',j\ominus j'}}
\bra{\Psi_{kj}\Psi_{D\ominus k',j\ominus j'}}
\end{split}
\label{g9}$$ We see that this state is already of the same form in the source qubits as the original $\rho$ (\[g7\]).
[*Step 3/*]{}. Alice and Bob measure their target qubits in (\[g9\]). To see the result of this measurement, let us write the explicit form of the target qubits, namely $$\Psi_{D\ominus k',j\ominus j'} =
\frac{1}{\sqrt{D}}\sum_{z=0}^{D-1}\ee^{\frac{-2\pi \ii k' z}{D}}
\ket{z}\ket{z\ominus (j\ominus j')}.
\label{g10}$$ Therefore, coincidences between Alice’s and Bob’s target qubits will happen only when the following condition is satisfied $$z = z\ominus (j\ominus j') \Longleftrightarrow j = j'.
\label{g11}$$
[*Step 4/*]{}. After their measurement, Alice and Bob communicate classically their result so that they retain the resulting source Bell pairs only when they have coincidences, and discard them otherwise. The resulting net effect of this process is to produce a Kronecker delta function $\delta_{jj'}$ in the target qubits. More precisely, the resulting unnormalized mixed state is given by $$\rho' \sim \sum_{k,j=0}^{D-1} \sum_{k',j'=0}^{D-1}
q_{k\ominus k',j} q_{k'j'} \delta_{jj'}
\ket{\Psi_{kj}} \bra{\Psi_{kj}}.
\label{g12}$$ Therefore, we end up with a diagonal mixed state of the same form as the starting one $$\rho' = \sum_{k,j=0}^{D-1}
q'_{kj} \ket{\Psi_{kj}} \bra{\Psi_{kj}}.
\label{g13}$$ with the new probabilities given by $$q'_{kj} = \frac{\sum_{k'=0}^{D-1}q_{k\ominus k',j} q_{k'j}}
{\sum_{k,j=0}^{D-1}\sum_{k'=0}^{D-1}q_{k\ominus k',j} q_{k'j}}.
\label{g14}$$ This is a generalized recursion relation that includes eq. (\[s6\]) as a particular instance.
[*Step 5/*]{}. Alice and Bob start all over again the same process with the initial state now being $\rho'$ in (\[g13\]-\[g14\]).
The nice feature of these generalized distillation protocols for dealing with qudits is the fact that we have at our disposal explicit analytical formulas (\[g14\]) for the evolution (flow) of the different weights (probabilities) of the generalized mixed states to be purified. As these distillation protocols are too general, it is worthwhile to consider some particular cases of interest separately. The general solution to the distillation recursion relations (\[g14\]) is presented in the appendix \[app\]. We hereby provide the following analysis of some examples:
[*i).*]{} Let us investigate the closest generalization of the simple protocols introduced in Sect. II. Thus, let us consider the following type of initial mixed state $$\begin{split}
\rho &:= \sum_{i=0}^{M} q_i \ket{\Psi_{0i}}\bra{\Psi_{0i}}, \ M\leq D-1,\\
1 &=: \sum_{i=0}^{M}q_i, \ q_i \geq 0,
\end{split}
\label{g15}$$ This corresponds to working with the subset of all possible generalized Bell states of the form $\{ \ket{\Psi_{0i}} \}_{i=0}^{D-1}$. Interestingly enough, this includes the case of the state $\rho_{++}$ in (\[s2\]) for $D=2$. The recursion relations (\[g14\]) for this special subset of states takes the following simpler form $$q'_i = \frac{q_i^2}{\sum_{j=0}^M q_j^2}.
\label{g16}$$
For $M=2$, i.e., considering a mixed state formed of just two Bell states of the form $\ket{\Psi_{0i}}$, the protocol has the following recursion relation $$q'_i = \frac{q_i^2}{q_i^2 + (1-q_i)^2},
\label{g17}$$ where here the index $i$ stands for any possible pair of Bell states of the type $\ket{\Psi_{0i}}$. In other words, we have found a direct $D$-dimensional generalization of the distillation protocols for qubits in Sect. II, with $q_i:=F$.
[*ii).*]{} For $M=D-1$ and taking $q_0:=F$ and $q_i:=\frac{1-F}{D-1}, i=1,\ldots,D-1$ we can find a more advantageous protocol than the previous one. In fact, in this case we find that $$q'_0:=F'= \frac{F^2}{F^2 + \frac{(1-F)^2}{D-1}}.
\label{g18}$$ The fixed points of this recursion relation are now given by $F_c=0,\frac{1}{D},1$. Despite being a non-linear recursion relation, (\[g18\]) admits an explicit analytical solution for the general term of the series $F_{k}$ given by $$F_{k} = \frac{F^{(2^k)}}{F^{(2^k)} +
(D-1) \left[\frac{1-F}{D-1}\right]^{(2^k)}}, \ k\geq 1, F_0:=F.
\label{g18b}$$ From this solution, we inmediately find that the fixed points $F_c=0,1$ are stable while $F_c=\frac{1}{D}$ is unstable.
In Fig. \[recursionFD\] we plot the function $F'=F'(F)$ for several values of the dimension $D$. From the analysis of these curves we inmediately obtain the corresponding flow diagram that we represent in Fig. \[flowdiagramD\]. We check that for $D=2$ we recover the flow diagram corresponding to standard qubits (Fig. \[flowdiagram\]).
\[Bc\]\[Bc\]\[1\]\[0\][$F$]{} \[Bc\]\[Bc\]\[1\]\[0\][$F'$]{}
\[Bc\]\[Bc\]\[1\]\[0\][0]{} \[Bc\]\[Bc\]\[1\]\[0\][1]{} \[Bc\]\[Bc\]\[1\]\[0\][$F_c$]{} \[Bc\]\[Bc\]\[1\]\[0\][$\frac{1}{D}$]{} \[Bc\]\[Bc\]\[1\]\[0\][$F$]{} ![Flow diagram for the fidelity $F$ of the generalized distillation protocol for qudits given by the recursion relation (\[g18\]).[]{data-label="flowdiagramD"}](flowdiagramD.eps "fig:"){width="6"}
We see from Fig. \[flowdiagramD\] that the stability basin is increased with respect to the case of standard qubits, as in Fig. \[flowdiagram\]. This means that we can start with a mixed state having a fidelity $F$ with respect to the Bell state $\ket{\Psi_{00}}$ lower that $\half$ and we still will succeed in purifiying that state towards fidelity close to 1. Thus, we have found that it is more advantageous to distill a given Bell state $\ket{\Psi_{00}}$ if we prepare the mixed state $\rho$ in (\[g15\]) in the form $$\rho := F \ket{\Psi_{00}} \bra{\Psi_{00}} +
\frac{1-F}{D-1} \sum_{i=1}^{D-1} \ket{\Psi_{0i}} \bra{\Psi_{0i}},
\label{g19}$$ rather than using just one single of those states $$\rho := F \ket{\Psi_{00}} \bra{\Psi_{00}} +
(1-F) \ket{\Psi_{0i}} \bra{\Psi_{0i}}, \ i\neq 0.
\label{g20}$$
We may wonder how is it likely for Alice and Bob to obtain the same values (coincidences) after measuring the target qudits in the step 3/ of the distillation protocol. Let us denote by ${\cal P}_{\rm AB}$ this probability which will depend on the value $F$ of the fidelity. From equations (\[g12\]) and (\[g18\]) we find this probability of coincidences to be $${\cal P}_{\rm AB}(F) = F^2 + \frac{(1-F)^2}{D-1}.
\label{g20b}$$ The minimum of this probability is at $F_0=\frac{1}{D}$ and its value is ${\cal P}_{\rm AB}(\frac{1}{D})=\frac{1}{D}$. Likewise, ${\cal P}_{\rm AB}(1)=1$. Thus, we find that the probability is lower and upper bounded as ${\frac{1}{D}\leq\cal P}_{\rm AB}(F)\leq F$ for $F\in[\frac{1}{D},1]$.
One is also interested in knowing the number of steps $K(\epsilon, F_0)$ needed to achieve a certain final fidelity close to 1, say $1-\epsilon$, starting from an appropriate initial fidelity $F_0>\frac{1}{D}$. This number can be computed from our analytical solution (\[g18b\]) from the condition $$F_{K(\epsilon, F_0)}:=1-\epsilon.
\label{g21}$$ Thus, we find the following analytical formula for the number of steps needed to obtain a certain degree of fidelity $\epsilon$ as a function of the initial fidelity $F_0>\frac{1}{D}$,i.e., $$K(\epsilon, F_0) = \left\lceil \log_2 \left(
\frac{\ln (\frac{\epsilon}{(1-\epsilon)(D-1)})}
{\ln (\frac{1-F_0}{(D-1)F_0})} \right) \right\rceil .
\label{g22}$$
\[Bc\]\[Bc\]\[1\]\[0\][$F_0$]{} \[Bc\]\[Bc\]\[1\]\[0\][$K$]{} ![Plot of the number of iterations $K(\epsilon, F_0)$ (\[g22\]) to achieve a final fidelity of $F=0.99$ as a function of the initial fidelity $F_0$ and for several values of the dimension $D$ of the qudits: $D=2,4,10$.[]{data-label="iterations"}](iterations2.eps "fig:"){width="8"}
In Fig. \[iterations\] we plot the number of iterations (\[g22\]) for a given value of the final fidelity $1-\epsilon$ that we take as the fixed value of $0.99$, and then we find how is the dependence on the initial fidelity $F_0$. We see that for a given admissible value of $F_0$, the lowest number of iterations corresponds to the protocol with the higher value of the qudits dimension $D$: $$K(\epsilon_0, F_0)_{D_1} \geq K(\epsilon_0, F_0)_{D_2}, \ {\rm for} \ D_1>D_2.
\label{g23}$$
[*iii).*]{} For qutrits, $D=3$, the most general diagonal mixed state with the allowed Bell states taking values on the set $\{ \ket{\Psi_{0i}} \}$ is $$\begin{split}
\rho &:= q_0\ket{\Psi_{00}} + q_1\ket{\Psi_{01}} + q_2\ket{\Psi_{02}} \\
1 &=: q_0 + q_1 + q_2
\end{split}
\label{g24}$$ Let us assume that the state we want to purify is $\ket{\Psi_{00}}$. Now, our recursion relation for our fidelity $q_0$ depends on two variables, namely, $$q'_0 = \frac{q_0^2}{q_0^2 + q_1^2 + (1-q_0-q_1)^2},
\label{g25}$$ and a similar equation for $q_1$ with $q_0 \leftrightarrow q_1$. In Fig. \[qutrits\] the dependence of the function fidelity $q'_0=q'_0(q_0,q_1)$ for qutrits is plotted. We observe that it is a monotonous incresing function which gurantees that the initial fidelity will flow towards 1, under certain conditions.
\[Bc\]\[Bc\]\[1\]\[0\][$q_0$]{} \[Bc\]\[Bc\]\[1\]\[0\][$q_1$]{} \[Bc\]\[Bc\]\[1\]\[0\][$q'_0$]{} ![The desired fidelity $q'_0$ for amplification after a single application of the distillation protocol for qutrits (\[g25\]) as a function of the previous fidelitites $q_0,q_1$.[]{data-label="qutrits"}](qutrits3.eps "fig:"){width="8"}
To find these conditions, we find that the set of fixed points of these recursion relations is given by $$(q_0,q_1)_c=\{ (0,0), (\half,0), (0,\half), (\frac{1}{3},\frac{1}{3}),
(\half,\half), (1,0), (0,1) \}.$$ We have also found the flow diagram associated to these recursion relations which is now two-dimensional and we show it in Fig. \[2Dflow\]. From this diagram we see that the purification protocol is successful in arriving to the maximum fidelity $q_0=1$ provided the initial fidelity lies in the stability basin of the fixed point $(1,0)$ which is given by the trapezoid formed by the set of points $(\half,\half),(\frac{1}{3},\frac{1}{3}),(\half,0),(1,0)$.
\[Bc\]\[Bc\]\[1\]\[0\][$(0,0)$]{} \[Bc\]\[Bc\]\[1\]\[0\][$(\half,0)$]{} \[Bc\]\[Bc\]\[1\]\[0\][$(1,0)$]{} \[Bc\]\[Bc\]\[1\]\[0\][$(\half,\half)$]{} \[Bc\]\[Bc\]\[1\]\[0\][$(0,1)$]{} \[Bc\]\[Bc\]\[1\]\[0\][$(0,\half)$]{} \[Bc\]\[Bc\]\[1\]\[0\][$q_0$]{} \[Bc\]\[Bc\]\[1\]\[0\][$q_1$]{} ![Two-dimensional flow diagram associated to the distillation protocol for qutrits (\[g25\]).[]{data-label="2Dflow"}](2Dflow.eps "fig:"){width="6"}
Distillation of Non-Diagonal Mixed States {#sec4}
=========================================
So far, we have been investigating the properties of distillation protocols applied to mixed bipartite state of diagonal form such as those in (\[g13\]), (\[g14\]). It is apparent that once we have a general result for the operation of the $U_{\rm BCNOT}$ gate of generalized Bell states (\[g8\]), we can also deal with non-diagonal mixed states, namely, $$\begin{split}
\rho := \sum_{k,j=0}^{D-1}\sum_{k',j'=0}^{D-1}
q_{kjk'j'}\ket{\Psi_{kj}}\bra{\Psi_{k'j'}},\\
1 =: \sum_{k,j=0}^{D-1}\sum_{k',j'=0}^{D-1}
q_{kjk'j'}, \ q_{kjk'j'}\geq 0.
\end{split}
\label{g26}$$ Since this is a too much general state, we prefer to extract from this class of non-diagonal states one type which we believe it may have potential applications.
Let us imagine that Alice and Bob are manipulating bipartite qudit states that are [*diagonal*]{} in the computational basis. More explicitly, the entangled state they want to purify is of the form $$\ket{\Psi_{\rm d}} := \frac{1}{\sqrt{D}}\sum_{i=0}^{D-1}\ket{ii},
\label{g27}$$ while states which are [*non-diagonal*]{} are considered as acting as disturbing noise that they want to get rid of. Specifically, this noise will be represented by the state $$\ket{\Psi_{\rm o}} := \frac{1}{\sqrt{D(D-1)}}\sum_{i\neq j=0}^{D-1}\ket{ij}.
\label{g28}$$ Then, in order to achieve their goal of purifying states of the diagonal form $\ket{\Psi_{\rm d}}$ with respect to non-diagonal states $\ket{\Psi_{\rm o}}$, they set up a distillation protocol based on sharing copies of the following mixed state $$\rho := F \ket{\Psi_{\rm d}}\bra{\Psi_{\rm d}}
+ (1-F) \ket{\Psi_{\rm o}}\bra{\Psi_{\rm o}}.
\label{g29}$$ We envisage that this scenario is physically feasible since we can imagine that the computational basis is realized in terms of some physical property taking values on $i=0,\ldots,D-1$ and that Alice and Bob have a mechanism to select when they have $\ket{ii}$ coincident qudits (or diagonal) from $\ket{ij}$ non-coincident qubits (non-diagonal).
To proceed with the distillation of the state $\rho$ in (\[g29\]), we first must express the states $\ket{\Psi_{\rm d}}$ and $\ket{\Psi_{\rm o}}$ in the basis of the generalized Bell states, with the result $$\begin{split}
\ket{\Psi_{\rm d}} &= \ket{\Psi_{00}},\\
\ket{\Psi_{\rm o}} &= \frac{1}{\sqrt{D-1}} \sum_{i=1}^{D-1} \ket{\Psi_{0i}}.
\end{split}
\label{g30}$$ Next, Alice and Bob share two pairs of non-diagonal mixed states $$\begin{split}
&\rho \otimes \rho = F^2 \ket{\Psi_{00} \Psi_{00}}\bra{\Psi_{00} \Psi_{00}} \\
&+ \frac{F(1-F)}{D-1}\sum_{i,j=1}^{D-1}
[\ket{\Psi_{00} \Psi_{0i}}\bra{\Psi_{00} \Psi_{0j}} +
\ket{\Psi_{0i} \Psi_{00}}\bra{\Psi_{0j} \Psi_{00}}]\\
&+ \frac{(1-F)^2}{(D-1)^2}\sum_{i,j,k,l=1}^{D-1}
\ket{\Psi_{0i} \Psi_{0k}}\bra{\Psi_{0j} \Psi_{0l}},
\end{split}
\label{g31}$$ and they apply bilaterally the CNOT gate to it (\[g8\]) with the result $$\begin{split}
&U_{\rm BCNOT}\rho \otimes \rho U_{\rm BCNOT} = F^2 \ket{\Psi_{00} \Psi_{00}}\bra{\Psi_{00} \Psi_{00}} \\
&+ \frac{F(1-F)}{D-1}\sum_{i,j=1}^{D-1}
[\ket{\Psi_{00} \Psi_{0,\ominus i}}\bra{\Psi_{00} \Psi_{0,\ominus j}} +
\ket{\Psi_{0i} \Psi_{0i}}\bra{\Psi_{0j} \Psi_{0j}}]\\
&+ \frac{(1-F)^2}{(D-1)^2}\sum_{i,j,k,l=1}^{D-1}
\ket{\Psi_{0i} \Psi_{0i\ominus k}}\bra{\Psi_{0j} \Psi_{0j\ominus l}}.
\end{split}
\label{g32}$$ The process of measuring the target qudits and retaining the source qudits when upon classical communication Alice and Bob find coincidences in their measures amounts to retaining the terms in (\[g32\]) that have the state $\ket{\Psi_{00}}$ in the target qudits. This means that only the first term and part of the last term in (\[g32\]) contribute to the final source mixed state, which takes the following form without normalization $$\begin{split}
\rho' &\sim F^2 \ket{\Psi_{00}} \bra{\Psi_{00}}
+ \frac{(1-F)^2}{(D-1)^2}\left(\sum_{i=1}^{D-1} \ket{\Psi_{0i}}\right)
\left(\sum_{j=1}^{D-1} \bra{\Psi_{0j}}\right) \\
&= F^2 \ket{\Psi_{00}} \bra{\Psi_{00}} + \frac{(1-F)^2}{D-1}
\ket{\Psi_{\rm o}} \bra{\Psi_{\rm o}}.
\end{split}
\label{g33}$$ Upon normalization, we arrive again at a non-diagonal mixed state of same form as the one we started with $\rho' = F' \ket{\Psi_{\rm d}}\bra{\Psi_{\rm d}}
+ (1-F') \ket{\Psi_{\rm o}}\bra{\Psi_{\rm o}}$, but with a new fidelity $F'$ given by $$F'= \frac{F^2}{F^2 + \frac{(1-F)^2}{D-1}}.
\label{g34}$$ Let us notice that this is precisely the same recursion relation that we found in Sect. III in a different context (\[g18\]).
Continuum Limit of Qudit Protocols {#sec5}
==================================
For the general case represented by the recursion relations (\[g16\]) we can also find the general solution for the $k$-th iteration $q_i^{(k)}, i=0,\ldots,D-1$ starting from their initial values $q_i^{(0)}$ satisfying $\sum_{i=0}^{D-1} q_i^{(0)}:=1$. We find the following solution $$q_i^{(k)} = \frac{\left[q_i^{(0)}\right]^{2^k}}
{\sum_{j=0}^M \left[q_j^{(0)}\right]^{2^k}}.
\label{g16b}$$ Let us assume that the maximum initial value is $M:={\rm max}\ \{ q_i^{(0)}\}$ and it is $p$ times degenerate. Then, using the general solution (\[g16b\]) we can inmeditely find the fixed points after the evolution with the recursion relations. We find $$\begin{split}
\lim_{k\rightarrow \infty} q_i^{(k)} = 0, \ {\rm if} \ q_i^{(0)}<M,\\
\lim_{k\rightarrow \infty} q_i^{(k)} = \frac{1}{p}, \ {\rm if} \ q_i^{(0)}=M.
\end{split}
\label{g16c}$$
From the analysis of these distillation protocols and the way they operate we arrive at the conclusion that they resemble a sort of amplitude amplification quite similar to what happens in the Grover algorithm where there exists what is called [*quantum amplitude amplification*]{}. However, there is an important distintion between both procedures: in the distillation method, the maximum amplification is attained asymptotically, while in Grover algorithm it is achieved periodically. The reason for this difference relies on the fact that the distillation process is not unitary (since we make measurements and discard states), while Grover is unitary. Thus, we propose to refer to the distillation protocol as [*weight amplification*]{}, since it is certain probability weights of the intial mixed states, and not amplitudes, what are being amplified.
When $D$ is very large, we can approximate the probability weights $q_i^{(0)}$ taking values on the discrete set $\{ 0,1,\ldots,D-1\}\ni i$, by a density function $q(x)^{(0)}$ defined on the real interval $[0,1]$. This is achieved by introducing the variable $x\in[0,1]$ defined as $x:=i\Delta x$ with $\Delta x:=\frac{1}{D-1}$. Thus, in the limit $D\rightarrow \infty (\Delta x \rightarrow 0)$, we get a probability density as $q_i^{(0)}(i\Delta x)\rightarrow q^{(0)}(x)dx$. It is also normalized as
$$\int_{0}^1 q^{(0)}(x) dx = 1.
\label{g35}$$
Likewise, we can take the continuum limit of the general recursion equation (\[g16b\]) in order to obtain the probability density $q(x)^{(k)}$ after $k$ steps of the distillation protocol. This is given by $$q^{(k)}(x) = \frac{\left[q^{(0)}(x)\right]^{2^k}}
{\int_{0}^1 \left[q^{(0)}(y)\right]^{2^k} dy}.
\label{g36}$$
\[Bc\]\[Bc\]\[1\]\[0\][$x$]{} \[Bc\]\[Bc\]\[1\]\[0\][$q^{(k)}(x)$]{} ![Evolution of the probability density under the iterative application of the distillation protocol in the continuum limit (\[g36\]). Starting with a parabolic distribution $k=0$, we show the resulting profiles after steps $k=1,2,\ldots,6$.[]{data-label="evolution"}](evolution2.eps "fig:"){width="8"}
This is a closed analytical equation that provides us with the evolution of the probability density for any initial probability profile $q^{(0)}(x)$. In Fig. \[evolution\] we plot this evolution for an initial distribution of a parabolic form $q^{(0)}(x) = 6(x-x^2)$. We see how as we increase the step $k$ of the distillation, the new distributions get peaked around the highest value of the initial distribution, which is $x=\half$ in this particular case. This behaviour illustrates the idea of the weight amplification and is in agreement with the results (\[g16c\]) for the fixed points of the flow equations.
Quantum Distillation and Quantum Renormalization {#sec6}
================================================
[**Quantum Distillation**]{} [**Quantum RG**]{}
------------------------------ ---------------------------
Mixed State $\rho$ Quantum Hamiltonian $H$
Computational Basis Local Site Basis
Bell Basis Energy Basis
Alice & Bob Tensor Product Blocking Method
L.O.C.C. Truncation Operator
Maximum Fidelity Minimum Energy
RG-Flow Diagram Distillation-Flow Diagram
It is interesting to notice the analogy between the recursive distillation process represented by the equation (\[s6\]) and Fig. \[flowdiagram\] and the truncation process in the Renormalization Group analysis of certain quantum lattice Hamiltonian models, specifically, the ITF model (Ising in a Transverse Field) [@jaitisi],[@analytic]. The basic idea of a QRG method is: i/ elimination of high energy states plus, ii/ iterative process. This is precisely what happens in a quantum distillation process which we have seen in the preceeding sections, achieving a purification of a mixed state by means of discarding states and a recursive procedure. This relationship can be made even closer if we briefly recall what a quantum renormalization group (QRG) method is. The subject of the distillation is a mixed state operator $\rho$, while that of the renormalization is a quantum Hamiltonian operator $H$. A summary of these relations is presented in Table \[tableQDQRG\] that will be deduced along the way.
![Block decomposition of the Heisenberg chain in 3-site blocks.[]{data-label="chainbrg"}](chainbrg)
\[Bc\]\[Bc\]\[0.751\]\[0\][$F_0$]{} \[Bc\]\[Bc\]\[0.751\]\[0\][$F_1$]{} \[Bc\]\[Bc\]\[0.751\]\[0\][$F_2$]{} \[Bc\]\[Bc\]\[0.751\]\[0\][$F_3$]{} ![Example of distillation process starting with 8 pairs of mixed states for Alice (and the same amount for Bob). After 3 steps the original fidelity $F_0$ is improved up to a final value of $F_3$ (we assume full success for simplicity).[]{data-label="binarydistill"}](binarydistill "fig:"){width="9"}
\[Bc\]\[Bc\]\[1\]\[0\][Initial chain with 8 sites]{} \[Bc\]\[Bc\]\[1\]\[0\][ QRG]{} ![Example of renormalization QRG process for an initial chain with 8 sites in 3 steps (making blocks of 2 sites each).[]{data-label="binaryflow"}](binaryflow "fig:"){width="8"}
The easiest way to present the QRG method is with an example of quantum lattice Hamiltonian like the isotropic Heisenberg model on a 1D chain: $$H = J \sum_{i=0}^N \vec{S}_i\cdot \vec{S}_{i+1},
\label{blocking12}$$ with $\vec{S}_i$ spin-$\half$ operators at site $i$ of the chain. The local site basis $\{ \ket{\!\downarrow}, \ket{\!\uparrow} \}$ corresponds to the computational basis $\{\ket{0}, \ket{1}\}$. Much like this latter basis is not enough for doing the distillation, the local site basis needs to be complemented with another type of basis. To see this, let us start the RG process with the block decomposition of the chain in blocks of $n_B=3$ sites as shown in Fig. \[chainbrg\]. This blocking method in QRG corresponds to the tensor product of Alice and Bob’s shared states at the begining of the distillation process, as shown in Fig. \[binarydistill\]. This is to be compared with the similar iterative process in the QRG method in Fig. \[binaryflow\].
The block Hamiltonian is then $$\begin{split}
H_{B} &= J (\vec{S}_1\cdot \vec{S}_{2} + \vec{S}_2\cdot \vec{S}_{3}) \\
&=\frac{{ J}}{2} \left[ (\vec{S}_1 + \vec{S}_{2} + \vec{S}_{3})^2 -
\vec{S}_{2}^2 - (\vec{S}_1 + \vec{S}_{3})^2\right]
\end{split}
\label{blocking13}$$ The label $B$ here stands for Block and not for Bob. The diagonalization of $H_{B}$ is straightforward using the Clebsch-Gordan decomposition of the tensor product of 3 irreducible representations of spin $S=\half$, $$\half { \otimes} \half { \otimes} \half = \half { \oplus} \half { \oplus} \frac{3}{2}.
\label{blocking14}$$ In particular, the ground state (GS) is given by $$\ket{\!\Uparrow}_{\rm GS} = \frac{1}{\sqrt{6}}\left[ 2\ket{\uparrow \downarrow \uparrow}
- \ket{\downarrow \uparrow \uparrow} - \ket{\uparrow \uparrow \downarrow}
\right],
\label{blocking15}$$ which is a spin doublet (with a similar expresion for the other state $\ket{\!\Downarrow}_{\rm GS}$, with the spins reversed). This fact is peculiar of the 3-site block and it is the main underlying reason for using a block of that size in the QRG (this fact is model dependent: for the ITF model, the blocking is with $n_B=2$ sites [@jaitisi],[@analytic], Fig. \[binaryflow\]). In the energy basis, the block Hamiltonian is diagonal and this corresponds to the Bell basis for the mixed state $\rho$ in the distillation process.
Now, the truncation of states amounts to retaining the state of lowest energy (doublet) and discarding the remaining 2 excited states. This reduction scheme is of the form $2^3=8\longrightarrow 2$. This truncation corresponds to discarding unwanted states of non-coincidences in the distillation process. The new effective site is again a spin-$\half$ site as shown in Fig. \[chain3spins\].
![QRG renormalization of $H_B$.[]{data-label="chain3spins"}](chain3spins)
The RG-truncation is implemented by means of a truncation operator $O$ constructed from the lowest energy eigenvalues of $H_B$ retained during the renormalization process. In this example, $O$ is constructed from the lowest energy doublet in the Clebsch-Gordan decomposition (\[blocking14\]), namely (\[blocking15\]). Similarly, in the distillation process we have that the tensor product of Alice and Bob’s states can be decomposed into states with coincident qubits in the target, denoted by $\rho^{a}_{\rm C}$ (\[s3\]), and states with non-coincident qubits in the target, denoted by $\rho^{a}_{\rm NC}$ (\[s3\]), i.e., $$\rho_{AB}\otimes \rho_{AB} =
(\sum_{a} \rho^{a}_{\rm C})\oplus (\sum_{a} \rho^{a}_{\rm NC}),
\label{distill14}$$ where the sum in $a$ runs over a certain number of mixed states of 4 parties. Notice that this stage is similar to the RG-stage represented by equation (\[blocking14\]). Next, an elimination process similar to the RG-truncation is performed by means of LOCC operations (measurements and classical communication) that retains only the bipartite states embedded in the $\rho^{a}_{\rm C}$ states.
Then, the renormalization of the block Hamiltonian is simply $${ H_{B'}} = { O} { H_B} { O^{\dagger}} = E_0 = -{ J}.
\label{blocking16}$$ Similarly, we could have expressed the state-elimination of the distillation in previous sections in terms of an truncation operator, say $O_D$, such that the new mixed state $\rho'_{AB}$ is obtained as $${\rho'_{AB}} = { O_D} (\rho_{AB}\otimes \rho_{AB}) { O^{\dagger}_D}.
\label{distill16}$$
In the case of the quantum Hamiltonian, we still need extra work since there are interaction links between blocks (see Fig. \[chainbrg\]). These are absent in the distillation protocol. However, the renormalization of the interblock Hamiltonian $H_{BB}$ follows also the same prescription as in (\[blocking16\]) and we arrive at $${ J} \vec{S}_r^{n} \cdot \vec{S}_l^{n+1} \xrightarrow{\rm { RG}}
{ J} (\frac{2}{3})^2 \vec{S}'_{n} \cdot \vec{S}'_{n+1},
\label{blocking18}$$ where we denote by $n$ and $n+1$ two successive blocks in the original lattice (see Fig. \[chainrlblocks\]) that become two successive sites (see Fig. \[chain3spins\]) in the new lattice after the renormalization. We can collect all these steps in Table \[tableQRG\] [@white], [@jaitisi]. This table should be contrasted with the similar table for the distillation process that can be formed with the steps explained in Sect. \[sec2\].
1/ Block Decomposition: $H=H_B+H_{BB}$.
--------------------------------------------------------------------------------
2/ Diagonalization of $H_B$.
3/ Truncation within each Block: $O$.
4/ Renormalization: $H'_B= O H_B O^{\dagger}$, $H'_{BB}= O H_{BB} O^{\dagger}$
5/ Iteration: Go to 1/ with $H'=H'_B+H'_{BB}$.
: Steps of the quantum renormalization group method (QRG) for lattice Hamiltonians.[]{data-label="tableQRG"}
The outcome of the RG-method is that we obtain the correct RG-flow for the coupling constant $J\longrightarrow 0$, signalling a gapless system plus an approximate estimation for the ground state energy, which by means of the variational principle, it is an upper bound for the exact energy. Therefore, the QRG is an energy minimization procedure. Likewise, the purification process produces a protocol for fidelity maximization along with a distillation-flow diagram.
\[Bc\]\[Bc\]\[1\]\[0\][$\vec{S}_r$]{} \[Bc\]\[Bc\]\[1\]\[0\][$\vec{S}_l$]{} \[Bc\]\[Bc\]\[0.75\]\[0\][r]{} \[Bc\]\[Bc\]\[0.75\]\[0\][l]{} \[Bc\]\[Bc\]\[1\]\[0\][$n$]{} \[Bc\]\[Bc\]\[1\]\[0\][$n+1$]{} ![QRG renormalization of the interblock Hamiltonian $H_{BB}$.[]{data-label="chainrlblocks"}](chainrlblocks "fig:")
This completes the relationship established in Table \[tableQDQRG\] between quantum distillation and quantum renormalization.
Conclusions {#sec7}
===========
The field of quantum distillation protocols has become very active in the theory of quantum information due to the central role played by entanglement in the quantum communication procedures and its tendency to degradation.
In this work we have been interested in several extensions of the purification protocols when dealing with multilevel systems (qudits) instead of the more usual qubit protocols. We have seen the various advantages of having distillation methods for qudits systems as compared with the simple case of qubits. We have also obtained the general form of the solution to the distillation recursion relations and several particular solutions have been studied explicitly. We have developed the relationship between quantum distillation protocols and quantum renormalization group methods, something which is interesting in itself and could serve as a guide for possible extension of purification methods.
We would like to mention that the possibility of working with qudits systems has become quite realistic in the recent years. For instance, it is possible to realize multilevel systems in terms of the orbital angular momentum of photons, instead of the more standard polarization (qubit) degree of freedom [@zeilinger2], [@arnaut], [@barnett]. Yet another possibility is to use the so called multiport beam splitters [@reck], [@zukowski], [@molina-terriza], [@padgett].
There are several ways in which this work can be extended. One is the consideration of noise as as source of errors during the distillation protocol itself. Another one is to allow the possibility of having these distillation protocols for qudits be embedded into a quantum repeater protocol [@repeaters-pur], [@repeaters-com].
[*Acknowledgments*]{}. This work is partially supported by the DGES under contract BFM2000-1320-C02-01.
General Solution of the Distillation Recursion Relations {#app}
========================================================
In this appendix, we look for more general solutions to the general distillation recursion relations (\[g14\]) than those studied in section \[sec3\]. To this end, it is convenient to introduce auxiliary variables $g_{kj}^{(n)}$ defined by $$g_{kj}^{(n)} = \sum_{k'=0}^{D-1} g_{k\ominus k'j}^{(n-1)} g_{k'j}^{(n-1)},\
g_{kj}^{(0)} = q_{kj}^{(0)},
\label{ap1}$$ so that the real weights $q_{kj}^{(n)}$ are related to these auxiliary variables as $$q_{kj}^{(n)} := \frac{g_{kj}^{(n)}}{\sum_{l,i=0}^{D-1} g_{li}^{(n)}}.
\label{ap2}$$ Thus, $g_{kj}^{(n)}$ are unnormalized probability weights. The recursion relations they satisfy can be read as follows (\[ap1\]): for a fixed second index $j$, the unnormalized weights $g_{kj}^{(n)}$ at the step $n$ of the distillation process are obtained as the convolution over the first indices $k$ of the unnormalized weights $g_{kj}^{(n-1)}$ in an earlier step. This fact calls for the introduction of the Fourier transform in order to analyze the relations (\[ap1\]). Let us introduce the new variables $R_j^{(n)}$ defined as $$R_{\hat{k}j}^{(n)} := \sum_{k=0}^{D-1} \ee^{\frac{2\pi \ii \hat{k} k}{D}} g_{kj}^{(n)}.
\label{ap3}$$ Now, using the properties of the convolution and the Fourier transform it is inmediate to arrive at a simpler recursion relation $$R_{\hat{k}j}^{(n)} = \left[R_{\hat{k}j}^{(n-1)}\right]^2,
\label{ap4}$$ which can be iterated all the way down to the initial step $$R_{\hat{k}j}^{(n)} = \left[R_{\hat{k}j}^{(0)}\right]^{(2^n)}.
\label{ap5}$$ Fourier transforming back to the unnormalized variables, we get $$g_{kj}^{(n)} = \frac{1}{D}\sum_{\hat{k}=0}^{D-1}
\ee^{-\frac{2\pi \ii k \hat{k}}{D}}
\left[ \sum_{k'=0}^{D-1} \ee^{\frac{2\pi \ii \hat{k}k'}{D}} q_{k'j}^{(0)}
\right]^{(2^n)},
\label{ap6}$$ from which we also obtain the normalized probability weights $q_{kj}^{(n)}$ upon normalization (\[ap2\]). In particular, for the case of qubits treated in Sect. \[sec2\], $D=2$ and if we also restrict ourselves to weights of the form $q_{00}=F, q_{01}=1-F, q_{10}=q_{11}=0$, we again obtain from the general solution (\[ap6\]) the simple recursion relation in equation (\[s6\]).
M. Lewenstein, D. Bruss, J.I. Cirac, B. Kraus, M. Kus, J. Samsonowicz, A. Sanpera, R. Tarrach, “Separability and distillability in composite quantum systems -a primer-”, J. Mod. Opt. [**47**]{}, 2841 (2000).
D. Bruss, J.I. Cirac, P. Horodecki, F. Hulpke, B. Kraus, M. Lewenstein, A. Sanpera, “Reflections upon separability and distillability”, J. Mod. Opt. [**49**]{}, 1399-1418 (2002).
C.H. Bennett, G. Brassard, S. Popescu, B. Schumacher, J.A. Smolin, W.K. Wootters, “Purification of Noisy Entanglement and Faithful Teleportation via Noisy Channels”, Phys. Rev. Lett. [**76**]{}, 722-725 (1996).
J.W. Pan, C. Simon, C. Brukner, A. Zeilinger, “Feasible Entanglement Purification for Quantum Communication”, Nature [**410**]{}, 1067 (2001).
N. Gisin, “Hidden quantum nonlocality revealed by local filters”, Phys. Lett. [**A**]{}210, 151, (1996).
P.G. Kwiat, S. Barraza-Lopez, A. Stefanov, N. Gisin, “Experimental entanglement distillation and ’hidden’ non-locality," Nature 409, 1014 (2001).
P.W. Shor, “Scheme for reducing decoherence in quantum computer memory”, Phys. Rev. [**A**]{}52, 2493-2496, (1995).
A.M. Steane, “Error correcting codes in quantum theory”, Phys. Rev. Lett. [**77**]{}, 793, (1996).
D. Gottesman, “A theory of fault-tolerant quantum computation," Phys. Rev. [**A**]{} 57, 127-137 (1998).
J. Preskill, “Fault-tolerant quantum computation." quant-ph/9712048.
C.H. Bennett, D.P. DiVincenzo, J. Smolin, W.K. Wooters, “Mixed-state entanglement and quantum error correction”, Phys. Rev. A[**54**]{}, 3824 (1996).
D. Deutsch, A. Ekert, R. Jozsa, C. Macchiavello, S. Popescu, A. Sanpera, “Quantum Privacy Amplification and the Security of Quantum Cryptography over Noisy Channels”, Phys. Rev. Lett. [**77**]{}, 2818-2821 (1996).
C. Macchiavello, “On the analytical convergence of the QPA procedure”, Phys. Lett. [**A**]{}246, 385 (1998).
C. Macchiavello, “Quantum Privacy Amplification”, in D. Bouwmeester, A. Ekert, A. Zeilinger (eds.), [*The physics of quantum information*]{}, Springer-Verlag 2000.
W. Dür, H.-J. Briegel, J.I. Cirac, P. Zoller, “Quantum repeaters based on entanglement purification”, Phys. Rev. A. [**59**]{}, 169 (1999).
W. Dür, H.-J. Briegel, J.I. Cirac, P. Zoller, “Quantum repeaters for communication”, quant-ph/9803056.
M. Horodecki, P. Horodecki, “Reduction criterion of separability and limits for a class of distillation protocols”, Phys. Rev. [**A**]{}59, 4206-4216, (1999).
G. Alber, A. Delgado, N. Gisin, I. Jex, “Generalized quantum XOR-gate for quantum teleportation and state purificationin arbitrary dimensional Hilbert spaces”, quant-ph/000802.
G. Alber, A. Delgado, N. Gisin, I. Jex, “Efficient bipartite quantum state purification in arbitrary dimensional Hilbert spaces”, quant-ph/0102035.
H. Bechmann-Pasquinucci, A. Peres, “Quantum Cryptography with 3-State Systems”, Phys. Rev. Lett. [**85**]{}, 3313-3316, (2000).
H. Bechmann-Pasquinucci, W. Tittel, “Quantum cryptography using larger alphabets”, Phys. Rev. [**A**]{} 61, 062308 (2000).
M. Bourennane, A. Karlsson, G. Björk, “Quantum key distribution using multilevel encoding”, Phys. Rev. [**A**]{} 64, 012306 (2001).
C.H. Bennett, H.J. Berstein, S. Popescu, B. Schumacher, “Concentrating partial entanglement by local operations”, Phys. Rev. A[**53**]{}, 2046 (1996).
J. González, M.A. Martín-Delgado, G. Sierra, A.H. Vozmediano, [*Quantum Electron Liquids and High-$T_c$ Superconductivity*]{}, Lecture Notes in Physics, Monographs vol. [**38**]{}. Springer-Verlag 1995.
M.A. Martín-Delgado, G. Sierra; [*“Analytic Formulations of the Density Matrix Renormalization Group"*]{}; Int. J. Mod. Phys. [**A11**]{}, 3147-3174 (1996).
S.R. White, [*“Density-matrix algorithms for quantum renormalization groups”*]{} Phys. Rev. B[**48**]{}, 10345 (1993).
A. Mair, A. Vaziri, G. Weihs, A. Zeilinger, “Entanglement of Orbital Angular Momentum States of Photons”, Nature [**412**]{}, 3123-316 (2001).
H.H. Arnaut, G.A. Barbosa, “Orbital and Intrinsic Angular Momentum of Single Photons and Entangled Pairs of Photons Generated by Parametric Down-Conversion” Phys. Rev. Lett. [**85**]{}, 286 (2000).
S. Franke-Arnold, S.M. Barnett, “Two-photon entanglement of orbital angular momentum states”, Phys. Rev. [**A**]{}65, 033823 (2002).
M. Reck, A. Zeilinger, H.J. Bernstein, P. Bertani, “Experimental realization of any discrete unitary operator”, Phys. Rev. Lett. [**73**]{}, 58-61 (1994).
M. Zukowski, A. Zeilinger, M.A. Horne, “Realizable higher-dimensional two-particle entanglements via multiport beam splitters”, Phys. Rev. [**A 55**]{}, 2564-2579 (1997).
G. Molina-Terriza, J.P. Torres, L. Torner, “Management of the Angular Momentum of Light: Preparation of Photons in Multidimensional Vector States of Angular Momentum”, Phys. Rev. Lett. [**88**]{}, 013601 (2002).
J. Leach, M.J. Padgett, S.M. Barnett, S. Franke-Arnold, J. Courtial, “Measuring the Orbital Angular Momentum of a Single Photon”, Phys. Rev. Lett. [**88**]{}, 257901 (2002).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We determine the impact of resource renewal on the lifetime of a forager that depletes its environment and starves if it wanders too long without eating. In the framework of a minimal starving random walk model with resource renewal, there are three universal classes of behavior as a function of the renewal time. For sufficiently rapid renewal, foragers are immortal, while foragers have a finite lifetime otherwise. In the specific case of one dimension, there is a third regime, for sufficiently slow renewal, in which the lifetime of the forager is independent of the renewal time. We outline an enumeration method to determine the mean lifetime of the forager in the mortal regime.'
author:
- 'M. Chupeau'
- 'O. Bénichou'
- 'S. Redner'
title: Universality classes of foraging with resource renewal
---
Introduction
============
What is the impact of renewal of resources on the state of a forager? If the environment is harsh and resources regenerate slowly, foragers may be confronted by perpetual scarcity. Thus a forager may often go hungry or even starve. Conversely, in an abundant environment where resources are quickly replenished, a forager may never experience starvation risk. Our goal is to map out the states of a forager as a function of the renewal time within the framework of the minimal “starving random walk” model [@Benichou:2014] that we define below.
The random walk model has been frequently invoked to describe the motion of foraging animals [@Berg:1983; @Bartumeus:2005; @Codling:2008], as well as a wide variety of classic applications [@Weiss; @Hughes; @BenAvraham; @Okubo; @Bell:1970]. In the wild, foraging animals can die from many causes, such as diseases or old age [@Fey:2015]. The underlying dynamics from these causes of death can be described by a “mortal” random walk that dies according to a specified lifetime distribution [@Yuste:2006; @Yuste:2013; @Abad:2013]. This model also a variety of applications to diverse fields, such as the diffusion of light in human tissue [@Bonner:1987], and biologically-inspired search problems [@Meerson:2015]. In the context of foraging, an important contributor to forager mortality is the possibility that it is unsuccessful in its search for food within its habitat [@Sinclair:1985]. Thus the age at which a forager dies is also determined by its trajectory and the amount of available resources. This coupling between the lifetime of a living organism and its trajectory, along which environmental resources are depleted, defines a nontrivial class of random-walk problems [@PW97; @D99; @BW03; @ABV03; @Z04; @AR05]—including the starving random walk model—that are relatively unexplored.
![Starving random walk with probabilistic resource renewal in two dimensions: (a) initial state and later times (b) & (c). Each site initially contains food that is eaten when found by the forager. Food reappears on empty sites after a random renewal time (represented here as glowing circles). The forager starves if it wanders $\mathcal{S}$ steps without eating.[]{data-label="model"}](model){width="47.50000%"}
In the original starving random walk model [@Benichou:2014], a random walk irreversibly depletes its environment during its wanderings, and starves if it wanders too long in a resource-depleted region. Initially, each lattice site contains one food unit. Whenever a forager, which undergoes a random walk, arrives at a full site, the food is completely eaten. Once depleted, a site remains empty. Whenever the forager arrives at an empty site, it does not eat. If the forager goes $\mathcal{S}$ steps without eating, it starves. We can think of $\mathcal{S}$ as the metabolic capacity of the forager—the amount of time it can live without food before starving. Asymptotic expressions for the walker lifetime and the territory visited at the starvation time were given in one dimension [@Benichou:2014]. Estimates for these two quantities in dimensions $d\geqslant 3$ and a lower bound for the territory visited at starvation in $d=2$ were also given. These results provide a first step in quantifying the interplay between the trajectory of a forager and the consumption of food and their effect on the lifetime of the forager.
However, the natural resources being consumed—preys, plants, water, and nutrients—typically obey their own dynamics. In particular, consumed resources usually do not disappear permanently [@Evans:1972; @Lemaire:2001]. Instead, they typically regenerate a certain time after they have been depleted [@McCallum:1905; @Harrison:1980; @Fenner]. This basic fact motivates our study of starving random walks with the possibility of renewal of resources (Fig. \[model\]).
As we will discuss, resource renewal substantially modifies the properties of a starving random walk. Using extreme trajectory arguments, we will argue that the correlations induced by the coupling between the trajectory of a forager, its metabolic capacity $\mathcal{S}$, and the dynamics of the resources lead to three universal regimes of behavior that are determined by $\mathcal{S}$ and the bounds $\mathcal{R}_1$ and $\mathcal{R}_2$ of the support of the renewal time distribution, but are insensitive to the shape of this distribution (Fig. \[diag\_phase\]). We will first demonstrate the existence of a transition between an immortal regime, in which the forager cannot starve, no matter what its trajectory, and a mortal regime, where the forager must eventually starve. Both regimes arise in any spatial dimension. We will also show that a third regime arises in one dimension only, in which renewal is so slow that the forager lifetime is the same as in the case of no renewal. We also develop an enumeration method that yields, in principle, the exact value of the mean lifetime of a forager in the mortal regime.
We will first address the case of starving random walks with deterministic renewal in one dimension (Sec. II). In Sec. III, we extend our basic results to the case of probabilistic renewal and to higher dimensions. Some brief conclusions are given in Sec. IV.
Deterministic renewal in 1D
============================
Let us first investigate starving random walks with *deterministic* renewal in one dimension. Such a deterministic mechanism describes plants that grow at a fixed rate to reach an edible size a fixed time after having been previously defoliated [@Erickson:1976]. We posit that food that is eaten at time step $t$ reappears at time $t+\mathcal{R}$, with $\mathcal{R}$ an integer. In each time step, the time elapsed since an empty site was depleted is increased by one and food appears at any site where this time equals $\mathcal{R}$. As part of this same time step, the walker hops to one of its nearest neighbors. If the site contains food, which may have appeared just before the walker arrives, the food is eaten. We call the set of empty sites, which may or may not be connected, the “desert”.
Immortality
-----------
For a forager with metabolic capacity $\mathcal{S} \in \mathbb{N}$, we now determine the range of the renewal times for which immortality arises. A forager is immortal if it survives forever on any trajectory, and, in particular, on the most unfavorable trajectories. The set of such trajectories is infinite, but they all possess the common pattern (Fig. \[limit\]) that the walker depletes two sites in a row at one end of the desert, before wandering within the desert as long as possible until being certain, because of resource renewal, to land on a food-containing site (see Appendix A for details). It takes $\mathcal{R}$ time steps before food reappears on one of these two sites (Fig. \[limit\]). Thus, roughly speaking, when $\mathcal{R}<\mathcal{S}$, the forager necessarily survives. The precise criterion actually is $\mathcal{R}\leq \mathcal{R}^*$, with $$\label{bound}
\mathcal{R}^* =
\begin{cases}
\mathcal{S}&\quad \mathcal{S} ~ \mathrm{even}\,,\\
\mathcal{S}+1&\quad \mathcal{S} ~ \mathrm{odd}\,,
\end{cases}$$ due to the even-odd oscillations of a nearest-neighbor random walk (Appendix A).
![Illustration of the common pattern (inside the dashed rectangle) of all extremal trajectories for the case of renewal time $\mathcal{R}=9$. This pattern consists of depleting two consecutive sites at one end of the desert. The forager wanders inside this desert which gradually shortens until it reaches length 2. The pattern ends when the walker is certain to land on a food-containing site. Green dots denote food-containing sites and the numbers give the ages of each site.[]{data-label="limit"}](alternating){width="45.00000%"}
Mean Lifetime in the Mortal Regime
----------------------------------
On the other hand, when $\mathcal{R}> \mathcal{R}^*$, the forager is mortal and thus eventually starves with probability 1 because we can construct trajectories that necessarily lead to starvation. These trajectories again involve the forager first carving out a desert and then wandering strictly within this desert so that renewal does not reach the forager before it starves. In contrast to the situation in Fig. \[bound\], the walker starves before it reaches a site where the resource has been regenerated. By their very existence, such trajectories are achieved with non-zero probability. From classical results about Markov chains (see, e.g., [@Grinstead:2012]), every trajectory will eventually generate a configuration for which the forager starves. Thus there are two regimes of behavior—immortality and mortality. The boundaries between these regimes depend only on the metabolic capacity $\mathcal{S}$ of the forager and the renewal time $\mathcal{R}$.
Using the Markov chain formalism, we can, in principle, determine the mean lifetime of the forager by enumerating the configurations of the system as the forager wanders. A configuration is defined as the location and age of each empty site in the desert, the position of the forager in the desert, and the number of time steps elapsed since the forager last ate. Here the age of an empty site is the time since the food was last consumed at this site. Thus a newly empty site has age 0, while a site that will regenerate at the next step has age $\mathcal{R}-1$. Because the desert has a finite size, the number of configurations is finite. We can therefore write the transition matrix that describes the evolution of the system at each step of the forager and thereby extract its mean lifetime.
Let us illustrate this enumeration for the simple case of $\mathcal{S}=2$ and $\mathcal{R}=3$. For this example, there are five distinct configurations (Fig. \[diagram\_trans\]). State arises after the first step, and the evolution of the system from one state to another is shown in the figure. The associated transition matrix is $$\begin{pmatrix}
0 & 0 & 1 & 0 & 0 \\
1/2 & 0 & 0 & 1/2 & 0 \\
0 & 1/2 & 0 & 0 & 0 \\
1/2 & 0 & 0 & 1/2 & 0 \\
0 & 1/2 & 0 & 0 & 1
\end{pmatrix}
\equiv \begin{pmatrix}
Q & 0 \\
V & I
\end{pmatrix}$$ where the states are listed in order –, with $$Q=\begin{pmatrix}
0 & 0 & 1 & 0 \\
1/2 & 0 & 0 & 1/2 \\
0 & 1/2 & 0 & 0 \\
1/2 & 0 & 0 & 1/2
\end{pmatrix},$$ and $V=(0,\frac{1}{2},0,0)$. We define the matrix (see [@Grinstead:2012]) $$N\equiv (I- Q^+)^{-1} = \begin{pmatrix}
2 & 2 & 1 & 2 \\
1 & 2 & 1 & 1 \\
2 & 2 & 2 & 2 \\
1 & 2 & 1 & 3
\end{pmatrix},$$ where $Q^+$ is the transpose of $Q$. Each entry $N_{ij}$ in this matrix is the average time that a system, which ultimately reaches starvation, spends in configuration $j$ when it starts from configuration $i$ [@Grinstead:2012]. From this matrix we can extract the mean absorption time $t_i$ starting from the state $i$. These are given by $$\begin{pmatrix}
t_1 \\
t_2 \\
t_3 \\
t_4
\end{pmatrix} = N \begin{pmatrix}
1 \\
1 \\
1 \\
1
\end{pmatrix} = \begin{pmatrix}
7 \\
5 \\
8 \\
7
\end{pmatrix}.$$
![State space of the system for metabolic capacity $\mathcal{S}=2$ and renewal time $\mathcal{R}=3$. The circled numbers denote the distinct states and $T$ is the time elapsed since the forager last ate.[]{data-label="diagram_trans"}](diag_S2_R3){width="230pt"}
Thus the mean lifetime of the forager is $t_1+1=8$, because after the first step, the system is necessarily in the state $\textcircled{1}$. In principle, this method can be extended to higher dimensions and also to probabilistic renewal with a bounded support of renewal times. However, in practice, this approach quickly becomes intractable because the number of possible configurations becomes prohibitively large when both $\mathcal{R}$ and $\mathcal{S}$ are large. Nevertheless, this approach gives a well-defined prescription for computing the average time until the forager starves.
Renewal Independent Regime
--------------------------
In one dimension, the mortal regime class can be further divided in two sub-regimes: (a) forager lifetime dependent on $\mathcal{R}$, and (b) renewal independence—lifetime independent of $\mathcal{R}$. Clearly, as $\mathcal{R}$ increases, the forager lifetime decreases and approaches the no-renewal limiting value as $\mathcal{R}\to\infty$. Does this decrease stop when $\mathcal{R}$ reaches a finite critical value $\mathcal{R}^{\dagger}$, or does the decrease continue as $\mathcal{R}\to\infty$? To resolve this question, we determine if there is at least one trajectory for which the forager can return to a replenished site without starving. If there is such a trajectory, then renewal is relevant, as the lifetime of the forager depends on the renewal time.
For a forager to return to a site on which food is renewed requires: (i) living long enough for such a renewal to occur and (ii) staying sufficiently close to this site to reach it without starving. These two attributes are most easily satisfied for a site (which we take as the origin) at which food has just been eaten and is surrounded by full sites. We determine the largest value of $\mathcal{R}$ for which the forager can return to this origin after the food at this site has been renewed.
![Optimal trajectory for a forager to live the longest and still be able to return the origin O and consume the replenished resource on this site, for the case $\mathcal{S}=5$. (i) The forager eats a site on the right edge of the desert every $\mathcal{S}$ steps until carving the largest semi-desert that it can cross. (ii) The forager crosses the desert. (iii) Reflection of stage (i). (iv) The forager crosses the left semi-desert to reach the origin. Such an excursion lasts $\mathcal{R}^{\dagger}$ steps (Eq. ). []{data-label="optimal"}](regime3){width="160pt"}
To return to the origin without starving imposes the constraint that the forager does not stray more than $\mathcal{S}$ steps from the origin. Moreover, to maximize the time that the forager wanders, it should eat only when it really needs to, that is, every $\mathcal{S}$ steps. The trajectory on which the forager lives the longest while staying within $\mathcal{S}$ steps of the origin therefore consists of the following components (Fig. \[optimal\]): (i) The forager creates a semi-desert of length $\mathcal{S}$ on one side, say to the right, of the origin. During this creation of the semi-desert, the forager eats by moving to a previously unvisited site only every $\mathcal{S}$ steps. (ii) The forager successfully crosses this semi-desert, which is the longest possible for which a successful traversal is possible. (iii) The forager creates a mirror image semi-desert of length $\mathcal{S}$ to the left of the origin. (iv) The forager crosses this left semi-desert and fetches the regenerated food at the origin.
The duration of this excursion is roughly $2 \mathcal{S}^2$, as the forager has eaten $2 \mathcal{S}$ times when it returns to the origin. The enumeration of the above sequence of moves (Appendix B) gives the maximal renewal time $\mathcal{R}^{\dagger}$ with $$\label{class3}
\mathcal{R}^{\dagger}=
\begin{cases}
2 \mathcal{S}^2-3\mathcal{S}+4 \qquad \textrm{$\mathcal{S}$ even}, \\
2 \mathcal{S}^2-\mathcal{S}+1 \qquad~\; \textrm{$\mathcal{S}$ odd}.
\end{cases}$$ For $\mathcal{R}\leq \mathcal{R}^\dagger$, a forager has a non-zero probability to eat food at a site where renewal has occurred. Conversely, for $\mathcal{R}>\mathcal{R}^{\dagger}$, a forager cannot reach any replenished site, so that renewal has no impact on the forager lifetime. Hence the mean lifetime does not gradually converge to the limiting no-renewal value, but rather reaches this value for $\mathcal{R}=\mathcal{R}^{\dagger}+1$. Thus we infer that there exists a renewal-independent regime (Fig. \[diag\_phase\]) for the lifetime of the forager.
Extensions
==========
We now extend the starving random walk model with resource renewal to accommodate two ecologically realistic features: (i) probabilistic renewal, in which the resource is regenerated at a random time after depletion, rather than after a fixed time $\mathcal{R}$, and (ii) starving random walks with resource renewal in higher dimensions.
Probabilistic renewal
---------------------
Suppose that each empty site is replenished a time $\tau$ after the resource at that site has been consumed, with $\tau$ drawn from a continuous distribution with support $[\mathcal{R}_1,\mathcal{R}_2]\subset \mathbb{R}^+$. This means that for a given site renewal cannot happen before a time $\mathcal{R}_1$ and also that replenishment necessarily occurs within a time $\mathcal{R}_2$ after depletion. We make no assumption on the shape of this distribution. In particular, $\mathcal{R}_1$ can be zero and $\mathcal{R}_2$ can be infinite.
In the case of deterministic renewal (Sec. II), we saw that the random walker is immortal when it is certain to land on a food-containing site before starving, even on the most unfavorable trajectories. The criterion for immortality in this case is determined by $\mathcal{R} \leq \mathcal{R}^*$, with $\mathcal{R}^*$ given in Eq. . For probabilistic renewal, the walker is sure to land on a food-containing site before starving if every renewal time in the support of the renewal-time distribution is smaller than $\mathcal{R}^*$. Therefore, immortality is assured when $\mathcal{R}_2 \leq \mathcal{R}^*$. However, if the upper bound $\mathcal{R}_2$ is infinite, that is, if replenishment on some sites can take an arbitrarily long time, immortality cannot occur.
On the other hand, if $\mathcal{R}_2>\mathcal{R}^*$, there exist patterns of steps for the random walker that lead to starvation, as in the case of deterministic renewal. Hence the walker is mortal. Additionally, the enumeration method presented above for deterministic renewal can be implement in a similar manner for the case of probabilistic renewal. In the probabilistic case, however, food does not reappear at a fixed time after depletion but at a time that is drawn from the renewal time distribution. What this means practically is that the number of configurations in probabilistic renewal is larger than that for deterministic renewal. Moreover, if the support of the distribution of renewal times is unbounded, the enumeration approach fails because the number of configurations is infinite.
We also argued in Sec. II that there exists a second transition inside the mortal regime between a sub-regime in which dynamics of the renewal controls the lifetime of the walker and a sub-regime where the lifetime becomes independent of the renewal dynamics. We inferred the existence of this transition by constructing the extremal trajectory that demarcates this second transition (Fig. \[optimal\]). If renewal has not occurred at the origin when the walker reaches this site at the end of the pattern of steps of Fig. \[optimal\], then renewal has no impact on the trajectory. Thus, the renewal independent regime arises if every depleted site remains empty for at least $\mathcal{R}^{\dagger}$ steps. In the case of probabilistic renewal, this second transition occurs when [$\mathcal{R}_1 > \mathcal{R}^{\dagger}$]{}.
These results are summarized in the phase diagram of Fig. \[diag\_phase\]; this is our key result.
Higher dimensions
-----------------
We now turn to starving random walks on higher-dimensional lattices for the general situation of probabilistic renewal. The class of trajectories that are the least favorable for the survival of walker (see Fig. \[limit\]), still arises in higher dimensions. Hence the immortality criterion $\mathcal{R}_2 \leq \mathcal{R}^*$ derived in the previous subsection remains valid, independent of the spatial dimension. Moreover, in the mortal regime, the enumeration method still works and can be used to determine the mean lifetime of a random walker.
In contrast to the transition to immortality, a transition to a renewal-independent regime does not occur in higher dimensions. The unique feature of one dimension is that the walker *must* traverse the desert that was carved by its previous trajectory to reach replenished sites. In contrast, in higher dimensions, there always exist trajectories on which a forager can stay alive for an arbitrarily long time and still return to the replenished sites without starving, because it can avoid the desert instead of having to cross it. Thus the renewal time—no matter how long—always affects the lifetime of the forager in greater than one dimension. Thus in higher dimensions there is only an immortal regime and a mortal regime in which the lifetime is function of the renewal dynamics. The transition between these two regimes is determined only by the upper bound of the distribution of renewal times, and not by the shape of this distribution, or by the spatial dimension.
Summary and conclusion
======================
To summarize, the renewal of resources has a dramatic effect on starving random walks. There exist three regimes of behavior as a function of the renewal time $\mathcal{R}$: (i) an immortal regime where a forager can live forever, (ii) a mortal regime where the forager lifetime is finite and depends on $\mathcal{R}$, and (iii) a renewal-independent mortal regime where renewal does not affect the lifetime of a forager. The latter arises only in one dimension, in which the average forager lifetime equals the value obtained in the absence of any renewal. In contrast, regimes (i) and (ii) arise in any spatial dimension and are universal with respect to the distribution of renewal times. The transitions between these regimes depend only on the bounds of the support of the renewal-time distribution and not on its shape. Much of this new phenomenology is controlled by the *times between visits* to distinct sites in a random walk, an apparently unexplored feature of site visitation statistics of random walks. Finally, we outlined an enumeration method to determine the mean forager lifetime in the mortal regimes (ii) and (iii). Average values of other basic observables, such as the number of distinct sites visited at starvation, the number of sites in the desert, the time spent in a certain configuration, can also be extracted from this approach.
Immortality is the main new feature that arises as a result of resource renewal. If a wandering organism can survive without food longer than the time needed for resources to be replenished, its lifetime is no longer limited by starvation but rather by external constraints (such as predation, diseases, life expectancy of the species, etc.). We speculate that perhaps the metabolic capacity of a given species is determined by the characteristic time for renewal of resources.
This work represents a first step to provide insight of the impact of resource renewal on the fate of a forager that depletes its environment by consumption. While most of our qualitative analysis was specific to the case of one dimension, our approach applies for any spatial dimension and also to arbitrary renewal dynamics. A basic question that we have not fully addressed is the analytic determination of the mean lifetime of the walker in the mortal regime. This calculation is of particular importance in two dimensions where it should be directly applicable to the modeling of animal behavior. In addition to developing a more complete theory in two dimensions, the inclusion of additional realistic features to this model, such as sensory awareness of the forager, or interactions between several foragers, such as sharing resources, are needed to give a more realistic description of ecosystems.
We acknowledge NSF Grant No. DMR-1205797 (S. R.) and ERC starting Grant No. FPTOpt-277998 (O. B.) for partial support of this research.
[99]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [****, ()]{} @noop [**]{} (, ) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [**]{} (, ) @noop [**]{} (, ) @noop [**]{}, Vol. (, ) @noop [**]{}, (, ) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} M. Perman and W. Werner, Probab. Theory Relat. Fields [ **108**]{}, 357 (1997).
B. Davis, Probab. Theory Relat. Fields [**113**]{}, 501 (1999).
I. Benjamini and D. B. Wilson, Electr. Commun. Probab. [ **8**]{}, 86 (2003).
O. Angel O, I. Benjamini, and B. Virag, Electr. Commun. Probab. [**8**]{}, 66 (2003).
M. P. W. Zerner, Probab. Theory Relat. Fields [**133**]{}, 98 (2005).
T. Antal and S. Redner, J. Phys. A [**38**]{}, 2555 (2005).
@noop [****, ()]{} @noop [**]{}, Vol. (, ) in @noop [**]{}, Vol. (, ) pp. @noop [****, ()]{} in @noop [**]{} (, ) pp. @noop [**]{} (, ) @noop [****, ()]{} @noop [**]{} (, )
The bound $\mathcal{R}^*$
=========================
We provide the details for the determination of Eq. , in the case of deterministic renewal. This bound for $\mathcal{R}^*$ depends on the parity of the metabolic capacity $\mathcal{S}$ because of the even-odd oscillations of the nearest-neighbor random walk. As described in the main text, the most unfavorable trajectories—on which the walker remains the longest without eating—possess the common pattern of eating two consecutive sites at one edge of the desert and then wandering as long as possible inside the desert, without depleting any additional site. This last feature implies that the desert gradually shortens as renewal happens, finally confining these most unfavorable trajectories to a two-site desert, made of the first two depleted sites of the pattern (see Fig. \[pattern\]).
![More detail of the common pattern (inside the dashed rectangle) of all extremal trajectories for renewal time $\mathcal{R}=9$. This pattern starts by depleting two consecutive sites at one end of the desert (shaded in blue). The forager then wanders inside this desert which gradually shortens (shaded in red) until it reaches length 2 (shaded in yellow). The pattern ends when the walker is sure to land on a food-containing site. Green dots denote food-containing sites and the numbers give the ages of each site.[]{data-label="pattern"}](alternatingSI.pdf){width="240pt"}
Depending on the value of $\mathcal{S}$, the walker either survives long enough to eat when renewal happens on these two sites (in the immortal regime), or starves (in the mortal regime). We determine the maximal value of the renewal time $\mathcal{R}^*$ that corresponds to the immortal regime for an example of most unfavorable trajectory for two consecutive values of $\mathcal{S}$ (Fig. \[threshold1\]). The walker is sure to eat before starving when the renewal time is $\mathcal{R}=\mathcal{R}^*$, even on this unfavorable trajectory (left column of Fig. \[threshold1\]), but can die if $\mathcal{R}=\mathcal{R}^*+1$ (right column of Fig. \[threshold1\]). For this example, which can be generalized to every value of $\mathcal{S}$, we see that $$\label{bound1}
\mathcal{R}^* =
\begin{cases}
\mathcal{S}&\qquad \mathcal{S} ~ \mathrm{even}\,,\\
\mathcal{S}+1&\qquad \mathcal{S} ~ \mathrm{odd}\,.
\end{cases}$$
![Destiny of the walker on the most unfavorable trajectories, for: (left column) the maximal renewal time $\mathcal{R}^*$ that leads to immortality, and (right column) for the minimal value of $\mathcal{R}$ that leads to mortality. The cases of odd and even metabolic capacities $\mathcal{S}$ are shown in (a) and (b), respectively. For both even and odd $\mathcal{S}$, the maximal value of $\mathcal{R}$ that yields immortality is the smallest even integer equal to or greater than $\mathcal{S}$. Here $T$ denotes the number of steps since the last meal. The dots represent food and the cross indicates the walker. Empty sites are labeled by the time elapsed since the food on these sites was eaten.[]{data-label="threshold1"}](limit.pdf){width="250pt"}
The bound $\mathcal{R}^{\dagger}$
=================================
We now give the details to derive Eq. (6) for the case of deterministic renewal. Note that the bound for $\mathcal{R}^{\dagger}$ also depends on the parity of the metabolic capacity $\mathcal{S}$ because of the even-odd oscillations of the nearest-neighbor random walk. If renewal is sufficiently quick, there exist trajectories for which the walker can return to a replenished site, in particular the origin of the walk, before starving. On the other hand, if renewal is too slow, then the walker either dies before this renewal happens, or carves a desert that is too large to be crossed without starving.
As mentioned in the main text, the maximal value $\mathcal{R}^{\dagger}$ of the renewal time for which the walker has a chance to return to the origin after the resource on this site has been renewed requires that: (i) the walker lives long enough for this renewal to occur, and (ii) the walker must stay sufficiently close to the origin to be able to reach it without starving. The walker can satisfy these two constraints by eating approximately every $\mathcal{S}$ steps and by staying within a segment of size $2 \mathcal{S}$ centered on the origin.
The trajectories on which the walker lives the longest while remaining within $\mathcal{S}$ steps of the origin consist of the following (Fig. \[limit\]): (i) The walker creates a desert of $\mathcal{S}$ sites on one side of the origin, say the right. During this creation of the desert, the walker waits as long as possible between each meal, that is to say $\mathcal{S}$ steps if $\mathcal{S}$ is odd, or $\mathcal{S}-1$ steps if $\mathcal{S}$ is even. Indeed, starting from the right edge of the desert, the walker needs an even number of steps to come back to this edge; thus an odd number of steps is required to eat (and deplete) the resource at the next site on the right side. (ii) The walker crosses the desert and reaches the site to the left of the origin after $\mathcal{S}$ steps. (iii) The walker creates a mirror image desert on the left side of the origin, by depleting $\mathcal{S}-1$ new sites. (iv) The walker crosses the left desert and reaches the origin after $\mathcal{S}$ steps. For the maximal value $\mathcal{R}^{\dagger}$ of the renewal time, the walker finally reaches the origin at the end of stage (iv) at the time step where the origin regenerates.
We now determine $\mathcal{R}^{\dagger}$ by counting the number of steps on this trajectory. The walker eats the site on the right of the origin at time step 1, and then takes a time $\mathcal{S}(\mathcal{S}-2)$ if $\mathcal{S}$ is odd, and $(\mathcal{S}-1)(\mathcal{S}-2)$ if $\mathcal{S}$ is even to complete the phase (i). Similarly, the phase (iii) lasts $\mathcal{S}(\mathcal{S}-1)$ if $\mathcal{S}$ is odd and $(\mathcal{S}-1)^2$ if $\mathcal{S}$ is even. Moreover, the phases (ii) and (iv) both last $\mathcal{S}$ steps independent of the parity of $\mathcal{S}$. Assembling these results yields the critical value: $$\mathcal{R}^{\dagger}=
\begin{cases}
2 \mathcal{S}^2-3\mathcal{S}+4 \qquad \textrm{$\mathcal{S}$ even}, \\
2 \mathcal{S}^2-\mathcal{S}+1 \qquad~\; \textrm{$\mathcal{S}$ odd}.
\end{cases}$$
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The Shahshahani geometry of evolutionary game theory is realized as the information geometry of the simplex, deriving from the Fisher information metric of the manifold of categorical probability distributions. Some essential concepts in evolutionary game theory are realized information-theoretically. Results are extended to the Lotka-Volterra equation and to multiple population systems.'
address: University of California Los Angeles
author:
- Marc Harper
bibliography:
- 'ref.bib'
title: Information Geometry and Evolutionary Game Theory
---
Introduction
============
The replicator equation is a widely-used model of natural selection. This paper explains the realization of the geometry of evolutionary game theory in terms of information theoretical principles, giving a purely mathematical and statistical origination of the replicator equation. Under this interpretation, the replicator equation models the information dynamics of a population of replicating entities. Additionally, the Kullback-Liebler information divergence, which serves as a Lyapunov function for the replicator dynamic, can be interpreted as a measure potential information, characterizing the concept of evolutionary stable state informatically.
Continuous Replicator Dynamic
-----------------------------
Consider a categorical distribution $X$ on $n$ categories of entities in a population. This is discrete probability distribution represented by a unit vector of $n$ variables $x = (x_1, \ldots, x_n)$ under the normalization $|x| = x_1 + \cdots + x_n = 1$, where $x_i$ denotes the proportion of the $i$-th type in the population. The replicator equation on this distribution is the differential equation $$\dot{x}_i = x_i\left(f_i(x) - \bar{f}(x)\right),$$ where $f(x) = (f_1(x), \ldots, f_n(x))$ is a fitness landscape and $\bar{f}(x)= x_1 f_1(x) + \cdots + x_n f_n(x)$ is the mean fitness.
Geometric Aspects of Evolutionary Game Theory
=============================================
The information theoretic interpretations of the previous chapter have a unified basis in information geometry. We begin with a description of the geometry of the simplex and geometric results known in evolutionary game theory.
The Geometry of the Simplex
---------------------------
Let $S^n$ be the interior of the $n$-simplex $\Delta^n$, which is $(n-1)$-dimensional. Each point $x$ of the simplex has the property that $x_1 + \cdots + x_n = 1$, so the tangent space at any point on the interior is the $(n-1)$-dimensional vector space described by $n$ vectors $v_1, \ldots, v_n$ such that $v_1 + \cdots + v_n = 0$. The orthogonal complement of the tangent space is the one dimensional line with direction vector $\mathbf{1} = (1, 1, \ldots, 1)$. Indeed $\mathbf{1} \cdot v = 0$ for any $v$ in the tangent space and the complement is necessarily one-dimensional. The faces of $\Delta^n$ are isomorphic to a simplex of one lower dimension, which can be seen by setting one of the $x_i$ to zero and indicates the absence of that type in the population. The replicator equation is *forward-invariant* on the simplex (and hence each of its faces), since if $x_i = 0$ then $\dot{x}_i = 0$. Because of this property, the replicator equation is called *non-innovative* since new types cannot arise, in contrast to evolutionary dynamics in which this is possible (notably the replicator-mutator equation [@Nowak06] and the orthogonal projection dynamic [@Sandholm09]).
Shahshahani Geometry
--------------------
Shahshahani introduced two Riemannian manifolds into mathematical biology[@Shahshahani79]: the positive orthant of ${\mathbb{R}}^n$, denoted ${\mathbb{R}}^n_+$, with the metric $$g_{ij}(x) = \frac{|x|}{x_i}\delta_{ij},$$ where $|x| = \sum_{i}{x_i}$ and the restriction to the simplex $\Delta^n = \{ x \in {\mathbb{R}}^n \, | \, |x| = 1, x_i \geq 0 \, \forall i \} \subset {\mathbb{R}}^n_+$, with the metric $$g_{ij}(x) = \frac{1}{x_i}\delta_{ij}.$$ Call the latter manifold the Shahshahani manifold; its metric is known as the Shahshahani metric. There is a normalization map $N: {\mathbb{R}}^n_{+} \to \Delta^n$ given by $x \mapsto \frac{x}{|x|}$. For each $\tau \in {\mathbb{R}}_{+}$, there is a map $\varphi_{\tau}$ mapping the simplex into ${\mathbb{R}}^n_{+}$ by $x \mapsto \tau x$. These maps are sections of the normalization map since $N \circ \varphi_{\tau} = \text{id}_{\Delta^n}$. The Shahshahani metric diverges on the boundary of the simplex so the metric is valid only on the interior $S^n$. Dynamics that are forward-invariant, such as the replicator dynamic, are not affected by the discontinuity at the boundary.
The Replicator Dynamic, Geometrically
-------------------------------------
The geometry of the Shahshahani manifold yields an elegant interpretation of the replicator equation: it is the gradient flow of the Shahshahani metric. Shahshahani proved the result for a special case of the replicator equation; the following more general theorem comes from [@Hofbauer98].
If the differential equation $\dot{x_i} = f_i(x)$ is a Euclidean gradient with $f_i = \frac{\partial V}{\partial x_i}$ then the replicator equation $x_i = {\hat{f}}_i(x) = x_i(f_i(x) - \bar{f}(x))$ is a gradient with respect to the Shahshahani metric.
In the case that the fitness landscape is a Euclidean gradient, the Shahshahani gradient gives a Lyapunov function for the dynamic. The classical case is that of a symmetric matrix $A$ and fitness landscape $f(x) = Ax$, where $A$ is the matrix of Malthusian fitness parameters given by the difference in birth rates and death rates $a_{ij} = b_{ij} - d_{ij}$ of an individual having alleles $i$ and $j$, where the alleles are of a single gene locus. In this case the Shahshahani potential is the mean fitness $\frac{1}{2} x \cdot f(x) = \frac{1}{2} x \cdot Ax$, with $Ax$ the Euclidean gradient[@Shahshahani79; @Hofbauer98].
Fisher’s Fundamental Theorem and Kimura’s Maximal Principle
-----------------------------------------------------------
Fisher’s fundamental theorem is a consequence of the geometric approach.
\[fft\_egt\] The rate of change of the Shahshahani potential is equal to the variance of the fitness landscape [@Hofbauer98]. $$\dot{V}(x) = \text{Var}_{x}[f(x)].$$
This is a general version of Fisher’s Fundamental Theorem of Natural Selection, specializing to the traditional result in the case of a symmetric and linear fitness landscape[@Hofbauer98]. Kimura’s maximal principle follows from the fact that the replicator equation is a gradient flow[@Shahshahani79]. As these are both important results in mathematical biology emerging from the geometry, an interpretation is desired of the Shahshahani metric that provides intuition for the introduced geometry on the simplex in the context of modeling natural selection.
The Information Geometry of Natural Selection
=============================================
An intuitive interpretation of the Shahshahani geometry is provided by information theory. Information geometry[@Amari93] studies manifolds of probability distributions $p(s, x)$ on a set $S$ depending on parameters $x$, which are the coordinates of the manifold. The manifold is endowed with the Fisher information metric, $$g_{ij}(x) = \mathbb{E}\left[ \frac{\partial \log p}{\partial x^i} \frac{\partial \log p}{\partial x^j} \right]$$ which can be shown to be the unique (up to a constant) metric respecting sufficient statistics [@Chentsov].
The Fisher Information Metric is the Shahshahani Metric
-------------------------------------------------------
The manifold of immediate interest is $P(X)$, the set of categorical probability distributions on a finite set $X$, with the Fisher information metric. In this case, it is convenient to abuse notation by allowing the parameters $x$ and distribution variables $s$ to have the same symbolic representation. [^1] There is a natural mapping $\varphi: P(X) \to \Delta^n$, where $|X| = n$, given by $p \to (p(1), p(2), \ldots, p(n)) = (x_1, \ldots, x_n)$. This maps $P(X)$ isometrically onto the simplex, and an easy computation shows that that the Fisher information metric is induced by the Shahshahani metric under this mapping. Simply observe that $$g_{ij}(x) = \mathbb{E}\left[ \frac{\partial \log x}{\partial x^i} \frac{\partial \log x}{\partial x^j} \right]
= \sum_{k}{x_k \frac{1}{x_i}\delta_{ik} \frac{1}{x_j}\delta_{jk}} = \frac{1}{x_i} \delta_{ij}$$ This result was recognized in [@Amari95] and [@Nihat05].
Fisher’s Fundamental Theorem
----------------------------
Fisher’s fundamental theorem is built into the geometry of $P(X)$[@Amari93]. Define the maps $E[g]$ on $P(X)$ by $p \mapsto E_{p}[g]$, where $g$ is from the set of functions ${{\mathbb{R}}}^X = \{ g : X \to {\mathbb{R}}\}$ and $E_{p}$ is the mean taken at the distribution $p$. Similarly, let $V_p[g]$ denote the variance of the function $g$ at $p$.
For any $g: X \to {\mathbb{R}}\in {{\mathbb{R}}}^X$, $$V_{p} = ||(\text{d}E[g])_p||_{p}^{2} = || (\nabla E[g])_p||_{p}^{2} ,$$ where the norm is induced by the Fisher information metric, for all $p \in P(X)$.
Information Divergences and Metrics on $P(X)$
---------------------------------------------
Some Riemannian metrics on $P(X)$ can be derived from information divergences[@Amari93]. Information geometry defines an *information divergence* as a smooth function $D(\cdot || \cdot): P(X) \times P(X)$ such that $D(x||y) \geq 0$ with equality iff $x = y$. The second order Taylor expansion in either variable evaluated along the diagonal $x=y$ begins with the Hessian term $H$. Indeed, $$\begin{aligned}
D(x||y) &= D(x || y)|_{x=y} + (\nabla D(x||y)|_{x=y})\cdot(x-y) + \frac{1}{2}\cdot(x-y)^{T}H(x)\cdot(x-y) + \cdots\\
&= 0 + 0 + \frac{1}{2}\cdot(x-y)^{T}H(x)\cdot(x-y) + \cdots\\\end{aligned}$$ because the gradient is parallel to $\mathbf{1}$ and $\mathbf{1} \cdot (x-y) = \mathbf{1} \cdot x - \mathbf{1} \cdot y = 1 -1 = 0$.
In the case that the Hessian is positive definite, it can be used to define a metric, $$g_{ij}^{(D)} = \left(\frac{\partial^2 D}{\partial x_i \partial y_j} \right)|_{x=y}.$$
A metric then defines a gradient flow, hence a global information divergence yields a dynamical system on the simplex. Importantly, the Hessian of the Kullback-Liebler divergence (in either variable, evaluated on the diagonal) is the Fisher information matrix, yielding the local to global connection of these two measures of information.
The Kullback-Liebler divergence localizes to the Fisher information metric. In coordinates we obtain the Shahshahani metric since $$\frac{\partial^2}{\partial x_i \partial y_j}{D_{KL}(x || y)}|_{x=y} = \frac{1}{x_i} \delta_{ij}.$$ Hence the induced gradient flow is the replicator equation. This allows the interpretation of Fisher’s Fundamental theorem and Kimura’s maximal principle in terms of Fisher information: natural selection forms a gradient with respect to an informatic measure, and hence locally has the direction of maximal information increase. The rate of change of the mean fitness of the population is given by the informatic variance.
Kullback-Liebler Divergence is a Lyapunov function for the Replicator Dynamic
-----------------------------------------------------------------------------
The following theorem shows that the Kullback-Liebler information divergence forms a Lyapunov function for the replicator dynamic, given an evolutionarily stable state. In fact, evolutionary stability is characterized by this property. A version of this theorem was proved in [@Akin79] and in [@Akin90].
\[ess\_Lyapunov\] The state ${\hat{x}}$ is an interior ESS for the replicator dynamic if and only if $D_{KL}({\hat{x}} || x)$ is a local Lyapunov function.
Let $V(x) = D_{KL}({\hat{x}} || x) = \sum_{i}{{\hat{x}}_i \log{{\hat{x}}_i}} - \sum_{i}{{\hat{x}}_i \log{x_i}}$. Then we have that $$\begin{aligned}
\dot{V}(x) &= -\sum_{i}{ {\hat{x}}_i \frac{\dot{x}_i}{x_i}} = -\sum_{i}{ {\hat{x}}_i (f_i(x) - \bar{f}(x)) }\\
&= -\sum_{i}{ {\hat{x}}_i f_i(x)} + \sum_{i}{ {\hat{x}}_i\bar{f}(x)} = -\sum_{i}{ {\hat{x}}_i f_i(x)} + \left(\sum_{i}{ {\hat{x}}_i}\right)\bar{f}(x)\\
&= -\sum_{i}{ {\hat{x}}_i f_i(x)} + \bar{f}(x) = -({\hat{x}} \cdot f(x) - x \cdot f(x)) < 0.\end{aligned}$$ The last inequality holds if and only if ${\hat{x}}$ is an ESS. Finally, by Jensen’s inequality, $D_{KL}$ is minimized when $x = {\hat{x}}$, so it is a local Lyapunov function.
A similar result is proven in [@Hofbauer98], with the Lyapunov function $V(x) = \prod_{i}{x_{i}^{ \hat{x}_i}}$, but the informatic origin is not apparent in this form, although the quantity $V$ can be interpreted as the probability of finding a categorical distribution on $x$ in the state $\hat{x}$. The quantity $D_{KL}({\hat{x}} || x)$ can be described as the *potential information* of the replicator system. The above result can then be interpreted information theoretically – natural selection acts to minimize the potential information.
Theorem \[ess\_Lyapunov\] holds for a class of ecological dynamics. A dynamic of the form $\dot{x}_i = x_i g_i(x), \, i=1,\ldots ,n$ (an *ecological dynamic*) is called *aggregate monotone* with respect to a fitness landscape $f$ if $g = (g_1, \ldots, g_n)$ has the property that $y \cdot f(x) > z \cdot f(x)$ if and only if $y \cdot g(x) > z \cdot g(x)$, for all distributions $x,y,z$. An aggregate monotone dynamic is the replicator dynamic up to a change in velocity [@Ritzberger95]. In particular, the replicator equation with a convex function applied to the fitness landscape is aggregate monotone. Consider the following extension of Theorem \[ess\_Lyapunov\].
\[ess\_Lyapunov\_aggregate\] For an aggregate monotone ecological dynamic $\dot{x}_i = x_i g_i(x)$, $D_{KL}({\hat{x}} || x)$ is a Lyapunov function for the dynamic if ${\hat{x}}$ is an interior ESS.
Let $V(x) = D_{KL}({\hat{x}} || x) = \sum_{i}{{\hat{x}}_i \log{{\hat{x}}_i}} - \sum_{i}{{\hat{x}}_i \log{x_i}}$. Note that since $\dot{x}_i = x_i g_i(x)$ is a dynamic on the simplex, $0 = \sum_{i}{x_i g_i(x)} = x \cdot g(x)$. Then we have that $$\begin{aligned}
\dot{V}(x) &= -\sum_{i}{ {\hat{x}}_i \frac{\dot{x}_i}{x_i}} = -\sum_{i}{ {\hat{x}}_i g_i(x) }\\
&= -{\hat{x}} \cdot g(x) = -({\hat{x}} \cdot g(x) - x \cdot g(x))\end{aligned}$$ Applying aggregate monotonicity to the last equality completes the proof.
Since a change of velocity does not alter the orbits of the dynamic, Theorem \[ess\_Lyapunov\_aggregate\] shows that the replicator equation is essentially the only aggregate monotone ecological dynamic in which evolutionary stability corresponds to minimizing the Kullback-Liebler divergence. For exactly which class of evolutionary dynamics this property holds for is an open question. From the proof it is clear that the assumption of aggregate monotonicity is too strong for a full characterization since it is only needed that ${\hat{x}} \cdot g(x) - x \cdot g(x) > 0$ if ${\hat{x}} \cdot f(x) - x \cdot f(x) > 0$, which quantifies over two distributions rather than three.
Exponential Families as Solutions of the Replicator Equation
------------------------------------------------------------
The exponential map on the Shahshahani manifold is $$exp(x,v) = \sum_{i}{ \frac{x_i e^{v_i}}{ \sum_{j}{x_j e^{v_j}} } \hat{e_i} },$$ where $\hat{e_i}$ is the $i$-th coordinate vector [@Nihat05]. The exponential map reduces to the exponential family at the barycenter $b = (\frac{1}{n}, \ldots, \frac{1}{n})$, $$exp(b,v) = \sum_{i}{ \frac{\frac{1}{n}e^{v_i}}{ \sum_{j}{\frac{1}{n} e^{v_j}} } \hat{e_i} } = \frac{1}{\sum_{j}{ e^{v_j}} }(e^{v_1}, \ldots, e^{v_n}).$$
The solutions of the replicator equation can be realized as exponential families [@Karev09; @Nihat05; @Akin82]. Let $x_i = \exp (v_i - G)$ with $\dot{v_i} = f_i(x)$ and $G(x)$ a normalization constant to ensure that the distribution sums to one. From the fact that $\sum_{i}{x_i} = 1$, $0 = \sum_{i}{\dot{x_i}}$ and so $$\begin{aligned}
0 = \sum_{i}{\dot{x}_i} &= \sum_{i}{\exp (v_i(x) - G(x)) (\dot{v}_i(x) - \dot{G}(x))}\\
&= \sum_{i}{x_i (\dot{v}_i(x) - \dot{G}(x))} = \sum_{i}{(x_i f_i(x))} - \dot{G}(x)\\
&= \bar{f}(x) - \dot{G}(x)\end{aligned}$$ Hence $\dot{G} = \bar{f}(x)$. Now $x_i$ satisfies $$\dot{x_i} = \exp (v_i(x) - G(x)) (\dot{v_i}(x) - \dot{G}(x)) = x_i (f_i(x) - \bar{f}(x)),$$ which is the replicator equation. In the case of a log-linear fitness landscape, explicit solutions can be derived [@Nihat05]. In this case, the equation for the variable $v$ can be reduced to a linear differential equation, which can be solved with eigenvalue methods.
Denormalization
---------------
Information geometry defines the *denormalized manifold* $\tilde{P}(X) = \{ \tau p | \tau \in {\mathbb{R}}^+, p \in P(X)\}$, which can be thought of as non-normalized discrete probability distributions. As with $P(X)$, $\tilde{P}(X)$ has an information metric. The denormalized manifold embeds into the reals as ${\mathbb{R}}^{n}_{+}$, with the denormalized information metric induced by the the metric given by Shahshahani, where the mapping back onto $P(X)$ realizes $\tau$ as $|x|$. In coordinates, the metric is given by $$\tilde{g}_{ij}(x) = \tau g_{ij}(x) = \frac{\tau}{x_i}\delta_{ij},$$ Akin uses the metric $$g_{ij}(x) = \frac{1}{x_i}\delta_{ij}$$ on ${\mathbb{R}}^n_{+}$ rather than the metric $$g_{ij}(x) = \frac{|x|}{x_i}\delta_{ij}$$ given by Shahshahani[@Shahshahani79; @Akin82]. Both metrics restrict to the same metric given by Shahshahani on the simplex. From the point of view of information geometry, the metric given by Shahshahani is the natural choice. The choice affects the form of the gradient on ${\mathbb{R}}^n_{+}$, which is in the case of Akin’s metric is the Lotka-Volterra predator-prey equation.
The Lotka-Volterra Equations and the Replicator Equation
--------------------------------------------------------
The Lotka-Volterra equations $$\label{lotka_volterra}
\dot{x}_i = x_i f_i(x)$$ descend from ${\mathbb{R}}^{n}_{+}$, through a normalization map onto the simplex, to a replicator equation with an altered landscape. To see this, let $|x| = x_1 + \cdots + x_n$, $\dot{x}_i = x_i f_i(x)$ and $y_i = \frac{x_i}{|x|}$. Rearrange to $|x|y_i = x_i$ and note that $\frac{d}{dt}{|x|} = \sum_{i}{\dot{x}_i} = \sum_{i}{x_i f_i(x)} = x \cdot f(x)$. By the product rule, $\frac{d}{dt}{|x|}y_i + |x|\dot{y}_i = \dot{x}_i$ and so $$\begin{aligned}
\frac{d}{dt}{y_i} &= \frac{\dot{x}_i - \frac{d}{dt}{|x|}\dot{y}_i }{|x|}\\
&= \frac{x_i f_i(x) - x \cdot f(x) y_i}{|x|}\\
&= y_i ( f_i(x) - y \cdot f(x) )\\
&= y_i ( g_i(y) - y \cdot g(y) ),\end{aligned}$$ where $g_i(y) = f_i(x)$ is an alteration of the fitness landscape.
The Lotka-Volterra equations are the gradient flow with respect to the metric given by Akin on ${\mathbb{R}}^n_{+}$[@Hofbauer98]. The gradient of the metric given by Shahshahani differs by a factor of $|x|$: $$\label{shifted_lotka_volterra}
\dot{x}_i = \frac{x_i}{|x|} f_i(x),$$
This system is transformable to Equation \[lotka\_volterra\] after a change of velocity eliminating the scalar function $B(x) = \frac{1}{|x|}$ because $B$ is strictly positive on ${\mathbb{R}}^{n}_{+}$. Equation \[shifted\_lotka\_volterra\] transforms to a replicator equation via the normalization map [@Shahshahani79].
The Lotka-Volterra equations can be interpreted as a gradient of the denormalized Fisher information metric, in the case that $f$ is an Euclidean metric, in analogy to the replicator equation. This allows denormalized analogues of earlier results, such as the following, which is true for the denormalized version of the Lotka-Volterra equation.
\[lv\_lyapunov\] Let ${\hat{x}}$ in ${\mathbb{R}}^n_{+}$ be such that $$\frac{{\hat{x}} \cdot f(x)}{|{\hat{x}}|} > \frac{x \cdot f(x)}{|x|}$$ in some neighborhood of ${\hat{x}}$ (a denormalized ESS). Suppose that the trajectory of $\dot{x}_i = \frac{x_i}{|x|} f_i(x)$ lies in a set that contains no point parallel to ${\hat{x}}$. Then the denormalized Kullback-Liebler divergence $$D_{KL}\left(\frac{{\hat{x}}}{|{\hat{x}}|} || \frac{x}{|x|} \right)$$ is a local Lyapunov function for Equation \[shifted\_lotka\_volterra\].
The divergence is minimal (and equal to zero) when $x = c {\hat{x}}$ for some constant $c$. Hence if the line through the origin and the point ${\hat{x}}$ intersects the the trajectory at most once, the divergence is zero if and only if ${\hat{x}} = x$. The time derivative is $$\begin{aligned}
\frac{d}{dt}\left[D_{KL}\left(\frac{{\hat{x}}}{|{\hat{x}}|} || \frac{x}{|x|}\right)\right] &= 0 -\frac{d}{dt}\left[ \sum_{i}{\frac{{\hat{x}}_i}{|{\hat{x}}|} (\log{x_i} - \log{|x|})} \right]\\
&= -\sum_{i}{\frac{{\hat{x}}_i}{|{\hat{x}}|} \frac{\dot{x}_i}{x_i}} + \sum_{i}{\frac{{\hat{x}}_i}{|{\hat{x}}|} \frac{\dot{|x|}}{|x|}} \\
&= -\sum_{i}{\frac{{\hat{x}}_i}{|{\hat{x}}|} \frac{1}{x_i} \frac{x_i}{|x|} f_i(x) } + \frac{x \cdot f(x) }{|x|^2} \\
&= -\frac{1}{|x|} \sum_{i}{\frac{{\hat{x}}_i}{|{\hat{x}}|} f_i(x)} + \frac{1}{|x|^2} x \cdot f(x)\\
&= -\frac{1}{|x|}\left( \frac{{\hat{x}} \cdot f(x)}{|{\hat{x}}|} - \frac{x \cdot f(x)}{|x|} \right) < 0.\end{aligned}$$
Informatics of Multiple Population Replicator Dynamics
======================================================
The information-theoretic approach easily extends to multiple population replicator equations such as bimatrix games. As before, the potential information plays a crucial role. It is the sum of the potential informations of all populations that plays the role of the Lyapunov function and gives rise to the geometry. It suffices to discuss the two population case as it is clear that the results extend inductively to finitely-many populations.
Two Populations
---------------
Consider two categorical distributions $p = (p_1, \ldots, p_n)$ and $q = (q_{n+1}, \ldots, q_{n+m})$ with fitness landscapes $f(p,q) = (f_1(p, q), \ldots, f_n(p, q))$ and $g(p, q) = (g_{n+1}(p,q), \ldots, g_{n+m}(p,q))$. Define the coupled replicator system
$$\begin{aligned}
\dot{p_i} &= p_i(f_i(p,q) -{\mathbb{E}_{p}\left[f(p,q)\right]}) \\
\dot{q_j} &= q_j(g_j(p,q) -{\mathbb{E}_{q}\left[g(p,q)\right]})\end{aligned}$$
where $i$ runs from 1 to $n$ and $j$ runs from $n+1$ to $n+m$. Note carefully that the expected values are taken with each distribution respectively.
This system is the gradient flow of the Riemannian metric defined on the interior of $\Delta^n \times \Delta^{m}$ given by $$G_{i,j}(p, q) = \begin{cases}
\frac{1}{p_i} & \text{if $i = j \leq n$ } \\
\frac{1}{q_i} & \text{if $i = j > n$ } \\
0 & \text{else}
\end{cases}$$ That is, the matrix is the direct sum matrix of the usual metric for each equation. As in the single population case, we can use potential information to form a Lyapunov function for the system. Given states $\hat{p}$ and $\hat{q}$, let $L$ be the sum of the potential information of each categorical distribution. That is, $$\begin{aligned}
L &= D_{p}(\hat{p}, p) + D_{q}(\hat{q}, q)\\
&= \sum_{i}{\hat{p}_i \log{\hat{p}_i}} - \sum_{i}{\hat{p}_i \log{p_i}} + \sum_{j}{\hat{q}_j \log{\hat{q}_j}} - \sum_{j}{\hat{q}_j \log{q_j}}\end{aligned}$$
The metric can be obtained as the localization of the sum of the divergence functions. All the usual calculations follow from the fact that the system is a gradient, e.g. Fisher’s Fundamental Theorem.
Potential Information is a Lyapunov Function
--------------------------------------------
Recall that $\hat{p}$ is an ESS in the single population case if $\hat{p} \cdot f(p) > p \cdot f(p)$ for all $p$ in a neighborhood of $\hat{p}$.
If $\hat{p}$ and $\hat{q}$ are ESS for each system respectively then $L$ is a Lyapunov function for the coupled system.
A straight-forward computation shows that (up to a negative) $$\dot{L} = \hat{p} \cdot f(p, q) - p \cdot f(p, q) + \hat{q} \cdot g(p, q) - q \cdot g(p, q).$$ $L$ is positive everywhere and has minimum at $(\hat{p}, \hat{q})$. Since $\hat{p}$ and $\hat{q}$ are ESS, $\dot{L} < 0$, so $L$ is a local Lyapunov function.
Notice that the hypothesis that both $\hat{p}$ and $\hat{q}$ are ESS is too strong. Indeed, all that is required is that $$\hat{p} \cdot f(p,q) + \hat{q} \cdot g(p,q) > p \cdot f(p, q) + q \cdot g(p,q).$$ Call this condition a *coupled ESS* (as well as its obvious higher dimensional analogs) and note that any ESS is a coupled ESS. Then $L$ is a Lyapunov for the system if and only if ($\hat{p}$ and $\hat{q}$) is a coupled ESS for the two population system.
Solutions
---------
We can again show that the solutions are exponential families. Let $\dot{v} = f(p, q)$ and $\dot{w} = g(p, q)$. Let $\dot{N} = {\mathbb{E}_{p}\left[f(p,q)\right]}$ and $\dot{M} = {\mathbb{E}_{q}\left[g(p,q)\right]}$. Then $p_i = \exp{(v_i - N)}$ and $q_j = \exp{(w_j - M)}$ is a solution to the coupled system. Indeed, $$\dot{p}_i = \exp{(v_i - N)} (\dot{v}_i - \dot{N}) = p_i( f_i(p, q) - {\mathbb{E}_{p}\left[f(p, q)\right]}),$$ and similarly for $q_j$.
Multiple Populations
--------------------
The above generalizes by induction to show that for a coupled system of multiple interacting populations, the sum of the respective potential informations gives a Lyapunov function for a coupled ESS.
Discussion
==========
The Shahshahani geometry can be interpreted within the framework of information theory as the information geometry of the simplex. This explains the origin of several quantities in evolutionary game theory including the Shahshahani metric and the Kullback-Liebler information divergence. An important feature of the approach is that the information-geometric reasoning extends to the Lotka-Volterra equation and the multiple population replicator equation easily within the framework. Additionally, the replicator dynamic arises intuitively from purely mathematical and statistical concepts such as Fisher information. This shows that the replicator equation models the information dynamics of natural selection.
[^1]: Different coordinates are sometimes chosen in information geometry for $P(X)$, letting (for instance) $x_{n+1} = 1 - \sum_{i=1}^{n}{x_i}$ to enforce $\sum_{i}{x_i} = 1$. This yields an asymmetric set of replicator equations, so a different set of coordinates is chosen in this exposition.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Recent experiments have shown that the superconducting energy gap in some cuprates is spatially inhomogeneous. Motivated by these experiments, and using exact diagonalization of a model $d$-wave Hamiltonian, combined with Monte Carlo simulations of a Ginzburg-Landau free energy functional, we have calculated the single-particle density of states LDOS$(\omega,r)$ of a model high-T$_c$ superconductor as a function of temperature. Our calculations include both quenched disorder in the pairing potential and thermal fluctuations in both phase and amplitude of the superconducting gap. Most of our calculations assume two types of superconducting regions: $\alpha$, with a small gap and large superfluid density, and $\beta$, with the opposite. If the $\beta$ regions are randomly embedded in an $\alpha$ host, the LDOS on the $\alpha$ sites still has a sharp coherence peak at $T = 0$, but the $\beta$ component does not, in agreement with experiment. An ordered arrangement of $\beta$ regions leads to oscillations in the LDOS as a function of energy. The model leads to a superconducting transition temperature $T_c$ well below the pseudogap temperature $T_{c0}$, and has a spatially varying gap at very low $T$, both consistent with experiments in underdoped Bi2212. Our calculated LDOS$(\omega,r)$ shows coherence peaks for $T < T_c$, which disappear for $T > T_c$, in agreement with previous work considering phase but not amplitude fluctuations in a homogeneous superconductor. Well above $T_c$, the gap in the LDOS disappears.'
author:
- 'Daniel Valdez-Balderas'
- David Stroud
title: 'Single-Particle Density of States of a Superconductor with a Spatially Varying Gap and Phase Fluctuations'
---
Introduction
============
According to low temperature scanning tunneling microscopy (STM) experiments, the local density of states (LDOS) of some cuprate materials have spatial variations [@cren; @pan_davis; @howald1; @lang_davis; @howald2; @fang; @kato_sakata; @mashima]. Among the cuprates, Bi$_{2}$Sr$_{2}$CaCu$_{2}$O$_{8+x}$ (Bi2212) is one of the most extensively studied in STM experiments. The LDOS spectrum shows that some regions of that material, which we will call $\alpha$-regions, have a small energy gap with large and narrow coherence peaks (reminiscent of the spectra observed in bulk superconducting materials), while other regions, which we will call $\beta$-regions, have a larger gap, but smaller and broadened peaks (which are reminiscent of the spectra seen in bulk pseudogap phase of some materials . These inhomogeneities occur on length scales of order 30$\AA$. Because at low doping concentrations $\alpha$ regions with “good” superconductivity are immersed in more metallic or semiconducting $\beta$ regions, some workers have made an analogy between these materials and granular superconductors [@lang_davis; @joglekar]: superconducting domains spatially separated from one another by non-superconducting regions, but connected through proximity effect or Josephson tunneling.
At present there is no general agreement regarding the origin of the inhomogeneities in the cuprate superconductors —whether they are in charge density, spin density, LDOS, or other properties [@scalapino_nunner_hirschfeld]. One hypothesis is that these inhomogeneities originate in a process of self organization due to competing orders [@zaanen; @emery_lin; @emery_kivelson; @low_emery; @emery_kivelson_tranquada; @jamei; @valdez_stroud]. In another approach, the spatially varying properties of the cuprates are attributed to crystal defects or impurities. In particular, it has been suggested that the inhomogeneities in the LDOS originate in the random spatial distribution of dopant atoms near the copper oxide (CuO$_2$) planes [@pan_davis; @martin_balatsky; @wang_pan; @atkinson; @nunner_andersen_hirscfeld].
Several workers have studied the LDOS of inhomogeneous superconductors at low $T$. For example, Ghosal [*et al*]{}[@ghosal] have calculated the LDOS of a strongly disordered $s$-wave superconducting layer in two dimensions, solving the Bogoliubov-de Gennes equations self-consistently. They have also done similar work on a model of $d$-wave superconductivity[@ghosal1]. Fang [*et al.*]{}[@fang], using a Green’s function approach, computed the zero temperature LDOS of a model lattice Hamiltonian in which one small region of the lattice has an different (either suppressed or enhanced) pairing strength than the rest; they find good agreement with experiments. Cheng and Su [@cheng_su] have also explored how the LDOS is affected by a single spatial inhomogeneity in the pairing strength of a BCS Hamiltonian; they find that an inhomogeneity with an LDOS most closely resembling the experimental results is produced by an inhomogeneity with a cone-shaped distribution of the pairing strength; this work thus suggests that it is the small-length-scale variation of the pairing strength that causes incoherence in the LDOS. Mayr [*et al.*]{} [@mayr_dagotto] have studied a phenomenological model with quenched disorder and observed a pseudogap in the LDOS caused by a mixture of antiferromagnetism and superconductivity, while Jamei [*et al.*]{} [@jamei_kapitulnik_kivelson] have investigated the low order moments of the LDOS and their relation to the local form of the Hamiltonian.
In this paper we propose a phenomenological approach to study the effect of inhomogeneities on the LDOS in a model for cuprate superconductors. The model is a mean-field BCS Hamiltonian with $d$-wave symmetry, in which the pairing-field is inhomogeneous and also undergoes thermal fluctuations in both phase and amplitude at finite temperatures $T$. It has been argued [@scalapino_nunner_hirschfeld] that the superconducting state of optimally doped to overdoped cuprates is well described by the BCS theory which includes a $d$-wave gap and scattering from defects outside the Cu0$_2$ plane. Instead of including such defects explicitly in our BCS Hamiltonian, we implicitly include their possible effects through inhomogeneities of the pairing-field amplitude. Furthermore, instead of self-consistently solving the Bogoliubov-de Gennes equations resulting from this model, we obtain the magnitude and phase of the complex pairing-field from Monte Carlo simulations based on a Ginzburg Landau free energy functional. Thus the procedure is as follows. First, we set the parameters of the Ginzburg Landau free energy functional from experiments. Next, using Monte Carlo simulations of this free energy, we obtain the pairing-field amplitudes which we then include in the BCS Hamiltonian. Finally, we diagonalize the latter in order to obtain the LDOS.
Now in optimally or nearly optimally doped Bi2212, the layers consist of randomly distributed $\beta$-regions immersed in a majority background of $\alpha$-regions [@pan_davis]. We therefore choose Ginzburg-Landau parameters so as to reproduce this morphology at $T
= 0$, then carry out simulations at both zero and finite $T$ to obtain the LDOS in the different spatial regions.
At $T = 0$ we compare these simulation results to those obtained using ordered instead of random arrangements of inhomogeneities. We find that the LDOS of the random systems much more closely resemble experiment. Specifically, regions with a small gap have sharp coherence peaks, while large-gap regions show lower and broader peaks. By contrast, systems with ordered inhomogeneities have LDOS spectra with sharp coherence peaks which oscillate as a function of energy. In the ordered systems, the coherence peaks in the small-gap regions strongly resemble those observed in a homogeneous small-gap system. But the spectral peaks in the large-gap regions dramatically differ from those in the corresponding homogeneous and disordered cases. Because the spectra of disordered systems more closely resemble experiments, we have also studied the evolution of the LDOS in these systems with increasing $T$. We consider both $T < T_c$ and $T >
T_c$, where $T_c$ is the phase-ordering transition temperature (equivalent to the Kosterlitz-Thouless transition temperature for this two-dimensional system). In both the $\alpha$ and $\beta$ regions, we find that the spectral gap starts to fill in as $T$ increases, and the spectral peaks broaden and are reduced in height. However, even above $T_{c}$, the LDOS is still suppressed at low energies, in comparison to the normal state. This result agrees with a previous study [@eckl; @eckl2] which considered thermal fluctuations of the phase but not of the magnitude of the complex pairing-field, and included no quenched disorder.
We have also studied the $T$-dependence of the magnitude of the pairing field, its thermal fluctuations, and the effective superfluid density of our disordered system. We find that the phase ordering temperature is greatly reduced from the spatial average of the mean-field transition temperatures appearing in the Ginzburg Landau free energy functional. This reduction is due to both thermal fluctuations and quenched disorder in our model.
Although our work involves a non-self-consistent solution of a $d$-wave BCS Hamiltonian, it differs from previous studies of this kind[@fang; @ghosal; @ghosal1; @cheng_su] because it includes thermal fluctuations as well as quenched disorder in the pairing-field amplitude. For our model, quenched disorder is crucial in obtaining LDOS spectra which depend smoothly on energy and are also consistent with the observed low and broad peaks in the $\beta$ regions. The rest of this article is organized as follows: In Section II, we present the BCS model Hamiltonian. In Section III, we derive the discrete form of the Ginzburg Landau free energy functional used in our calculations. We also discuss simple estimates of the phase ordering temperature, our choice of model parameters and our method of introducing inhomogeneities into our model. Section IV describe the computational methods used at both zero and finite temperature. These methods include a classical Monte Carlo approach to treat thermal fluctuations, exact diagonalization to obtain the LDOS, and the reduction of finite size effects on the LDOS by the inclusion of a magnetic field. Section V presents our numerical results at both $T = 0$ and finite $T$. A concluding discussion and summary are given in Section VI.
Model
=====
Microscopic Hamiltonian
-----------------------
We consider the following Hamiltonian: $$H_{BCS} = 2 \sum_{\langle i,j \rangle,\sigma}t_{ij}c_{i\sigma}^{\dagger}c_{j\sigma}
+2\sum_{\langle i,j \rangle}(\Delta_{ij}c_{i\downarrow}c_{j\uparrow} +
\text{c.c.}) -\mu \sum_{i,\sigma}c_{i\sigma}^{\dagger}c_{i\sigma}
\label{eq:hamil_bcs}$$ Here, $\sum_{\langle i, j \rangle }$ denotes a sum over distinct pairs of nearest neighbors on a square lattice with $N$ sites, $c_{j\sigma}^{\dagger}$ creates an electron with spin $\sigma$ ($\uparrow$ or $\downarrow$) at site $j$, $\mu$ is the chemical potential, $\Delta_{ij}$ denotes the strength of the pairing interaction between sites $i$ and $j$, and $t_{ij}$ is the hopping energy, which we write as $$t_{ij} = -t_{hop}. %\, e^{i A_{ij}}.
\label{eq:hopp_const}$$ where $t_{hop} > 0$. Following a similar approach to that of Eckl [*et al.*]{} [@eckl], we take $\Delta_{ij}$ to be given by $$\Delta_{ij} = \frac{1}{4}\frac{|\Delta_i|+|\Delta_j|}{2} e^{i
\theta_{ij}},
\label{delta}$$ where $$\theta_{ij} =
\begin{cases}
(\theta_i+\theta_j)/2, & \text{if bond $\langle i, j \rangle$ is in $x$-direction,}\\
(\theta_i+\theta_j)/2 + \pi, & \text{if bond $\langle i, j \rangle$ is in $y$-direction,}
\end{cases}
\label{eq:thetaij}$$ and $$\Delta_{j}=|\Delta_j|e^{i\theta_{j}},
\label{eq:psi_complex}$$ is the value of the complex superconducting order parameter at site $j$. We will refer to the lattice over which the sums in (\[eq:hamil\_bcs\]) are carried out as the [*atomic*]{} lattice (in order to distinguish it from the [*XY*]{} lattice, which will be described in the next section.) The first term in Eq. (\[eq:hamil\_bcs\]) thus corresponds to the kinetic energy, the second term is a BCS type of pairing interaction with $d$-wave symmetry, and the third term is the energy associated with the chemical potential. Eq. (\[eq:hamil\_bcs\]) may also be written $$H_{BCS} = \Psi^{\dagger} \hat A \ \Psi - N \mu,
\label{eq:hamil_bcs2}$$ where $$\Psi \equiv
\left(
\begin{array}{c}
c_{i\uparrow} \\
c_{i\downarrow}^{\dagger}
\end{array}
\right), \quad \text{$i=1,N$}
\label{eq:Psi}$$ and $$\hat A= \left[
\begin{array}{cc}
\hat t & \hat\Delta^{\ast} \\
\hat\Delta & -\hat t^{\ast}
\end{array}
\right].
\label{eq:A}$$ Here $\hat t$ and $\hat\Delta$ are $N \times N$ matrices with elements $\hat t_{ij}$ \[$\hat t_{ij}=t_{ij}$, as given by eq.(\[eq:hopp\_const\]) if $i$ and $j$ are nearest-neighbors, $\hat
t_{ij}=-\mu$ if $i=j$, and $\hat t_{ij}=0$ otherwise\] and $\hat
\Delta_{ij}$ \[$\hat \Delta_{ij} = \Delta_{ij}$, as given by eq.(\[delta\]) if $i$ and $j$ are nearest-neighbors, and $\hat
\Delta_{ij} = 0 $ otherwise\].
Let $\hat U$ be the unitary matrix that diagonalizes $\hat A$, i. e., $$\hat B = \hat U^{\dagger} \hat A\ \hat U, \quad\quad \hat B \quad \text{diagonal.}
\label{eq:B}$$ We can then rewrite (\[eq:hamil\_bcs2\]) as $$H_{BCS} = \Phi^{\dagger} \hat B \ \Phi - N \mu,
\label{eq:hamil_bcs3}$$ with $\Phi$ defined by $$\Psi = \hat U\ \Phi.
\label{eq:phi}$$ If we make the following definitions: $$\Phi \equiv
\left(
\begin{array}{c}
\gamma_{i\uparrow} \\
\gamma_{i\downarrow}^{\dagger}
\end{array}
\right), \quad \text{$i=1,N$}
\label{eq:phi2}$$ and $$\hat U \equiv \left[
\begin{array}{cc}
u_{j}(r_{i}) & -v_{j}^{\ast}(r_{i}) \\
v_{j}(r_{i}) & u_{j}^{\ast}(r_{i})
\end{array}
\right], \quad \text{$i$, $j=1,N$,}
\label{eq:U}$$ where $i$ labels the [*row*]{} and $j$ the [*column*]{} of $N \times
N$ matrices, then we can see that (\[eq:phi\]) is the typical Bogoliubov - de Gennes transformation [@tinkham; @degennes]: $$\begin{aligned}
c_{i\uparrow} = \sum_{j=1}^{N}[ \gamma_{j\uparrow}u_{j}(r_{i}) - \gamma_{j\downarrow}^{\dagger}v_{j}^{\ast}(r_{i})],\nonumber \\
c_{i\downarrow} = \sum_{j=1}^{N}[ \gamma_{j\downarrow}u_{j}(r_{i}) + \gamma_{j\uparrow}^{\dagger}v_{j}^{\ast}(r_{i})].
\label{eq:degennes}\end{aligned}$$ Thus $\Phi$ is a $2N$-dimensional column matrix and $\hat U$ is a $2N
\times 2N$-dimensional square matrix.
Denoting the diagonal elements of the matrix $\hat B$ by $E_n$, we can use (\[eq:A\]), (\[eq:B\]) and (\[eq:U\]) to obtain $$\left[
\begin{array}{cc}
\hat t & \hat\Delta^{\ast} \\
\hat\Delta & -\hat t^{\ast}
\end{array}
\right]
\left[
\begin{array}{c}
u_{n}(r_{i}) \\
v_{n}(r_{i})
\end{array}
\right]
= E_{n}
\left[
\begin{array}{c}
u_{n}(r_{i}) \\
v_{n}(r_{i})
\end{array}
\right].
\label{eq:matrix_eigenvalue}$$ Eq. (\[eq:matrix\_eigenvalue\]) is the eigenvalue problem which must be solved in order to compute the local density of states, as we describe next.
Explicit expression for the local density of states
---------------------------------------------------
We wish to compute the local density of states, denoted LDOS$(\omega,
r_i)$, as a function of the energy $\omega$ and lattice position $r_i
= (x_i, y_i)$ at both zero and finite temperature $T$. Given the value of the of the superconducting order parameter $\Delta_{i}$ at each lattice site, the matrix $\hat \Delta$ can be constructed and the $\text{LDOS}(\omega,r,\{\Delta_{i}\})$ can be computed through [@ghosal] $$\begin{aligned}
\text {LDOS} (\omega,r_{i},\{ \Delta_{i} \} ) = \sum_{n, E_{n} \geq 0} [
|u_{n}(r_{i})|^2 \delta(\omega-E_{n}) + |v_{n}(r_{i})|^2 \delta(\omega+E_{n}) ]
%
% \ [\ |u_{n}(r_{i})|^2(1-f(E_{n},T))+|v_{n}(r_{i})|^2f(E_{n},T)\ ]\ \delta(\omega-E_{n}) \ + \nonumber \\
% \ [\ |u_{n}(r_{i})|^2f(E_{n},T)+|v_{n}(r_{i})|^2(1-f(E_{n},T))\ ]\ \delta(\omega+E_{n}) \ \rbrace,
\label{eq:finite_temp_dos}\end{aligned}$$ At $T=0$ all the phases $\theta_{i}$ are the same, since this choice minimizes the energy of the superconducting system. Thus, in this case, once we know $\{|\Delta_{i}|\}$ we can solve eq. (\[eq:matrix\_eigenvalue\]) for $u_{n}(r_{i})$, $v_{n}(r_{i})$ and $E_{n}$, and use this solution in (\[eq:finite\_temp\_dos\]). At finite $T$, since $\Delta_{i}$ will thermally fluctuate, we need a procedure to obtain an average of (\[eq:finite\_temp\_dos\]) over the relevant configurations of $\{\Delta_{i}\}$. We explain that procedure next.
Model for thermal fluctuations
==============================
At finite $T$ we compute LDOS$(\omega,r_i)$ by performing an average of $\text {LDOS} (\omega,r_{i},\{\Delta_{i}\})$ over different configurations $\{\Delta_{i}\}$. Those configurations are obtained assuming that the thermal fluctuations of $\{\Delta_{i}\}$ are governed by a Ginzburg-Landau (GL) free energy functional $F$, which is treated as an effective classical Hamiltonian. The Ginzburg Landau free energy functional has been widely studied and applied to a variety of systems. It has been extensively used to study granular conventional superconductors [@muhlschlegel_scalapino_denton; @fisher_barber_jasnow; @deutscher_imry_gunter; @patton_lamb_stroud; @ebner_stroud23; @ebner_stroud25; @ebner_stroud28; @ebner_stroud39]. Other studies have focused on the use of GL theory to describe the phase diagram of extreme type II superconductors [@nguyen_sudbo], the influence of defects on the structure of the order parameter of $d$-wave superconductors [@xu_ren_ting; @alvarez_buscaglia_balseiro], and the effect of thermal fluctuations on the heat capacity of high temperature superconductors [@ebner_stroud39; @ramallo_vidal]. Yet other researchers have derived the GL equations for vortex structures from microscopic theories [@ren_xu_ting]. There has also been interest studying the nature of the transition in certain parameter ranges for this type of model [@alvarez_fort; @bittner_janke].
In this section we discuss a procedure for obtaining a suitably discrete form of $F$, and determining its coefficients from experiments. \[The final form of $F$ is given by eq.(\[eq:GLgeneral4\]).\] We also discuss a way to estimate the phase ordering temperature using this model, the choice of the parameters that determine the GL coefficients, and finally a method of introducing inhomogeneities into the model.
Discrete form of the Ginzburg-Landau free energy
------------------------------------------------
For a continuous superconductor in the absence of a vector potential, the Ginzburg-Landau free energy density has the form $$F^{\prime}= \alpha \left(\frac{T}{T_{c0}}-1\right)|\psi^{\prime}|^2 +
\frac{b}{2} |\psi^{\prime}|^4 + \frac{\hbar}{2m^{\ast}}|\nabla\psi^{\prime}|^2.
\label{eq:energy_density}$$ Since $|\psi^{\prime}|^2$ and F$^\prime$ have dimensions of inverse volume, and energy per unit volume, it follows that $\alpha$ and $b$ have dimensions of energy, and (energy $\times$ volume), respectively.
The squared penetration depth $\lambda^2(T)$ and zero-temperature Ginzburg-Landau coherence length $\xi_{0}$ are related to the coefficients of $F$ by [@tinkham] $$\alpha = \frac{\hbar^2}{2m^{\ast}\xi_{0}^{2}},
\label{eq:alpha}$$ and $$b = 8 \pi \mu_{B}^2\left(\frac{\lambda(0)}{\xi_{0}}\right)^2
\label{eq:b}$$ where $\mu_{B}^2\simeq 5.4\times 10^{-5}\text{eV}$-$\AA ^3$ is the square of the Bohr magneton.
Let us assume that the position-dependent superconducting energy gap $\Delta_{i}$ at $r_i$ is related to $\psi^{\prime}_{i}$, as in conventional BCS theory, through $$|\psi^{\prime}_{i}|^2 =
\frac{\alpha_{i}}{9.38b_{i}}\left|\frac{\Delta_{i}}{k_{B}T_{c0i}}\right|^2,
\label{eq:psi_delta}$$ where we have also assumed that $T_{c0}$, $b$, and $\alpha$ are functions of position. The validity of (\[eq:psi\_delta\]) can be verified by noting that in the absence of fluctuations $F^{\prime}$ is minimized by $$|\psi^{\prime}_{i}|^2 =
\frac{\alpha_{i}}{b_{i}}\left(1-\frac{T}{T_{c0i}}\right).
\label{eq:minimized}$$ Combining (\[eq:minimized\]) with (\[eq:psi\_delta\]), we obtain, at $T=0$, $$|\Delta_{i}(0)|^2 = 9.38 (k_{B}T_{c0i})^2.
\label{eq:zero_temp_gap}$$ This result agrees well with experiment provided (i) $T_{c0}$ is interpreted as the temperature at which an energy gap opens according to ARPES experiments, and (ii) $\Delta(0)$ is taken as the low-temperature ($T<<T_c$) magnitude of the gap observed in ARPES and tunneling experiments [@mourachkine].
In order to obtain a discrete version of the free energy functional, we integrate the free energy density (\[eq:energy\_density\]) over volume to yield the free energy $$F=\int F^{\prime}dV.
\label{eq:integrate_Fprime}$$ Assuming that $\psi^{\prime}\sim \text{constant}$ within a volume $\xi_{0}^{2}d$ (where $\xi_{0}$ is the zero-temperature coherence length and $d$ is thickness of the superconducting layer), we can discretize the layer into $M$ cells of volume $\xi_{0}^{2}d$. Using (\[eq:psi\_delta\]), we can then write $$\begin{aligned}
\frac{F}{K_{1}} =
\sum_{i=1}^{M} \left(\frac{T}{T_{c0i}}-1\right)\frac{1}{\lambda^{2}_{i}(0)} \left|\frac{\Delta_{i}}{k_{B}T_{c0i}}\right|^2 +
\sum_{i=1}^{M} \frac{1}{2(9.38)}\frac{1}{\lambda^{2}_{i}(0)} \left|\frac{\Delta_{i}}{k_{B}T_{c0i}}\right|^4 \nonumber \\
+ \sum_{\langle ij \rangle}\left| \frac{\Delta_{i}}{\lambda_{i}(0)k_{B}T_{c0i}} -
\frac{\Delta_{j}}{\lambda_{j}(0)k_{B}T_{c0j}} \right|^2,
\label{eq:general1}\end{aligned}$$ where $$K_{1} \equiv \frac{\hbar^4 d}{32 (9.38) \pi m^{\ast 2} \mu_{B}^{2}}$$ ‘$K_{1}\simeq 2866$ eV-$\AA^2$ if $d=10\AA$. Except for $d$, $K_{1}$ is independent of material-specific parameters.
In (\[eq:general1\]) the sums are performed over what we will call the [*XY*]{} lattice, which is not necessarily the same as the [ *atomic*]{} lattice used in (\[eq:hamil\_bcs\]). In (\[eq:general1\]), $\Delta_{i} = |\Delta_{i}|e^{-i\theta_{i}}$ is the value of the superconducting order parameter on the $i$th cell of the [*XY*]{} lattice. The third sum is carried out over [*distinct*]{} pairs of nearest-neighbors cells $\langle ij \rangle$. In order to see how the [*XY*]{} lattice and the atomic lattice are related, we now analyze some of the relevant length scales in our problem. Typically, the linear dimension of the [*XY*]{} lattice cell is taken to be the $T = 0$ coherence length $\xi_{0}$ of the material in the superconducting layer. In a cuprate superconductor, e. g., Bi2212, $\xi_{0} \approx 15 \AA$, while the lattice constant of the microscopic (atomic) Hamiltonian of eq. (1) - i. e., the distance between the Cu sites in the CuO$_2$ plane - is $a_{0}
\approx 5.4 \AA$ [@mourachkine]. Thus, in this case, a single [*XY*]{} cell would contain about nine sites of the atomic lattice, on each of which the superconducting order parameter would have the same value $\Delta_i$. It is convenient to introduce a dimensionless superconducting gap $$\psi_{i} \equiv \frac{\Delta_{i}}{E_{0}},
\label{eq:psi_def}$$ and a dimensionless temperature $$t \equiv \frac{k_{B}T}{E_{0}},
\label{eq:t_def}$$ where $E_{0}$ is an arbitrary energy scale which will be specified below. We can then rewrite (\[eq:general1\]) as $$\begin{aligned}
\frac{F}{K_{1}} =
\sum_{i=1}^{M} \left(\frac{t}{t_{c0i}}-1\right)\frac{1}{\lambda^{2}_{i}(0)t_{c0i}^2} \left|\psi_{i}\right|^2 +
\sum_{i=1}^{M} \frac{1}{2(9.38)}\frac{1}{\lambda^{2}_{i}(0)t_{c0i}^4} \left|\psi_{i}\right|^4 \nonumber \\
+ \sum_{\langle ij \rangle}
\left[
\left| \frac{\psi_{i}}{\lambda_{i}(0)t_{c0i}} \right|^2 +
\left| \frac{\psi_{j}}{\lambda_{j}(0)t_{c0j}} \right|^2 -
\frac{2|\psi_{i}||\psi_{j}|}{\lambda_{i}(0)t_{c0i}\lambda_{j}(0)t_{c0j}}
\cos(\theta_{i}-\theta_{j}).
\right]
\label{eq:GLgeneral3}\end{aligned}$$ In our calculations, we will employ periodic boundary conditions. In that case, sums of the form $\sum_{\langle ij \rangle}(a_i+a_j)$ can be replaced by $4\sum_{i} a_i$, and $$\begin{aligned}
\frac{F}{K_{1}} =
\sum_{i=1}^{M} \left(\frac{t}{t_{c0i}}+3\right)\frac{1}{\lambda^{2}_{i}(0)t_{c0i}^2} \left|\psi_{i}\right|^2 +
\sum_{i=1}^{M} \frac{1}{2(9.38)}\frac{1}{\lambda^{2}_{i}(0)t_{c0i}^4} \left|\psi_{i}\right|^4 \nonumber \\
-\sum_{\langle ij \rangle}
\frac{2|\psi_{i}||\psi_{j}|}{\lambda_{i}(0)t_{c0i}\lambda_{j}(0)t_{c0j}}
\cos(\theta_{i}-\theta_{j}).
\label{eq:GLgeneral4}\end{aligned}$$ Eq. (\[eq:GLgeneral4\]) is the most general form for the Ginzburg-Landau free energy functional considered in our calculations. In our simulations we allow both the amplitude $|\psi |$ and the phase $\theta$ of $\psi$ to undergo thermal fluctuations.
Thermal averages
----------------
As mentioned at the beginning of this section, at finite $T$ we compute LDOS$(\omega,r_i)$ by performing an average of $\text {LDOS}
(\omega,r_{i},\{\psi_{i}\})$ over different configurations $\{\psi_{i}\}$. Those configurations are obtained assuming that the thermal fluctuations of $\{\psi_{i}\}$ are governed by the Ginzburg-Landau (GL) free energy functional $F$ described above. $F$ is treated as an effective classical Hamiltonian, and thermal averages $\langle... \rangle$ of quantities $Q$, such as LDOS$(\omega,r_i)$, are obtained through $$\langle Q \rangle = \frac{\int \prod_{i=1}^{N} d^2\psi_{i}\, e^{ -F/ k_{B}T} Q(\{\psi_i \}) }{Z},
\label{eq:Q_thermal_avg}$$ where $Z$ is the canonical partition function, $$Z = \int \prod_{i=1}^{N} d^2\psi_{i}\, e^{ -F /k_{B}T}.
\label{eq:partition_function}$$
Estimate of the Kosterlitz-Thouless transition
----------------------------------------------
If amplitude fluctuations are neglected, the Hamiltonian (\[eq:GLgeneral4\]) would correspond to an XY model on a square lattice. If the system is homogeneous, this XY model undergoes a Kosterlitz-Thouless transition at a temperature $$k_{B}T_{c}\simeq 0.89 J_{XY},
\label{eq:KT_transi}$$ where $J_{XY}$ is the coupling constant between spins: $$H_{XY} = - J_{XY} \sum_{\langle ij \rangle}
\cos(\theta_{i}-\theta_{j}).
\label{eq:xy}$$ From eqs. (\[eq:GLgeneral4\]) and (\[eq:xy\]), the XY coupling between sites $i$ and $j$ is given by $$J_{XY,ij}(t) \equiv \frac{2 K_{1} |\psi_{i}||\psi_{j}|}{\lambda_{i}(0)t_{c0i}\lambda_{j}(0)t_{c0j}}.
\label{eq:JXY_deff}$$ If we approximate $\psi_{i}(t)$ by the value that minimizes $F^{\prime}$ when fluctuations are neglected, $$|\psi_{i}(t)|\simeq \sqrt{9.38 (1-t/t_{c0i})}\, t_{c0i},
\label{eq:psi}$$ then $$J_{XY,ij}(t) \simeq \frac{2(9.38)\sqrt{ (1-t/t_{c0i}) (1-t/t_{c0j}) } }{\lambda_{i}(0)\,\lambda_{j}(0)}.
\label{eq:JXY1}$$ which in the homogeneous case reduces to $$J_{XY}(t) \simeq \frac{18.76(1-t/t_{c0}) }{\lambda^{2}(0)}.
\label{eq:JXY2}$$ This result and eq. (\[eq:KT\_transi\]) give $$T_{c} \simeq \frac{T_{c0}}{1 + T_{c0}/\gamma_{1}},
\label{eq:Tc}$$ where $$\gamma_1 = \frac{(0.89)(18.76)K_{1}}{\lambda^{2}(0)k_{B}}.$$ Eq. (\[eq:Tc\]) can also be rewritten as $$t_{c} \simeq \frac{t_{c0}}{1 + t_{c0}\gamma_2},
\label{eq:tc}$$ where $\gamma_2 = \frac{E_{0}}{k_{B}\gamma_1}$. Using $\lambda(0)$ = 1800 $\AA$ and $d$ = 10 $\AA$, we obtain $\gamma_1=172K$. Finally, if we choose $E_{0}=200$ meV (for reasons given below), we obtain $\gamma_2 = 13.54$.
Expressions (\[eq:Tc\]) and (\[eq:tc\]) will typically overestimate the phase-ordering (or Kosterlitz-Thouless) transition temperature $T_c$. Both thermal fluctuations of $|\psi|$ and quenched disorder will generally reduce $T_{c}$ below these estimates.
Choice of parameters
--------------------
Next, we describe our choice of parameters entering both the microscopic model \[Eq. (\[eq:hamil\_bcs\])\], and that for thermal fluctuations \[Eq. (\[eq:GLgeneral4\])\]. In a typical cuprate, such as underdoped Bi2212, the low-$T$ superconducting gap is $\sim
50$ meV, the hopping integral $t_{hop} \sim 200$ meV, $\lambda(0)
\sim 1800\AA$, and the pseudogap opens at $T_{c0} \sim 200K \simeq
20meV/k_{B}$. Also, the lattice constant of the CuO$_2$ lattice plane is $a_{0}\sim 5.4 \AA$, while $\xi_{0} \sim 15 \AA$. If in eqs.(\[eq:psi\_def\]) and (\[eq:t\_def\]) we choose $E_{0} = t_{hop} =
200$meV, then, using those expressions, we obtain $|\psi(0)| = 0.25$ and $t_{c0} = 0.1$. We can substitute these values into eq.(\[eq:Tc\]) to obtain an estimate for the phase ordering temperature, namely $T_{c}=130K$. Our actual simulations, carried out in the presence of thermal fluctuations of the gap magnitude and quenched disorder, actually yield a lower $T_{c}$, as expected.
We have carried out calculations using this set of parameters, but also with smaller values of $\xi_0$, in order to treat larger XY lattices. Suppose we wish to carry out a simulation on a $16\times
16$ XY lattice. If we use the parameters values described above, we would have a $48 \times 48$ atomic lattice. To compute the density of states on this lattice, we would have to diagonalize $4608 \times
4608$ matrices \[see Eq. (\[eq:A\])\]. Each such diagonalization takes $\sim$ 1 hour on a node for serial jobs of the OSC Pentium 4 Cluster, which has a 2.4 GHz Intel P4 Xeon processor. Since thermal averages require several hundred diagonalizations, a $16\times 16$ XY lattice is too large using these parameters. If, however, we choose a smaller coherence length, we will have fewer atomic sites per XY cell, and hence a smaller matrix to diagonalize for a $16 \times 16$ XY lattice. In the BCS formalism, $\xi_{0}\propto v_{F}/|\Delta|$, where $v_{F}$ is the Fermi velocity. Thus, if $\xi_{0}$ is $n$ times smaller than the experimental value, then, for fixed $v_F$, $\Delta$, and hence $t_{c0}$, will be $n$ times larger than that value.
Inhomogeneities
---------------
As noted above, experiments show that in some cuprates the energy gap is spatially inhomogeneous [@cren; @pan_davis; @howald1; @lang_davis; @howald2; @kato_sakata; @fang; @mashima]. Typically, in some spatial regions, which we call $\alpha$-regions, the LDOS has a small gap and large coherence peaks, while in other regions, the $\beta$-regions, the LDOS has a larger gap and reduced coherence peaks. The percentage of the area occupied by $\alpha$ and $\beta$ regions, respectively, depends on the doping concentration. In Bi2212, for example, a nearly optimally doped sample (hole dopant level $\sim 0.18$) has $\sim 10$ % of the area occupied by $\beta$ regions, while for an underdoped sample (hole dopant level $\sim
0.14$), the areal fraction of the $\beta$ regions is about $\sim 50$ % [@lang_davis].
We introduce spatial inhomogeneities into our model by including a binary distribution of $t_{c0i}$’s. Typically, we chose the smaller value of $t_{c0i}$ so that, for a homogeneous system, the gap $\Delta_{i}(0)$ resulting from our model \[eq. \[eq:psi\]\] approximately equals that observed in experiments (for further details, see discussion in the subsection entitled “choice of parameters”). We refer to XY cells with this small $t_{c0i}$ as $\alpha$-cells. For the $\beta$-cells, on the other hand, we assume a value $t_{c0i}$ $K$ times large than that of the $\alpha$-cells. We obtained our best results by choosing $K = 3$. We have carried out simulations considering both an ordered and a random distribution of $\beta$-cells. To determine the distribution of $\lambda_i(0)$, we use the connection between the local superfluid density $n_{s,i}(T)$ and $\lambda_i(T)$ implied by eqs. (\[eq:alpha\]), (\[eq:b\]) and (\[eq:psi\_delta\]): $$n_{s,i}(T) = |\psi^{\prime}_{i}(T)|^2 = \frac{\hbar^2}{(9.38)16\pi\mu_B^2 m^\ast}
\frac{|\Delta_i(T)|^2}{(k_BT_{c0i})^2}\frac{1}{\lambda_i^2(0)}.$$ Thus, at fixed but very low $T$, since $|\Delta_i(0)/k_BT_{c0i}|^2$ is independent of position according to our model \[see Eq. (\[eq:zero\_temp\_gap\])\], $n_{s,i}(T)\propto 1/\lambda_i^2(0)$. Since the coherence peaks in the local density of states are observed to be lower where the gap is large, we will assume that $t_{c0i}$ and $\lambda_i^2(0)$ are correlated according to the equation $$\lambda_i^2(0) = \frac{ \lambda^2(0) }{t_{c0}}t_{c0i}
\label{eq:lambda2_correl}$$ where $t_{c0}$ and $\lambda^2(0)$ are obtained from the observed bulk properties of the material under consideration. \[For example, we typically obtain $t_{c0}$ from (\[eq:zero\_temp\_gap\]) where we take $|\Delta_{i}(0)|$ as the average of the low temperature gap observed in experiments, and $\lambda(0)=1800\AA$.\] Substituting (\[eq:lambda2\_correl\]) into eq. (\[eq:JXY1\]) gives, for $t<<t_{c0i}$ and $t<<t_{c0j}$, $$J_{XY,ij} \propto \frac{1}{\sqrt{t_{c0i}\,t_{c0j}}}.
\label{eq:JXY3}$$
Computational method
====================
Monte Carlo
-----------
We compute thermal averages of several quantities, including $\text
{LDOS} (\omega,r_{i}, T)$, using a Monte Carlo (MC) technique. Thus, we estimate integrals of the form (\[eq:Q\_thermal\_avg\]) using $$\langle Q \rangle = \frac{1}{N_{m}} \sum_{j=1}^{N_{m}} Q( \{ \psi_{i} \} ),
\label{eq:Q_thermal_avg_estimator}$$ where $N_{m}$ is the number of configurations $\{ \psi_{i} \}$ used to compute the average, and the configurations $\{ \psi_{i} \}$ are obtained using the standard Metropolis algorithm [@barkema_newman; @thijssen] as we now describe. We first set the values of the $t_{c0i}$ and $\lambda_{i}(0)$ in each XY lattice cell as described in the previous section. This completely determines the GL free energy functional $F$ \[Eq.(\[eq:GLgeneral4\]).\] We then set the initial values of $\psi_{i}$ so as to minimize $F$. Next we perform attempts to change the value of each $\psi_{i}$ by $\delta_i$, where $\delta_{i}$ is the complex number $\delta_i= \delta_{i, re} + i \delta_{i,im}$, and $\delta_{i,
re}$ and $i \delta_{i,im}$ are random numbers with a uniform distribution in the range $[-\delta_{0},\delta_{0}]$. We define a *MC step* as an attempt to change the value $\psi_{i}$ on each of the XY cells. The value of $\delta_{0}$ is in turn adjusted at each temperature so that attempts to change $\psi$ have a success rate of 50%. Attempts to change $\psi_{i}$ are accepted with a probability $\exp(-\Delta F/k_{B}T)$, where $\Delta F =
F[\psi_{1},\psi_{2},\ldots,\psi_{i}+ \delta_{i},\ldots,\psi_{M}] -
F[\psi_{1},\psi_{2},\ldots,\psi_{i},\ldots,\psi_{M}]$. In this way, different configurations $\{ \psi_{i} \}$ are obtained.
In order to select which of those configurations $\{ \psi_{i} \}$ to use in (\[eq:Q\_thermal\_avg\_estimator\]), we first made an estimate the phase autocorrelation time $\tau$ [@eckl2], in units of MC steps, at each temperature. We chose $\tau=\text{min}[\tau^{\prime},500]$, where $\tau^{\prime}$ is implicitly defined by $$\frac{c(\tau^{\prime})}{c(0)} = \frac{1}{e},
\label{eq:autocorrel1}$$ and $c(\tau^{\prime})$ is an space average of the phase autocorrelation function [@thijssen]: $$c(\tau^{\prime}) = \frac{1}{M} \sum_{j=1}^{M}
\left[
\langle e^{i\theta_{j}(\tau^{\prime})} e^{-i\theta_{j}(0)} \rangle -
\langle e^{i\theta_{j}(\tau^{\prime})} \rangle \langle e^{-i\theta_{j}(0)} \rangle
\right].
\label{eq:autocorrel2}$$ Once we estimated $\tau$, we performed 20$\tau$ MC steps to allow the system to equilibrate, then we carried out an additional 100$\tau$ MC steps at each $T$ for each disorder realization. During those 100$\tau$ MC steps, we sampled $\{ \psi_{i} \}$ every $\tau$ MC step, thus obtaining $N_{m}=100$ configurations to use in (\[eq:Q\_thermal\_avg\_estimator\]) to estimate the quantities of interest. We also performed longer simulations, averaging over $N_{m}=300$ configurations to compute the LDOS, and $N_{m}=5000$ configurations to compute $\gamma$, $|\psi|$ and the root-mean-square fluctuations $[\sigma_{|\psi|}]$ (defined below), obtaining virtually the same results as with $N_{m}=100$ configurations. When carrying out the simulation, we need a mapping between the sites of the XY lattice and those of the atomic lattice. To do this mapping, we divide the atomic lattice into regions of area $\xi_{0}
\times \xi_{0}$. Each such region constitutes an XY cell. All atomic sites within such a cell are assigned the same value of the order parameter $\psi_i$. Clearly, the lattice constant $\xi_0$ of the XY lattice must be an integer multiple of the atomic lattice constant $a_0$. Thus, if our XY lattice has $L^2 $ sites, then the atomic lattice has $[L\xi_{0}/a_{0}]^2$ sites.
We diagonalize all matrices numerically using LAPACK [@lapack] subroutine “zheev”, which can find all of the eigenvectors and eigenvalues of a complex, hermitian matrix. We calculate the density of states by distributing the eigenvalues into bins of width $\Delta\omega$. The delta function appearing in (\[eq:finite\_temp\_dos\]) is approximated by $$\delta(x) = \frac{1}{\pi}\frac{\epsilon}{\epsilon^2 + x^2},$$ where we choose $\epsilon \sim \Delta\omega \sim 0.01 t_{hop}$.
Reducing finite size effects through inclusion of a magnetic field
------------------------------------------------------------------
To reduce finite size effects on LDOS$(\omega,r)$, we use a method introduced by Assaad [@assaad]. The basic idea of this method is to break the translational invariance of $t_{hop}$ through the substitution $t_{hop}\rightarrow t_{ij}(L)$ in Eq. (\[eq:hopp\_const\]). This is done so as to improve convergence of the quantities of interest, such as LDOS$(\omega,r)$, as a function of the size of the atomic lattice $N$ [@eckl2]. However, $t_{ij}(N)$ must still satisfy $$\lim_{N\rightarrow \infty} t_{ij}(N) = -t_{hop},
\label{eq:hopp_const2}$$ so that the original form of $t_{ij}$ is recovered in the thermodynamic limit.
Assaad showed that if one makes the substitution $t_{hop}\rightarrow
t_{ij}(N)$ through the inclusion of a finite magnetic field, the convergence of the density of states is greatly improved. The magnetic field enters through the Peierls phase factor:
$$t_{ij} = -t_{hop}\, e^{i A_{\vec i \vec j}},
\label{hopp_const2}$$
with $$A_{\vec i \vec j} = \frac{2\pi}{\Phi_0}\int_{\vec i}^{\vec j}\vec A(\vec r)\cdot d\vec
r.$$ Here $\vec A(\vec r)$ is the vector potential at $\vec r$, $\Phi_0 =
hc/e$ is the flux quantum corresponding to one electronic charge $e$, and the integral runs along the line from site $i$ to site $j$.
We use a gauge which allows periodic boundary conditions, and with which the flux through the atomic lattice can be chosen to be any integer multiple of $\Phi_0$ [@assaad; @yu_stroud]. Let $\vec i = (x\hat e_{x}a_{0}, y\hat e_{y}a_{0})$, $\hat e_{x}$ and $\hat e_{y}$ are unit vectors in the $x$ and $y$ directions, and $x$ and $y$ are integers in the range $[0, N-1]$. Then $$A_{\vec i \vec j} =
\begin{cases}
\pm \frac{2\pi m}{N^2}x, & \text{if} \quad \vec j = \vec i \pm a_{0} \hat e_{y},\\
- \frac{2\pi m}{N}y, & \text{if} \quad \vec j = \vec i + a_{0} \hat e_{x} \quad \text{and}\quad x=N-1,\\
\frac{2\pi m}{N}y, & \text{if} \quad \vec j = \vec i - a_{0} \hat e_{x} \quad \text{and}\quad x=0,\\
0, & \text{otherwise}
\end{cases}
\label{eq:delta1}$$ where $m$ is the number of flux quanta through the atomic lattice. We have chosen $m=1$, so that the magnetic field in our system has the smallest non-zero value possible.
Results
=======
Zero temperature
----------------
Fig. \[fig:ldos\_pure\_diff\_gap\_values\] shows the spatially averaged density of states, DOS$(\omega)$, obtained by summing the local density of states, LDOS$(\omega,r)$, over all sites $r$ on a $48 \times 48$ atomic lattice with homogeneous $t_{c0}$ at zero temperature. The zero temperature pairing strength is given by $|\psi(0)|=\sqrt{9.38}t_{c0}$, as shown by eqs. (\[eq:zero\_temp\_gap\]) and (\[eq:psi\_def\]). For the case $t_{c0}=0$, the pairing strength is zero, and we observe the standard Van Hove peak [@van_hove] for a two-dimensional tight-binding band at $\omega = 0$. For finite pairing strength we observe a suppression of the density of states near $\omega = 0$, while strong coherence peaks occur at $\omega \simeq |\psi(0)|$. In Fig. 2, we compare the density of states DOS$(\omega)$ for a $32
\times 32$ atomic lattice containing a single quantum of magnetic flux ($q=1$), and a larger ($48 \times 48$) atomic lattice containing no magnetic field ($q=0$). Both systems are assumed homogeneous with $t_{c0} = 0.14$. As can be seen, the two are very similar except at low $|\omega|$, where the magnetic field is known to induce a change in the density of states [@lages_sacramento_tesanovic]. Note also that the zero-field DOS$(\omega)$ is less smooth than that of the lattice with one quantum of flux, even though the zero-field lattice is larger. In zero-field case we have determined the density of states using a bin width $\Delta \omega = 0.09$, while in the finite-field case we used $\Delta \omega = 0.01$. (The frequencies and widths are given in units of $t_{hop}$.) We have also carried out a similar calculation for $q = 1$ and a $48 \times 48$ atomic lattice; the results are similar to those shown for the $32 \times 32$ lattice except that the density of states at $\omega = 0$ is reduced by about a third. This Figure, and the results just mentioned, show that including the magnetic field is very useful in smoothing the density of states plots. Before presenting our results for inhomogeneous systems, we briefly describe our method of introducing inhomogeneities into our model. We work with atomic lattices of size $L\times L$, in which the sites are divided into groups of $2 \times 2$. Each of these groups forms an XY cell, within which the superconducting order parameter $\psi$ is kept uniform. The value of $\psi$ in each cell is determined by the GL free energy Eq. (\[eq:GLgeneral4\]), which in turn depends on the set of values $\{t_{c0i}\}$ and $\{\lambda_{c0i}\}$. Because $t_{c0i}$ and $\lambda_{c0i}$ are correlated in our model, once we have the set $\{t_{c0i}\}$, the GL free energy is completely determined and $\psi$ at each cell can be computed through the MC method described above.
In Fig. \[fig:ldos\_ord\_dis\_low\_b\] (a) and (b), we show results for two inhomogeneous systems. Both systems consist of $48 \times
48$ atomic lattices in which a fraction $c_{\beta}=0.11$ of the XY cells are of the $\beta$ type with $t_{c0}=0.42$, while the remainder of the cells are of the $\alpha$ type, with $t_{c0}=0.14$. The curves are spatial averages of the LDOS$(\omega,r)$ over the $\alpha$ and $\beta$ cells. In (a), they correspond to a system in which the $\beta$ cells form an ordered array , while the curves in part (b) correspond to a system in which the $\beta$ cells are distributed randomly through the lattice. For comparison, Fig. \[fig:ldos\_ord\_dis\_low\_b\](c) shows results of two homogeneous systems: one with all $\alpha$ cells and one with all $\beta$ cells.
The dotted line in Fig. \[fig:ldos\_ord\_dis\_low\_b\](a) represents an average of LDOS$(\omega,r)$ over the $\beta$ cells. It differs significantly from the $\beta$ curve of the homogeneous case, \[dotted curve in Fig. \[fig:ldos\_ord\_dis\_low\_b\](c)\]. Specifically, instead of the single sharp, and much higher, peak in the homogeneous $\beta$ case, there is a lower peak which is shifted slightly to smaller $|\omega|$ and also has strong oscillations as a function of $\omega$ (probably because of the ordered arrangement of the $\beta$ cells). The largest maximum of this oscillating peak is quite sharp, however, and occurs at a distinctly smaller energy than in the homogeneous case.
The solid line in Fig. \[fig:ldos\_ord\_dis\_low\_b\](a) corresponds to an average of the LDOS$(\omega,r)$ over $\alpha$ cells. It differs less from the homogeneous $\alpha$ system \[solid curve in Fig. \[fig:ldos\_ord\_dis\_low\_b\](c)\] than in the $\beta$ case: the main peak is not much shifted in energy, and it is slightly lower and broader than the homogeneous case. However, an additional peak does appear at the same position as the larger peak of the inhomogeneous $\beta$ curve described above.
In Fig. \[fig:ldos\_ord\_dis\_low\_b\](b), we show the corresponding density of states plots for a system with randomly distributed $\beta$ cells. In this case we observe that the LDOS$(\omega,r)$, averaged over $\alpha$ cells, has slightly lower and broader peaks than that of the homogeneous $\alpha$ system shown in Fig. \[fig:ldos\_ord\_dis\_low\_b\](c), but the peaks still occur at the same energy in both cases: $\omega \sim 0.42$. However, the average of the LDOS$(\omega,r)$ over the $\beta$ cells is drastically different from the homogeneous $\beta$ case: the main peak is greatly broadened, compared to the homogeneous $\beta$ case.
In Figs. \[fig:lattice\_ord\] and \[fig:lattice\_dis\], we show representative ordered and disordered arrangements of $\alpha$ and $\beta$ cells (for an $18\times 18$ XY lattice), similar to those used in the calculations of Fig. 3(a) and 3(b). In our density of states calculations for disordered arrangements, we typically average over about five realizations of the disorder, and use $24
\times 24$ XY lattices rather than the $18\times 18$ shown in the schematic picture.
In Fig. \[fig:ldos\_ord\_dis\_low\_a\], we show plots analogous to Fig. \[fig:ldos\_ord\_dis\_low\_b\], but for a much larger concentration of $\beta$ cells ($c_\beta = 0.89$). Part (a) shows results for an ordered array of $\alpha$ cells immersed in a background of $\beta$ cells. The simple, sharp peaks of the homogeneous $\beta$ case \[dotted curves in Fig. \[fig:ldos\_ord\_dis\_low\_b\](c) and Fig. \[fig:ldos\_ord\_dis\_low\_a\](c)\] are split into two sharp peaks at a slightly smaller energy, while the sharp peaks of the homogeneous alpha regions, \[solid curve in Fig. \[fig:ldos\_ord\_dis\_low\_b\](c)\] become even sharper and shifted toward higher energies, leading to a reduction in the density of states near $\omega = 0$. Also, in the inhomogeneous $\beta$ curve of Fig. \[fig:ldos\_ord\_dis\_low\_a\](a), a weak second peak appears at the same energy as one of the peaks in the inhomogeneous $\alpha$ curve. The case of a disordered distribution of $\alpha$ regions immersed in a background of $\beta$ regions is shown in Fig. \[fig:ldos\_ord\_dis\_low\_a\] (b). The peaks of the curves corresponding to both the $\alpha$ and $\beta$ regions become lower and broader than in the homogeneous cases, Fig. \[fig:ldos\_ord\_dis\_low\_a\](c). The peak in the $\beta$ curve occurs at approximately the same energy as in the homogeneous case. The corresponding inhomogeneous $\alpha$ peak, on the other hand, occurs at a higher energy relative to the homogeneous case.
Finite temperatures
-------------------
We have carried out a finite-$T$ study for the system topology most similar to the experimental one[@lang_davis]: a random distribution of $\beta$ regions immersed in a background of $\alpha$ regions. Calculated results for such a system at $T = 0$ are shown in Fig. \[fig:ldos\_ord\_dis\_low\_b\](b). Because more matrix diagonalizations are required at finite $T$ to obtain the relevant thermal averages, we work with $32 \times 32$ atomic lattices, instead of the $48 \times 48$ used at $T = 0$. Since the computational time needed for one diagonalization scales with the linear size $L$ of the system like $L^{6}$, each diagonalization takes about one-tenth the time in these smaller system. Fortunately, the reduction of finite-size effects achieved by introducing a magnetic field leads to good results even for this relatively small system size. This can be seen by comparing the $t=0$ results in Fig. \[fig:ldos\_allT\], which are obtained for a $32 \times 32$ atomic lattice, to the corresponding results shown Fig. \[fig:ldos\_ord\_dis\_low\_b\](b) for a $48 \times 48$ atomic lattice. Besides the partial densities of states, we calculate several additional quantities at finite $t$: the effective superfluid density $\gamma(t)$, the thermal- and space-averaged values of $|\psi|$ in the $\alpha$ and $\beta$ regions, and the relative fluctuations $\sigma_{|\psi|}$ of $|\psi|$ averaged over each of those regions.
We compute the superfluid density $\gamma$ by averaging the diagonal elements $\gamma_{\alpha\alpha}$ ($\alpha = x,y$) of the helicity modulus tensor $\hat \gamma$. Thus, we compute $\gamma =
(\gamma_{xx}+\gamma_{yy})/2$, where [@ebner_stroud28] $$\begin{aligned}
\gamma_{xx} =
\frac{1}{M} \langle \sum_{\langle i,j\rangle} (x_i-x_j)^2 J_{XY,ij} \cos(\theta_{i}-\theta_{j}) \rangle -
\frac{1}{M t} \langle [ \sum_{\langle i,j\rangle} (x_i-x_j) J_{XY,ij} \sin(\theta_{i}-\theta_{j})]^{2} \rangle \nonumber \\
+ \frac{1}{M t} \langle \sum_{\langle i,j\rangle}(x_i-x_j) J_{XY,ij} \sin(\theta_{i}-\theta_{j}) \rangle^{2}.
\label{helicity}\end{aligned}$$ Here $x_i$ is the $x$ coordinate of $i$th XY cell $i$, $M$ is the total number of XY cells, $J_{XY,ij}$ is the effective XY coupling between XY cells and is given by Eq. (\[eq:JXY\_deff\]), $\theta_i$ is the phase of $\psi_i$ and $\langle \rangle$ denotes a canonical average. $\gamma_{yy}$ is defined by the analogous expression with $x_i$ replaced by $y_i$. In our computations, we have set the lattice constant $a_{XY}$ of the XY lattice to be unity.
The mean-square order parameter averaged over the $\alpha$ region is computed from $$[\langle|\psi|^{2}\rangle]_{\alpha} =
\frac{1}{M_{\alpha}} \sum_{i\in\alpha} \langle|\psi_{i}|^{2}\rangle,
\label{eq:psi_avg}$$ where the sum is carried out over all $M_{\alpha}$ XY cells of type $\alpha$. $\left[\langle|\psi|^{2}\rangle\right]_{\beta}$ is defined similarly. The mean magnitude of the order parameter in the $\alpha$ and $\beta$ regions, denoted $[\langle|\psi|\rangle]_\alpha$ and $[\langle|\psi|\rangle]_\beta$, are defined by an equation analogous to eq. (\[eq:psi\_avg\]). We compute the relative fluctuations $[\sigma_{|\psi|}]_\alpha$ of $|\psi|$ within XY cells of type $\alpha$ from the definition $$\left[\sigma_{|\psi|}\right]_\alpha =
\left[\sqrt{ \frac{\langle|\psi_i|^{2}\rangle-\langle|\psi_i|\rangle^{2}}
{\langle|\psi_i|\rangle^{2}} }\right]_\alpha,
\label{eq:sigma_avg}$$ where the triangular brackets denote a thermodynamic average, and $\left[...\right]_{\alpha}$ denotes a space average over the $\alpha$ sites. $\left[\sigma_{|\psi|}\right]_\beta$ is computed analogously. In systems with disorder, the square brackets denote a disorder average as well as a space average.
Fig. \[fig:ldos\_allT\] shows the partial LDOS$(\omega,r)$ averaged over $\alpha$ and $\beta$ cells, at both $t = 0$ and finite $t$. The systems shown have a fraction $c_{\beta}=0.1$ of $\beta$ sites randomly distributed. At $t=0$ the $\alpha$ regions show strong, sharp coherence peaks while the $\beta$ regions have a larger gap but lower and broader peaks. When the temperature is increased to $t=0.015$, the heights of both peaks are reduced, and their widths are increased, but the $\alpha$ peak is still quite sharp, because the system still has phase coherence. This temperature is still below the phase ordering temperature of $t_{c}\simeq 0.03$, as discussed below. As $t$ is increased still further, to $t = 0.035$ and $t =
0.055$, the two density of states peaks broaden still further, there is scarcely any residue of a gap in the density of states, and there is now no sign of a real coherence peak in either the $\alpha$ or the $\beta$ regions.
In Fig. \[fig:hel\], we show the superfluid density $\gamma(t)$, for the model just described but for various concentrations $c_\beta$ of the (randomly distributed) $\beta$ cells. For $c_\beta
= 0.1$, the phase-ordering transition temperature $t_c \sim 0.03$ in these units. Thus, of the plots in the previous Figure, two are below and two are above the phase-ordering transition.
In Figs. \[fig:psi\_abs\_a\] and \[fig:psi\_abs\_b\], we show the thermal, spatial, and disorder averages of $|\psi|$ over the $\alpha$ and $\beta$ regions, denoted $[\langle|\psi|\rangle]_\alpha$ and $[\langle|\psi|\rangle]_\beta$, while Figs. \[fig:rel\_fluct\_a\], and \[fig:rel\_fluct\_b\] show the corresponding averages of the root-mean-square fluctuations $\sigma_{|\psi|}$. A number of features deserve mention. First, the average $|\psi|$ is, of course, larger in the $\beta$ regions than in the $\alpha$ regions, but the root-mean-square fluctuations are comparable in each of the two regions. Second, the increases in the averages of $|\psi|$ above the phase-ordering temperature is an artifact of a Ginzburg-Landau free energy functional as we now explain in detail. The asymptotic behavior of $|\psi|^2$ as $t\rightarrow \infty$, for homogeneous systems, can be obtained in the following way : At very high temperatures $t \gg t_{c0}$, the first term in (\[eq:GLgeneral4\]) goes like $\sim|\psi|^2$, the second goes like $\sim|\psi|^4/t$ and the third (coupling) term goes like $\sim|\psi|^2/t$. We can then neglect the contribution of the third term, whence at large $t$ the XY cells are effectively decoupled. The thermal average of $|\psi|^2$ for an isolated cell is given by $$\langle|\psi|^2\rangle = \frac{ \int_{0}^{\infty}|\psi|d|\psi| |\psi|^2 \exp(-f|\psi|^2-g|\psi|^4) }
{ \int_{0}^{\infty}|\psi|d|\psi| \exp(-f|\psi|^2-g|\psi|^4) }.
\label{eq:isolated1}$$ In our case, $$f=\frac{K_{1}}{t_{c0}^{3}\lambda^{2}(0)E_{0} }
\label{eq:f}$$ and $$g=\frac{K_{1}}{2\,(9.38) t_{c0}^{4}\lambda^{2}(0)E_{0} t}.
\label{eq:g}$$ If $f$ and $g$ are real and positive, as in the present case, the integrals appearing in (\[eq:isolated1\]) can be carried out, with the result $$\langle|\psi|^2\rangle = -\frac{f}{2g} +
\frac{\exp(-f^2/4g)}{\sqrt g \pi \text{erfc}\,\,(f/2\sqrt{g})}.% \text{if $N_{i}\neq0$}.
\label{eq:isolated2}$$ Here $\text{erfc}(z)=1-\text{erf}(z)$, $\text{erf}(z)$ being the gaussian error function. Using an asymptotic expansion [@arfken] for $\text{erf}(f/2\sqrt{g}$), applicable when $f/\sqrt{g}>>1$ as in the present case, we can show that $$\lim_{t\rightarrow\infty}\langle|\psi|^2\rangle \rightarrow 1 / f.
\label{eq:asymptotic}$$ Substituting $f$ from eq. (\[eq:f\]) leads to $$\lim_{t\rightarrow\infty}\langle|\psi|^2\rangle \rightarrow \frac{t_{c0}^{3}\lambda^{2}(0)E_{0} }{K_{1}} \simeq 0.6
\label{high_t_psi2}$$ where the last approximate equality is obtained using the parameters we have discussed above, namely $K_{1}\simeq 2866$ eV-$\AA^2$, $\lambda(0)$ = 1800 $\AA$, $t_{c0}=0.14$, $E_{0}=200$meV.
On the other hand, using eq. (\[eq:psi\]), we obtain $$\lim_{t\rightarrow\ 0}\langle|\psi|^2\rangle \simeq 0.2
\label{low_t_psi2}$$ Thus, our model introduces an unphysical finite value of $\langle|\psi|^2\rangle$ at large $t$. This behavior has been observed in other studies of similar models[@bittner_janke], while in other investigations this feature is less obvious because of the parameters used[@ebner_stroud25]. In our case, since we are interested in temperatures $t < t_{c} < t_{c0}$ or $t\sim
t_{c}<t_{c0}$, this unphysical high-temperature behavior should not be relevant to our calculations.
Our results for $\left[\langle |\psi| \rangle\right]_\alpha $ and $\left[\langle |\psi|\rangle\right]_\beta$ suggest an explanation for one feature in the plots of the LDOS (see Fig. 7). Namely, the $\beta$ peak generally occurs at [*higher*]{} $\omega$ than it would in a superconductor made entirely of $\beta$ material. This shift occurs because, when the $\beta$ and $\alpha$ regions are mixed, $\left[\langle|\psi|\rangle\right]_\beta$ is larger than its value in a homogeneous $\beta$ system (as we further discuss below). This behavior of $\left[\langle|\psi|\rangle\right]_\beta$ can be seen in Fig. \[fig:psi\_abs\_b\], where this quantity is plotted for different values of $c_{\beta}$. Clearly, at low $t$, $|\psi|$ increases as $c_{\beta}$ decreases. For the homogeneous $\beta$ system, $|\psi(t=0)| = 1.29$, as can be obtained directly from eq.(\[eq:psi\]); this value is shown as an open circle at $t = 0$. This upward shift in the $\left[\langle|\psi|\rangle\right]_\beta$ would be difficult to measure, since a pure $\beta$ material may not exist.
The behavior of $\left[\langle|\psi|\rangle\right]_\beta$ has an analog, in our model, in the corresponding behavior in the $\alpha$ cells. Specifically, if $\alpha$ cells are the minority component in $\beta$ host, $\langle|\psi|\rangle$ tends to be substantially smaller than in pure $\alpha$ systems: the smaller the concentration $c_\alpha = 1 - c_\beta$, the smaller the value of $\langle
|\psi|\rangle$ in those regions \[see Fig. \[fig:psi\_abs\_a\]\]. The behavior of $\langle |\psi|\rangle$ in both $\alpha$ and $\beta$ regions basically follows from our earlier discussion, according to which $|\psi|^2$ is larger in regions with a small gap.
Discussion
==========
We have presented a phenomenological model for the temperature-dependent single-particle density of states in a BCS superconductor with a $d_{x^2-y^2}$ order parameter. Our model includes both inhomogeneities in the gap magnitude and fluctuations in the phase and amplitude of the gap. While some of these features have been included in previous models for the density of states (e. g. phase fluctuations in a homogeneous d-wave superconductor, inhomogeneities in the gap magnitude at $T = 0$), our model is more general, and thus potentially more realistic for some cuprate superconductors.
Our main goal is to examine the properties of an inhomogeneous superconductor, including many effects which are likely to be significant in real cuprate materials. The amplitude and phase fluctuations are treated by a discretized Ginzburg-Landau free energy functional, while the density of states is obtained by solving the Bogoliubov-de Gennes equations for a superconductor with a tight-binding density of states and a $d_{x^2-y^2}$ energy gap.
In all our calculations, we have assumed that the superconductor has two types of regions: $\alpha$, with a small gap but a high superfluid density; and $\beta$, with a large gap and small superfluid density. This assumption appears consistent with many experiments on the high-T$_c$ cuprates, especially in the underdoped regime [@lang_davis]. If we assume that the minority component is of type $\beta$, embedded in an $\alpha$ host, we find that the local density of states at $T = 0$ at the $\alpha$ sites has sharp coherence peaks, whereas that of the $\beta$ sites is substantially broadened. This behavior is similar to experiment [@lang_davis; @pan_davis]. This description applies to a [*disordered*]{} distribution of $\beta$ sites in an $\alpha$ host. If the $\beta$ sites are, instead, arranged on a lattice, the local density of states on the $\beta$ sites is sharper, but also has distinct oscillations as a function of energy. Since such oscillations are absent in experiments, the actual $\beta$ regions, if they exist as a minority component, are probably distributed randomly.
In the reverse case of $\alpha$ regions embedded randomly in a $\beta$ host, neither component has an extremely sharp density of states peak. While the $\alpha$ peak is still quite sharp, it is broader than the $\alpha$ peak in the $\beta$-minority case. This result suggests that, if one component occurs only as isolated regions, its minority status tends to broaden its coherence peaks.
Our results also show that the local density of states is strongly affected by phase fluctuations. This feature has already been found for a [*homogeneous*]{} d-wave superconductor[@eckl], but here we demonstrate it in an inhomogeneous superconductor. The most striking effect of finite $T$ is that the coherence peak in the $\alpha$ component disappears above the phase-ordering transition temperature $T_c$. The $\beta$ component does not show a coherence peak even at very low temperatures, but nonetheless this peak too is significantly broadened above $T_c$. For $T$ well above $T_c$, there is no appreciable gap in the local density of states either at the $\alpha$ or the $\beta$ sites.
Our calculations include thermal fluctuations in the amplitude as well as the phase of $\psi$. In general, thermal amplitude fluctuations seem to have only a minor influence on the local density of states. By contrast, the variations in $|\psi|$ due to [*quenched*]{} disorder (i. e., the presence of $\alpha$ and $\beta$ regions in our model) strongly affect the local density of states, as we have already described. To check on the influence of purely thermal amplitude fluctuations, we have calculated the density of states of a homogeneous $\alpha$ superconductor with both phase and amplitude fluctuations, and have compared this to a similar calculation with only phase fluctuations. We find that the additional presence of amplitude fluctuations has little effect on the density of states. To smooth the local density of states, we include in our density of states calculations a magnetic field equal to a flux $hc/e$ in the entire sample area, following the method of Assaad[@assaad]. This field greatly smooths the local density of states, which otherwise varies extremely sharply with energy, because of the many degenerate states of a finite sample at zero field. Our calculated density of states does, of course, correspond to a physical magnetic field, and thus differs slightly from that at zero field. For example, in a homogeneous system with a finite $d$-wave gap, the LDOS$(\omega)$ goes to zero as $|\omega| \rightarrow
0$. By contrast, at finite field, the LDOS approaches a constant value at low $|\omega|$. With no gap, our calculated DOS with nonzero field is indistinguishable from that of a conventional 2D tight-binding band (see Fig. 1), because the field is low (typically around 0.002 flux quanta per atomic unit cell). We conclude that the weak magnetic field very effectively smooths the calculated LDOS in a finite two-dimensional sample with a d-wave gap, but produces a density of states similar to that at zero field, except at very low $|\omega|$.
Although we have included this magnetic field in the Bogoliubov-de Gennes equations, we have omitted it from the Ginzburg-Landau free energy functional, which is thus that of a zero-field system. As we now discuss, we believe that this numerical scheme should indeed converge to the correct physical result for zero magnetic field in the limit of a large computational sample.
As noted earlier, we introduce the vector potential into the LDOS calculation in order to smooth the resulting density of states. In the limit of a large system, the effect of the vector potential, corresponding to a single quantum of flux, should become negligible, since the flux density becomes very small. This is already suggested by our calculated results for the two system sizes we consider (see Fig. 2 and the corresponding discussion). Even with a finite superconducting gap, the vector potential affects the LDOS very little, except at low energies; moreover, even this effect becomes smaller as the sample size increases. Therefore, in the limit of a large enough sample, our approach should give a very similar LDOS to one calculated with no vector potential. Hence, it is reasonable to use this approach in combination with a zero-field Ginzburg-Landau free energy functional to calculate the LDOS at finite temperatures.
If we were to introduce a similar field into the Ginzburg-Landau functional, we believe that it would have a substantial effect, not due to smoothing, on the phase ordering. There would be, not only the XY-like phase transition, as at zero field, but also additional phase fluctuations arising from the extra field-induced vortex. Since this extra vortex is absent at zero field, these effects would be irrelevant to the zero-field system we wish to model. By contrast, introducing a field into the Bogoliubov-de Gennes equations, as we do, provides desirable smoothing with little change in the calculated LDOS; moreover, even this slight change decreases with increasing sample size. Therefore, we believe that the best way to obtain a smooth LDOS at both zero and finite temperatures is to introduce the vector potential into the Bogoliubov-de Gennes equations, for smoothing purposes, but not to include it in the Ginzburg-Landau free energy functional. Our numerical results suggest that this procedure is indeed justified.
One feature of our numerical results may seem counterintuitive. In our model, the $\alpha$ component is assumed to have a gap three times smaller than that of $\beta$, but has a larger local superfluid density, i. e., a smaller penetration depth. We then find that the gap in the $\alpha$ region is [*smaller*]{} in a two-component system with both $\alpha$ and $\beta$ regions, than it is in a pure $\alpha$ system. This counterintuitive result, however, emerges naturally from our discrete Ginzburg-Landau model, which is minimized if the quantities $\Delta_i/(\lambda_i(0)T_{c0i})$ are equal. For our model, $\lambda_i(0)$ is smaller in the small-gap material. It would be of interest if experimental evidence of this behavior were found in a real material.
In our calculation, the LDOS is obtained from a Bogoliubov-de Gennes Hamiltonian whose parameters are determined by the Ginzburg-Landau functional. In fact, it should be possible to proceed in the opposite direction, and obtain the parameters of the functional from the LDOS. Specifically, the energy required to change a phase difference by a given amount depends on an integral over the LDOS. Thus, the calculation we have presented in this paper can, in principle, be made fully self-consistent.
Acknowledgments
===============
We are grateful for support through the National Science Foundation grant DMR04-13395. The computations described here were carried out through a grant of computing time from the Ohio Supercomputer Center. We thank E. Carlson, S. Kivelson, M. Randeria, N. Trivedi, R. Sensarma, D. Tanaskovic, and K. Kobayashi for valuable discussions.
[99]{}
T. Cren, D. Roditchev, W. Sacks, J. Klein, J.-B. Moussy, C. Deville-Cavellin, and M. Lagues, *Phys. Rev. Lett* [**84**]{}, 147 (2000).
C. Howald, P. Fournier, and A. Kapitulnik, *Phys. Rev. B* [**64**]{}, 100504(R) (2001).
S. H. Pan, J. P. O’Neal, R. L. Badzey, C. Chamon, H. Ding, J. R. Engelbrecht, Z. Wang, H. Eisaki, S. Uchida, A. K. Gupta, K.-W. Ng, E. W. Hudson, K. M. Lang, and J. C. Davis, *Nature* [**413**]{}, 282 (2001).
K. M. Lang, V. Madhavan, J. E. Hoffman, E. W. Hudson, H. Eisaki, S. Uchida, and J. C. Davis, *Nature* [ **415**]{}, 412 (2002).
C. Howald, H. Eisaki, N. Kaneko, M. Greven, and A. Kapitulnik, *Phys. Rev. B* [**67**]{}, 014533 (2003).
T. Kato, S. Okitsu, and H. Sakata, *Phys. Rev. B* [**72**]{}, 144518 (2005).
A. C. Fang, L. Capriotti, D. J. Scalapino, S. A. Kivelson, N. Kaneko, M. Greven, and A. Kapitulnik, *Phys. Rev. Lett.* [**96**]{}, 017007 (2006).
H. Mashima, N. Fukuo, Y. Matsumoto, G. Kinoda, T. Kondo, H. Ikuta, T. Hitosugi and T. Hasegawa, *Phys. Rev. B* [**73**]{}, 060502(R) (2006).
Yogesh N. Joglekar, A. H. Castro Neto, and Alexander V. Balatsky, *Phys. Rev. Lett.* [**92**]{},037004 (2004).
D. J. Scalapino, T. Nunner, and P. J. Hirschfeld, *Journal of Physics and Chemistry of Solids* [**67**]{}, 2006.
Jan Zaanen and Olle Gunnarsson, *Phys. Rev. B* [**40**]{}, 7391 (1989).
V. J. Emery, S. A. Kivelson, and H. Q. Lin, *Phys. Rev. Lett.* [**64**]{}, 475 (1990).
V. J. Emery and S. A. Kivelson, *Physica C* [**209**]{}, 597 (1993).
U. Löw, V. J. Emery, K. Fabricius, and S. A. Kivelson, *Phys. Rev. Lett.* [**72**]{}, 1918 (1994).
See V. J. Emery, S. A. Kivelson and J. M. Tranquada, *Proc. Natl. Acad. Sci. USA* [**96**]{}, 8814 (1999).
Reza Jamei, Steven Kivelson, and Boris Spivak, *Phys. Rev. Lett.* [**94**]{}, 056805 (2005).
Daniel Valdez-Balderas and David Stroud, *Phys. Rev. B* [**72**]{}, 214501 (2005).
Ivar Martin and Alexander V. Balatsky, *Physica C* [**357-360**]{}, 46 (2001).
Ziqiang Wang, Jan R. Engelbrecht, Shancai Wang, Hong Ding, and Shuheng H. Pan, *Phys. Rev. B* [**65**]{}, 064509 (2002).
W. A. Atkinson, *Phys. Rev. B* [**71**]{}, 024516 (2005).
Tamara S. Nunner, Brian M. Andersen, Ashot Melikyan, and P. J. Hirschfeld, *Phys. Rev. Lett.* [**95**]{}, 177003 (2005). A. Ghosal, M. Randeria, and N. Trivedi, Phys.Rev. Lett. [**81**]{}, 3940 (1998); A. Ghosal, M. Randeria, and N. Trivedi, Phys. Rev. [**B65**]{}, 014501 (2002).
A. Ghosal, M. Randeria, and N. Trivedi, Phys. Rev. [**B63**]{}, 020505(R) (2000). Arti Garg, Mohit Randeria, and Nandini Trivedi, cond-mat/0609666.
Ming Cheng and W. P. Su, *Phys. Rev. B* [**72**]{}, 094512 (2005).
Matthias Mayr, Gonzalo Alvarez, Adriana Moreo, and Elbio Dagotto, *Phys. Rev. B* [**73**]{}, 014509 (2006).
R. Jamei, J. Robertson, E-A. Kim, A. Fang, A. Kapitulnik, and S. A. Kivelson, cond-mat/0608318.
T. Eckl, D. J. Scalapino, E. Arrigoni, and W. Hanke *Phys. Rev. B* [**66**]{}, 140510(R) (2002).
Doctoral thesis, Universität Würzburg, Würzburg, 2004. See, e. g., Michael Tinkham, *Introduction to superconductivity, 2nd edition* (Dover Publications, Mineola, New York, 1996). P. G. de Gennes, *Superconductivity of metals and alloys* (Addison-Wesley, New York, 1989).
B. Muhlschlegel, D. J. Scalapino, and R. Denton, *Phys. Rev. B* [**6**]{}, 1767 (1972).
Michael E. Fisher, Michael N. Barber, and David Jasnow, *Phys. Rev. A* [**8**]{}, 1111 (1973).
G. Deutscher, Y. Imry, L. Gunther, *Phys. Rev. B* [**10**]{}, 4598 (1974).
B. R. Patton, W. Lamb and D. Stroud, *Inhomogeneous Superconductors–1979*, AIP Conf. Proc. 58, edited by D. U. Gubser et al, pp. 13-22.
C. Ebner and D. Stroud, *Phys. Rev. B* [**23**]{}, 6164 (1981).
C. Ebner and D. Stroud, *Phys. Rev. **B*** [**25**]{}, 5711 (1982).
C. Ebner and D. Stroud, *Phys. Rev. **B*** [**28**]{}, 5053 (1983).
C. Ebner and D. Stroud, *Phys. Rev. **B*** [**39**]{}, 789 (1989).
A. K. Nguyen and A. Sudbo, *Phys. Rev. B* [**60**]{}, 15307 (1999).
Wang Xu, Yong Ren and C. S. Ting, *Phys. Rev. B* [**53**]{}, 12481 (1996).
J. J. Vicente Alvarez, Gustavo C. Buscaglia, and C. A. Balseiro, *Phys. Rev. B* [**54**]{}, 16168 (1996).
Manuel V. Ramallo and Felix Vidal, *Phys. Rev. B* [**59**]{}, 4475 (1999).
Yong Ren, Ji-Hai Xu, and C. S. Ting, *Phys. Rev. Lett.* [**74**]{}, 3680 (1995).
G. Alvarez and H. Fort, *Phys. Rev. B* [ **64**]{}, 092506 (2001).
Elmar Bittner and Wolfhard Janke, *Phys. Rev. Lett.* [**89**]{}, 130201 (2002).
Andrei Mourachkine, *High-Temperature Superconductivity in the Cuprates* (Kluwer Academic Publishers, 2002). See plots in pages 67 and 84.
M.E.J. Newman and G.T. Barkema, *Monte Carlo Methods in Statistical Physics* (Oxford: Clarendon Press, 1999.)
J. M. Thijssen, *Computational physics* (Cambridge Univ. Press, Cambridge, 1999).
E. Anderson, Z. Bai, C. Bischof, S. Blackford, J. Demmel, J. Dongarra, J. Du Croz, A. Greenbaum, S. Hammarling, A. McKenney, A. and D. Sorensen, *[LAPACK]{} Users’ Guide*, 3rd Ed., (Society for Industrial and Applied Mathematics, Philadelphia, PA, 1999). F. F. Assaad *Phys. Rev. **B*** [**65**]{}, 115104 (2002).
Wenbin Yu, K. H. Lee and D. Stroud, *Phys. Rev. **B*** [**47**]{}, 5906 (1993).
Leon Van Hove *Phys. Rev.* [**89**]{}, 1189 (1953).
J. Lages, P. D. Sacramento, and Z. Tesanovic, *Phys. Rev. B* [**69**]{}, 094503 (2004).
George B. Arfken and Hans J. Weber, *Mathematical Methods for Physicists* (Academic Press, 1995).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A necessary and sufficient condition for a Lévy process $X$ to stay positive, in probability, near 0, which arises in studies of Chung-type laws for $X$ near 0, is given in terms of the characteristics of $X$.'
address: |
School of Finance, Actuarial Studies and Statistics, Australian National University, Canberra, ACT,\
Australia.
author:
-
title: 'Conditions for a Lévy process to stay positive near 0, in probability'
---
./style/arxiv-general.cfg
Introduction {#s1}
============
Let $(X_t)_{t\ge0}$ be a real valued Lévy process with canonical triplet $(\gamma,\sigma^2,\Pi)$, thus having characteristic function $Ee^{\mathrm{i}\theta X_t}=
e^{t\Psi
(\theta)}$, $t\ge0$, $\theta\in\mathbb{R}$, with characteristic exponent $$\label{ce}
\Psi(\theta):= \mathrm{i}\theta\gamma- \frac{1} 2
\sigma^2 \theta^2 +\int_{\mathbb{R}\setminus\{0\}}
\bigl(e^{\mathrm{i}\theta x}-1-\mathrm{i}\theta x{\mathbf{1}}_{\{
|x|\le1\}} \bigr) \Pi(\mathrm{d}x).$$ Here, $\gamma\in\mathbb{R}$, $\sigma^2\ge0$, and $\Pi$ is a Borel measure on $\mathbb{R}\setminus\{0\}$ such that $\int_{\mathbb{R}\setminus\{0\}}(x^2\wedge1)\Pi(\mathrm
{d}x)<\infty$.
The condition $$\label{bothpos}
\mathop{\liminf}_{t\downarrow0}P(X_t\le0)\wedge
P(X_t\ge0)>0$$ was shown by Wee [@wei] to imply a Chung-type law at 0 for $X$. Attention is drawn to this in a recent paper of Aurzada, Döring and Savov [@ads], who give extended and refined versions of the Chung law using a quite different approach to that of Wee. The difference between (\[bothpos\]) and the conditions imposed by Aurzada et al. [@ads] is not at all clear, though based on some examples they suggest that theirs are weaker than (\[bothpos\]). Our aim in this paper is to give necessary and sufficient conditions for $X$ to stay positive near 0, or to stay negative near 0, and hence to characterise (\[bothpos\]). We need some more notation. The positive, negative and two-sided tails of $\Pi$ are $$\begin{aligned}
\overline{\Pi}^+(x)&:=& \Pi\bigl\{(x,\infty)\bigr\},\qquad \overline{
\Pi}^-(x):= \Pi\bigl\{(-\infty ,-x)\bigr\}\quad \mbox{and}
\nonumber
\\[-8pt]
\label{pidef}
\\[-8pt]
\nonumber
\overline{\Pi}(x) &:=&
\overline{\Pi}^+(x)+\overline{\Pi }^-(x), \qquad x>0.\end{aligned}$$ The restriction of $\Pi$ to $(0,\infty)$ is denoted by $\Pi^{(+)}$, and we define $\Pi^{(-)}$ on $(0,\infty)$ by $\Pi^{(-)}(\mathrm{d}x):= -\Pi(-\mathrm{d}x)$, for $x>0$. We are only interested in small time behaviour of $X_t$, and we eliminate the compound Poisson case by assuming $\Pi(\mathbb
{R})=\infty$ throughout. Define truncated and Winsorised moments as $$\begin{aligned}
\nu(x)&=& \gamma-\int_{x<|y|\le1}y\Pi(\mathrm{d}y),
\nonumber
\\[-8pt]
\label{Adef}
\\[-8pt]
\nonumber
A(x)&=&
\gamma+\overline{\Pi}^+(1)-\overline{\Pi}^-(1) -\int_x^1
\bigl(\overline{\Pi}^+(y)-\overline{\Pi}^-(y)\bigr)\,\mathrm{d}y\end{aligned}$$ and $$\label{Vdef}
V(x)=\sigma^2+\int_{0<|y|\le x}y^2
\Pi(\mathrm{d}y),\qquad U(x)=\sigma^2+2\int_0^xy
\overline{\Pi}(y)\,\mathrm{d}y,\qquad x>0.$$ These functions are defined and finite for all $x>0$ by virtue of property $\int_{0<|y|\le1}y^2\Pi(\mathrm{d}y)<\infty$ of the Lévy measure $\Pi$ but only their behaviour as $x\downarrow0 $ will be relevant for us. Integration by parts shows that $$\label{IP}
A(x)= \nu(x)+x\bigl(\overline{\Pi}^+(x)-\overline{\Pi}^-(x)\bigr),
\qquad x>0.$$
Doney [@doney2004], Lemma 9, gives the following version of the Itô decomposition of $X$ which caters for positive and negative jumps separately. Take constants $h_+>0$ and $ h_->0$. Then for $t\ge0$, $$\begin{aligned}
X_t& =& t\gamma-t\nu_+(h_+) +t\nu_-(h_-)
\nonumber
\\[-8pt]
\label{decomp3}
\\[-8pt]
\nonumber
&&{}+\sigma
Z_t+ X_t^{(S,h_+,+)}+ X_t^{(S,h_-,-)}+
X_t^{(B,h_+,+)} + X_t^{(B,h_-,-)},\end{aligned}$$ where $\gamma$ and $\sigma$ are as in (\[ce\]), and the functions $\nu_\pm$ are $$\label{nupmdef} \nu_+(h_+):=\int_{(h_+,1]}x\Pi(\mathrm{d}x) \quad
\mbox{and}\quad \nu_-(h_-):=\int_{(h_-,1]}x\Pi^{(-)}(\mathrm{d}x).$$ Again, only their behaviour for small values of $h_\pm$ will be relevant. We can keep $ h_\pm\in(0,1)$. Note that $\nu(x)=\gamma-\nu_+(x)+\nu_-(x)$. In (\[decomp3\]), $(X_t^{(S,h_+,+)})_{t\ge0}$ is a compensated sum of small *positive* jumps, that is, $$\begin{aligned}
X_t^{(S,h_+,+)}&=& \mathrm{a.s.}\lim_{\varepsilon\downarrow0}
\biggl( \sum_{0<s\le t}\Delta X_s
1_{\{\varepsilon<\Delta X_s\le h_+\}} -t\int_{\varepsilon<x\le h_+} x\Pi(\mathrm{d}x) \biggr),\end{aligned}$$ $(X_t^{(S,h_-,-)})_{t\ge0}$ is a compensated sum of small *negative* jumps, that is, $$\begin{aligned}
&& X_t^{(S,h_-,-)}= \mathrm{a.s.}\lim_{\varepsilon\downarrow0}
\biggl( \sum_{0<s\le t}\Delta X_s
1_{\{-h_-\le\Delta X_s< -\varepsilon\}} -t\int_{-h_-\le x< -\varepsilon} x\Pi(\mathrm{d}x) \biggr),\end{aligned}$$ where the almost sure limits exist; and $(X_t^{(B,h_\pm, \pm)})_{t\ge0}$ are the processes of positive and negative big jumps, thus, $$\begin{aligned}
X_t^{(B,h_+,+)}&=& \sum_{0<s\le t} \Delta
X_s 1_{\{\Delta X_s>h_+\}
} \quad \mbox{and}\quad X_t^{(B,h_-,-)}= \sum
_{0<s\le t} \Delta X_s 1_{\{\Delta X_s<-h_-\}},\qquad t>0.\end{aligned}$$ Finally, $(Z_t)_{t\ge0}$ is a standard Brownian motion independent of the jump processes, all of which are independent from each other.
To motivate our approach, we quote part of a result due to Doney [@doney2004]. It gives an equivalence for $X$ to remain positive at small times, with probability approaching 1, in terms of the functions $A(x)$, $U(x)$ and the negative tail of $\Pi$. The condition reflects the positivity of $X$ at small times in that the function $A(x)$ remains positive for small values of $x$, and dominates $U(x)$ and the negative tail of $\Pi$ in a certain way.
\[th3.3\] Suppose $\Pi(\mathbb{R})=\infty$. Suppose also that $\overline{\Pi}^-(0+)>0$. Then $$\label{3.7}
\lim_{t\downarrow0}P(X_t> 0)=1$$ if and only if $$\label{3.6a}
\lim_{x\downarrow0} \frac{A(x)}{\sqrt{U(x)\overline{\Pi}^-(x)}}=\infty.$$
Suppose alternatively that $X$ is spectrally positive, that is, $\overline{\Pi}^-(x)=0$ for all $x>0$. Then (\[3.7\]) is equivalent to $$\label{3.8}
\sigma^{2}=0\quad \mbox{and}\quad A(x)\ge0 \qquad \mbox{for all small }x,$$ and this happens if and only if $X$ is a subordinator. Furthermore, we then have $A(x)\geq0$, not only for small $x$, but for all $x>0$.
\(i) Other equivalences for (\[3.7\]) are in Theorem 1 of Doney [@doney2004] (and his remark following the theorem). He assumes a priori that $\sigma^2=0$ but this is not necessary as it follows from the inequality: $$\label{4.1x}
\mathop{\limsup}_{x\downarrow0}\frac{A(x)}{\sqrt{\overline{\Pi
}^-(x)}}<\infty,$$ whenever $\overline{\Pi}^-(0+)>0$, which is proved in Buchmann, Fan and Maller [@bfm].
\(ii) When $\Pi(\mathbb{R})<\infty$, $X$ is compound Poisson and its behaviour near 0 is simply determined by the sign of the shift constant $\gamma$. We eliminate this case throughout.
The next section contains our main result which is essentially a subsequential version of Theorem \[th3.3\].
Staying positive near 0, subsequential version {#s2a}
==============================================
Denote the jump process of $X$ by $(\Delta X_t)_{t\ge0}$, where $\Delta X_t = X_t-X_{t-}$, $t>0$, with , and define $\Delta X_t^+=\max(\Delta X_t,0)$, $\Delta X_t^-=\max(-\Delta X_t,0)$, $(\Delta X^+)_t^{(1)}=\sup_{0<s\le t}\Delta X_s^+$, $(\Delta
X^-)_t^{(1)}=\sup_{0<s\le t}\Delta X_s^-$.
\[th3.3s\] Assume $\Pi(\mathbb{R})=\infty$.
Suppose $\overline{\Pi}^-(0+)>0$. Then the following are equivalent:\
there is a non-stochastic sequence $t_k\downarrow0$ such that $$\label{3.7s}
P(X_{t_k}>0)\to1; $$ there is a non-stochastic sequence $t_k\downarrow0$ such that $$\begin{aligned}
\label{Xdels}
\frac{X_{t_k}}{(\Delta X^-)_{t_k}^{(1)}} &\stackrel{\mathrm {P}} {\longrightarrow}& \infty\qquad
\mbox{as } k\to\infty;
\\
\label{3.6as}
\limsup_{x\downarrow0} \frac{A(x)}{\sqrt{U(x)\overline{\Pi}^-(x)}} &=& \infty.\end{aligned}$$
Suppose alternatively that $X$ is spectrally positive, that is, $\overline{\Pi}^-(x)=0$ for all $x>0$. Then (\[3.7s\]) is equivalent to $\lim_{t\downarrow0} P(X_t>0)\to1$, thus to (\[3.8\]), equivalently, $X_t$ is a subordinator, and $A(x)\geq0$ for all $x>0$.
Suppose $\overline{\Pi}^-(0+)>0$. Then $X_{t_k}/t_k\stackrel{\mathrm{P}}{\longrightarrow}\infty$ for a non-stochastic sequence $t_k\downarrow0$ if and only if $$\label{3.6as+}
\limsup_{x\downarrow0} \frac{A(x)}{1+\sqrt{U(x)\overline{\Pi}^-(x)}}=\infty.$$
\(i) When $\overline{\Pi}^-(0+)>0$, $\sup_{0<s\le t} \Delta X_s^->0$ a.s. for all $t>0$, so the ratio in (\[Xdels\]) is well defined.
\(ii) Sato [@S99], page 65, shows that $P(X_t\le x)$ is a continuous function of $x$ for all $t>0$ when $\Pi(\mathbb{R})=\infty$. So $P(X_t>0)=P(X_t\ge0)$ for all $t>0$ and $P(X_{t_k}>0)$ can be replaced by $P(X_{t_k}\ge0)$ in (\[3.7s\]) without changing the result (and similarly in Theorem \[th3.3\]).
\(iii) Assuming $\overline{\Pi}^+(0+)=\infty$ and $\overline{\Pi
}^-(0+)>0$, the contrapositive of (\[3.7s\]) shows that there is no sequence $t_k\downarrow0$ such that $P(X_{t_k}>0)\to1$, or, equivalently, $\liminf_{t\downarrow0}P(X_t\le0)>0$, if and only if $$\label{3.6ask}
\limsup_{x\downarrow0} \frac{A(x)}{\sqrt{U(x)\overline{\Pi
}^-(x)}}<\infty.$$ By a symmetrical argument, when $\overline{\Pi}^-(0+)=\infty$ and $\overline{\Pi}
^+(0+)>0$, then $\liminf_{t\downarrow0}P(X_t\ge 0)>0$ if and only if $$\label{3.6ask-}
\liminf_{x\downarrow0} \frac{A(x)}{\sqrt{U(x)\overline{\Pi}^+(x)}}>-\infty.$$ Combining these gives the following.
\[cor1\] Assume $\overline{\Pi}^+(0+)=\overline{\Pi}^-(0+)=\infty$. Then (\[bothpos\]) holds if and only if $$\label{3.6askb}
-\infty< \liminf_{x\downarrow0} \frac{A(x)}{\sqrt{U(x)\overline{\Pi}^+(x)}}\quad
\mbox{and} \quad \limsup_{x\downarrow0} \frac{A(x)}{\sqrt{U(x)\overline{\Pi}^-(x)}}<\infty.$$
When one of $\overline{\Pi}^+(0+)$ or $\overline{\Pi}^-(0+)$ is infinite but the other is zero, conditions for (\[bothpos\]) can also be read from Theorem \[th3.3s\].
\(ii) A random walk version of Theorem \[th3.3s\] is in Kesten and Maller [@KM]. Andrew [@PA], Theorem 4, has results related to Theorem \[th3.3s\], including the equivalence of (\[3.7s\]) and (\[Xdels\]).
Some inequalities for the distribution of $X$ {#subs2.5a}
=============================================
For the proof of Theorem \[th3.3s\], some lemmas are needed. The first gives a non-uniform Berry–Esseen bound for a small jump component of $X$. The proof is rather similar to that of Lemma 4.3 of Bertoin, Doney and Maller [@bdm], so we omit details.
\[ron2lem1\] Fix $h_-\ge0$, $h_+\ge0$, $h_-\vee h_+>0$. Let $(X_t^{(-h_-,h_+)})_{t\ge0}$ be the small jump martingale obtained from $X$ as the compensated sum of jumps with magnitudes in $(-h_-,h_+)$: $$\begin{aligned}
&&
X_t^{(-h_-,h_+)}= \mathrm{a.s.}\lim_{\varepsilon\downarrow0}
\biggl( \sum_{0<s\le t}\Delta X_s
1_{\{ \Delta X_s\in(-h_-,-\varepsilon)\cup(\varepsilon,h_+) \}} -t\int_{x\in(-h_-,-\varepsilon)\cup(\varepsilon,h_+)}x\Pi(\mathrm {d}x) \biggr).\end{aligned}$$ (Interpret integrals over intervals of the form $(0,-\varepsilon)$, and $(\varepsilon,0)$, $\varepsilon>0$, as 0.) Define absolute moments $m_k^{(-h_-,h_+)}:=\int_{-h_-<x<h_+}|x|^k\Pi(\mathrm{d}x)$, $k=2,3,\ldots,$ and assume $\sigma^2+m_2^{(-h_-,h_+)}>0$. Then we have the non-uniform bound: for any $x\in\mathbb{R}$, $t>0$, $$\begin{aligned}
\label{be1}
&& \biggl{\vert}P \biggl( \frac{\sigma Z_t+X_t^{(-h_-,h_+)}}{\sqrt{t(\sigma
^2+m_2^{(-h_-,h_+)})}}\le x \biggr)-\Phi(x)\biggr{\vert}\leq
\frac{C m_3^{(-h_-,h_+)} }{\sqrt{t}( \sigma^2+m_2^{(-h_-,h_+)}
)^{3/2}(1+|x|)^{3}},\end{aligned}$$ where $C$ is an absolute constant and $\Phi(x)$ is the standard normal c.d.f.
Next, we use Lemma \[ron2lem1\] to develop other useful bounds. Define $$\label{Vpmdef}
V_+(x)=\int_{0<y\le x} y^2 \Pi(\mathrm{d}y) \quad \mbox{and}\quad V_-(x)=\int_{-x\le y<0} y^2
\Pi(\mathrm{d}y), \qquad x>0.$$ In the next lemma, the “$+$” and “$-$” signs are to be taken together. When $\overline{\Pi}^+(0+)=0$ we have $V_+\equiv0$, and interpret $(X_t^{(S,d_+,+)})_{t\ge0}$ as 0; similarly with “$-$” replacing “$+$” when $\overline{\Pi}^-(0+)=0$.
\[dolem11\] Suppose $d_\pm>0$, $\kappa_\pm>0$ and $K_\pm$ are constants satisfying $$\label{Kpm}
K_\pm\ge4C\max \biggl(\frac{\kappa_\pm}{\Phi(-\kappa_\pm)},
\frac{1}{\Phi(-\kappa_\pm)\sqrt{1-\Phi(-\kappa_\pm)/2}} \biggr),$$ where $C$ is the absolute constant in (\[be1\]). Then for each $t>0$ $$\label{an}
P \bigl(X_t^{(S,d_\pm,\pm)}\le K_\pm
d_\pm-\kappa_\pm\sqrt{tV_\pm (d_\pm)}
\bigr) \ge\Phi(-\kappa_\pm)/2.$$
Suppose, for each $t>0$, $d_\pm=d_\pm(t)>0$ satisfy $$\label{Pilb}
t\overline{\Pi}^+(d_+)\le c_+ \quad \mbox{and}\quad t\overline{\Pi}^-(d_--)\ge
c_-$$ for some $c_+>0$, $c_->0$. Assume $\kappa_\pm>0$ and $K_\pm$ are constants satisfying (\[Kpm\]).
Suppose $\overline{\Pi}^+(0+)>0$. Then for each $t>0$ and $L\ge0$ $$\begin{aligned}
&& P \bigl(X_t\le t\gamma-t\nu_+(d_+)+t\nu_-(d_-)
+K_+d_+-Ld_- -\kappa_+
\sqrt{tV_+(d_+)} -\kappa_-\sqrt{tV_-(d_-)} \bigr)\qquad
\nonumber
\\[-8pt]
\label{Xlb2}
\\[-8pt]
\nonumber
&&\quad\ge e^{-c_+}\Phi(-\kappa_+) \Phi(-\kappa_-)P\bigl(N(c_-)\ge K_-+L
\bigr)/8,\end{aligned}$$ where $N(c_-)$ is a Poisson rv with expectation $c_-$.
When $\overline{\Pi}^+(0+)=0$, (\[Xlb2\]) remains true with $\nu_+(d_+)=V_+(d_+)=d_+=c_+=0$.
Suppose $0\le\overline{\Pi}^-(0+)<\infty=\overline{\Pi}^+(0+)$ and, for $t>0$, $d_+=d_+(t)>0$ is such that $t\overline{\Pi}^+(d_+(t))\le c_+$. Suppose $\kappa_+>0$ and $K_+$ are constants satisfying (\[Kpm\]). Then $$\label{Xlb3}
P \bigl(X_t\le t\gamma-t\nu_+(d_+)+t\nu_-(0) +K_+d_+-
\kappa_+\sqrt{tV_+(d_+)} \bigr) \ge e^{-c_+} \Phi(-\kappa_+)/4,$$ where $\nu_-(0)\equiv0$ when $\overline{\Pi}^-(0+)=0$.
\(i) We give the proof just for the “$+$” signs. Fix $t>0$ and take any constants $d_+>0$, $\kappa_+>0$ and $K_+$, with $\kappa_+$ and $K_+$ satisfying (\[Kpm\]).
Assume $V_+(d_+)>0$. Apply the bound (\[be1\]) in Lemma \[ron2lem1\] to $X_t^{(S,d_+,+)} $, which has Lévy measure $\Pi$ restricted to $(0,d_+)$. Noting that $\int_{0 <y\le x}y^3\Pi(\mathrm{d}y) \le xV_+(x)$, $x>0$, (\[be1\]) then gives, for each $t>0$, $$\label{norbd}
\sup_{x\in\mathbb{R}}\bigl{\vert}P \bigl(X_t^{(S,d_+,+)}
\le x\sqrt{tV_+(d_+) } \bigr) -\Phi(x)\bigr{\vert}\leq
\frac{Cd_+}{\sqrt{tV_+(d_+)}}.$$ Substitute $x=-\kappa_+$ in this to get $$P \bigl(X_t^{(S,d_+,+)}\le-\kappa_+\sqrt{tV_+(d_+)} \bigr) \ge
\Phi(-\kappa_+) -\frac{Cd_+}{\sqrt{ tV_+(d_+)}}.$$ When $2Cd_+\le\Phi(-\kappa_+)\sqrt{ tV_+(d_+)}$, this inequality implies $$\label{ac0}
P \bigl(X_t^{(S,d_+,+)}\le-\kappa_+
\sqrt{tV_+(d_+)} \bigr) \ge\tfrac{1}{2} \Phi(-\kappa_+).$$ When $2Cd_+> \Phi(-\kappa_+)\sqrt{ tV_+(d_+)}$, we have $$\begin{aligned}
2\kappa_+\sqrt{tV_+(d_+)}< 4Cd_+\kappa_+/\Phi(-\kappa_+)\le K_+d_+,\end{aligned}$$ since $K_+$ satisfies (\[Kpm\]). Apply Chebychev’s inequality, noting that $X_t^{(S,d_+,+)}$ has mean 0 and variance $tV_+(d_+)$, to get $$\begin{aligned}
P \bigl(X_t^{(S,d_+,+)}\le K_+d_+-\kappa_+\sqrt{tV_+(d_+)}
\bigr) &\ge& 1-\frac{tV_+(d_+)}{ (K_+d_+-\kappa_+\sqrt
{tV_+(d_+)} )^2}
\nonumber
\\
&\ge& 1- \frac{4tV_+(d_+)}{K_+^2d_+^2}.\end{aligned}$$ Also when $2Cd_+>\Phi(-\kappa_+)\sqrt{ tV_+(d_+)}$, by choice of $K_+$ in (\[Kpm\]) we have $$\frac{4tV_+(d_+)}{K_+^2d_+^2} \le\frac{16C^2}{\Phi^2(-\kappa_+)K_+^2} \le1-\frac{\Phi(-\kappa_+)}{2},$$ giving $$\label{ac2}
P \bigl(X_t^{(S,d_+,+)}\le K_+d_+-\kappa_+
\sqrt{tV_+(d_+)} \bigr) \ge\tfrac{1}{2} \Phi(-\kappa_+).$$ The same inequality holds when $2Cd_+\le\Phi(-\kappa_+)\sqrt{
tV_+(d_+)}$, by (\[ac0\]), so it holds in general.
When $V_+(d_+)=0$, $\Pi(\cdot)$ has no mass in $(0,d_+)$, and (\[an\]) with a “$+$” sign remains valid in the sense that $X_t^{(S,d_+,+)}=0$ a.s. and the left-hand side of (\[an\]) equals 1. This proves (\[an\]) with a “$+$” sign, and the same argument goes through with “$-$” in place of “$+$”.
\(ii) We use the Itô representation in (\[decomp3\]). Fix $t>0$ and take any constants $d_\pm>0$ satisfying (\[Pilb\]). Let $\kappa_\pm>0$ be any constants and choose $K_\pm$ to satisfy (\[Kpm\]). For the small jump processes, we have the bounds in (\[an\]). Note that these remain true if $\overline{\Pi}^+(0+)=0$ or $\overline
{\Pi}^-(0+)=0$. For the big positive jumps, we have $$\begin{aligned}
P\bigl(X_t^{(B,d_+,+)}=0\bigr) &\ge& P(\mbox{no }
\Delta X_s \mbox{ exceeds } d_+\mbox{ up\ till time } t)
\nonumber
\\
\label{sldiff}
&=& e^{-t\overline{\Pi}^+(d_+)}
\\
&\ge & e^{-c_+} \qquad \bigl(\mbox{by }(\ref{Pilb})\bigr).\nonumber\end{aligned}$$ Equation (\[sldiff\]) remains true with $c_+=0$ when $\overline{\Pi}^+(0+)=0$. By (\[decomp3\]), the probability on the left-hand side of (\[Xlb2\]) is, for any $L\ge0$, $$\begin{aligned}
&& P\bigl(\sigma Z_t+ X_t^{(S,d_+,+)}+X_t^{(B,d_+,+)}+
X_t^{(S,d_-,-)}+X_t^{(B,d_-,-)}
\nonumber
\\
&& \qquad {}\le K_+d_+ -Ld_- -\kappa_+\sqrt{tV_+(d_+)} -\kappa_-\sqrt{tV_-(d_-)} \bigr)
\nonumber
\\
\label{ac4}
&&\quad\ge P \bigl(Z_t\le0, X_t^{(S,d_+,+)}\le K_+d_+-
\kappa_+\sqrt{tV_+(d_+)},\ X_t^{(B,d_+,+)}=0,
\\
&&\qquad X_t^{(S,d_-,-)}\le K_-d_--\kappa_-\sqrt{tV_-(d_-)},\
X_t^{(B,d_-,-)}\le-(K_-+L)d_- \bigr)
\nonumber
\\
&&\quad \ge e^{-c_+} \Phi(-\kappa_+) \Phi(-\kappa_-) P \bigl(
X_t^{(B,d_-,-)}\le-(K_-+L)d_- \bigr)/8.\nonumber\end{aligned}$$ In the last inequality, we used (\[an\]) (twice; once with “$+$” and once with “$-$”), (\[sldiff\]) and the independence of the $Z_t$ and the $X_t^{(\cdot)}$ processes. No jump in $X_t^{(B,d_-,-)}$ is larger than $-d_-$, so we have the upper bound $X_t^{(B,d_-,-)}\le-d_-N_t^-(d_-)$, where $N_t^-(d_-)$ is the number of jumps of $X_t$ less than or equal in size to $-d_-$ which occur by time $t$. $N_t^-(d_-)$ is distributed as Poisson with expectation $t\overline{\Pi}^-(d_--)$, and $t\overline{\Pi}^-(d_--)\ge c_-$ by (\[Pilb\]). (Note that this implies $\overline{\Pi}^-(0+)>0$.) The Poisson distribution is stochastically monotone in the sense that if $N(\mu_1)$ and $N(\mu_2)$ are Poisson rvs with means $\mu_1>\mu_2$, then $P(N(\mu_1)\ge x) \ge P(N(\mu_2)\ge x)$ for all $x\ge0$. So, letting $N(c_-)$ be a Poisson rv with expectation $c_-$, we have $$\label{nc}
P\bigl(N_t^-(d_-)\ge K_-+L\bigr) \ge P\bigl(N(c_-)\ge
K_-+L\bigr).$$ Then using $$\label{ncn}
P \bigl( X_t^{(B,d_-,-)}\le-(K_-+L)d_- \bigr) \ge
P \bigl(N_t^-(d_-)\ge K_-+L \bigr)$$ and (\[ac4\]) we arrive at (\[Xlb2\]). When $\overline{\Pi}^+(0+)=0$, we can take all the “$+$” terms in (\[ac4\]) as 0 to get (\[Xlb2\]) with all the “$+$” terms 0.
\(iii) Assume $0\le\overline{\Pi}^-(0+)<\infty=\overline{\Pi}^+(0+)$. In this case, we do not define $d_-$ but still have $d_+=d_+(t)>0$ and assume $t\overline{\Pi}^+(d_+)\le c_+$ as in (\[Pilb\]). From (\[decomp3\]), write $$\label{decomp5}
X_t=t\gamma-t\nu_+(d_+)+t\nu_-(0)
+X_t^{(S,d_+,+)} +X_t^{(B,d_+,+)} +
X_t^{(0,-)},$$ where the negative jump components have been amalgamated into $$\begin{aligned}
&& X_t^{(0,-)}:= \sum_{0<s\le t} \Delta
X_s 1_{\{\Delta X_s\le0\}}, \qquad t>0,\end{aligned}$$ which is a compound Poisson process comprised of non-positive jumps. This term and the term $t\nu_-(0)$ are absent from (\[decomp5\]) when $\overline{\Pi}^-(0+)=0$. Using (\[an\]), (\[sldiff\]) and (\[decomp5\]), write $$\begin{aligned}
&& P \bigl(X_t\le t\gamma-t\nu_+(d_+)+t\nu_-(0+) +K_+d_+-
\kappa_+\sqrt{tV_+(d_+)} \bigr)
\nonumber
\\
&&\quad\ge P \bigl(Z_t\le0, X_t^{(S,d_+,+)}\le K_+d_+-
\kappa_+\sqrt{tV_+(d_+)}, X_t^{(B,d_+,+)}=0, X_t^{(0,-)}
\le0 \bigr)
\nonumber
\\
&&\quad\ge e^{-c_+}\Phi(-\kappa_+) P \bigl( X_t^{(0,-)}
\le0 \bigr)/4= e^{-c_+} \Phi(-\kappa_+)/4\end{aligned}$$ and this gives (\[Xlb3\]).
Proof of Theorem 2.1 {#s2ac}
====================
*Part* (i). Assume $\overline{\Pi}^-(0+)>0$ throughout this part.
(\[3.6as\]) $\Longrightarrow$ (\[3.7s\]): Assume (\[3.6as\]). $\overline{\Pi}^-(0+)>0$ implies $\overline{\Pi}^-(x)>0$ in a neighbourhood of 0 so we can assume $\overline{\Pi}^-(x)>0$ for all $0<x<1$. Choose $1>x_k\downarrow0$ such that $$\frac{A(x_k)}{\sqrt{U(x_k)\overline{\Pi}^-(x_k)}}\to\infty$$ as $k\to\infty$. This implies $\sigma^2=0$ by (\[4.1x\]) (because $U(x)\ge\sigma^2$). It also means that $A(x_k)>0$ for all large $k$, and without loss of generality we may assume it to be so for all $k$. Let $$s_k:= \sqrt{\frac{U(x_k)}{\overline{\Pi}^-(x_k)A^2(x_k)}},$$ then $$s_k\overline{\Pi}^-(x_k) = \frac{\sqrt{U(x_k)\overline{\Pi}^-(x_k)}}{A(x_k)}\to0$$ and since $\overline{\Pi}^-(0+)>0$, also $s_k\to0$ as $k\to\infty$. In addition, we have $$\frac{U(x_k)}{s_kA^2(x_k)} = \frac{\sqrt{U(x_k)\overline{\Pi}^-(x_k)}}{A(x_k)}\to0$$ and $$\frac{s_kA(x_k)}{x_k}= \sqrt{\frac{U(x_k)}{x_k^2\overline{\Pi}^-(x_k)}}\ge1.$$ Set $$t_k:= \sqrt{\frac{s_k}{\overline{\Pi}^-(x_k)}},$$ so $t_k/s_k\to\infty$, but still $t_k\overline{\Pi}^-(x_k)\to0$, as $k\to\infty$. Then $$\label{Ut0}
\frac{U(x_k)}{t_kA^2(x_k)} =\frac{s_k}{t_k}\frac{U(x_k)}{s_kA^2(x_k)} \to0,$$ and $$\label{Ati}
\frac{t_kA(x_k)}{x_k} =\frac{t_k}{s_k}\frac{s_kA(x_k)}{x_k} \to
\infty,$$ as $k\to\infty$.
Recall (\[IP\]) and use the Itô decomposition in (\[decomp3\]) with $\sigma^2=0$ and $h_+=h_-=h>0$ to write $$\label{dec2} X_t=tA(h)+X_t^{(S,h)}+X_t^{(B,h,+)}-th
\overline{\Pi}^+(h)+ X_t^{(B,h,-)}+th\overline{\Pi}^-(h), \qquad
t>0.$$ Here, $X_t^{(S,h)}=X_t^{(S,h,+)}+X_t^{(S,h,-)}$ is the compensated small jump process, and $X_t^{(B,h,\pm)}$ are the positive and negative big jump processes.
Suppose $\overline{\Pi}^+(0+)>0$. Since each jump in $X_t^{(B,h,+)}$ is at least $h$, we have the lower bound $X_t^{(B,h,+)}\geq hN_t^+(h)$, where $N_t^+(h)$ is Poisson with expectation $t\overline{\Pi}^+(h)$ (and variance $t\overline{\Pi}^+(h)$). Using this and substituting in (\[dec2\]) with $t=t_k$ and $h=x_k$ we get $$\label{dec4} X_{t_k} \ge t_kA(x_k)
+X_{t_k}^{(S,x_k)}+x_k \bigl(N^+_{t_k}(x_k)-t_k
\overline{\Pi }^+(x_k) \bigr) + X_{t_k}^{(B,x_k,-)}.$$ Since $t_k\overline{\Pi}^-(x_k)\to0$, we have $P(X_{t_k}^{(B,x_k,-)}=0)\to1$ as $k\to\infty$. Also, for $\varepsilon\in(0,1)$, $$\begin{aligned}
P \bigl( X_{t_k}^{(S,x_k)}+x_k \bigl(N^+_{t_k}(x_k)-t_k
\overline{\Pi }^+(x_k) \bigr) \le-\varepsilon t_kA(x_k)
\bigr) &\le& \frac{t_kV(x_k)+t_kx_k^2\overline{\Pi}^+(x_k)}{\varepsilon^2t_k^2A^2(x_k)}
\nonumber
\\
&\le& \frac{U(x_k)}{\varepsilon^2t_kA^2(x_k)} \to0,\end{aligned}$$ as $k\to\infty$ by (\[Ut0\]), so $$\label{Xot}
P \biggl(\frac{X_{t_k}}{x_k} \ge(1-\varepsilon)\frac
{t_kA(x_k)}{x_k}
\biggr) \to1,$$ and hence, by (\[Ati\]), $X_{t_k}/x_k\stackrel{\mathrm
{P}}{\longrightarrow}\infty$ as $k\to\infty$. Thus, (\[3.7s\]) holds.
Alternatively, if $\overline{\Pi}^+(0+)=0$, we can omit the term containing $N^+_{t_k}(x_k)-t_k\overline{\Pi}^+(x_k)$ in (\[dec4\]) and in what follows it, and again obtain (\[Xot\]), and hence (\[3.7s\]).[^1]
(\[3.6as\]) $\Longrightarrow$ (\[Xdels\]): Continuing the previous argument, $t_k\overline{\Pi}^-(x_k)\to0$ implies $$P \bigl( \bigl(\Delta X^-\bigr)_{t_k}^{(1)} >x_k
\bigr) = P \Bigl( \sup_{0<s\le t_k}\Delta X_s^-
>x_k \Bigr)= 1-e^{-t_k \overline{\Pi}^-(x_k)} \to0 \qquad \mbox{as } k\to\infty,$$ so, using (\[Xot\]), (\[Xdels\]) also holds when (\[3.6as\]) holds and $\overline{\Pi}^-(0+)>0$.
(\[Xdels\]) $\Longrightarrow$ (\[3.7s\]): This is obvious when $\overline{\Pi}^-(0+)>0$.
(\[3.7s\]) $\Longrightarrow$ (\[3.6as\]): Assume $\Pi(\mathbb{R})=\infty$ as well as $\overline{\Pi}^-(0+)>0$, and that (\[3.7s\]) holds. Suppose (\[3.6as\]) fails, so we can choose $1<a<\infty$, $x_0>0$, such that $$\label{Ab}
A(x) \le a\sqrt{U(x)\overline{\Pi}^-(x)},$$ for all $0<x\le x_0$. We will obtain a contradiction. Note that (\[3.7s\]) implies $\sigma^2=0$, because $X_t/\sqrt
{t}\stackrel{\mathrm{D}}{\longrightarrow}N(0,\sigma^2)$, a non-degenerate normal rv, when $\sigma^2>0$. So we assume $\sigma^2=0$ in what follows. We consider 3 cases.
Assume in fact that $\overline{\Pi
}^-(0+)=\infty
=\overline{\Pi}^+(0+)$. In this situation, we can introduce quantile versions for the $d_\pm$ in (\[Pilb\]). Define the non-decreasing function $$\label{ddef}
d_+(t):=\inf\bigl\{x>0: \overline{\Pi}^+(x)\le t^{-1}
\bigr\},\qquad t>0,$$ and set $d_+(0)=0$. Since $\overline{\Pi}^+(0+)=\infty$, we have $0<d_+(t)<\infty$ for all $t>0$, $d_+(t)\downarrow0$ as $t\downarrow0$, and $$\label{dprop}
t\overline{\Pi}^+\bigl(d_+(t)\bigr) \le1\le t\overline{\Pi}^+
\bigl(d_+(t)-\bigr)\qquad \mbox{for all } t>0.$$ Analogously, define $d_-(0)=0$, and $$\label{d-def} d_-(t):=\inf\bigl\{x>0: \overline{\Pi}^-(x)\le t^{-1}
\bigr\}, \qquad t>0,$$ having, since $\overline{\Pi}^-(0+)=\infty$, $0<d_-(t)<\infty$, $d_-(t)\downarrow0$ as $t\downarrow0$, and $$\label{d-prop}
t\overline{\Pi}^-\bigl(d_-(t)\bigr) \le1 \le t\overline{\Pi}^-
\bigl(d_-(t)-\bigr).$$ With $a$ as in (\[Ab\]), set $\kappa_+=\kappa_-=\kappa=2a$, then choose $K_\pm$ to satisfy (\[Kpm\]). Then (\[3.7s\]) together with (\[Xlb2\]) shows that we must have $$\label{Ac}
0\le t_k \bigl(\gamma -\nu_+(d_+)+\nu_-(d_-) \bigr)
+K_+d_+-Ld_- -\kappa\sqrt{t_k\bigl(V_+(d_+)+ V_-(d_-)\bigr)},$$ for all large $k$. Here, $d_+$ and $d_-$ are any positive numbers and we used the inequality $\sqrt{a} +\sqrt{b} \ge\sqrt{a+b}$, $a,b>0$, in (\[Xlb2\]). Take $\lambda>0$ and set $$d_+= d_+(\lambda t_k)\quad \mbox{and}\quad d_-=d_-(t_k),$$ where $d_+(\cdot)$ and $d_-(\cdot)$ are defined in (\[ddef\]) and (\[d-def\]). By (\[dprop\]) and (\[d-prop\]), we then have $$t\overline{\Pi}^+ \bigl(d_+(\lambda t)\bigr)\le\lambda^{-1}\quad\mbox{and} \quad t
\overline{\Pi}^-\bigl(d_-(t)-\bigr)\ge1,$$ so (\[Pilb\]) holds with $c_+=\lambda^{-1}$ and $c_-=1$. With $t_k$ as the sequence in (\[3.7s\]), let $d=d(t_k):=\max(d_+(\lambda
t_k),d_-(t_k))$.
Equation (\[Ac\]) implies $$\begin{aligned}
&& t_k \biggl(\gamma-\int_{(d_+,1]}y\Pi(\mathrm{d}y)+\int_{(d_-,1]}y\Pi ^{(-)}(\mathrm{d}y)
\biggr)
\nonumber
\\[-8pt]
\label{Ac0}
\\[-8pt]
\nonumber
&&\quad\ge -K_+d_+ + Ld_- +\kappa\sqrt{t_k\bigl(V_+(d_+)+ V_-(d_-)
\bigr)}.\end{aligned}$$ Adding the quantity $$t_k \biggl(\int_{(d_+,d]}y\Pi(\mathrm{d}y)-\int
_{(d_-,d]}y\Pi ^{(-)}(\mathrm{d} y) \biggr)$$ to both sides of (\[Ac0\]) gives $t_k\nu(d)$ on the left, and a quantity no smaller than $$t_kd_+ \bigl(\overline{\Pi}^+(d_+)-\overline{\Pi}^+(d) \bigr)
-t_kd \bigl(\overline{\Pi}^-(d_-)-\overline{\Pi}^-(d) \bigr) -K_+d_+
+ Ld_- +\kappa\sqrt{t_k\bigl(V_+(d_+)+ V_-(d_-)\bigr)}$$ on the right. Further adding $t_kd(\overline{\Pi}^+(d)-\overline{\Pi}^-(d))$ to both sides gives $t_kA(d)$ on the left (see (\[IP\])), and then after some cancellation we arrive at $$\label{Ac1} t_kA(d)\ge t_kd_+\overline{
\Pi}^+(d_+)-t_kd\overline{\Pi}^-(d_-) -K_+d_+ + Ld_- + \kappa
\sqrt{t_k\bigl(V_+(d_+)+ V_-(d_-)\bigr)}.$$
At this stage, it is helpful to assume that $\overline{\Pi}^+(x)$ is a continuous function on $(0,\infty)$. It then follows from (\[dprop\]) that $t_k\overline{\Pi}^+(d_+(\lambda t_k))=1/\lambda$, while $t_k\overline{\Pi}^-(d_-(t_k))\le1$ by (\[d-prop\]). Also, $d\le d_++d_-$. Thus, we deduce $$\label{Ad}
t_kA(d)\ge (1/\lambda-K_+- 1 )d_+ + (L- 1 )d_- +
\kappa\sqrt{t_k\bigl(V_+(d_+)+ V_-(d_-)\bigr)}.$$ Next, write $$\begin{aligned}
V_+(d_+)+ V_-(d_-) &=& V_+(d)-\int_{(d_+,d]}y^2
\Pi(\mathrm{d}y) +V_-(d)-\int_{(d_-,d]}y^2
\Pi^{(-)}(\mathrm{d}y)
\nonumber
\\
\label{Vest}
&\ge& V(d)-d^2 \bigl(\overline{\Pi}^+(d_+)-\overline{\Pi}^+(d)
\bigr) -d^2 \bigl(\overline{\Pi}^-(d_-)-\overline{\Pi}^-(d) \bigr)
\\
&=& U(d)-d^2 \bigl(\overline{\Pi}^+(d_+)+\overline{\Pi}^-(d_-)
\bigr).\nonumber\end{aligned}$$ So $$t_k \bigl( V_+(d_+)+ V_-(d_-) \bigr) \ge t_kU(d)-d^2
(1/\lambda+1 )$$ giving $$\sqrt{t_k\bigl(V_+(d_+)+ V_-(d_-)\bigr)} \ge \sqrt{t_kU(d)}-d
(1/\sqrt{\lambda}+1 ).$$ Substituting into (\[Ad\]), we obtain $$\begin{aligned}
t_kA(d) &\ge& \kappa\sqrt{t_kU(d)} + (1/\lambda-K_+- 1 -
\kappa/\sqrt{\lambda}-\kappa )d_+
\nonumber
\\
&& {}+ (L-1-\kappa/\sqrt{\lambda}-\kappa )d_-.\end{aligned}$$ Choose $\lambda$ small enough for the first expression in brackets on the right-hand side to be positive. Then choose $L$ large enough for the second expression in brackets on the right-hand side to be positive. This gives $$\label{Ag}
t_kA(d)\ge\kappa\sqrt{t_kU(d)},$$ for all large $k$, where $d=d(t_k)= \max(d_+(\lambda
t_k),d_-(t_k))\downarrow0$ as $k\to\infty$. Inequality (\[Ag\]) implies $$\label{bn}
\frac{A(d)}{\sqrt{U(d)\overline{\Pi}^-(d)}} \ge \frac{\kappa}{\sqrt{t_k\overline{\Pi}^-(d)}} \ge \frac{\kappa}{\sqrt{t_k\overline{\Pi}^-(d_-)}}
\ge\kappa,$$ giving a contradiction with (\[Ab\]), since $\kappa=2a$.
This proves (\[3.6as\]) from (\[3.7s\]) in case $\overline{\Pi}
^+(0+)=\overline{\Pi}^-(0+)=\infty$ and $\overline{\Pi}^+(x)$ is continuous for $x>0$. To complete the proof of part (i), case (a), of the theorem we remove the assumption of continuity made in deriving (\[Ad\]). This can be done using the following lemma.
\[lemexvaguenicepis\] Let $\Pi$ be any Lévy measure with $\overline{\Pi}^+(0+)=\infty$. Then there exists a sequence of Lévy measures $\Pi_n$, absolutely continuous with respect to Lebesgue measure, and having strictly positive $C^\infty$-densities on $\mathbb
{R}_*:=\mathbb{R}
\setminus\{0\}$, satisfying $\overline{\Pi}_n^+(0+)=\infty$ and $\Pi_n\stackrel
{v}{\longrightarrow}\Pi$ as $n\to
\infty$.
($\stackrel{v}{\longrightarrow}$ refers to vague convergence in $\overline\mathbb{R}_*$; see, for example, Chapter 15 in Kallenberg [@Kall].) We extend $\Pi$ to a Borel measure on $\mathbb{R}$ by setting $\Pi(\{
0\}
):=0$. Assume $\Pi\neq0$, so $C:=\int x^2 \Pi(\mathrm{d}x)/(1+x^2)\in(0,\infty)$. Observe that $P(\mathrm{d}x):=x^2 \Pi(\mathrm{d}x)/C(1+x^2)$ defines a Borel probability measure on $\mathbb{R}$. For all $n\in\mathbb
{N}$, the convolved probability measure $P_n:=P\star N(0,1/n)$ admits a strictly positive $C^\infty$-Lebesgue density, when $N(0,1/n)$ is a normal rv with expectation 0 and variance $1/n$. Set $\Pi_n(\mathrm{d}x):= C(1+x^2)
P_n(\mathrm{d}x)/x^2$, $n=1,2,\ldots.$ It is easily verified that $(\Pi_n)_{n\in\mathbb{N}}$ is a sequence of L' evy measures with the desired properties.
Now to complete the proof of part (i), assume (\[3.7s\]), that is, $P(X_{t_k}\ge0)\to1$ for some $t_k\downarrow0$, for a general $X$ with L' evy measure $\Pi$. Using Lemma \[lemexvaguenicepis\], construct a sequence of approximating Lévy measures $\Pi_n$, converging vaguely to $\Pi$, such that their positive tails $\overline{\Pi}_n^+$ are continuous on $(0,\infty)$ with $\overline{\Pi}_n^+(0+)=\infty$. On the negative side, let $\overline{\Pi}_n^-(x)=\overline{\Pi
}^-(x)$, $x>0$. Let $(X_t(n))_{t\ge0}$ be Lévy processes with measures $\Pi_n$ and other characteristics the same as for $X$. Define $\nu_n$, $A_n$, $V_n$, $U_n$, $\nu_{n,\pm}$, $V_{n,\pm}$, as in (\[Adef\]), (\[Vdef\]), (\[nupmdef\]) and (\[Vpmdef\]), but with $\Pi_n$ replacing $\Pi
$. The subscript $n$ functions converge to the original functions at points of continuity of the latter. $X_t(n)$ has characteristic exponent given by (\[ce\]) with $\Pi_n$ replacing $\Pi$, so as $n\to\infty$ we have $X_t(n)\stackrel{\mathrm{D}}{\longrightarrow
}X_t$ for each $t>0$. Under assumption (\[3.7s\]), $\lim_{n\to\infty}P(X_{t_k}(n)>0)
=P(X_{t_k}>0)>1-\delta$ for arbitrary $\delta\in(0,1/2)$ and $k$ large enough. Thus, $P(X_{t_k}(n)>0)>1-2\delta$ for $n\ge n_0(k)$ and $k\ge k_0$. So (\[Ac\]) holds for the subscript $n$ quantities with probability $1-2\delta$. But (\[Ac\]) is deterministic so it holds in fact for the subscript $n$ quantities (with probability 1) whenever $n\ge n_0(k)$ and $k\ge k_0$. The proof using continuity of $\Pi_n$ then shows that (\[bn\]) holds with $A$, $U$ and $\overline{\Pi}^-$ replaced by $A_n$, $U_n$ and $\overline{\Pi}_n^-$. Then letting $n\to\infty$ shows that (\[bn\]) itself holds as stated for $k\ge k_0$. Again we get a contradiction, and thus complete the proof that (\[3.7s\]) implies (\[3.6as\]) for case (a).
Assume that $0\le\overline{\Pi
}^+(0+)<\infty
=\overline{\Pi}^-(0+)$. As in the proof for case (a), we take $\kappa_+=\kappa_-=\kappa>2a$, $K_\pm$ to satisfy (\[Kpm\]), define $d_-(t)>0$ by (\[d-def\]), and write $d_-=d_-(t)$ for $t>0$. But for $d_+$ we set $d_+(t)\equiv d_-(t)>0$. We take $c_+=1$ in (\[Pilb\]) as we may since $t\overline{\Pi}
^+(d_+)=t\overline{\Pi}^+(d_-)\le t\overline{\Pi}^+(0+)\to0$ as $t\downarrow0$. With this set-up, (\[Ac\]) is true (with $d_-$ replacing $d_+$) and we can follow the proof of case (a) through to get (\[Ac1\]) with $d$ and $d_+$ replaced by $d_-$; thus, $$\label{Ac10} t_kA(d_-)\ge t_kd_- \bigl(\overline{
\Pi}^+(d_-)-\overline{\Pi}^-(d_-) \bigr) -K_+d_- + Ld_- + \kappa
\sqrt{t_k\bigl(V_+(d_-)+V_-(d_-)\bigr)}.$$ Estimating $V_\pm$ along the lines of (\[Vest\]) we find the right-hand side of (\[Ac10\]) is not smaller than $$\begin{aligned}
&&\kappa\sqrt{t_kU(d_-)} + (L-K_+-1-\kappa )d_-.\end{aligned}$$ Choose $L$ large enough in this to get (\[Ag\]), and hence (\[bn\]) with $d_-$ in place of $d$, hence (\[3.6as\]) again.
Assume that $0<\overline{\Pi
}^-(0+)<\infty=\overline{\Pi}
^+(0+)$. Define $d_+(t)$ by (\[ddef\]), so we have (\[dprop\]). Then (\[Xlb3\]) with $c_+=1$ and $\kappa_+=\kappa=2a$, together with (\[3.7s\]), shows that we must have $$\label{Ac6}
0\le t_k \bigl(\gamma-\nu_+(d_+)+\nu_-(0)
\bigr)+K_+d_+ -\kappa\sqrt{t_kV_+(d_+)},$$ for all large $k$. Here again we write $d_+= d_+(\lambda t_k)$ for $\lambda>0$. Inequality (\[Ac6\]) implies $$\label{Ac7}
t_k \biggl(\gamma-\int_{(d_+,1]}y\Pi(\mathrm{d}y)+\int_{(0,1]}y\Pi ^{(-)}(\mathrm{d}y)
\biggr) \ge -K_+d_+ +\kappa\sqrt{t_kV_+(d_+)}.$$ Subtracting the quantity $$t_k\int_{(0,d_+]}y\Pi^{(-)}(\mathrm{d}y)\le
t_kd_+ \bigl(\overline {\Pi} ^-(0+)-\overline{\Pi}^-(d_+) \bigr)$$ from both sides of (\[Ac7\]) gives $t_k\nu(d_+)$ on the left, and a quantity no smaller than $$-t_kd_+ \bigl(\overline{\Pi}^-(0+)-\overline{\Pi}^-(d_+) \bigr)
-K_+d_+ +\kappa\sqrt{t_kV_+(d_+)}$$ on the right. Further adding $t_kd_+(\overline{\Pi}^+(d_+)-\overline
{\Pi}^-(d_+))$ to both sides gives (see (\[IP\])) $$\begin{aligned}
\label{Ac8} t_kA(d_+)\ge t_kd_+\overline{
\Pi}^+(d_+)-t_kd_+\overline{\Pi}^-(0+)-K_+d_+ + \kappa
\sqrt{t_kV_+(d_+)}.\end{aligned}$$
At this stage, as before, assume $\overline{\Pi}^+(x)$ is continuous. It then follows from (\[dprop\]) that $t_k\overline{\Pi}^+(d_+(\lambda t_k))=1/\lambda$, while $t_k\overline{\Pi}^-(0+)\le1$ for large $k$. Thus, from (\[Ac8\]) we deduce $$\label{Ad1}
t_kA(d_+)\ge (1/\lambda-K_+- 1 )d_+ +\kappa
\sqrt{t_kV_+(d_+)},$$ for large enough $k$. Further, $$\begin{aligned}
t_kV_+(d_+)&=& t_k \bigl(U(d_+)-V_-(d_+)-d_+^2
\overline{\Pi}^+(d_+)-d_+^2\overline {\Pi}^-(d_+) \bigr)
\cr
&\ge&
t_kU(d_+)-t_kd_+^2 \bigl(2\overline{
\Pi}^-(0+)+\overline{\Pi }^+(d_+) \bigr)
\cr
&\ge& t_kU(d_+)-4d_+^2,\end{aligned}$$ using that $V_-(d_+)\le d_+^2\overline{\Pi}^-(0+)$. So $$\sqrt{t_kV_+(d_+)}\ge\sqrt{t_kU(d_+)}-2d_+.$$ Substituting into (\[Ad1\]), we get $$\begin{aligned}
t_kA(d_+) \ge \kappa\sqrt{t_kU(d_+)}+ (1/\lambda-K_+-1-2
\kappa )d_+,\end{aligned}$$ for large $k$. We can choose $\lambda$ small enough for the expression in brackets on the right-hand side to be positive. This gives $$\label{Ac9}
t_kA(d_+)\ge\kappa\sqrt{t_kU(d_+)},$$ which, since $\overline{\Pi}^-(0+)>0$ is assumed, implies $$\begin{aligned}
&& \frac{A(d_+)}{\sqrt{U(d_+)\overline{\Pi}^-(d_+)}} \ge \frac{\kappa}{\sqrt{t_k\overline{\Pi}^-(0+)}} \to\infty \qquad \mbox{as } k\to\infty,\end{aligned}$$ a contradiction with (\[Ab\]). We can remove the continuity assumption as before. So (\[3.6as\]) is proved when $0<\overline{\Pi}^-(0+)<\infty
=\overline{\Pi}^+(0+)$.
Now we will deal with the case when $\overline{\Pi}^-(0+)=0$ but $\overline{\Pi}^+(0+)=\infty$. Again assume $\overline{\Pi}^+(x)$ is continuous. The working in case (c) is still valid from (\[Ac6\]) to (\[Ac9\]). The negative jump process is now absent from $X_t$ and (\[Ac6\]) gives, with $d_+=d_+(\lambda t)$, $$\begin{aligned}
0&\le& t_k \bigl(\gamma-\nu_+(d_+) \bigr)+K_+d_+\nonumber
\\
\label{Ac2}
&=& t_k \biggl(\gamma-\int_{d_+}^1
\overline{\Pi}^+(y)\,\mathrm{d}y -d_+\overline{\Pi}^+(d_+)+\overline{\Pi}^+(1)
\biggr)+K_+d_+
\\
&\le& t_k \biggl(\gamma-\int_{d_+}^1
\overline{\Pi}^+(y)\,\mathrm{d}y +\overline{\Pi}^+(1) \biggr)-(1/\lambda-K_+)d_+,\nonumber\end{aligned}$$ using $t_k\overline{\Pi}^+(d_+(\lambda t_k))= 1/\lambda$ in the last inequality (since $\overline{\Pi}^+(x)$ is continuous). Then choosing $\lambda<1/K_+$ we get $$\begin{aligned}
&&\int_{d_+}^1\overline{\Pi}^+(y)\,\mathrm{d}y \le
\gamma+\overline {\Pi}^+(1).\end{aligned}$$ Letting $k\to\infty$ (so $t_k\downarrow0$ and $d_+=d_+(\lambda
t_k)\downarrow
0$) shows that $\int_0^1\overline{\Pi}^+(y)\,\mathrm{d}y<\infty$. Since $X_t$ has no negative jumps, from this we deduce that $X$ is of bounded variation with drift $\mathrm{d}_X=A(0+)$ (see Doney and Maller [@dm1], Theorem 2.1 and Remark 1), which is non-negative by (\[Ac9\]). Thus, $X$ is a subordinator with non-negative drift. It follows that $$A(x)= \gamma+\overline{\Pi}^+(1)-\int_x^1
\overline{\Pi }^+(y)\,\mathrm{d}y = \mathrm{d}_X+\int
_0^x \overline{\Pi}^+(y)\,\mathrm{d}y$$ is non-negative for all $x>0$. This is proved assuming continuity of $\Pi_n$ but that assumption can be removed as before. Then $A(x)\ge0$ together with $\sigma^2=0$ implies $\lim_{t\downarrow
0}P(X_t>0)\to1$ as $t\downarrow0$ by Theorem \[th3.3\], hence (\[3.7s\]).
Finally, suppose $\overline{\Pi}^-(0+)>0$ and (\[3.6as+\]) holds. Then we can choose $x_k\downarrow0$ such that $$\frac{A(x_k)}{\sqrt{U(x_k)\overline{\Pi}^-(x_k)}}\to\infty \quad \mbox{and}\quad A(x_k) \to\infty,$$ as $k\to\infty$. Following exactly the proof of part (i), we get (\[Xot\]), and this implies $X_{t_k}/t_k\stackrel{\mathrm{P}}{\longrightarrow}\infty$, since $A(x_k)\to\infty$.
Conversely, suppose $X_{t_k}/t_k\stackrel{\mathrm{P}}{\longrightarrow}\infty$ as $k\to
\infty$ for a non-stochastic sequence $t_k\downarrow0$. Then $\lim_{k\to\infty}P(X_{t_k}^M>0)=1$ for every $M>0$, where $X_t^M$ is Lévy with triplet $(\gamma-M,\sigma
^2,\Pi)$. Consequently, (\[3.6as\]) holds with $A$, $U$, $\Pi$ replaced by $A^M(\cdot)=A(\cdot)-M$, $U^M=U$, $\Pi^M=\Pi$, and this modified version implies (\[3.6as+\]). This completes part (iii), and the proof of the theorem.
Acknowledgements {#acknowledgements .unnumbered}
================
I am grateful to a referee for a close reading of the paper and helpful suggestions, and to Boris Buchmann for supplying Lemma \[lemexvaguenicepis\].
Research partially supported by ARC Grant DP1092502.
[9]{}
(). . .
, (). . .
, (). . .
, ().
(). . .
(). . .
(). , ed. . : .
(). . .
(). . : .
(). . .
[^1]: Observe that the assumption $\Pi(\mathbb
{R})=\infty$ was not used in this part of the proof. The trivial case, $X_t=t\gamma$, $\gamma>0$, when $A(x)\equiv\gamma$, is included if we interpret (\[3.6as\]) as holding then.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Recently, Convolutional Neural Networks have shown promising results for 3D geometry prediction. They can make predictions from very little input data such as a single color image. A major limitation of such approaches is that they only predict a coarse resolution voxel grid, which does not capture the surface of the objects well. We propose a general framework, called hierarchical surface prediction (HSP), which facilitates prediction of high resolution voxel grids. The main insight is that it is sufficient to predict high resolution voxels around the predicted surfaces. The exterior and interior of the objects can be represented with coarse resolution voxels. Our approach is not dependent on a specific input type. We show results for geometry prediction from color images, depth images and shape completion from partial voxel grids. Our analysis shows that our high resolution predictions are more accurate than low resolution predictions.'
author:
-
bibliography:
- 'paper\_arxiv.bib'
title: Hierarchical Surface Prediction for 3D Object Reconstruction
---
Introduction
============
We live in a world composed of 3D objects bounded by 2D surfaces. Fundamentally, this means that we can either represent geometry implicitly as 3D volume or explicitly as 2D mesh surface which lives within the 3D space. When working with 3D geometry in practice for many tasks a specific representation is more suited than the other.
\
Recently, 3D prediction approaches which directly learn a function to map an input image to the output geometry have emerged [@choy20163d; @girdhar2016learning; @yan2016perspective]. The function is represented as Convolutional Neural Network (CNN) and the geometry is represented as voxel grid, which is a natural choice for CNNs. The advantage of such an approach is that it can represent arbitrary topology and allows for large shape variations. However, such approaches have a major limitation. Due to the cubic growth of the volume with increasing resolution only a coarse voxel grid is predicted. A common resolution is $32\times32\times32$. Using higher resolutions very quickly becomes computationally infeasible. Moreover, due to the 2D nature of the surface, the ratio between surface or close to the surface and non-surface voxels becomes smaller and smaller with increasing resolution. Voxels far from the surface are generally predicted correctly quite early in the training and hence do not induce any gradients which can be used to teach the system where the surface lies. Therefore, with increasing resolution also the work done on computing gradients which are uninformative increases. The underlying reason why this problem exists is that current systems do not exploit the fact that surfaces are only two dimensional. Can we build a system that does?
In this paper we introduce a general framework, hierarchical surface prediction (HSP), for high resolution 3D object reconstruction which is organized around the observation that only a few of the voxels are in the vicinity of the object’s surface the boundary between free and occupied space, while most of the voxels will be “boring”, either completely inside or outside the object. The basic principle therefore is to only predict voxels around the surface, which we can now afford to do at higher resolution. The key insight to enable our method to achieve this is to change the standard prediction of free and occupied space into a three label prediction with the labels *free space*, *boundary* and *occupied space*. Furthermore, we do not directly predict high resolution voxels. Instead, we hierarchically predict small blocks of voxels from coarse to fine resolution in an octree, which we call voxel block octree. Thereby we have at each resolution the signal of the *boundary* label which tells us that the descendants of the current voxel will contain both, *free space* and *occupied space*. Therefore, only nodes containing voxels with the *boundary* label assigned need higher resolution prediction. By hierarchically predicting only those voxels we build up our octree. This effectively reduces the computational cost and hence allows us to predict significantly higher resolution voxel grids than previous approaches. In our experiments we predict voxels at a resolution of $256\times256\times256$ (Fig. \[fig:teaser\]).
One of the fundamental questions which arises once higher resolution predictions are feasible, is whether or not a higher resolution prediction leads to more accurate results. In order to analyze this we compare our high resolution predictions to upsampled low resolution predictions in our quantitative evaluation and determine that our method outperforms these baselines. Our results are not only quantitatively more accurate, they are also qualitatively more detailed and have a higher surface quality.
The remainder of the paper is organized as follows. In Sec. \[sec:relWork\] we discuss the related work. We then introduce our framework in Sec. \[sec:formulation\]. Quantitative and qualitative evaluations are done in Sec. \[sec:exp\] and eventually we draw the conclusions in Sec. \[sec:conclusion\].
Related Work {#sec:relWork}
============
Traditionally dense 3D reconstruction from images has been done using a large collection of images. Geometry is extracted by dense matching or direct minimization of reprojection errors. The common methods can be broadly grouped as implicit volumetric reconstruction [@curless1996volumetric; @kolev2007continuous; @kutulakos2000theory; @lempitsky2007global; @zach2007globally] or explicit mesh based [@delaunoy2008minimizing; @gargallo2007minimizing] approaches. All these approaches facilitate a reconstruction of arbitrary geometry. However, in general a lot of input data is required to constrain the geometry enough. In difficult cases it is not always possible to recover the geometry of an object with multi-view stereo matching. This can be due to challenging materials such as transparent and reflective surfaces or lack of images from all around an object. For such cases object shape priors have been proposed [@dame2013dense; @hane2014class; @yingze2013dense].
All the methods mentioned above in general use multiple images as input data. Approaches which are primarily designed to reconstruct 3D geometry from a single color image were also studied. Vetter [@blanz1999morphable] propose to build a morphable shape model for faces, which allows for the reconstruction of faces from just a single image. In order to build such a model a lot of manual interaction and high quality scanning of faces is required. Other works propose to learn a shape model from silhouettes [@cashman2013shape] or silhouettes and keypoints [@kar2015category].
Recently, CNNs have also been used to predict geometry from single images. Choy [@choy20163d] propose to use an encoder decoder architecture in a recurrent network to facilitate prediction of coarse resolution voxel grids from a single or multiple color images. An alternative approach [@girdhar2016learning] proposes to first train an autoencoder on the coarse resolution voxel grids and then train a regressor which predicts the shape code from the input image. It has also been shown that CNNs can be leveraged to predict alternative primitive based representations [@abstractionTulsiani17]. While these approaches rely on ground truth shapes, CNNs can also be trained to predict 3D shapes using weaker supervision [@rezende2016unsupervised; @yan2016perspective; @tulsiani2017ray] for example image silhouettes. These types of supervisions utilize ray formulations which originally have been proposed for multi-view reconstruction [@liu2011complete; @savinov2016semantic].
Predicting geometry from color images has also been addressed in a 2.5D setting where the goal is, given a color image to predict a depth map [@eigen2014depth; @ladicky2014pulling; @saxena2005learningdepth]. These approaches can generally produce high resolution output but do only capture the geometry from a single viewpoint. Tatarchenko [@tatarchenko2016multi] propose to predict a collection of RGBD (color and depth) images from a single color image and fuse them in a post-processing step.
The volumetric approach has the major limitation that using high resolution voxel grids is computationally demanding and memory intensive. Multi-view approaches alleviate this, by the use of data adaptive discretization of the volume with Delaunay tedrahedrization [@labatut2007efficient], voxel block hashing [@niessner2013real] or octrees [@chen2013scalable; @steinbrucker2014volumetric]. Our work is inspired by these earlier works on octrees. A crucial difference is that in these works the structure of the octree can be determined from the input geometry but in our case we are looking at a more general problem where the structure of the tree needs to be predicted. Similarly, a very recent work [@riegler2016octnet] which proposes to use octrees in a CNN, also assumes that the structure of the octree is given as input. We predict the structure of the tree together with its content.
Our hierarchical surface prediction framework also relates to coarse-to-fine approaches in optical flow computation. Similar to volumetric 3D reconstruction also dense optical flow is computationally demanding due the the large label space. By computing optical flow hierarchically on a spatial pyramid, e.g. [@brox2004high], real-time computation of optical flow is feasible [@zach2007duality]. The benefits of using a coarse-to-fine approach in learning based optical flow prediction has also recently been pointed out [@ranjan2016optical; @ilg2016flownet].
Formulation {#sec:formulation}
===========
In this Section we describe our hierarchical surface prediction framework, which allows for high resolution voxel grid predictions. We first briefly discuss how the state-of-the-art, coarse resolution baseline works and then introduce the voxel block octree structure which we are predicting.
{width="0.95\linewidth"}
Voxel Prediction {#sec:voxPred}
----------------
The basic voxel prediction framework is adapted from [@choy20163d; @girdhar2016learning]. We consider an encoder/decoder architecture which takes an input $\mathcal{I}$, which in our experiments is either a color image, depth image or a partial low resolution voxel grid. A convolutional encoder encodes the input to a feature vector or shape code $\mathcal{C}$ which, in all our experiments, is a 128 dimensional vector. An up-convolutional decoder then decodes $\mathcal{C}$ into a predicted voxel grid $\mathcal{V}$. A labeling of the voxel space into free and occupied space can be determined using a suitable threshold $\tau$ (Sec. \[sec:quantEval\]) or alternatively a mesh can be extracted using marching cubes [@lorensen1987marching] directly on the predicted voxel occupancies at an iso value $\sigma$.
The main problem which prevents us from directly utilizing this formulation with high resolutions for $\mathcal{V}$ is that each voxel, even if it is far from the surface, needs to be represented and a prediction is made for it. Given that the surface area grows quadratically and the volume cubically with respect to edge division, the ratio of surface to non-surface voxels becomes smaller with increasing resolution.
Voxel Block Octree
------------------
In our hierarchical surface prediction method, we propose to predict a data structure with an up-convolutional decoder architecture, which we call ‘voxel block octree’. It is inspired from octree formulations [@chen2013scalable; @steinbrucker2014volumetric] used in traditional multi-view reconstruction approaches. The key insight which allows us to use such a data structure in a prediction framework is to extend the standard two label formulation to a three label formulation with labels *inside*, *boundary* and *outside*. As we will see later our data structure allows us to generate a complete voxel grid at high resolution while only making predictions around the surface. This leads to an approach which facilitates efficient training of an encoder/decoder architecture end-to-end.
An octree is a tree data structure which is used to partition the 3D space. The root node describes the cube of interest in the 3D space. Each internal node has up to 8 child nodes which describe the subdivision of the current node’s cube into the eight octants. Note that we slightly deviate from the standard definition of the octree where either none or all the 8 child nodes are present.
We consider a tree with levels $\ell \in \{1,\ldots,L\}$. Each node of the tree contains a ‘voxel block’ – a subdivision of the node’s 3D space into a voxel grid of size $b^3$. Each voxel in the voxel block at tree levels $\ell < L$ contains classifier responses for the three labels *occupied space*, *boundary* and *free space*. The nodes in layers $\ell \in \{1, \ldots, L-1 \}$, therefore, contain voxel blocks $\mathcal{B}^{\ell,s} \in [0,1]^{b^3 \times 3}$ where the index $s$ describes the position of the voxel block with respect to the global voxel grid. In the lowest level $L$ we do not need the boundary label, therefore it is composed of nodes which contain a voxel block $\mathcal{B}^{\ell,s} \in [0,1]^{b^3}$, where each element of the block describes the classifier response for the binary classification into free or occupied space. Note that each child node describes a space which has a cube side length of only half of the current node. By keeping the voxel block resolution fixed at $b^3$ we also divide the voxel side length by a factor of two. This means that the prediction in the child nodes are of higher resolution. The voxel block resolution $b$ is chosen such that it is big enough for efficient prediction using an up-convolutional decoder network and at the same time small enough such that local predictions around the surface can be made. In our experiments we use voxel blocks with $b=16$ and a tree with depth $L = 5$. A visualization of the voxel block octree is given in Fig. \[fig:treeVis\].
![Visualization of the voxel block octree. The (left) part depicts the space subdivision which is induced by the octree in the (middle). The (right) side is a 2D visualization of parts of the voxel blocks with a voxel block size of $b=4$.[]{data-label="fig:treeVis"}](figures/tree/spaceSubdivision.pdf "fig:"){width="0.31\linewidth"} ![Visualization of the voxel block octree. The (left) part depicts the space subdivision which is induced by the octree in the (middle). The (right) side is a 2D visualization of parts of the voxel blocks with a voxel block size of $b=4$.[]{data-label="fig:treeVis"}](figures/tree/octree.pdf "fig:"){width="0.31\linewidth"} ![Visualization of the voxel block octree. The (left) part depicts the space subdivision which is induced by the octree in the (middle). The (right) side is a 2D visualization of parts of the voxel blocks with a voxel block size of $b=4$.[]{data-label="fig:treeVis"}](figures/tree/voxelBlocks.pdf "fig:"){width="0.19\linewidth"}
The predictions stored at any given level of the tree might be sparse. In order to reconstruct a complete model in high resolution, we upsample all the voxels which have not been predicted on the highest resolution from the closest predicted resolution. As we will see later, the voxels which get upsampled are in the inside and outside of the object not directly next to the boundary. Therefore the resolution of the actual surface remains high. In order to be able to extract a smooth high quality surface using marching cubes it is crucial that all the voxels which are close to the surface are predicted at high resolution and the predicted values are smooth. Therefore we aim to always evaluate the highest resolution around the boundary. At this point we would like to note that we can also extract a model at intermediate resolution by considering the boundary label as occupied space. This choice is consistent with the choice of considering any voxel as occupied space which intersects the ground truth mesh during voxelization (Sec. \[sec:voxelization\]).
Network Architecture
--------------------
In the previous section we introduced our voxel block octree data structure. We will now describe the encoder / decoder architecture which we are using to predict voxel block octrees using CNNs. An overview of our architecture is depicted in Fig. \[fig:architecture\].
The encoder part is identical to the one of the basic voxel prediction pipeline from Sec. \[sec:voxPred\], the input $\mathcal{I}$ gets first transformed into a shape code $\mathcal{C}$ by a convolutional encoder network. An up-convolutional also called de-convolutional decoder architecture [@dosovitskiy2017learning; @zeiler2010deconvolutional] predicts the voxel block octree. In order to be able to predict the blocks $\mathcal{B}^{\ell,s}$ and at the same time also being able to capture the information which allows for predictions of the higher resolution levels $\ell+1, \ldots$, we introduce feature blocks $\mathcal{F}^{\ell,s} \in \mathbb{R}^{(b+2p)^3 \times c}$. The spatial extent of the feature blocks is bigger or equal to the ones of the voxels blocks to allow for a padding $p \geq 0$. As we will see later the padding allows for an overlap when predicting neighbouring octants which leads to smoother results. The fourth dimension is the number of feature channels which can be chosen depending on the application. In our experiments we use $c=32$ and $p=2$.
The most important part for the understanding of our decoder architecture is how we predict level $\ell+1$ given level $\ell$. In the following we will discuss this procedure in detail. It consist of three basic steps.
1. Feature Cropping
2. Upsampling
3. Output Generation
**Feature Cropping.** At this point we assume that we have given the feature block $\mathcal{F}^{\ell,s}$. The goal is to generate the output for the child node corresponding to a specific octant $\mathcal{O}$. The feature block $\mathcal{F}^{\ell,s}$ contains information to generate the output for all the child nodes. In order to only process the information relevant for the prediction of the child node corresponding to $\mathcal{O}$ we extract a $((b/2 + 2p)^3 \times c)$ region out of the four dimensional tensor $\mathcal{F}^{\ell,s}$ spatially centered around $\mathcal{O}$. An illustration of this process is given in Fig. \[fig:cropUpsampling\]. If neighboring octants are processed, the extracted feature channels will have some overlap, which helps to generate a smoother output (see Fig. \[fig:overlap\]).
**Upsampling.** The upsampling module takes input from the cropping module and predicts a new $(b+2p)^3$ feature block $\mathcal{F}^{\ell+1,r}$ via up-convolutional and convolutional layers.
**Output Generation.** The output network takes the prediction of the feature block $\mathcal{F}^{\ell+1,r}$ from the upsampling module and generates the voxel block $\mathcal{B}^{\ell+1,r}$. This is done using a sequence of convolutional layers. The supervision is given in form of the three ground truth labels for the voxel block $\mathcal{B}^{\ell+1,r}$, there is supervision for each level of the tree. Once the output is generated the child nodes for the next level get generated. The decision on whether to add a child node and hence a higher resolution prediction is based on the boundary prediction in the corresponding octant of the voxel block $\mathcal{B}^{\ell+1,r}$. We compute the maximum boundary prediction response of the corresponding octant $\mathcal{O}'$ $$C_{\mathcal{O}'}^{\ell+1,r} =\max_{i,j,k \in \mathcal{O}'} \mathcal{B}^{\ell+1,r}_{i,j,k,2} \enspace,
\label{eq:childCriterion}$$ with label $2$ the *boundary* label. The child node is generated if $C_{\mathcal{O}'}^{\ell+1,r}$ is above a small threshold $\gamma$. The intuition behind this choice is that as soon as there is some evidence that the surface is going through a specific block we should predict a higher resolution output. On the other hand if the prediction on a specific level is very certain that there is no boundary within that subtree, a higher resolution prediction is not necessary. For the classes aeroplane, chair and car in practice only around 5 to 10% of the voxels are predicted at high resolution, in average (Fig. \[fig:slices\]), which shows that our approach effectively reduces the number of voxels that need to be predicted.
![Responses at the highest resolution, the checkered areas indicate not predicted at that resolution, (left) slice through airplane, (middle) slice through front legs of a chair, (right) slice through a car. []{data-label="fig:slices"}](figures/highResSlices/aero_checker.png "fig:"){width="0.31\linewidth"} ![Responses at the highest resolution, the checkered areas indicate not predicted at that resolution, (left) slice through airplane, (middle) slice through front legs of a chair, (right) slice through a car. []{data-label="fig:slices"}](figures/highResSlices/chair_checker.png "fig:"){width="0.31\linewidth"} ![Responses at the highest resolution, the checkered areas indicate not predicted at that resolution, (left) slice through airplane, (middle) slice through front legs of a chair, (right) slice through a car. []{data-label="fig:slices"}](figures/highResSlices/car_checker.png "fig:"){width="0.31\linewidth"}
In the first level of the tree, which predicts the root node from the shape code $\mathcal{C}$, a small decoding network directly predicts the first feature block without a cropping module in between. Likewise, in the deepest level of the tree, no explicit feature block is needed. Therefore the output is directly generated from the cropped features of the previous level. Also note that the output and upsampling modules have their individual filters at each level of the tree. Furthermore, thanks to this architecture all the convolutions and up-convolutions are standard layers and no special versions for the octree are required. Detailed layer configurations can be found in the appendix.
Efficient Training with Subsampling
-----------------------------------
As we have seen above, only predicting voxels on the boundary at each level largely reduces the complexity of the prediction task. In the beginning of the training, this is not the case because the boundaries are not yet predicted correctly. Moreover, even once the training has advanced enough and the boundaries are mostly placed correctly, an evaluation of the complete tree is still slow for training (approx. $0.7s$ for only a forward pass).
Therefore, we propose to utilize a subsampling of the child nodes during training. This is possible thanks to our hierarchical structure of the prediction. The tree gets traversed in a depth first manner. Each time the boundary label is present in an octant according to Eq. \[eq:childCriterion\] the child node is traversed with a certain probability $\rho$. Different schedules for $\rho$ can be used. We experimented with both, having a fixed probability $\rho$ or start with a very low $\rho$ and gradually increase it during training. The first version leads to a simpler training procedure and was hence used in our experiments. Thanks to the depth first traversal the memory consumption (weights and filter responses) grows linearly with the number of levels if the same upsampling and output architecture is used for each level.
In order to be able to train with mini-batches we sum up all the gradients during traversal of the tree and only do a gradient step as soon as we have done a forward and backward traversal of the subsampled tree for the training examples of the whole mini-batch. As loss function we use Cross-Entropy.
Experiments {#sec:exp}
===========
For our evaluation we use the synthetic dataset ShapeNetCore [@shapenet2015], which is commonly used to evaluate gemetry prediction networks. It is composed of Computer Aided Design (CAD) models of objects which are organized in categories. We use three categories for our evaluation, aeroplanes, chairs and cars. Our system is implemented in Torch[^1] and we use Adam [@kingma2015adam] for stochastic optimization, with a mini-batch size of 4.
Voxelization of the Ground Truth {#sec:voxelization}
--------------------------------
In a prepossessing step we voxelize the ground truth data. A common approach is to consider all the voxels which intersect the ground truth mesh as occupied space and then fill in the interior by labeling all voxels which are not reachable through free space from the boundary as occupied space. While this works well for the commonly used $32^3$ resolution it does not directly generalize to high resolutions. The reason is that the CAD models are generally not watertight meshes and with increasing resolution the risk that a hole in the mesh prevents filling the interior increases. In our case we are voxelizing the models at a resolution of $256^3$. In order to have filled interiors of the objects at high resolution we utilize a multi-scale approach. We first fill the interior at $32^3$ resolution. Then erode the occupied space by one voxel. We then fix every high resolution voxel which falls within this eroded low resolution occupied space as occupied space and also fix every high resolution voxel which falls within the original low resolution free space as free space. Furthermore, we fix all the high resolution voxels which intersect the original mesh surface as occupied space. To determine the ground truth label for the remaining high resolution voxels we run a graph-cut based regularization with a small smoothness term and a preference to keep the unlabeled voxels as free space.
Given the high resolution voxels at $256^3$ resolution we can now build the ground truth for all the remaining levels of the pyramid. At each level of the pyramid we label the voxels which contain the boundary at highest resolution as boundary and the other labels either free or occupied space, our coarsest resolution is $16^3$.
Baselines
---------
We consider two baselines. Both of the baselines use the state-of-the-art coarse resolution voxel prediction framework from Sec. \[sec:voxPred\]. For both baselines we use a uniform prediction resolution of $32^3$ (except for one evaluation where we use $64^3$ resolution), they differ in the way the ground truth is computed. The first baseline follows the standard approach of labeling all voxels which intersect the ground truth mesh surface as occupied space and then fill in the interior. This can be achieved by downsampling our high resolution ground truth and label all low resolution voxels which contain at least one high resolution occupied space voxel as occupied space and all the other ones as free space. We call this baseline “Low Resolution Hard" (LR Hard). The other baseline uses a soft assignment for the low resolution, the label is given by the ratio of high resolution free to occupied space voxels that the low resolution voxel contains. Therefore these labels can have fractional assignments. We call this baseline “Low Resolution Soft" (LR Soft). Note that the baseline LR Soft makes use of the high resolution voxelization but the baseline LR Hard is equivalent to voxelizing at low resolution. As our goal is prediction of high resolution geometry we trininearly upsample the raw classification output of the baselines from $32^3$ to $256^3$ and conduct the evaluation at high resolution.
![Evaluation on the validation set at different extraction thresholds for the class aeroplane at a resolution of $256^3$. For IoU higher is better and for CD lower is better.[]{data-label="fig:theshEvalAero"}](figures/plots/rgbAeroIoU.pdf "fig:"){width="0.9\linewidth"}\
![Evaluation on the validation set at different extraction thresholds for the class aeroplane at a resolution of $256^3$. For IoU higher is better and for CD lower is better.[]{data-label="fig:theshEvalAero"}](figures/plots/rgbAeroCD.pdf "fig:"){width="0.9\linewidth"}
------------------------------------- --------- ------------ ------------ ------------
Metric Method Car Chair Aero
(lr)[1-1]{} (lr)[2-2]{} (lr)[3-5]{} LR Hard 0.649 0.368 0.458
LR Soft 0.679 0.360 0.518
HSP **0.701** **0.378** **0.561**
LR Hard 0.0142 0.0275 0.0190
LR Soft 0.0122 0.0321 0.0190
HSP **0.0108** **0.0266** **0.0121**
------------------------------------- --------- ------------ ------------ ------------
: Quantitative analysis for the task of predicting geometry from RGB input.[]{data-label="tab:rgbPredictionQuant"}
\
\
\
\
\
\
{width="0.115\linewidth"} {width="0.115\linewidth"} {width="0.115\linewidth"} {width="0.115\linewidth"}
Input Data
----------
To obtain the RGB/Depth images which are used as input for training and testing our CNNs, we render the CAD models in the ShapeNet dataset using Blender[^2]. For each CAD model, we render 10 images from random viewpoints obtained via uniformly sampling azimuth from $[0,360)$ degrees and elevation from $[-20,30]$ degrees. We also use random lighting variations to render the RGB images.
We also use partial voxel grids as input. In practice such an input can arise when multiple depth maps of an object are available but they are all from the same side of the object. To imitate this task we use the ground truth of the LR Soft baseline and randomly zero the data in half of the voxel grid. The network then learns to predict the complete high resolution geometry from the partial low resolution input.
Quantitative Evaluation {#sec:quantEval}
-----------------------
We conducted a quantitative evaluation on the task of predicting high resolution geometry from a single RGB image. We assign each of the 3D models from the dataset randomly to the train, validation or test split with probabilities 70%, 10% and 20%, respectively. We trained category specific networks for our method Hierarchical Surface Prediction (HSP) and both baselines for the three categories, aeroplanes, chairs and cars. To quantify the accuracy we use use the Intersection over Union (IoU) score and the symmetric Chamfer Distance (CD). We compute the measures for each model and average over the dataset. To compute the measures we binarize the non-binary predictions with a suitable threshold. In order to determine the best threshold for each measure and category we test an exhaustive set of thresholds using the validation set. The plots given in Fig. \[fig:theshEvalAero\] show that the choice of a good threshold has a very drastic influence to the error measure. As a first experiment, we study predictions at a coarse resolution of $64^3$. For this experiment both the baseline and HSP are trained to predict $64^3$ resolution and the evaluation is conducted at the same resolution. We demonstrate that our hierarchical prediction performs similar to the LR hard baseline the performance of HSP does not degrade compared to uniform prediction at the same resolution. For the baseline LR Hard we get an IoU of 43.47% and 43.12% for HSP, respectively on the validation set for the class chair. However, unlike HSP, the uniform prediction baselines cannot scale to finer resolutions and we now empirically show the benefits of being able to predict at finer resolutions. The evaluation is conducted at $256^3$ resolution. We compute the IoU and Chamfer distance for each of the models in the test set using the per category and per error measure best threshold determined using the validation set. The results are given in Table \[tab:rgbPredictionQuant\]. We outperform the baselines for all categories and evaluation metrics.
Qualitative Evaluation
----------------------
In Figs. \[fig:chairQualitative\],\[fig:aeroQualitative\] and \[fig:carQualitative\] we show example results on the task of geometry prediction from a single color image for the three categories. The results are selected to demonstrate the variablility of shapes within a category. Random results are shown in the appendix. The results are predicted with the same networks which we used in the quantitative analysis above. The meshes are extracted with marching cubes [@lorensen1987marching] using the same thresholds as for the quantitative analysis using the Chamfer Distance. For chair the LR Hard baseline produces visually pleasant results. However, this same baseline does not achieve detailed results on the categories aeroplane and car. Due to the fractional labels the LR Soft baseline often misses thin structures. The results generated with our proposed HSP have high quality surfaces for all the three categories.
We also downloaded two images of cars with a white background from the Internet using Google image search and processed them with our car prediction network, which was trained only on synthetic data. As can be seen in Fig. \[fig:addRes\], the network is able to generalize to this type of input. Additionally, we trained a network on the category car which takes a partial low resolution voxel grid as input and predicts a high resolution model of the whole object (c.f. Fig. \[fig:addRes\]).
Conclusion and Future Work {#sec:conclusion}
==========================
In this work, we presented a hierarcical surface prediction framework which facilitates high resolution 3D reconstruction of objects from very little input data. We presented results on the task of predicting geometry from a single color image, a single depth image and from a partial low resolution voxel grid. Our quantitative evaluation shows that the high resolution prediction leads to more accurate results than low resolution baselines. Qualitatively, our predicted surfaces are superior to the low resolution baselines.
In this work we have demonstrated that high resolution geometry prediction is feasible therefore future work will investigate the potential to use prediction based reconstruction methods in multi-view stereo. Furthermore, for almost symmetric objects methods which explicitly take the symmetry into account are an interesting further direction which might improve the quality of prediction based approaches.
Acknowledgments {#acknowledgments .unnumbered}
===============
C. H[ä]{}ne received funding from the “Early Postdoc.Mobility” fellowship No. 165245 from the Swiss National Science Foundation. The project received funding from the Intel/NSF VEC award IIS-153909 and NSF award IIS-1212798.
[^1]: http://torch.ch/
[^2]: http://www.blender.org
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study the Ginzburg-Landau model of type-I superconductors in the regime of small external magnetic fields. We show that, in an appropriate asymptotic regime, flux patterns are described by a simplified branched transportation functional. We derive the simplified functional from the full Ginzburg-Landau model rigorously via $\Gamma$-convergence. The detailed analysis of the limiting procedure and the study of the limiting functional lead to a precise understanding of the multiple scales contained in the model.'
author:
- 'Sergio Conti[^1]'
- 'Michael Goldman[^2]'
- 'Felix Otto[^3]'
- 'Sylvia Serfaty[^4]'
bibliography:
- 'CGOS.bib'
title: '[A branched transport limit of the Ginzburg-Landau functional]{}'
---
Introduction
============
In 1911, K. Onnes discovered the phenomenon of superconductivity, manifested in the complete loss of resistivity of certain metals and alloys at very low temperature. W. Meissner discovered in 1933 that this was coupled with the expulsion of the magnetic field from the superconductor at the critical temperature. This is now called the Meissner effect. After some preliminary works of the brothers F. and H. London, V. Ginzburg and L. Landau proposed in 1950 a phenomenological model describing the state of a superconductor. In their model (see below), which belongs to Landau’s general theory of second-order phase transitions, the state of the material is represented by the order parameter $u: \Omega\to {\mathbb{C}}$, where $\Omega$ is the material sample. The density of superconducting electrons is then given by $\rho:= |u|^2$. A microscopic theory of superconductivity was first proposed by Bardeen-Cooper-Schrieffer (BCS) in 1957, and the Ginzburg-Landau model was derived from BCS by Gorkov in 1959 (see also [@frankseiringer] for a rigorous derivation).
One of the main achievements of the Ginzburg-Landau theory is the prediction and the understanding of the mixed (or intermediate) state below the critical temperature. This is a state in which, for moderate external magnetic fields, normal and superconducting regions coexist. The behavior of the material in the Ginzburg-Landau theory is characterized by two physical parameters. The first is the coherence length $\xi$ which measures the typical length on which $u$ varies, the second is the penetration length $\lambda$ which gives the typical length on which the magnetic field penetrates the superconducting regions. The Ginzburg-Landau parameter is then defined as $\kappa:= \frac{\lambda}{\xi}$. The Ginzburg-Landau functional is given by $$\label{GLorigin}
\int_{\Omega} |\nabla_A u|^2 +\frac{\kappa^2}{2} (1-|u|^2)^2 dx +\int_{\R^3} |\nabla\times A- {B_{\mathrm{ext}}}|^2 dx$$ where $A: \R^3\to \R^3$ is the magnetic potential (so that $B:=\nabla\times A$ is the magnetic field), $ \nabla_A u:= \nabla u -i Au$ is the covariant derivative of $u$ and ${B_{\mathrm{ext}}}$ is the external magnetic field. In these units, the penetration length $\lambda$ is normalized to $1$. As first observed by A. Abrikosov this theory predicts two types of superconductors. On the one hand, when $\kappa < 1/\sqrt{2}$, there is a positive surface tension which leads to the formation of normal and superconducting regions corresponding to $\rho\simeq 0$ and $\rho\simeq 1$ respectively, separated by interfaces. These are the so-called type-I superconductors. On the other hand, when $\kappa>1/\sqrt{2}$, this surface tension is negative and one expects to see the magnetic field penetrating the domain through lines of vortices. These are the so-called type-II superconductors. In this paper we are interested in better understanding the former type but we refer the interested reader to [@tinkham; @SanSerf; @Serfrev] for more information about the latter type. In particular, in that regime, there has been an intensive work on understanding the formation of regular patterns of vortices known as Abrikosov lattices.
In type-I superconductors, it is observed experimentally [@ProzHo; @Proz; @Proal] that complex patterns appear at the surface of the sample. It is believed that these patterns are a manifestation of branching patterns inside the sample. Although the observed states are highly history-dependent, it is argued in [@ChokKoOt; @Proal] that the hysteresis is governed by low-energy configurations at vanishing external magnetic field. The scaling law of the ground-state energy was determined in [@ChokConKoOt; @ChokKoOt] for a simplified sharp interface version of the Ginzburg-Landau functional and in [@CoOtSer] for the full energy, these results indicate the presence of a regime with branched patterns at low fields.
This paper aims at a better understanding of these branched patterns by going beyond the scaling law. Starting from the full Ginzburg-Landau functional, we prove that in the regime of vanishing external magnetic field, low energy configurations are made of nearly one-dimensional superconducting threads branching towards the boundary of the sample. In a more mathematical language, we prove $\Gamma-$convergence [@braides; @dalmaso] of the Ginzburg-Landau functional to a kind of branched transportation functional in this regime. We focus on the simplest geometric setting by considering the sample $\Om$ to be a box $Q_{\Lz,T}:=(-\frac{\Lz}{2}, \frac{\Lz}{2})^2\times(-T,T)$ for some $T,\Lz>0$ and consider periodic lateral boundary conditions. The external magnetic field is taken to be perpendicular to the sample, that is ${B_{\mathrm{ext}}}:= {b_{\mathrm{ext}}}e_3$ for some ${b_{\mathrm{ext}}}>0$ and where $e_3$ is the third vector of the canonical basis of $\R^3$. After making an isotropic rescaling, subtracting the bulk part of the energy and dropping lower order terms (see and the discussion after it), minimizing can be seen as equivalent to minimizing
$$\begin{gathered}
\label{Etilde}
E_T(u,A):=\frac{1}{L^2}
\int_{Q_{L,1}} |\nabla_{TA} u |^2 + \left(B_3 - \a (1-\rho)\right)^2 + |B'|^2 dx
+ \| B_3-\a \b\|_{H^{-1/2}(\{x_3=\pm 1\})}^2,\,\end{gathered}$$
where we have let $B:=(B',B_3):=\nabla \times A$, $$\kappa T:=\sqrt{2}\alpha, \qquad {b_{\mathrm{ext}}}:=\frac{\beta \kappa}{\sqrt{2}} \qquad \textrm{ and } \qquad L:=\Lz/T.$$ If $u=\rho^{1/2} \exp(i\theta)$, since $|\nabla_{TA} u |^2 =|\nabla \rho^{1/2}|^2 +\rho|\nabla \theta-T A|^2$, in the limit $T\to +\infty$ we obtain, at least formally, that $A$ is a gradient field in the region where $\rho>0$ and therefore the Meissner condition $\rho B=0$ holds. Moreover, in the regime $\alpha\gg 1$, from we see that $B_3 \simeq \a (1-\rho)$ and $\rho$ takes almost only values in $\{0,1\}$. Hence $\Div B=0$ can be rewritten as $$ \partial_3 \chi + \frac1\alpha\Divp \chi B'=0,$$ where $\chi:= (1-\rho)$ and $\Divp$ denotes the divergence with respect to the first two variables. Therefore, from the Benamou-Brenier formulation of optimal transportation [@AGS; @villani] and since from the Meissner condition, $B'\simeq \frac{1}{\chi} B'$, the term $$\int_{Q_{L,1}} |B'|^2 dx\simeq \int_{Q_{L,1}} \frac{1}{\chi}|B'|^2 dx$$ in the energy can be seen as a transportation cost. We thus expect that inside the sample (this is, in $Q_{L,1}$), superconducting domains where $\rho\simeq1$ and $B\simeq 0$ alternate with normal ones where $\rho\simeq 0$ and $B_3\simeq \alpha$. Because of the last term $\| B_3-\a \b\|_{H^{-1/2}(\{x_3=\pm 1\})}^2$ in the energy , one expects $B\simeq \a \b e_3$ outside the sample. This implies that close to the boundary the normal domains have to refine. The interaction between the surface energy, the transportation cost and the penalization of an $H^{-1/2}$ norm leads to the formation of complex patterns (see Figure \[figscales\]).
It has been proven in [@CoOtSer] that in the regime $T\gg 1$, $\a \gg 1$ and $\b \ll 1$, $$\label{scallawintro}
\min E_T(u,A)\sim \min\{ \a^{4/3} \b^{2/3}, \a^{10/7} \b\}\,.$$ The scaling $\min E_T(u,A)\sim \a^{4/3} \b^{2/3}$ (relevant for $\a^{-2/7}\ll \b$) corresponds to uniform branching patterns whereas the scaling $\min E_T(u,A)\sim \a^{10/7} \b$ corresponds to non-uniform branching ones. We focus here for definiteness on the regime $\min E_T(u,A)\sim \a^{4/3} \b^{2/3}$, although we believe that our proof can be extended to the other one. Based on the construction giving the upper bounds in , we expect that in the first regime there are multiple scales appearing (see Figure \[figscales\]): $$\label{scalesepar}\stackrel{\displaystyle\textrm{penetration}}{\textrm{ length}}\,\stackrel{\displaystyle\ll}{\, } \, \stackrel{\displaystyle\textrm{ coherence}}{\textrm{ length}}\, \stackrel{\displaystyle\ll}{\,}
\, \stackrel{\displaystyle\textrm{ diameter of the}}{\textrm{ threads in the bulk}} \, \, \stackrel{\displaystyle\ll}{\,} \, \, \stackrel{\displaystyle\textrm{distance between the}}{\textrm{ threads in the bulk}},$$ which amounts in our parameters to $$T^{-1} \ll \alpha^{-1}\ll \alpha^{-1/3}\beta^{1/3}\ll \alpha^{-1/3}\beta^{-1/6}.$$
\[figscales\]
In order to better describe the minimizers we focus on the extreme region of the phase diagram $T,T \a^{-1}, \beta^{-1}, \a \b^{7/2}\to +\infty$, with $L={\widetilde}L \alpha^{-1/3}\beta^{-1/6}$ for some fixed ${\widetilde}L>0$. In this regime, we have in particular $\a^{-1}\ll \a^{-1/3}\b^{1/3}$ so that the separation of scales holds. We introduce an anisotropic rescaling (see Section \[Sec:GL\]) which leads to the functional
$$\begin{aligned}
\label{eqdefwe}
\wE(u,A)&:=\frac{1}{\widetilde{L}^2}\Big[\int_{Q_{{\widetilde}L,1}} \a^{-2/3}\b^{-1/3}\lt|\nabla_{\a^{1/3}\b^{-1/3}TA}'u\rt|^2+
\a^{-4/3}\b^{-2/3}\lt|(\nabla_{\a^{1/3}\b^{-1/3}TA} u)_3\rt|^2\\
\nonumber
+ & \lt.\a^{2/3}\b^{-2/3}\left({B_3} -(1-|{u}|^2)\right)^2
+ \b^{-1}|{B'}|^2 dx +\a^{1/3}\b^{7/6}\|\b^{-1}{B_3}-1\|_{H^{-1/2}(x_3=\pm1)}^2\rt].\end{aligned}$$
Our main result is a $\Gamma-$convergence result of the functional $\wE$ towards a functional defined on measures $\mu$ living on one-dimensional trees. These trees correspond to the normal regions in which $\rho\simeq 0$ and where the magnetic field $B$ penetrates the sample. Roughly speaking, if for a.e. $x_3\in (-1,1)$ the slice of $\mu=\mu_{x_3}\otimes dx_3$ has the form $\mu_{x_3}=\sum_i {\varphi}_i \delta_{X_i(x_3)}$ where the sum is at most countable, then we let (see Section \[Sec:limiting\] for a precise definition) $$\label{eqdefIintro}
I(\mu):=\frac1{\widetilde{L}^2}\int_{-1}^1 K_* \sum_i \sqrt{{\varphi}_i}+ {\varphi}_i |{\dot{X}}_i|^2 dx_3,$$ where $K_*=8\sqrt\pi/3$ and ${\dot{X}}_i$ denotes the derivative (with respect to $x_3$) of $X_i(x_3)$. The $X_i$’s represent the graphs of each branch of the tree (parametrized by height) and the ${\varphi}_i$’s represent the flux carried by the branch. We can now state our main result
\[maintheo\] Let $T_n, \a_n, \b_n^{-1}\to+\infty$ with $T_n \a_n^{-1}, \a_n \b_n^{7/2}\to +\infty$, ${\widetilde}L>0$, then:
1. \[maintheolb\] For every sequence $(u_n, A_n)$ with $\sup_n \wE(u_n,A_n)<+\infty$, up to subsequence , $\beta_n^{-1}(1-|u_n|^2)$ weakly converges to a measure $\mu$ of the form $\mu=\mu_{x_3}\otimes dx_3$ with $\mu_{x_3}=\sum_i {\varphi}_i \delta_{X_i}$ for a.e. $x_3\in (-1,1)$, $\mu_{x_3} \weaklim dx'$ (where $dx'$ denotes the two dimensional Lebesgue measure on $Q_{{\widetilde}L}$) when $x_3\to \pm 1$ and such that $$\liminf_{n\to +\infty} \wE(u_n,A_n)\ge I(\mu).$$
2. \[maintheoub\] If in addition $L_n^{2} \a_n \b_n T_n\in 2\pi \Z$, where $L_n:= {\widetilde}L \a_n^{-1/3}\b_n^{-1/6}$, then for every measure $\mu$ such that $I(\mu)<+\infty$ and $\mu_{x_3} \weaklim dx'$ as $x_3\to \pm1$ , there exists $(u_n,A_n)$ such that $\b_n^{-1}(1-|u_n|^2)\weaklim \mu$ and $$\limsup_{n\to +\infty} \, \wE(u_n,A_n)\le I(\mu).$$
By scaling, it suffices to consider the case ${\widetilde}L=1$. The first assertion follows from Proposition \[gammaliminf\], the second one from Proposition \[gammalimsup-v2\].
Let us stress once again that our result could have been equivalently stated for the full Ginzburg-Landau energy instead of $\wE$ (see Section \[Sec:GL\]).\
Within our periodic setting, the quantization condition $L_n^2 \a_n \b_n T_n\in 2\pi \Z$ for the flux is a consequence of the fact that the phase circulation of the complex-valued function in the original problem is naturally quantized. It is necessary in order to make our construction but we believe that it is also a necessary condition for having sequences of bounded energy (see the discussion in Section \[Sec:GL\] and the construction in Section \[recoveryGL\]). We remark that scaling back to the original variables this condition is the physically natural one $L_0^2{b_{\mathrm{ext}}}\in 2\pi\Z$.\
Before going into the proof of Theorem \[maintheo\] we address the limiting functional $I(\mu)$, which has many similarities with irrigation (or branched transportation) models that have recently attracted a lot of attention (see [@BeCaMo] and more detailed comments in Section \[Sec:reliri\] or the recent papers [@BraWirth; @BraWir16] where the connection is also made to some urban planning models). In Section \[Sec:limiting\], we first prove that the variational problem for this limiting functional is well-posed (Proposition \[existmu\]) and show a scaling law for it (Proposition \[branchingmu\] and Proposition \[branchinglowerbound\]). In Proposition \[def:subsystem\], we define the notion of subsystems which allows us to remove part of the mass carried by the branching measure. This notion is at the basis of Lemma \[noloop\] and Proposition \[finitebranch\] which show that minimizers contain no loops and that far from the boundary, they are made of a finite number of branches. From the no-loop property, we easily deduce Proposition \[reg\] which is a regularity result for minimizers of $I$. The main result of Section \[Sec:limiting\] is Theorem \[theodens\] which proves the density of “regular” measures in the topology given by the energy $I(\mu)$. As in nearly every $\Gamma-$convergence result, such a property is crucial in order to implement the construction for the upper bound \[maintheoub\]. We now comment on the proof of Theorem \[maintheo\]. Let us first point out that if the Meissner condition $\rho B =0$ were to hold, and $A$ could be written as a gradient field in the set $\{\rho>0\}$, then $|\nabla_{TA} u|^2= |\nabla \rho^{1/2}|^2$ and we would have $$\label{introlowerbound}\int_{Q_{L,1}} |\nabla_{TA} u |^2 + \left(B_3 - \a (1-\rho)\right)^2dx=\int_{Q_{L,1}} |\nabla \rho^{1/2} |^2 + \a^2 \chi_{\rho>0} (1-\rho)^2 +\chi_{\rho=0} (B_3-\a)^2 dx.$$ This is a Modica-Mortola [@ModMort] type of functional with a degenerate double-well potential given by $W(\rho):= \chi_{\rho>0} (1-\rho)^2$. Thanks to Lemma \[lemmameissner\], one can control how far we are from satisfying the Meissner condition. From this, we deduce that almost holds (see Lemma \[lemmafirstlowerbound\]). This implies that the Ginzburg-Landau energy gives a control over the perimeter of the superconducting region $\{\rho>0\}$. In addition, $\b\ll 1$ imposes a small cross-area fraction for $\{\rho>0\}$. Using then isoperimetric effects to get convergence to one-dimensional objects (see Lemma \[lowerbound2d\]), we may use Proposition \[gammaliminf\] to conclude the proof of \[maintheolb\].
In order to prove \[maintheoub\], thanks to the density result in Theorem \[theodens\], it is enough to consider regular measures. Given such a measure $\mu$, we first approximate it with quantized measures (Lemma \[lemmaquantize\]). Far from branching points the construction is easy (see Lemma \[lemmacurve3\]). At a branching point, we need to pass from one disk to two (or vice-versa); this is done passing through rectangles (see Lemma \[Lembranch3\] and Figure \[fig1\]). Close to the boundary we use instead the construction from [@CoOtSer], which explicitly generates a specific branching pattern with the optimal energy scaling; since the height over which this is done is small the prefactor is not relevant here (Proposition \[propboundraylayer\]). The last step is to define a phase and a magnetic potential to get back to the full Ginzburg-Landau functional. This is possible since we made the construction with the Meissner condition and quantized fluxes enforced, see Proposition \[thirdupperbound\].
From and the discussion around , for type-I superconductors, the Ginzburg-Landau functional can be seen as a non-convex, non-local (in $u$) functional favoring oscillations, regularized by a surface term which selects the lengthscales of the microstructures. The appearance of branched structures for this type of problem is shared by many other functionals appearing in material sciences such as shape memory alloys [@KohnMuller92; @KohnMuller94; @Conti00; @KnuepferKohnOtto2013; @BelGol; @Zwicknagl2014; @ChanConti2015; @ContiZwicknagl], uniaxial ferromagnets [@ChoKoOtmicro; @ViehOtt; @KnMu] and blistered thin films [@BCDM00; @JinSternberg; @BCDM02]. Most of the previously cited results on branching patterns (including [@ChokConKoOt; @ChokKoOt; @CoOtSer] for type-I superconductors) focus on scaling laws. Here, as in [@ViehOtt; @CDZ], we go one step further and prove that, after a suitable anisotropic rescaling, configurations of low energy converge to branched patterns. The two main difficulties in our model with respect to the one studied in [@ViehOtt] are the presence of an additional lengthscale (the penetration length) and its sharp limit counterpart, the Meissner condition $\rho B=0$ which gives a nonlinear coupling between $u$ and $B$. Let us point out that for the Kohn-Müller model [@KohnMuller92; @KohnMuller94], a much stronger result is known, namely that minimizers are asymptotically self-similar [@Conti00] (see also [@Viehmanndiss; @AlChokOt] for related results). In [@Gol], the optimal microstructures for a two-dimensional analogue of $I(\mu)$ are exactly computed.
The paper is organized as follows. In section \[sec:not\], we set some notation and recall some notions from optimal transport theory. In Section \[Sec:GL\], we recall the definition of the Ginzburg-Landau functional together with various important quantities such as the superconducting current. We also introduce there the anisotropic rescaling leading to the functional $\widetilde{E}_T$. In Section \[Sec:intermediate\], we introduce for the sake of clarity intermediate functionals corresponding to the different scales of the problem. Let us stress that we will not use them in the rest of the paper but strongly believe that they help understanding the structure of the problem. In Section \[Sec:limiting\], we carefuly define the limiting functional $I(\mu)$ and study its properties. In particular we recover a scaling law for the minimization problem and prove regularity of the minimizers. We then prove the density in energy of ’regular’ measures. This is a crucial result for the main $\Gamma-$ convergence result which is proven in the last two sections. As customary, we first prove the lower bound in Section \[Sec:gammaliminf\] and then make the upper bound construction in Section \[Sec:upperbound\].
Notation and preliminary results {#sec:not}
=================================
In the paper we will use the following notation. The symbols $\sim$, $\ges$, $\les$ indicate estimates that hold up to a global constant. For instance, $f\les g$ denotes the existence of a constant $C>0$ such that $f\le Cg$, $f\sim g$ means $f\les g$ and $g\les f$. In heuristic arguments we use $a\simeq b$ to indicate that $a$ is close (in a not precisely specified sense) to $b$. We use a prime to indicate the first two components of a vector in $\R^3$, and identify $\R^2$ with $\R^2\times\{0\}{\subseteq}\R^3$. Precisely, for $a\in \R^3$ we write $a'=(a_1,a_2,0)\in\R^2{\subseteq}\R^3$; given two vectors $a,b\in\R^3$ we write $a'\times b'=(a\times b)_3 =
(a'\times b')_3$. We denote by $(e_1,e_2,e_3)$ the canonical basis of $\R^3$. For $L>0$ and $T>0$, $Q_L:=(-\frac{L}{2},\frac{L}{2})^2$ and $Q_{L,T}:=Q_L\times(-T,T)$. For a function $f$ defined on $Q_{L,T}$, we denote $f_{x_3}$ the function $f_{x_3}(x'):=f(x',x_3)$ and we analogously define for $\Omega {\subseteq}Q_{L,T}$, the set $\Omega_{x_3}$. For $x=(x',x_3)$ and $r>0$ we let $\B_r(x)=\B(x,r)$ be the ball of radius $r$ centered at $x$ (in $\R^3$) and $\B'_r(x')=\B'(x',r)$ be the analogue two-dimensional ball centered at $x'$. Unless specified otherwise, all the functions and measures we will consider are periodic in the $x'$ variable, i.e., we identify $Q_L$ with the torus $\R^2/L\Z^2$. In particular, for $x',y'\in Q_L$, $|x'-y'|$ denotes the distance for the metric of the torus, i.e., $|x'-y'|:=\min_{k\in \Z^2} |x'-y'+Lk|$. We denote by $\H^k$ the $k-$dimensional Hausdorff measure. We let $\cP(Q_L)$ be the space of probability measures on $Q_L$ and $\M(Q_{L,T})$ be the space of finite Radon measures on $Q_{L,T}$, and similarly $\M(Q_L)$. Analogously, we define $\M^+(Q_{L,T})$ and $\M^+(Q_L)$ as the spaces of finite Radon measures which are also positive. For a measure $\mu$ and a function $f$, we denote by $f\sharp \mu$ the push-forward of $\mu$ by $f$.
We recall the definition of the (homogeneous) $H^{-1/2}$ norm of a function $f\in L^2(Q_{L})$ with $\int_{Q_{L}} f dx'=0$, $$\label{defHdem}
\|f\|_{H^{-1/2}}^2:=\inf\lt\{ \int_{Q_{L}\times [0,+\infty)} |B|^2 dx \ : \ \Div B=0, \ B_3(\cdot,0)=f \rt\},$$ which can be alternatively given in term of the two-dimensional Fourier series as (see, e.g., [@ChokKoOt]) $$ \|f\|_{H^{-1/2}}^2=\frac{1}{2\pi} \sum_{ k'\in \lt(\frac{2\pi}{L}\Z\rt)^2\backslash\{0\}} \frac{ |\hat f (k')|^2}{|k'|}.$$ We shall write $\|f\|^2_{H^{-1/2}(Q_L\times\{\pm T\})}$ for $\|f\|^2_{H^{-1/2}(Q_L\times\{T\})}+
\|f\|^2_{H^{-1/2}(Q_L\times\{-T\})}$.
The $2$-Wasserstein distance between two measures $\mu$ and $\nu\in {\mathcal{M}}^+(Q_L)$ with $\mu(Q_L)=\nu(Q_L)$ is given by $$W_2^2(\mu,\nu):=\min\lt\{ \mu(Q_L)\int_{Q_L\times Q_L} |x-y|^2 \, d\Pi(x,y) \, : \, \Pi_1=\mu, \ \Pi_2=\nu\rt\},$$ where the minimum is taken over measures on $Q_L\times Q_L$ and $\Pi_1$ and $\Pi_2$ are the first and second marginal of $\Pi$[^5], respectively. For measures $\mu,\nu\in{\mathcal{M}}^+(Q_{L,T})$, the $2$-Wasserstein distance is correspondingly defined. We now introduce some notions from metric analysis, see [@AGS; @villani] for more detail. A curve $\mu:(a,b)\to \cP(Q_L)$, $z\mapsto \mu_z$ belongs to $AC^2(a,b)$ (where $AC$ stands for absolutely continuous) if there exists $m\in L^2(a,b)$ such that $$\label{metricderiv}
W_2(\mu_{z},\mu_{{\widetilde}{z}})\le \int_{z}^{{\widetilde}{z}} m(t) dt \ \qquad \forall a<z\le {\widetilde}{z} <b.$$ For any such curve, [the [*speed*]{}]{} $$|\mu'|(z):=\lim_{{\widetilde}{z}\to z}\frac{W_2(\mu_{{\widetilde}{z}},{\mu}_{z})}{|z-{\widetilde}{z}|}$$ exists for $\H^1-$a.e. $z\in (a,b)$ and $|\mu'|(z)\le m(z)$ for $\H^1$-a.e. $z\in (a,b)$ for every admissible $m$ in . Further, there exists a Borel vector field $B$ such that $$\label{eqexistsB}
B(\cdot,z)\in L^2(Q_L,\mu_{z}), \qquad \| B(\cdot,z)\|_{L^2(Q_L,\mu_{z})}\le |\mu'|(z) \quad \textrm{for } \H^1\text{-a.e. } \ z\in (a,b)$$ and the continuity equation $$\label{conteqini}
\partial_3 \mu_{z}+\Divp (B \mu_{z})=0$$ holds in the sense of distributions [@AGS Th. 8.3.1]. Conversely, if a weakly continuous curve $\mu_{z} : (a,b)\to \cP(Q_L)$ satisfies the continuity equation for some Borel vector field $B$ with $\| B(\cdot,z)\|_{L^2( Q_L,\mu_{z})}\in L^2(a,b)$ then $\mu\in AC^2(a,b)$ and $|\mu'|(z)\le \| B(\cdot,z)\|_{L^2(Q_L,\mu_{z})}$ for $\H^1$-a.e. $z\in (a,b)$. In particular, we have $$\label{eqw22b}
W_2^2(\nu,\hat\nu)= \min_{\mu,B} \lt\{ 2T\int_{Q_{L,T}} |B|^2 d\mu_{z} dz \ :\ \mu_{-T}=\nu, \ \mu_T= \hat\nu \textrm{ and \eqref{conteqini} holds}\rt\}\,,$$ where by scaling the right-hand side does not depend on $T$.\
For a (signed) measure $\mu\in \M(Q_L)$, we define the Bounded-Lipschitz norm of $\mu$ as $$\label{BLdefin}
\|\mu\|_{BL}:=\sup_{\|\psi\|_{Lip}\le 1} \int_{Q_L} \psi d\mu,$$ where for a $Q_L-$periodic and Lipschitz continuous function $\psi$, $\|\psi\|_{Lip}:=\|\psi\|_{\infty}+\|\nabla \psi\|_{\infty}$. By the Kantorovich-Rubinstein Theorem [@villani Th. 1.14], the $1-$Wasserstein and the Bounded-Lipschitz norm are equivalent.
The Ginzburg-Landau functional {#Sec:GL}
==============================
In this section we recall some background material concerning the Ginzburg-Landau functional and introduce the anisotropic rescaling leading to $\wE$.\
For a (non necessarily periodic) function $u:Q_{\Lz,T}\to{\mathbb{C}}$, called the order parameter, and a vector potential $A: Q_{\Lz}\times\R\to \R^3$ (also not necessarily periodic), we define the covariant derivative $$\nabla_A u:= \nabla u-iAu,$$ the magnetic field $$B:=\nabla\times A,$$ and the [superconducting]{} current $$\label{defj}
j_A:= \frac12\left( -i\bar u (\nabla_A u)+ i u (\overline{\nabla_A u})\right)=\textrm{Im}(iu \overline{\nabla_A u})\,.$$ Let us first notice that $|\nabla_A u|^2$ and the observable quantities $\rho$, $B$ and $j_A$ are invariant under change of gauge. That is, if we replace $u$ by $u e^{i\varphi}$ and $A$ by $A+\nabla \varphi$ for any function $\varphi$, they remain unchanged. We also point out that if $u$ is written in polar coordinates as $u=\rho^{1/2} e^{i\theta}$, then $$ |\nabla_A u|^2=|\nabla \rho^{1/2}|^2+\rho|\nabla \theta-A|^2.$$ For any admissible pair $(u,A)$, that is such that $\rho$, $B$ and $j_A$ are $Q_{\Lz}$-periodic, we define the Ginzburg-Landau functional as $$ {E_\mathrm{GL}}(u,A):=\int_{Q_{\Lz,T}} |\nabla_A u|^2
+ \frac{\kappa^2}{2} (1-|u|^2)^2 dx
+ \int_{Q_{\Lz}\times \R} |\nabla\times A-{B_{\mathrm{ext}}}|^2 dx\,.$$ We remark that $u$ and $A$ need not be (and, if ${B_{\mathrm{ext}}}\ne 0$, cannot be) periodic. See [@CoOtSer] for more details on the functional spaces we are using. Here ${B_{\mathrm{ext}}}:={b_{\mathrm{ext}}}e_3$ is the external magnetic field and $\kappa\in(0,1/\sqrt2)$ is a [material constant, called the Ginzburg-Landau parameter]{}. From periodicity and $\Div B=0$ it follows that $ \int_{Q_{\Lz}\times\{x_3\}} B_3 \, dx'$ does not depend on $x_3$ and therefore, if the energy is finite, necessarily $$\label{B2bex}
\int_{Q_{\Lz}\times\{x_3\}} B_3 \, dx'=\Lz^2
{b_{\mathrm{ext}}}\hskip1cm\text{ for all }x_3\in\R\,.$$ We first remove the bulk part from the energy ${E_\mathrm{GL}}$. In order to do so, we introduce the quantity $${\mathcal{D}}_A^3u := (\partial_2 u - i A_2 u) - i (\partial_1 u
- i A_1 u)
= (\nabla_A u)_2 - i (\nabla_A u)_1$$ and, more generally, $${\mathcal{D}}_A^ku := (\partial_{k+2} u - i A_{k+2} u) - i (\partial_{k+1} u
- i A_{k+1} u) = (\nabla_A u)_{k+2} - i (\nabla_A u)_{k+1} \, ,$$ where components are understood cyclically (i.e., $a_k=a_{k+3}$). The operator ${\mathcal{D}}_A$ (which corresponds to a creation operator for a magnetic Laplacian) was used by Bogomol’nyi in the proof of the self-duality of the Ginzburg-Landau functional at ${\kappa}=\frac{1}{\sqrt{2}} $ (cf. e.g. [@jaffetaubes]). His proof relied on identities similar to the next ones, which will be crucial in enabling us to separate the leading order part of the energy.
Expanding the squares, one sees (for details see [@CoOtSer Lem. 2.1]) that (recall that $\rho:=|u|^2$) $$\label{magic}
|\nabla'_A u|^2 = |{\mathcal{D}}_A^3u|^2 + \rho B_3 + \nabla'\times
j'_A$$ and, for any $k=1,2,3$, $$\label{magic2}
|(\nabla_A u)^{k+1}|^2 + |(\nabla_A u)^{k+2}|^2= |{\mathcal{D}}_A^ku|^2
+ \rho B_k + (\nabla\times j_A)_k \,.$$ This implies $$|\nabla_A' u |^2 = (1-{\kappa}\sqrt2) |\nabla_A' u |^2 + {\kappa}\sqrt2
|{\mathcal{D}}_A^3u|^2
+{\kappa}\sqrt2\rho B_3 + {\kappa}\sqrt2\nabla'\times j_A'\,.$$ The last term integrates to zero by the periodicity of $j_A$. Therefore, for each fixed $x_3$, using , we have $$\int_{Q_{\Lz}} |\nabla_A' u |^2\, dx' = \int_{Q_{\Lz}} (1-{\kappa}\sqrt2) |\nabla_A' u |^2 + {\kappa}\sqrt2
|{\mathcal{D}}_A^3u|^2 +{\kappa}\sqrt2(\rho-1) B_3 \, dx'+ \Lz^2 {\kappa}\sqrt2 {b_{\mathrm{ext}}}\,.$$ We substitute and obtain, using $\int_{Q_{\Lz}} (B_3-{b_{\mathrm{ext}}})^2 dx'=\int_{Q_{\Lz}} B_3^2-{b_{\mathrm{ext}}}^2 dx'$ [and completing squares]{}, $$\label{eqEGLE}
{E_\mathrm{GL}}(u,A)=2T\Lz^2\left({\kappa}\sqrt2 {b_{\mathrm{ext}}}-{b_{\mathrm{ext}}}^2\right) + E(u,A)
+ {\kappa}\sqrt2
\int_{Q_{\Lz,T}}|{\mathcal{D}}_A^3u|^2-|\nabla_A'u|^2 dx,$$ where $$\begin{aligned}
E(u,A)&:=&
\int_{Q_{\Lz,T}} |\nabla_A u |^2 + \left(B_3 - \frac{{\kappa}}{\sqrt2} (1-\rho)\right)^2 dx
\\&&
+ \int_{Q_{\Lz}\times \R} |B'|^2 dx+
\int_{Q_{\Lz}\times(\R\setminus(-T,T))} (B_3 -{b_{\mathrm{ext}}})^2 dx\,.\end{aligned}$$ In particular, the bulk energy is $2{\Lz}^2T({\kappa}\sqrt2 {b_{\mathrm{ext}}}-{b_{\mathrm{ext}}}^2) $. Since we are interested in the regime $\kappa\ll 1$ and since $ |{\mathcal{D}}_A^3u|^2\le 2 |\nabla_A'u|^2$, the contribution of the last term in (\[eqEGLE\]) to the energy is (asymptotically) negligible with respect to the first term in $ E$, and therefore it can be ignored in the following.
Applying to $B-{b_{\mathrm{ext}}}e_3 $ and minimizing outside $Q_{\Lz} \times [-T,T]$ if necessary, the last two terms in $E(u,A)$ can be replaced by $$\int_{Q_{{\Lz},T}} |B'|^2 dx+ \| B_3-{b_{\mathrm{ext}}}\|_{H^{-1/2}(\{x_3= T\})}^2+\| B_3-{b_{\mathrm{ext}}}\|_{H^{-1/2}(\{x_3=-T\})}^2\,,$$ so that $ E(u,A)$ becomes $$\label{E}
E(u,A)=
\int_{Q_{\Lz,T}} |\nabla_A u |^2 + \left(B_3 - \frac{{\kappa}}{\sqrt2} (1-\rho)\right)^2 + |B'|^2 dx+\| B_3-{b_{\mathrm{ext}}}\|_{H^{-1/2}(\{x_3=\pm T\})}^2\,.$$
Let us notice that the normal solution $\rho=0$, $B={b_{\mathrm{ext}}}e_3$ (for which we can take $A(x_1,x_2,x_3)={b_{\mathrm{ext}}}x_1 e_2$) is always admissible but has energy equal to $${E_\mathrm{GL}}(u,A)= \Lz^2 T \kappa^2\gg 2\Lz^2T({\kappa}\sqrt2 {b_{\mathrm{ext}}}-{b_{\mathrm{ext}}}^2),$$ in the regime $\kappa\gg {b_{\mathrm{ext}}}$ that we consider here.
The following scaling law is established in [@CoOtSer].
\[theoCoOtSer\] For ${b_{\mathrm{ext}}}<\kappa/8$, $\kappa\le 1/2$, $\kappa
T \ge 1$, $\Lz$ sufficiently large, if the quantization condition $$\label{quantificationini}
{b_{\mathrm{ext}}}\Lz^2 \in 2\pi \Z,$$ holds then $$\label{minE}
\min E(u,A) \sim \min \left\{{b_{\mathrm{ext}}}^{2/3} \kappa^{2/3} T^{1/3} , {b_{\mathrm{ext}}}\kappa^{3/7} T^{3/7} \right\}\Lz^2.$$
We believe that is also a necessary condition for to hold. Indeed, we expect that if $(u,A)$ is such that $$E(u,A) \les \left\{{b_{\mathrm{ext}}}^{2/3} \kappa^{2/3} T^{1/3} \Lz^2, {b_{\mathrm{ext}}}\kappa^{3/7} T^{3/7} \Lz^2 \right\},$$ then the normal phase $\rho\simeq 0$ is the minority phase (typically disconnected on every slice) and there exist $x_3\in(-T,T)$ and (periodic) curves $\gamma_1$ and $\gamma_2$ such that $\Gamma_1:=\{(\gamma_1(s), s, x_3) : s\in[ 0, 1]\} {\subseteq}\{\rho\simeq 1\}$ and $\Gamma_2:=\{(s,\gamma_2(s), x_3) : s\in[ 0, 1]\} {\subseteq}\{\rho\simeq 1\}$, with $\gamma_i(1)=\gamma_i(0)+L_0 e_i$. If this holds then using Stokes Theorem on large domains the boundary of which is made of concatenations of the curves $\Gamma_i$, it is possible to prove that must hold. As in [@CoOtSer], we will need to assume in order to build the recovery sequence in Section \[recoveryGL\].\
The first regime in corresponds to uniform branching patterns while the second corresponds to well separated branching trees (see [@ChokConKoOt; @ChokKoOt; @CoOtSer]). We focus here on the first regime, that is ${\kappa}^{5/7}\ll {b_{\mathrm{ext}}}T^{2/7}$, and replace $\kappa$ and ${b_{\mathrm{ext}}}$ by the variables $\alpha$, $\beta$, defined according to $$\kappa T = \sqrt{2}\a \qquad {b_{\mathrm{ext}}}= \frac{\beta \kappa }{\sqrt{2}}=\frac{\alpha\beta}{T}\,,$$ and then rescale
------------------------------------------------------- ---------------------------
$\hat{x}:= x/T$, $ L:=\Lz/T$,
\[6pt\] $\hat{u}(\hat{x}):= u(x)$, $\hat{A}(\hat{x}):= A(x)$
\[6pt\] $E_T(\hat u,\hat A):=\frac{1}{T L^2} E(u,A)$,
------------------------------------------------------- ---------------------------
so that in particular $ \hat{B}(\hat{x})= \widehat{{\nabla}}\times \hat{A}(\hat{x})= TB(x)$ and ${\nabla}_A u(x)= T^{-1} \widehat{{\nabla}}
\hat{u}( x/T)-i \hat{A}(x/T)\hat u (x/T)$. Changing variables and removing the hats yields $$E_T(u,A)= \frac{1}{L^2}\lt(\int_{Q_{L,1}}
|{\nabla}_{T{A}} {u}|^2+ \left({B_3} - \a(1-\rho)\right)^2 + |{B'}|^2 dx+\|{B_3}- \a\b\|_{H^{-1/2}(\{x_3=\pm1\})}^2\rt)\,.$$ as was anticipated in (\[Etilde\]). In these new variables, the scaling law becomes $E_T\sim \min\{\alpha^{4/3}\beta^{2/3},\a^{10/7}\beta \}$ and the uniform branching regime corresponds to $E_T\sim \alpha^{4/3}\beta^{2/3}$ which amounts to $\a^{-2/7}\ll \b\ll 1$, see also (\[scallawintro\]). Constructions (leading to the upper bounds in [@CoOtSer; @ChokKoOt; @ChokConKoOt]), suggest that in this regime, typically, the penetration length of the magnetic field inside the superconducting regions is of the order of $T^{-1}$, the coherence length (or domain walls) is of the order of $\a^{-1}$, the width of the normal domains in the bulk is of the order of $\a^{-1/3} \b^{1/3}$ and their separation of order $\a^{-1/3}\b^{-1/6}$. These various lengthscales motivate the anisotropic rescalings that we will introduce in Section \[Sec:intermediate\].
In closing this section we present the anisotropic rescaling that will lead to the functional defined in (\[eqdefwe\]), postponing to the next section a detailed explanation of its motivation. We set for $x \in Q_{L,1}$,
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------
$\begin{pmatrix}{\widetilde}{x}'\\ {\widetilde}{x}_3\end{pmatrix}:= \begin{pmatrix}\a^{1/3} \b^{1/6} x'\\x_3\end{pmatrix}$, $ {\widetilde}L:= \a^{1/3}\b^{1/6}L$,
\[8pt\] $\begin{pmatrix}{\widetilde}{A}'\\ {\widetilde}{A}_3\end{pmatrix}({\widetilde}{x}):=\begin{pmatrix}\a^{-2/3}\b^{1/6} A'\\ \a^{-1/3} \b^{1/3}A_3\end{pmatrix}(x)$, $ {\widetilde}{u}({\widetilde}{x}):=u(x),$
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------
to get ${\widetilde}B_3({\widetilde}x)=\alpha^{-1}B_3(x)$, ${\widetilde}B'({\widetilde}x)=\alpha^{-2/3}\beta^{1/6}B'(x)$ inside the sample. Outside the sample, i.e. for $|x_3|\ge 1$, we make the isotropic rescaling ${\widetilde}{x}:= \pm e_3+\a^{1/3}\b^{1/6}( x \mp e_3)$ to get ${\widetilde}B({\widetilde}x)=\a^{-1} B(x)$. A straightforward computation leads to $\wE({\widetilde}u, {\widetilde}A)=\a^{-4/3}\b^{-2/3} E_T(u,A)$, where $$\begin{gathered}
\wE(u,A):=\frac{1}{\widetilde{L}^2}\Big[\int_{Q_{{\widetilde}L,1}} \a^{-2/3}\b^{-1/3}\lt|\nabla_{\a^{1/3}\b^{-1/3}TA}'u\rt|^2+
\a^{-4/3}\b^{-2/3}\lt|(\nabla_{\a^{1/3}\b^{-1/3}TA} u)_3\rt|^2\\
+ \a^{2/3}\b^{-2/3}\left({B_3} -(1-|{u}|^2)\right)^2
+ \b^{-1}|{B'}|^2 dx +\a^{1/3}\b^{7/6}\|\b^{-1}{B_3}-1\|_{H^{-1/2}(x_3=\pm1)}^2\Big],\end{gathered}$$ with ${\nabla}\times A=B$ (and in particular $\Div B=0$). We assume that ${\widetilde}L$ is a fixed quantity of order 1. For simplicity of notation, the detailed analysis is done only for the case ${\widetilde}L=1$.\
Let us point out that in these units, the penetration length is of order $T^{-1}\a^{1/3}\b^{1/6}$, the coherence length of order $\a^{-2/3}\b^{1/6}$, the width of the normal domains in the bulk of order $\b^{1/2}$ and the distance between the threads of order one. That is, the scale separation reads now $$\label{scalesepar2}
T^{-1}\a^{1/3}\b^{1/6}\ll \a^{-2/3}\b^{1/6}\ll \b^{1/2}\ll1.$$
The intermediate functionals {#Sec:intermediate}
============================
In this section we explain the origin of the rescaling leading from $E_T$ to ${\widetilde}E_T$, and the different functionals which appear at different scales. This material is not needed for the proofs but we think it is important to illustrate the meaning of our results. We carry out the scalings in detail but the relations between the functionals are here discussed only at a heuristic level.
We want to successively send $T\to+\infty$, $\alpha\to +\infty$ and $\beta\to 0$. For this we are going to introduce a hierarchy of models starting from $E_T(u,A)$ and finishing at $I(\mu)$. When sending first $T\to +\infty$ with fixed $\alpha$ and $\beta$, the functional $E_T$ approximates $$ F_{\alpha,\beta}(\rho,B):=\frac{1}{L^2} \lt(\int_{Q_{L,1}} |\nabla \rho^{1/2}|^2 +(B_3-\alpha (1-\rho))^2 +|B'|^2 dx +\|{B_3}- \a\b\|_{H^{-1/2}(x_3=\pm1)}^2\rt),$$ with the constraints $$\label{meissner4}
\Div B=0 \qquad\textrm{and } \qquad \rho B=0.$$ The main difference between $E_T$ and $F_{\alpha,\beta}$ is that for the latter, since the penetration length (which corresponds to $T^{-1}$) was sent to zero, the Meissner condition is enforced. We now want to send the coherence length (of order $\a^{-1}$) to zero at fixed $\beta$, while keeping superconducting domains of finite size. Since the typical domain diameter is of order $\a^{-1/3}\b^{1/3}$ and their distance is of order $\a^{-1/3}\b^{-1/6}$, we are led to the anisotropic rescaling:
-------------------------------------------------------------------------------------------------------------------------------- ---------------------------------
$\begin{pmatrix}\hat{x}'\\ \hat{x}_3\end{pmatrix}:= \begin{pmatrix}\a^{1/3} x'\\x_3\end{pmatrix}$, $\hat{L}:= \a^{1/3}L$,
\[6pt\] $\begin{pmatrix}\hat{B}'\\ \hat{B}_3\end{pmatrix}(\hat{x}):=\begin{pmatrix}\a^{-2/3} B'\\ \a^{-1}B_3\end{pmatrix}(x)$, $\hat{\rho}(\hat{x}):=\rho(x)$,
\[6pt\] $\wF=\a^{-4/3} F_{\a,\b}$.
-------------------------------------------------------------------------------------------------------------------------------- ---------------------------------
In these variables, the coherence length is of order $\a^{-2/3}\ll 1$ (at least horizontally) while the diameter of the normal domains is of order $\beta^{1/3}$ and their separation of order $\beta^{-1/6}$. Dropping the hats (we just keep them on the functional and on $L$ to avoid confusion) we obtain $$\begin{gathered}
\wF(\rho,B):=\frac{1}{\hL^2} \lt(\int_{Q_{\hL,1}} \a^{-2/3}\left|\begin{pmatrix} \nabla' \rho^{1/2}\\ \a^{-1/3} \partial_3 \rho^{1/2}
\end{pmatrix}\right|^2
+\a^{2/3}\lt(B_3- (1-\rho)\rt)^2 +|B'|^2 dx\rt.\\
\lt. +\a^{1/3}\|{B_3}- \b\|_{H^{-1/2}(x_3=\pm1)}^2\rt),\end{gathered}$$ with the constraints . The scaling of Theorem \[theoCoOtSer\] indicates that $\wF$ behaves as $ \min\{\beta^{2/3},\a^{2/21}\beta \}$ which is of order $\beta^{2/3}$ if $\a\gg 1$ and $\beta$ is fixed. We remark that, letting $\eta:= \alpha^{-1/3}$ and $\delta:=\eta^2=\alpha^{-2/3}$, one has $$\begin{gathered}
\wF(\rho,B)= \frac{1}{\hL^2} \left(\int_{Q_{\hL,1}} \delta \lt|\begin{pmatrix}\nabla'\rho^{1/2}\\ \eta \partial_3 \rho^{1/2}\end{pmatrix}\rt|^2+\frac{1}{\delta} \lt(B_3-(1-\rho)\rt)^2
+|B'|^2 dx \rt.\\
\lt.+ \eta^{-1} \|B_3-\beta\|_{H^{-1/2}(x_3=\pm1)}^2\rt).\end{gathered}$$ In this form, $ \wF(\rho,B)$ is very reminiscent of the functional studied in [@ViehOtt]. Notice however that besides the Meissner condition which makes our functional more rigid, the scaling $\delta=\eta^2$ is borderline for the analysis in [@ViehOtt].
Recalling that $\lt(B_3-(1-\rho)\rt)^2 = \chi_{\rho>0} (1-\rho)^2$, the corresponding term in $\wF$ has the form of a double well-potential, and so in the limit $\alpha\to+\infty$ the functional $\wF$ approximates $$ G_\beta(\chi,B'):=\frac{1}{\hL^2} \lt( \int_{Q_{\hL,1}} \frac{4}{3}|D'\chi| +|B'|^2 dx \rt),$$ with the constraints $\chi\in \{0,1\}$, $\chi(\cdot,x_3) \weaklim \beta dx'$ when $x_3\to \pm1$ and $$ \partial_3 \chi +\Divp B'=0 \qquad \textrm{and } \qquad \chi B'=B'.$$ This is similar to the simplified sharp-interface functional that was studied in [@ChokKoOt; @ChokConKoOt]. In the definition of $G_\beta$, we used the notation $$ \int_{Q_{\hL,1}} |D' u|:= \sup_{\stackrel{\xi \in C^{\infty}(Q_{\hL,1}),}{ |\xi|_\infty \le 1}} \, \int_{Q_{\hL,1}} u \Divp \xi \, dx,$$ for the horizontal $BV$ norm of a function $u\in L^1(Q_{\hL,1})$. By definition it is lower semicontinuous for the $L^1$ convergence and it is not hard to check that if we let $u_{x_3}(x'):= u(x',x_3)$, then $$\int_{Q_{\hL,1}} |D'u|=\int_{-1}^1 \left(\int_{Q_{\hL}} |D' u_{x_3}|\right) dx_3,$$ where $\int_{Q_{\hL}} |D' u_{x_3}|$ is the usual $BV$ norm of $u_{x_3}$ in $Q_{\hL}$ [@AFP]. From this and the usual co-area formula [@AFP Th. 3.40], we infer that $$\label{coarea}
\int_{Q_{\hL,1}} |D'u|=\int_{\R} \int_{-1}^1 \H^1( \partial \{u_{x_3}>s\}) dx_3 ds.$$ In , $\partial \{u_{x_3}>s\}$ represents the measure-theoretic boundary of $\{u_{x_3}>s\}$ in $Q_{\hL}$.
We finally want to send the volume fraction of the normal phase to zero and introduce the last rescaling in $\beta$ for which we let
-------------------------------------------------------------------------------------------------------------------- ----------------------------------------------
$\begin{pmatrix}{\widetilde}{x}'\\ {\widetilde}{x}_3\end{pmatrix}:= \begin{pmatrix}\b^{1/6} x'\\x_3\end{pmatrix}$, $
{\widetilde}L:= \beta^{1/6}\hL$,
\[8pt\] $\tB'({\widetilde}{x}):=\beta^{1/6} B'(x)$ $\tchi({\widetilde}{x}):=\beta^{-1}\chi(x)$,
\[6pt\] $\wG:=\b^{-2/3} G_{\b}$.
-------------------------------------------------------------------------------------------------------------------- ----------------------------------------------
After this last rescaling, the domain width is of order $\b^{1/2}\ll1$, the separation between domains of order $1$. We obtain (dropping the tildas again) the order-one functional $$ \wG(\chi,B'):= \frac{1}{\widetilde{L}^2}\int_{Q_{{\widetilde}L,1}} \frac{4}{3} \beta^{1/2}|D'\chi| +\chi|B'|^2 dx$$ under the constraints $\chi\in \{0,\beta^{-1}\}$, $\chi(\cdot,x_3) \weaklim dx'$ when $x_3\to \pm1$, and $$ \partial_3 \chi +\Divp (\chi B')=0 \qquad \textrm{and } \qquad \chi B'=\beta^{-1}B'.$$ This functional converges to $I(\mu)$ as $\beta\to 0$.
Let us point out that since we are actually passing directly from the functional $\wE$ to $I$ in Theorem \[maintheo\], we are covering the whole parameter regime of interest. In particular, our result looks at first sight stronger than passing first from $ E_T$ to $F_{\a,\b}$, then from $\wF$ to $G_\beta$ and finally from $\wG$ to $I$. However, because of the Meissner condition, we do not have a proof of density of smooth objects for $F_{\a,\b}$ and $G_\b$. Because of this, we do not obtain the $\Gamma-$convergence of the intermediate functionals (the upper bound is missing).
The limiting energy {#Sec:limiting}
===================
Before proving the $\Gamma$-limit we study the limiting functional $I$ that was mentioned in (\[eqdefIintro\]) and motivated in the previous section. We give here a self-contained treatment of the functional $I$, which is motivated by the analysis discussed above, and will be crucial in the proofs that follow. However, in this discussion we do not make use of the relation to the Ginzburg-Landau functional.
For $L,T>0$ we denote by ${\mathcal{A}}_{L,T}$ the set of pairs of measures ${\mu}\in{\mathcal{M}}^+(Q_{L,T})$, $m\in {\mathcal{M}}(Q_{L,T};\R^2)$ with $m\ll\mu$, satisfying the continuity equation $$\label{conteq0}
\pa_3 {\mu}+\Divp m=0 \qquad \textrm{in } Q_{L,T},$$ and such that ${\mu}={\mu}_{x_3}\otimes d{x_3}$ where, for a.e. $x_3\in(-T,T)$, ${\mu}_{x_3}=\sum_i {\varphi}_i \delta_{X_i}$ for some ${\varphi}_i> 0$ and $X_i\in Q_L$. We denote by ${\mathcal{A}}_{L,T}^*:=\{\mu: \exists m, (\mu,m)\in{\mathcal{A}}_{L,T}\}$ the set of admissible $\mu$.
We define the functional $I:{\mathcal{A}}_{L,T}\to[0,+\infty]$ by $$\label{Imum}
I(\mu,m):=
\frac{K_*}{L^2} \int_{-T}^T \sum_{x'\in Q_L} \left(\mu_{x_3}(x')\right)^{1/2} \, dx_3 +
\frac{1}{L^2}\int_{Q_{L,T}}
\left(\frac{dm}{d{\mu}} \right)^2 d{\mu},$$ where $K_*:= \frac{8 \sqrt{\pi}}{3}$ and (with abuse of notation) $I:{\mathcal{A}}_{L,T}^*\to[0,\infty]$ by $$\label{Imu}
I({\mu}):=\min \{ I(\mu,m)\ : \ m\ll{\mu}, \ \pa_3 {\mu}+ {\Divp} m=0\}.$$
Condition (\[conteq0\]) is understood in a $Q_L$-periodic sense, i.e., for any $\psi\in C^1(\R^3)$ which is $Q_L$-periodic and vanishes outside $\R^2\times (0,T)$ one has $\int_{Q_{L,T}} \partial_3\psi d\mu + \nabla'\psi \cdot dm=0$. If $\mu_{x_3}=\sum_i {\varphi}_i \delta_{X_i}$ then $\sum_{x'\in Q_L} \left(\mu_{x_3}(x')\right)^{1/2}=\sum_i {\varphi}_i^{1/2}$. Because of , $\mu_{x_3}(Q_L)$ does not depend on $x_3$.
Let us point out that the minimum in (\[Imu\]) is attained thanks to (\[eqexistsB\]). Moreover, the minimizer is unique by strict convexity of $m\to\int_{Q_{L,T}} \left(\frac{dm}{d{\mu}} \right)^2 d{\mu}$. As proven in Lemma \[lemmacurves\] below, if $\mu$ is made of a finite union of curves then there is actually only one admissible measure $m$ for . More generally, Since every measure $\mu$ with finite energy is rectifiable (see Corollary \[coroXia\]), we believe that it is actually always the case. For $\mu$ an admissible measure and $z,{\widetilde}{z}\in[-T,T]$, we let $$\label{Iztz}
I^{(z,{\widetilde}{z})}(\mu):= \frac{K_*}{L^2} \int_{z}^{{\widetilde}{z}} \sum_{x'\in Q_L} \left(\mu_{x_3}(x')\right)^{1/2} \, dx_3 +
\frac{1}{L^2}\int_{Q_{L}\times[z,{\widetilde}{z}]}
\left(\frac{dm}{d{\mu}} \right)^2 d{\mu},$$ where $m$ is the optimal measure for ${\mu}$ on $[z,{\widetilde}{z}]$ (which coincides with the restriction to $[z,{\widetilde}{z}]$ of the optimal measure on $[-T,T]$).
From (\[eqw22b\]) one immediately deduces for every measure $\mu$, and every $x_3, {\widetilde}x_3\in[-T,T]$, the following estimate on the Wasserstein distance $$\label{HolderW2}
W_2^2(\mu_{x_3}, {\mu}_{{\widetilde}x_3})\le L^2 I(\mu) |x_3-{\widetilde}x_3|.$$ In particular for every measure $\mu$ with $I(\mu)<+\infty$, the curve $x_3\mapsto \mu_{x_3}$ is Hölder continuous with exponent $1/2$ in ${\mathcal{M}}^+( Q_L)$ (endowed with the metric $W_2$) and the traces $\mu_{\pm T}$ are well defined.
Existence of minimizers
-----------------------
Given two measures $\bar {\mu}_{\pm}$ in $\M^+(Q_L)$ with $\bar \mu_+(Q_L)=\bar \mu_-(Q_L)$, we are interested in the variational problem $$\label{limitProb}
\inf\lt\{ I(\mu) \ : \ \mu_{\pm T}= \bar {\mu}_{\pm} \rt\}.$$ We first prove that any pair of measures with equal flux can be connected with finite cost and that there always exists a minimizer. The construction is a branching construction which gives the expected scaling (see [@ChokConKoOt; @CoOtSer]) if the boundary data is such that $\bar {\mu}_+=\bar{\mu}_-$.
\[branchingmu\] For every pair of measures $ \bar {\mu}_{\pm}\in {\mathcal{M}}^+(Q_L)$ with $\bar\mu_+(Q_L)=\bar\mu_-(Q_L)=\Phi$, there is $\mu\in {\mathcal{A}}^*_{L,T}$ such that $\mu_{\pm T}=\bar\mu_\pm$ and $$I(\mu)\les \frac{T\Phi^{1/2}}{L^2}+\frac{\Phi}{T}.$$ If $\bar\mu_+=\bar\mu_-$, then there is a construction with $$I(\mu)\les \frac{T \Phi^{1/2}}{L^2} +\frac{ T^{1/3} \Phi^{2/3} }{L^{4/3}},$$ and such that the slice at $x_3=0$ is given by $\mu_0=\Phi N^{-2}\sum_j \delta_{X_j}$, with $X_j$ the $N^2$ points in $[-L/2,L/2)^2\cap ((L/N)\Z)^2$, and $N:=\lfloor1+ \Phi^{1/6}L^{2/3}/T^{2/3}\rfloor$. The measure $\mu$ is supported on countably many segments, which only meet at triple points.
By periodicity we can work on $[0,L)^2$ instead of $[-L/2,L/2)^2$. We first perform the construction for $x_3\ge 0$. The idea is to approximate $\bar\mu_+$ by linear combinations of Dirac masses, which become finer and finer as $x_3$ approaches $T$. Fix $N\in\N$, chosen below. For $n\in\N$, fix $x_{3,n}:=T(1-3^{-n})$, and let $T_n:= x_{3,n}-x_{3,n-1}=\frac{2}{3^{n}}T$ be the distance between two consecutive planes. At level $x_{3,n}$ we partition $Q_L$ into squares of side length $L_n:= \frac{L}{2^nN}$. More precisely, for $i,j=0,...,2^nN-1 $, we let $x'_{ij,n}:=\lt(L_n\, i, L_n \, j\rt)$ be a corner of the square $Q_{ij,n}:= x'_{ij,n}+ [0,L_n)^2$, and we let $\Phi_{ij,n}:=\bar\mu_+(Q_{ij,n})$ be the flux associated to this square.
We define the measures $\mu^\mathrm{br}$ and $m^\mathrm{br}$ (here the suffix $\mathrm{br}$ stands for branching) by $$\mu^\mathrm{br}_{x_3}:= \sum_{ij} \Phi_{ij,n} \delta_{X_{ij,n}(x_3)} \text{ and }
m^\mathrm{br}_{x_3}:= \sum_{ij} \frac{d X_{ij,n}}{dx_3}(x_3) \Phi_{ij,n} \delta_{X_{ij,n}(x_3)} \quad
\text{for } x_3\in [x_{3,n-1}, x_{3,n}),$$ where $X_{ij,n}:[x_{3,n-1}, x_{3,n}]\to Q_L$ is a piecewise affine function such that $X_{ij,n}(x_{3,n})=x'_{ij,n}$, $X_{ij,n}(x_{3,n}-\frac12 T_n)=x'_{i_*j,n}$, and $X_{ij,n}(x_{3,n-1})=x'_{i_*j_*,n}$, where $i_*=2\lfloor i/2\rfloor$, $j_*=2\lfloor j/2\rfloor$. Four such curves end in every $i_*$, $j_*$ (which corresponds to the pair $i_*/2$, $j_*/2$ at level $n-1$), but they are pairwise superimposed for $x_3\in [x_{3,n-1},x_{3,n}-\frac12 T_n]$, therefore all junctions are triple points (one curve goes in, two go out).
Using that $\sum_{i j} \Phi_{ij,n}=\Phi$ and $\sum_{ij}\sqrt{\Phi_{ij,n}} \le
(\sum_{ij} \Phi_{ij,n})^{1/2}(\sum_{ij} 1)^{1/2} = \Phi^{1/2} 2^nN$, we get that the energy of $\mu^\mathrm{br}$ is given by $$\begin{aligned}
I(\mu^\mathrm{br})&=\frac{1}{L^2} \sum_{n=1}^{+\infty} \sum_{ij} \left( K_* T_n \sqrt{\Phi_{ij,n}}+ \Phi_{ij,n} T_n \frac{2 L_n^2}{ T_n^2}\rt)\\
& \les L^{-2} TN\Phi^{1/2} \sum_{n=0}^{+\infty} \lt(\frac{2}{3}\rt)^n + \frac{\Phi}{TN^2}\sum_{n=0}^{+\infty} \lt(\frac{3}{4}\rt)^n\,.\end{aligned}$$ If we choose $N=1$, then there is only one point in the central plane, $\mu_0=\Phi \delta_0$. Therefore the top and bottom constructions can be carried out independently, since by assumption the total flux is conserved, and we obtain the first assertion.
If $\bar\mu_+=\bar\mu_-$, we can choose the value of $N$ which makes the energy minimal. Up to constants this is the value given in the statement. Inserting in the estimate above gives the second assertion.
If the boundary densities are maximally spread, in the sense that they are given by the Lebesgue measure, the scaling is optimal, as the following lower bound shows.
\[branchinglowerbound\] For every measure $\mu\in {\mathcal{A}}^*_{L,T}$ such that $\mu_{\pm T}=\Phi L^{-2} dx'$ one has $$\label{lower}
I(\mu)\ges \frac{T \Phi^{1/2}}{L^2}+ \frac{ T^{1/3}\Phi^{2/3}}{ L^{4/3}}\, .$$
The bound $I(\mu)\ge L^{-2} T\Phi^{1/2}$ follows at once from the subadditivity of the square root. Hence we only need to prove the other one. We give two proofs of this bound. The first uses only elementary tools while the second is based on an interpolation inequality.
[*First proof:*]{} Let $I:=I(\mu,v\mu)$, where $v:=dm/d\mu$. Fix $\lambda>0$, chosen below. Choose $x_3\in (-T,T)$ such that $\mu_{x_3}=\sum_i\varphi_i \delta_{X_i}$ obeys $$\label{eqlbchoicex3}
\sum_i \varphi_i^{1/2} \le \frac{L^2 I}{T}\,.$$ For some set ${\mathcal{I}}{\subseteq}\N$ to be chosen below, let $\psi:Q_L\to \R$ be a mollification of the function $\max\{ (\lambda-\dist(x',X_i))_+: i\in {\mathcal{I}}\}$, where as usual the distance is interpreted periodically. By the divergence condition, $$\int_{Q_L} \psi d\mu_{x_3} = \frac{\Phi}{L^{2}} \int_{Q_L} \psi dx'+\int_{-T}^{x_3} \int_{Q_L} \nabla'\psi\cdot v d\mu$$ (to prove this, pick $\xi_\eps\in C^1_c((-T,x_3))$ which converge pointwise to $1$ and use $\xi_\eps\psi$ as a test function in (\[conteq0\]) and then pass to the limit). Since $|\nabla'\psi|\le 1$, $$\sum_{i\in {\mathcal{I}}} \lambda \varphi_i \le \frac{\Phi}{L^{2}} \sum_{i\in {\mathcal{I}}}\frac{ \pi}{3} \lambda^3+ \int_{-T}^{x_3} \int_{Q_L} |v| d\mu
\le \frac{\Phi}{L^2}\sum_{i\in {\mathcal{I}}} \pi \lambda^3+ L (T\Phi)^{1/2} I^{1/2},$$ where in the second step we used Hölder’s inequality and flux conservation. We choose ${\mathcal{I}}=\{i\in \N: \varphi_i\ge 4 \Phi \lambda^2/L^2\}$. From the definition of ${\mathcal{I}}$, we have $$\frac{\Phi}{L^2}\sum_{i\in {\mathcal{I}}} \pi \lambda^3\le \frac{\pi}{4} \sum_{i\in {\mathcal{I}}} \lambda {\varphi}_i.$$ Therefore, since $\pi<4$, we obtain $$\label{E1}
\sum_{i\in {\mathcal{I}}} \lambda \varphi_i {\stackrel{<}{\sim}}L(T\Phi I)^{1/2}\,.$$ At the same time, again by the definition of ${\mathcal{I}}$ and (\[eqlbchoicex3\]), $$\label{E2}
\sum_{i\not\in {\mathcal{I}}} \varphi_i\le \frac{2\lambda\Phi^{1/2}}{L} \sum_{i\not\in {\mathcal{I}}} \varphi_i^{1/2} \le 2\lambda L \Phi^{1/2}\, \frac{I}{T}\,.$$Adding and , we obtain $$\sum_{i\in {\mathcal{I}}} \varphi_i {\stackrel{<}{\sim}}\frac{1}{\lambda} L(T\Phi I)^{1/2}+ \lambda L \Phi^{1/2}\, \frac{I}{T}\, ,$$ hence $$\Phi^{1/2}{\stackrel{<}{\sim}}\frac{1}{\lambda} L(T I)^{1/2}+ \lambda L \, \frac{I}{T}\, ,$$ and optimizing over $\lambda$ by choosing $\lambda = \frac{T^{3/4}}{I^{1/4}}$ yields $I\ges \Phi^{2/3} T^{1/3}L^{-4/3}$.
[*Second Proof:*]{} As before, let $x_3\in (-T,T)$ be such that $\mu_{x_3}=\sum_i\varphi_i \delta_{X_i}$ obeys . By Young’s inequality and , we have $$I\ges \frac{T}{L^2} \left( \sum_i \varphi_i^{1/2} \right) +\frac{W_2^2(\mu_{x_3}, \Phi L^{-2} dx')}{L^2 T}\ges L^{-2} T^{1/3}\left( \sum_i \varphi_i^{1/2} \right)^{2/3} \left(W_2^2(\mu_{x_3}, \Phi L^{-2} dx')\right)^{1/3}.$$ The desired lower bound would then follow if we can show that for every measure $\mu\in\M^+(Q_L)$ with $\mu=\sum_i {\varphi}_i \delta_{X_i}$ and $\sum_i {\varphi}_i = \Phi$, $$\label{interpolBVW2}
\left( \sum_i \varphi_i^{1/2} \right)^{2/3} \left(W_2^2(\mu, \Phi L^{-2} dx')\right)^{1/3}\ges \Phi^{2/3} L^{-4/3}.$$ By rescaling it is enough considering $\Phi=L=1$. The optimal transport map is necessarily of the form $\psi(x')=X_i$ if $x'\in E_i$, where $E_i$ is a partition of $Q_1$ with $|E_i|={\varphi}_i$ (the corresponding transport plan is $(Id\times \psi)\sharp dx'$). By definition, it holds $$W_2^2(\mu, dx')\ge \sum_{i} \int_{E_i} |x'-X_i|^2 dx'.$$ But since $|E_i|={\varphi}_i=|\B'(X_i,({\varphi}_i/\pi)^{1/2})|$, $$\sum_{i} \int_{E_i} |x'-X_i|^2 dx'\ge \sum_{i} \int_{\B(X_i,({\varphi}_i/\pi)^{1/2})} |x'|^2 dx'\ges \sum_{i} {\varphi}_i^2 .$$ By Hölder’s inequality, we conclude that $$1=\sum_i {\varphi}_i \le \left(\sum_i {\varphi}_i^{1/2}\right)^{2/3} \left(\sum_i {\varphi}_i^2\right)^{1/3}\les \left( \sum_i \varphi_i^{1/2} \right)^{2/3} \left(W_2^2(\mu, dx')\right)^{1/3},$$ as desired.
The lower bound can also be obtained as a consequence of the scaling law proven in [@CoOtSer] for the Ginzburg-Landau model combined with our lower bound in Section \[Sec:gammaliminf\] (which does not use this lower bound). However, since the proof here is much simpler and contains some of the main ideas behind the proofs of [@ChokConKoOt; @CoOtSer], we decided to include it. Similarly, the interpolation inequality can be obtained by approximation from a similar inequality proven in [@CintiOt] (where it is used in the same spirit as here to re-derive the lower bounds of [@ChokConKoOt]).
We end this section by proving existence of minimizers.
\[existmu\] For every pair of measures $ \bar {\mu}_{\pm }$ with $\bar {\mu}_+(Q_L)=\bar {\mu}_-(Q_L)$, the infimum in is finite and attained.
In this proof we assume $L=T=1$ and $\bar\mu_+(Q_{1})=1$. By Proposition \[branchingmu\] the infimum is finite. Let now $\mu^n$ be a minimizing sequence for $I$. Since $\sup_n I(\mu^n)<+\infty$, thanks to , the functions $x_3\mapsto \mu^n_{x_3}$ are equi-continuous in $\cP(Q_1)$ (recall that $W_2$ metrizes the weak convergence in $\mathcal{P}(Q_1)$) hence by the Arzelà-Ascoli theorem there exists a subsequence, still denoted $\mu^n$, uniformly converging (in $x_3$) to some measure $\mu$ which also satisfies the given boundary conditions. Moreover, if $m^n$ is an optimal measure in for $\mu^n$, since by the Cauchy-Schwarz inequality we have $$\int_{Q_{1,1}} d|m^n|\le \lt(\int_{Q_{1,1}} \lt(\frac{dm^n}{d\mu^n}\rt)^2 d\mu^n\rt)^\hal \lt(\int_{Q_{1,1}} d\mu^n\rt)^\hal\les 1,$$ there also exists a subsequence $m^n$ converging to some measure $m$ satisfying . By [@AFP Th. 2.34 and Ex. 2.36] we deduce that $m\ll \mu$ and $$\liminf_{n\to +\infty} \int_{Q_{1,1}} \lt(\frac{dm^n}{d\mu^n}\rt)^2 d\mu^n\ge \int_{Q_{1,1}} \lt(\frac{dm}{d\mu}\rt)^2 d\mu.$$ It remains to prove that $\mu_{x_3}=\sum_i {\varphi}_i(x_3) \delta_{X_i(x_3)}$ for a.e. $x_3$ and that $$\label{semicontperi}
\liminf_{n\to +\infty} \int_{-1}^1 \sum_{x'\in Q_{1}} \left(\mu^n_{x_3}(x')\right)^{1/2} \, dx_3 \ge \int_{-1}^1 \sum_{x'\in Q_{1}} \left(\mu_{x_3}(x')\right)^{1/2} \, dx_3.$$ If $\mu^n_{x_3}=\sum {\varphi}_i^n(x_3) \delta_{X^n_i}(x_3)$, with ${\varphi}_i^n$ ordered in a decreasing order, we let $f_n(x_3):= \sum_i \sqrt{{\varphi}_i^n(x_3)}$ and observe that $\int_{-1}^1 f_n dx_3 \le I(\mu^n)\les 1 $. Hence, by Fatou’s lemma, $$\label{fatouf}
1\ges \liminf_{n\to +\infty} \int_{-1}^1 f_n(x_3) dx_3\ge \int_{-1}^1 \liminf_{n\to +\infty} f_n(x_3) dx_3,$$ from which we infer that $g(x_3):=\liminf_{n\to+\infty} f_n(x_3)$ is finite for a.e. $x_3$. Consider such an $x_3$ and let $\psi(n)$ be a subsequence (which depends on $x_3$) such that $g(x_3)=\lim_{n\to+\infty} f_{\psi(n)}(x_3)$. Up to another subsequence, still denoted $\psi(n)$, we may assume that for every $i\in \N$, ${\varphi}_i^{\psi(n)}(x_3)$ converges to some ${\varphi}_i(x_3)$ and $X_i^{\psi(n)}(x_3)$ converges to some $X_i(x_3)$. By Lemma \[lemsqrt\] (see below), for every $N\in \N$, $$\sum_{i\le N} {\varphi}_i^{\psi(n)}(x_3)\ge 1- \frac{f_{\psi(n)}(x_3)}{\sqrt{N}}.$$ This implies, by tightness, $\mu_{x_3}^{\psi(n)}\weaklim \sum_i {\varphi}_i(x_3) \delta_{X_i(x_3)}$ and $\sum_i ({\varphi}_i(x_3))^{1/2}\le g(x_3)$. But since $\mu_{x_3}^{\psi(n)}\weaklim {\mu}_{x_3}$, we have ${\mu}_{x_3}=\sum_i {\varphi}_i(x_3) \delta_{X_i(x_3)}$. Finally, by the subadditivity of the square root, and the definition of $g$ we obtain .
\[lemsqrt\] If a nonincreasing sequence of positive numbers $\gamma_i$ is such that $$\sum_i \gamma_i = c_0 \qquad \text{ and }\qquad \sum_i
\sqrt{\gamma_i} \le C_0,$$ then for all $N\in \N$ one has $$\sum_{i\le N} \gamma_i \ge c_0- C_0 \sqrt{\frac{c_0}{N}}.$$
Indeed $\sum_{i >N} \gamma_i \le \sqrt{\gamma_N} \sum_{i >N}
\sqrt{\gamma_i} \le C_0 \sqrt{\gamma_N},$ while $c_0 \ge \sum_{i\le
N} \gamma_i \ge N \gamma_N$.
Regularity of minimizers
------------------------
We now want to prove regularity of the minimizing measures $\mu$. In order to prove that we can restrict our attention to measures containing no loops, we first define the notion of subsystem.
\[Existence of subsystems\]\[def:subsystem\] Given a point $x:=(X, x_3) \in Q_{L,T}$ and $\mu\in{\mathcal{A}}^*_{L,T}$ with $I(\mu)<+\infty$, there exists a subsystem ${\widetilde}{{\mu}}$ of ${\mu}$ emanating from $x$, meaning that there exists ${\widetilde}{{\mu}} \in {\mathcal{A}}^*_{L,T}$ such that
1. \[def:subsystempos\] $ {\widetilde}{{\mu}}\le {\mu}$ in the sense that ${\mu}-{\widetilde}{{\mu}}$ is a positive measure,
2. \[def:subsystemdelta\] ${\widetilde}{{\mu}}_{x_3}=a\delta_{X}$, where $a={\mu}_{x_3}(X)$,
3. \[def:subsystemconti\] $$\pa_3 {\widetilde}{{\mu}}+ \div' \lt(\frac{dm}{d{\mu}} {\widetilde}{{\mu}}\rt)=0.$$
In particular, \[def:subsystemdelta\] implies that $({\mu}_{x_3} - {\widetilde}{{\mu}}_{x_3}) \perp \delta_{X}$ in the sense of the Radon-Nikodym decomposition.
Let us for notational simplicity assume that $x_3=0$, $L=T=1$, $\mu(Q_{1,1})=2$. Let us denote $v= \frac{d m}{d{\mu}}$. According to [@AGS Th. 8.2.1 and (8.2.8)], since $v\in L^2(Q_{1,1},\mu)$, there exists a positive measure $\sigma$ on $ C^0([-1, 1];Q_1)$ (endowed with the sup norm), whose disintegration [@AGS Th. 5.3.1] with respect to $\mu_0$, i.e. $\sigma=\int_{Q_1} \sigma_{x'} d\mu_0(x')$, is made of probability measures $\sigma_{x'}$ concentrated on the set of curves $ \gamma$ solving $$\left\{ \begin{array}{l}
\dot{\gamma}(x_3)= v(\gamma(x_3))\\
\gamma(0)= x',\end{array}\right.$$ and such that for every $x_3\in [-1,1]$, ${\mu}_{x_3}= (e_{x_3})_\# \sigma$, where $e_{x_3}$ denotes the evaluation at $x_3$, in the sense that $$\int_{Q_1}\varphi d\mu_{x_3} = \int_{ C^0([-1, 1];Q_1)} \varphi(\gamma(x_3)) d\sigma (\gamma)
\hskip1cm \text{ for all }\varphi\in C^0(Q_1)\,.$$ Then, the measure ${\widetilde}{{\mu}}= {\widetilde}{{\mu}}_{x_3}\otimes dx_3$ with ${\widetilde}{\mu}_{x_3}= (e_{x_3})_\# (a \sigma_{X})$, where $a=\mu_{0}(X)$, satisfies all the required properties.
\[noloop\] Let ${\mu}$ be a minimizer for the Dirichlet problem (\[limitProb\]), $\bar x_3\in (-T,T)$. Let $x_1=(X_1,\bar{x}_3)$, $x_2=(X_2,\bar{x}_3)$ be two points in the plane $\{x:x_3=\bar{x}_3\}$. Let ${\mu}_1$ and ${\mu}_2$ be subsystems of ${\mu}$ emanating from $x_1$ and $x_2$. Let $x_+:=(X_+,z_+)$ be a point with $ z_+ > \bar{x}_3$ and $x_-:=(X_-,z_-)$ a point with $ z_-<\bar{x}_3$, and such that ${\mu}_{1,z_+}, {\mu}_{2,z_+}$ both have Diracs at $X_+$, and ${\mu}_{1,z_-}, {\mu}_{2,z_-}$ both have Diracs at $X_-$ (with nonzero mass). Then $X_1=X_2$.
Let ${\varphi}_1:=\mu_{\bar x_3}(X_1)$ be the mass of ${\mu}_1$ and ${\varphi}_2$ be the mass of $\mu_2$. Let ${\varphi}_{1, +}:={\mu}_{1,\bar x_3}(X_+)$ be the mass of ${\mu}_1$ at $x_+$, ${\varphi}_{2, +}$ the mass of ${\mu}_2$ at $x_+$, ${\varphi}_{1, - } $ the mass of ${\mu}_1$ at $x_-$, ${\varphi}_{2,-}$ the mass of ${\mu}_2$ at $x_-$. Let ${\varphi}:= \min \{{\varphi}_{1, +},{\varphi}_{1, -}, {\varphi}_{2, +}, {\varphi}_{2,-}\}$ which by assumption is positive.
We define ${\mu}_{1, +}$ as the subsystem of ${\mu}_1$ coming from $x_+$, it is thus of mass ${\varphi}_{1, +}$, and at level $\bar{x}_3$ all its mass is at $X_1$ (since it is a subsystem of ${\mu}_1$ for which this is the case). Similarly with ${\mu}_{1, -}$, ${\mu}_{2, +}, {\mu}_{2,-}$. We can now define ${\widetilde}{{\mu}}_1:= \frac{{\varphi}}{{\varphi}_{1, +} } {\mu}_{1, +}$ for $x_3\ge\bar{x}_3$ and ${\widetilde}{{\mu}}_1:=\frac{{\varphi}}{{\varphi}_{1, -} }{\mu}_{1, -}$ for $x_3 <\bar{x}_3$, and the same with ${\widetilde}{{\mu}}_2$. The measures ${\widetilde}{{\mu}}_1$ and ${\widetilde}{{\mu}}_2$ are “systems” of mass ${\varphi}$ that join $X_-$ and $X_+$. By construction, we have $$\partial_{x_3} ( {\widetilde}{{\mu}}_1- {\widetilde}{{\mu}}_2 ) \restr[z_-,z_+]+\div' \lt( ( {\widetilde}{{\mu}}_1- {\widetilde}{{\mu}}_2 ) \restr[z_-,z_+]\frac{dm}{d\mu}\rt)=0$$ and $$-\frac{\varphi}{\min(\varphi_{2,+}, \varphi_{2,-})}\mu \le ( {\widetilde}{{\mu}}_1- {\widetilde}{{\mu}}_2 ) \restr[z_-,z_+]\le \frac{\varphi}{\min(\varphi_{1,+}, \varphi_{1,-})}\mu.$$
We now define $\hat{{\mu}}_\eta:= {\mu}+ \eta ( {\widetilde}{{\mu}}_1- {\widetilde}{{\mu}}_2 ) \restr[z_-,z_+]$, which is admissible for $\eta$ small enough [(and different from $\mu$ unless $X_1=X_2$)]{}, and evaluate $$\begin{gathered}
I(\hat{{\mu}}_\eta)- I({\mu}) =
K_*\int_{z_-}^{z_+}
\sum_{x'\in Q_1}\left( \mu_{x_3} (x') +\eta ({\widetilde}{\mu}_1)_{x_3}(x')-\eta ({\widetilde}{\mu}_2)_{x_3}(x') \right)^\hal
- \sum_{x'\in Q_1}\left( \mu_{x_3} (x') \right)^\hal \, dx_3 \\
+\eta \int_{Q_1 \times [z_-, z_+] }
\left(\frac{dm}{d{\mu}} \right)^2
( d{\widetilde}{{\mu}}_1- \, d{\widetilde}{{\mu}}_2)\,.
\end{gathered}$$ But the function $\eta \mapsto \sqrt{a+\eta b}$ is strictly concave for $a>0$ and $b\ne0$, therefore $I(\hat{{\mu}}_\eta)+I(\hat{{\mu}}_{-\eta})<2I({\mu})$ for any $\eta\ne 0$, a contradiction with the minimality of ${\mu}$.
A consequence of this lemma is the following. Consider a minimizing measure $\mu$ of $I$. Let $z_-$ and $z_+$ be any two slices and let $X_-$ be one of the Diracs at slice $z_-$. Let ${\widetilde}{{\mu}}$ be a subsystem emanating from $(X_-,z_-)$. Let $X_+$ be any point in the slice $z_+$ where ${\widetilde}{{\mu}}$ carries mass. Then, there is a unique “path" connecting $X_-$ to $X_+$ (otherwise there would be a loop). Since this is true for any couple of “sources" in two different planes, this means that there are at most a countable number of absolutely continuous curves (absolutely continuous because of the transport term) on which ${\mu}\restr[z_-, z_+]$ is concentrated. So we have a representation of the form $$\label{representationmu}
{\mu}= \sum_i \frac{\varphi_i}{\sqrt{1+|{\dot{X}}_i|^2}} \, \mathcal{H}^1\restr\Gamma_i \, ,$$ where the sum is countable and $\Gamma_i=\{ (X_i(x_3),x_3) : x_3\in [a_i,b_i]\} $ with $X_i$ absolutely continuous and almost everywhere non overlapping.
Another consequence is that if there are two levels at which ${\mu}$ is a finite sum of Diracs, then it is the case for all the levels in between. So, if there is a slice with an infinite number of points, then either it is also the case for all the slices below or for all the slices above.
For measures which are concentrated on finitely many curves we obtain a simple representation formula for $I(\mu)$.
\[lemmacurves\] Let $\mu=\sum_{i=1}^N \frac{\varphi_i}{\sqrt{1+|{\dot{X}}_i|^2}} \, \mathcal{H}^1\restr \Gamma_i \in{\mathcal{A}}^*_{L,T}$ with $\Gamma_i=\{ (X_i(x_3),x_3) : x_3\in [a_i,b_i]\}$ for some absolutely continuous curves $X_i$, almost everywhere non overlapping. Every ${\varphi}_i$ is then constant on $[a_i,b_i]$ and we have conservation of mass. That is, for $x:=(x',x_3)$, letting $$\begin{aligned}
{\mathcal{I}}^+(x)&:=\{ i\in[1,N] \ : \ x_3=b_i, \ X_i(b_i)=x'\}, \\ {\mathcal{I}}^-(x)&:=\{ i\in[1,N] \ : \ x_3=a_i, \ X_i(a_i)=x'\}, \end{aligned}$$ it holds $$\sum_{i\in{\mathcal{I}}^-(x)} {\varphi}_i=\sum_{i\in{\mathcal{I}}^+(x)} {\varphi}_i.$$ Moreover, $m=\sum_i \frac{\varphi_i}{\sqrt{1+|{\dot{X}}_i|^2}}
{\dot{X}}_i \, \mathcal{H}^1\restr\Gamma_i$ and $$\label{Iparticul}
I(\mu)=\frac{1}{L^2}\sum_i \int_{a_i}^{b_i} K_* \sqrt{{\varphi}_i}+ {\varphi}_i |{\dot{X}}_i|^2 dx_3.$$
Let $\bar x=(\bar x',\bar x_3)$ with $\bar x_3 \in (-T,T)$ be such that $\mu_{\bar x_3}(\bar x')\neq 0$. Then, by continuity of the $X_i$’s, there exist $\delta>0, \eps>0$ such that every curve $\Gamma_i$ with $\Gamma_i\cap\lt( \B_\eps'(\bar x')\times [\bar x_3-\delta,\bar x_3+\delta]\rt) \neq \emptyset$ satisfies $X_i(\bar x_3)=\bar x',$ and such that $\mu\restr (\B'_{2\eps}(\bar x')\backslash \B_\eps'(\bar x'))\times [\bar x_3-\delta,\bar x_3+\delta]=0$ (and thus also $m\restr (\B'_{2\eps}(\bar x')\backslash \B_\eps'(\bar x'))\times [\bar x_3-\delta,\bar x_3+\delta]=0$ since $m\ll {\mu}$). Consider then $\psi_1\in C^\infty_c(\B'_{2\eps}(\bar x'))$ with $\psi_1=1$ in $\B_\eps(\bar x')$ and $\psi_2\in C^\infty_c( \bar x_3-\delta,\bar x_3+\delta)$ and test with $\psi:=\psi_1 \psi_2$ to obtain $$\begin{aligned}
\int_{\bar x_3-\delta}^{\bar x_3+\delta} \frac{d\psi_2}{dx_3} (x_3)\lt(\sum_{X_i(x_3)\in \B_\eps'(\bar x')} {\varphi}_i\rt) dx_3&=\int_{\bar x_3-\delta}^{\bar x_3+\delta}\int_{\B'_\eps(\bar x')} \frac{d\psi_2}{dx_3}(x_3)\psi_1(x') d\mu\\
&=\int_{Q_{1,1}} \frac{\partial \psi}{\partial x_3} d\mu=-\int_{Q_{1,1}} \nabla'\psi \cdot dm\\
&=-\int_{\bar x_3-\delta}^{\bar x_3+\delta}\int_{\B_{2\eps}(\bar x')\backslash \B_\eps(x')} \psi_2 \nabla' \psi_1 \cdot dm =0,\end{aligned}$$ from which the first two assertions follow. It can be easily checked that this implies that $\bar m:= \sum_i \frac{\varphi_i}{\sqrt{1+|{\dot{X}}_i|^2}}
{\dot{X}}_i \, \mathcal{H}^1\restr\Gamma_i$ satisfies . Let $m$ be any other measure satisfying and let us prove that $\nu:=\bar m -m=0$. Since $\Divp \nu_{x_3}=0$, we have for every $\psi\in C^{\infty}(Q_1)$ $$\sum_i \nabla' \psi(X_i(x_3))\cdot \nu_i(x_3)=0,$$ where $\nu_{x_3}=\sum_i \nu_i (x_3) \delta_{X_i(x_3)}$, from which the claim follows.
We remark that Corollary \[coroXia\] below will imply that representation holds for every measure $\mu$ with $I(\mu)<+\infty$. The previous results lead to the following.
\[reg\] A minimizer of the Dirichlet problem (\[limitProb\]) with boundary conditions $\bar{\mu}_+= \sum_{i =1}^N {\varphi}_i^+\delta_{X_i^+}$ and $\bar{\mu}_{-} =
\sum_{i=1}^N {\varphi}_i^- \delta_{X_i^-} $ (some ${\varphi}_i$ may be zero) satisfies
1. $\mu=
\sum_{i=1}^M \frac{\varphi_i}{\sqrt{1+|{\dot{X}}_i|^2}} \, \mathcal{H}^1\restr\Gamma_i $ for some $M\in\N$, where $\Gamma_i=\{ (X_i(x_3),x_3) : x_3\in [a_i,b_i]\}$ are disjoint up to the endpoints, and the $X_i$ are absolutely continuous.
2. Each $X_i$ is affine.
3. If $\bar{\mu}_{-}= \bar{\mu}_{+} $ then there exists a symmetric minimizer with respect to the $x_3=0$ plane.
Let $\mu^{ij}$ be the subsystem emanating from $(X_i^-,-T)$ of the subsystem emanating from $(X_j^+,T)$ of $\mu$, so that $\mu=\sum_{ij}\mu^{ij}$ by $\mu^{ij}_{x_3}\le \mu_{x_3}$ and conservation of mass. By Lemma \[noloop\] we have $\mu^{ij}_{x_3}={\varphi}_{ij}(x_3)\delta_{X^{ij}}(x_3)$ for all $x_3$, otherwise there would be loops. By Lemma \[lemmacurves\], ${\varphi}_{ij}(x_3)$ does not depend on $x_3$. By (\[HolderW2\]), if ${\varphi}_{ij}>0$ then $X^{ij}$ is absolutely continuous. After a relabeling, (i) is proven.
Assertion (ii) follows from minimizing $I({\mu})$ as given by with respect to ${\dot{X}}_i$.
Let now $\bar{\mu}_{-}= \bar{\mu}_{+}$. If $I(\mu,(-T,0))\le I(\mu,(0,T))$ we obtain a symmetric minimizer $\hat\mu$ by reflection of $\mu\LL(-T,0)$ across $\{x_3=0\}$, and analogously in the other case. This proves (iii).
We now show that for symmetric minimizers, at arbitrarily small distance from the boundary we have a finite number of Diracs. We already know that at arbitrarily small distance we have a countable number, and then that we have a representation of ${\mu}$ of the form . Let us point out that we will not use this proposition but rather include it for its own interest.
\[finitebranch\] Fix $\bar\mu\in {\mathcal{M}}^+(Q_L)$. Let $\mu$ be a symmetric minimizer of $I$ subject to $\mu_{\pm T}=\bar\mu$. Then for any $\delta>0$ sufficiently small, the number of Diracs in each slice $x_3 \in [-T +
\delta T , T- \delta T]$ is ${\stackrel{<}{\sim}}\delta^{-4}$.
We may assume $L=T=1$, $\mu(Q_{1,1})=2$. By symmetry, we need only to consider the interval $[0,1-\delta]$. If $\mu_{1-\delta}=\sum_{i} \p_i \delta_{X_i}$, it suffices to prove that ${\varphi}_i{\stackrel{>}{\sim}}\delta^{4}$ for every $i$. For the rest of the proof we fix a point $X_i$ and in order to ease notation we write ${\varphi}:={\varphi}_i$ and $X:= X_i$. Let ${\widetilde}{{\mu}}$ be the subsystem emanating from $(X,1-\delta)$. Thanks to the symmetry of ${\mu}$ and to the no-loop condition, $\mu$ and $\mu-{\widetilde}{{\mu}}$ are disjoint for $x_3> 1-\delta$. Indeed, if this was not the case, by symmetry they would meet also for $x_3<- 1+\delta$, and there would be a loop, which is excluded by Lemma \[noloop\]. Therefore $$\label{secondestimI}
I({\mu})- I({\mu}- {\widetilde}{{\mu}})\ge \int_{1-\delta}^1 \sum_{x'\in Q_1} \left( {\widetilde}{{\mu}}_{x_3}(x')\right)^\hal \, dx_3\ge \delta \sqrt{\p},$$ where in the second step we used subadditivity of the square root.
Let now ${\widetilde}{\mu}_{1}$ be the trace of ${\widetilde}{{\mu}}$ on $x_3={1}$ and for $z:={\varphi}^{1/4}$, let $\hat\mu$ be the symmetric comparison measure constructed as follows:
- in $[0, 1-z]$, $\hat\mu_{x_3}:= (1+\frac{\p}{1-\p})(\mu_{x_3}-{\widetilde}{\mu}_{x_3})$ and
- in $[1-z,1]$, $\hat\mu_{x_3}:=\mu_{x_3}-{\widetilde}{\mu}_{x_3}+\nu_{x_3}$ where $\nu$ is a measure connecting $\frac{\p}{1-\p} (\mu-{\widetilde}{{\mu}})_{1-z}$ to ${\widetilde}{{\mu}}_{1}$ constructed in Proposition \[branchingmu\], so that (recall ) $$ I^{(1-z,1)}(\nu)\les z\sqrt{\p}+\frac{\p}{z}\sim \p^{3/4}.$$
Since ${\mu}$ is a minimizer it follows by subadditivity of the energy that, for some universal (but generic) constant $C$, $$\begin{aligned}
I(\mu)\le I(\hat\mu)&\le \lt(1+\frac{{\varphi}}{1-\p}\rt) I(\mu-{\widetilde}{{\mu}})+ C\p^{3/4}\\
&\le I(\mu-{\widetilde}{{\mu}})+C\p^{3/4}.
\end{aligned}$$ Indeed, $I(\mu-{\widetilde}{{\mu}}) \le I(\mu) {\stackrel{<}{\sim}}1$ while $\varphi \ll 1$ without loss of generality. Recalling , we deduce that $C\p^{3/4}\ge \delta \p^{1/2}$, which yields the result.
\[defpolygonal\] We say that a measure $\mu$ is polygonal if $$\mu= \sum_i \frac{\varphi_i}{\sqrt{1+|{\dot{X}}_i|^2}} \, \mathcal{H}^1\restr\Gamma_i,$$ where the sum is countable, $\Gamma_i$ are segments of the form $\Gamma_i=\{ (X_i(x_3),x_3) : x_3\in [a_i,b_i]\}$ disjoint up to the endpoints, and for any $z\in (0,T)$ only finitely many segments intersect $Q_L\times (-z,z)$. We say it is finite polygonal if the total number of segments is finite.
For any polygonal measure, the representation formula holds.
Let $\bar {\mu}\in {\mathcal{M}}^+(Q_L)$. Then, every symmetric minimizer of $I$ with boundary data $\mu_{\pm T}=\bar {\mu}$ is polygonal.
It suffices to show that for any $z\in (0,T)$ the measure $\mu$ is polygonal in $Q_L\times (-z,z)$. By Proposition \[finitebranch\] the measures $\mu_{z}$ and $\mu_{-z}$ are finite sums of Diracs. Since $\mu$ is a minimizer, it minimizes $I$ restricted to $(-z,z)$ with boundary data $\mu_{\pm z}$. By Proposition \[reg\] we conclude.
Density of regular and quantized measures
-----------------------------------------
In this section we want to prove that when $\bar \mu_{\pm T }= \Phi L^{-2} dx'$, the set of “regular" measures is dense in energy.
\[defregular\] We denote by $\Mreg(Q_{L,T}){\subseteq}{\mathcal{A}}^*_{L,T}$ the set of regular measures, i.e., of measures $\mu$ such that:
- The measure $\mu$ is finite polygonal, according to Def. \[defpolygonal\].
- All branching points are triple points. This means that any $x\in Q_{L,T}$ belongs to the closures of no more than three segments.
For $N\in \N$, we say that $\mu$ is $N$-regular, $\mu\in \MNreg(Q_{L,T}){\subseteq}\Mreg(Q_{L,T})$, if in addition
- The traces obey $\mu_{T}=\mu_{-T}=\Phi N^{-2}\sum_{j} \delta_{X_j}$, where the $X_j$ are $N^2$ points on a square grid, spaced by $L/N$, and $\Phi\ge0$.
We can now state the main theorem of this section:
\[theodens\] For every measure $\mu\in {\mathcal{A}}^*_{L,T}$ with $I(\mu)<+\infty$ and $ \mu_{\pm T}= \Phi L^{-2} dx'$, there exists a sequence of measures $\mu_N
\in\MNreg$, with $\mu_N \weaklim \mu$ and such that $\limsup_{N \to +\infty} I(\mu_N)\le I(\mu)$, $\mu_N(Q_{L,T})=\mu(Q_{L,T})$.
The proof will be based on the following intermediate result.
\[lemmamudensityinside\] For every measure $\mu\in {\mathcal{A}}^*_{L,T}$ with $I(\mu)<+\infty$, and such that the traces $\mu_T$ and $\mu_{-T}$ are finite sums of Diracs, there exists a sequence of regular measures $\mu^{(N)}\in \Mreg(Q_{L,T})$ with $\mu^{(N)} \weaklim \mu$, $\mu^{(N)}_{\pm T}=\mu_{\pm T}$, and such that $\limsup_{N \to +\infty} I(\mu^{(N)})\le I(\mu)$.
We shall modify $\mu$ in two steps to make it polygonal: first on finitely many layers, to have finitely many Diracs on each of them, and then in the rest of the volume, using local minimization.
Fix $N\in\N$ and $\delta>0$, both chosen later. We choose levels $z_j\in (Tj/N, T(j+1)/N)$, for $j=-N+1,\dots, N-1$, with the property that $\mu_{z_j}=\sum_{k\in\N} \varphi_{j,k}\delta_{x'_{j,k}}$, with $\sum_{k\in \N} \varphi_{j,k}^{1/2}<+\infty$ for every $j$. We shall iteratively truncate the measure at these levels so that it is supported on finitely many points; for notational simplicity we also define $z_{\pm N}:=\pm T$. The measures $\mu^j$, $j=-N,\dots, N$ will all satisfy $\mu^j\ll\mu$ and $I(\mu^j)\le I(\mu)$.
We start with $\mu^{-N}:=\mu$. In order to construct $\mu^{j+1}$ from $\mu^j$, we first choose $K_j$ such that $\sum_{k\ge K_j} \varphi_{j,k}^{1/2}\le \delta/N$. Then we define $\mu^{j+1}$ as the sum of the subsystems of $\mu^j$ originating from the points $(x'_{j,k},z_j)$ with $k<K_j$. Clearly $I(\mu^{j+1})\le I(\mu^j)\le I(\mu)$. At the same time, since $\sum_{k=K_j}^{+\infty} {\varphi}_{j,k} \le \Phi^{1/2} \sum_{k=K_j}^{+\infty} {\varphi}_{j,k}^{1/2}$, $$\mu^{j+1}_{z_j}(Q_L)
= \mu^{j}_{z_j}(Q_L)- \sum_{k=K_j}^{+\infty} {\varphi}_{j,k}\ge \mu^{j}_{z_j}(Q_L)-\frac{ \Phi^{1/2} \delta}{N},$$ so that $|\mu^{j+1}-\mu^j|(Q_{L,T})\le 2\Phi^{1/2}\delta T/N$. Therefore $|\mu^N-\mu|(Q_{L,T})\les \delta T \Phi^{1/2}$.
We define $\hat\mu^N$ as the minimizer with boundary data $\mu^N_{z_j}$ and $\mu^N_{z_{j+1}}$ in each stripe $Q_L\times (z_j, z_{j+1})$. Then $I(\hat\mu^N)\le I(\mu^N)\le I(\mu)$, and $\hat\mu^N_{\pm T}=\mu^N_{\pm T}$.
At this point we fix the boundary data. For this, we let ${\widetilde}{\mu}$ be the minimizer on $Q_{L,T}$ with boundary data ${\widetilde}{\mu}_\pm:=\mu_{\pm T}-\mu^N_{\pm T}$. These boundary data are finite sums of Diracs, and their flux is $|{\widetilde}{\mu}_\pm|(Q_L)=|\mu_T-\mu^N_T|(Q_L)\les \delta \Phi^{1/2}$. By Proposition \[reg\] the minimizer is finite polygonal, by Proposition \[branchingmu\] it has energy no larger than a constant times $\delta \Phi^{1/2}T^{-1} + \delta^{1/2} T \Phi^{1/4} L^{-2}$. Finally we set $\mu^{(N)}:=\hat\mu^N+{\widetilde}{\mu}$. Then $\mu^{(N)}_{\pm T}=\mu_{\pm T}$, and $$I(\mu^{(N)})\le I(\hat\mu^N)+I({\widetilde}{\mu}) \le I(\mu) + C\left(\delta \Phi^{1/2}T^{-1} + \delta^{1/2} T \Phi^{1/4} L^{-2}\rt) .$$
Up to a small perturbation, we may further assume that all junctions are triple. We can now choose for instance $\delta=1/N$. It only remains to show that $\mu^{(N)}\weaklim\mu$ as $N\to+\infty$. Recalling that $|\mu^N-\mu|(Q_{L,T})+|{\widetilde}{\mu}|(Q_{L,T})\les \delta \Phi^{1/2}T$, we only need to show that $\hat\mu^N-\mu^N\weaklim 0$.
For $x_3\in (z_j,z_{j+1})$ we have by definition of $\hat \mu^N$ $$W_2(\mu_{x_3}^N, \hat \mu_{x_3}^N) \le W_2(\mu_{x_3}^N, \mu_{z_j}^N) +W_2(\hat \mu_{x_3}^N, \hat\mu_{z_j}^N),$$ and by $$W_2^2(\mu_{x_3}^N, \mu_{z_j}^N)+W_2^2(\hat \mu_{x_3}^N, \hat \mu_{z_j}^N) \le L^2 (z_{j+1}-z_j) \lt( I(\mu^N)+I(\hat \mu^N)\rt),$$ so that $$\max_{x_3} W_2^2(\mu_{x_3}^N, \hat \mu_{x_3}^N) {\stackrel{<}{\sim}}\frac{L^2T}{N}I(\mu).$$ For $x_3\in (-T,T)$, let $\Pi_{x_3}$ be an optimal transport plan from $\mu^N_{x_3}$ to $\hat{\mu}^N_{x_3}$. Considering then the transport plan $\Pi:=\Pi_{x_3}\otimes dx_3$ between $\mu^N$ and $\hat{\mu}^N$, we get $$W_2^2(\mu^N,\hat{\mu}^N)\le \int_{-T}^T W^2_2(\mu_{x_3}^N,\hat\mu_{x_3}^N) dx_3\les \frac{L^2T^2}{N}I(\mu)
\,,$$ which yields that indeed $\hat\mu^N-\mu^N\weaklim 0$.
Let $\eps\in (0,1/4)$, chosen such that it tends to zero as $N\to\infty$. We define $\hat\mu$ in $Q_{L,(1-2\eps)T}$ as a rescaling by $(1-2\eps)$ in the vertical direction, $\hat\mu_{(1-2\eps)x_3}=\mu_{x_3}$. An easy computation shows that $I(\hat\mu, -(1-2\eps)T, (1-2\eps)T)\le \frac{1}{1-2\eps} I(\mu)$. In particular, $\hat\mu_{(1-2\eps)T}=\Phi L^{-2} dx'$. For $x_3\in ((1-2\eps)T, T)$ we define $\hat\mu$ as the result of Proposition \[branchingmu\]. Then we set ${\widetilde}{\mu}=\hat\mu$ on $(-(1-\eps)T,(1-\eps)T)$ and extend it constant outside, in the sense that ${\widetilde}{\mu}_{x_3}=\hat\mu_{(1-\eps)T}$ for $x_3\in ((1-\eps)T,T)$, and the same on the other side. Since $\hat \mu_{(1-\eps)T}$ is the midplane configuration of the branching measure constructed in Proposition \[branchingmu\], ${\widetilde}{\mu}_{x_3}$ is a finite sum of Diracs for $|x_3|\ge (1-\eps) T$. We obtain $$I({\widetilde}{\mu}) \le \frac{1}{1-2\eps} I(\mu) + C \eps^{1/3} \left(\Phi^{2/3}T^{1/3} L^{-4/3}+\eps^{2/3} \Phi^{1/2} T L^{-2} \right)\,.$$ By Lemma \[lemmamudensityinside\] applied to the inner domain $(-(1-\eps)T,(1-\eps)T)$ there is a finite polygonal measure $\check\mu$ which is close to ${\widetilde}{\mu}$ and has the same boundary data at $x_3=\pm T(1-\eps)$. The measure given by $\check{\mu}$ inside, and ${\widetilde}{\mu}$ outside, has the required properties.
We now turn to the quantization of the measures.
We say that a regular measure $\mu\in\Mreg(Q_{L,T}){\subseteq}{\mathcal{A}}_{L,T}^*$ is $k$-quantized, for $k>0$, if for all $(x',z)\in Q_{L,T}$ one has $k\mu_z(\{x'\})\in 2\pi\N$.
\[lemmaquantize\] Let $\mu\in \MNreg(Q_{L,T})$ and $\Phi=\mu_T(Q_L)$. For any $k>0$ such that $k\Phi\in 2\pi\N$ there is a $k$-quantized regular measure $\mu^{k}\in \Mreg$ such that $\mu^{k}(Q_{L,T})=\mu(Q_{L,T})$ and $$\left(1-\frac{ C(\mu)}k\right) \mu \le \mu^k \le \left(1+\frac{ C(\mu)}k\right)\mu\,.$$ This implies in particular $\mu^k\ll\mu$, $\mu^k\to\mu$ strongly, $W_2^2(\mu,\mu_k)\les C(\mu)k^{-1}$ and $I(\mu^k)\to I(\mu)$ as $k\to\infty$.
The measure $\mu$ consists of finitely many segments, each with a flux. To prove the assertion it suffices to round up or down the fluxes to integer multiples of $2\pi/k$ without breaking the divergence condition, and without changing the total flux.
Since $\mu\in\MNreg(Q_{L,T})$, we have $\mu_{T}=\Phi N^{-2}\sum_i \delta_{X_i}$. We select $\varphi_i^k$ as $2\pi \lfloor k \Phi /(2\pi N^2)\rfloor /k$ or $2\pi \lfloor k \Phi /(2\pi N^2)+1\rfloor /k$ , depending on $i$. Precisely, we choose the first value for $i=0$ and then, at each $i$, we choose the lower one if $\sum_{j<i} ({\varphi}_j^k- \Phi N^{-2})>0$, and the upper one otherwise. This concludes the definition of $\mu^k_T$.
The fluxes in the interior of the sample are defined by propagating the rounding. At each point where a bifurcation occurs, if there is more then one outgoing branch we distribute the rounding as discussed for $\mu_T^k$. This increases the maximal error by at most $2\pi/k$, at each branching point. Since $\mu$ is finite polygonal, there is a finite number of branching points, hence the total error is bounded by a constant times $1/k$. Precisely $|{\varphi}_i^k-{\varphi}_i|\le C(\mu)/k$ for any segment $i$. Since ${\varphi}_i$ only takes finitely many values, $|{\varphi}_i^k-{\varphi}_i|\le {\varphi}_i C(\mu)/k$ for any segment $i$, which concludes the proof.
Relation with irrigation problems {#Sec:reliri}
---------------------------------
The functional $I(\mu)$ bears similarities with the so-called irrigation problems which have attracted a lot of interest (see for instance [@Xia; @BeCaMo]). Besides their applications to the modeling of communication networks and other branched patterns (see again [@BeCaMo] and the references therein), they have also been recently used in the study of Sobolev spaces between manifolds [@bethuel]. Let us recall their definition and for this, follow the notation of [@BeCaMo]. For $E(G)$ a set of oriented straight edges and $\p: E(G)\to (0,+\infty)$ we define the irrigation graph $G$ as the vector measure $$G:=\sum_{e\in E(G)} {\varphi}(e) \mathbf{e} \, \H^1\restr e \,$$ where $\mathbf{e}$ is the unit tangent vector to $e$. For $\a\in [0,1]$, we then define the Gilbert energy of $G$ by $$M^{\a}(G):=\sum_{e\in E(G)} \p(e)^\a \H^1(e).$$ Given two atomic probability measures $\mu^+=\sum_{i=1}^k a_i \delta_{X_i}$ and $\mu^-=\sum_{j=1}^l b_j \delta_{Y_j}$, we say that $G$ irrigates $(\mu^+,\mu^-)$ if $\div G= \mu^+-\mu^-$ in the sense of distributions (this implies in particular that $G$ satisfies Kirchoff’s law). If we are now given any two probability measures $(\mu^+,\mu^-)$ and a vector measure $G$, with $\div G=\mu^+-\mu^-$ (sometimes called an irrigation path between $\mu^+$ and $\mu^-$), we define $$M^{\a}(G):=\inf \{ \liminf_{i\to +\infty} M^{\a}(G_i)\},$$ where the infimum is taken among all the sequences of irrigation graphs $G_i$ with $G_i\weaklim G$ in the sense of measures and such that $\div G= \mu^+_i-\mu^-_i$ for some atomic measures $\mu^{\pm}_i$ tending to $\mu^{\pm}$. If no such sequence exists then we set $M^\a(G)=+\infty$. The irrigation problem then consists in minimizing $M^{\a}(G)$ among all the transport paths $G$ between $\mu^+$ and $\mu^-$. For $\a=0$ this is a generalization of the famous Steiner problem while for $\a=1$ it is just the Monge-Kantorovich problem.
Using some powerful rectifiability criterion of B. White, the following theorem was proven by Q. Xia [@Xia].
\[theoXia\] Given $0<\a<1$, any transport path $G$ with $M^{\a}(G)+M^1(G)<+\infty$ is rectifiable in the sense that $$G={\varphi}\tau \, \H^1\restr\Gamma$$ for some density function $\p$ and some $1-$rectifiable set $\Gamma$ having $\tau$ as tangent vector.
For minimal irrigation paths, much more is known about their interior and boundary regularity [@BeCaMo]. For instance, as for our functional $I(\mu)$ (see Proposition \[reg\]), it can also be proven that minimal irrigation paths contain no loops and that for $\a> 1-\frac{1}{n}$ (where $n$ is the dimension of the ambient space i.e. $n=3$ for us), any two probability measures $\mu^{\pm}$ can be irrigated at a finite cost (compare with Proposition \[existmu\]).
Using Theorem \[theoXia\] and Lemma \[lemmamudensityinside\], we can obtain the following rectifiability result.
\[coroXia\] Every measure $\mu$ for which $I(\mu)<+\infty$ is rectifiable.
Using the construction of Lemma \[lemmamudensityinside\], we can find a sequence $\mu^n$ such that $\mu^n\weaklim \mu$, $\limsup_{n\to +\infty} I(\mu^n)\le I(\mu)$ and $\mu^n=\sum_{i=1}^N \frac{{\varphi}_i}{\sqrt{1+|{\dot{X}}_i|^2}} \H^1\restr \Gamma_i$ for some straight edges $\Gamma_i=\{(x_3,X_i(x_3)) \, :\, x_3\in(a_i,b_i)\}$. Letting $\widetilde{\mu}^n:=\sum_{i=1}^N \frac{{\varphi}_i}{\sqrt{1+|{\dot{X}}_i|^2}} \begin{pmatrix} {\dot{X}}_i\\ 1\end{pmatrix} \H^1\restr \Gamma_i$, we have for $\a\ge\frac{3}{4}$, $$\begin{aligned}
M^\a({\widetilde}\mu^n)&=\sum_i \int_{a_i}^{b_i} {\varphi}_i^\alpha \sqrt{1+ |{\dot{X}}_i|^2} dx_3\\
&\les \sum_i \int_{a_i}^{b_i} {\varphi}_i (1+|{\dot{X}}_i|^2) + {\varphi}_i^{2\alpha-1} dx_3\\
&\les \sum_i \int_{a_i}^{b_i} {\varphi}_i (1+|{\dot{X}}_i|^2) + \sqrt{{\varphi}_i} dx_3\les I(\mu^n) +1\end{aligned}$$ so that $\liminf_{n\to +\infty} M^{\a}(\widetilde{\mu}^n)\les I(\mu)+1<+\infty$ and by Theorem \[theoXia\], the claim follows.
In [@OudSant], an approximation of the functional $M^\a$ in the spirit of the Modica-Mortola [@ModMort] approximation of the perimeter was proposed. Even though their proofs and constructions are completely different from ours, this approach bears some similarities with our derivation of the functional $I({\mu})$ from the Ginzburg-Landau functional $E_T(u,A)$.
Lower bound {#Sec:gammaliminf}
===========
In the rest of the paper we consider sequences with $$\label{eqlbassumptcoeff}
T_n\to +\infty, \hskip5mm \alpha_n\to +\infty,\hskip5mm \beta_n\to 0, \hskip5mm \frac{T_n}{\alpha_n} \to +\infty,\hskip5mm \alpha_n \beta_n^{7/2}\to +\infty\,.$$ No constant appearing in the sequel will depend on the specific choice of the sequence. We observe that (\[eqlbassumptcoeff\]) immediately implies $\alpha_n\beta_n^2\to+\infty$ and $\alpha_n^2\beta_n\to+\infty$. Let us recall that in this proof we set ${\widetilde}L=1$ and that (see (\[eqdefwe\])) $$\begin{gathered}
\wE(u,A)=\int_{Q_{1,1}} \a^{-2/3}\b^{-1/3}\lt|\nabla_{\a^{1/3}\b^{-1/3}TA}'u\rt|^2+
\a^{-4/3}\b^{-2/3}\lt|(\nabla_{\a^{1/3}\b^{-1/3}TA} u)_3\rt|^2\\
+ \a^{2/3}\b^{-2/3}\left({B_3} -(1-|{u}|^2)\right)^2
+ \b^{-1}|{B'}|^2 dx +\a^{1/3}\b^{7/6}\|\b^{-1}{B_3}-1\|_{H^{-1/2}(x_3=\pm1)}^2.\end{gathered}$$ In this section, we prove the following compactness and lower bound result.
\[gammaliminf\] Fix sequences of positive numbers $\alpha_n$, $\beta_n$, $T_n$ such that (\[eqlbassumptcoeff\]) holds, and let $(u_n, A_n)$ be such that $\sup_{n} \wE(u_n,A_n)<+\infty$. Then up to a subsequence, the following holds :
1. $\b_n^{-1}(1-\rho_n)\weaklim \mu$ for some measure $\mu$, $\b_n^{-1} B'_n\weaklim m$ for some vector-valued measure $m\ll {\mu}$ satisfying the continuity equation .
2. For almost every $x_3\in (-1,1)$, there exists some probability measure $\mu_{x_3}$ on $Q_1$ with $\mu=\mu_{x_3}\otimes dx_3$ and such that $\mu_{x_3}\weaklim dx'$ as $x_3\to \pm 1$.
3. For almost every $x_3\in (-1,1)$, ${\mu}_{x_3}=\sum_{i\in {\mathcal{I}}} \p_i \delta_{X_i}$ with ${\mathcal{I}}$ at most countable and $\p_i>0$.
4. One has $(\mu,m)\in {\mathcal{A}}_{1,1}$ with $$\liminf_{n\to +\infty} \wE(u_n,A_n)\ge I({\mu},m).$$
Let us first show that the energy gives a quantitative control on the failure of the Meissner condition $\rho B=0$ in a weak sense.
\[lemmameissner\] For every $Q_1$-periodic test function $\psi\in H^1_\mathrm{per}(Q_{1,1})$, if $\|\rho\|_\infty\le 1$ then $$\label{estimmeissner}
\lt|\int_{Q_{1,1}} \rho B_3 \psi dx\rt|\les
\frac{\a^{1/3}\b^{2/3}}{T}\wE(u,A)\|\psi\|_{L^\infty}+
\frac{\beta^{1/2}}{T}\wE(u,A)^\hal \|{\nabla}'
\psi\|_{L^2},$$ and, if additionally $\psi(x',\pm1)=0$, for $k=1,2$ and $\alpha^2\beta\ge 1$, $$\label{estimmeissnerprime}
\lt|\int_{Q_{1,1}} \rho B_k \psi dx\rt|\les
\frac{\a^{2/3}\b^{5/6}}{T}\wE(u,A)\|\psi\|_{L^\infty}+
\frac{\alpha^{1/3}\beta^{2/3}}{T}\wE(u,A)^\hal \|{\nabla}\psi\|_{L^2}.$$ Moreover, if $\xi\in H^1_0(-1,1)$ and $\psi$ is a periodic Lipschitz continuous function on $Q_1$ then $$\label{estimmeissnerter}
\lt|\int_{Q_{1,1}} \rho B'\cdot \nabla' \psi \, \xi dx\rt|\les
\frac{\a^{2/3}\b^{5/6}}{T}\wE(u,A)\|\xi \nabla'\psi\|_{L^\infty}+
\frac{\b^{1/2}}{T}\wE(u,A)^\hal \| \partial_3 \xi \nabla'\psi \|_{L^2}.$$
Let $\lambda:=\a^{1/3}\b^{-1/3}T$. For (\[estimmeissner\]) we use formula with $A$ substituted by $\lambda A$, that is $|{\nabla}'_{\lambda A} u|^2 = |{\mathcal{D}}_{\lambda A}^3 u|^2 +\rho \lambda B_3+
{\nabla}'\times j'_{\lambda A}$. We integrate against a test function $\psi$, $$\begin{aligned}
\lt|\int_{Q_{1,1}} \rho B_3 \psi dx\rt| & = & \frac{\a^{1/3}\b^{2/3}}{T}\lt|\int_{Q_{1,1}}\a^{-2/3}\b^{-1/3} \left(|{\nabla}_{\lambda A}'u|^2 - |{\mathcal{D}}_{\lambda A}^3 u|^2 - {\nabla}'\times j'_{\lambda A}\right) \psi dx\rt|\\
& \les & \frac{\a^{1/3}\b^{2/3}}{T}\lt(\wE(u,A)\|\psi\|_{L^\infty} + \a^{-1/3}\b^{-1/6}\int_{Q_{1,1}} \a^{-1/3}\b^{-1/6}|j'_{\lambda A}| |{\nabla}' \psi|dx\rt)\\
& \les & \frac{\a^{1/3}\b^{2/3}}{T}\lt(\wE(u,A)\|\psi\|_{L^\infty}+ \a^{-1/3}\b^{-1/6}\wE(u,A)^\hal \|{\nabla}'
\psi\|_{L^2} \rt),\end{aligned}$$ [where we have used that $|j'_{\lambda A}|\le |\nabla'_{\lambda A} u| $ in view of the definition and the upper bound $\rho\le 1$.]{} We obtain similarly : One first checks from the definition of ${\mathcal{D}}_{\lambda A}u$ that $$\left| |{\mathcal{D}}^1_{\lambda A} u|^2- |(\nabla_{\lambda A} u)_3|^2 - |(\nabla_{\lambda A} u)_2|^2\right|
\le 2 |(\nabla_{\lambda A} u)_2| \, |(\nabla_{\lambda A} u)_3|.$$ Testing (\[magic2\]) with $\psi$ and integrating by parts the term with $j_{\lambda A}$ as above gives $$\lambda \lt|\int_{Q_{1,1}} \rho B_k \psi dx\rt|\les
\int_{Q_{1,1}} 2 |\psi| \, |(\nabla_{\lambda A} u)_2| \, |(\nabla_{\lambda A} u)_3|
+ |\nabla \psi| \, |\nabla_{\lambda A} u| dx\,.$$ Estimating $ 2|(\nabla_{\lambda A} u)_2| \, |(\nabla_{\lambda A} u)_3| \le \alpha^{1/3}\beta^{1/6}|(\nabla_{\lambda A} u)_2|^2 + \alpha^{-1/3}\beta^{-1/6}|(\nabla_{\lambda A} u)_3|^2 $ and $\|\nabla_{\lambda A} u\|_2^2\le \alpha^{4/3}\beta^{2/3} \wE(u,A)$ concludes the proof of (\[estimmeissnerprime\]).\
The proof of is very similar to the proof of . Arguing as above with $\xi\partial_{k}\psi$ playing the role of $\psi$, we get $$\lambda \lt|\int_{Q_{1,1}} \rho B'\cdot \nabla'\psi \xi dx\rt|\les
\int_{Q_{1,1}} 2 |\xi \nabla' \psi | \, |(\nabla_{\lambda A} u)_2| \, |(\nabla_{\lambda A} u)_3| dx
+ \lt|\int_{Q_{1,1}} (\nabla \times j_{\lambda A})\cdot \nabla \psi \xi dx\rt|\,.$$ The first term is estimated exactly as before, while the second one gives after integration by parts of $\nabla \times $ and using $\nabla \times \nabla=0$, $$\lt|\int_{Q_{1,1}} (\nabla \times j_{\lambda A})\cdot \nabla \psi \xi dx\rt|= \lt|\int_{Q_{1,1}} \nabla \xi\cdot (j_{\lambda A}\times \nabla \psi) dx\rt|\le \|\partial_3 \xi \nabla' \psi\|_{L^2} \lt(\int_{Q_{1,1}} |j'_{\lambda A}|^2\rt)^{1/2},$$ from which we conclude the proof.
We now prove that for admissible pairs $(u,A)$ of bounded energy, the corresponding curves $x_3\mapsto \beta^{-1}B_3(\cdot,x_3)$ satisfy a sort of uniform Hölder continuity. This is the analog of for the limiting energy.
\[HolderE\] For every admissible pair $(u,A)$ with $\|\rho\|_\infty\le 1$, and every $x_3,{\widetilde}{x}_3\in(-1,1)$, letting $E:=\wE(u,A)$ it holds (recall ) $$\begin{gathered}
\label{HolderE2}
\|\b^{-1}B_3(\cdot,x_3) - \b^{-1}B_3(\cdot,{\widetilde}{x}_3) \|_{BL}\les E^{1/2} |x_3-{\widetilde}{x}_3|^{1/2}
+ \sigma(\alpha,\beta,T)(E^{1/2}+E),
\end{gathered}$$ where $\sigma(\alpha,\beta,T):= \lt(\frac{\a}{T}\rt)^{1/2} \lt(\a^2\b^{5/2}\rt)^{-1/6}+ (\a^{1/2}\b)^{-1/3} + \lt(\frac{\alpha^{1/3}}{T\b^{5/6}} \rt)^{1/2}$, which goes to zero in the regime . In particular, in that regime, if $E\les 1$, for every $x_3,{\widetilde}{x}_3\in(-1,1)$ with $|x_3-{\widetilde}{x}_3|\ge \sigma^{1/2}(\alpha,\b,T)$, there holds $$\label{equiconB}
\|\beta^{-1}B_3(\cdot,x_3)-\beta^{-1}B_3(\cdot,{\widetilde}{x}_3)\|_{BL}\les |x_3-{\widetilde}{x}_3|^{1/2}.$$
The proof resembles that of [@CoOtSer Lem. 3.13]. First, we show that for every $Q_1-$periodic and Lipschitz continuous function $\psi$ with $\|\psi\|_{Lip}\le 1$, $$\label{firstestimHolder}
\lt|\int_{Q_1\times\{x_3\} } \b^{-1}B_3 \psi dx' -\int_{Q_1\times\{{\widetilde}{x}_3\} } \b^{-1}B_3 \psi dx'\rt|\le|x_3-{\widetilde}{x}_3|^{1/2} \b^{-1/2} E^{1/2}.$$ This follows from $\Div B=0$ and integration by parts, which yields $$\begin{aligned}
\lt|\int_{Q_1\times\{x_3\} } \b^{-1}B_3 \psi dx' -\int_{Q_1\times\{{\widetilde}{x}_3\} } \b^{-1}B_3 \psi dx'\rt|&=\lt|\int_{Q_1\times(x_3,{\widetilde}{x}_3)} \b^{-1}\partial_3 B_3 \psi dx\rt|\\
&=\lt|\int_{Q_1\times(x_3,{\widetilde}{x}_3)}\beta^{-1} B' \cdot \nabla' \psi dx\rt|\\
&\le |x_3-{\widetilde}{x}_3|^{1/2} \beta^{-1/2} \lt(\int_{Q_1\times(x_3,{\widetilde}{x}_3)} \beta^{-1}|B'|^2 dx\rt)^{1/2}\\
&\le |x_3-{\widetilde}{x}_3|^{1/2} \beta^{-1/2}E^{1/2}.\end{aligned}$$ For $|x_3-{\widetilde}{x}_3|\le T^{-1} \a^{1/3}\b^{1/6}$, this implies that $$\|\b^{-1}B_3(\cdot,x_3) - \b^{-1}B_3(\cdot,{\widetilde}{x}_3) \|_{BL}\le \lt(\a^{1/3} T^{-1} \b^{-5/6}\rt)^{1/2} E^{1/2}\le \sigma(\a,\b,T) E^{1/2},$$ and is proven. Letting $\hat{\sigma}(\a,\b,T):= \lt(\frac{\a}{T}\rt)^{1/2} \lt(\a^2\b^{5/2}\rt)^{-1/6}+ (\a^{1/2}\b)^{-1/3}$, we are left to prove that for $Q_1-$periodic and Lipschitz continuous $\psi$ with $\|\psi\|_{Lip}\le 1$ and $|x_3-{\widetilde}{x}_3|\ge T^{-1}\a^{1/3}\b^{1/6}$, $$\label{toproveHolder}
\lt|\int_{Q_{1}\times\{x_3\}} \beta^{-1}B_3 \psi dx' -\int_{Q_{1}\times\{{\widetilde}{x}_3\}} \beta^{-1}B_3 \psi dx' \rt|\les |x_3-{\widetilde}{x}_3|^{1/2}E^{1/2}+ \hat{\sigma}(\alpha,\beta,T)(E^{1/2}+E).$$ Up to translation we may assume ${\widetilde}{x}_3=0$ and $x_3>0$. Let $\delta\le x_3/2$ and define $\xi: \R\to \R$ by $$\xi(z):=\begin{cases}
\frac{z}{\delta} & \textrm{if } 0<z<\delta\\
1 & \textrm{if } \delta\le z \le x_3-\delta\\
\frac{x_3-z}{\delta} & \textrm{if } x_3-\delta\le z\le x_3\\
0 & \textrm{otherwise}.
\end{cases}$$ We then have using again $\Div B=0$ and integration by parts $$\label{integpart}
\int_{Q_{1,1}} \beta^{-1} B_3 \psi \partial_3\xi dx=-\int_{Q_{1,1}} \b^{-1}\rho B'\cdot \nabla'\psi \, \xi dx -\int_{Q_{1,1}} \b^{-1}(1-\rho) B'\cdot\nabla'\psi\, \xi dx.$$ The first term on the right-hand side of is estimated by . For the second term, we now estimate $$\label{firsttermright}
\int_{Q_{1,1}} \b^{-1}(1-\rho)|B'|\xi dx\le \lt(\int_{Q_1\times(0,x_3)} \b^{-1}(1-\rho) dx\rt)^{1/2} \lt(\int_{Q_{1,1}}\b^{-1}(1-\rho)|B'|^2 dx \rt)^{1/2}.$$ We rewrite the first factor as $$\int_{Q_1\times(0,x_3)} \b^{-1}(1-\rho) dx=\int_{Q_1\times(0,x_3)} \b^{-1}B_3 dx +\int_{Q_1\times(0,x_3)} \b^{-1}(B_3-(1-\rho)) dx,$$ from which we obtain $$\int_{Q_1\times(0,x_3)} \b^{-1}(1-\rho) dx\le \lt|\int_{Q_1\times(0,x_3)} \b^{-1}B_3 dx \rt| +|x_3|^{1/2} \lt(\int_{Q_{1,1}} \b^{-2} (B_3-(1-\rho))^2 dx\rt)^{1/2}.$$ This allows to make use of $$\int_{Q_{1,1}} \b^{-2} (B_3-(1-\rho))^2 dx\le \a^{-2/3}\b^{-4/3} E,$$ and $\int_{Q_1\times\{z\}} \beta^{-1} B_3 dx' =1$, yielding $$\label{firstestimrightbis}
\int_{Q_1\times(0,x_3)} \b^{-1}(1-\rho) dx\le |x_3|+|x_3|^{1/2}\a^{-1/3}\b^{-2/3} E^{1/2}.$$ The second factor in is directly estimated by $$\int_{Q_{1,1}}\b^{-1}(1-\rho)|B'|^2 dx\le E,$$ so that inserting this and into gives $$\label{firstestimrightter}
\lt| \int_{Q_{1,1}} \b^{-1}(1-\rho) B'\cdot\nabla'\psi \xi dx\rt|\les |x_3|^{1/2} E^{1/2} +|x_3|^{1/4} \a^{-1/6}\b^{-1/3}E^{3/4}.$$
Letting $f(z):=\int_{Q_1\times\{z\}}\beta^{-1} B_3 \psi dx'$, we thus obtain from , and , $$\label{firstestimf}
\lt|\int_0^{x_3} f \partial_3 \xi dz\rt|\les \frac{\a^{2/3}\b^{-1/6}}{T} E +\frac{\a^{1/3}\b^{-1/3}}{T} E^{1/2} \|\partial_3\xi\|_{L^2}+ |x_3|^{1/2} E^{1/2} +|x_3|^{1/4} E^{3/4} \a^{-1/6}\b^{-1/3}.$$ Since by definition of $\xi$, $\int_0^{x_3} f \partial_3\xi dz=\frac{1}{\delta} \int_0^\delta f dz-\frac{1}{\delta}\int_{x_3-\delta}^{x_3} f dz$, we have $$f(x_3)-f(0)=\int_0^{x_3} f \partial_3 \xi dz +\frac{1}{\delta}\int_{0}^\delta (f-f(0)) dz+\frac{1}{\delta}\int_{x_3-\delta}^{x_3} (f(x_3)-f) dz,$$ so that $$|f(x_3)-f(0)|\le \lt|\int_0^{x_3} f \partial_3 \xi dz\rt|+\sup_{(0,\delta)}|f-f(0)| +\sup_{(x_3-\delta,x_3)}|f-f(x_3)|.$$ In view of this elementary inequality, the estimates and combine to $$\begin{gathered}
|f(x_3)-f(0)|\les |x_3|^{1/2} E^{1/2} +\frac{\a^{1/3}\b^{-1/3}}{T} E^{1/2} \delta^{-1/2} + \delta^{1/2}\b^{-1/2}E^{1/2} +\frac{\a^{2/3}\b^{-1/6}}{T} E\\ + |x_3|^{1/4} E^{3/4} \a^{-1/6}\b^{-1/3}. \end{gathered}$$ We now optimize in $\delta $ by choosing $\delta=T^{-1}\a^{1/3}\b^{1/6}$, which combined with $\frac{\a^{2/3}\b^{-1/6}}{T}\ll \a^{-1/6}\b^{-1/3}$, yields in the form of $$|f(x_3)-f(0)|\les |x_3|^{1/2} E^{1/2} + (E^{1/2}+E)\lt( \lt(\frac{\a}{T}\rt)^{1/2} \lt(\a^2\b^{5/2}\rt)^{-1/6}+ \a^{-1/6}\b^{-1/3}\rt).$$
\[remarkequi\] We notice that thanks to the Kantorovich-Rubinstein Theorem [@villani Th. 1.14], if $B_3$ is non-negative then we can substitute the Bounded-Lipschitz norm in by a $1-$Wasserstein distance. In particular, it would imply that if $(\alpha_n,\beta_n,T_n)$ satisfy and if $(u_n,A_n)$ are admissible with $\|\rho_n\|_{\infty}\le 1$, $\wE(u_n,A_n)\les 1$, and $B^n_3\ge 0$, then the corresponding curves $x_3\mapsto \beta_n^{-1} B^n_3(\cdot,x_3)$ would be in some sense equi-continuous in the space of probability measures endowed with the Wasserstein metric.
For $\eps>0$ fixed, we define the following regularization of the singular double well potential $\chi_{\rho>0}(1-\rho)^2$ : $$\label{doublewell}
W_\eps(\rho):=\eta_\eps(\rho)(1-\rho)^2 \ \text{with } \ \eta_\eps(\rho):=\min\{\rho/\eps,1\}
\,,$$ see and Figure \[figweps\]. We next show that the energy controls $W_\eps(\rho)$. Similar ideas have been used in the context of Bose-Einstein condensates [@GolMer].
![The cutoff function $\eta_\eps$ and the two-well potential $W_\eps$ used in Lemma \[lemmafirstlowerbound\].[]{data-label="figweps"}](fig-weps-crop){height="4cm"}
\[lemmafirstlowerbound\] For every $\eps>0$ there exists $C_\eps>0$ such that for every $(u,A)$ with $\|\rho\|_\infty\le 1$ it holds, $$\label{estimlower}
\int_{Q_{1,1}} \a^{2/3}\b^{-2/3}W_\eps(\rho) dx
\le \int_{Q_{1,1}} \a^{2/3}\b^{-2/3}\left(B_3 -(1-\rho)\right)^2 dx
+C_\eps \wE(u,A) \frac{\a}{T}.$$
As above, to lighten notation, we let $E:= \wE(u,A)$. Writing $(1-\rho)= B_3 -(B_3-(1-\rho))$, we obtain by Young’s inequality $$\begin{aligned}
(1-\rho)^2&= B_3 (1-\rho)-(B_3-(1-\rho)) (1-\rho)\\
&\le B_3 (1-\rho) +\frac{1}{2} (1-\rho)^2+\frac{1 }{2} (B_3-(1-\rho))^2.\end{aligned}$$ Multiplying by $2\eta_\eps(\rho)$ and using that $0\le\eta_\eps\le 1$ we obtain for $W_\eps(\rho)$ the estimate $$\label{wer}
W_\eps(\rho) =\eta_\eps{(\rho)} (1-\rho)^2\le (B_3-(1-\rho))^2+2\eta_\eps{(\rho)} (1-\rho)B_3.$$ Let $\psi_\eps(s):=2\frac{\eta_\eps(s)}{s}(1-s)= 2\min \{ \frac{1}{\eps}, \frac{1}{s}\} (1-s) $ then $\psi_\eps$ is bounded by $1/\eps$ and is Lipschitz continuous in $s^{1/2} $ with constant of order $\eps^{-3/2}$ i.e. $\sup_t |(\psi_\eps(t^2))'|\les \eps^{-3/2}$. Since $2\eta_\eps {(\rho)} (1-\rho)B_3= \rho B_3 \psi_\eps(\rho)$, using with $\psi=\psi_\eps{(\rho)} $, we get $$\begin{aligned}
\left| \int_{Q_{1,1}} 2\eta_\eps {(\rho)} (1-\rho)B_3 dx\right|&\les \frac{\a^{1/3}\b^{2/3}}{T}\lt(\eps^{-1} E+ \a^{-1/3}\b^{-1/6} E^{1/2} \|{\nabla}' (\psi_\eps(\rho))\|_{L^2} \rt) \\
&\les \frac{\a^{1/3}\b^{2/3}}{T}\lt(\eps^{-1}E+ \eps^{-3/2} \a^{-1/3}\b^{-1/6} E^{1/2} \|{\nabla}'\rho^{1/2}\|_{L^2} \rt) \\
&\les C_\eps \frac{\a^{1/3}\b^{2/3}}{T}\lt(E+ \int_{Q_{1,1}} \a^{-2/3}\b^{-1/3} |\nabla' \rho^{1/2}|^2 dx\rt) \\
&\les C_\eps \frac{\a^{1/3}\b^{2/3}}{T} E,\end{aligned}$$ where we used that $|\nabla' \rho^{1/2}|\le |\nabla'_{\a^{1/3}\b^{1/3}TA} u|$ and thus $\int_{Q_{1,1}} \a^{-2/3}\b^{-1/3} |\nabla' \rho^{1/2}|^2 dx\le E$. Estimate follows from inserting this estimate into .
To prove the lower bound, we will need the following two dimensional result.
\[lowerbound2d\] Let $\chi_n\in BV(Q_1, \{0,\b_n^{-1}\})$ be such that $\lim_{n\to +\infty}\int_{Q_1}\chi_n dx'=1$ and $$\sup_{n} \int_{Q_1} \b_n^{1/2} |D' \chi_n| <+\infty.$$ Then, up to a subsequence, $\chi_n\weaklim \sum_i {\varphi}_i \delta_{X_i}$ for some at most countable family of ${\varphi}_i> 0$ and $X_i \in Q_1$, and $$\label{liminfperibeta}
\liminf_{n\to +\infty} \int_{Q_1} \b_n^{1/2} |D' \chi_n|\ge 2\sqrt{\pi}\sum_i \sqrt{{\varphi}_i}.$$
[[*Step 1 (Compactness):*]{}]{} For each $n$ we split the cube $Q_1$ into small cubes of side length $3\b_n^{1/2}$. Let $Q_i^{n}$ be an enumeration of theses cubes such that $$\varphi_i^{n} :=\int_{Q_i^{n}} \chi_n dx'$$ is nonincreasing in $i$. Since $|\{\chi_n=\b_n^{-1}\}|=\b_n \int_{Q_1} \chi_n dx'=\b_n+o({\beta_n})$, we have $|{Q_i^{n}}\cap\{\chi_n=\b_n^{-1}\}|\le\b_n+o(\b_n)\le \frac{1}{2}|{Q_i^{n}}|$ and thus by the relative isoperimetric inequality [@AFP Th. 3.46], we have on each ${Q_i^{n}}$ $$\varphi_i^{n}=\int_{{Q_i^{n}}} \chi_n dx'\les\left( \int_{{Q_i^{n}}} \b_n^{1/2} |D'
\chi_n|\right)^2\,.$$ It follows that $$\sum_i \sqrt{\varphi_i^{n}} \les \int_{Q_1} \b_n^{1/2}|D' \chi_n| \le C,$$ by the energy bound. Arguing as in the proof of Proposition \[existmu\], we deduce from Lemma \[lemsqrt\] that up to extracting a subsequence, $\chi_n \weaklim \sum_{i\in {\mathcal{I}}}{\varphi}_i \delta_{X_i}$ for some ${\varphi}_i> 0$ and $X_i\in Q_1$, pairwise distinct.
[[*Step 2 (Lower bound):*]{}]{} Assume now that the ${\varphi}_i$ are labeled in a decreasing order and fix $N\in \N$. Choose $r\in(0,1/4)$ sufficiently small so that $$\label{bxr}
\B'(X_i,r)\cap \B'(X_j,r)=\emptyset \qquad \qquad \forall i,j \le N \text{ with } i\neq j,$$ and let $\psi\in C^\infty_c(\B'(X_i, r);[0,1])$ be a smooth function such that $\psi=1$ in $\B'(X_i, r/2)$ and $|{\nabla}\psi|\le C/r$. Let $C_\mathrm{iso}=(2\sqrt{\pi})^{-1}$ be the isoperimetric constant in dimension 2, then we may write $$\begin{aligned}
\nonumber
\left(\int_{\B'(X_i, r/2)
} \b_n^{-1} \chi_n dx'\right)^\hal &
\le &\left( \int_{Q_1} ( \psi \chi_n)^2 dx'\right)^{1/2}
\le C_\mathrm{iso} \int_{Q_1} |D' (\psi \chi_n)|\\
\nonumber & \le & C_\mathrm{iso} \left( \int_{\B'(X_i, r) }
|D' \chi_n| + \frac{C}{r} \int_{\B'(X_i,r)\backslash \B'(X_i, r/2)}
\chi_n dx'\right)\\
\nonumber & \le & C_\mathrm{iso} \int_{\B'(X_i, r) }
|D' \chi_n| + \frac{C}{r}
\end{aligned}$$ since $\int_{Q_1} \chi_n dx'\to 1$. Multiplying by $\beta_n^{1/2}$ and summing over $N$, we get $$\label{estimisop}
\sum_{i=1}^N \left(\int_{\B'(X_i, r/2)
} \chi_n dx'\right)^{1/2} \le C_\mathrm{iso} \int_{Q_1} \beta_n^{1/2}|D' \chi_n| + \beta_n^{1/2}\frac{CN}{r}.$$
Next observe that since $\chi_n\rightharpoonup
\sum_{i} {\varphi}_i \delta_{X_i}$, we have for every $i=1,\dots,N$, $$\liminf_{n\to +\infty}\lt(\int_{\B'(X_i, r/2)} \chi_n dx'\right)^{1/2}= {\varphi}_i^{1/2}
\,.$$ Therefore, passing to the limit in , we obtain $$\sum_{i=1}^N {\varphi}^{1/2}_i\le C_\mathrm{iso} \liminf_{n \to +\infty} \int_{Q_1} \beta_n^{1/2}|D' \chi_n|.$$ Since $N$ was arbitrary this implies .
With this lemma at hand, we can prove the compactness and lower bound result.
We fix for the proof a sequence $(u_n,A_n)$ with $\wE(u_n,A_n)\les 1$. We then let $B_n:=(B'_n,B_3^n):=\nabla \times A_n$ and $\rho_n:=|u_n|^2$.\
[[*Step 1 (Compactness):*]{}]{} Notice first that $$\label{bounddoublewell}
\int_{Q_{1,1}} \lt( \frac{B^n_3-(1-\rho_n)}
{\beta_n}
\rt)^2 dx\le \a_n^{-2/3} \b_n^{-4/3} \wE(u_n,A_n)
\to0\,.$$ By [@CoOtSer Lem. 3.7] there is $\hat u_n$ with $\hat\rho_n:=|\hat u_n|^2=\min\{\rho_n,1\}$ such that $\wE(\hat u_n,A_n)\le (1+2\alpha_n/T_n)\wE(u_n,A_n)$ (the error comes from the last two terms in (\[eqEGLE\])). In particular, also $ (B^n_3-(1-\hat\rho_n))/\beta_n\to0$ in $L^2(Q_{1,1})$. Using this and $|B^n_3|\le |B^n_3-(1-\hat\rho_n)|$ on $\{ B^n_3\le 0\}$ we obtain $$\label{convneg}\lt|\int_{\{B^n_3<0\}} \b_n^{-1} B^n_3 dx\rt|\les \lt( \b_n^{-2} \int_{Q_{1,1}} (B^n_3-(1-\hat\rho_n))^2 dx\rt)^\hal \le \frac{1}{(\alpha_n\beta_n^2)^{2/3} } \wE(u_n,A_n) \to 0,$$ and since $\int_{Q_{1,1}} \b_n^{-1}B^n_3 dx=2$ the sequence $\b_n^{-1}B^n_3$ is bounded in $L^1$ and, after extracting a subsequence, $\b_n^{-1} B^n_3 \weaklim \mu$ for some measure $\mu$. From we also get $\b_n^{-1}(1-\rho_n)\weaklim \mu$, and the same for $\hat\rho_n$. It also follows from that $$\int_{Q_{1,1}} \b_n^{-1} (1- \hat\rho_n) dx= 2+\left|\beta_n^{-1} \int_{Q_{1,1}} \left(B_3^n-(1-\hat \rho_n) \right) dx\right|\to 2.$$ Moreover, since $\int_{Q_{1,1}} \b_n^{-1}(1-\hat\rho_n) |B_n'|^2 dx\le\int_{Q_{1,1}} \b_n^{-1} |B'_n|^2 dx\le \wE(u_n,A_n)$ it holds $$\int_{Q_{1,1}} \b_n^{-1}(1-\hat\rho_n) | B_n'| dx\le \lt(\int_{Q_{1,1}} \b_n^{-1}(1-\hat\rho_n) |B_n'|^2 dx\rt)^\hal \lt(\int_{Q_{1,1}} \b_n^{-1}(1-\hat\rho_n) dx\rt)^{\hal}\les 1,$$ thus (up to a subsequence) $\b_n^{-1}(1-\hat\rho_n) B_n'\weaklim m$ for some vector-valued measure $m$. By [@AFP Th. 2.34], $$\label{liminftrans}
\liminf_{n\to +\infty} \int_{Q_{1,1}} \b_n^{-1}|B_n'|^2 dx\ge\liminf_{n\to +\infty} \int_{Q_{1,1}} \b_n^{-1}(1-\hat\rho_n)|B_n'|^2 dx\ge \int_{Q_{1,1}} \lt(\frac{dm}{d{\mu}}\rt)^2 d\mu,$$ and $m\ll \mu$. Moreover, from Lemma \[lemmameissner\] we have that $\b_n^{-1}\hat\rho_n B'_n\weaklim 0$ in a distributional sense and therefore $\b^{-1}_n B_n'$ itself converges to $m$. Letting $n\to +\infty$ in $$\b_n^{-1} \Div B_n=\partial_3 \left[\b_n^{-1} B^n_3\right]+\Divp \lt[\b_n^{-1}B_n'\rt]=0,$$ we obtain $\partial_3 \mu +\Divp m=0$. This proves (i).\
We now prove that $\mu=\mu_{x_3}\otimes dx_3$, that $\beta^{-1}_nB^n_3(\cdot,x_3)\weaklim \mu_{x_3}$ for a.e. $x_3\in(-1,1)$ and that $\mu_{x_3}\weaklim dx'$ as $x_3\to \pm 1$. By we have that, up to a subsequence in $n$, for a.e. $x_3\in (-1,1)$, $$\lim_{n\to +\infty} \int_{Q_1\cap \{B^n_3<0\}} \beta^{-1}_n B^n_3 dx=0.$$ Let $\mathcal{G}{\subseteq}(-1,1)$ be the set of $x_3$ for which this hold. For every $x_3\in \mathcal{G}$, the $L^1$ norm of $\beta^{-1}_nB^n_3(\cdot,x_3)$ is bounded thus if we fix a countable dense set $\mathcal{G}_d{\subseteq}\mathcal{G}$, we can assume up to extraction that for every $x_3\in \mathcal{G}_d$, $\beta^{-1}_n B^n_3(\cdot,x_3)\weaklim \nu_{x_3}$ for some probability measure $\nu_{x_3}$. For $x_3,{\widetilde}{x}_3\in \mathcal{G}_d$, thanks to the weak lower semi-continuity of $\|\cdot \|_{BL}$ and , $$\begin{aligned}
\|\nu_{x_3}-\nu_{{\widetilde}{x}_3}\|_{BL}&\le \limsup_{n\to +\infty} \|\beta_n^{-1} B_3^n (\cdot, x_3) - \beta_n^{-1} B_3^n(\cdot, {\widetilde}{x}_3)\|_{BL} \\
&\les \lim_{n\to +\infty}\lt( |x_3-{\widetilde}{x}_3|^{1/2} +\sigma(\alpha_n,\beta_n,T_n)\rt) \\
&= |x_3-{\widetilde}{x}_3|^{1/2}. \end{aligned}$$ Therefore, there exists a unique Hölder-continuous extension of $\nu_{x_3}$ to $(-1,1) \ni x_3 \mapsto \nu_{x_3} \in \mathcal{P}(Q_1)$. We claim that $\mu=\nu_{x_3}\otimes dx_3$. For $K\to +\infty$, let $\{z^K_j\}_{j=1}^{K+1} \in \mathcal{G}_d$ be an increasing sequence such that $|z^K_j-z^K_{j+1}|\les K^{-1}$, $|z_1^K+1|\les K^{-1}$ and $|z_{K+1}^K-1|\les K^{-1}$. Notice that for $n$ large enough, we have for every $j$ that $|z^K_j-z^K_{j+1}|\ge \sigma^{1/2}(\a_n,\b_n,T_n)$ (where $\sigma$ is defined in Lemma \[HolderE\]) so that applies. Let $\psi$ be a $Q_1-$periodic and Lipschitz continuous function on $Q_{1,1}$ with $\|\psi\|_{Lip}\le 1$. By the continuity of $x_3 \mapsto \nu_{x_3}$, we have $$\begin{aligned}
\int_{Q_{1,1}} \psi (d\mu- d\nu_{x_3}\otimes dx_3)&=\lim_{K\to +\infty} \int_{Q_{1,1}} \psi d\mu-\sum_{j=1}^K (z^K_{j+1}-z^K_j) \int_{Q_1} \psi(\cdot ,z_j^K) d\nu_{z_j^K}\\
&=\lim_{K\to +\infty} \lim_{n\to +\infty} \sum_{j=1}^K \int_{z_j^K}^{z_{j+1}^K} \int_{Q_1} \psi(x',x_3) \beta_n^{-1} B^n_3(x',x_3)\\
&\qquad \qquad \qquad \qquad \qquad -\psi(x',z_j^K)\beta_n^{-1} B^n_3(x',z_j^K) dx' dx_3.\end{aligned}$$ Using the finite difference version of Leibniz’ rule, $$\begin{gathered}
\psi(x',x_3) B^n_3(x',x_3)-\psi(x',z_j^K) B^n_3(x',z_j^K)\\
= B^n_3(x',x_3)(\psi(x',x_3)-\psi(x', z_j^K))+ \psi(x',z_j^K)( B^n_3(x',x_3) - B^n_3(x',z_j^K))\end{gathered}$$ and using that $\|\psi\|_{Lip}\le 1$, we can estimate for fixed $K,j$ and $n$ large enough, $$\begin{aligned}
\lt| \int_{z_j^K}^{z_{j+1}^K} \rt.& \lt.\int_{Q_1} \psi(x',x_3) \beta_n^{-1} B^n_3(x',x_3)-\psi(x',z_j^K)\beta_n^{-1} B^n_3(x',z_j^K) dx' dx_3\rt|\\
&\le \int_{z_j^K}^{z_{j+1}^K} \int_{Q_1} \beta_n^{-1} |B^n_3| |x_3-z^K_j| dx + \int_{z_j^K}^{z_{j+1}^K} \|\beta^{-1}_n( B^n_3(x',x_3) - B^n_3(x',z_j^K))\|_{BL}\\
&\les K^{-1} \int_{z_j^K}^{z_{j+1}^K} \int_{Q_1} \beta_n^{-1} |B^n_3|dx +K^{-1} K^{-1/2},\end{aligned}$$ where in the last line we have used that $|x_3-z^K_j|\les K^{-1}$ and . Summing this estimate over $j$, we obtain $$\left|\int_{Q_{1,1}} \psi (d\mu- d\nu_{x_3}\otimes dx_3)\right| \les \lim_{K\to +\infty} K^{-1}\lt[\lim_{n\to +\infty} \int_{Q_{1,1}}\beta_n^{-1} |B^n_3| dx + K^{-1/2}\rt]=0.$$ This establishes that $\mu= \nu_{x_3}\otimes dx_3$. Moreover, this proves that for every $x_3\in \mathcal{G}_d$, the whole sequence $\beta_n^{-1} B^n_3(\cdot,x_3)$ weakly converges to $\mu_{x_3}$. Since the set $\mathcal{G}_d$ was arbitrary, this proves the above convergence for all $x_3\in \mathcal{G}$.\
We finally show that the boundary conditions hold. For this we focus on $x_3=1$. For $x_3\in \mathcal{G}$, it holds by the weak lower semi-continuity of $\|\cdot \|_{BL}$, $$\begin{aligned}
\|1-\mu_{x_3}\|_{BL}&\le \liminf_{n\to +\infty} \|1-\beta_n^{-1}B_3^n(\cdot, x_3)\|_{BL}\\
& \le \liminf_{n\to +\infty} \|1-\beta_n^{-1}B_3^n(\cdot, 1)\|_{BL}+\limsup_{n\to +\infty} \|\beta_n^{-1}B_3^n(\cdot, 1)-\beta_n^{-1}B_3^n(\cdot, x_3)\|_{BL}.\end{aligned}$$ By , the second right-hand side term is controlled by $|1-x_3|^{1/2}$. For the first right-hand side term we note that because of $\|\psi\|_{H^{1/2}}{\stackrel{<}{\sim}}\|\psi\|_{Lip}$, we have $$\|\beta_n^{-1}B_3^n (\cdot, 1)- 1\|_{BL}^2 {\stackrel{<}{\sim}}\|\beta_n^{-1} B_3^n (\cdot, 1)-1\|_{H^{1/2}}^2 \le \alpha_n^{-1/3}\beta_n^{-7/6} \wE(\hat u_n, A_n).$$ Hence, it is the last assumption in that ensures that this term vanishes in the limit $n\to +\infty$. We thus obtain the desired estimate $$\| 1- \mu_{x_3}\|_{BL}{\stackrel{<}{\sim}}|1-x_3|^{1/2}.$$
[[*Step 2 (Lower bound and structure of $\mu$):*]{}]{} The starting point is an application of the usual Modica-Mortola trick. In this step we only deal with $\hat u_n$ and $\hat\rho_n$, and drop the hats for brevity. By and $|\nabla' \rho^{1/2}|\le |\nabla'_{\lambda A} u|$ we obtain from the Cauchy-Schwarz inequality : $$\label{estimMM}
\left(1 + C_\eps \frac{\alpha_n}{T_n}\right)\wE(u_n,A_n)\ge \int_{Q_{1,1}} \b_n^{-1/2}2\sqrt{W_\eps(\rho_n)} |\nabla' \rho_n^{1/2}| +\b_n^{-1}|B_n'|^2 dx,$$ for any $\eps>0$. We momentarily fix a small $\delta>0$ and estimate by the co-area formula , $$\begin{aligned}
\int_{Q_{1,1}} 2\sqrt{W_\eps(\rho_n)}|\nabla' \rho_n^{1/2}| dx&\ge \int_{\delta}^{1-\delta} \int_{-1}^1 2\sqrt{W_\eps(s^2)} \H^1(\partial \{\rho_n(\cdot,x_3)>s^2\}) dx_3 ds.\end{aligned}$$ In particular there exists $s_n\in[\delta,1-\delta]$ depending on $n$ such that $$\begin{gathered}
\int_{\delta}^{1-\delta} 2\sqrt{W_\eps(s^2)} \int_{-1}^1 \H^1(\partial \{\rho_n(\cdot,x_3)>s^2\}) dx_3 ds\\
\ge \lt(\int_{\delta}^{1-\delta} 2\sqrt{W_\eps(s^2)} ds\rt)\int_{-1}^1 \H^1(\partial \{\rho_n(\cdot,x_3)>s^2_n\}) dx_3.\end{gathered}$$ Letting $\chi_n(x',x_3):=\b_n^{-1}(1-\chi_{\{\rho_n(\cdot,x_3)>s_n^2\}}(x'))$ this reads $$\label{boundBV}
\int_{Q_{1,1}} \b_n^{-1/2}2\sqrt{W_\eps(\rho_n)}|\nabla' \rho_n^{1/2}| dx\ge C_{\delta,\eps}\int_{Q_{1,1}} \b_n^{1/2} |D' \chi_n| dx_3,$$ where $C_{\delta,\eps}:=\int_{\delta}^{1-\delta} 2\sqrt{W_\eps(s^2)} ds$.
Let $\gamma_n\to 0$ to be chosen later. For $n$ large enough, if $\rho_n\le \gamma_n$, then $\chi_n=\beta_n^{-1}$ while if $\rho_n\ge 1-\gamma_n$, $\chi_n=0$ so that $$\begin{aligned}
\int_{Q_{1,1}} |\chi_n-\beta_n^{-1}(1-\rho_n)| dx&\le\beta_n^{-1} \int_{\{\rho_n\le \gamma_n\}} \rho_n dx +\beta_n^{-1} \int_{\{\rho_n\ge 1-\gamma_n\}} (1-\rho_n) dx \\
& \qquad +\beta_n^{-1}|\{\gamma_n<\rho_n<1-\gamma_n\}|
\\ &\le 2\beta_n^{-1}\gamma_n +\beta_n^{-1}|\{\gamma_n<\rho_n<1-\gamma_n\}|.\end{aligned}$$ By definition of $W_\eps$ (recall , $\min_{[\gamma_n,1-\gamma_n]} W_\eps=\min(\frac{\gamma_n}{\eps}, \gamma_n^2)=\gamma_n^2$ so that using that $\int_{Q_{1,1}} \alpha_n^{2/3}\beta_n^{-2/3} W_\eps(\rho_n) dx\les 1$, $$\beta_n^{-1}|\{\gamma_n<\rho_n<1-\gamma_n\}|\le \beta_n^{-1} \gamma_n^{-2} \int_{Q_{1,1}} W_\eps(\rho_n) dx \les \beta^{-1/3}_n\gamma_n^{-2}\alpha_n^{-2/3}.$$ Therefore, if we choose $\gamma_n$ such that $\beta_n\gg \gamma_n\gg \alpha_n^{-1/3}\beta_n^{-1/6}$, which is possible since by hypothesis $\alpha_n \beta_n^{7/2}\to +\infty$, we obtain that $$\lim_{n\to +\infty}\int_{Q_{1,1}} |\chi_n-\beta_n^{-1}(1-\rho_n)| dx=0.$$ Combining this with , we obtain that $$\lim_{n\to +\infty} \int_{Q_{1,1}}|\chi_n-\b_n^{-1}B_3^n| dx=0.$$ By Fubini, this implies that, after passing to a subsequence in $n$, for a.e. $x_3\in (-1,1)$, if $ \b_n^{-1}B_3^n(\cdot,x_3)\weaklim \mu_{x_3}$ then also $\chi_n(\cdot,x_3)\weaklim \mu_{x_3}$. Moreover, from $\int_{Q_1} \b_n^{-1} B^n_3(x',x_3) dx'=1$ for a.e. $x_3\in(-1,1)$, we obtain that $\lim_{n\to +\infty}\int_{Q_1} \chi_n(x',x_3) dx'=1$ for a.e. $x_3\in(-1,1)$. We thus can use Lemma \[lowerbound2d\] to prove that $\mu_{x_3}=\sum_i \p_i \delta_{X_i}$ for some $\p_i>0$ and $$\label{lowerboundperim}
\liminf_{n\to +\infty} \int_{Q_{1,1}} \b_n^{1/2} |D'\chi_n| dx_3 \ge 2\sqrt{\pi} \int_{-1}^1 \sum_i \sqrt{\p_i} dx_3\, ,$$ where we used Fatou lemma. This shows (iii). Putting , , and together we find $$\liminf_{n\to +\infty} \wE(u_n,A_n)\ge \int_{-1}^1 2\sqrt{\pi} C_{\delta,\eps} \sum_i \sqrt{{\varphi}_i} dx_3 +\int_{Q_{1,1}} \lt(\frac{dm}{d\mu}\rt)^2 d\mu,$$ for any $\eps$ and $\delta$. Since $$\lim_{\eps\to0}\lim_{\delta\to0}C_{\delta,\eps}= 2\int_0^1 (1-t^2) dt=\frac{4}{3},$$ and $K_*=\frac{8\sqrt{\pi}}{3}$, this concludes the proof of (iv).
Upper bound {#Sec:upperbound}
===========
In this section we construct a recovery sequence for any sequences $T_n, \alpha_n, \beta_n$ which obey (\[eqlbassumptcoeff\]) and additionally the condition of quantization of the total flux $$\label{defKtotal}
L_n^2 T_n\a_n \b_n\in 2\pi \N\,,$$ where $L_n:={\widetilde}L \a^{-1/3}_n \b_n^{-1/6}$. Recalling the form of the gradient term in the functional $\wE$ defined in (\[eqdefwe\]), to discuss quantization of the flux of individual domains it is convenient to introduce $$\label{defK}
k_n:=\alpha_n^{1/3}\beta_n^{2/3}T_n\,.$$ The global flux quantization (\[defKtotal\]) then reads $k_n {\widetilde}L^ 2\in 2\pi\N$ and, in the ${\widetilde}L=1$ case we are considering here, simplifies to $k_n\in 2\pi \N$. Condition (\[eqlbassumptcoeff\]) implies $k_n\to +\infty$ and in particular $k_n/\beta_n\to+\infty$, so that the quantization condition becomes less and less stringent with increasing $n$. Aim of this section is to prove the following:
\[gammalimsup-v2\] Assume ${\widetilde}L=1$, (\[eqlbassumptcoeff\]) and (\[defKtotal\]). Then, for every $\mu$ with $I(\mu)<+\infty$ and $\mu_1=\mu_{-1}=dx'$ there exist sequences $u_n: Q_{1,1}\to {\mathbb{C}}$ and $A_n: Q_{1,1}\to \R^3$ such that $$\limsup_{n\to +\infty} \, \wE(u_n,A_n)\le I(\mu).$$ The fields $\rho_n:=|u_n|^2$ and $B_n:= \nabla \times A_n$ are $Q_1$-periodic and it holds $$\b_n^{-1}(1-\rho_n)\weaklim \mu \qquad \textrm{and } \qquad \b_n^{-1} B_n'\weaklim m,$$ where $m$ is the measure such that $I(\mu)=I({\mu},m)$.
The idea of the construction is to use the density result Theorem \[theodens\] to separate the construction in two regions. In the bulk, the measure will be approximated by a finite polygonal measure for which the construction is made in Section \[secconstrbulk\]. In the boundary layer, we plug in the construction of [@CoOtSer], see Section \[secconstrboundary\], which is optimal up to a factor. Since the energy in the boundary layer is small, its suboptimal effect disappears in the limit.
We shall first construct the density $\rho$ and the magnetic field $B$. The appropriate energy is ${\widetilde}F_{\alpha,\beta}:={\widetilde}F_{\alpha,\beta}^{(-T,T)}+{\widetilde}F_{\alpha,\beta}^{\Ext}$, where $$\begin{gathered}
\label{eqdeftildaF}
{\widetilde}F_{\alpha,\beta}^{(a,b)}(\rho,B):=\int_{Q_1\times(a,b)} \Bigl( \a^{-2/3}\b^{-1/3}\lt|\nabla'\rho^{1/2}\rt|^2+
\a^{-4/3}\b^{-2/3}\lt|\partial_3\rho^{1/2}\rt|^2\\
+ \a^{2/3}\b^{-2/3}\left({B_3} -(1-\rho)\right)^2
+ \b^{-1}|{B'}|^2 \Bigr)dx\,,\end{gathered}$$ and $$\label{eqdeftildaFext}
{\widetilde}F_{\alpha,\beta}^{\Ext}(B):=
\alpha^{1/3}\beta^{7/6}
\| \beta^{-1}B_3-1\|^2_{H^{-1/2}(Q_1\times \{\pm1\})}\,.$$ Their sum corresponds to the energy $\wE$, up to a reconstruction of $u$ and $A$ that will be discussed in Section \[recoveryGL\]. A pair $(\rho,B)$ is admissible for ${\widetilde}F_{\a,\b}^{(a,b)}$ if $\rho$ and $B$ are $Q_1$-periodic, $\div B=0$ and $\int_{Q_1} B_3(x',x_3)dx'=\beta$ for all $x_3$.
We say that a pair $(\rho,B)$ is $k$-quantized if there is a closed set $\omega{\subseteq}Q_{1,1}$ such that $B=0$ outside $\omega$, $\rho=0$ in $\omega$, and the flux of $ \b^{-1}B_3$ over every connected component of $\omega\cap\{x_3=z\}$ is an integer multiple of $2\pi /k$, for all $z\in(-1,1)$.
Construction in the bulk {#secconstrbulk}
------------------------
This section is concerned with the local construction of flux tubes. For notational simplicity we present this construction in $\R^2\times(a,b)$, without the periodicity assumption; since $\rho=1$ and $B=0$ outside a small region, its periodic extension is immediate (see proof of Proposition \[propubregular\] below). We start from the optimal profile at the boundary of the individual tubes, with the lengths measured in units of the coherence length (recall ). For the purpose of the upcoming constructions, we cut the profile at a lengthscale $R\gg1$ towards the normal region.
\[lemmaVR\] Consider the functional $\displaystyle G(v):=2\sqrt\pi\int_0^{+\infty} |\dot{v}|^2+(1-v^2)^2 dt$. Then, $$\inf\left\{ G(v): v(0)=0, \lim_{t\to+\infty} v(t)=1\right\}=K_*=\frac83\sqrt\pi\,.$$ Furthermore, for all $R\ge3$ there is $v_R\in C^\infty(\R;[0,1])$ with $v_R(t)=0$ for $t\le 0$, $v_R(t)=1$ for $t\ge R$, $|\dot{v}_R|\le 2$, and, setting $K_R:=G(v_R)$, one has $$\lim_{R\to\infty} K_R=\frac83\sqrt\pi,$$ and $\int_0^{+\infty} t \left(|\dot{v}_R|^2+(1-v_R^2)^2\right) dt{\stackrel{<}{\sim}}1$.
The lower bound follows from the usual Modica-Mortola type computation $$\int_0^{+\infty} |\dot{v}|^2+(1-v^2)^2dt\ge 2\int_0^{+\infty} (1-v^2)\dot{v}dt=2\int_0^1 (1-s^2) ds = \frac43\,.$$ To prove the upper bound, we recall that $v(t):= \tanh t$ is the minimizer of $G$ under the constraint $v(0)=0$. A direct computation shows that $G(v)=K_*$. We then define for $R>0$, $$\hat v_R(t):=
\begin{cases}
0 & \text{ if } t<1/R,\\
\displaystyle\frac{\tanh (t-1/R)}{\tanh (R-2/R)}&\text{ if } t\in[1/R,R-1/R],\\
1 & \text{ if } t>R-1/R,
\end{cases}$$ and $v_R:=\psi_{1/R}\ast \hat v_R$, with $\psi_{1/R}\in C^\infty_c(-1/R,1/R)$ a mollifier. By construction, this has the desired properties and verifies $G(v_R)\to G(v)$ as $R\to\infty$.
We start with the simple case in which the limiting measure comes from a Lipschitz curve; in practice this will be used only for affine or piecewise affine curves.
\[lemmacurve3\] Let $X: (a,b)\to \R^2$ be a Lipschitz curve, ${\varphi}>0$, $R>0$. We define $$\rho(x):= v_R^2\left(\frac{|x'-X(x_3)|-\sqrt{\beta{\varphi}/\pi}}{\eta}\right),$$ where $\eta:=\alpha^{-2/3}\beta^{1/6}\ll 1$ is the coherence length (see ) and $v_R$ was introduced in Lemma \[lemmaVR\], and define $B$ through $$B_3(x):=\chi_{\B'(X(x_3),\sqrt{\beta {\varphi}/\pi} )}(x')\,, \hskip1cm
B'(x):=B_3(x) \dot{X}(x_3) \,.$$ Then, $\rho B=0$ almost everywhere, $\div B=0$, and $${\widetilde}F^{(a,b)}_{\alpha,\beta}(\rho,B)\le \left(1+ \frac{C}{({\varphi}\alpha^{4/3}\beta^{2/3})^{1/2}}
+ \frac{C}{{\varphi}\alpha^{4/3}\beta^{2/3}}\right)
\int_a^b \bigl( K_R\sqrt{{\varphi}}+{\varphi}|\dot{X}|^2\bigr) dx_3.$$ The constant $C$ is universal. Moreover, if $\mu:={\varphi}\delta_{X(x_3)}\otimes dx_3$, $$\label{distestim}
W_2^2(\b^{-1} B_3, \mu)\le |b-a|\frac{\beta {\varphi}^2}{2\pi}.$$
The condition $\rho B=0$ follows from $v_R(t)=0$ for $t\le 0$. To check the divergence condition, pick $\psi\in C^1_c(\R^2\times(a,b))$ and compute $$\begin{aligned}
\int_{\R^2\times(a,b)} (B_3\partial_3 \psi + B'\cdot \nabla'\psi) dx
&=\int_a^b \int_{\B'(X(x_3), \sqrt{\beta{\varphi}/\pi})}
( \partial_3 \psi + \nabla'\psi \cdot \dot{X}(x_3)) dx'dx_3\\
& =\int_a^b \int_{\B'(0, \sqrt{\beta{\varphi}/\pi})}
\frac{d}{dx_3} \psi(X(x_3)+y', x_3) dy' dx_3=0\,.\end{aligned}$$ We now estimate the energy. We start from the interfacial energy at fixed $x_3$ (see ), $$\begin{aligned}
E_1(x_3):=\int_{\R^2} \alpha^{-2/3}\beta^{-1/3} |\nabla'\rho^{1/2}|^2 + \alpha^{2/3}\beta^{-2/3}
(B_3-(1-\rho))^2 dx'\,.\end{aligned}$$ If $|x'-X(x_3)|< \sqrt{\beta{\varphi}/\pi}$ then $\rho=0$ and $B_3=1$, whereas if $|x'-X(x_3)|>\eta R +\sqrt{\beta{\varphi}/\pi}$ then $\rho=1$, $B_3=0$. In the intermediate region, we use $|\nabla'\rho^{1/2}|=|\dot{v}_R|/\eta$. Passing to polar coordinates and using $r=|x'-X(x_3)|$ as an integration variable, $$\begin{aligned}
E_1(x_3)=
\int_{\sqrt{\beta{\varphi}/\pi}}^{\sqrt{\beta{\varphi}/\pi}+\eta R} \Bigl[
\frac{\alpha^{-2/3}\beta^{-1/3}}{\eta^2} |\dot{v}_R|^2
+ \alpha^{2/3}\beta^{-2/3} (1-v_R^2)^2\Bigr]
\lt(\frac{r-\sqrt{\beta{\varphi}/\pi}}\eta\rt)2\pi r dr \,.\end{aligned}$$ We change variables according to $r=\sqrt{\beta{\varphi}/\pi}+s\eta$, insert the definition of $\eta$, and by Lemma \[lemmaVR\] obtain $$\begin{aligned}
E_1(x_3)&=
\int_0^R \beta^{-1/2}
\left[|\dot{v}_R|^2+ (1-v_R^2)^2\right]2\pi (\sqrt{\beta{\varphi}/\pi}+s\eta) ds \\
& \le K_R \sqrt{\varphi}+ C \alpha^{-2/3}\beta^{-1/3}
\le \sqrt{\varphi}\left(K_R+\frac{C}{{\varphi}^{1/2}\alpha^{2/3}\beta^{1/3}}\right)\,.\end{aligned}$$ The other contributions to the energy are the cost of transport and the vertical part of the gradient, $$\begin{aligned}
E_2(x_3):=\int_{\R^2} \alpha^{-4/3}\beta^{-2/3} |\partial_3\rho^{1/2}|^2 + \beta^{-1}|B'|^2 dx'\,.\end{aligned}$$ By definition of $B'$, we have for the second term, $$\begin{aligned}
\int_{\R^2} \beta^{-1}|B'|^2 dx'= |\dot{X}(x_3)|^2 {\varphi}\,.\end{aligned}$$ For the first one we use $|\partial_3\rho^{1/2}|\le |\dot{v}_R||\dot{X}|/\eta$ and change variables as above to obtain $$\begin{aligned}
\int_{\R^2} \alpha^{-4/3}\beta^{-2/3} |\partial_3\rho^{1/2}|^2 dx'
&{\stackrel{<}{\sim}}\frac{ \alpha^{-4/3}\beta^{-2/3}}{\eta} |\dot{X}(x_3)|^2
\int_0^R ( \sqrt{\beta{\varphi}/\pi}+s\eta)|v_R'|^2 ds\\
&{\stackrel{<}{\sim}}\alpha^{-4/3}\beta^{-2/3} |\dot{X}(x_3)|^2 (1 + \sqrt{\beta{\varphi}}/\eta)\\
&= |\dot{X}(x_3)|^2 {\varphi}\left(
\frac{1}{{\varphi}\alpha^{4/3}\beta^{2/3}}+
\frac{1}{{\varphi}^{1/2}\alpha^{2/3}\beta^{1/3}}
\right)
\,.\end{aligned}$$ To prove we consider the transport map $T(x',x_3)=(X(x_3),x_3)$ which gives $$W_2^2(\b^{-1}B_3,\mu)\le \int_a^b \b^{-1} \int_{\B'(X(x_3),\sqrt{\b {\varphi}/\pi})} |x'-X(x_3)|^2 dx'dx_3=|b-a| \frac{\b {\varphi}^2}{2\pi}.$$
We now turn to the construction around branching points. Since the total length around branching points is small, the construction here does not need to achieve the optimal constant but only the optimal scaling. The idea of the construction is the following. We first transform disks into squares, then split the square into two rectangles and then retransform each rectangle into a disk. The construction is sketched in Figure \[fig1\]. We start with the transformation from a rectangle to a disk.
![Construction around a branching point[]{data-label="fig1"}](figsep2-crop){width="6cm"}
\[lemmacurveendpoints\] Let $a\le b\in \R$, $\gamma>0$, $X\in \R^2$, ${\varphi}>0$, and $R\ge 3$. Let then $Q {\subseteq}\R^2$ be a rectangle with side lengths $w$ and $h$ centered in $X$, such that $wh=\beta{\varphi}$, $w/h+h/w\le \gamma$ and $\a^{-4/3}\b^{-2/3}\le {\varphi}$. Let as before $\eta:=\alpha^{-2/3}\beta^{1/6}$ be the coherence length.
Then there are $\rho\in L^\infty(\R^2\times [a,b];[0,1])$ and $B\in L^2(\R^2\times [a,b];\R^3)$ such that $\div B=0$, $\rho B=0$, with $B=0$ and $\rho=1$ on $(\R^2\setminus \B'(X,r))\times[a,b]$ for some $r\sim \eta R + \sqrt{\beta{\varphi}}$, and $${\widetilde}F^{(a,b)}_{\alpha,\beta}(\rho, B) {\stackrel{<}{\sim}}\sqrt{\varphi} |a-b| R^2+ \frac{r^2}{|a-b|}\varphi R^2,$$ where the implicit constants only depend on $\gamma$. Further, $\rho$ and $B$ satisfy the boundary conditions $$\rho(x',a)= v_R^2\left(\frac{|x'-X|-\sqrt{\beta{\varphi}/\pi}}{\eta}\right),
\hskip1cm
\rho(x',b)= \min\left\{1,\frac{\dist^2(x', Q)}{\eta^2}\right\}\,,$$ and $$B_3(x',a)= \chi_{\B'(X, \sqrt{\beta{\varphi}/\pi})}(x')\,,
\hskip3mm\text{ }\hskip3mm
B_3(x',b)= \chi_{Q}(x')\,.$$
We note that our assumptions on the parameters $w, h, {\varphi},$ just mean $w\sim h\sim \sqrt{\beta {\varphi}}\ge \eta$. That is, the thread diameter is large compared to the coherence length. Note that $r\sim \eta R+\sqrt{\beta {\varphi}}$ behaves like the maximum between the thread diameter $w\sim h$ and the cut-off scale $\eta R$. The proof of Lemma \[lemmacurveendpoints\] is based on an explicit construction for a bilipschitz bijection with unit determinant that transforms a rectangle into a circle, which we first present.
![The construction in Lemma \[lemmasquarecircle\] transforms a rectangle into a circle, keeping $\det \nabla'u=1$. Each curve corresponds to a different value of $x_3$, and plots the solution of $\hat r(r,\theta)=t$, which corresponds to $r(\theta)=t/[\lambda(x_3)\cos(\theta x_3)]$ for $\theta\in (-\pi/4,\pi/4)$.[]{data-label="fig-square"}](fig-square-crop){width="5cm"}
\[lemmasquarecircle\] Assume that $z_-,z_+,h,w>0$, $X\in\R^2$ are given, $z_-< z_+$. Then there is $u:\R^2\times[z_-,z_+]\to\R^2$ such that $$u(x',z_-)=x'\,,\hskip5mm
u(X+(-\frac12w,\frac12w)\times(-\frac12h,\frac12h),z_+)=\B'(X,\sqrt{hw/\pi})\,,\hskip5mm
\det \nabla' u =1 \text{ a.e.}\,.$$ The function $x\mapsto (u(x),x_3)$ is bilipschitz, its inverse is of the form $y\mapsto (U(y),y_3)$, and $u(X,x_3)=X$ for all $x_3$. If additionally $h/w+w/h\le \gamma$, then the bounds $|\nabla'u|+|\nabla'U|{\stackrel{<}{\sim}}1$, $|\partial_3 u|+|\partial_3 U|{\stackrel{<}{\sim}}h/|z_+-z_-|$ hold, with constants which only depend on $\gamma$.
By scaling and translation we may assume that $hw=\pi$, $z_+=1$, $z_-=0$, $X=0$. We can further assume $h=w=\sqrt\pi$, as the general case is obtained by taking the composition at each $x_3$ with the linear map ${\mathop{\mathrm {diag}}}(g(x_3), 1/g(x_3))$, where $g(x_3):=\sqrt{h/w}(1-x_3)+x_3$. We work in polar coordinates, and construct functions $\hat r$, $\hat\theta$ of $r$, $\theta$ and $x_3$ such that $$u(r\cos\theta,r\sin\theta,x_3)=(\hat r\cos\hat\theta, \hat r\sin\hat\theta),$$ with $r\ge 0$ and $0\le\theta\le \pi/4$, and then extend by symmetry. We set $$\hat \theta = f(\theta, x_3) \,,\hskip5mm \hat r = r \lambda(x_3) \cos(x_3\theta),$$ where $\lambda$ and $f$ are two functions still to be determined (see Figure \[fig-square\]). The extension of the $\frac{1}{8}-$sectors by reflection is feasible provided that $f(0,x_3)=0$ and $f(\pi/4,x_3)=\pi/4$ for all $x_3$, the boundary data are attained provided that $f(\theta,0)=\theta$, $\lambda(0)=1$, $\lambda(1)=2/\sqrt\pi$. The latter ensures that indeed the straight segment $r \cos \theta=\frac{w}{2}=\frac{\sqrt{\pi}}{2}$ is mapped into the unit circle $\hat{r}=1$. The determinant condition is equivalent to $$1=\frac{\hat r}{r}\partial_r \hat{r} \partial_\theta f= \lambda^2(x_3)\cos^2(x_3\theta) \partial_\theta f(\theta, x_3),$$ which can be solved (using $f(0,x_3)=0$) to give $$f(\theta,x_3):=\frac{1}{x_3\lambda^2(x_3)}\tan (\theta x_3)\,,$$ smoothly extended to $f(\theta,0)=\theta/\lambda^2(0)$. The condition $f(\pi/4,x_3)=\pi/4$ determines $\lambda$, $$\lambda(x_3):=\left(\frac{\tan (\pi x_3/4)}{\pi x_3/4}\right)^{1/2},$$ which obeys $\lambda(1)=2/\sqrt\pi$ and smoothly extends to $\lambda(0)=1$. Clearly $\lambda\sim 1$ so that $\partial_\theta f\sim 1$. This implies that the change of variables defines a smooth deformation of the $\frac{1}{8}-$sector which smoothly depends on $x_3\in[0,1]$.
After a rotation and a translation, we may assume that $X=0$, $a=0$ and $b>0$; so that $Q=(-w/2,w/2)\times (-h/2,h/2)$. We treat two regions separately.
In the lower region $\R^2\times [0,b/2]$, we interpolate between $\B'_0:=\B'(0, \sqrt{\beta{\varphi}/\pi})$ and $Q$. To do this, let $u$ and $U$ be the functions from Lemma \[lemmasquarecircle\], using it for the rectangle $Q$ and $z_-=0$, $z_+=b/2$. The magnetic field is defined by $$B_3(x):=\chi_{\B'_0}(u(x))\,,\hskip1cm
B'(x):=\chi_{\B'_0}(u(x))\partial_3 U(u(x),x_3)\hskip3mm \text{ for } x_3\in[0,b/2]\,.$$ The density is defined by $$\rho(x):=v_R^2\left(\frac{|u(x)|-\sqrt{\beta{\varphi}/\pi}}{\eta}\right) \hskip3mm \text{ for } x_3\in[0,b/2]\,.$$ The condition $\rho B=0$ follows immediately, as well as the boundary data at $x_3=0$. Since $B=0$ and $\rho=1$ whenever $|u(x)|\ge \eta R + \sqrt{\beta{\varphi}/\pi}$ and since $|u(x)|\ge |x'|/\|\nabla'U\|_\infty$, we have $B=0$ and $\rho=1$ whenever $|x'| \ge \|\nabla'U\|_\infty (\eta R + \sqrt{\beta{\varphi}})$. In order to check $\div B=0$, we fix $\psi\in C^1_c(\B'_r\times(0,b/2))$. Performing then a change of variables at each $x_3$ gives, since $\det \nabla'U=1$, $$\begin{aligned}
1
\int_{\R^2\times(0,b/2)} \left[\partial_3 \psi B_3 + \nabla'\psi\cdot B' \right]dx &=
\int_{\R^2\times(0,b/2)} (\partial_3\psi B_3 + \nabla' \psi \cdot B')(U(y),y_3) dy
\\
&= \int_{\R^2\times(0,b/2)} \chi_{\B'_0} \left[\partial_3 \psi( U(y),y_3) + \partial_3 U(y)\cdot \nabla'\psi( U(y),y_3)) \right] dy \\
&= \int_{\B'_0} \int_{(0,b/2)} \frac{d}{dx_3} (\psi(U(y),y_3)) dy_3 dy'=0\,.\end{aligned}$$ By the properties of $u$ we obtain $B_3(x',b/2)=\chi_Q(x')$.
In the upper region $\R^2\times[b/2,b]$, we keep $B(x',x_3):=e_3 \chi_Q(x')$ concentrated on $Q$ and linearly interpolate the profile between $\rho^{1/2}(x',b/2)$ and the profile $\rho^{1/2}(x',b)$ given in the statement: $$\rho^{1/2}(x):=\frac{2}{b}\lt[
\rho^{1/2}(x',b/2) (b-x_3)+ \rho^{1/2}(x',b)(x_3-b/2)\rt]
\hskip3mm \text{ for } x_3\in(b/2,b)\,.$$ It is immediate to check that $B_3$ and $\rho$ match continuously at all interfaces, $\rho B=0$ and $\div B=0$. This concludes the construction.
We estimate the energy similarly to the proof of Lemma \[lemmacurve3\]. Lemma \[lemmasquarecircle\] gives $|\nabla'u|{\stackrel{<}{\sim}}1$, $|\partial_3 u|{\stackrel{<}{\sim}}h/b$, $|\partial_3 U|{\stackrel{<}{\sim}}h/b$, with constants depending only on $\gamma$. This yields $|B'|\les h/b$. Furthermore, by Lemma \[lemmaVR\] $|v_R|\le1$ and $|\dot{v}_R|\le 2$, so that $\rho\le 1$, $|\nabla'\rho^{1/2}|{\stackrel{<}{\sim}}1/\eta $ and $|\partial_3\rho^{1/2}|{\stackrel{<}{\sim}}h/(\eta b)$.
We start with the region $x_3\in [0,b/2]$. All integrals in $x'$ can be restricted to the set $$\Omega_{x_3}:=\{x': \sqrt{
\beta{\varphi}/\pi}\le |u(x',x_3)|\le \eta R+\sqrt{\beta{\varphi}/\pi}\}.$$ Since $\det \nabla u'=1$, we have $$|\Omega_{x_3}|=|\B'(0,\eta R+\sqrt{\beta{\varphi}/\pi})\setminus \B'_0|
= \pi \eta^2R^2+2\sqrt{\beta{\varphi}\pi}\eta R{\stackrel{<}{\sim}}\eta \sqrt{\beta{\varphi}} R^2,$$ (in the last step we used the assumption on ${\varphi}$ and $R\ge 3$). Therefore $$\begin{aligned}
{\widetilde}F_{\alpha,\beta}^{(0,b/2)}(\rho,B)
&{\stackrel{<}{\sim}}\int_{0}^{b/2}\int_{\Omega_{x_3}} \left( \frac{\alpha^{-2/3}\beta^{-1/3}}{\eta^2}
+ \frac{h^2\alpha^{-4/3}\beta^{-2/3}}{\eta^2 b^2} + \alpha^{2/3}\beta^{-2/3} + \frac{h^2}{\beta b^2}\right) dx' dx_3\\
&{\stackrel{<}{\sim}}\int_{0}^{b/2}|\Omega_{x_3}| dx_3 \left( \alpha^{2/3}\beta^{-2/3} + \frac{h^2}{\beta b^2}\right) \\
&{\stackrel{<}{\sim}}\left( b + \frac{h^2}b \alpha^{-2/3}\beta^{-1/3}\right) R^2 \sqrt{{\varphi}}\,\\
&{\stackrel{<}{\sim}}\sqrt{{\varphi}} b R^2 +\frac{r^2}{b} {\varphi}R^2,
\end{aligned}$$ where in the last line we have used that $h^2\les r^2$ and $\a^{-2/3}\b^{-1/3}\les \sqrt{{\varphi}}$. The region $(b/2,b)$ is simpler, as the $|B'|^2$ term does not appear, the others are the same with the exception of $|\partial_3 \rho^{1/2}|\les 1/b$, which is smaller by the factor $\eta/h\les 1$.
Using this building block we can finally produce the construction that will be used at branching points.
\[Lembranch3\] Let $a<b\in \R$, $\gamma>0$, $\ell>0$, $R\ge 3$, $X_i\in\R^2$ and $\varphi_i>0$ for $i=0,1,2$ with ${\varphi}_0={\varphi}_1+{\varphi}_2$, ${\varphi}_0/{\varphi}_1+{\varphi}_0/{\varphi}_2\le \gamma$ and $\sqrt{\beta {\varphi}_0}\ge \eta$, where as above $\eta:=\alpha^{-2/3}\beta^{1/6}$.
Then, if $|X_0-X_1|,|X_0-X_2|\le \ell/4$, $\sqrt{\beta \varphi_0}\ll |X_1-X_2|$ and $ \eta R\ll \ell$, then there are $\rho\in L^\infty(\R^2\times [a,b];[0,1])$ and $B\in L^2(\R^2\times [a,b];\R^3)$ such that $\div B=0$, $\rho B=0$, $B=0$ and $\rho=1$ on $(\R^2\setminus \B'(X_0,3\ell/4))\times[a,b]$, $${\widetilde}F^{(a,b)}_{\alpha,\beta}(\rho, B) {\stackrel{<}{\sim}}\sqrt{\varphi_0} |a-b| R^2+ \frac{\ell^2}{|a-b|}\varphi_0R^2,$$ where the implicit constants only depend on $\gamma$, and which satisfy the boundary conditions$$\rho(x',a)=v_R^2\left(\frac{|x'-X_0|-\sqrt{\beta{\varphi}_0/\pi}}{\eta}\right)
\hskip1mm\text{ }\hskip1mm
\rho(x',b)=\min_{i=1,2}v_R^2\left(\frac{|x'-X_i|-\sqrt{\beta{\varphi}_i/\pi}}{\eta}\right)\,,$$ and $$B_3(x',a)= \chi_{\B'(X_0, \sqrt{\beta{\varphi}_0/\pi})}(x')\,,
\hskip3mm\text{ }\hskip3mm
B_3(x',b)=\sum_{i=1,2} \chi_{\B'(X_i, \sqrt{\beta{\varphi}_i/\pi})}(x')\,.$$
Note that our assumptions on the parameters ${\varphi}_0, {\varphi}_1, {\varphi}_2$ and $h$ just mean that $\eta\les \sqrt{\beta {\varphi}_i}\ll |X_1-X_2|\les \ell$ and $\eta R\ll \ell$. That is, the thread diameter is at least as large as the coherence length $\eta$ but small compared to the distance between the threads. Likewise, $\ell$ is large compared to the cut-off scale $\eta R$.
After a translation and a rotation we may assume that $a=0$, $b>0$, $X_0=0$, $X_2-X_1=\zeta e_1$, with $\zeta>0$. Let $$w_0:=h:=(\b {\varphi}_0)^{1/2}, \qquad w_1:= \frac{{\varphi}_1}{{\varphi}_0} w_0 \qquad \textrm{ and } \qquad w_2:= \frac{{\varphi}_2}{{\varphi}_0}w_0.$$ Let then $Q^i:=X_i+[-w_i/2,w_i/2]\times[-h/2,h/2]$, for $i=0,1,2$. Notice that since $|X_i|\le \ell/4$ and $w_i\le w_0=h\ll \ell$, $Q^i{\subseteq}\B'(3\ell/8)$.\
We divide the interval $(0,b)$ in three parts (see again Figure \[fig1\]). In $(0,b/3)$, we apply Lemma \[lemmacurveendpoints\] to transform $\B'(\sqrt{\beta{\varphi}_0/\pi})$ into $Q^0$ (in particular we have $\rho=1$ on $\lt(\R^{2}\backslash \B'(\frac{3\ell}{4})\rt)\times (0,b/3)$). In $(b/3,2b/3)$, we connect $Q^0$ to $Q^1$ and $Q^2$ by an explicit construction (see below). Finally, in $(2b/3,b)$ we apply again Lemma \[lemmacurveendpoints\] to transform $Q^1$ and $Q^2$ back into $\B'(X_1, \sqrt{\beta {\varphi}_1/\pi})$ and $\B'(X_2, \sqrt{\beta {\varphi}_2/\pi})$. Notice that in $(2b/3,b)$, if $\rho(x',x_3)\ne 1$ then by Lemma \[lemmacurveendpoints\] and our hypothesis on the parameters, necessarily$$|x'|\le \max_i |X_i|+O(\eta R+\sqrt{\beta{\varphi}_0})\le \frac{\ell}{4} +o(\ell).$$ Thus, $\rho=1$ on $\lt(\R^{2}\backslash \B'(\frac{3\ell}{4})\rt)\times (2b/3,b)$.\
It only remains to discuss the construction in the central region. Let $y_1:=X_1+w_2e_1/2$, $y_2:=X_2-w_1e_1/2$ and, for $i=1,2$, ${\widetilde}Q^i:=Q^i-y_i$, so that up to a null set, $Q^0$ is the disjoint union of ${\widetilde}Q^1$ and ${\widetilde}Q^2$ (see Figure \[figseparsquares\]). Since $y_2-y_1=X_2-X_1-w_0e_1/2$, and by assumption $w_0=(\beta{\varphi}_0)^{1/2} \ll|X_1-X_2|$, we have $(y_2-y_1)\cdot e_1>0$.\
For $i=1,2$ and $x_3\in(b/3,2b/3)$ we set ${\widetilde}{Q}^i(x_3):= {\widetilde}{Q}_i+\frac{x_3-b/3}{b/3} y_i$. Since $(y_2-y_1)\cdot e_1>0$, we have ${\widetilde}{Q}^1(x_3)\cap{\widetilde}{Q}^2(x_3)=\emptyset$ for all $x_3$. Furthermore, since $Q^i{\subseteq}\B'(3\ell/8)$, also ${\widetilde}{Q}^i(x_3){\subseteq}\B'(3\ell/8)$ for $x_3\in(b/3,2b/3)$. We finally let $$\rho^{1/2}(x):=\min \lt\{ 1,\eta^{-1}\dist(x',{\widetilde}{Q}^1(x_3)\cup {\widetilde}{Q}^2(x_3))\rt\}\hskip3mm \text{ for } x_3\in[b/3,2b/3],$$ and correspondingly $$(B',B_3)(x):=
\sum_{i=1,2}
\lt( \frac{y_i}{b/3},1\rt)\chi_{{\widetilde}{Q}^i(x_3)}(x')\,.$$ All admissibility conditions are easily checked. In particular, $\rho=1$ if $\dist(x',{\widetilde}{Q}^1\cup{\widetilde}{Q}^2(x_3))\ge \eta $ which holds if $|x'|\ge 3\ell/4$. The energy estimate is immediate.
Boundary layer {#secconstrboundary}
--------------
\[propboundraylayer\] Let $N\in\N$, $\alpha,\beta,T,t>0$ be given. Let $k:=\a^{1/3}\b^{2/3}T$ and let ${\varphi}_1, \dots, {\varphi}_{N^2}$ be positive numbers such that $$k {\varphi}_i \in 2\pi \N \, \qquad \textrm{and } \qquad\sum_{i=1}^{N^2} {\varphi}_i=1.$$ Assume that in the regime $\beta\le \alpha$, $\alpha \sqrt{2}/T\le 1$, we have $t\gg \alpha^{-1}$, $1/N\ges \alpha^{-2/3}\beta^{-1/3}$ and ${\varphi}_i\sim 1/N^2$ for every $i$. Then there are $\rho$ and $B$, admissible for ${\widetilde}F_{\alpha,\beta}^{(0,t)}$ and $k$-quantized, such that $$\label{eqboundarylint}
{\widetilde}F_{\alpha,\beta}^{(0,t)}(\rho,B)
\les tN + \frac{1}{N^2t},$$ $$\label{eqboundarylext}
\alpha^{1/3}\beta^{7/6}
\|\beta^{-1}{B_3}- 1\|_{H^{-1/2}(x_3=0)}^2\les \frac{\beta^{1/3}}{\alpha^{1/3}} + \frac{\beta^{1/2}}{N^2t}
+\alpha^{1/3}\beta^{7/6},$$ and, denoting by $Q^i$ the squares centered in the $N^2$ points of the square grid of spacing $N^{-1}$ with $|Q^i|=\beta {\varphi}_i$, $$\label{eqrandwb3bdly}
B_3(x',t)=\sum_i \chi_{Q^i}(x') \text{ and }
\rho^{1/2}(x',t)=\min\{ 1,\eta^{-1} \dist(x',\cup_i Q^i) \}\,.$$
Note that the assumptions on the parameters $N$ and ${\varphi}_i$ mean that the sidelength $\sqrt{\beta {\varphi}_i}$ of the squares is larger than the coherence length $\a^{-2/3}\b^{1/6}$ (see ) and that both of them are small compared to the distance (equal to $1/N$) between the squares. The assumption on the parameter $t$ means that the thickness $t$ of the boundary layer is large with respect to the coherence length, where we recall that vertical and horizontal lengths have different units.
The key construction is described in [@CoOtSer Lem. 4.7]. However, the notation and the scalings are different in that paper. Indeed, since there was no need to rescale $x_3$ by the thickness and $x'$ by the distance between the threads $T \alpha^{-1/3} \beta^{-1/6}$, in that paper length was measured in terms of the penetration length. We first let $r_i$ be a square of side $1/N$ centered on the square grid of spacing $N^{-1}$, $Q^i{\subseteq}r_i$ be a square with the same center and area given by $ |Q^i|=\beta{\varphi}_i$, and $b_i:= \beta{\varphi}_iN^2$. Denoting with a star the quantities from [@CoOtSer], we set $T_*:=Tt/2$, $\kappa_*:=\alpha \sqrt2/T$, $L_*:=T\alpha^{-1/3}\beta^{-1/6}$, $r_i^*:=L_*r_i$, $\hat r_i^*:=L_*Q^i$, $d_0^*:=L_*/N$, $\rho_0^*:=\beta^{1/2}L_*/N$, $N_*=N$, and $ b_i^*:=(|\hat r_i^*|/|r_i^*|)\kappa_*/\sqrt2 =b_i \kappa_*/\sqrt2$.
[@CoOtSer Lem. 4.7] gives $Q_{L_*}$-periodic fields $\chi_*\in BV_\loc(\R^2\times (0,T_*);\{0,1\})$ and $B_*=(B'_*,B_3^*)\in L^2_\loc(\R^3;\R^3)$ such that $\div B_*=0$, $B_*(1-\chi_*)=0$, and, $$\chi_*(x',T_*)=\sum_i \chi_{\hat r_i^*}(x') \qquad \textrm{and} \qquad B^*_3(x',T_*)=\frac{\kappa_*}{\sqrt2} \sum_i \chi_{\hat r_i^*}(x') \hskip5mm\text{ for } x'\in Q_{L_*}\,,$$ with energy $$\label{estimstar}
\frac{1}{L_*^2} \int_{Q_{L_*}\times(0,T_*)} \kappa_* |D\chi_*|+|B_*'|^2+\chi_* \lt(B_3^*-\frac{\kappa_*}{\sqrt2}\rt)^2 dx \les
E_*^{int}:= \lt( \frac{\kappa_* \rho_0^* T_*}{\sqrt2 } + \frac{\kappa_*^2(\rho_0^*d_0^*)^2}{2T_*}\rt)\lt(\frac{N}{L_*}\rt)^2$$ and $$\frac{1}{L_*^2} \|B_3^*-\sum_i b_i^*\chi_{q_j^*}\|^2_{H^{-1/2}(Q_{L_*})}\les
E_*^{ext}:=\lt(
\kappa_*(\rho_0^*)^2+
\frac{\kappa_*^2(\rho_0^*)^3 d_0}{T_*}\rt)\lt(\frac{N}{L_*}\rt)^2\,.$$ We extend $\chi_*$ and $B_*$ to $Q_{L_*}\times(T_*,2T_*)$ by $$\chi_*(x):=\chi_*(x',T_*) \qquad \textrm{and} \qquad B_*=(0,B_3^*(x',T_*)) \hskip5mm\text{ for } x'\in Q_{L_*}\times(T_*,2T_*),$$ so that holds in $Q_{L_*}\times(0,2T_*)$. We set $$\rho_*(x):=\min\{1,\kappa_*^2d(x,\omega_*)^2\},$$ where $\omega_*=\{x: \chi_*(x)=1\}$ (here $d(\cdot,\omega_*)$ is the 3D distance, periodic in the tangential directions). Since $\omega_*$ is invariant in the $x_3-$direction inside $(T_*,2T_*)$ and since $t\gg \a^{-1}$, giving $T_*\gg \kappa_*^{-1}$, for $x_3=2T_*$, $$\min\{1,\kappa_*^2d(x,\omega_*)^2\}=\min\{1,\kappa_*^2\dist(x',\omega_*\cap\{x_3=2T_*\})^2\} \qquad \textrm{for } x=(x',2T_*).$$ Hence, $$\rho_*(x',2T_*)=\min_{i} \lt\{\min(1,\kappa_*^2\dist(x',\hat{r}^i_*)\rt\}.$$ The same computation as in the proof of [@CoOtSer Th. 4.9] leads to $$\begin{aligned}
1
\frac{1}{L_*^2} \int_{Q_{L_*}\times(0,2T_*)} |\nabla\rho_*^{1/2}|^2+|B_*'|^2+ (B_3^*-\frac{\kappa_*}{\sqrt2}(1-\rho_*))^2 dx \les
E_*^{int}\,.
\end{aligned}$$ We scale back in the tangential direction to obtain $$\rho(x',x_3):=\rho^*(L_*x',x_3)\quad \text{ and } \quad
(B',B_3)(x',x_3):=\frac{\sqrt2}{\kappa_*}\lt(\frac{B'_*}{L_*},B_3^*\rt)(L_*x',x_3)\,.$$ We have $\div B=0$ and (\[eqrandwb3bdly\]) holds. Changing variables gives $\nabla'\rho^{1/2}(x)=L_*\nabla'\rho_*^{1/2}(L_*x',x_3)$ and $\partial_3\rho^{1/2}(x)=\partial_3 \rho^{1/2}_*(L_*x',x_3)$, so that $$\begin{gathered}
{\widetilde}F^{(0,t)}_{\alpha,\beta}(\rho,B)\le
\alpha^{-4/3}\beta^{-2/3} \frac{1}{L_*^2} \int_{Q_{L_*}\times(0,2T_*)} |\nabla \rho_*^{1/2}|^2 + |B'_*|^2 +
\lt(B_3^*-\frac{\kappa_*}{\sqrt2}(1-\rho_*)\rt)^2 dx\\
\les\alpha^{-4/3}\beta^{-2/3} E_*^{int}.\end{gathered}$$ Since $E_*^{int}=\alpha^{4/3}\beta^{2/3} (Nt/2+2/(N^2t))$, this concludes the estimate for ${\widetilde}F^{(0,t)}_{\alpha,\beta}$.
We now estimate the boundary term. By the embedding of $L^\infty$ into $H^{-1/2}$ (recall that since $\sum_i {\varphi}_i =1$, $\int_{Q_1} \sum_i(b_i-\beta) \chi_{r_i} dx'=0$) we have $$\|\sum_i (b_i-\beta)\chi_{r_i}\|^2_{H^{-1/2}(Q_{1})}\les
\|b_i-\beta\|_\infty^2 \,.$$ Rescaling the boundary estimate for $B_3^*$ leads to $$\|B_3-\sum_i b_i\chi_{r_i}\|^2_{H^{-1/2}(Q_1)}\les
\frac{1}{L_*^3} \frac2{\kappa_*^2}\|B_3^*-\sum_i b_i^*\chi_{r_j^*}\|^2_{H^{-1/2}(Q_{L_*})}\les
\frac{\beta^{7/6}}{\alpha^{2/3}} + \frac{\beta^{4/3}}{\alpha^{1/3} N^2t}$$ Adding terms and using that $\sum_i \chi_{r_i}\equiv 1$, we conclude that $$\alpha^{1/3}\beta^{7/6}
\|\beta^{-1}{B_3}- 1\|_{H^{-1/2}(Q_1)}^2\les \frac{\beta^{1/3}}{\alpha^{1/3}} + \frac{\beta^{1/2}}{N^2t}
+\frac{\alpha^{1/3}}{\beta^{5/6}} \|b_i-\beta\|_\infty^2 \,.$$
Back to the full GL functional {#recoveryGL}
------------------------------
In this section, we relate the functional ${\widetilde}F_{\alpha,\beta}$, as defined in (\[eqdeftildaF\]–\[eqdeftildaFext\]), with the functional ${\widetilde}E_T$, as defined in (\[eqdefwe\]).
\[thirdupperbound\] Let $k=\alpha^{1/3}\beta^{2/3}T$ and let $\rho,B$ be $k$-quantized admissible functions for ${\widetilde}F_{\alpha,\beta}$ with $\rho B=0$. Let $\omega:=\{\rho=0\}$ and then for $z\in(-1,1)$, $\omega_{z}:=\omega\cap \{x_3=z\}$. Assume that $\omega$ is closed, that $\omega^c$ is connected and that for every $x_3\in(-1,1)$, $\omega_{x_3}^c$ is also connected. Then, there is a pair $(u,A)$, admissible for ${\widetilde}E_T$, such that $\rho= |u|^2$, $B=\nabla\times A$, and $$\label{limsupT}
{\widetilde}E_T(u,A)={\widetilde}F_{\a,\b}(\rho,B).$$
We shall construct a function $A$ on $\R^2\times (-1,1)$ so that $B=\nabla\times A$ everywhere, and then a multivalued function $\theta$ on the set $\Omega^c:=\Z^2\times\{0\}+\omega^c$ such that $\nabla\theta=\b^{-1}kA$. Here $B$ and $\rho$ are $Q_1$-periodic, but $A$ and $\theta$ not necessarily. Setting $u:=\rho^{1/2} e^{i\theta}$ we then have $\nabla_{\b^{-1}kA}u= e^{i\theta}\nabla\rho^{1/2}$, which directly implies (\[limsupT\]). Notice that thanks to the hypothesis on $\omega$, the set $\Omega^c$ is a connected open set.
We start from the construction of $A$. By [@CoOtSer Lem. 4.8] applied to $B-\beta e_3$ there is a $Q_1$-periodic potential $A_\mathrm{per}$ such that $\nabla \times A_\mathrm{per}= B- \b e_3$ and $\Div A=0$. We define $ A(x):= A_\mathrm{per}(x)+\b x_1 e_2$ so that $\nabla \times A=B$. We remark that on any open set where $B=0$ the vector field $A$ is curl-free and divergence-free, therefore harmonic, and in particular smooth. In particular, since $\Omega^c$ is open and since by the Meissner condition, $B=0$ in $\Omega^c$, $A$ is smooth in $\Omega^c$.
We now turn to the existence of $\theta$. For a fixed level $x_3$, let $h+\omega^i_{x_3}$, $h\in\Z^2$, $i=1,..., I(x_3)$ be the connected components of $(\R^2\times\{x_3\}) \cap (\Z^2+\omega_{x_3})$. Denote the flux going through $\omega^i_{x_3}$ by $$\Phi^i(x_3):=\int_{\omega^i_{x_3}} B_3\, dx'.$$ By assumption we have $$\label{condfluxi}
\b^{-1}k \Phi^i(x_3)\in 2\pi \Z \qquad \textrm{ for every $i$. }$$
Fix a smooth curve $\Gamma_0:=\{(\gamma_0(x_3),x_3) \ : \ x_3\in(-1,1)\}{\subseteq}\Omega^c$. For $x_3,y_3\in(-1,1)$, let $\Gamma_0^{x_3,y_3}:=\{(\gamma_0(t),t) \ : \ t\in(x_3,y_3)\}{\subseteq}\Gamma_0$. For $x=(\gamma_0(x_3),x_3)$, let $\theta(x):= \b^{-1}k \int_{\Gamma_0^{0,x_3}} A\cdot \tau d\H^1$. Now, for a generic $x=(x',x_3)\in \Omega^c$, let $$\theta(x):= \theta(\gamma_0(x_3),x_3)+\b^{-1}k\int_{\Gamma^x} A\cdot \tau d\H^1,$$ where $\Gamma^x$ is any (horizontal) curve in $\Omega^c\cap(\R^2\times\{x_3\})$ connecting $x$ to $(\gamma_0(x_3),x_3)$. This gives a well defined $\theta:\Omega^c\to\R/2\pi\Z$ since for every closed (horizontal) curve $\Gamma$ in $\Omega^c\cap(\R^2\times\{x_3\})$, $$\label{condfluxcurvebis}
\b^{-1}k\int_{\Gamma} A\cdot \tau d\H^1\in 2\pi \Z.$$ Indeed, this follows by Stokes’ Theorem and . Let us show that $\nabla \theta=\b^{-1}k A $ in $\Omega^c$. Let $x=(x',x_3)$ be fixed and let $\Gamma^x$ be a fixed simple smooth curve joining $x$ to $(\gamma_0(x_3),x_3)$ inside $\Omega^c\cap(\R^2\times\{x_3\})$. Since $\Omega^c$ is open, there is a simply connected neighborhood $V$ of $\Gamma^x$ such that $V{\subseteq}\Omega^c$. Let then $y=(y',y_3)\in V$. Upon shrinking $V$, we may assume that $(\gamma_0(y_3),y_3)\in V$. Let $\Gamma^y{\subseteq}V$ be a smooth curve joining $y$ to $ (\gamma_0(y_3),y_3)$ and let $\Gamma^{x,y}{\subseteq}V$ be a smooth curve joining $x$ to $y$. By definition, we have $$\theta(x)=\theta(\gamma_0(x_3),x_3)+ \b^{-1}k \int_{\Gamma^x}A\cdot \tau d\H^1, \quad \theta(y)=\theta(\gamma_0(y_3),y_3)+ \b^{-1}k\int_{\Gamma^y} A\cdot \tau d\H^1$$ and $$\theta(\gamma_0(y_3),y_3)=\theta(\gamma_0(x_3),x_3)+ \b^{-1}k\int_{\Gamma_0^{x_3,y_3}} A\cdot \tau d\H^1,$$ so that $$\theta(y)-\theta(x)= \b^{-1}k\int_{\Gamma^y} A\cdot \tau d\H^1+ \b^{-1}k\int_{\Gamma_0^{x_3,y_3}} A\cdot \tau d\H^1-\b^{-1}k\int_{\Gamma^x} A\cdot \tau d\H^1.$$ However, by Stokes Theorem (and $B=0$ in $V$), $$\int_{\Gamma^y}A\cdot \tau d\H^1+\int_{\Gamma_0^{x_3,y_3}} A\cdot \tau d\H^1-\int_{\Gamma^x} A\cdot \tau d\H^1=\int_{\Gamma^{x,y}} A\cdot \tau d\H^1,$$ so that $$\theta(y)-\theta(x)=\b^{-1}k\int_{\Gamma^{x,y}} A\cdot \tau d\H^1,$$ proving that indeed, $\nabla \theta=\b^{-1}k A$ in $\Omega^c$.
Proof of the upper bound
------------------------
We start from a construction for an $N$-regular measure and finite $R$.
\[propubregular\] Let $\mu\in \MNreg(Q_{1,1})$ for some $N\in\N$, $R\ge 3$, and assume (\[eqlbassumptcoeff\]) and (\[defKtotal\]) hold. Then, there exist sequences $u_n: Q_{1,1}\to {\mathbb{C}}$ and $A_n: Q_{1,1}\to \R^3$ such that $$\label{estimenerregular}
\limsup_{n\to +\infty} \, \wE(u_n,A_n)\le
\frac{K_R}{K_*}
I(\mu)
+ \frac{C}{N^{1/2}}(1+I(\mu)),$$ with $|u_n|$ and $\nabla\times A_n$ $Q_1-$periodic. Here $K_R$ and $K_*$ are as in Lemma \[lemmaVR\], $C$ is universal. Moreover, $$\label{estimW2regular}
\limsup_{n\to +\infty} W_2^2(\b^{-1}_n B^n_3,\mu)\les N^{-3/2}.$$
We first modify slightly $\mu$ in order to be able to use Proposition \[propboundraylayer\] for the boundary layer. Set $\varepsilon_N:=N^{-3/2}$. For $x_3\in [-1+\eps_N,1-\varepsilon_N]$ let $\hat{\mu}_{x_3}:=\mu_{x_3/(1-\varepsilon_N)}$ and for $x_3\in [-1,-1+\eps_N]\cup[1-\varepsilon_N,1]$, let $\hat{\mu}_{x_3}:=\mu_1=\mu_{-1}=N^{-2}\sum_l \delta_{X_l}$, where the $\{X_l\}_{l=1}^{N^2}$ form a regular grid with spacing $1/N$ (recall Def. \[defregular\]). We then have $$I(\hat{\mu})\le \frac{1}{1-\varepsilon_N} I(\mu)+2K_*\varepsilon_N N.$$ Moreover, since $|\hat \mu|(Q_1\times (1-\varepsilon_N,1))= \varepsilon_N$, the same on the other side, and $$W_2^2((1-\varepsilon_N)\mu,\hat{\mu}\LL(Q_1\times (-1+\varepsilon_N,1-\varepsilon_N))\les \eps_N^2,$$ so that $$W_2^2(\mu, \hat{\mu})\les \eps_N+\eps^2_N\les N^{-3/2},$$ it is enough to prove the estimates for $\hat{\mu}$ instead of $\mu$. We start by characterizing the geometry of the construction, which will not depend on $n$. The measure $\hat{\mu}$ is supported on finitely many polygonal curves, parametrized by $X_i:[a_i,b_i]\to Q_1$, which are disjoint up to the endpoints and carry a flux ${\varphi}_i>0$. The endpoints are either on the boundary of $Q_{1,1}$, or they are triple points. Those on the boundary constitute a regular grid. We let $\phimin:=\min_i{\varphi}_i$, $\phimax:=\max_i {\varphi}_i$, and $\gamma:=8\phimax/\phimin$. Since there are finitely many curves, these quantities are finite and positive. We define as before $\eta_n:=\alpha_n^{-2/3}\beta_n^{1/6}$ to be the coherence length; by (\[eqlbassumptcoeff\]) we have $\eta_n\to0$.
Let $y_j:=(Y_j,z_j)\in Q_{1,1}$ denote the internal endpoints of the curves. For $\ell>0$ sufficiently small one has that, for any $j$, the only curves which intersect $\B'_\ell(Y_j)\times(z_j-\ell,z_j+\ell)$ are those with an endpoint in $y_j$. Since $y_j$ is a triple point, there are three such curves $\Gamma_0$, $\Gamma_1$ and $\Gamma_2$, intersecting only at $y_j$. If we let $M$ be the maximal slope of all curves and $t_\ell:=\ell/(8M)$, then no curve intersects $\partial\B'_{\ell/8}(Y_j)\times (z_j-t_\ell,z_j+t_\ell)$ (this means, they all “exit” from the top and bottom faces). Without loss of generality, we can assume that $(X_0,z_j-t_\ell)=\Gamma_0\cap\lt( \B'(Y_j,\ell)\times\{z_j-t_\ell\}\rt)$, $(X_1,z_j+t_\ell)=\Gamma_1\cap \lt(\B'(Y_j,\ell)\times\{z_j+t_\ell\}\rt)$ and $(X_2,z_j+t_\ell)=\Gamma_2\cap \lt(\B'(Y_j,\ell)\times\{z_j+t_\ell\}\rt)$ (see Figure \[figtriplepoint\]). By definition of $t_\ell$, it holds $|X_0-Y_j|\le \frac{\ell}{8}$ and $|X_i-X_0|\le \frac{\ell}{4}$ for $i=1,2$. We set $\omega_j:=\B'_{\ell}(Y_j)\times (z_j-t_\ell,z_j+t_\ell)$ and let $\delta_\ell$ be the minimum distance between any two curves outside $\cup_j\omega_j$.
To explain the strategy, we first carry out a construction which ignores quantization. For sufficiently large $n$ we have $ \eta_nR+\sqrt{\phimax \beta_n/\pi}<\delta_\ell/2$ and can thus use the construction of Lemma \[lemmacurve3\] in a $\delta_\ell/2$-neighborhood of each curve (outside of $\cup_j \omega_j$), extending by $\rho_n=1$ and $B_n=0$ to the complement. In the cylinders $\omega_j$, since the geometry is fixed, for $n$ sufficiently large the conditions $\sqrt{\beta_n \phimin}\ge \eta_n$, $\sqrt{\beta_n\phimax}\ll |X_1-X_2|$ and $ \eta_n R\ll \ell$ are satisfied and we can use Lemma \[Lembranch3\]. We then have that $\rho_n=1$ outside $\B'(X_0, 3\ell/4)\times(z_j-t_\ell,z_j+t_\ell){\subseteq}\omega_j$.
In order to obtain a quantized field, we define $k_n$ as in (\[defK\]), which obeys $ k_n\in 2\pi\N$ and $k_n\to+\infty$. Let $\hat{\mu}^n$ be the $k_n$-quantized approximation of $\hat{\mu}$, as given by Lemma \[lemmaquantize\]. For sufficiently large $n$ we have $\frac12 \hat{\mu}\le \hat{\mu}^n\le 2\hat{\mu}$ and for $|x_3|\ge 1-\varepsilon_N$, $$\hat{\mu}_{x_3}^n=\sum_l{\varphi}_l^n\delta_{X_l},$$ where $|{\varphi}_l^n-1/N^2|\le C(\hat{\mu})/k_n$, ${\varphi}_l^n\le 2/N^2$. The fluxes ${\varphi}_i^n$ obey $\frac12{\varphi}_i\le {\varphi}_i^n\le 2{\varphi}_i$.
We then construct $B_n$ and $\rho_n$ using Lemma \[lemmacurve3\] and Lemma \[Lembranch3\] as discussed above. The geometry is the one determined by $\hat{\mu}$. In particular, the points $y_j$, and the constants $M$, $\phimin$, $\phimax$, $\delta_\ell$ and $t_\ell$ do not depend on $n$, and only $\delta_\ell$ and $t_\ell$ depend on $\ell$. Adding terms gives for sufficiently large $n$ (as a geometry dependent function of $\ell$ and $R$) $$\begin{aligned}
{\widetilde}F_{\alpha_n,\beta_n}^{(-1,1)}(\rho_n,B_n)\le &
\sum_i \left(1+ \frac{C}{({\varphi}_i^n \alpha_n^{4/3}\beta_n^{2/3})^{1/2}}
+ \frac{C}{{\varphi}_i^n \alpha_n^{4/3}\beta_n^{2/3}}\right)
\int_{a_i}^{b_i} \bigl( K_R\sqrt{{\varphi}_i^n}+{\varphi}_i^n |\dot{X}_i|^2\bigr) dx_3
\\
&+C(\gamma) \sum_j \sqrt{\phimax} t_\ell R^2+ \frac{\ell^2}{t_\ell}\phimax R^2,\end{aligned}$$ where the first sum runs over all curves and the second over the cylinders and where $C(\gamma)>0$ is a constant depending on $\gamma$. In the limit $n\to+\infty$ we have ${\varphi}_i^n\to{\varphi}_i$, $\alpha_n^2\beta_n\to+\infty$ and therefore, inserting the definition of $t_\ell$, $$\begin{aligned}
\limsup_{n\to+\infty} {\widetilde}F_{\alpha_n,\beta_n}^{(-1,1)}(\rho_n,B_n)\le &
\sum_i \int_{a_i}^{b_i} \bigl( K_R\sqrt{{\varphi}_i}+{\varphi}_i |\dot{X}_i|^2\bigr) dx_3
+C(\gamma) R^ 2 \ell \sum_j \left(\frac{\sqrt{\phimax}}{M} +M\phimax\right)\,.\end{aligned}$$ The sum over $j$ depends only on $\hat{\mu}$. Therefore if $\ell$ is chosen sufficiently small we have $$\limsup_{n\to +\infty} \, {\widetilde}F_{\alpha,\beta}^{(-1,1)}(\rho_n,B_n)\le \frac{K_R}{K_*} I(\mu)+\frac1{N^{1/2}}\,.$$ Moreover, thanks to and Lemma \[lemmaquantize\], we have $$\begin{aligned}
W_2^2(\beta^{-1}_n B^n_3,\hat{\mu})&\les W_2^2(\beta^{-1}_n B^n_3,\hat{\mu}^n)+W_2^2(\hat{\mu},\hat{\mu}^n)\\
&\les C(\hat{\mu})\b_n + t_\ell \, \sharp\{Y_j\}+ C(\hat{\mu})k_n^{-1}\\
&\les C(\hat{\mu})\b_n + \frac{\ell}{M} \, \sharp\{Y_j\}+ C(\hat{\mu})k_n^{-1}.\end{aligned}$$ Again, since $ \sharp\{Y_j\}/M$ only depends on $\hat\mu$, if we choose $\ell$ sufficiently small then $$\label{estimW2inside}
\limsup_{n\to+\infty} W_2^2(\beta^{-1}_n B^n_3,\hat{\mu})\le N^{-3/2}.$$ We finally address the boundary layer, focusing for definiteness on the side $x_3>0$. We first apply once more Lemma \[lemmacurveendpoints\] to each curve for $x_3\in (1-\eps_N, 1-\eps_N/2)$, so that the resulting fields $B_n$ and $\rho_n$ obey for all $x_3\in (1-\eps_N/2,1)$, $$B_n(x)=\sum_l e_3 \chi_{Q^l}(x') \text{ and }
\rho_n(x)= \min\{1,\eta_n^{-1} \dist(x',\cup_l Q^l)\},$$ where $Q^l$ are squares centered in the $N^2$ points $X_l$ with $|Q^l|= \b{\varphi}_l^n $ , and $|{\varphi}_l^n-1 /N^2|\le C(\hat{\mu})/k_n$. Then we modify $\rho_n$ and $B_n$ in the set $x_3\in (1-\eps_N/2,1)$ using Proposition \[propboundraylayer\]. This results in new fields $\hat \rho_n$, $\hat B_n$ which obey $${\widetilde}F_{\alpha_n,\beta_n}^{(1-\eps_N/2,1)}(\hat\rho_n,\hat B_n)+
\alpha_n^{1/3}\beta_n^{7/6}
\|\beta_n^{-1}{(\hat{B}_n)_3}- 1\|_{H^{-1/2}(x_3=1)}^2\les
\frac{1}{N^{1/2}}+
\frac{\beta_n^{1/3}}{\alpha_n^{1/3}} + \frac{\beta_n^{1/2}}{N^{1/2}}
+\frac{\alpha_n^{1/3}}{\beta_n^{5/6}} \frac{C(\hat\mu)N^4}{k_n^2} \,.$$ By (\[eqlbassumptcoeff\]) and (\[defK\]) all terms up to the first one tend to zero as $n\to+\infty$. Therefore $$\limsup_{n\to +\infty} \, {\widetilde}F_{\alpha_n,\beta_n}(\hat\rho_n,\hat B_n)\le \frac{K_R}{K_*}
I(\hat{\mu})
+\frac{C}{N^{1/2}}\,.$$ Finally, we pass from $(\hat\rho_n,\hat B_n)$ to $(u_n,A_n)$ via Proposition \[thirdupperbound\] and conclude the proof of . We also obtain since $$W^2_2(\beta^{-1}_n (\hat{B}_n)_3, \hat{\mu})\les W^2_2(\beta^{-1}_n B^n_3, \hat{\mu})+N^{-3/2}.$$
It only remains to combine the different steps.
By density (Theorem \[theodens\]) there is a sequence $\mu^N\in \MNreg(Q_{1,1})$ of $N$-regular measures converging weakly to $\mu$, with $\limsup_{N\to+\infty} I(\mu^N)\le I(\mu)$. Fix $R\ge 3$ and let $(u_n^N,A_n^N)$ be as in Proposition \[propubregular\]. Taking a diagonal subsequence (first with $N$, then with $R$) we obtain $$\limsup_{n\to +\infty} \, \wE(u_n^{N(n)},A_n^{N(n)})\le I(\mu),$$ and $\b_n^{-1}(B_n^{N(n)})_3\weaklim \mu $. From the compactness statement of Proposition \[gammaliminf\] and uniqueness of $m$ one obtains $\b_n^{-1} (B_n^{N(n)})'\weaklim m$ and $\b_n^{-1}(1-|u_n^{N(n)}|^2)\weaklim \mu $.
Acknowledgment {#acknowledgment .unnumbered}
==============
M. G. thanks E. Esselborn, F. Barret and P. Bella for stimulating discussions about optimal transportation, and the FMJH PGMO fundation for partial support through the project COCA. The work of S.C. was partially supported by the Deutsche Forschungsgemeinschaft through the Sonderforschungsbereich 1060 [*“The mathematics of emergent effects”*]{}.
[^1]: Institut für Angewandte Mathematik, Universität Bonn, Germany, email: [email protected]
[^2]: LJLL, Université Paris Diderot, CNRS, UMR 7598, France email: [email protected], part of his research was funded by a Von Humboldt PostDoc fellowship
[^3]: Max Planck Institute for Mathematics in the Sciences, Leipzig, Germany, email: [email protected]
[^4]: Courant Institute, NYU & Institut Universitaire de France & UPMC, Paris, France, email: [email protected]
[^5]: for $p\ge 1$, we analogously define $W_p^p(\mu,\nu):=\min\lt\{ \mu(Q_L)\int_{Q_L\times Q_L} |x-y|^p \, d\Pi(x,y) \, : \, \Pi_1=\mu, \ \Pi_2=\nu\rt\}$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Stochastic volatility modelling of financial processes has become increasingly popular. The proposed models usually contain a stationary volatility process. We will motivate and review several nonparametric methods for estimation of the density of the volatility process. Both models based on discretely sampled continuous time processes and discrete time models will be discussed.
The key insight for the analysis is a transformation of the volatility density estimation problem to a deconvolution model for which standard methods exist. Three type of nonparametric density estimators are reviewed: the Fourier-type deconvolution kernel density estimator, a wavelet deconvolution density estimator and a penalized projection estimator. The performance of these estimators will be compared.\
[*Key words:*]{} stochastic volatility models, deconvolution, density estimation, kernel estimator, wavelets, minimum contrast estimation, mixing\
[*AMS subject classification:*]{} 62G07, 62G08, 62M07, 62P20, 91G70
author:
- |
Bert van Es, Peter Spreij\
[*Korteweg-de Vries Institute for Mathematics*]{}\
[*Universiteit van Amsterdam*]{}\
[*PO Box 94248*]{}\
[*1090GE Amsterdam*]{}\
[*The Netherlands*]{}\
- |
Harry van Zanten\
[*Department of Mathematics*]{}\
[*Eindhoven University of Technology*]{}\
[*PO Box 513*]{}\
[*5600 MB Eindhoven*]{}\
[*The Netherlands*]{}
title: Nonparametric methods for volatility density estimation
---
Introduction
============
We discuss a number of nonparametric methods that come into play when one wants to estimate the density of the volatility process, given observations of the price process of some asset. The models that we treat are mainly formulated in continuous time, although we pay some separate attention to discrete time models. The observations of the continuous time models will always be in discrete time however and may occur at low frequency (fixed lag between observation instants), or high frequency (vanishing time lag). In this review, for simplicity we focus on the univariate marginal distribution of the volatility process, although similar results can be obtained for multivariate marginal distributions.
Although the underlying models differ in the sense that they are formulated either in continuous or in discrete time, in all cases the observations are given by a discrete time process. Moreover, as we shall see, the observation scheme can always (approximately) be cast as of ‘signal plus noise’ type $$Y_i = X_i + {\varepsilon}_i,$$ where $X_i$ is to be interpreted as the ‘signal’. If for fixed $i$ the random variables $X_i$ and ${\varepsilon}_i$ are independent, the distribution of the $Y_i$ is a convolution of the distributions of $X_i$ and ${\varepsilon}_i$. The density of the ‘signal’ $X_i$ is the object of interest, while the density of the ‘noise’ ${\varepsilon}_i$ is supposed to be known to the observer. The statistical problem is to recover the density of the signal by [*deconvolution*]{}. Classically for such models it was often also assumed that the processes $(X_i)$ and $({\varepsilon}_i)$ are i.i.d. Under these conditions Fan [@104] gave lower bounds for the estimation of the unknown density $f$ at a fixed point $x_0$ and showed that kernel-type estimators achieve the optimal rate. An alternative estimation method was proposed in the paper Pensky and Vidakovic [@103], using wavelet methods instead of kernel estimators and where global $L^2$-errors were considered instead of pointwise errors.
However, for the stochastic volatility models that we consider, the i.i.d. assumption on the $X_i$ is violated. Instead, the $X_i$ may be modelled as stationary random variables, that are allowed to exhibit some form of weak dependence, controlled by appropriate mixing properties, strongly mixing or $\beta$-mixing. These mixing conditions are justified by the fact that they are satisfied for many popular GARCH-type and stochastic volatility models (see e.g. Carrasco and Chen [@Car02]), as well as for continuous time models, where $\sigma^2$ is solves a stochastic differential equation, see e.g. Genon-Catalot et al. [@Gen00]. The estimators that we discuss are based on kernel methods, wavelets and penalized contrast estimation, also referred to as penalized projection estimation. We will review the performance of these deconvolution estimators under weaker than i.i.d. assumptions and show that this essentially depends on the smoothness and mixing conditions of the underlying process and the frequency of the observations. For a survey of other nonparametric statistical problems for financial data we refer to Franke et al. [@FKM09]
The paper is organized as follows. In Section \[section:ctsmodel\] we introduce the continuous time model. In Section \[section:kernel\] we consider a kernel type estimator of the invariant volatility density and apply it to a set of real data. Section \[section:wavelet\] is devoted to a wavelet density estimator and in Section \[section:ppe\] a minimum contrast estimator is discussed. Some related results for discrete time models are reviewed in Section \[section:edtm\] and Section \[section:remarks\] contains some concluding remarks.
The continuous time model {#section:ctsmodel}
=========================
Let $S$ denote the log price process of some stock in a financial market. It is often assumed that $S$ can be modelled as the solution of a stochastic differential equation or, more general, as an Itô diffusion process. So we assume that we can write $$\label{eq:sde}
{\mathrm{d}}S_t = b_t\, {\mathrm{d}}t + \sigma_t \,{\mathrm{d}}W_t, \ \ \ S_0=0,$$ or, in integral form, $$\label{eq:si}
S_t = \int_0^t b_s\, {\mathrm{d}}s + \int_0^t \sigma_s \,{\mathrm{d}}W_s,$$ where $W$ is a standard Brownian motion and the processes $b$ and $\sigma$ are assumed to satisfy certain regularity conditions (see Karatzas and Shreve [@KS91]) to have the integrals in (\[eq:si\]) well-defined. In a financial context, the process $\sigma$ is called the volatility process. One often takes the process $\sigma$ independent of the Brownian motion $W$.
Adopting this common assumption throughout the paper, unless explicitly stated otherwise, we also assume that $\sigma$ is a strictly stationary positive process satisfying a mixing condition, for example an ergodic diffusion on $(0,\infty)$. We will assume that the one-dimensional marginal distribution of $\sigma$ has an invariant density with respect to the Lebesgue measure on $(0,\infty)$. This is typically the case in virtually all stochastic volatility models that are proposed in the literature, where the evolution of $\sigma$ is modelled by a stochastic differential equation, mostly in terms of $\sigma^2$, or $\log \sigma^2$ (cf. e.g. Wiggins [@Wiggins87], Heston [@Heston93]). Often $\sigma^2_t$ is a function of a process $X_t$ satisfying a stochastic differential equation of the type $$\label{eq:sigmasde}
{\mathrm{d}}X_t = b(X_t)\, {\mathrm{d}}t + a(X_t)\, {\mathrm{d}}B_t,$$ with $B_t$ a Brownian motion. Under regularity conditions, the invariant density of $X$ is up to a multiplicative constant equal to $$\label{eq: piet} x \mapsto \frac{1}{a^2(x)}\,{\exp
\left(2\int_{x_0}^x \frac{b(y)}{a^2(y)}\,{\mathrm{d}}y\right)},$$ where $x_0$ is an arbitrary element of the state space, see e.g. Gihman and Skorohod [@GS72] or Skorokhod [@Skorohod89]. From formula (\[eq: piet\]) one sees that the invariant distribution of the volatility process (take $X$ for instance equal to $\sigma^2$ or $\log \sigma^2$) may take on many different forms, as is the case for the various models that have been proposed in the literature. In absence of parametric assumptions on the coefficients $a$ and $b$, we will investigate nonparametric procedures to estimate the corresponding densities, even refraining from an underlying model like (\[eq:sigmasde\]), partly aimed at recovering possible ‘stylized facts’ exhibited by the observations.
For instance, one could think of [*volatility clustering*]{}. This may be cast by saying that for different time instants $t_1,t_2$ that are close, the corresponding values of $\sigma_{t_1},\sigma_{t_2}$ are close again. This can partly be explained by assumed continuity of the process $\sigma$, but it might also result from specific areas around the diagonal where the multivariate density of $(\sigma_{t_1},\sigma_{t_2})$ assumes high values if $t_1$ and $t_2$ are relatively close. It is therefore conceivable that the density of $(\sigma_{t_1},\sigma_{t_2})$ has high concentrations around points $(\ell,\ell)$ and $(h,h)$, with $\ell<h$, a kind of bimodality of the joint distribution, with the interpretation that clustering occurs around a low value $\ell$ or around a high value $h$. This in turn may be reflected by bimodality of the univariate marginal distribution of $\sigma_t$.
A situation in which this naturally occurs is the following. Consider a regime switching volatility process. Assume that for $i=0,1$ we have two stationary processes $X^i$ having stationary densities $f^i$. We assume these two processes to be independent, and also independent of a two-state stationary homogeneous Markov chain $U$ with states $0,1$. The stationary distribution of $U$ is given by $\pi_i:=P(U_t=i)$. The process $\xi$ is defined by $$\xi_t=U_tX^1_t+(1-U_t)X^0_t.$$ Then $\xi$ is stationary too and it has a stationary density $f$ given by $$\begin{aligned}
f(x) & = \pi_1f^1(x) + \pi_0f^0(x).\end{aligned}$$ Suppose that the volatility process is defined by $\sigma^2_t=\exp(\xi_t)$ and that the $X^i$ are both Ornstein-Uhlenbeck processes given by $${\mathrm{d}}X^i_t = -b_i(X^i_t-\mu_i)\,{\mathrm{d}}t + a_i\,{\mathrm{d}}W^i_t,$$ with $W^1$, $W^2$ independent Brownian motions, $\mu_1\neq \mu_2$ and $b_1,b_2>0$. Suppose that the $X^i$ start in their stationary $N(\mu_i,\frac{a_i^2}{2b_i})$ distributions. Then the stationary density $f$ is a bimodal mixture of normal densities with $\mu_1$ and $\mu_2$ as the locations of the local maxima. Nonparametric procedures are able to detect such a property and are consequently by all means sensible tools to get some first insights into the shape of the invariant density.
A first object of study is the marginal univariate distribution of the stationary volatility process $\sigma$. The standing assumption in all what follows is that this distribution admits a density w.r.t. Lebesgue measure. We will also consider the invariant density of the integrated squared volatility process over an interval of length $\Delta$. By stationarity of $\sigma$ this is the density of $\int_0^\Delta \sigma_t^2\,{\mathrm{d}}t$. We will consider density estimators and assess their quality by giving results on their mean squared or integrated mean squared error. For kernel estimators, we rely on Van Es et al. [@bert03], where this problem has been studied for the marginal [*univariate*]{} density of $\sigma$. In Van Es and Spreij [@VanEsSpreij09] one can find results for multivariate density estimators. Results on wavelet estimators will be taken from Van Zanten and Zareba [@VanzantenZareba08]. Penalized contrast estimators have been treated in Comte and Genon-Catalot [@com].
The observations of log-asset price $S$ process are assumed to take place at the time instants $0, \Delta, 2\Delta, \ldots,
n\Delta$. In case one deals with low frequency observations, $\Delta$ is fixed. For high frequency observations, the time gap satisfies $\Delta = \Delta_n \to 0$ as $n \to \infty$. To obtain consistency for the estimators that we will study in the latter case, we will make the additional assumption $n\Delta_n \to
\infty$.
To explain the origin of the estimators that we consider in this paper, we often work with the simplified model, which is obtained from (\[eq:sde\]) by taking $b_t=0$. We then suppose to have discrete-time data $S_0, S_\Delta, S_{2\Delta}, \ldots$ from a continuous-time stochastic volatility model of the form $${\mathrm{d}}S_t = \sigma_t\,{\mathrm{d}}W_t.$$ Under this additional assumption, we will see that we (approximately) deal with stationary observations $Y_i$ that can be represented as $Y_i=X_i+{\varepsilon}_i$, where for each $i$ the random variables $X_i$ and ${\varepsilon}_i$ are independent.
Kernel deconvolution {#section:kernel}
====================
In this section we consider kernel deconvolution density estimators. We construct them, give expressions for bias and variance and give an application to real data.
Construction of the estimator
-----------------------------
To motivate the construction of the estimator, we first consider (\[eq:sde\]) without the drift term, so we assume to have the simplified model $$\label{eq:sdesimple}
{\mathrm{d}}S_t = \sigma_t \,{\mathrm{d}}W_t, \ \ \ S_0=0.$$ It is assumed that we observe the process $S$ at the discrete time instants $0$, $\Delta$, $2\Delta, \ldots, n\Delta$, satisfying $\Delta\to 0$, $n\Delta\to\infty$. For $i=1, 2, \ldots$ we work, as in Genon-Catalot et al. [@GCJL98; @GCJL99], with the normalized increments $$X^{\Delta}_i = \tfrac{1}{\sqrt{\Delta}}(S_{i\Delta} -
S_{(i-1)\Delta}).$$ For small $\Delta$, we have the rough approximation $$\begin{aligned}
X^{\Delta}_i & = \tfrac{1}{\sqrt{\Delta}}
\int_{(i-1)\Delta}^{i\Delta}\sigma_t\,{\mathrm{d}}W_t \nonumber \\
& \approx
\sigma_{(i-1)\Delta} \tfrac{1}{\sqrt{\Delta}}(W_{i\Delta} -
W_{(i-1)\Delta}) \label{eq:approx}\\
& = \sigma_{(i-1)\Delta}Z^\Delta_i,\nonumber\end{aligned}$$ where for $i=1, 2, \ldots$ we define $$Z^\Delta_i= \frac{1}{\sqrt{\Delta}}(W_{i\Delta} -
W_{(i-1)\Delta}).$$ By the independence and stationarity of Brownian increments, the sequence $Z^\Delta_1, Z^\Delta_2, \ldots$ is an i.i.d. sequence of standard normal random variables. Moreover, the sequence is independent of the process $\sigma$ by assumption.
Writing $Y_i=\log (X_i^{\Delta})^2$, $\xi_i=\log\sigma_{(i-1)\Delta}^2$, ${\varepsilon}_i=\log (Z_i^{\Delta})^2$ and taking the logarithm of the square of $X_i^\Delta$ we get $$\begin{aligned}
Y_i & \approx \xi_i +{\varepsilon}_i,\end{aligned}$$ where the terms in the sum are independent. Assuming that the approximation is sufficiently accurate we can use this approximate convolution structure to estimate the unknown density $f$ of $\log
\sigma_{i\Delta}^2$ from the transformed observed $Y_i=\log
(X^\Delta_i)^2$. The characteristic functions involved are denoted by $\phi_Y$, $\phi_\xi$ and $\phi_k$, where $k$ is the density of the ‘noise’ $\log (Z_i^\Delta)^2$. One obviously has $\phi_Y=\phi_\xi\phi_k$ and one easily sees that the density $k$ is given by $$k(x)= \frac{1}{\sqrt{2\pi}}\, e^{\tfrac{1}{2}x} e^{-
\tfrac{1}{2}e^{x}},$$ and its characteristic function by $$\phi_k(t)= \frac{1}{\sqrt{\pi}}\, 2^{i t} \Gamma\Big(\frac12 + it\Big).$$
The idea of getting a deconvolution estimator of $f$ is simple. Using a kernel function $w$, a bandwidth $h$, and the $Y_i$, the density $g$ of the $Y_i$ is estimated by $$g_{nh}(y)=\frac{1}{nh}\sum_jw\big(\frac{y-Y_j}{h}\big).$$ Denoting $\phi_{g,nh}$ the characteristic function of $g_{nh}$, one estimates $\phi_Y$ by $\phi_{g,nh}$ and $\phi_\xi$ by $\phi_{g,nh}/\phi_k$. Following a well-known approach in statistical deconvolution theory (see e.g. Section 6.2.4 of Wand and Jones [@WandJones95]), Fourier inversion then yields the density estimator of $f$. By elementary calculations one obtains from this procedure $$\label{eq:fnhc}
f_{nh}(x)=\frac{1}{nh}\sum_{j=1}^n
v_h\left(\frac{x-\log(X^\Delta_j)^2}{h}\right),$$ where $v_h$ is the kernel function, depending on the bandwidth $h$, $$\label{fourkernel}
v_h(x)=\frac{1}{2\pi}\infint \frac{\phi_w(s)}{\phi_k(s/h)}\
e^{-{{\rm i}}sx}\,{\mathrm{d}}s.$$ One easily verifies that the estimator $f_{nh}$, is real-valued.
To justify the approximation in (\[eq:approx\]), we quantify a stochastic continuity property of $\sigma^2$. In addition to this we make the mixing condition explicit. We impose
\[cond:sigma\] The process $\sigma^2$ satisfies the following conditions.
1. It is $L^1$-Hölder continuous of order one half, $\ex|\sigma^2_t -
\sigma_0^2| = O(t^{1/2})$ for $t \to 0$.
2. It is strongly mixing with coefficient $\alpha(t)$ satisfying, for some $0<q<1$, $$\label{eq:alpha^q}
\int_0^\infty \alpha(t)^q\,{\mathrm{d}}t<\infty.$$
The kernel function $w$ is assumed to satisfy the following conditions (an example of such a kernel is given in (\[simkernel\]) below, see also Wand [@Wand98]), which includes in particular the behavior of $\phi_{w}$ at the boundary of its domain.
\[cond:w\] Let $w$ be a real symmetric function with real valued symmetric characteristic function $\phi_w$ with support \[-1,1\]. Assume further
1. $\infint |w(u)|du < \infty$ , $\infint w(u)du=1$ , $\infint u^2|w(u)|du<\infty$ ,
2. $\phi_w(1-t)=At^\rho + o(t^\rho),\quad\mbox{as}\ t\downarrow 0 $ for some $\rho >0$, $A\in{\mathbb{R}}$.
The first part of Condition \[cond:sigma\] is motivated by the situation where $X=\sigma^2$ solves a SDE like (\[eq:sde\]). It is easily verified that for such processes it holds that $\ex |\sigma^2_t-\sigma^2_0| = O(t^{1/2})$, provided that $b \in
L_1(\mu)$ and $a \in L_{2}(\mu)$, where $\mu$ is the invariant probability measure. Indeed we have $\ex |\sigma^2_t-\sigma^2_0| \leq \ex
\int_0^t |b(\sigma^2_s)|\,ds + (\ex \int_0^t a^2(\sigma^2_s)\, ds)^{1/2} =
t||b||_{L_1(\mu)} + \sqrt{t} ||a||_{L_2(\mu)}$.
The main result we present for this estimator concerns its mean squared error at a fixed point $x$. Although the motivation of the estimator was based on the simplified model (\[eq:sdesimple\]), the result below applies to the original model (\[eq:sde\]). For its proof and additional technical details, see Van Es et al. [@bert03].
\[contasthm\] Assume that $\ex b_t^2$ is bounded. Let the process $\sigma$ satisfy Condition \[cond:sigma\], and let the kernel function $w$ satisfy Condition \[cond:w\]. Moreover, let the density $f$ of $\log\sigma^2_t$ be twice continuously differentiable with a bounded second derivative. Also assume that the density of $\sigma^2_t$ is bounded in a neighbourhood of zero. Suppose that $\Delta=n^{-\delta}$ for given $0<\delta <1$ and choose $h=\gamma\pi/\log n$, where $\gamma > 4/\delta$. Then the bias of the estimator (\[eq:fnhc\]) satisfies $$\label{contasthm:1}
\ex f_{nh}(x) - f(x) = \tfrac{1}{2}h^2f''(x)\int u^2w(u)du +
o(h^2),$$ whereas, the variance of the estimator satisfies the order bounds $$\label{contasthm:2}
\var f_{nh}(x) =
O\Big(\frac{1}{n}\,h^{2\rho}e^{\pi/h}\Big)+O\Big(\frac{1}{{nh^{1+q}\Delta}}\Big).$$
\[remark:mse\] The choices $\Delta=n^{-\delta}$, with $0<\delta<1$ and $h=\gamma\pi/\log n$, with $\gamma>4/\delta$ render a variance that is of order $n^{-1+1/\gamma}(1/\log n)^{2\rho}$ for the first term of (\[contasthm:2\]) and $n^{-1+\delta}(\log n)^{1+q}$ for the second term. Since by assumption $\gamma>4/\delta$ we have $1/\gamma<\delta/4<\delta$ so the second term dominates the first term. The order of the variance is thus $n^{-1+\delta}(\log
n)^{1+q}$. Of course, the order of the bias is logarithmic, hence the bias dominates the variance and the mean squared error of $f_{nh}(x)$ is of order $(\log n)^{-4}$.
It can then be shown that for the characteristic function $\phi_k$ one has the behavior $$|\phi_k(s)| = \sqrt{2}\,e^{-\frac{1}{2}\pi
|s|}(1+O(\tfrac{1}{|s|})),\,|s|\to\infty.$$ This means that $k$ is supersmooth in the terminology of Fan [@104] which explains the slow logarithmic rate at which the bias vanishes. Sharper results on the variance can be obtained when $\sigma^2$ is strongly mixing, see Van Es et al. [@bert04] for further details. The orders of the bias and of the MSE remain unchanged though.
An application to the Amsterdam AEX index
-----------------------------------------
In this section we present an example using real data of the Amsterdam AEX stock exchange. We have estimated the volatility density from 2600 daily closing values of the Amsterdam stock exchange index AEX from 12/03/1990 until 14/03/2000. These data are represented in Figure \[fig:10\]. We have centered the daily log returns, i.e we have subtracted the mean (which equaled 0.000636), see Figure \[fig:11\]. The deconvolution estimator is given as the left hand picture in Figure \[fig:13\]. Observe that the estimator strongly indicates that the underlying density is unimodal. Based on computations of the mean and variance of the estimate, with $h=0.7$, we have also fitted a normal density by hand and compared it to the kernel deconvolution estimator. The result is given as the right hand picture in Figure \[fig:13\]. The resemblance is remarkable.
![AEX. Left: daily closing values. Right: log of the daily closing values. \[fig:10\]](aexclose.pdf "fig:") ![AEX. Left: daily closing values. Right: log of the daily closing values. \[fig:10\]](logaexclose.pdf "fig:")
![AEX. Left: the values of $X_t$, i.e. the centered daily log returns. Right: $\log(X_t^2)$ . \[fig:11\]](meancorr-loglogreturns.pdf "fig:") ![AEX. Left: the values of $X_t$, i.e. the centered daily log returns. Right: $\log(X_t^2)$ . \[fig:11\]](meancorr-loglogreturnssq.pdf "fig:")
![AEX. Left: The estimate of the density of $\log(\sigma_t^2)$ with $h=0.7$. Right: The normal fit to the $\log(\sigma_t^2)$. The dashed line is the normal density and the solid line the kernel estimate. \[fig:13\]](meancorr-aex-h07.pdf "fig:") ![AEX. Left: The estimate of the density of $\log(\sigma_t^2)$ with $h=0.7$. Right: The normal fit to the $\log(\sigma_t^2)$. The dashed line is the normal density and the solid line the kernel estimate. \[fig:13\]](meancorr-normalfit.pdf "fig:")
The kernel used to compute the estimates is a kernel from Wand [@Wand98], with $\rho=3$ and $A=8$, $$\label{simkernel}
w(x)={\frac{48x(x^2-15)\cos x - 144(2x^2-5)\sin x}{\pi x^7}}.$$ It has characteristic function $$\label{simchar}
\phi_w(t)=(1-t^2)^3,\quad |t|\leq 1.$$ The bandwidths are chosen by hand. The estimates have been computed by fast Fourier transforms using the Mathematica 4.2 package.
This is actually the same example as in our paper Van Es et al. [@bert04] on volatility density estimation for discrete time models. The estimator (\[eq:fnhc\]) presented here is, as a function of the sampled data, exactly the same as the one for the discrete time models. The difference lies in the choice of underlying model. In the present paper the model is a discretely sampled continuous time process, while in Van Es et al. [@bert04] it is a discrete time process. For the latter type of models the discretization step in the beginning of this section is not necessary since these models satisfy an exact convolution structure.
Wavelet deconvolution {#section:wavelet}
=====================
As an alternative to kernel methods, in this section we consider estimators based on wavelets. Starting point is again the simplified model (\[eq:sdesimple\]). Contrary to the previous section, we are now interested in estimating the accumulated squared volatility over an interval of length $\Delta$. We assume having observations of $S$ at times $i\Delta$ to our disposal, but now with $\Delta$ fixed (low frequency observations). Let, as before, $X^\Delta_i=\Delta^{-1/2}(S_{i\Delta} - S_{(i-1)\Delta})$ and let $\bar{\sigma}_i^2=\Delta^{-1}\int_{(i-1)\Delta}^{i\Delta}\sigma^2_t\,{\mathrm{d}}t$. Denote by $\mathcal{F}_\sigma$ the $\sigma$-algebra generated by the process $\sigma$. By the assumed independence of the processes $\sigma$ and $W$, we have for the characteristic function of $X^\Delta_i$ given $\mathcal{F}_\sigma$ $${\mathbb{E}}[\exp({{\rm i}}sX^\Delta_i)| \mathcal{F}_\sigma]=\exp(-{\frac{1}{2}}\bar{\sigma}_i^2s^2).$$ Consider also the model $\tilde X^{\Delta}_i=\bar{\sigma_i}Z_i$, with $\bar{\sigma_i}$ and $Z_i$ independent for each $i$ and $Z_i$ a standard Gaussian random variable. Then $${\mathbb{E}}[\exp({{\rm i}}s\tilde X^{\Delta}_i)| \mathcal{F}_{\sigma_i}]=\exp(-{\frac{1}{2}}\bar{\sigma}_i^2s^2).$$ It follows that $X^\Delta_i$ and $\tilde X^{\Delta}_i$ are identically distributed. From this observation we conclude that the transformed increments $\log\big(\Delta^{-1}(S_{i\Delta} - S_{(i-1)\Delta})^2\big)$ are then distributed as $Y_i = \xi_i + {\varepsilon}_i$, where $$\xi_i = \log \bar{\sigma}^2_i, \ \ \
{\varepsilon}_i = \log Z_i^2,$$ and $Z_i$ is an i.i.d. sequence of standard Gaussian random variables, independent of $\sigma$. The sequence $\xi_i$ is stationary and we assume that its marginal density $g$ exists, i.e. $g$ is the density of $\log \big(\Delta^{-1}\int_0^\Delta \sigma^2_u\,{\mathrm{d}}u\big)$. The density of the ${\varepsilon}_i$ is again denoted by $k$. Of course, estimating $g$ is equivalent to estimating the density of the aggregated squared volatility $\int_0^\Delta
\sigma^2_u\,du$.
In the present section the main focus is on the quality of the estimator in terms of the mean integrated squared error, as opposed to establishing results for the (pointwise) mean squared error as in Section \[section:kernel\]. At the end of this section we compare the results presented here to those of Section \[section:kernel\].
First we recall the construction of the wavelet estimator proposed in Pensky and Vidakovic [@103]. For the necessary background on wavelet theory, see for instance Blatter [@Bla98], Jawerth and Sweldens [@102], and the references therein. For the construction of deconvolution estimators we need to use band-limited wavelets. As in Pensky and Vidakovic [@103] we use a Meyer-type wavelet (see also Walter [@107], Walter and Zayed [@108]). We consider an orthogonal scaling function and wavelet $\varphi$ and $\psi$, respectively, associated with an orthogonal multiresolution analysis of $L^2({\mathbb{R}})$. We denote in this section the Fourier transform of a function $f$ by $\tilde f$, i.e.$$\tilde f ( \omega )= \int_{\mathbb{R}}e^{-{{\rm i}}\omega x} f(x)\,{\mathrm{d}}x,$$ and suppose that for a symmetric probability measure $\mu$ with support contained in $\left[ - \pi/3, \pi/3 \right]$ it holds that $$\tilde \varphi (\omega) = \Big(\mu(\omega-\pi, \omega+\pi] \Big)^{1/2}, \ \
\tilde \psi (\omega) = e^{-i\omega/2} \Big(\mu(|\omega|/2-\pi, |\omega| -\pi]\Big)^{{1}/{2}}.$$ Observe that the assumptions imply that $\varphi$ and $\psi$ are indeed band-limited. For the supports of their Fourier transforms we have $\text{supp}\,\tilde\varphi \subset \left[ -4\pi/3,4\pi/3\right]$ and $\text{supp}\,\tilde\psi \subset \left[-8\pi/3,-2\pi/3\right] \cup
\left[ 2\pi/3,8\pi/3\right]$. By choosing $\mu$ smooth enough we ensure that $\tilde\varphi$ and $\tilde\psi$ are at least twice continuously differentiable.
For any integer $m$, the unknown density $g$ can now be written as $$\label{213610}
g(x)=\sum_{l\in {\mathbb Z}} a_{m,l}\varphi_{m,l}(x) +
\sum_{l\in {\mathbb Z}} \sum_{j=m}^{\infty} b_{j,l}\psi_{j,l}(x),$$ where $\varphi_{m,l}(x) = 2^{m/2}\varphi (2^m x-l)$, $\psi_{j,l}(x) = 2^{j/2}\psi (2^jx-l)$ and the coefficients are given by $$a_{m,l}=\int_{{\mathbb{R}}}\varphi_{m,l}(x)g(x)\, {\mathrm{d}}x \, ,\quad \quad
b_{j,l}=\int_{\mathbb{R}}\psi_{j,l}(x)g(x)\,{\mathrm{d}}x.$$ The idea behind the linear wavelet estimator is simple. We first approximate $g$ by the orthogonal projection given by the first term on the right-hand side of (\[213610\]). For $m$ large enough the second term will be small, and can be controlled by using the approximation properties of the specific family of wavelets that is being used. The projection of $g$ is estimated by replacing the coefficients $a_{m,l}$ by consistent estimators and truncating the sum. Using the fact that the density $p$ of an observation $Y_i$ is the convolution of $g$ and $k$ it is easily verified that $$a_{m,l}=\int_{\mathbb{R}}2^{m/2}U_m(2^mx-l)p(x)\,{\mathrm{d}}x = 2^{m/2}{\mathbb{E}}U_m(2^mY_i-l),$$ where $U_m$ is the function with Fourier transform $$\label{eq: um}
\tilde U_m (\omega) = \frac{\tilde \varphi (\omega) }{\tilde k(-2^m \omega)}.$$ We estimate the coefficient $a_{m,l}$ by its empirical counterpart $$\hat a_{m,l,n}=\frac{1}{n} \sum_{i=1}^{n} 2^{{m}/{2}} U_m(2^mY_i-l).$$ Under the mixing assumptions that we will impose on the sequence $Y$, it will be stationary and ergodic. Hence, by the ergodic theorem, $\hat a_{m,l,n}$ is a consistent estimator for $a_{m,l}$. The wavelet estimator is now defined by $$\label{213619}
\hat g_{n}(x)= \sum_{\left|l\right| \le L_n} \hat a_{m_n,l,n} \varphi_{m_n,l}(x),$$ where the detail level $m_n$ and the truncation point $L_n$ will be chosen appropriately later.
The main results in the present section are upper bounds for the mean integrated square error of the wavelet estimator $\hat
g_{n}$, which is defined as usual by $$\text{MISE}\left( \hat g_{n} \right)={\mathbb{E}}\int_{{\mathbb{R}}} \left( \hat
g_{n} (x) - g(x) \right)^2\, {\mathrm{d}}x.$$ We will specify how to choose the detail level $m_n$ and the truncation point $L_n$ in (\[213619\]) optimally in different cases, depending on the smoothness of $g$ and $k$. The smoothness properties of $g$ are described in terms of $g$ belonging to certain Sobolev balls and by imposing a weak condition on its decay rate. The Sobolev space $H^\alpha$ is defined for $\alpha > 0$ by $$\label{eq:sobolev}
H^{\alpha}= \Big\{g : \;\, \|g\|_{\alpha}=\Big( \int_{\mathbb{R}}|\tilde
g(\omega)|^2(\omega^2+1)^{\alpha}\,{\mathrm{d}}\omega \Big)^{1/2} \! <\infty \Big\}.$$ Roughly speaking, $g \in H^\alpha$ means that the first $\alpha$ derivatives of $g$ belong to $L^2({\mathbb{R}})$. The Sobolev ball of radius $A$ is defined by $$\mathscr{S}_{\alpha}(A)=\left\{ g \in
H^{\alpha}: \, \|g\|_{\alpha} \le A\right\}.$$ The additional assumption on the decay rate is reflected by $g$ belonging to $$\mathscr{S}^*_{\alpha}(A, A')=\mathscr{S}_{\alpha}(A) \cap
\Big\{g: \sup_x |xg(x)| \le A'\Big\}.$$ We now have the following result, see Van Zanten and Zareba [@VanzantenZareba08], for the wavelet density estimator $\hat g_n$ of $g$ defined by (\[213619\]).
\[thm: sv\] Suppose that the volatility process $\sigma^2$ is strongly mixing with mixing coefficients satisfying $$\label{eq: karel}
\sum_{k \ge 0} \alpha_{k\Delta}^p < \infty$$ for some $p \in (0,1)$. Then with the choices $$2^{m_n} = \frac{\log n}{1+(4\pi^2/3)}, \ \ \
L_n = \Big({\log n}\Big)^{r}, \ \ \ r \ge 1+2\alpha$$ the mean square error of the wavelet estimator satisfies $$\sup_{g \in \mathscr{S}^*_{\alpha}(A, A')} \mathrm{MISE} \left( \hat g_{n} \right)
= O\Big({(\log n)^{-2\alpha}}\Big)$$ for $\alpha, A, A' > 0$. If (\[eq: karel\]) is satisfied for [*all*]{} $p \in (0,1)$, the same bound is true if the choice for $L_n$ is replaced by $L_n = n$.
Let us point out the relation with the results of Section \[section:kernel\] and with those in Van Es et al. [@bert04], see also Section \[section:dtm\]. In that paper kernel-type deconvolution estimators for discrete time stochastic volatility models were considered. When applied to the present model the results say that under the same mixing condition and assuming that $g$ has two bounded and continuous derivatives, the (pointwise) mean squared error of the kernel estimator is of order $(\log n)^{-4}$. The analogue of $g$ having two bounded derivatives in our setting is that $g \in \mathscr{S}^*_{2}(A,
A')$ for some $A,A' > 0$. Indeed, the theorem yields the same bound $(\log n)^{-4}$ for the MISE in this case. The same bound is valid for the MSE when estimating the marginal density for continuous time models, see Theorem \[contasthm\] and its consequences in Remark \[remark:mse\]. Theorem \[thm: sv\] is more general, because the smoothness level is not fixed at $\alpha=2$, but allows for different smoothness levels of order $\alpha\neq 2$ as well. Moreover, the wavelet estimator is adaptive in the sense that it does not depend on the unknown smoothness level, if the condition on the mixing coefficients holds for all $p \in
(0,1)$.
Penalized projection estimators {#section:ppe}
===============================
The results of the preceding sections assume that the true (integrated) volatility density has a finite degree of regularity, either in Hölder or in Sobolev sense. Under this assumption the nonparametric estimators have logarithmic convergence rates, cf.Remark \[remark:mse\] and Theorem \[thm: sv\]. Although admittedly slow, the minimax results of Fan [@104] show that these rates are in fact optimal in this setting. In the paper Pensky and Vidakovic [@103] it was shown however that if in a deconvolution setting the density of the unobserved variables has the same degree of smoothness as the noise density, the rates can be significantly improved, cf. also the lower bounds obtained in Butucea [@but1] and Butucea and Tsybakov [@but2]. This observation forms the starting point of the paper Comte and Genon-Catalot [@com], in which a nonparametric volatility density estimator is developed that achieves better rates than logarithmic if the true density is super smooth.
In the latter paper it is assumed that there are observations $S_\Delta, S_{2\Delta}, \ldots, S_{n\Delta}$ of a process $S$ satisfying the simple equation (\[eq:sdesimple\]), with $V = \sigma^2$ a $(0,\infty)$-valued process independent of the Brownian motion $W$. It is assumed that we deal with high frequency observations, $\Delta\to 0$ and $n\Delta\to\infty$. We impose the following condition on $V$.
\[cond:v\] The process $V$ is a time-homogenous, continuous Markov process, strictly stationary and ergodic. It is either $\beta$-mixing with coefficient $\beta(t)$ satisfying $$\int_0^\infty \beta(t)\,{\mathrm{d}}t < \infty,$$ or is $\rho$-mixing. Moreover, it satisfies the Lipschitz condition $${\mathbb{E}}\Big( \log\Big(\frac1\Delta
\int_0^\Delta V_t\,{\mathrm{d}}t\Big) - \log V_0\Big)^2 \le C\Delta,$$ for some $C > 0$.
In addition to this a technical assumption is necessary on the density $f$ of $\log V_0$ we are interested in and on the density $g_\Delta$ of $\log\Big(\frac1\Delta \int_0^\Delta
V_t\,{\mathrm{d}}t\Big)$, which is assumed to exist. Contrary to the notation of the previous section, we write $g_\Delta$ instead of $g$, since now $\Delta$ is not fixed.
\[cond:t\] The invariant density $f$ is bounded and has a second moment and $g_\Delta \in L^2({\mathbb{R}})$.
As a first step in the construction of the final estimator a preliminary estimator $\hat f_L$ is constructed for $L \in {\mathbb{N}}$ fixed. Note that Condition \[cond:t\] implies that $f \in
L^2({\mathbb{R}})$, hence we can consider its orthogonal projection $f_L$ on the subspace $S_L$ of $L^2({\mathbb{R}})$, defined as the space of functions whose Fourier transform is supported on the compact interval $[-\pi L, \pi L]$. An orthonormal basis for the latter space is formed by the Shannon basis functions $\psi_{L,j}(x) =
\sqrt{L}\psi(Lx-j)$, $j \in {\mathbb{Z}}$, with $\psi(x) = \sin(\pi
x)/(\pi x)$ the sinc kernel. For integers $K_n \to \infty$ to be specified below, the space $S_L$ is approximated by the finite-dimensional spaces $S^n_L = {\rm span}\{\psi_{L, j}: |j|
\le K_n\} $. The function $f_L$ is estimated by $\hat f_L =
\mathop{\rm argmin}_{h \in S^n_L}\gamma_n(h)$, where the contrast function $\gamma_n$ is defined for $h\in L^2({\mathbb{R}})\cap L^1({\mathbb{R}})$ by $$\gamma_n(h) = \|h\|^2_2 - \frac2n\sum_{i=1}^n u_h(\log
(X^\Delta_i)^2), \qquad u_h(x) = \frac1{2\pi}\int_{-\infty}^\infty
e^{{{\rm i}}xs}\frac{\tilde h(-s)}{\phi_k(s)}\,{\mathrm{d}}s.$$ Here, as before, $\phi_k$ is the characteristic function of $\log{\varepsilon}^2$, with ${\varepsilon}$ standard normal and $\tilde{h}$ is the Fourier transform of $h$. It is easily seen that $$\hat f_L = \sum_{|j| \le K_n} \hat a_{L,j}\psi_{L,j}, \qquad \hat
a_{L,j} = \frac1n\sum_{j=1}^n u_{\psi_{L,j}}(\log (X^\Delta_i)^2).$$ Straightforward computations show that, with ${\left< \cdot,\cdot \right>}$ the $L^2({\mathbb{R}})$ inner product, ${\mathbb{E}}u_h(\log (X^\Delta_i)^2) =
{\left< h,g_\Delta \right>}$, and hence ${\mathbb{E}}\gamma_n(h) = \|h-g_\Delta\|^2_2
- \|g_\Delta\|^2_2$. So in fact, $\hat f_L$ is an estimator of the element of $S^n_L$ which is closest to $g_\Delta$. Since $S^n_L$ approximates $S_L$ for large $n$ and $g_\Delta$ is close to $f$ for small $\Delta$, the latter element should be close to $f_L$.
Under Conditions \[cond:v\] and \[cond:t\], a bound for the mean integrated square error, or quadratic risk MISE$(\hat f_L) =
{\mathbb{E}}\|\hat f_L - f\|^2_2$ can be derived, depending on the approximation error $\|f-f_L\|_2$, the bandwidth $L$ and the truncation point $K_n$, see Comte and Genon-Catalot [@com], Theorem 1. The result implies that if $f$ belongs to the Sobolev space $H^\alpha$ as defined in (\[eq:sobolev\]), then the choices $K_n = n$ and $L
= L_n \sim \log n$ yield a MISE of the order $(\log n
)^{-2\alpha}$, provided that $\Delta = \Delta_n = n^{-\delta}$ for some $\delta \in (0,1)$. Not surprisingly, this is completely analogous to the result obtained in Theorem \[thm: sv\] for the wavelet-based estimator in the fixed $\Delta$ setting. In particular the procedure is adaptive, in that the estimator does not depend on the unknown regularity parameter $\alpha$.
To obtain faster than logarithmic rates and adaptation in the case that $f$ is supersmooth, a data-driven choice of the bandwidth $L$ is proposed. Define $$\hat L = \mathop{\rm argmin}_{L \in \{1, \ldots, \log n\}}
\Big(\gamma_n(\hat f_L) + {\rm pen}_n(L)\Big),$$ where the penalty term is given by $${\rm pen}_n(L) = \kappa \frac{(1+L)\Phi_k(L)}{n},$$ for $\kappa > 0$ a calibration constant and $$\Phi_k(L) = \int_{-\pi L}^{\pi L}\frac{1}{|\phi_k(s)|^2}\,ds.$$ For the quadratic risk of the estimator $\hat f_{\hat L}$, the following result holds (Comte and Genon-Catalot [@com]).
Under Conditions \[cond:t\] and \[cond:v\] we have $$\begin{aligned}
{\rm MISE}(\hat f_{\hat L}) & \le C_1 \inf_{L \in \{1, \ldots, \log n\}} \Big(\|f-f_L\|^2_2 + \frac{(1+L)\Phi_k(L)}{n}\Big) \\
& \quad + C_2\frac{\log^2 n}{K_n} + C_3\frac{\log n}{n\Delta} +
C_4\Delta \log^3 n,\end{aligned}$$ for constants $C_1, C_2, C_3, C_4 > 0$.
It can be seen that this bound is worse than the corresponding bound for the estimator $\hat f_L$ by a factor of the order $L$. This is at worst a logarithmic factor which, as usual in this kind of setting, has to be paid for achieving adaptation. The examples in Section 6 of Comte and Genon-Catalot [@com] show that indeed, the estimator $\hat f_{\hat L}$ can achieve algebraic convergence rates in case the true density $f$ is supersmooth.
Estimation for discrete time models {#section:edtm}
===================================
Although the main focus of the present paper is on estimation procedures for continuous time models, in the present section we also highlight some analogous results for discrete time models. These deal with both density and regression function estimation.
Discrete time models {#section:dtm}
--------------------
The discrete analogue of (\[eq:sdesimple\]) is $$\label{eq:s}
X_t = \sigma_t Z_t,\,t=1,2,\ldots .$$ Here we denote by $X$ the detrended or demeaned log-return process. Stochastic volatility models are often described in this form. The sequence $Z$ is typically an i.i.d. noise (e.g. Gaussian) and at each time $t$ the random variables $\sigma_t$ and $Z_t$ are independent. See the survey papers by Ghysels et al. [@GHR96] or Shephard [@Shephard96]. Also in this section we assume that the process $\sigma$ is strictly stationary and that the marginal distribution of $\sigma$ has a density with respect to the Lebesgue measure on $(0,\infty)$. We present some results for a nonparametric estimator of the density of $\log \sigma^2_t$, as well as results for a nonparametric estimator of a nonlinear regression function, in case $\sigma^2$ is given by a nonlinear autoregression. The standing assumption in all what follows is that for each $t$ the random variables $\sigma_t$ and $Z_t$ are independent, the noise sequence is standard Gaussian and $\sigma$ is a strictly stationary, positive process satisfying a certain mixing condition.
In principle one can distinguish two classes of models. The way in which the bivariate process $(\sigma,Z)$, in particular its dependence structure, is further modelled offers different possibilities. In the first class of models one assumes that the process $\sigma$ is predictable with respect to the filtration ${\mathcal{F}}_t$ generated by the process $Z$, and obtains that $\sigma_t$ is independent of $Z_t$ for each fixed time $t$. We furthermore have that (assuming that the unconditional variances are finite) $\sigma^2_t$ is equal to the conditional variance of $X_t$ given $\mathcal{F}_{t-1}$. This class of models has become quite popular in the econometrics literature. It is well known that this class also contains the (parametric) family of GARCH-models, introduced by Bollerslev [@Bollerslev86].
In the second class of models one assumes that the whole process $\sigma$ is independent of the noise process $Z$, and one commonly refers to the resulting model as a stochastic volatility model. In this case, the natural underlying filtration ${\mathcal{F}}=\{\mathcal{F}_t\}_{t\geq 0}$ is generated by the two processes $Z$ and $\sigma$ in the following way. For each $t$ the $\sigma$-algebra $\mathcal{F}_t$ is generated by $Z_s$, $s\leq t$ and $\sigma_s$, $s\leq t+1$. This choice of the filtration enforces $\sigma$ to be predictable. As in the first model the process $X$ becomes a martingale difference sequence and we have again (assuming that the unconditional variances are finite) that $\sigma^2_t$ is the conditional variance of $X_t$ given $\mathcal{F}_{t-1}$. An example of such a model is given in De Vries [@DeVries98], where $\sigma$ is generated as an AR(1) process with $\alpha$-stable noise ($\alpha\in(0,1)$).
As in the previous sections we refrain from parametric modelling and review some completely nonparametric approaches. We will mainly focus on results for the second class, as it is the discrete time analogue of the stochastic volatility models of the previous sections. At the heart of all what follows is again the convolution structure that is obtained from (\[eq:s\]) by squaring and taking logarithms, $$\log X_t^2 = \log \sigma_t^2 + \log Z_t^2.$$
Density estimation
------------------
The main result of this section gives a bias expansion and a variance bound of a kernel density type estimator of the density $f$ of $\log\sigma^2_t$, which chosen to be, analogously to (\[eq:fnhc\]), $$\label{eq:fnhd}
f_{nh}(x)=\frac{1}{nh}\sum_{j=1}^n v_h\left(\frac{x-\log
(X_j)^2}{h}\right),$$ where $v_h$ is the kernel function of (\[fourkernel\]).
The next theorem is derived from Van Es et al. [@bert04], where a multivariate density estimator is considered. It establishes the expansion of the bias and an order bound on the variance of our estimator under a strong mixing condition. Under broad conditions this mixing condition is satisfied if the process $\sigma$ Markov, since then convergence of the mixing coefficients to zero takes place at an [*exponential rate*]{}, see Theorems 4.2 and Theorem 4.3 of Bradley [@Bradley86] for precise statements. Similar behaviour occurs for ARMA processes with absolutely continuous distributions of the noise terms (Bradley [@Bradley86], Example 6.1).
\[discrasthmp\] Assume that the process $\sigma$ is strongly mixing with coefficient $\alpha_k$ satisfying $$\sum_{j=1}^\infty
\alpha_j^{\beta}<\infty,$$ for some $\beta\in (0,1)$. Let the kernel function $w$ satisfy Condition \[cond:w\] and let the density $f$ of $\log\sigma^2_t$ be bounded and twice continuously differentiable with bounded second order partial derivatives. Assume furthermore that $\sigma$ and $Z$ are independent processes. Then we have for the estimator of $f$ defined as in (\[eq:fnhd\]) and $h\to 0$ $$\label{discrasthm:1p}
\ex f_{nh}(x)= f(x)+\tfrac{1}{2}h^2f''(x)\int u^2 \, w(u)\,{\mathrm{d}}u+o(h^2)$$ and $$\label{discrasthm:3p}
\var f_{nh}(x) = O\big(\tfrac{1}{n}\,h^{2\rho}\,e^{\pi/h}\big).$$
Comparing the above results to the ones in Theorem \[contasthm\], we observe that in the continuous time case, the variance has an additional $O\Big(\frac{1}{{nh^{1+q}\Delta}}\Big)$ term.
Regression function estimation
------------------------------
In this section we assume the basic model (\[eq:s\]), but in addition we assume that the process $\sigma$ satisfies a nonlinear autoregression and we consider nonparametric estimation of the regression function as proposed in Franke et al. [@FHK03]. In that paper a discrete time model was proposed as a discretization of the continuous time model given by (\[eq:sde\]). In fact, Franke et al. include a mean parameter $\mu$, but since they assume it to be known, without loss of generality we can still assume (\[eq:s\]). Assume that the volatility process is strictly positive and consider $\log
\sigma_{t}^2$. It is assumed that its evolution is governed by $$\label{eq:ar}
\log\sigma_{t+1}^2=m(\log\sigma_{t}^2)+\eta_t,$$ where the $\eta_t$ are i.i.d. Gaussian random variables with zero mean. The regression function $m$ is assumed to satisfy the stability condition $$\label{eq:mstable}
\limsup_{|x|\to\infty}|\frac{m(x)}{x}|<1.$$ Under this condition the process $\sigma$ is exponentially ergodic and strongly mixing, see Doukhan [@Doukhan94] and these properties carry over to the process $X$ as well. Moreover, the process $\log \sigma^2_t$ admits an invariant density $f$.
Denoting $Y_t=\log X_t^2$, we have $$Y_t=\log\sigma_{t}^2+\log Z_t^2.$$ It is common to assume that the processes $Z$ and $\eta$ are independent, the second class of models described in Section \[section:dtm\], but dependence between $\eta_t$ and $Z_t$ for fixed $t$ can be allowed for (first model class) without changing in what follows, see Franke et al. [@FHK03].
The purpose of the present section is to estimate the function $m$ in (\[eq:ar\]). To that end we use the estimator $f_{nh}$ as defined in (\[eq:fnhd\]). Since this estimator resembles an ordinary kernel density estimator, the important difference being that the kernel function $v_h$ now depends on the bandwidth $h$, the idea is to mimic the classical Nadaraya-Watson regression estimator similarly, in order to obtain an estimator of $m(x)$. Doing so, one obtains the estimator $$\label{eq:estim}
m_{nh}(x)= \frac{\frac{1}{nh}\sum_{j=1}^n v_h\big(\frac{x-Y_j}{h}\big)Y_{j+1}}{f_{nh}(x)}.$$ It follows that $$m_{nh}(x)-m(x)=\frac{p_{nh}(x)}{f_{nh}(x)},$$ where $$p_{nh}(x)=\frac{1}{nh}\sum_{j=1}^n v_h\big(\frac{x-Y_j}{h}\big)(Y_{j+1}-m(x)).$$ In Franke et al. [@FHK03] bias expansions for $p_{nh}(x)$ and $f_{nh}$ are given that fully correspond to those in Theorem \[discrasthmp\]. They are again of order $h^2$, under similar assumptions. It is also shown that the variances of $p_{nh}$ and $f_{nh}$ tend to zero. The main result concerning the asymptotic behavior then follows from combining the asymptotics for $p_{nh}$ and $f_{nh}$.
Assume that $m$ satisfies the stability condition (\[eq:mstable\]), that $m$ and $f$ are twice differentiable and the first of Condition \[cond:w\] on the kernel $w$. The estimator $m_{nh}(x)$ satisfies $(\log
n)^2(m_{nh}(x)-m(x))=O_p(1)$ if $h=\gamma/\log n$ with $\gamma>\pi$.
Following the proofs in Franke et al. [@FHK03], one can conclude that e.g. the variance of $p_{nh}$ is of order $O(\frac{\exp(\pi/h)}{nh^4})$, which tends to zero for $h=\gamma/\log n$, with $\gamma>\pi$. For the variance of $f_{nh}$ a similar bound holds. Comparing these order bounds to the ones in Theorem \[discrasthmp\], we see that the latter ones are sharper. This is partly due to the fact that Franke et al. [@FHK03], don’t impose conditions on the boundary behavior of the function $\phi_w$ (the second of Condition \[cond:w\]), whereas their other assumptions are the same as in Theorem \[discrasthmp\].
Concluding remarks {#section:remarks}
==================
In recent years, many different parametric stochastic volatility models have been proposed in the literature. To investigate which of these models are best supported by observed asset price data, nonparametric methods can be useful. In this paper we reviewed a number of such methods that have recently been proposed. The overview shows that ideas from deconvolution theory can be instrumental in dealing with this statistical problem and that both for high and for low frequency data, methods are now available for nonparametric estimation of the (integrated) volatility density at optimal convergence rates.
On a critical note, the methods available so far all assume that the volatility process is independent of the Brownian motion driving the asset price dynamics. This is a limitation, since in several interesting models non-zero correlations are assumed between the Brownian motions driving the volatility dynamics and the asset price dynamics.
[99]{}
Blatter, C. (1998). [*Wavelets, a primer*]{}. A. K. Peters Ltd.
Bollerslev, T. (1986), Generalized autoregressive conditional heteroscedasticity, [*J. Econometrics*]{} [**31**]{}, 307–321.
Bradley, R.C. (1986), Basic properties of strong mixing conditions, in [*Dependence in Probability and Statistics, E.Eberlein and M.S. Taqqu Eds.*]{}, Birkhaüser.
Butucea, C. (2004). Deconvolution of supersmooth densities with smooth noise. [*Can. J. Statist.*]{} [**32**]{}, 181–192.
Butucea, C. and Tsybakov, A.B. (2008). Sharp optimality in density deconvolution with dominating bias. I. [*Theory Probab. Appl.*]{} [**52**]{}(1), 24–39.
Carrasco, M. and Chen, X. (2002). Mixing and moment properties of various [GARCH]{} and stochastic volatility models. [*Econometric Theory*]{} [**18**]{}, 17–39.
Comte, F. and Genon-Catalot, V. (2006), Penalized projection estimator for volatility density, [*Scand. J. Statist.*]{} [**33(4)**]{}, 875–893.
Doukhan, P. (1994). [*Mixing: properties and examples*]{}. Springer.
Van Es, A.J. and Spreij, P.J.C. (2009), Multivariate Nonparametric Volatility Density Estimation, [*preprint Mathematics ArXiv*]{} 0910.4337.
Van Es, A.J., Spreij, P.J.C. and Van Zanten, J.H. (2003). Nonparametric volatility density estimation. [*Bernoulli*]{} [**9**]{}(3), 451–465.
Van Es, A.J., Spreij, P.J.C. and Van Zanten J.H. (2005), Nonparametric volatility density estimation for discrete time models, [*J. Nonparametr. Stat.*]{} [**17**]{}, 237–251.
Fan, J. (1991). [On the optimal rates of convergence of nonparametric deconvolution problems.]{} [*Ann. Statist.*]{} [**19**]{}, 1257–1272.
Franke, J., Härdle, W. and Kreiss, J.P. (2003), Nonparametric estimation in a stochastic volatility model, In: [*Recent Advances and Trends in Nonparametric Statistics*]{}, M.G. Akritas and D.N. Politis Eds, Elsevier.
Franke, J., Kreiss, J.P. and Mammen E. (2009), Nonparametric modelling in financial time series, In: [*Handbook of Financial Time Series*]{}, T. Mikosch, J.P. Kreiss, R.A. Davis and T.G. Andersen Eds, Springer.
Genon-Catalot, V., Jeantheau, T. and Larédo, C. (1998), Limit theorems for discretely observed stochastic volatility models, [*Bernoulli*]{} [**4**]{}, 283–303.
Genon-Catalot, V., Jeantheau, T. and Larédo, C. (1999), Parameter estimation for discretely observed stochastic volatility models, [*Bernoulli*]{} [**5**]{}, 855-872.
Genon-Catalot, V., Jeantheau, T. and Lar[é]{}do, C. (2000). [Stochastic volatility models as hidden [M]{}arkov models and statistical applications]{}. [*Bernoulli*]{} [**6**]{}(6),[1051–1079]{}.
Ghysels, E., Harvey, A. and Renault, E. (1996), Stochastic Volatility, in Maddala, G.S. and Rao, C.R. (eds), [*Handbook of Statistics*]{}, Vol. 14, Statistical Methods in Finance, North-Holland, Amsterdam, 119–191.
Gihman, I.I. and Skorohod A.V. (1972), [*Stochastic Differential Equations*]{}, Springer.
Heston, S.L. (1993), A closed-form solution for options with stochastic volatility with applications to Bond and Currency options, [*The Review of Finacial Studies*]{} [**6**]{} (2), 327–343.
Jawerth, B. and Sweldens, W. (1994). [An overview of wavelet based multiresolution analyses]{}. [*SIAM Rev.*]{} [**36**]{}, 377–412.
Karatzas, I. and S.E. Shreve (1991), [*Brownian Motion and Stochastic Calculus*]{}, Springer Verlag, New York.
Pensky, M. and Vidakovic, B. (1999). [Adaptive wavelet estimator for nonparametric density deconvolution.]{} [*Ann. Statist.*]{} [**27**]{}, 2033–2053.
Shephard, N. (1996), Statistical aspects of ARCH and stochastic volatility, in D.R. Cox, D.V Hinkley and O.E. Barndorff-Nielsen (eds.), [*Time Series Models in Econometrics, Finance and Other Fields*]{}, Chapman & Hall, London, 1–67.
Skorokhod, A.V. (1989), [*Asymptotic Methods in the Theory of Stochastic Differential Equations*]{}, AMS.
De Vries, C.G. (1991), On the relation between GARCH and stable processes, [*J. Econometrics*]{} [**48**]{}, 313–324.
Walter, G.G. (1994). *Wavelets and Other Orthogonal Systems with Applications.* CRC Press.
Walter, G.G. and Zayed, A.I. (1996). [Characterization of analytic functions in terms of their wavelet coefficients.]{} [*Complex Variables*]{} [**29**]{}, 265–276.
Wand, M.P. (1998), Finite sample performance of deconvolving kernel density estimators, [*Statist. Probab. Lett.*]{} [**37**]{}, 131–139.
Wand, M.P. and Jones, M.C. (1995), Kernel Smoothing, Chapman and Hall, London.
Wiggins, J. B. (1987), Option valuation under stochastic volatility, [*Journal of Financial Economics*]{} [**19**]{}, 351–372.
Van Zanten, H. and Zareba, P. (2008), A note on wavelet density deconvolution for weakly dependent data, [*Stat. Inference Stoch. Process.* ]{} [**11**]{}, 207–219.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'R. Wijnands, J.J.M. in ’t Zand, M. Rupen, T. Maccarone, J. Homan, R. Cornelisse, R. Fender, J. Grindlay, M. van der Klis, E. Kuulkers, C.B. Markwardt, J.C.A. Miller-Jones, Q.D. Wang,'
subtitle: Description of the program and preliminary results
title: 'The [*XMM-Newton*]{}/[ *Chandra*]{} monitoring campaign of the Galactic center region'
---
Introduction
============
Many X-ray sources (the so-called X-ray transients) exhibit orders of magnitude variability in their X-ray luminosities. Normally they are too dim to be detected and they are only discovered when they experience one of their bright outbursts. The brightest transients can be identified with Galactic neutron stars and black holes accreting matter from a companion star (the “X-ray binaries”). The outbursts are ascribed to a huge increase in the accretion rate onto the compact object. Several other types of sources can also manifest themselves as X-ray transients (e.g. accreting white dwarfs, magnetars, $\gamma$-ray bursts, flare stars, young stellar objects, active binaries), however, their peak luminosities are many orders of magnitude lower than those of the transient X-ray binaries (except for $\gamma$-ray bursts and bursts from magnetars). One usually refers only to transient X-ray binaries when talking about X-ray transients.
Classifying the X-ray transients \[subsection:classifying\]
-----------------------------------------------------------
Early satellites were mostly able to detect only the brightest transients. As the instruments became more sensitive, fainter transients were found which exhibit qualitatively different behavior from the brighter systems. Consequently, in this paper we classify the different X-ray transients based on their maximum observed peak X-ray luminosities (from here on we will quote 2–10 keV luminosities unless otherwise noted):
#### [**Bright to very bright X-ray transients:**]{}
These transients have peak X-ray luminosities of $10^{37-39}$ erg s$^{-1}$ (e.g., [@1997ApJ...491..312C]). Monitoring campaigns with several satellites (e.g., [*BeppoSAX*]{}, [*RXTE*]{}, [*INTEGRAL*]{}; see, e.g., [@2001egru.conf..463I; @1996ApJ...469L..33L; @2001ASPC..251...94S; @2004AstL...30..382R]; see [@2001egru.conf..463I] for an overview of past and current monitoring campaigns focusing on the Galactic center region) have been very successful in discovering many such bright X-ray transients and monitoring their X-ray properties. Intensive studies of the large amount of available data have yielded a good understanding of their behavior and it has been found that a large fraction of them harbor accreting black holes if they accrete from a companion star with a mass $\la$1 solar mass, but a neutron star if the companion has a mass $\ga$10 solar masses. In systems with a low-mass companion (the low-mass transients), the mass transfer occurs because the donor star overflows its Roche-lobe. To explain the outbursts, the disk instability model is the most widely accepted (e.g., [@2001NewAR..45..449L]). In the massive transients the mass transfer occurs because of the strong stellar wind of the companion star or through a decretion disk. The most common type are the so-called Be/X-ray transients in which matter is accreted from the circumstellar decretion disks around rapidly spinning B or sometimes late O-type stars. The physics behind the irregular outbursts observed for the bright high-mass transients are not yet fully understood but it has been suggested that these systems might be Be/X-ray binaries which have relatively low eccentricities (e.g., ).
#### [**Faint X-ray transients:**]{}
These transients have peak X-ray luminosities of $10^{36-37}$ erg s$^{-1}$. Although the faint high-mass transients also contain mostly neutron-star accretors, their outbursts are markedly different from their brighter siblings since the faint outbursts occur usually in series separated by the orbital period while the brighter outbursts do not repeat every orbit, clearly indicating different physics involved (). It is thought that these periodic outbursts occur when an accretor is in a wide eccentric orbit around a Be star and only at minimum distance from the companion does the accretor come inside the decretion disk of the Be star where its is able to accrete matter and thus become X-ray active (e.g, ). Similarly, several detections of low-mass faint transients have been made in the past (e.g., [@1990IAUC.5104....1S; @1994ApJ...425..110P]), but only recently it was realized that they might form a distinct class from the brighter low-mass systems (). This realization came when many such faint low-mass systems were detected by the Wide Field Cameras (WFC) aboard [*BeppoSAX*]{}. Several global characteristics of the faint low-mass transients sets them apart from the brighter systems. First, contrary to what is observed for the brighter low-mass systems, a very large fraction of the faint ones contain neutron-star accretors as determined by the detection of type-I X-ray bursts or millisecond X-ray pulsations[^1] from them. Second, their Galactic distribution is different from that of the brighter systems (), with the faint low-mass transients more concentrated toward the Galactic center. [@2000MNRAS.315L..33K] argued that the faint low-mass transients are indeed different from the bright systems and that they are mainly neutron star X-ray binaries in very compact binaries with orbital periods $<$80 minutes. However, clearly not all systems have such short orbital period since, e.g., the faint transients and millisecond X-ray pulsars SAX J1808.4–3658, XTE J1814–338, and IGR J00291+5934 have periods $>$80 minutes. Also not all systems harbor neutron stars since the faint transient XTE J1118+480 is a strong black hole candidate ([@2001ApJ...556...42W]).
#### [**Very faint X-ray transients (VFXTs):**]{}
These transients have peak X-ray luminosities of $10^{34-36}$ erg s$^{-1}$. Until recently, only limited evidence was available for the existence of this class because detecting such very faint transients is challenging due to sensitivity and/or angular resolution limits of many X-ray instruments. Despite the difficulties in finding VFXTs, pointed observations with relatively sensitive X-ray satellites (e.g., [*Granat*]{}; [*ASCA*]{}) resulted in the detection of several VFXTs near the Galactic center (e.g., [@1994ApJ...425..110P; @1998ApJ...508..854T; @1996PASJ...48..417M]). More recently, a significant number of systems have been found (), thanks to the sensitive [*Chandra*]{} and [*XMM-Newton*]{} X-ray satellites, supporting the claim (e.g., [@2005MNRAS.357.1211S]) that a class of VFXTs exists in our Galaxy. It is very likely that these VFXTs are accreting neutron stars and black holes since only one accreting white dwarf has exhibited outbursts above $10^{34}$ erg s$^{-1}$ ([@1985MNRAS.212..917W]). For this reason we use $10^{34}$ erg s$^{-1}$ as the lower luminosity border for the VFXTs class (see also [@1984MNRAS.210..899V]). In section \[subsection:VFXTs\] we discuss the VFXTs further.\
Our classification of [*bright to very bright*]{}, [*faint*]{}, and [*very faint*]{} X-ray transients is somewhat arbitrary and it is clear that systems straddling these classes do exist. For example, the neutron star X-ray transient in the globular cluster NGC 6440 was classified as a faint X-ray transient () but later it was found to also exhibit bright X-ray outbursts ([@2003ApJ...598..481K]). Furthermore, the bright neutron star X-ray transient SAX J1747.0–2853 (e.g., ) was seen on several occasions at luminosities of only a few times $10^{35}$ erg s$^{-1}$ (e.g., [@2002ApJ...579..422W]). If the bright outbursts of this source had been missed because, for example, no X-ray satellite was pointed in the source direction, then it would have been misclassified as a VFXT. Nevertheless, in this paper such a classification will prove quite useful in talking about the different types of transients.
Recently, another type of accreting neutron star has been identified in the Galaxy: the so-called ’burst-only sources’ (). The burst-only sources are a group of nine objects detected by the [*BeppoSAX*]{}/WFC when they exhibited a type-I X-ray burst. No detectable accretion emission around the bursts could be detected with that instrument, with typical upper limits on the accretion luminosities of the order of $10^{36}$ erg s$^{-1}$. Subsequent X-ray observations of these systems using a variety of satellites revealed that one source is a persistent X-ray source at very faint luminosities ([@intzandetal2005]), two are faint X-ray transients (e.g., ), and one is a VFXT (e.g., [@2004MNRAS.351...31H]). The other five sources could not yet be classified, although they are all X-ray transients since follow-up observations with [*Chandra*]{} could not detect any persistent accretion luminosities from them (). Presumably, they experienced their X-ray bursts during brief accretion outbursts which had peak luminosities below $\sim$$10^{36}$ erg s$^{-1}$. Therefore, these systems are good candidates to be classified as VFXTs, although definitive proof has to come from detecting these systems during such very faint outbursts (see also [@2004NuPhS.132..518C] for a discussion on the classification of burst-only sources).
VFXTs in more detail \[subsection:VFXTs\]
-----------------------------------------
Currently, little is known about the properties of the VFXTs due to the low number of systems known and the scarcity of observations during their outburst episodes. Some of them might be intrinsically bright transients at large distances; however, many are observed near the Galactic center, indicating source distances of $\sim$8 kpc and therefore very low peak intrinsic luminosities. In addition, some VFXTs might be intrinsically brighter than observed due to inclination effects (e.g., [@2005ApJ...633..228M]), but this can be argued for only a small fraction of the VFXTs (see the appendix; see also [@kingwijnands2005]). Therefore, we consider it likely that most VFXTs indeed have very faint intrinsic X-ray luminosities.
The characteristics of the VFXTs (e.g., their spectra, their outburst light curves, timing properties; e.g., [@2005MNRAS.357.1211S; @2005ApJ...622L.113M; @1998ApJ...508..854T]) indicate that they are not a homogeneous class of sources but that different types of accreting neutron stars and black holes show themselves as VFXTs. The detection of slow pulsations in some VFXTs (e.g., [@1998ApJ...508..854T]) indicates that at least some VFXTs could have a high mass donor star since a slow X-ray pulsar is usually associated with high-mass X-ray binaries. The detection of a significant number of high-mass systems would be very important since the currently identified (brighter) systems form likely only a small fraction of the total Galactic population of high-mass X-ray binaries. However, it is possible that some of these systems might in fact not be high-mass X-ray binaries but instead relatively close-by accreting magnetic white dwarfs (i.e., the intermediate polars). From the X-ray data it is difficult to ascertain the exact nature of these systems and identifications of their optical or IR counterparts are needed to distinguish the high-mass X-ray binary systems from the white dwarf systems.
No significant pulsations have so far been detected for the other VFXTs. More information about the nature of these systems could be obtained by studying them in optical or infrared. However, none of the VFXTs has so far been detected at optical or infrared wavelengths. For the VFXTs in the Galactic center region it was found that they harbor companions fainter than B2 IV stars ([@2005ApJ...622L.113M]). Therefore, it is likely that a significant fraction of the VFXTs are neutron stars and black holes accreting matter at a very low rate from a low-mass companion star. Moreover, at least one VFXT has exhibited type-I X-ray bursts (SAX J1828.5–1037; ) which are usually identified with low-mass X-ray binaries. Such low-mass X-ray binaries have very low time-averaged accretion rates, which could become a challenge for our understanding of their evolution ([@kingwijnands2005]).
To improve on the limited amount of knowledge about the observational properties of VFXTs, monitoring campaigns are required with instruments which:
- are sensitive enough to detect the very faint X-ray luminosities of these systems
- have a large field-of-view (FOV) to monitor a large region allowing for the discovery and monitoring of many systems
- have rapid data turn-around time to allow fast follow-up observations at all wavelengths to study the VFXTs when they are still active
- have (sub-)arcsecond resolution to limit source confusion and to allow for unique determination of the counterparts at other wavelengths which will help to establish the nature of the systems (e.g., IR observations might help to determine the type of companion star).
Such a monitoring instrument does not exist and in practice might be difficult to achieve since, for example, a large FOV and very good sensitivity are usually mutually exclusive. Fortunately, [*XMM-Newton*]{} and [*Chandra*]{} are excellent to perform such a monitoring program but only within a limited FOV. Therefore, we have secured such a program using both satellites for a limited region close to Sgr A\*. This region was chosen for two reasons: there is a high concentration of stars along the line of sight and most known VFXTs have been detected in this region (e.g., [@2005ApJ...622L.113M]). We will first describe the details of the program and then present some initial preliminary results.
Description of our program
==========================
In proposal cycle 4 of [*XMM-Newton*]{}, we have secured a program using this satellite in combination with the [*Chandra*]{} X-ray Observatory to monitor a 1.7 square degree region centered around Sgr A\*. The FOV of our program is shown in Figure \[fig:fov\]. We will have 4 X-ray epochs (two with [*XMM-Newton*]{} and two with [*Chandra*]{}) with each epoch consisting of seven observations (their pointing directions are indicated in Figure \[fig:fov\]). Each pointing lasts $\sim$5 ksec and will reach an overall sensitivity of $\sim$$5\times 10^{33}$ erg s$^{-1}$ (at 8 kpc; from here on we assume a distance of 8 kpc when quoting X-ray luminosities) for the [*Chandra*]{} observations to an order of magnitude better for the [*XMM-Newton*]{} observations. The FOVs of the different pointings overlap by several arcminutes (Fig. \[fig:fov\]) to compensate for the loss in sensitivity toward the edge of the FOV of the instruments. For the [*Chandra*]{} pointings we have chosen to use the HRC-I detector because it has the largest FOV of any instrument aboard [*Chandra*]{} (the FOV of the HRC-I is approximately equal to the FOV of the instruments aboard [*XMM-Newton*]{}), although this comes at a loss of sensitivity for hard sources and the loss of any spectral information. All detectors will be on during the [*XMM-Newton*]{} observations but we will mostly focus on the data obtained with the EPIC detectors (the two MOS and the pn chips). The RGS will only be useful when a single bright source is in the FOV.
The purpose of our monitoring campaign is to study the variability behavior of the transient and persistent sources in the survey region at levels not reachable by other monitoring instruments in orbit. Although our interests focus on the accreting neutron stars and black holes in the FOV, our campaign will also be used to study the X-ray properties of X-ray active stars, dense star clusters (e.g., the Arches cluster), flares from Sgr A\*, accreting white dwarfs, and any other object in the FOV which might exhibit persistent or transient X-ray emission at levels detectable during our observations. The different observation epochs are separated in time from each other by at least one month so that we can monitor the X-ray behavior of the detected sources on timescales of about 1 month to almost a year. The first epoch data were gathered on 5 June 2005 ([*Chandra*]{}) and the remaining three epochs are currently scheduled for the week of 17–24 October 2005 ([*Chandra*]{}), and at the end of February 2006 and early April 2006 ([*XMM-Newton*]{}). Here we report on the initial results obtained during the first epoch [*Chandra*]{} data.
Data analysis
=============
The first data of our monitoring campaign were obtained on 5 June 2005, using the [*Chandra*]{} satellite (see Tab. \[table:observations\] for a log of the observations) and the HRC-I detector. We processed the data using the [*Chandra*]{} CIAO tools (version 3.2.1) and the standard [*Chandra*]{} analysis threads[^2]. We checked for background flares during our observations, but none were found, allowing us to use all available data. We merged the 7 different HRC-I observations into one image which we show in Figure \[fig:images\]. Clearly, several bright sources are visible by eye. We used the tool [*wavdetect*]{} to search for point sources in our data and to obtain the coordinates of each source that was detected. We ran the tool on the combined image as well as on each individual observation. Due to variations in the size and shape of the point-spread-function as a function of offset angle from the pointing directions, we ran [*wavdetect*]{} on images with different binning factors. We are still exploring ways to optimize our detection method so it is likely that in our final analysis we will find more sources than the ones we report on in this paper. However, we expect that in the current analysis we are complete for sources with inferred X-ray luminosities $>$$10^{34}$ erg s$^{-1}$.
The errors on the source positions are difficult to estimate for the sources found at relatively large offset angles. The asymmetries in the point-spread-function at large off-axis angles can result in large systematic uncertainties when using [*wavdetect*]{} (e.g., [@hong2005]). We are pursuing extensive simulations to investigate the effects on the positional errors for a large range of offset angles as well as different source luminosities. A similar investigation has already been performed for the [*Chandra*]{}/ACIS-I combination by [@hong2005]. Although they investigated the ACIS-I (and not the HRC-I), used only offset angles up to $10'$ (instead of $>$20$'$ as we sometimes encounter), and focused mainly on the faint sources, we will use their results as a first order approximation on the accuracy of the positions we obtain using [*wavdetect*]{}. We use equation 5 in [@hong2005] to estimate the uncertainties in our positions. If the sources were detected in multiple observations, we only give the positions and their errors obtained from the data set in which the sources had the smallest offset in order to minimize systematic uncertainties. However, we urge caution when using our positional uncertainties.
For each detected source we extracted the background corrected count rate from the images using the standard CIAO tools for the full energy range of the HRC-I (0.08–10 keV). The count rates obtained can be converted into fluxes using PIMMS[^3], assuming particular spectral models and values for the interstellar absorption. Again, the analysis is complicated because many sources are detected at large off-axis angles and vignetting becomes a serious issue. It is currently difficult to correct the count rates for vignetting because the effects depend strongly on both the off-axis angle and the assumed source spectra (which are unknown since the HRC-I does not currently allow to extract energy information). Therefore, the count rates we quote are the uncorrected count rates and are therefore lower limits to the count rates the sources would have if they were located on-axis. Depending on the off-axis angles and the source spectra, the on-axis count rates could have been larger by a factor of a few. Again, if the sources were detected in multiple observations, we only give the count rates from the data set in which the sources had the smallest offset to minimize the systematic errors on the derived fluxes.
Results
=======
In total we have detected 21 sources so far. Two sources (the Sgr A\* complex and the Arches cluster) are known to embody a complex of point sources in combination with strong diffuse emission ([@2003ApJ...589..225M; @2002ApJ...570..665Y]). The analysis of these complex regions is still in progress. Ten of the remaining sources can be identified with known stars (e.g., HD 316314, HD 316224, HD 161274, TYC 06840-38-1, ALS 4400; Fig. \[fig:images\] left) or have clear counterparts in the Digital Sky Survey images indicating that they are foreground objects (and hence have relatively low X-ray luminosities). We will not discuss the detections of the foreground objects further in this paper, instead we will focus on the detected X-ray binaries.
The persistent sources
----------------------
We detected the two persistent X-ray binaries known to be present in the surveying region: 1E 1743.1–2843 and 1A 1742–294.
### 1E 1743.1–2843
1E 1743.1–2843 is a persistent X-ray binary for which the type of accreting object is not yet known. The source was in the FOV of two of the seven HRC-I pointings and was detected during both observations but we detected no bursts from the source. The position obtained from our [*Chandra*]{}/HRC-I data is consistent with that derived from a previous [*XMM-Newton*]{} observation () although, due to the systematic uncertainties in our positional errors, it cannot currently be determined if our position is better than the [*XMM-Newton*]{} one. We used PIMMS to convert the obtained count rate (see Tab. \[table:binaries\]). We assumed an absorbed power-law model similar to what was found by when fitting the [*XMM-Newton*]{} observation of the source (they obtained an equivalent hydrogen column density $N_{\rm H}$ of $2\times10^{23}$ cm$^{-2}$ and a photon index of 1.8). This results in unabsorbed fluxes of $1.8\times 10^{-10}$ (2–10 keV) and $3\times 10^{-10}$ erg cm$^{-2}$ s$^{-1}$ (0.5–10 keV) and X-ray luminosities of 1.4 and $2.3\times 10^{36}$ erg s$^{-1}$, respectively. These X-ray luminosities are very similar to what has been seen before for this source (e.g., ).
### 1A 1742–294
1A 1742–294 is a persistent X-ray binary harboring a neutron-star accretor as evidenced by the type-I X-ray bursts observed from this system (see, e.g., [@1994ApJ...425..110P]). We detected this source during both HRC-I pointings in which the source was in the FOV. During the GC-10 pointing we detected an X-ray burst. Our [*Chandra*]{} position is fully consistent with the best position so far reported on this source (using [*ROSAT*]{}; ) and despite the possible unknown systematic uncertainty in our errors, our position is better. We again used PIMMS to convert the obtained count rate (see Tab. \[table:binaries\]) and used the absorbed power-law model ($N_{\rm H} \sim$$6\times10^{22}$ cm$^{-2}$ ; photon index $\sim$1.8) found when fitting the [*BeppoSAX*]{} and [*ASCA*]{} data of the source ([@1999ApJ...525..215S; @2002ApJS..138...19S]). This results in unabsorbed fluxes of $2.3\times 10^{-10}$ (2–10 keV) and $3.7\times 10^{-10}$ erg cm$^{-2}$ (0.5–10 keV) s$^{-1}$. The corresponding X-ray luminosities are 1.8 and $2.8\times 10^{36}$ erg s$^{-1}$, consistent with what has been observed before for this source (e.g., [@1999ApJ...525..215S]).
The transient sources
---------------------
Two transients were clearly visible during our observations: GRS 1741.9–2853 and XMM J174457–2850.3. We made a preliminary announcement of the detection of these new outbursts on 6 June 2005 ([@2005ATel..512....1W]). Following these detections, we obtained an additional [*Chandra*]{} observation of both sources (using the ACIS-I detector) on 1 July 2005 (see Tab. \[table:observations\] for details). Because the two transients were only $\sim$4.6$'$ away from each other, we could observe both sources with only one ACIS-I pointing. We placed both sources at an off-axis angle of 7$'$ in order to limit pile-up in case the sources were as bright as seen during the HRC-I observations. The ACIS-I data were also analyzed using CIAO and the standard threads. Again, all data could be used since no episodes of high background emission occurred during our observation.
### GRS 1741.9–2853
GRS 1741.9–2853 is a neutron star X-ray transient (it exhibits type-I bursts; e.g., ) which has been detected several times in outburst since its original discovery in 1990 ([@1990IAUC.5104....1S]). Its peak luminosity is typically a few times $10^{36}$ erg s$^{-1}$ making it a faint X-ray transient (see [@2003ApJ...598..474M] for more details). This source was detected during two of our pointings (Tab. \[table:binaries\]) but we detected no bursts. The position of the source was consistent with, but not better than the one obtained by [@2003ApJ...598..474M]. The observed count rate was converted into fluxes using PIMMS and assuming an absorbed power-law with $N_{\rm H} = 9.7\times10^{22}$ cm$^{-2}$ and a photon index of 1.88 ([@2003ApJ...598..474M]). This results in unabsorbed fluxes of $1.1\times 10^{-10}$ (2–10 keV) and $1.8\times 10^{-10}$ erg cm$^{-2}$ s$^{-1}$ (0.5–10 keV), yielding X-ray luminosities of 0.8 and $1.4\times 10^{36}$ erg s$^{-1}$, respectively (for comparison with previous [*Chandra*]{} data on this source reported by [@2003ApJ...598..474M], we also list the 2–8 keV luminosity of $7.0\times10^{35}$ erg s$^{-1}$). GRS 1741.9–2853 was also detected during the additional [*Chandra*]{}/ACIS-I observation (Fig. \[fig:extra\_image\]). We extracted the source spectrum using a source extraction region of 10$''$ and a background extraction circle of 50$''$ from a source-free region close to GRS 1741.9–2853. The spectrum was rebinned to have at least 15 counts per bin to allow the $\chi^2$ fitting method. The resulting spectrum is shown in Figure \[fig:spectra\]. We used XSPEC to fit the spectrum and the fit results obtained are listed in Table \[table:spectral\_fits\]. Clearly, the source flux had decreased by almost an order of magnitude within about a month (i.e., since 5 June 2005). The long-term light curve of the source is plotted in Figure \[fig:lc\] showing the multiple outbursts of the source in the last 15 years.
### XMM J174457–2850.3
XMM J174457–2850.3 is also clearly detected during our HRC-I observations (Fig. \[fig:images\]). This source has been detected only once before in outburst in 2001 (using [*XMM-Newton*]{}; [@2005MNRAS.357.1211S]). During that outburst the source was seen at a peak luminosity of $5 \times 10^{34}$ erg s$^{-1}$, justifying a classification as a VFXT. We detected it during two of our pointings (Tab. \[table:binaries\]) but saw no bursts. Our source position is consistent with that obtained by [@2005MNRAS.357.1211S], although the exact uncertainty on our HRC-I position is currently unclear. However, the source was also detected during our additional ACIS-I observation yielding a more reliable position even though the source was relatively weak (see Tab. \[table:binaries\]). This position is significantly better than the [*XMM-Newton*]{} one. The observed HRC-I count rate was converted into fluxes using PIMMS and assuming an absorbed power-law with $N_{\rm H} = 6\times10^{22}$ cm$^{-2}$ and a photon index of 1.0 ([@2005MNRAS.357.1211S]). This resulted in unabsorbed fluxes of $1.1\times 10^{-10}$ (2–10 keV) and $1.3\times 10^{-10}$ erg cm$^{-2}$ s$^{-1}$ (0.5–10 keV) and in X-ray luminosities of 0.8 and $1.0\times 10^{36}$ erg s$^{-1}$, respectively. This is significantly brighter than what was previously found for the source and makes it a borderline case as a VFXT. As stated above, XMM J174457–2850.3 was also detected during the additional [*Chandra*]{}/ACIS-I observation (Fig. \[fig:extra\_image\]). We extracted the source spectrum using a source extraction region of 5$''$. Due to the rather low number of source photons (26 counts in the 0.3–7.0 keV energy range) we did not rebin the spectrum or subtract the background (which was $<$0.3 photon in the source region and therefore negligible) so that we could use the Cash statistics ([@1979ApJ...228..939C]) when fitting the spectrum in XSPEC. The fit results obtained for this observation are also listed in Table \[table:spectral\_fits\] and the resulting spectrum is shown in Figure \[fig:spectra\]. Clearly, the source flux had decreased by nearly three orders of magnitude within approximately a month (i.e., since 5 June 2005). The long-term light curve of the source is plotted in Figure \[fig:lc\].
### A possible new VFXT
None of the other known transients in the FOV of our observations (see Tab. \[table:sources\_in\_FOV\]) were conclusively detected in our HRC-I data. The upper limits on their luminosities depend strongly on their spectral shape and their off-axis positions, with a rough estimate of $\sim$$10^{34}$ erg s$^{-1}$. Several additional weak sources were detected during our observations which could not be identified with a star in the Digital Sky Survey database. Only one of these had a large enough count rate (see Tab. \[table:binaries\]) that its X-ray luminosity exceeded $10^{34}$ erg s$^{-1}$ if it had a ’prototypical X-ray binary’ spectrum (power-law model with photon index of 1.8 and a typical $N_{H}$ of $6\times10^{22}$ cm$^{-2}$). Using such a spectral shape, the source had unabsorbed X-ray fluxes of 1.9 and $3.1\times10^{-12}$ erg s$^{-1}$ for 2–10 keV and 0.5–10 keV, respectively, and thus luminosities of $1.5\times
10^{34}$ erg s$^{-1}$ (2–10 keV) and $2.4\times10^{34}$ erg s$^{-1}$ (0.5–10 keV). We note that we do not know the intrinsic source spectrum and therefore these fluxes and luminosities could be significantly off if the real source spectrum is considerably different. We investigated the [*Chandra*]{} and [*XMM-Newton*]{} archives and found that the source was in the FOV of one previous [*XMM-Newton*]{} observation. The source was not detected during this [*XMM-Newton*]{} observation, but it was at the edge of its FOV making it difficult to obtain a reliable upper limit on the flux, especially because we do not know the spectral shape of the source. We estimate that the luminosity of the source was at least a factor of a few fainter during the [*XMM-Newton*]{} observation compared with our HRC-I data. Although this is suggestive of a transient nature for this source, it could also be a highly variable persistent source. Currently, we will refer to this source as a possible new VFXT.
### Observations at other wavelengths
We obtained VLA observations at 4 and 6 cm on 8–9 June 2005 of GRS 1741.9–2853, XMM J174457–2850.3, and the possible new VFXT. The analysis of these radio data is complicated by the strong side-lobes of Sgr A\* and we are still in the process of fully analyzing these data. A preliminary analysis of the 4 cm data shows that none of the sources were conclusively detected with radio fluxes of $0.003\pm0.060$, $-0.002\pm0.046$, and $0.032\pm0.043$ mJy/beam, respectively. On 8 June 2005, [@2005ATel..522....1L] obtained I-band images of GRS 1741.9–2853 and XMM J174457–2850.3 using the Magellan-Baade telescope but could not detect the I-band counterparts of the sources. This is not surprising when considering the high absorption column in front of both sources.
Discussion
==========
We have presented our initial results of the first observations taken as part of our [*XMM-Newton*]{}/[*Chandra*]{} monitoring campaign of the inner region of our Galaxy. Using our [*Chandra*]{}/HRC-I observations we detected mostly foreground objects (like X-ray active stars), but we also detected two persistent X-ray binaries, two X-ray transients, and one possible very faint X-ray transient (but its transient nature requires further confirmation). Clearly, our monitoring [*XMM-Newton*]{}/[*Chandra*]{} campaign is detecting transients in outburst which are being missed by the other monitoring instruments in orbit. Our campaign therefore complements, as designed, other monitoring campaigns using satellites currently in orbit (e.g., [@1996ApJ...469L..33L; @2001ASPC..251...94S; @2004AstL...30..382R; @2005ATel..438....1K]). These programs find mainly the brighter transients or the faint transients far away from the crowded fields near Sgr A\*.
The faint X-ray transient GRS 1741.9–2853 was detected at a level of $\sim$$10^{36}$ erg s$^{-1}$, very similar to what has been observed previously for this source. A month after our initial HRC-I observations this source could still be detected at $\sim$$10^{35}$ erg s$^{-1}$ with the ACIS-I. The parameters obtained for the source spectrum during this observation were consistent with those found by [@2003ApJ...598..474M] when the source was an order of magnitude brighter, indicating that the source spectrum is not very dependent on source luminosity. Although we did not detect X-ray bursts during our observations, this source is known to exhibit such phenomena making it very likely to be a neutron star accreting from a low-mass companion star. Even though no optical/infrared counterpart has so far been found for this source, type-I X-ray bursts have only been seen for low-mass X-ray binaries making it very likely that GRS 1741.9–2853 is also such a system. The fact that GRS 1741.9–2853 harbors a neutron star also is consistent with the non-detection of the source in our radio data since neutron-star low-mass X-ray binaries are known to exhibit very low radio luminosities (e.g., [@2001MNRAS.324..923F; @2005ApJ...626.1020M] )
Figure \[fig:lc\] shows that the source has been seen to be in outburst at least 5 times with X-ray luminosities above $10^{34}$ erg s$^{-1}$. Its recurrence time can be estimated to be between 2 and 5 years, making GRS 1741.9–2853 one of the most active transients in our FOV, with a duty cycle of about 50% (as estimated from Fig. \[fig:lc\])[^4] . Its peak luminosity is very similar to the accreting millisecond X-ray pulsar SAX J1808.4–3658. For that system and the other accreting millisecond X-ray pulsars, it has been suggested that their pulsating nature is related to their rather low time averaged accretion rates (e.g., [@2001ApJ...557..958C]). Although the time averaged accretion rate of GRS 1741.9–2853 seems to be higher than for the accreting millisecond pulsars due to its higher duty cycle, GRS 1741.9–2853 could still be a millisecond X-ray pulsar as well (see also [@2003ApJ...598..474M]), especially if its duty cycle has been overestimated (see footnote \[footnote:duty\]). Unfortunately, its faintness and its location in the Sgr A\* region make it very difficult to detect these pulsations using [*RXTE*]{} because of significant contribution to the detected count rate from other sources in the FOV. However, with [*XMM-Newton*]{} pulsations could be detected within several tens of ksec (depending on the actual fluxes of the source) if they have similar strengths as the pulsations seen in the known accreting millisecond pulsars.
The VFXT XMM J174457–2850.3 was also detected during our HRC-I observations at an X-ray luminosity close to $10^{36}$ erg s$^{-1}$. This is about a factor of 20 higher than what was previously seen for this source ([@2005MNRAS.357.1211S]). This demonstrates that VFXTs can exhibit a large range of X-ray luminosities (similar to what has been observed for the brighter systems) and XMM J174457–2850.3 is at the border between faint and very faint X-ray transients, clearly demonstrating that our luminosity boundaries are somewhat arbitrary as discussed in the introduction. It is possible that the previous detection of this source was made either during the rise or decay of a full outburst and that the maximum luminosity reached at the time was closer to what we have observed for the source during our HRC-I observations. Within a month the source luminosity has decreased by nearly 3 orders of magnitude. Its X-ray spectrum at this low X-ray luminosity was consistent with that found by [@2005MNRAS.357.1211S] demonstrating that for this source the shape of its spectrum is not strongly dependent on luminosity for luminosities below $5\times 10^{34}$ erg s$^{-1}$. Since we cannot extract any spectral information from the HRC-I data we cannot determine if the spectrum was significantly different at times when the source had X-ray luminosities close to $10^{36}$ erg s$^{-1}$. Since only very few observations have been performed of this source (see Fig. \[fig:lc\]), it is difficult to estimate its recurrence time (at most of order 3 years according to Fig. \[fig:lc\]) and its time-averaged accretion rate. The non-detection at radio wavelengths might indicate that the source harbors a neutron-star accretor since according to the radio-X-ray correlation found for low luminosity black hole binaries ([@2003MNRAS.344...60G]) the source should have had (if it harbors a black-hole accretor which was accreting at $\sim$$10^{36}$ erg s$^{-1}$) a radio flux of $\sim$1 mJy, significantly higher than our radio upper limit. Alternatively, XMM J174457–2850.3 could still harbor a black hole, but one which does not follow this correlation.
We did not detect any unambiguous new VFXTs during our observations, although we detected a possible new VFXT whose transient nature must be confirmed. This will be possible with our next sets of monitoring observations. Our three additional epochs will also be very important to find further VFXTs, either previously unknown transients or recurrent ones. Our observations will allow us to set tighter constraints on the time averaged accretion rates of these systems than is currently possible with the available data. Such constraints are especially important for the low-mass X-ray binaries among the VFXTs because if their time-averaged accretion rates is very low then our theories of the evolution of such systems will have a very hard time explaining their existence without invoking exotic scenarios such as accretion from a brown dwarf or planet or intermediate mass black hole accretors (e.g., [@kingwijnands2005]). The latter option cannot be invoked if type-I X-ray bursts have been observed for these systems since this establishes the existence of a neutron star accretor. Potential candidates for such systems are the burst-only sources mentioned in the introduction. Monitoring observations of these sources with sensitive X-ray telescopes would be very useful to constrain their time-averaged accretion rates to determine if indeed these rates are very low for these systems.
Finding new VFXTs and determining their time averaged accretion rate using our monitoring campaign is only one way forward to increase our understanding of these enigmatic transients. We now discuss other avenues that can be explored as well to achieve that goal. First, a search in the data archives for previously unnoticed VFXTs (e.g., by comparing different exposures of the same fields) might lead to the detection of several more systems. Second, larger regions of our Galaxy need to be monitored at the desired sensitivity to detect the low fluxes observed from VFXTs. It is especially important to determine if a large number of VFXTs also exist outside the inner region of our Galaxy. [@2005ApJ...622L.113M] found that the excess of VFXTs within 10 arcminutes of Sgr A\* is significant and might point to an unusual formation history of these systems. However, if a large number of VFXTs are also found further away from Sgr A\* (e.g., systems like XMM J174716–2810.7 or SAX J1828.5–1037; [@2003ATel..147....1S; @2004MNRAS.351...31H]), then any production mechanism which requires the high stellar density near Sgr A\* cannot be invoked for these VFXTs. There is currently no monitoring satellite in orbit which can perform that task mainly due to a lack of sensitivity and angular resolution of the instruments. However, it is possible to derive a first approximation to the number density of VFXTs at large distances from the center of a spiral galaxy by performing several deep pointings of the core of other spiral galaxies. The most obvious choice is the nearest large spiral galaxy to our own, M31. Within a $\sim$250 ksec exposure it is possible to observe a 3.7 kpc $\times$ 3.7 kpc region of M31 using the [*Chandra*]{}/ACIS-I detector with a limit sensitivity of $1-4 \times
10^{34}$ erg s$^{-1}$ (depending on the spectral properties of the sources). Several such deep pointings would detect all but the faintest X-ray transients in a large region of M31. Alternatively, such programs can also be performed for the smaller spiral galaxy M33 or for galaxies further away. In the latter case, the limiting sensitivity will be of course less.
RW thanks Michael Muno and Andrew King for useful discussions about very faint X-ray transients. We also thank Michael Muno for the information on the absence of eclipses in CXOGC J17535.5–290124.
Boirin, L., Barret, D., Olive, J. F., Bloser, P. F., & Grindlay, J. E. 2000, , 361, 121
Branduardi, G., Ives, J. C., Sanford, P. W., Brinkman, A. C., & Maraschi, L. 1976, , 175, 47P
Cash, W. 1979, , 228, 939
Chen, W., Shrader, C. R., & Livio, M. 1997, , 491, 312
Cocchi, M., Bazzano, A., Natalucci, L., Ubertini, P., Heise, J., Muller, J. M., & in ’t Zand, J. J. M. 1999, , 346, L45
Cornelisse, R., et al. 2002a, , 392, 885
Cornelisse, R., Verbunt, F., in’t Zand, J. J. M., Kuulkers, E., & Heise, J. 2002b, , 392, 931
Cornelisse, R., et al. 2004, Nuclear Physics B Proceedings Supplements, 132, 518
Cremonesi, D. I., Mereghetti, S., Sidoli, L., & Israel, G. L. 1999, , 345, 826
Cumming, A., Zweibel, E., & Bildsten, L. 2001, , 557, 958
Fender, R. P., & Kuulkers, E. 2001, , 324, 923
Gallo, E., Fender, R. P., & Pooley, G. G. 2003, , 344, 60
Galloway, D. K., Chakrabarty, D., Muno, M. P., & Savov, P. 2001, , 549, L85
Giles, A. B., Swank, J. H., Jahoda, K., Zhang, W., Strohmayer, T., Stark, M. J., & Morgan, E. H. 1996, , 469, L25
Hands, A. D. P., Warwick, R. S., Watson, M. G., & Helfand, D. J. 2004, , 351, 31
Heise, J., in’t Zand, M. J. J., Smith, S. M. J., Muller, M. J., Ubertini, P., Bazzano, A., Cocchi, M., & Natalucci, L. 1999, Astrophysical Letters Communications, 38, 297
Homan, J., & van der Klis, M. 2000, , 539, 847
Hong, J., van den Berg, M., Schlegel, E.M., Grindlay, J.E., Koenig, X., Laycock, S., Zhao, P. 2005, ApJ, in press
in ’t Zand, J. J. M., et al. 1991, Advances in Space Research, 11, 187
in ’t Zand, J. J. M., et al. 1999, , 345, 100
in’t Zand, J. 2001, ESA SP-459: Exploring the Gamma-Ray Universe, 463
In ’t Zand, J.J.M., Cornelisse, R., & Mendez, M., 2005, A&A, in press
Kaaret, P., Zand, J. J. M. i., Heise, J., & Tomsick, J. A. 2003, , 598, 481
King, A. R. 2000, , 315, L33
King, A.R. & Wijnands, R. 2005, Letters, in press (astro-ph/0511486)
Kuulkers, E., et al. 2005, The Astronomer’s Telegram, 438, 1
Lasota, J.-P. 2001, New Astronomy Review, 45, 449
Laycock, S., Zhao, P., Torres, M. A. P., Wijnands, R., Steeghs, D., Grindlay, J., Hong, J., & Jonker, P. G. 2005, The Astronomer’s Telegram, 522, 1
Levine, A. M., Bradt, H., Cui, W., Jernigan, J. G., Morgan, E. H., Remillard, R., Shirey, R. E., & Smith, D. A. 1996, , 469, L33
Maeda, Y., Koyama, K., Sakano, M., Takeshima, T., & Yamauchi, S. 1996, , 48, 417
Muno, M. P., et al. 2003a, , 589, 225
Muno, M. P., Baganoff, F. K., & Arabadjis, J. S. 2003b, , 598, 474
Muno, M. P., Pfahl, E., Baganoff, F. K., Brandt, W. N., Ghez, A., Lu, J., & Morris, M. R. 2005a, , 622, L113
Muno, M. P., Belloni, T., Dhawan, V., Morgan, E. H., Remillard, R. A., & Rupen, M. P. 2005b, , 626, 1020
Muno, M. P., Lu, J. R., Baganoff, F. K., Brandt, W. N., Garmire, G. P., Ghez, A. M., Hornstein, S. D., & Morris, M. R. 2005, , 633, 228
Negueruela, I., Smith, D. M., Reig, P., Chaty, S., Miguel Torrejon, J. 2005, To appear in Proceedings of “The X-ray Universe 2005”, held in San Lorenzo de El Escorial (Madrid, Spain), 26-30 September 2005, ESA-SP 604 (astro-ph/0511088)
Okazaki, A. T., & Negueruela, I. 2001, A&A, 377, 161
Pavlinsky, M. N., Grebenev, S. A., & Sunyaev, R. A. 1994, , 425, 110
Porquet, D., Rodriguez, J., Corbel, S., Goldoni, P., Warwick, R. S., Goldwurm, A., & Decourchelle, A. 2003, , 406, 299
Porquet, D., Grosso, N., Burwitz, V., Andronov, I. L., Aschenbach, B., Predehl, P., & Warwick, R. S. 2005, , 430, L9
Revnivtsev, M. G., et al. 2004, Astronomy Letters, 30, 382
Sakano, M., Koyama, K., Murakami, H., Maeda, Y., & Yamauchi, S. 2002, , 138, 19
Sakano, M., Warwick, R. S., Decourchelle, A., & Wang, Q. D. 2005, , 357, 1211
Sidoli, L., Mereghetti, S., Israel, G. L., Chiappetti, L., Treves, A., & Orlandini, M. 1999, , 525, 215
Sidoli, L., Belloni, T., & Mereghetti, S. 2001, , 368, 835
Sidoli, L., & Mereghetti, S. 2003, The Astronomer’s Telegram, 147, 1
Sunyaev, R. 1990, , 5104, 1
Swank, J., & Markwardt, C. 2001, ASP Conf. Ser. 251: New Century of X-ray Astronomy, 251, 94
Torii, K., et al. 1998, , 508, 854
Verbunt, F., Elson, R., & van Paradijs, J. 1984, , 210, 899
Wagner, R. M., Foltz, C. B., Shahbaz, T., Casares, J., Charles, P. A., Starrfield, S. G., & Hewett, P. 2001, , 556, 42
Wang, Q. D., Gotthelf, E. V., & Lang, C. C. 2002, , 415, 148
Watson, M. G., King, A. R., & Osborne, J. 1985, , 212, 917
Werner, N., et al. 2004, , 416, 311
Wijnands, R. 2004, AIP Conf. Proc. 714: X-ray Timing 2003: Rossi and Beyond, 714, 209
Wijnands, R., Strohmayer, T., & Franco, L. M. 2001, , 549, L71
Wijnands, R., Miller, J. M., & Wang, Q. D. 2002, , 579, 422
Wijnands, R., et al. 2005, The Astronomer’s Telegram, 512, 1
Yusef-Zadeh, F., Law, C., Wardle, M., Wang, Q. D., Fruscione, A., Lang, C. C., & Cotera, A. 2002, , 570, 665
-------------------- ------- ---------------------- ---------- ------------
Field name ObsID Date Exposure Instrument
(June 5, 2005, UTC) (ksec)
GC-1 6188 01:44 – 03:32 5.08 HRC-I
GC-2 6190 03:32 – 05:05 5.15 ”
GC-3 6192 05:05 – 06:37 5.14 ”
GC-7 6194 06:37 – 08:09 5.14 ”
GC-8 6196 08:09 – 09:41 5.14 ”
GC-9 6198 09:41 – 11:13 5.15 ”
GC-10 6200 11:13 – 13:08 5.15 ”
GRS 1741.9–2853 & 6311 02:04 – 03:40 July 1 4.01 ACIS-I
XMM J174457–2850.3 ”
-------------------- ------- ---------------------- ---------- ------------
-------------------- ---------- ------------ -------------- ---------------- ------------------ ------------------- -----------------------------------------------
Source name In FOV Offset[^5] Count rate[^6] Comment
of frame ($'$) RA Dec Error ($''$)[^7] (counts s$^{-1}$)
1E 1743.1–2843 GC-1 6.6 17 46 21.094 -28 43 42.3 1.1 0.247$\pm$0.007 Persistent X-ray binary
GC-2 18.5
1A 1742–294 GC-9 11.1 17 46 05.201 -29 30 53.3 1.3 0.84$\pm$0.01 Persistent neutron star low-mass X-ray binary
GC-10 20.6 Burst detected during GC-10
GRS 1741.9–2853 GC-2 9.9 17 45 02.385 -28 54 50.2 1.5 0.267$\pm$0.008 Neutron star faint X-ray transient
GC-7 15.9
ACIS-I 6.9 17 45 02.350 -28 54 49.9 1.2 0.269$\pm$0.008
XMM J174457–2850.3 GC-7 11.4 17 44 57.440 -28 50 20.3 1.5 0.337$\pm$0.009 Very faint X-ray transient
GC-2 13.6
ACIS-I 7.0 17 44 57.451 -28 50 21.1 2.1 0.0064$\pm$0.0015
Possible VFXT GC-3 8.7 17 47 37.671 -29 08 09.6 2.5 0.007$\pm$0.002 Transient nature needs confirmation
-------------------- ---------- ------------ -------------- ---------------- ------------------ ------------------- -----------------------------------------------
Parameter GRS 1741.9–2854 XMM J174457–2850.3
------------------------------------------------------ ---------------------- ---------------------
$N_{\rm H}$ ($10^{22}$ cm$^{-2}$) $10.5^{+4.9}_{-3.7}$ 6[^8]
Photon index $1.8^{+1.0}_{-0.8}$ $1.3\pm1.1$
Flux ($10^{-12}$ erg cm$^{-2}$ s$^{-1}$, unabsorbed)
0.5–10.0 keV $37^{+93}_{-14}$ $0.5^{+0.4}_{-0.5}$
2.0–10.0 keV $22^{+10}_{-3}$ $0.4^{+0.2}_{-0.4}$
2.0–8.0 keV $17^{+12}_{-3}$ $0.3^{+0.1}_{-0.3}$
![Field-of-view of our monitoring campaign in Galactic coordinates. The circles indicate the [*XMM-Newton*]{}/MOS FOV which is similar to the [*Chandra*]{}/HRC-I FOV. The FOV is over-plotted on the [*Chandra*]{} Galactic center survey data as presented by [@2002Natur.415..148W], which is centered around Sgr A\* and covers a region of approximately 1$\times$ 2. The ’field name’ (see Tab. \[table:observations\]) of each pointing is indicated.[]{data-label="fig:fov"}](4129fig1.ps){width="\textwidth"}
![The merged image of the 7 [*Chandra*]{}/HRC-I observations. [**Left panel:**]{} the Arches cluster and a sub-set of the detected foreground objects are indicated (only those stars which are known in Simbad are labeled). [**Right panel:**]{} the detected X-ray binaries as well as the possible VFXT. Also indicated is the complex X-ray emission around Sgr A\*. []{data-label="fig:images"}](4129f2a.ps "fig:"){width="50.00000%"}![The merged image of the 7 [*Chandra*]{}/HRC-I observations. [**Left panel:**]{} the Arches cluster and a sub-set of the detected foreground objects are indicated (only those stars which are known in Simbad are labeled). [**Right panel:**]{} the detected X-ray binaries as well as the possible VFXT. Also indicated is the complex X-ray emission around Sgr A\*. []{data-label="fig:images"}](4129f2b.ps "fig:"){width="50.00000%"}
![[**Left panel:**]{} the [*Chandra*]{}/HRC-I image of XMM J174457–2850.3 and GRS 1741.9–2853 as obtained on June 5, 2005 (field GC-2), [**Right panel:**]{} the [*Chandra*]{}/ACIS-I image of both sources obtained on July 1, 2005. To show the two sources most clearly we have rebinned the images to a resolution of $\sim$2$''$. The apparent extended nature of the sources is due to their relatively large offset positions with respect to the pointing direction of the satellite.[]{data-label="fig:extra_image"}](4129fig3.ps){width=".45\textwidth"}
![The light curves of GRS 1741.9–2853 (top panel) and XMM J174457–2850.3 (bottom panel). In both panels, the triangles indicate the new data reported in this paper. In the top panel, the squares and the upper limits are taken from [@2005ApJ...622L.113M] (see this paper for the exact energy ranges for each point; our [*Chandra*]{} luminosities are for 2–8 keV). In the bottom panel, the squares and the upper limit are from [@2005MNRAS.357.1211S] and all luminosities are for 2–10 keV.[]{data-label="fig:lc"}](4129fig5.ps){width="45.00000%"}
Distribution of transients in the FOV \[section:appendix\]
==========================================================
In Table \[table:sources\_in\_FOV\] we have listed the known X-ray transients and persistent X-ray binaries which are in the FOV of our [*XMM-Newton*]{} and [*Chandra*]{} monitoring observations. This table also shows the distance of the sources with respect to Sgr A\* and the maximum reported X-ray luminosity for these sources as found in the literature. We have converted the luminosities to the 2–10 keV energy range using the reported spectral parameters of the sources. From this table and from Figure \[fig:Lx\_v\_distance\] it can clearly be seen that only one bright to very bright transient is present (1A 1742–289) within a distance of $<$$15'$ from Sgr A\*, but eight VFXTs[^9] and two faint X-ray transients (although the two faint transients are just barely brighter than $10^{36}$ erg s$^{-1}$ and they could just be the brightest VFXTs in the FOV). Clearly, the number of VFXTs within 15$'$ of Sgr A\* is significantly larger than that of the brighter transients. When extending the region out to 25$'$, we see eight VFXTs, three faint transients and only 3 bright to very bright transients. Note, however, that the region $>$10$'$ away from Sgr A\* has been less sampled with sensitive X-ray instruments than closer to Sgr A\*, meaning that the number of VFXTs within $25'$ of Sgr A\* is likely to grow thanks to our monitoring program.
This strongly suggests that the number density of VFXTs is indeed significantly higher than that of the brighter transients. The fact that the brighter systems were until recently much easier to detect than the fainter systems means that this discrepancy in number densities will only become larger in the future. However, one VFXT (CXOGC J174540.0–290031) and one faint transient (AX J1745.6–2901) have exhibited eclipses and there is strong evidence ([@2005ApJ...633..228M]) that at least CXOGC J174540.0–290031 was intrinsically much brighter ($>$10$^{36}$ erg s$^{-1}$) than what we observe because the inner part of the system is blocked from our line of sight and we only observe the scattered (e.g., in a corona) X-ray emission from the system, making it artificially seem very faint. One can argue that this holds true for the other eclipsing source and possibly for all VFXTs for which we have not yet observed the eclipses or the X-ray dips associated with high inclination. If we conservatively assume that sources with inclinations in the range of 60$^\circ$–90$^\circ$ could appear as VFXTs, then a random distribution of orbital inclinations should give roughly equal numbers of bright transients and VFXTs (see also [@kingwijnands2005]; the solid angle is proportional to the cosine of the inclination). Clearly, the distribution of observed transients in the FOV of our monitoring campaign (see Tab. \[table:sources\_in\_FOV\]) shows a lack of brighter systems, even when also considering the two persistent X-ray binaries.
Moreover, strong evidence exists that for the known dippers and eclipsing X-ray binaries in other parts of the Galaxy we [*do*]{} directly look at the inner part of the systems and not just indirectly via scattering (e.g., due to the detection of kHz QPOs or nearly coherent oscillations during type-I X-ray bursts in several high inclination sources; e.g., ). Therefore, the observed X-ray luminosity is indeed the intrinsic luminosity of these systems. Only for the so-called accretion-disk-corona sources (which have the highest inclination of all systems) do we have evidence that only the scattered emission is seen making these systems appear fainter than they are intrinsically. The inclination range for such sources is significantly smaller than what we used above, thus making the problem even worse. Clearly, it is very likely that most of the VFXTs do not appear very faint as a result of inclination effects but rather that they are intrinsically very faint.
The fact that VFXTs seem to be overabundant close to the Galactic center compared to brighter systems might also have consequences for certain types of models for the VFXTs. For example, [@kingwijnands2005] discussed briefly the possibility that the VFXTs harbor neutron stars and black holes which accrete from the weak wind of low-mass companion stars resulting in very faint outbursts. However, they argued that these systems must eventually evolve into brighter states because the companion stars will fill their Roche lobes at a certain time in the future. If indeed, as they suggested, these brighter states have longer durations than the wind accreting states, then the lack of brighter systems present among the very faint ones would suggest that this model cannot explain the nature of the VFXTs.
[l c l l c c]{} Source name & Distance from Sgr A\* & Maximum luminosity[^10] & Classification & Comment & Reference[^11]\
& ($'$) & (erg s$^{-1}$) & & &\
\
CXOGC J174540.0-290031 & 0.05 & $1\times 10^{35}$ & very faint & eclipser, radio counterpart & 1\
CXOGC J174541.0-290014 & 0.31 & $6\times 10^{33}$ & very faint & & 1\
CXOGC J174540.0-290005 & 0.37 & $4\times 10^{34}$ & very faint & & 1\
CXOGC J174538.0-290022 & 0.44 & $3\times 10^{34}$ & very faint & & 1\
1A 1742-289 & 0.92 & $7\times 10^{38}$ & very bright & radio counterpart & 2\
CXOGC J174535.5-290124[^12] & 1.35 & $4\times 10^{34}$ & very faint & & 1\
AX J1745.6-2901 & 1.37 & $2\times 10^{36}$ & faint & burster, eclipser & 3\
CXOGC J174554.3-285454 & 6.38 & $8\times 10^{34}$ & very faint & & 1\
GRS 1741.9-2853 & 10.00 & $2\times 10^{36}$ & faint & burster & 4\
XMM J174544-2913.0 & 12.56 & $5\times 10^{34}$ & very faint & & 5\
XMM J174457-2850.3 & 13.78 & $9\times 10^{35}$ & very faint & & 6\
SAX J1747.0-2853 & 19.55 & $4\times 10^{37}$ & bright & burster & 7\
GRO J1744-28 & 21.71 & $3\times 10^{38}$ & bright & X-ray pulsar & 8\
KS 1741-293 & 22.09 & $5\times 10^{36}$ & faint & & 9\
\
1E 1743.1-2843 &18.99 & $3\times 10^{36}$ & faint & & 10\
1A 1742-294 &30.95 & $1\times 10^{37}$ & bright & burster & 11\
![The observed maximum X-ray luminosities (2–10 keV) of the X-ray transients in our FOV as a function of the distance of the transients to Sgr A\*. \[fig:Lx\_v\_distance\] ](4129fA1.ps){width="45.00000%"}
[^1]: It is interesting to note that all but one of the seven currently known accreting millisecond X-ray pulsars (see [@2004AIPC..714..209W] for a review) are faint X-ray transients. The one which is not a faint transient (XTE J1751–305; peak X-ray luminosities of a few times $10^{37}$ erg s$^{-1}$) had a very short outburst duration (e-folding time of only $\sim$7 days) resulting in a low time-averaged accretion rate similar to the other accreting millisecond pulsars.
[^2]: Available from http://cxc.harvard.edu/ciao/
[^3]: Available from http://cxc.harvard.edu/toolkit/pimms.jsp
[^4]: We note that this is likely an upper limit on the duty cycle since actual source detections are more frequently reported in the literature than non-detections which will skew the data toward detections. For example, the data presented by [@2003ApJ...598..474M] of GRS 1741.9–2853 (as used in Fig. \[fig:lc\]) does not report the non-detections of the source as seen with [*ROSAT*]{} () or [*BeppoSAX*]{} ([@1999ApJ...525..215S]; using the Narrow Field Instruments). Since no upper limits on the source flux are given in these papers, we also do not include these non-detections in Fig. \[fig:lc\]. \[footnote:duty\]
[^5]: Offset between the source position and the pointing position
[^6]: Count rates are for the full [*Chandra*]{}/HRC-I energy range (0.08–10 keV) or the 0.3–7 keV energy range for the [*Chandra*]{}/ACIS-I; they are background corrected, but are not corrected for offset.
[^7]: The errors on the coordinates are calculated using equation 5 in [@hong2005] with the addition of a 0.7$''$ pointing uncertainty, corresponding to 95% confidence levels.
[^8]: Parameter fixed to the value found by [@2005MNRAS.357.1211S].
[^9]: The VFXT CXOGC J174535.5–290124 is located in the error circle of the faint transient and eclipsing source AX J1745.6–2901. It is possible that both sources are the same one, which would reduce the number of VFXTs to seven although this would not affect the conclusions in this appendix. However, no eclipses were found for CXOGC J174535.5–290124 in the [*Chandra*]{} data available for this source (M. Muno 2005, private communication), making it less likely that both sources are the same one.
[^10]: These maximum observed luminosities are taken from the references and converted to a 2–10 keV luminosity for a distance of 8 kpc
[^11]: References: 1: [@2005ApJ...622L.113M]; 2: [@1976MNRAS.175P..47B]; 3: [@1996PASJ...48..417M]; 4: [@2003ApJ...598..474M]; 5: [@2005MNRAS.357.1211S]; 6: this paper; 7: ; 8: [@1996ApJ...469L..25G]; 9: [@1991AdSpR..11..187I]; 10: ; 11: [@2002ApJS..138...19S]
[^12]: This source is located in the error circle of AX J1745.6–2901 and it is possible that both are the same source. However, no eclipses were found for CXOGC J174535.5–290124 in the [*Chandra*]{} data available for this source (M. Muno 2005, private communication) making this identification less likely.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'With the dramatic increase of dimensions in the data representation, extracting latent low-dimensional features becomes of the utmost importance for efficient classification. Aiming at the problems of unclear margin representation and difficulty in revealing the data manifold structure in most of the existing linear discriminant methods, we propose a new discriminant feature extraction framework, namely Robust Locality-Aware Regression (RLAR). In our model, we introduce a retargeted regression to perform the marginal representation learning adaptively instead of using the general average inter-class margin. Besides, we formulate a new strategy for enhancing the local intra-class compactness of the data manifold, which can achieve the joint learning of locality-aware graph structure and desirable projection matrix. To alleviate the disturbance of outliers and prevent overfitting, we measure the regression term and locality-aware term together with the regularization term by the $L_{2,1}$ norm. Further, forcing the row sparsity on the projection matrix through the $L_{2,1}$ norm achieves the cooperation of feature selection and feature extraction. Then, we derive an effective iterative algorithm for solving the proposed model. The experimental results over a range of UCI data sets and other benchmark databases demonstrate that the proposed RLAR outperforms some state-of-the-art approaches.'
author:
- 'Liangchen Hu and Wensheng Zhang [^1] [^2] [^3] [^4]'
title: |
Robust Locality-Aware Regression\
for Labeled Data Classification
---
Locality-aware graph learning, margin representation learning, joint $L_{2,1}$-norms sparsity, feature selection and extraction.
Introduction
============
dimensionality leads to high storage overhead, heavy computation and huge time consumption in the training process of machine learning. And, as the ambient space expands exponentially with the increase of dimensionality, the proportion of training data in the whole data space drops sharply, thus resulting in the worse generalization of the training model [@Ladla2011; @Ma2018]. A significant way to address these issues is dimensionality reduction (DR) [@Passalis2018; @Pang2019], which transforms the original high-dimensional spatial data into a low-dimensional subspace by some resultful means. Feature selection (FS) and feature extraction (FE) are two main techniques for processing the DR problems of high-dimensional data [@Khalid2014]. FS aims at learning a way for eliminating redundant features in the original space, while FE gains compact features with stronger recognition through recombination of original features in the process of spatial transformation. By contrast, FE is a more powerful means.
Among the many FE methods, principal component analysis (PCA) [@Hotelling1933] and linear discriminant analysis (LDA) [@Fisher1936] are two commonly-used unsupervised and supervised techniques respectively. PCA maximizes the divergence of the mapped data without considering label information during training, while LDA makes the mapped samples have better classification performance based on discriminability. Unfortunately, the mean dependence of LDA makes it incapable of revealing the data manifold structure, thus resulting in poor classification performance. Otherwise, multiple clusters usually happens to be formed in the same class [@Sugiyama2007], such as odd-even classification of handwritten digits, multiple triggers of a single disease, etc., such data conforms to non-Gaussian distribution [@Boutemedjet2009; @Luo2018; @Li2017], which challenges LDA of Gaussian distribution hypothesis. Although improved for diversification purposes, many variants of LDA still inherit this defect due to the problem of unchanged mean dependence, for example, orthogonal LDA (OLDA) [@Ye2006], maximum margin criterion (MMC) [@Li2006], sparse uncorrelated LDA (SULDA) [@Zhang2016], robust LDA (RLDA) [@Zhao2019], robust sparse LDA (RSLDA) [@Wen2019], etc. To appropriately reduce the dimensionality of data and improve the computational efficiency while maintaining high classification performance, preserving the local manifold structure is crucial to success. Inspired by this, Sugiyama improved LDA’s scatter loss into the form of sample pairs by combining the idea of locality preserving projections (LPPs) [@He2004], namely local Fisher discriminant analysis (LFDA) [@Sugiyama2007]. Cai et al. [@Cai2007] proposed locality sensitive discriminant analysis (LSDA), which aims to mine the potential subspace in the way of perceiving the local geometric structure of data manifold, in which the nearby samples having the same label are close to each other instead of being far away from each other. Similarly, Nie et al. [@Nie2007] proposed neighborhood minmax projections (NMMPs) formulated by pairwise samples and derived an effective strategy for solving trace ratio optimization problem. Besides, Fan et al. [@Fan2011] presented an improved LDA framework, the local LDA (LLDA), which can perform well without satisfying the assumption of Gaussian distribution.
Considering the aforementioned methods in a unified graph embedding framework, they differ in the way of constructing the graphs within and between classes, including the connection and affinity of the sample pairs. In the view of data manifold recovery [@Seung2000; @Belkin2002], the sample relationships in the intra-class graph constructed in the original space are expected to be preserved completely in the low-dimensional embedding space. However, the linear projections used in these methods make nonlinear manifold recovery almost impossible. In other words, the intra-class graph constructed in the original space is not optimal. Therefore, some new strategies have emerged to try to learn the optimal intra-class graphs while learning the optimal projections. On the premise that the affinity of intra-class samples satisfies the quadratic condition, Li et al. [@Li2017] studied the full-connection intra-class graph, and proposed locality adaptive discriminant analysis (LADA) which can well reveal the potential data manifold structure. Luo et al. [@Luo2018] proposed another adaptive discriminant analysis (ADA), which formulates the intra-class affinity loss in the form of heat kernel function and approximates it by quadratic model. Instead of investigating fully connected intra-class graph, Pang et al. [@Pang2019] aims at simultaneously learning neighborship and projection matrix (SLNP). Recently, Nie et al. [@Nie2019] put forward submanifold-preserving discriminant analysis (SPDA) with an auto-optimized k-nearest neighbor (KNN) graph, which differs from SLNP in that it considers only the connection of nearby samples. The data manifold has the property of local Euclidean homeomorphism, which makes the partial connected graph superior to the fully connected graph in revealing the manifold structure. Besides, considering only the connection information and not the affinity between samples, the model is easily affected by outliers.
Aside from preserving the intra-class structure, inter-class separability is also an indispensable part of achieving better classification performance. Establishing an effective margin representation facilitates the separability between classes. Here, we classify the commonly-used margin representations in the current mainstream DR strategies or classifiers into four categories, namely, average margin, weighted average margin, maximum margin and induced margin. Maximizing the inter-class scatter or global scatter in LDA and some of its variants is the pursuit of average margin. Some DR methods based on manifold learning, such as LSDA and stable orthogonal local discriminant embedding (SOLDE) [@Gao2013], with the construction of inter-class graph, aim to achieve the weighted average margin. The maximum margin is typically used in the design of classifiers, such as the maximum margin hyperplane determined by the support vectors in support vector machines (SVMs) [@Brereton2010; @Hsu2002]. Besides, least-squares regression (LSR), as a fundamental tool in statistics, can also be regarded as a strategy of margin representation. The purpose of margin representation learning can be achieved by guiding the samples of different classes towards disparate pre-set targets. Here, we define such an margin as the induced margin. Over the past decades, various regression analysis methods have been developed, such as ridge regression (RR) [@Hoerl1970], lasso regression [@Tibshirani1996], elastic net regression [@Zou2005], generalized robust regression (GRR) [@Lai2019], adaptive locality preserving regression (ALPR) [@Wen2020] and some kernel based regression methods [@Rosipal2001; @Liu2009; @Chen2016]. Most of these methods take the zero-one labels as the regression targets. However, the strict zero-one targets are too harsh on the marginal representation to yield superior classification performance. To remedy this deficiency, the trend is to learn relaxed regression targets instead of the original zero-one targets, with some representative methods such as discriminative LSR (DLSR) [@Xiang2012], retargeted LSR (ReLSR) [@Zhang2015], and groupwise ReLSR [@Wang2018]. Among them, ReLSR has been used for the marginally structured representation learning (MSRL) [@Zhang2018] and has been successful.
In real life, the collected data may be doped with some noise or outliers. However, in the conventional FE methods, the squared $L_2$ norm, which tends to enlarge the influence of outliers, is generally employed to measure the loss. Of course, we can mitigate this risk by measuring angles instead of distances, as the angle linear discriminant embedding (ALDE) [@Liu2015] does. Besides, to suppress the sensitivity of squared $L_2$ norm, some new evaluation criteria based on the $L_1$ norm are combined on PCA and LDA, including $L_{1}$-PCA [@Ke2005], $R_1$-PCA [@Ding2006], PCA-$L_{1}$ [@Kwak2008], LDA-$R_1$ [@Li2010], sparse discriminant analysis (SDA) [@Clemmensen2011], LDA-$L_1$ [@Zhong2013; @Wang2014], etc. Since the optimization of $L_1$-norm-based loss function is relatively troublesome, the more efficient $L_{2,1}$ norm gradually attracts the attention of researchers. By imposing the $L_{2,1}$ norm on both the regression loss term and the regularization term, Nie et al. [@Nie2010] proposed an efficient and robust feature selection method (RFS). Inspired by this, some new formulations of PCA and LDA with $L_{2,1}$ norm have been proposed successively, and joint sparse PCA (JSPCA) [@Yi2017], adaptive weighted sparse PCA (AW-SPCA) [@Yi2019], L21SDA [@Shi2014] and RLDA [@Zhao2019] are the representative ones. Unfortunately, RLDA only employs the $L_{2,1}$ norm for the measure of intra-class scatter, thus resulting in limited effect on suppressing outliers.
Taken together, preserving the intrinsic structure of the data manifold and forcing the separation between classes are crucial to generalization and classification performance of feature extraction. Combining with the excellent properties of $L_{2,1}$ norm on resisting outliers and removing redundant features, we propose a novel discriminant feature extraction algorithm with flexible learning of intra-class structure and margin representation, which has the following advantages.
1. By integrating locality-aware graph learning and flexible margin representation learning, we build a new discriminant learning criterion, which enhances intra-class compactness while allowing for flexible intra-class and inter-class differences.
2. The flexible locality-aware structural learning strategy formulated by us is capable of revealing the local adjacency structure of the intra-class samples in the desired subspace.
3. Joint $L_{2,1}$ norms in the proposed model can not only relieve the pressure brought by outliers or noise, but also conduct feature selection and subspace learning simultaneously.
4. We theoretically prove the convergence of the proposed algorithm and experimentally verify its superior classification and generalization performance on multiple databases.
The remainder of our paper is organized as follows. In Section \[related\_work\], we briefly describe some notations and review some of the work. In Section \[methodology\], we elaborate the process of establishing the model and derive an effective algorithm. In Section \[algorithm\_analysis\], we provide theoretical analysis of the proposed algorithm, including convergence proof and computational complexity analysis. In Section \[experiments\], we investigate the performance of the proposed algorithm through a series of comparative experiments. Section \[conclusion\] concludes the paper with some additional summary.
Related Work {#related_work}
============
Briefly, some notations in our writing are described in this section. Throughout, all the vectors and matrices we defined are in bold italics, and all other variables are in single italics. Given a data matrix $\bm X\!=\![\bm X_1,\bm X_2,\cdots,\bm X_n]\!\in\!\mathbb{R}^{d\times n}$, where $\bm X_i$ refers to a sample instance. Assuming that $\bm X$ can be classified into $c$ classes, we write $\bm X^i\!=\![\bm X^i_1,\bm X^i_2,\cdots,\bm X^i_{n_i}]\!\in\!\mathbb{R}^{d\times n_i}$ as the data submatrix of class $i$. Moreover, $\bm X^T$, $\bm X^{-1}$ and $Tr(\bm X)$ represent $\bm X$’s transpose, inverse and trace, respectively. And, we denote the matrix $[\bm A;\bm B]\!\in\!\mathbb{R}^{(p+q)\times t}$ as the composition of $\bm A$ and $\bm B$, where $\bm A\!\in\!\mathbb{R}^{p\times t}$ and $\bm B\!\in\!\mathbb{R}^{q\times t}$. Some commonly-used norms, such as the Frobenius norm, $L_2$ norm, and $L_{2,1}$ norm, are defined as $\|\bm X\|_F=\sqrt{\sum_{i,j}X_{ij}^2}$, $\|\bm X_{i,:}\|_2=\sqrt{\sum_jX_{ij}^2}$, and $\|\bm X\|_{2,1}=\sum_i\|\bm X_{i,:}\|_2$, respectively. Using these notations, we redescribe several of the work related to our research below.
LDA
---
LDA projects the high-dimensional data $\bm X$ into the low-dimensional latent space through a linear mapping $f(\bm X)=\bm W^T\bm X$, which aggregates the intra-class samples to the intra-class mean while maximizing the discrepancy between the inter-class means as follows $$\label{eq:LDA}
\max_{\bm W}\frac{\sum_{i=1}^cn_i\left\|\bm W^T(\bm M_i-\bm M)\right\|_2^2}{\sum_{i=1}^c\sum_{j=1}^{n_i}\left\|\bm W^T(\bm X^i_j-\bm M_i)\right\|_2^2}$$ where $\bm M_i=\frac{1}{n_i}\sum_{j=1}^{n_i}\bm X^i_j$ refers to the intra-class mean of class $i$ and $\bm M\!=\!\frac{1}{n}\sum_{i=1}^n\bm X_i$ denotes the mean of all the samples.
Denoting the intra-class scatter matrix and inter-class scatter matrix as $\bm S_w\!=\!\sum_{i=1}^c\sum_{j=1}^{n_i}(\bm X^i_j-\bm M_i)(\bm X^i_j-\bm M_i)^T$ and $\bm S_b\!=\!\sum_{i=1}^cn_i(\bm M_i-\bm M)(\bm M_i-\bm M)^T$ respectively, we can rewrite Eq. (\[eq:LDA\]) as the following trace ratio problem $$\label{eq:LDA:trace}
\max_{\bm W}\frac{Tr(\bm W^T\bm S_b\bm W)}{Tr(\bm W^T\bm S_w\bm W)}.$$ Solving problem (\[eq:LDA:trace\]) is equivalent to solving the following generalized eigendecomposition problem $$\label{eq:LDA:solution}
\bm S_b\bm W=\bm S_w\bm W\bm \Lambda.$$ Since Eq. (\[eq:LDA:solution\]) involves matrix inverse, LDA suffers from the small-sample-size problem [@Raudys1991]. From the definitions of intra-class scatter and inter-class scatter, LDA treats each sample equally, which causes the samples far from the mean to have a greater impact on the model. Moreover, LDA can handle the non-gaussian data [@Li2017; @Luo2018] hardly because of the mean dependence.
LSR and ReLSR
-------------
Here, we briefly review the classical least squares regression (LSR) model with a class indicator matrix $\bm Y\!=\![\bm Y_1,\bm Y_2,\cdots,\bm Y_n]^T\!\in\!\mathbb{R}^{n\times c}$ which is assigned to the data matrix $\bm X$. Assuming that the linear mapping between the row vectors in $\bm Y$ and the column vectors in $\bm X$ is $\bm Y_i=\bm W^T\bm X_i+\bm b$ where $\bm W\in\mathbb{R}^{d\times c}$ refers to the regression matrix and $\bm b\in\mathbb{R}^{c\times 1}$ is a bias vector, we can obtain unbiased estimates of $\bm W$ and $\bm b$ by solving the following objective function $$\label{eq:LSR}
\min_{\bm W,\bm b}\sum_{i=1}^n\left\|\bm W^T\bm X_i+\bm b-\bm Y_i\right\|_2^2.$$ Conventionally, the indicator matrix $\bm Y$ is a strict zero-one matrix in which only the $l_i$-th entry of each row is one, where $l_i\!\in\!\{1,2,\cdots,c\}$ is the class label of sample $\bm X_i$. In reality, however, due to the diversity of data sampled from various distributions, strict zero-one indicators do not make sense and may be detrimental to classification.
To overcome this problem, Zhang et al. proposed the retargeted least squares regression (ReLSR) [@Zhang2015], a method of learning targets flexibly from regression results, while maintaining a certain discriminant power. The joint learning framework of linear mapping and regression target of ReLSR is as follows $$\label{eq:ReLSR}
\begin{split}
&\min_{\bm W,\bm b,\bm T}\left\|\bm X^T\bm W+\bm 1_n\bm b^T-\bm T\right\|_F^2+\beta\|\bm W\|_F^2\\
&s.t.~T_{il_i}-\max_{j\neq l_i}T_{ij}\geq 1, i=1,2,\cdots,n
\end{split}$$ where $\bm 1_n\!\in\!\mathbb{R}^{n\times 1}$ refers to a column vector with all 1s, $\bm T\in\mathbb{R}^{n\times c}$ represents a retargeted matrix and $\beta$ is a regularization parameter that controls the degree of bias. Actually, (\[eq:ReLSR\]) can be regarded as a ridge regression (RR) [@Hoerl1970] of adaptive learning regression targets with a margin constraint on the true class and the most likely false class. By solving problem (\[eq:ReLSR\]), we can achieve a more accurate classification than solving problem (\[eq:LSR\]).
RFS
---
Besides the inflexible predefined targets, another drawback of LSR is that its loss function is in the form of squared Frobenius norm, which inevitably results in sensitivity to outliers or noise. To address this issue, Nie et al. proposed an efficient and robust feature selection (RFS) method [@Nie2010], which avoids the dilemma by jointly minimizing the $L_{2,1}$-norms. The objective function of RFS with our notations can be written as follows $$\label{eq:RFS}
\min_{\tilde{\bm W}}\frac{1}{\gamma}\left\|\tilde{\bm X}^T\tilde{\bm W}-\bm Y\right\|_{2,1}+\|\tilde{\bm W}\|_{2,1}$$ where $\tilde{\bm W}$ absorbing the offset vector $\bm b$ is denoted as $[\bm b^T;\bm W]$, and correspondingly $\tilde{\bm X}\!=\![\bm 1_n^T;\bm X]$. It can be seen from Eq. (\[eq:RFS\]) that simultaneously utilizing $L_{2,1}$-norms on both loss function and regularization term can not only resist outliers, but also jointly induce the row sparsity of regression matrix.
With $\bm E=\frac{1}{\gamma}(\tilde{\bm X}^T\tilde{\bm W}-\bm Y)$, $\bm A=[\tilde{\bm X}^T~\gamma \bm I]$ and $\bm U=[\bm W;\bm E]$, rewriting the above problem (\[eq:RFS\]) as $$\label{eq:re:RFS}
\min_{\bm U}\|\bm U\|_{2,1}~~~~s.t.~~~\bm A\bm U=\bm Y.$$ The constrained problems similar to Eq. (\[eq:re:RFS\]) can easily derive a closed-form solution by using the Lagrange multiplier method.
Methodology
===========
In this section, we analyze the irrationality of LDA optimization criteria, and establish more discriminative optimization criteria on the premise of ensuring higher generalization performance by replacing intra-class scatter and inter-class scatter. Besides, to mitigate the interference caused by outliers or noises, we focus on minimizing the joint $L_{2,1}$-norms of all modules in our model. Then, we deduce the process of model optimization and condense an effective algorithm.
Proposed Model
--------------
As it involves the performance of classification and generalization, there are two criteria for evaluating the quality of supervised dimensionality reduction, one is the maintenance of the intra-class structural information, the other is the preservation of the inter-class separability. Although LDA has both of these criteria, namely minimizing intra-class scatter and maximizing inter-class scatter, there are obvious deficiencies, as analyzed below.
Reformulates the loss function of the intra-class scatter as the form of sample pair as follows $$\label{eq:LDA:intraclass}
\begin{split}
&\sum_{i=1}^c\sum_{j=1}^{n_i}\left\|\bm W^T\left(\bm X^i_j-\bm M_i\right)\right\|_2^2\\
=&\sum_{i=1}^c\sum_{j,k=1}^{n_i}\!S_{jk}^i\!\left\|\bm W^T\bm X^i_j-\bm W^T\bm X^i_k\right\|_2^2
\end{split}$$ where $S_{jk}^i=\frac{1}{2n_i}$ refers to the similarity between embedded sample pairs (See appendix \[LDA:intra-class\] for details of proof). Obviously from Eq. (\[eq:LDA:intraclass\]), the fully connected intra-class graph is employed in LDA, and $S^i_{jk}$ of LDA is only connected with the number of intra-class samples, which means that there is no gap in the affinity between all intra-class embedded sample pairs. This is extremely unreasonable, because samples far from the intra-class mean contribute more to the change in the objective function.
Since non-orthogonal linear transformation cannot guarantee metric invariance, the affinity between sample pairs in the ambient high-dimensional space cannot be maintained in the embedded low-dimensional space. Here, we develop an adaptive locality-aware method for calculating the affinity of samples following the learning process of projection matrix, as shown below $$\label{eq:obj:1}
\begin{split}
&\min_{\bm V,\bm W}\sum_{i=1}^c\sum_{j,k=1}^{n_i}\frac{V_{jk}^i}{2K}\left\|\bm W^T\bm X^i_j-\bm W^T\bm X^i_k\right\|_2\\
&s.t.~\bm V_{j,:}^i\bm 1=K,\ V_{jk}^i\in\{0,1\}
\end{split}$$ where $V_{jk}^i$ refers to the connection indicator in the $j$-th KNN graph of class $i$. Instead of measuring the loss by the squared $L_2$ norm in (\[eq:LDA:intraclass\]), we can reduce the interference of outliers by directly using the $L_2$ norm in the modeling of (\[eq:obj:1\]). Although intuitively only the connections in the intra-class graph are optimized in (\[eq:obj:1\]), the sample affinity can be induced below. $$\label{eq:obj:11}
\begin{split}
&\min_{\bm S,\bm W}\sum_{i=1}^c\sum_{j,k=1}^{n_i}S_{jk}^i\left\|\bm W^T\bm X^i_j-\bm W^T\bm X^i_k\right\|_2^2\\
&s.t.~\|\bm S_{j,:}^i\|_0=K
\end{split}$$ where $\|\cdot\|_0$ denotes the number of non-zero entries in a vector, and $S_{jk}^i\!=\!\frac{1}{2K\|\bm W^T\bm X^i_j-\bm W^T\bm X^i_k\|_2}$ tends to penalize sample pairs that are far away from each other, which can be regarded as a significative definition of affinity between samples.
In terms of inter-class separability, LDA requires that the average margin between different classes should be maximally expanded, and classes far away from other classes will occupy a larger proportion. In other words, the definition of such inter-class separability makes different classes influenced by each other easily. The alternative strategy to eliminate this influence is to employ the margin representation in the form of regression. In the original regression, the strict zero-one target matrix cannot be approximated as an ideal low-dimensional embedding, thus we prefer the flexible formulation in problem (\[eq:ReLSR\]) which helps to realize the learning of margin representation. Combining the retargeted regression term and locality-aware term, we have the following learning model $$\label{eq:obj:2}
\begin{split}
&\min_{\bm W,\bm b,\bm V,\bm T}\left\|\bm X^T\bm W+\bm 1\bm b^T-\bm T\right\|_{F}^2\\
&~~~~~~~~~~~~~~+\beta\sum_{i=1}^c\sum_{j,k=1}^{n_i}S_{jk}^i\left\|\bm W^T\bm X_{j}^i-\bm W^T\bm X_{k}^i\right\|_2^2\\
&s.t.~T_{il_i}-\max_{j\neq l_i}T_{ij}\geq1,\|\bm S_{j,:}^i\|_0=K
\end{split}$$ where $\beta\!>\!0$ is a tradeoff coefficient.
In real life, the collected data often have higher dimensions and are mixed with various noises, which leads to the generation of many outliers and the appearance of redundant features. To alleviate the interference of outliers on the training process, we tend to employ $L_2$ norm instead of squared $L_2$ norm to measure the value of loss function. As indicated by problem (\[eq:RFS\]), for matrix variables, it should be $L_{2,1}$ norm instead of Frobenius norm. Besides, the projection matrix plays the role of feature loading and feature fusion. We can achieve the intention of feature selection by forcing the row sparsity of the projection matrix, which can be achieved by performing the $L_{2,1}$ norm [@Nie2010; @Lai2019; @Wen2020]. To sum up, we establish a unified learning framework, which covers the joint learning of adaptive graph structure, projection matrix with feature selection and margin representation. The objective function is as follows $$\label{eq:obj}
\begin{split}
&\min_{\bm W,\bm b,\bm V,\bm T}\left\|\bm X^T\bm W+\bm 1\bm b^T-\bm T\right\|_{2,1}+\alpha\left\|\bm W\right\|_{2,1}\\
&~~~~~~~~~~~~~+\beta\sum_{i=1}^c\sum_{j,k=1}^{n_i}\frac{V_{jk}^i}{2K}\left\|\bm W^T\bm X_{j}^i-\bm W^T\bm X_{k}^i\right\|_2\\
&s.t.~T_{il_i}-\max_{j\neq l_i}T_{ij}\geq1,\bm V_{j,:}^i\bm 1=K,\ V_{jk}^i\in\{0,1\}
\end{split}$$ where $\alpha\!>\!0$ is a regularization penalty parameter.
Optimization Strategy
---------------------
Obviously, since the regression term, regularization term and locality-aware term in (\[eq:obj\]) are all characterized by $L_{2,1}$ norm, problem (\[eq:obj\]) is a non-smooth optimization problem with multivariable coupling, which urgently needs us to obtain the optimal solution through the strategy of alternating iteration. In each iteration, we transform the problem into an smooth optimization problem that is jointly convex for all variables. The details are as follows.
Fix $\bm V$ and $\bm T$, Update $\bm W$ and $\bm b$: First, we re-formulate the problem (\[eq:obj\]) with Frobenius norm as $$\label{eq:re:obj}
\begin{split}
&\min_{\bm W,\bm b,\bm V,\bm T}\left\|\sqrt{\hat{\bm D}}\left(\bm X^T\bm W\!+\!\bm 1\bm b^T\!-\!\bm T\right)\right\|_F^2\!+\!\alpha\left\|\sqrt{\tilde{\bm D}}\bm W\right\|_F^2\\
&~~~~~~~~~~~+\beta\frac{1}{2}\sum_{i=1}^c\sum_{j,k=1}^{n_i}\frac{V_{jk}^i}{G_{jk}^i}\left\|\bm W^T\bm X_{j}^i\!-\!\bm W^T\bm X_{k}^i\right\|_2^2\\
\end{split}$$ where $\hat{\bm D}\in\mathbb{R}^{n\times n}$ and $\tilde{\bm D}\in\mathbb{R}^{d\times d}$ are diagonal matrices with $ii$-th entries $1/\|(\bm X^T\bm W+\bm 1\bm b^T-\bm T)_{i,:}\|_2$ and $1/\|\bm W_{i,:}\|_2$ respectively, $G^i_{jk}=K\|\bm W^T\bm X_{j}^i-\bm W^T\bm X_{k}^i\|_2$ and specifically, $\frac{0}{0}=0$. When $\bm S$ and $\bm T$ are known, and $\hat{\bm D}$, $\tilde{\bm D}$, and $\bm G$ are assumed to be constants, Eq. (\[eq:re:obj\]) can be written as the matrix trace optimization problem with respect to $\bm W$ and $\bm b$ is as follows $$\label{eq:re:tr:obj}
\begin{split}
&\min_{\bm W,\bm b}Tr\!\left(\!\left(\!\bm X^T\!\bm W\!+\!\bm 1\bm b^T\!-\!\bm T\!\right)^T\!\hat{\bm D}\!\left(\!\bm X^T\!\bm W\!+\!\bm 1\bm b^T\!-\!\bm T\!\right)\!\right)\!\\
&\!+\!\alpha Tr\!\left(\!\bm W^T\!\tilde{\bm D}\!\bm W\!\right)\!+\!\beta Tr\!\left(\!\bm W^T\!\bm X\!(\!\bm D\!-\!\frac{\bm S\!+\!\bm S^T}{2}\!)\!\bm X^T\bm\!\bm W\!\right)\!\\
\end{split}$$ where $\bm S\!=\!\bm V\oslash\bm G$ (Note that $\oslash$ is the element-wise division operator of matrices), $\bm D$ is a diagonal matrix with the $i$-th entry $D_{ii}=\sum_j(S_{ij}+S_{ji})/{2}$.
Taking the derivative of Eq. (\[eq:re:tr:obj\]) w.r.t. $\bm b$ and setting it to zero, we get $$\label{eq:b}
\begin{split}
&\bm W^T\bm X\hat{\bm D}\bm 1+\bm b\bm 1^T\hat{\bm D}\bm 1-\bm T^T\hat{\bm D}\bm 1=\bm 0\\
&\Rightarrow\bm b=\frac{\left(\bm T^T-\bm W^T\bm X\right)\hat{\bm D}\bm 1}{\bm 1^T\hat{\bm D}\bm 1}.
\end{split}$$ Then, similarly setting the derivative of $\bm W$ to zero, we arrive at $$\label{eq:der:W}
\begin{split}
&\bm X\!\hat{\bm D}\!\left(\!\bm X^T\!\bm W\!+\!\bm 1\bm b^T\!-\!\bm T\!\right)\!+\!\alpha\tilde{\bm D}\bm W\!+\!\beta\bm X\bm L\bm X^T\bm W=\bm 0\\
\end{split}$$ where $\bm L\!=\!\bm D-\frac{\bm S+\bm S^T}{2}$. Combining Eq. (\[eq:b\]) and Eq. (\[eq:der:W\]), we obtain the optimal $\bm W$ as follows $$\label{eq:W}
\bm W = \left(\bm X\bm H\bm X^T\!+\!\alpha\tilde{\bm D}\!+\!\beta\bm X\bm L\bm X^T\!\right)^{-1}\bm X\bm H\bm T$$ where $\bm H=\hat{\bm D}-\frac{\hat{\bm D}\bm 1\bm 1^T\hat{\bm D}}{\bm 1^T\hat{\bm D}\bm 1}$.
Fix $\bm W$, $\bm b$ and $\bm T$, Update $\bm V$: Since $\bm W$, $\bm b$ and $\bm T$ are fixed, Eq. (\[eq:obj\]) can be reduced to $$\label{eq:subobj:V}
\begin{split}
&\min_{\bm V}\sum_{i=1}^c\sum_{j,k=1}^{n_i}V_{jk}^i\left\|\bm W^T\bm X^i_j-\bm W^T\bm X^i_k\right\|_2\\
&s.t.~\bm V_{j,:}^i\bm 1=K,\ V_{jk}^i\in\{0,1\}
\end{split}$$ Eq. (\[eq:subobj:V\]) means that each subproblem of $i$ and $j$ is independent of each other. Then, the above problem can be further simplified to $$\label{eq:subobj:V2}
\begin{split}
&\min_{\bm V^i_{j,:}}\bm V^i_{j,:}{\bm G^i_{j,:}}^T~~s.t.~~\bm V_{j,:}^i\bm 1=K,\ V_{jk}^i\in\{0,1\},
\end{split}$$ where $\bm 1$ refers to a column vector with all entries 1. Accordingly, the optimal solution of Eq. (\[eq:subobj:V2\]) can be directly determined by $K$ non-zero minimum values in vector $\bm G^i_{j,:}$.
Fix $\bm V$, $\bm W$ and $\bm b$, Update $\bm T$: By fixing the regression matrix $\bm W$ and offset $\bm b$, Eq. (\[eq:obj\]) degenerates into a retargeting problem [@Zhang2015] $$\label{eq:T}
\begin{split}
&\min_{\bm T}\left\|\bm X^T\bm W+\bm 1\bm b^T-\bm T\right\|_{2,1}=\left\|\bm Y-\bm T\right\|_{2,1}\\
&s.t.~T_{il_i}-\max_{j\neq l_i}T_{ij}\geq1
\end{split}$$ where the regression result $\bm X^T\bm W+\bm 1\bm b^T$ is simply denoted as $\bm Y\in\mathbb{R}^{n\times c}$. As can be seen easily from Eq. (\[eq:T\]), there are $n$ mutually independent constrained convex subproblems, each of which is shown below $$\label{eq:sub:T}
\begin{split}
&\min_{\bm T_{i,:}}\left\|\bm Y_{i,:}-\bm T_{i,:}\right\|_{2}=\sqrt{\sum_{j=1}^c\left(Y_{ij}-T_{ij}\right)^2}\\
&s.t.~T_{il_i}-\max_{j\neq l_i}T_{ij}\geq1.
\end{split}$$
As in [@Zhang2015], we redefined the target vector $\bm T_{i,:}$ as follows $$\label{eq:re:T}
T_{ij}=\left\{\!\begin{array}{ll}
Y_{ij}+\triangle_i,&j=l_i \\
Y_{ij}+\min{(\triangle_i-v_j,0)}, &j\neq l_i
\end{array}
\right.$$ where $\triangle_i=T_{il_i}-Y_{il_i}$ is a step parameter, $v_j=Y_{ij}+1-Y_{il_i}$ is an indicator and $v_j\leq 0$ means that class $i$ and class $l_i$ satisfy the margin constraint. Using the new representation in Eq. (\[eq:re:T\]), we can rewrite optimization problem (\[eq:sub:T\]) as $$\label{eq:new:sub:T}
\min\Gamma(\triangle_i)=\sqrt{\triangle_i^2+\sum_{j\neq l_i}(\min{(\triangle_i-v_j,0)})^2}.$$ Denoting $\tau=\sqrt{\triangle_i^2+\sum_{j\neq l_i}(\min{(\triangle_i-v_j,0)})^2}$ and taking the derivative of Eq. (\[eq:new:sub:T\]) as follows $$\label{eq:der:sub:T}
\Gamma'(\triangle_i)=\frac{1}{\tau}\left(\triangle_i+\sum_{j\neq l_i}\left(\min{(\triangle_i-v_j,0)}\right)\right).$$ Obviously, when $\triangle_i+\sum_{j\neq l_i}\left(\min{(\triangle_i-v_j,0)}\right)=0$, problem (\[eq:new:sub:T\]) minimizes, and the optimal $\triangle_i$ is calculated as $$\label{eq:triangle}
\triangle_i=\frac{\sum_{j\neq l_i}v_j\Phi(v_j)}{1+\sum_{j\neq l_i}\Phi(v_j)}$$ where $\Phi(v_j)=\left\{\!\begin{array}{ll}
1,& \Gamma'(v_j)>0\\
0,& \text{other}
\end{array}
\right.$. Then, the optimal target matrix $\bm T$ can be derived from Eq. (\[eq:re:T\]).
Based on the above results, we develop an effective alternative iterative algorithm. The detailed steps are described in Algorithm \[alg:Framwork\].
Data matrix $\bm X\!\in\!\mathbb{R}^{d\!\times\!n}$, labels $\{l_i\}_{i=1}^n$, penalty parameter $\alpha$ and tradeoff coefficient $\beta$.\
Initialize $\hat{\bm D}\!\in\!\mathbb{R}^{n\!\times\!n}$, $\tilde{\bm D}\!\in\!\mathbb{R}^{d\!\times\!d}$ as the identity matrices.\
Initialize $G_{jk}^i\!=\!\|\bm X_{j}^i-\bm X_{k}^i\|_2$.\
Initialize target matrix $T_{ij}\!=\!\left\{\!\begin{array}{ll}
1,&j=l_i \\
0, &j\neq l_i
\end{array}\right.$. Update $V^i_{jk}$ by solving the problem (\[eq:subobj:V\]). Update the affinity matrix $\bm S=\bm V\oslash\bm G$. Calculate $\bm W=\left(\bm X\left(\bm H\!+\!\beta\bm L\right)\bm X^T\!+\!\alpha\tilde{\bm D}\!\right)^{-1}\!\bm X\bm H\bm T$, where $\bm H=\hat{\bm D}-\frac{\hat{\bm D}\bm 1\bm 1^T\hat{\bm D}}{\bm 1^T\hat{\bm D}\bm 1}$ and $\bm L=\!\bm D\!-\!\frac{\bm S+\bm S^T}{2}\!$. Calculate $\bm b=\frac{\left(\bm T^T-\bm W^T\bm X\right)\hat{\bm D}\bm 1}{\bm 1^T\hat{\bm D}\bm 1}$. Calculate $T_{ij}=\left\{\!\begin{array}{ll}
Y_{ij}+\triangle_i,&j=l_i \\
Y_{ij}+\min{(\triangle_i-v_j,0)}, &j\neq l_i
\end{array}\right.$, where $\bm Y=\bm X^T\bm W+\bm 1\bm b^T$, $\triangle_i=\frac{\sum_{j\neq l_i}v_j\Phi(v_j)}{1+\sum_{j\neq l_i}\Phi(v_j)}$ and $v_j=Y_{ij}+1-Y_{il_i}$. Update $\hat{D}_{ii}\!=\!\frac{1}{\|(\bm X^T\bm W+\bm 1\bm b^T-\bm T)_{i,:}\|_2+\varepsilon}$. Update $\tilde{D}_{ii}\!=\!\frac{1}{\|\bm W_{i,:}\|_2+\varepsilon}$. Update [$G_{jk}^i\overset{j\neq k}{=}\|\bm W^T\bm X_{j}^i-\bm W^T\bm X_{k}^i\|_2$]{}. $\bm W$, $\bm b$, $\bm V$, $\bm S$, $\bm T$.
Algorithm Analysis {#algorithm_analysis}
==================
Convergence Analysis
--------------------
\[lemma:con:ineq\] For any two sets of non-zero constants $\{a_i\}_{i=1}^n$ and $\{b_i\}_{i=1}^n$ where $a_i,b_i\in\mathbb{R}^{+}$, if the following inequality holds $$\sum_{i=1}^n\frac{a_i}{\sqrt{b_i}}\leq\sum_{i=1}^n\sqrt{b_i},$$ then we have $\sum_{i=1}^n\sqrt{a_i}\leq\sum_{i=1}^n\sqrt{b_i}$.
Combined with Cauchy inequality, the following result can be obtained $$\begin{split}
\sum_{i=1}^n\sqrt{a_i}&=\sum_{i=1}^n\frac{\sqrt{a_i}\sqrt{\sqrt{b_i}\sum_{j=1}^n\sqrt{b_j}}}{\sqrt{\sqrt{b_i}\sum_{j=1}^n\sqrt{b_j}}}\\
&\leq\sqrt{\sum_{i=1}^n\frac{a_i}{\sqrt{b_i}\sum_{j=1}^n\sqrt{b_j}}}\sqrt{\sum_{i=1}^n\sqrt{b_i}\sum_{j=1}^n\sqrt{b_j}}.
\end{split}$$ Since $\sqrt{\sum_{i=1}^n\frac{a_i}{\sqrt{b_i}\sum_{j=1}^n\sqrt{b_j}}}\leq1$ and $\sum_{i=1}^n\sqrt{b_i}>0$, then we have $$\sum_{i=1}^n\sqrt{a_i}\leq\sum_{i=1}^n\sqrt{b_i},$$ which completes the proof.
\[lemma:ineq\] For any two sets of $z$-dimensional non-zero vectors $\{\bm p_i\}_{i=1}^n$ and $\{\bm q_i\}_{i=1}^n$ where $\bm p_i,\bm q_i\in\mathbb{R}^{z\times 1}$, if the following inequality holds $$\sum_{i=1}^nk_i\frac{\|\bm p_i\|^2_2}{\|\bm q_i\|_2}\leq\sum_{i=1}^nk_i\|\bm q_i\|_2, k_i>0$$ then we have $\sum_{i=1}^nk_i\|\bm p_i\|_2\leq\sum_{i=1}^nk_i\|\bm q_i\|_2$.
By the definition of norm, we have $$\|\bm p_i\|_2=\sqrt{\sum_{j=1}^zp_{ij}^2}, \|\bm q_i\|_2=\sqrt{\sum_{j=1}^zq_{ij}^2}.$$ Just set $a_i=k_i^2\sum_{j=1}^zp_{ij}^2$, $b_i=k_i^2\sum_{j=1}^zq_{ij}^2$, and then the conclusion in Lemma \[lemma:con:ineq\] is easy to follow.
\[convergence\_theorem\] Algorithm \[alg:Framwork\] monotonically decreases the value of the objective function in Eq. (\[eq:obj\]) in each iteration, and ultimately converges to the local optimal solution.
We refer to $F(\bm W^t,\bm b^t,\bm V^t,\bm T^t,\hat{\bm D}^t,\tilde{\bm D}^t,\bm G^t)$ as the objective function of problem (\[eq:re:obj\]) at $t$-th iteration. Then, by solving subproblem (\[eq:re:tr:obj\]), we arrive at $$\label{eq:WbT:t+1}
\begin{split}
&F(\bm W^{t+1},\bm b^{t+1},\bm V^t,\bm T^t,\hat{\bm D}^t,\tilde{\bm D}^t,\bm G^t)\\
&\leq F(\bm W^t,\bm b^t,\bm V^t,\bm T^t,\hat{\bm D}^t,\tilde{\bm D}^t,\bm G^t).
\end{split}$$ Rewriting (\[eq:WbT:t+1\]) in the form of 2-norms will yield $$\label{eq:Wb:ineq}
\begin{split}
&\sum_{i=1}^n\!\frac{\|(\bm X^T\bm W^{t\!+\!1}\!+\!\bm 1{\bm b^{t\!+\!1}}^T\!-\!\bm T^t)_{i,:}\|_2^2}{\|(\bm X^T\bm W^{t}\!+\!\bm 1{\bm b^{t}}^T\!-\!\bm T^t)_{i,:}\|_2}\!+\!\alpha\sum_{i=1}^d\!\frac{\|\bm W^{t\!+\!1}_{i,:}\|_2^2}{\|\bm W^{t}_{i,:}\|_2}\\
&~~+\beta\sum_{i=1}^c\sum_{j,k=1}^{n_i}\frac{{V_{jk}^i}^t}{2K}\frac{\|{\bm W^{t+1}}^T\bm X_{j}^i-{\bm W^{t+1}}^T\bm X_{k}^i\|_2^2}{\|{\bm W^{t}}^T\bm X_{j}^i-{\bm W^{t}}^T\bm X_{k}^i\|_2}\\
&\leq\sum_{i=1}^n{\|(\bm X^T\bm W^{t}\!+\!\bm 1{\bm b^{t}}^T\!-\!\bm T^t)_{i,:}\|_2}\!+\!\alpha\sum_{i=1}^d{\|\bm W^{t}_{i,:}\|_2}\\
&~~~~+\beta\sum_{i=1}^c\sum_{j,k=1}^{n_i}\frac{{V_{jk}^i}^t}{2K}{\|{\bm W^{t}}^T\bm X_{j}^i\!-\!{\bm W^{t}}^T\bm X_{k}^i\|_2}.
\end{split}$$ Obviously, inequality (\[eq:Wb:ineq\]) satisfies the conditions in Lemma \[lemma:ineq\], and then we have $$\label{eq:Wb:ineq:2norm}
\begin{split}
&\sum_{i=1}^n{\|(\bm X^T\bm W^{t\!+\!1}\!+\!\bm 1{\bm b^{t\!+\!1}}^T\!-\!\bm T^t)_{i,:}\|_2}\!+\!\alpha\sum_{i=1}^d{\|\bm W^{t\!+\!1}_{i,:}\|_2}\\
&~~~+\beta\sum_{i=1}^c\sum_{j,k=1}^{n_i}\frac{{V_{jk}^i}^t}{2K}{\|{\bm W^{t+1}}^T\bm X_{j}^i\!-\!{\bm W^{t+1}}^T\bm X_{k}^i\|_2}\\
&\leq\sum_{i=1}^n{\|(\bm X^T\bm W^{t}\!+\!\bm 1{\bm b^{t}}^T\!-\!\bm T^t)_{i,:}\|_2}\!+\!\alpha\sum_{i=1}^d{\|\bm W^{t}_{i,:}\|_2}\\
&~~~+\beta\sum_{i=1}^c\sum_{j,k=1}^{n_i}\frac{{V_{jk}^i}^t}{2K}{\|{\bm W^{t}}^T\bm X_{j}^i\!-\!{\bm W^{t}}^T\bm X_{k}^i\|_2}.
\end{split}$$ From (\[eq:Wb:ineq:2norm\]), we easily know that $$\label{eq:WbDG:t+1}
\begin{split}
&F(\bm W^{t+1},\bm b^{t+1},\bm V^t,\bm T^t,\hat{\bm D}^{t+1},\tilde{\bm D}^{t+1},\bm G^{t+1})\\
&\leq F(\bm W^t,\bm b^t,\bm V^t,\bm T^t,\hat{\bm D}^t,\tilde{\bm D}^t,\bm G^t).
\end{split}$$ Furthermore, by solving subproblems (\[eq:subobj:V\]) and (\[eq:T\]), we obtain the following result $$\label{eq:V:t+1}
\begin{split}
&F(\bm W^{t+1},\bm b^{t+1},\bm V^{t+1},\bm T^{t+1},\hat{\bm D}^{t+1},\tilde{\bm D}^{t+1},\bm G^{t+1})\\
&\leq F(\bm W^{t+1},\bm b^{t+1},\bm V^{t+1},\bm T^t,\hat{\bm D}^{t+1},\tilde{\bm D}^{t+1},\bm G^{t+1})\\
&\leq F(\bm W^{t+1},\bm b^{t+1},\bm V^t,\bm T^t,\hat{\bm D}^{t+1},\tilde{\bm D}^{t+1},\bm G^{t+1}).
\end{split}$$ Combined with (\[eq:WbDG:t+1\]) and (\[eq:V:t+1\]), the final result holds $$\label{eq:final:t+1}
\begin{split}
&F(\bm W^{t+1},\bm b^{t+1},\bm V^{t+1},\bm T^{t+1},\hat{\bm D}^{t+1},\tilde{\bm D}^{t+1},\bm G^{t+1})\\
&\leq F(\bm W^{t},\bm b^{t},\bm V^t,\bm T^{t},\hat{\bm D}^{t},\tilde{\bm D}^{t},\bm G^{t}).
\end{split}$$
It is easy to conclude that, following the update step of each variable in Algorithm \[alg:Framwork\], the value of the modeled objective function (\[eq:obj\]) decreases monotonically with the increase of the number of iterations, and finally converges to a local optimal solution.
Computational Complexity Analysis
---------------------------------
The computational complexity of each step of the proposed algorithm RLAR is roughly estimated below. The complexity in step 2 of Algorithm \[alg:Framwork\] is less than $O(n^2d)$. The calculation of affinity matrix $\bm S$ in step 6 is at most $O(nK)$. Since $d\geq c$, updating $\bm W$ in step 7 is, at most, of order $O(nd^2+d^3)$. Updating $\bm b$ in step 8 takes $O(ndc)$. In step 9, calculating $\bm Y$ costs $O(ndc)$, and then calculating $\bm T$ costs $O(nc)$. It costs $O(nc)$ and $O(dc)$ to calculate $\hat{\bm D}$ and $\tilde{\bm D}$, respectively. Finally, updating $\bm G$ in step 12 is less than $O(n^2c)$. In summary, assuming the algorithm performs $t$ iterations, then the total cost of our RLAR is of order $O(n^2d\!+\!(nd^2\!+\!d^3\!+\!n^2c)t)$ at most. The experimental results show that the RLAR can converge in less iterations. Thus, for large-scale data with lower dimensions, our computational complexity is acceptable.
Experimental Results {#experiments}
====================
In this section, we investigate the performance of the proposed RLAR in terms of classification and robustness by comparing it with some state-of-the-art approaches performed on ten publicly benchmark databases. Besides, some visualization results, parameter sensitivity analysis and convergence study are employed to further evaluate the effectiveness of the proposed method.
Databases Instances Features Classes
------------- ----------- ---------- ---------
Dermatology 366 34 6
Diabetes 768 8 2
Ionosphere 351 34 2
Iris 150 4 3
Wine 178 13 3
Binalpha 1404 320 36
YaleB 2414 1024 38
AR 1400 1024 100
COIL20 1440 1024 20
Caltech101 9144 3000 102
: Brief description of all databases for classification
\[tab:data:description\]
Experimental Settings
---------------------
The databases involved in our experimental comparisons are from a variety of scenarios to highlight the adaptability of our proposed RLAR, including five UCI data sets and five relatively large-scale databases. Brief information about these databases is described in Table \[tab:data:description\]. All features of these data are normalized prior to the experiments. And, we repeat each experiment for 10 trials independently with different random splits of training and test data, and then record the mean accuracy and the standard deviation.
All the participating FE algorithms consist of some representative discriminant algorithms, excellent regression algorithms and manifold-inspired algorithms, specifically including RR [@Hoerl1970], LDA [@Fisher1936], MMC [@Li2006], LSDA [@Cai2007], LFDA [@Sugiyama2007], NMMP [@Nie2007], SDA [@Clemmensen2011], SULDA [@Zhang2016], L21SDA [@Shi2014], ALDE [@Liu2015], ReLSR [@Zhang2015], MPDA [@Zhou2017], RSLDA [@Wen2019] and ALPR [@Wen2020]. And, we perform cross validation to search the best parameters for each algorithm or directly accept the suggested default parameter settings. To be fair, the resulting feature dimension achieved by all algorithms is uniformly set to $c$, except for some LDA-based algorithms that only reduce dimension to $c-1$ at most. For the compact representations produced by running these algorithms, we simply utilize the 1-NN classifier to evaluate the classification performance.
For the number of neighbors used in locality-aware structure learning, we simply set it to 7 for the data split with a training sample size greater than 10 per class and 3 for other cases. Besides, our approach determines two hyper-parameters $\alpha$ and $\beta$ by searching from the grid coordinate set $\{0.001,0.005,0.01,0.05,0.1,0.5,1 ,10,100,1000\}$. And we cover the searching process in details later in the subsection of parameter sensitivity analysis. The termination condition of our algorithm, including all algorithms involved in iterative optimization, is uniformly set to $30$ iterations.
Classification Performance Evaluation
-------------------------------------
### UCI Classification
In the experiments conducted here, we employ five small-scale data sets taken from the UCI Machine Learning Repository , namely ’Dermatology’, ’Diabetes’, ’Ionosphere’, ’Iris’ and ’Wine’. These data sets belong to different domains, which helps verify the universality of our method. We randomly assign 20% of the samples from these databases to the training set and 80% to the test set, and repeat the process 10 times. The experimental results are listed in Table \[tab:UCI\]. And the best results are marked in bold.
Although some comparison methods have achieved superior performance on some data sets, they are not superior in all cases. On these data sets, working well in all of comparisons indicates that our model has strong universality and high efficiency. Meanwhile, achieving a relatively small standard deviation in most cases also suggests that our method is somewhat stable. To some extent, the comparisons with other discriminant methods proves that our method has stronger ability to extract discriminant information.
--------------------- ----------- ------ ----------- ------ ----------- ------ ----------- ------ ----------- ------
\[4\][\*]{}[Alg.]{}
ACC Std ACC Std ACC Std ACC Std ACC Std
RR 95.51 1.14 55.83 1.86 86.19 2.33 96.00 1.56 64.48 5.42
LDA 94.39 0.97 55.42 1.43 83.88 2.48 95.58 1.31 88.41 3.28
MMC 95.78 0.87 52.26 3.05 84.63 4.10 94.08 4.79 78.21 6.51
LSDA 94.39 1.08 55.83 1.51 83.70 3.80 95.17 1.66 87.93 3.38
LFDA 95.14 1.09 55.60 2.98 86.58 2.15 96.50 1.35 63.59 5.28
NMMP 88.61 2.88 52.84 2.25 83.70 3.49 94.42 4.29 81.03 6.47
SDA 93.64 1.34 51.19 2.34 74.41 7.10 95.67 1.70 63.31 5.92
SULDA 93.44 1.52 56.36 2.15 84.09 2.27 95.08 1.27 88.00 7.00
L21SDA 95.14 1.62 55.78 2.27 85.30 3.04 91.17 3.54 61.86 3.56
ALDE 95.51 0.88 52.61 1.79 76.65 3.74 94.42 4.30 71.72 8.82
ReLSR 94.01 1.71 55.52 1.54 86.58 3.59 95.25 1.62 70.07 6.64
MPDA 94.12 1.57 55.29 3.10 83.67 2.64 93.67 4.47 84.69 2.60
RSLDA 90.61 2.55 51.90 1.91 82.63 4.59 96.00 1.10 63.38 5.40
ALPR 94.49 1.67 55.49 2.03 83.02 3.15 94.83 2.25 65.72 5.76
RLAR **95.99** 0.92 **57.00** 2.81 **86.76** 4.42 **96.58** 1.54 **90.34** 2.82
--------------------- ----------- ------ ----------- ------ ----------- ------ ----------- ------ ----------- ------
\[tab:UCI\]
### Handwriting Recognition
Handwriting recognition is one of the classical tasks in pattern recognition and computer vision. To evaluate the performance of the proposed RLAR for this task, we perform a series of comparative experiments on the Binary Alphadigits database to demonstrate the effectiveness of our method. The database consists of 1404 samples belonging to 36 classes, each of which is a binary image of $20\times16$ pixels. Besides, the database contains not only digits of ’0’ through ’9’, but also capital letters of ’A’ through ’Z’, thus posing a challenge to classification.
For convenience, we simply denote the database as ’Binalpha’. Then, we randomly select $u$ ($u\!=\!10, 13, 16, 19$) images of each subject to form the training set, and the remaining samples to form the test set. The mean recognition results on the database are shown in Table \[tab:binalpha\], where ’\# number’ stands for the number of training samples in each class and is also used in the later recording of experimental results. It can be observed that the recognition rate of each method increases with the expansion of the training set. And, we found that LDA, LSDA, NMMP, SULDA, and MPDA that performed well in the UCI data sets fail on this database. Moreover, our method is superior to many other methods in recognition efficiency, which also indicates that our method has a strong ability of discrimination.
Alg. \# 10 \# 13 \# 16 \# 19
-------- -------------------- -------------------- -------------------- --------------------
RR 49.45$\pm$1.10 50.68$\pm$0.88 52.17$\pm$1.45 53.53$\pm$1.14
LDA 11.47$\pm$0.86 26.20$\pm$1.71 35.40$\pm$1.89 41.69$\pm$1.64
MMC 63.65$\pm$1.23 64.54$\pm$0.92 65.01$\pm$1.46 65.89$\pm$1.31
LSDA 11.59$\pm$0.97 26.15$\pm$1.81 35.40$\pm$1.88 41.65$\pm$1.61
LFDA 64.20$\pm$1.05 66.22$\pm$0.90 67.57$\pm$1.19 68.35$\pm$1.28
NMMP 23.14$\pm$2.56 31.18$\pm$2.65 35.89$\pm$2.54 37.01$\pm$2.10
SDA 51.70$\pm$1.34 56.05$\pm$0.88 59.32$\pm$1.82 62.61$\pm$1.94
SULDA 10.99$\pm$0.73 19.74$\pm$1.34 26.30$\pm$1.75 30.15$\pm$1.30
L21SDA 43.79$\pm$1.39 44.51$\pm$1.39 46.50$\pm$1.39 49.10$\pm$1.67
ALDE 64.64$\pm$1.13 66.55$\pm$1.01 67.24$\pm$1.31 68.39$\pm$1.57
ReLSR 53.66$\pm$1.11 54.64$\pm$1.15 55.69$\pm$1.02 56.72$\pm$1.12
MPDA 8.40$\pm$3.15 28.29$\pm$2.57 40.64$\pm$1.69 47.29$\pm$1.42
RSLDA 16.21$\pm$1.60 27.05$\pm$2.17 35.89$\pm$1.41 42.51$\pm$1.84
ALPR 51.80$\pm$1.70 52.04$\pm$1.42 53.45$\pm$1.60 55.74$\pm$0.97
RLAR **64.80$\pm$1.31** **66.92$\pm$0.85** **68.33$\pm$1.22** **69.40$\pm$0.87**
: Classification performance (mean$\pm$std %) of various approaches on the Binalpha database
\[tab:binalpha\]
### Face Recognition {#face:recognition}
In this recognition scenario, we employ two real commonly-used face databases to evaluate the performance of all algorithms, namely the extended YaleB database [@Georghiades2001] and the AR database [@Martinez1998]. The YaleB database contains 2414 samples from 38 subjects, while the AR database contains more than 4000 color face images of 126 individuals. These two databases are collected under the condition of illumination and expression changes, while the AR database also contains some occlusions. These changes are challenging the performance of our RLAR.
For the AR database, we extract a subset of 1400 images without any occlusion, including 50 female and 50 male subjects, for the experiments. Before implementing all the algorithms, the face images in both of databases are resized to $32\times32$ pixels. Each experiment is independently repeated for 10 times, and the average experimental results of various methods on the two databases are listed in Tables \[tab:YaleB\] and \[tab:AR\] respectively, in which the number of training samples in each class is included. And the best results in each set of comparisons are shown in bold. Although our method is slightly inferior to ALPR in Table \[tab:YaleB\], it is superior to the others. Moreover, in Table \[tab:AR\] our approach trumps all others. These are sufficient to confirm that the discriminant model we have established is efficient enough to yield desirable face recognition results.
Alg. \# 15 \# 20 \# 25 \# 30
-------- -------------------- -------------------- -------------------- --------------------
RR 93.63$\pm$0.99 95.87$\pm$0.66 97.22$\pm$0.58 98.21$\pm$0.60
LDA 90.68$\pm$0.83 89.99$\pm$0.94 81.91$\pm$0.53 86.51$\pm$0.97
MMC 92.44$\pm$1.23 94.90$\pm$0.74 96.15$\pm$0.74 97.31$\pm$0.59
LSDA 90.74$\pm$0.74 89.93$\pm$1.04 82.38$\pm$0.75 86.54$\pm$0.97
LFDA 88.45$\pm$1.18 90.56$\pm$0.94 91.81$\pm$0.75 93.08$\pm$1.00
NMMP 92.39$\pm$0.67 92.26$\pm$0.89 87.15$\pm$0.58 88.04$\pm$0.97
SDA 93.84$\pm$1.04 95.66$\pm$0.80 96.66$\pm$0.61 97.76$\pm$0.30
SULDA 89.50$\pm$0.81 89.50$\pm$1.04 81.71$\pm$0.72 85.33$\pm$0.84
L21SDA 94.67$\pm$0.81 96.84$\pm$0.59 97.72$\pm$0.45 98.75$\pm$0.26
ALDE 78.80$\pm$1.73 83.88$\pm$1.41 87.04$\pm$1.00 89.40$\pm$0.82
ReLSR 94.16$\pm$0.86 96.41$\pm$0.69 97.55$\pm$0.65 98.48$\pm$0.40
MPDA 92.41$\pm$0.67 92.27$\pm$0.87 87.53$\pm$0.56 85.93$\pm$1.06
RSLDA 86.20$\pm$1.27 89.67$\pm$0.68 91.81$\pm$0.74 93.41$\pm$0.67
ALPR **95.70$\pm$0.93** **97.52$\pm$0.52** **98.47$\pm$0.39** **99.16$\pm$0.32**
RLAR 95.07$\pm$0.80 97.04$\pm$0.72 97.84$\pm$0.48 98.65$\pm$0.33
: Classification performance (mean$\pm$std %) of various approaches on the YaleB database
\[tab:YaleB\]
Alg. \# 5 \# 6 \# 7 \# 8
-------- -------------------- -------------------- -------------------- --------------------
RR 95.07$\pm$0.73 96.64$\pm$0.42 97.74$\pm$0.67 98.18$\pm$0.75
LDA 90.83$\pm$1.12 90.98$\pm$0.41 90.63$\pm$0.82 88.83$\pm$1.54
MMC 91.71$\pm$0.74 94.15$\pm$0.75 95.57$\pm$0.95 96.55$\pm$0.72
LSDA 91.13$\pm$0.92 91.21$\pm$0.67 90.34$\pm$0.65 88.88$\pm$1.40
LFDA 81.71$\pm$1.03 85.20$\pm$0.63 88.17$\pm$0.83 90.47$\pm$1.04
NMMP 94.56$\pm$0.95 94.84$\pm$0.39 94.94$\pm$0.84 94.50$\pm$0.74
SDA 95.44$\pm$0.71 97.00$\pm$0.57 97.97$\pm$0.59 98.12$\pm$0.71
SULDA 90.34$\pm$1.05 90.34$\pm$0.59 90.19$\pm$0.84 88.10$\pm$1.48
L21SDA 95.11$\pm$0.74 97.09$\pm$0.53 97.90$\pm$0.50 98.17$\pm$0.62
ALDE 92.60$\pm$0.59 94.69$\pm$0.64 95.76$\pm$0.91 96.68$\pm$0.80
ReLSR 94.54$\pm$0.66 96.35$\pm$0.47 97.41$\pm$0.65 97.97$\pm$0.75
MPDA 93.84$\pm$0.90 94.89$\pm$0.76 95.81$\pm$0.86 96.03$\pm$1.00
RSLDA 87.16$\pm$1.93 90.68$\pm$1.48 93.01$\pm$1.15 94.73$\pm$0.75
ALPR 96.07$\pm$0.70 97.78$\pm$0.55 98.47$\pm$0.71 98.88$\pm$0.48
RLAR **97.33$\pm$0.67** **98.25$\pm$0.30** **98.71$\pm$0.47** **98.95$\pm$0.46**
: Classification performance (mean$\pm$std %) of various approaches on the AR database
\[tab:AR\]
Alternatively in Fig. \[fig:Visualization\], we illustrate the optimal visualization results for the case of 8 training samples per class on the AR database, including the retargeted matrix $\bm T$, the connection matrix $\bm V$, and the induced affinity matrix $\bm S$. It also shows intuitively that our method has obvious effect on revealing the intra-class local graph structure and learning the target representation with distinct margins.
### Object Recognition
To demonstrate the effectiveness of our method in dealing with the problem of object recognition, we conduct a series of comparison experiments on the COIL20 and Caltech101 databases. The COIL20 database contains 20 objects and a total of 1440 images. As the objects rotate on the turntable, images of each object are taken at 5-degree intervals, with 72 images per object. The Caltech101 database has images of 102 classes of objects containing a background class, each of which has about 40 to 800 images, and most classes have about 50 images.
For the COIL20 database we just use these gray-scale images that are resized to $32\times32$ pixels. For the samples in Caltech101 database, we employ spatial pyramid features with dimension 3000 for recognition in view of the differences in background, size and scale. Besides, $u$ ($u\!=\!15,20,25,30$) and $v$ ($v\!=\!10,15,20,25$) samples are selected from each class of the two databases as the training sets, and the remaining samples are used as the test sets. The average experimental results achieved by various methods are shown in Tables \[tab:COIL20\] and \[tab:Caltech101\]. As can be seen in Tables \[tab:COIL20\] and \[tab:Caltech101\], our approach achieves competitive performance compared to other approaches. In particular, with the exception of MMC, our method performs significantly better on the Caltech101 database than any other method.
Alg. \# 15 \# 20 \# 25 \# 30
-------- -------------------- -------------------- -------------------- --------------------
RR 94.50$\pm$1.15 96.68$\pm$0.76 97.59$\pm$0.72 98.20$\pm$0.56
LDA 87.28$\pm$0.86 89.28$\pm$0.92 89.88$\pm$0.79 90.49$\pm$1.03
MMC 96.73$\pm$1.01 98.11$\pm$0.50 98.78$\pm$0.45 99.25$\pm$0.42
LSDA 87.89$\pm$0.96 89.53$\pm$1.04 90.36$\pm$0.95 90.62$\pm$0.75
LFDA 96.39$\pm$0.76 97.62$\pm$0.28 98.93$\pm$0.42 99.35$\pm$0.45
NMMP 92.57$\pm$1.37 93.91$\pm$1.06 94.52$\pm$0.49 94.90$\pm$0.61
SDA 92.24$\pm$1.34 95.16$\pm$1.08 96.21$\pm$0.78 97.60$\pm$0.63
SULDA 85.61$\pm$1.10 87.66$\pm$1.20 88.60$\pm$1.06 88.98$\pm$0.96
L21SDA 94.27$\pm$0.84 96.13$\pm$0.94 96.86$\pm$0.68 97.74$\pm$0.71
ALDE 96.61$\pm$0.99 98.11$\pm$0.67 98.70$\pm$0.37 99.21$\pm$0.49
ReLSR 94.87$\pm$0.72 96.77$\pm$0.69 97.60$\pm$0.78 98.18$\pm$0.50
MPDA 92.57$\pm$1.37 93.91$\pm$1.06 94.52$\pm$0.49 94.90$\pm$0.61
RSLDA 91.77$\pm$1.51 93.73$\pm$1.20 94.33$\pm$1.05 95.08$\pm$0.81
ALPR 95.32$\pm$0.69 97.03$\pm$0.71 97.93$\pm$0.58 98.35$\pm$0.58
RLAR **96.92$\pm$0.87** **98.58$\pm$0.30** **99.10$\pm$0.41** **99.38$\pm$0.33**
: Classification performance (mean$\pm$std %) of various approaches on the COIL20 database
\[tab:COIL20\]
Alg. \# 10 \# 15 \# 20 \# 25
-------- -------------------- -------------------- -------------------- --------------------
RR 58.23$\pm$0.84 60.50$\pm$0.49 61.25$\pm$0.53 62.03$\pm$0.49
LDA 56.64$\pm$0.94 55.73$\pm$0.52 50.26$\pm$0.73 34.72$\pm$0.84
MMC 61.95$\pm$0.88 66.45$\pm$0.62 69.39$\pm$0.57 71.72$\pm$0.42
LSDA 56.56$\pm$0.96 55.80$\pm$0.41 50.33$\pm$0.64 34.69$\pm$0.82
LFDA 55.67$\pm$1.25 61.05$\pm$0.74 64.82$\pm$0.69 67.66$\pm$0.87
NMMP 54.25$\pm$0.96 54.46$\pm$0.65 51.52$\pm$0.80 42.92$\pm$0.78
SDA 56.76$\pm$0.75 61.50$\pm$0.83 64.48$\pm$0.72 66.93$\pm$0.37
SULDA 46.75$\pm$0.84 48.17$\pm$0.84 44.73$\pm$0.70 30.96$\pm$0.61
L21SDA 54.41$\pm$1.20 56.54$\pm$0.88 54.96$\pm$0.70 52.66$\pm$0.41
ALDE 58.44$\pm$0.79 62.57$\pm$0.68 65.28$\pm$0.50 67.53$\pm$0.49
ReLSR 61.68$\pm$1.03 65.25$\pm$0.58 67.32$\pm$0.61 69.01$\pm$0.51
MPDA 54.23$\pm$1.09 54.41$\pm$0.53 51.52$\pm$0.79 42.97$\pm$0.81
RSLDA 51.82$\pm$1.20 51.11$\pm$0.77 46.11$\pm$0.58 36.22$\pm$0.79
ALPR 61.50$\pm$1.07 65.11$\pm$0.46 66.45$\pm$0.66 67.09$\pm$0.44
RLAR **62.81$\pm$1.03** **67.64$\pm$0.68** **70.52$\pm$0.58** **72.61$\pm$0.41**
: Classification performance (mean$\pm$std %) of various approaches on the Caltech101 database
\[tab:Caltech101\]
Robustness Evaluation
---------------------
To investigate the sensitivity of our RLAR to outliers, we conduct two groups of comparative experiments involving the intensity and quantity of outliers on the AR database. The subset of AR database containing 1400 face images mentioned in Section \[face:recognition\] is denoted as ’Subset1’ here. And, we extract another subset with natural occlusion from the AR database, named as ’Subset2’, including 600 images blocked by glasses and 600 images blocked by scarves from 50 male and 50 female subjects. Besides, we artificially block facial images in ’Subset1’ by a ’baboon’ image with varying block sizes. Some sample images of natural occlusion and artificial occlusion are illustrated in Fig. \[fig:AR:Samples\] and Fig. \[fig:AR:CorruptedSamples\], respectively. The detailed experimental schemes and results are as follows.
![One image without blocking and two images with glasses and scarf blocking selected from the AR database.[]{data-label="fig:AR:Samples"}](AR_Sample_Images.eps "fig:")\
![Sample images corrupted by a ’baboon’ image with varying block sizes.[]{data-label="fig:AR:CorruptedSamples"}](AR_Corrupted_Images.eps "fig:")\
We first observe the variation of classification performance with the intensity of outliers. We randomly select 8 samples from each class of ’Subset1’ to form the training set, among which 3 images are corrupted by a randomly located square block of a ’baboon’ image, and the remaining samples were used as the test set. The block size determines the occlusion level of an image. Then, we evaluate the classification performance of various methods at four occlusion levels, and then list the average experimental results for 10 trials independently in Table \[tab:artificial:occlusion\]. For the sensitivity of quantitative outliers, we randomly select 1, 2, and 3 samples from the above 8 training samples, and then replace them with the glasses and scarf blocking images selected randomly in ’Subset2’. Similarly, we independently perform 10 trials for each evaluation, and present the average experimental results in Table \[tab:original:occlusion\].
In Tables \[tab:artificial:occlusion\] and \[tab:original:occlusion\], as the intensity and quantity of outliers increase, we observe that the classification accuracies achieved by various methods gradually decreases. It is worth noting that SULDA, who performs well in the above experiments, fails completely in the face of outlier interference. And in Table \[tab:artificial:occlusion\], NMMP also fails at the occlusion level of $30\times30$. It can be concluded from Table \[tab:artificial:occlusion\] that our method is superior to all methods in terms of recognition accuracy at different occlusion levels. In Table \[tab:original:occlusion\], the proposed method is slightly inferior to ALPR except when the number of scarf blocking images is 2, while it outperforms all of the compared methods in other cases. In general, with the increase of the intensity and quantity of outliers, the recognition accuracies of our RLAR does not decrease significantly, which also demonstrates that our method has excellent performance in resisting outliers.
--------------------- -------------- -------------- -------------- --------------
\[4\][\*]{}[Alg.]{}
15$\times$15 20$\times$20 25$\times$25 30$\times$30
RR 96.60 96.08 95.83 94.50
LDA 84.82 83.72 78.13 56.32
MMC 93.73 92.60 91.63 91.08
LSDA 85.22 83.88 78.27 56.23
LFDA 84.22 81.98 79.88 77.80
NMMP 92.27 91.55 88.65 2.40
SDA 96.57 95.82 95.72 93.90
SULDA 30.73 31.53 27.87 8.22
L21SDA 97.25 96.53 96.50 94.22
ALDE 93.73 92.18 90.63 88.60
ReLSR 96.55 95.72 95.40 93.53
MPDA 92.85 91.98 91.80 91.52
RSLDA 88.98 87.12 82.50 66.53
ALPR 98.25 97.80 97.43 94.90
RLAR **98.43** **98.02** **97.92** **95.23**
--------------------- -------------- -------------- -------------- --------------
: Classification performance (%) on the AR database with varying block sizes.
\[tab:artificial:occlusion\]
--------------------- ----------- ----------- ----------- ----------- ----------- -----------
\[4\][\*]{}[Alg.]{}
\# 1 \# 2 \# 3 \# 1 \# 2 \# 3
RR 97.60 97.13 96.45 97.57 97.05 96.42
LDA 87.13 86.48 85.28 87.47 86.20 84.13
MMC 95.35 94.47 93.20 95.98 95.63 94.98
LSDA 87.02 86.65 84.90 87.87 86.32 84.50
LFDA 87.37 85.08 82.05 88.57 87.43 85.42
NMMP 93.57 92.70 91.92 93.30 92.57 91.47
SDA 97.70 97.15 96.40 97.80 97.17 96.73
SULDA 41.50 37.72 29.48 40.72 36.98 28.40
L21SDA 97.90 97.55 96.90 98.02 97.73 97.32
ALDE 95.47 94.18 92.02 95.55 93.72 90.77
ReLSR 97.58 96.87 96.10 97.60 96.98 96.07
MPDA 94.80 93.65 92.38 95.00 93.95 92.12
RSLDA 91.92 90.67 87.67 92.83 91.82 90.78
ALPR 98.55 98.10 97.73 98.63 **98.33** 97.93
RLAR **98.67** **98.38** **98.05** **98.72** 98.32 **97.98**
--------------------- ----------- ----------- ----------- ----------- ----------- -----------
: Classification performance (%) on the AR database with diverse number of glasses and scarves blocking images.
\[tab:original:occlusion\]
Parameter Sensitivity Analysis
------------------------------
In this section, we examine the parameter sensitivity of the proposed RLAR, which involves three hyper-parameters, namely the number of neighbors $K$, the regularization parameter $\alpha$, and the tradeoff coefficient $\beta$. In all of the experiments above, we set $K$ to either a fixed 3 or a fixed 7, which we mentioned in our experimental settings. Here we mainly focus on discussing the impact of changes in $\alpha$ and $\beta$ on the performance of the proposed model. The predetermined adjustment coordinate set of these two parameters is set as $\{0.001,0.005,0.01,0.05,0.1,0.5,1 ,10,100,1000\}$. The recognition results versus two parameters on 10 databases are visualized in Fig. \[fig:Parameter:sensitivity\], where the value of horizontal and vertical coordinates represents the subscript of the coordinate set, and the number or proportion of training samples per class is indicated in parenthesis of the corresponding caption. It can be observed that the two parameters are not allowed to be too large and not allowed to be too small, and generalized classification performance can be achieved near $[0.1, 0.1]$.
![Classification performance evaluation (%) of the proposed RLAR versus the number of neighbor samples $K$ on the YaleB database.[]{data-label="fig:YaleB:K:Selection"}](YaleB_K_Selection.eps "fig:")\
Besides, we fix the optimal $\alpha$ and $\beta$ obtained through grid search, and observe the effect of the number of neighbors from 1 to 29 on the classification performance on the YaleB database. The recognition results are illustrated in the Fig. \[fig:YaleB:K:Selection\], from which it is observed that the classification performance varies slightly with $K$ and reaches the optimal at 3, 4, and 5 neighbors. Moreover, it is acceptable to set $K$ as a fixed value in all the previous experiments.
Convergence Study
-----------------
The model we built involves multiple variables and is non-smooth, which inspires us to develop an iterative optimization strategy for solving it. The convergence of the optimization algorithm is theoretically guaranteed in Section \[algorithm\_analysis\]. Here, we experimentally verify the convergence performance of the proposed optimization algorithm on 10 databases. Accordingly, we show the convergence curves in Fig. \[fig:Convergence\], from which we observe that all convergence curves are indeed monotonically decreasing and tend to flatten within 30 iterations. The validity of the proposed RLAR is also confirmed by the mutual support between theoretical proof and experimental results.
Analysis of Experimental Results
--------------------------------
Combining theoretical and experimental results, we concentrate on discussing the following insights.
1. It can be concluded from the above results in Tables \[tab:UCI\]-\[tab:original:occlusion\] that our RLAR surpasses other state-of-the-art approaches and can survive in multiple application scenarios with a relatively efficient and robust classification. And the tuning of the three hyper-parameters that our RLAR handles can be achieved through a simple grid search. The visualization results in Fig. \[fig:Visualization\] and the convergence curves in Fig. \[fig:Convergence\] experimentally support the aforementioned theoretical assumptions and derivations.
2. LFDA and NMMP et al. developed the point-to-center loss of LDA as the point-to-point loss, and endowed interlinked samples affinity, breaking through the limit of Gaussian distribution, and the above experimental results indeed confirmed the effectiveness of this move. Inspired by this, we convert the fully-connected intra-class scatter loss of LDA into the loss of partial connection, and effectively overcome the mean dependence and sensitivity to outliers of LDA and its variants by formalizing the affinity between neighbor samples with the non-squared $L_2$ norm.
3. Generally speaking, our approach performs more stable and efficient than these state-of-the-art discriminant FE methods in classification efficiency, which indicates that the proposed new discriminant criteria are indeed effective and have the stronger ability to extract the discriminant information. Besides, compared with the manifold-inspired method MPDA, the underlying structure of the data preserved more effectively by forcibly separating the submanifolds on which different classes of samples are attached while revealing the overall manifold structure of the data.
4. The above experimental results demonstrate that multi-classification regression methods RR and ReLSR are relatively resultful for multi-scene recognition, thus suggesting that they are indeed valid in separating samples belonging to different classes in the latent subspace. Moreover, ReLSR is more versatile than RR in most comparisons, which indicates that learning regression targets with large margins of different classes is beneficial to classification performance.
5. Differ from the conventional methods of depicting the loss function with the squared $L_2$ norm, all the modules in our model are directly measured with the $L_{2,1}$ norm of matrix, which not only enables our model to have the ability of anti-noise, but also can realize the joint process of subspace learning and feature selection. Specially designed sensitivity experiments to outliers also verify the robustness of this strategy.
Conclusion
==========
In this paper, we succeed in achieving a more robust and discriminative low-dimensional representation of data, which is suitable for labeled data classification in multiple scenarios. The proposed model can both adaptively reveal the local structure of the data manifold and flexibly learn the margin representation. All the modules in our model are measured by $L_{2,1}$ norms, thus achieving the joint robust subspace learning and feature selection. Furthermore, we derive an alternate iterative optimization algorithm which is theoretically proved to converge. Extensive experiments conducted on several UCI and other real-world databases have demonstrated the robustness to outliers and classification efficiency of the proposed method. Although we employ retargeted regression here as a way to induce the margin representation, there are certain limitations on the low-dimensional representations, which is where our future work needs to focus.
Proof of Eq. (\[eq:LDA:intraclass\]) {#LDA:intra-class}
====================================
$$\begin{split}
&\sum_{i=1}^c\sum_{j=1}^{n_i}\left\|\bm X^i_j-\bm M_i\right\|_2^2\\
=&\sum_{i=1}^c\sum_{j=1}^{n_i}(\bm X^i_j-\frac{1}{n_i}\sum_{k=1}^{n_i}\bm X^i_k)^T(\bm X^i_j-\frac{1}{n_i}\sum_{k=1}^{n_i}\bm X^i_k)\\
=&\sum_{i=1}^c\!\sum_{j=1}^{n_i}({\bm X^i_j}^T\!\bm X^i_j\!-\!\frac{2}{n_i}{\bm X^i_j}^T\!\sum_{k=1}^{n_i}\!\bm X^i_k\!+\!\frac{1}{n_i^2}\!\sum_{j,k=1}^{n_i}\!{\bm X^i_j}^T\!\bm X^i_k)\\
=&\sum_{i=1}^c(\sum_{j=1}^{n_i}{\bm X^i_j}^T\bm X^i_j-\frac{1}{n_i}\sum_{j,k=1}^{n_i}{\bm X^i_j}^T\bm X^i_k)\\
=&\sum_{i=1}^c\frac{1}{2n_i}\sum_{j,k=1}^{n_i}({\bm X^i_j}^T\bm X^i_j-2{\bm X^i_j}^T\bm X^i_k+{\bm X^i_k}^T\bm X^i_k)\\
=&\sum_{i=1}^c\frac{1}{2n_i}\sum_{j,k=1}^{n_i}(\bm X^i_j-\bm X^i_k)^T(\bm X^i_j-\bm X^i_k)\\
=&\sum_{i=1}^c\sum_{j,k=1}^{n_i}\frac{1}{2n_i}\|\bm X^i_j-\bm X^i_k\|_2^2\\
\end{split}$$
Acknowledgment {#acknowledgment .unnumbered}
==============
The authors would like to thank...
[1]{} L. Ladha and T. Deepa, “Feature Selection Methods And Algorithms,” Int. J. Comput. Eng., vol. 3, no. 5. pp. 1787-1797, 2011. X. Ma, F. Zhang, Y. Li, and J. Feng, “Robust sparse representation based face recognition in an adaptive weighted spatial pyramid structure,” Sci. China Inf. Sci., vol. 61, no. 1. pp. 86-98, 2018. N. Passalis and A. Tefas, “Dimensionality reduction using similarity-induced embeddings,” IEEE Trans. Neural Netw. Learn. Syst., vol. 29, no. 8, pp. 3429-3441, 2018. Y. Pang, B. Zhou, and F. Nie, “Simultaneously Learning Neighborship and Projection Matrix for Supervised Dimensionality Reduction,” IEEE Trans. neural networks Learn. Syst., vol. 30, no. 9, pp. 2779-2793, 2019.
S. Khalid, T. Khalil, and S. Nasreen, “A survey of feature selection and feature extraction techniques in machine learning,” in Proceedings of 2014 Science and Information Conference, SAI 2014, 2014, pp. 372-378.
H. Hotelling, “Analysis of a complex of statistical variables into principal components,” J. Educ. Psychol., vol. 24, no. 6, pp. 417-441, 1933. R. A. Fisher, “The use of multiple measurements in taxonomic problems,” Ann. Eugen., vol. 7, no. 2, pp. 179-188, 1936. M. Sugiyama, “Dimensionality reduction of multimodal labeled data by local fisher discriminant analysis,” J. Mach. Learn. Res., vol. 8, pp. 1027-1061, 2007. S. Boutemedjet, N. Bouguila, and D. Ziou, “A hybrid feature extraction selection approach for high-dimensional non-Gaussian data clustering,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 31, no. 8, pp. 1429-1443, 2009. T. Luo, C. Hou, F. Nie, and D. Yi, “Dimension Reduction for Non-Gaussian Data by Adaptive Discriminative Analysis,” IEEE Trans. Cybern., vol. 49, no. 3, pp. 933-946, 2019. X. Li, M. Chen, F. Nie, and Q. Wang, “Locality adaptive discriminant analysis,” in IJCAI International Joint Conference on Artificial Intelligence, 2017, pp. 2201-2207.
J. Ye and T. Xiong, “Computational and theoretical analysis of null space and orthogonal linear discriminant analysis,” J. Mach. Learn. Res., vol. 7, pp. 1183-1204, 2006. H. Li, T. Jiang, and K. Zhang, “Efficient and Robust Feature Extraction by Maximum Margin Criterion,” IEEE Trans. Neural Networks, vol. 17, no. 1, pp. 157-165, 2006.
X. Zhang, D. Chu, and R. C. E. Tan, “Sparse Uncorrelated Linear Discriminant Analysis for Undersampled Problems,” IEEE Trans. Neural Networks Learn. Syst., vol. 27, no. 7, pp. 1469-1485, 2016. H. Zhao, Z. Wang, and F. Nie, “A New Formulation of Linear Discriminant Analysis for Robust Dimensionality Reduction,” IEEE Trans. Knowl. Data Eng., vol. 31, no. 4, pp. 629-640, 2019. J. Wen et al., “Robust Sparse Linear Discriminant Analysis,” IEEE Trans. Circuits Syst. Video Technol., vol. 29, no. 2, pp. 390-403, 2019.
X. He and P. Niyogi, “Locality preserving projections,” in Advances in Neural Information Processing Systems, 2004, vol. 16, no. December, pp. 153-160. D. Cai, X. He, K. Zhou, J. Han, and H. Bao, “Locality sensitive discriminant analysis,” in IJCAI International Joint Conference on Artificial Intelligence, 2007, pp. 708-713. F. Nie, S. Xiang, and C. Zhang, “Neighborhood MinMax projections,” in IJCAI International Joint Conference on Artificial Intelligence, 2007, pp. 993-998. Z. Fan, Y. Xu, and D. Zhang, “Local linear discriminant analysis framework using sample neighbors,” IEEE Trans. Neural Netw., vol. 22, no. 7, pp. 1119-1132, 2011.
H. S. Seung and D. D. Lee, “The manifold ways of perception,” Science, vol. 290, no. 5500. pp. 2268-2269, 2000. M. Belkin and P. Niyogi, “Laplacian eigenmaps and spectral techniques for embedding and clustering,” in Advances in Neural Information Processing Systems, 2002, pp. 585-591.
F. Nie, Z. Wang, R. Wang, and X. Li, “Submanifold-preserving discriminant analysis with an auto-optimized graph,” IEEE Trans. Cybern., pp. 1-14, 2019.
Q. Gao, J. Ma, H. Zhang, X. Gao, and Y. Liu, “Stable orthogonal local discriminant embedding for linear dimensionality reduction,” IEEE Trans. Image Process., vol. 22, no. 7, pp. 2521-2531, 2013.
R. G. Brereton and G. R. Lloyd, “Support Vector Machines for classification and regression,” Analyst, vol. 135, no. 2. pp. 230-267, 2010. C. W. Hsu and C. J. Lin, “A comparison of methods for multiclass support vector machines,” IEEE Trans. Neural Networks, vol. 13, no. 2, pp. 415-425, 2002.
A. E. Hoerl and R. W. Kennard, “Ridge regression: Biased estimation for nonorthogonal problems,” Technometrics, vol. 12, no. 1, pp. 55-67, 1970. R. Tibshirani, “Regression shrinkage and selection via the lasso,” J. R. Stat. Soc. Ser. B, vol. 58, no. 1, pp. 267-288, 1996. H. Zou and T. Hastie, “Regularization and variable selection via the elastic net,” J. R. Stat. Soc. Ser. B Stat. Methodol., vol. 67, no. 2, pp. 301-320, 2005.
Z. Lai, D. Mo, J. Wen, L. Shen, and W. K. Wong, “Generalized Robust Regression for Jointly Sparse Subspace Learning,” IEEE Trans. Circuits Syst. Video Technol., vol. 29, no. 3, pp. 756-772, 2019. J. Wen, Z. Zhong, Z. Zhang, L. Fei, Z. Lai, and R. Chen, “Adaptive Locality Preserving Regression,” IEEE Trans. Circuits Syst. Video Technol., vol. 30, no. 1, pp. 75-88, 2020. R. Rosipal and L. J. Trejo, “Kernel partial least squares regression in reproducing kernel hilbert space,” J. Mach. Learn. Res., vol. 2, no. 2, pp. 97-123, 2002. W. Liu, I. Park, Y. Wang, and J. C. Principe, “Extended kernel recursive least squares algorithm,” IEEE Trans. Signal Process., vol. 57, no. 10, pp. 3801-3814, 2009. B. Chen, J. Liang, N. Zheng, and J. C. Pr¨ªncipe, “Kernel least mean square with adaptive kernel size,” Neurocomputing, vol. 191, pp. 95-106, 2016.
S. M. Xiang, F. P. Nie, G. F. Meng, C. H. Pan, and C. S. Zhang, “Discriminative least squares regressions for multiclass classification and feature selection,” IEEE Trans. Neural Netw. Learn. Syst., vol. 23, no. 11, pp. 1738-1754, 2012. X. Zhang, L. Wang, S. Xiang and C. Liu, “Retargeted Least Squares Regression Algorithm,” IEEE Transactions on Neural Networks and Learning Systems, vol. 26, no. 9, pp. 2206-2213, 2015. L. Wang and C. Pan, “Groupwise retargeted least-squares regression,” IEEE Trans. Neural Netw. Learn. Syst., vol. 29, no. 4, pp. 1352-1358, 2018.
S. Liu, L. Feng, and H. Qiao, “Scatter balance: An angle-based supervised dimensionality reduction,” IEEE Trans. Neural Networks Learn. Syst., vol. 26, no. 2, pp. 277-289, 2015. Q. Ke and T. Kanade, “Robust $L_1$ norm factorization in the presence of outliers and missing data by alternative convex programming,” in IEEE Computer Society Conference on Computer Vision and Pattern Recognition, CVPR 2005, 2005, vol. I, pp. 739-746. C. Ding, D. Zhou, X. He, and H. Zha, “R1-PCA: Rotational invariant L1-norm principal component analysis for robust subspace factorization,” in ACM International Conference Proceeding Series, 2006, vol. 148, pp. 281-288. N. Kwak, “Principal component analysis based on L1-norm maximization,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 30, no. 9, pp. 1672-1680, 2008. X. Li, W. Hu, H. Wang, and Z. Zhang, “Linear discriminant analysis using rotational invariant L1 norm,” Neurocomputing, vol. 73, no. 13¨C15, pp. 2571-2579, 2010. L. Clemmensen, T. Hastie, D. Witten, and B. Ersboll, “Sparse discriminant analysis,” Technometrics, vol. 53, no. 4, pp. 406-413, 2011. F. Zhong and J. Zhang, “Linear discriminant analysis based on L1-norm maximization,” IEEE Trans. Image Process., vol. 22, no. 8, pp. 3018-3027, 2013. H. Wang, X. Lu, Z. Hu, and W. Zheng, “Fisher discriminant analysis with L1-norm,” IEEE Trans. Cybern., vol. 44, no. 6, pp. 828-842, 2014.
S. Yi, Z. Lai, Z. He, Y. ming Cheung, and Y. Liu, “Joint sparse principal component analysis,” Pattern Recognit., vol. 61, pp. 524-536, 2017. S. Yi, Z. He, X.-Y. Jing, Y. Li, Y.-M. Cheung, and F. Nie, “Adaptive Weighted Sparse Principal Component Analysis for Robust Unsupervised Feature Selection,” IEEE Trans. Neural Networks Learn. Syst., pp. 1-11, 2019.
X. Shi, Y. Yang, Z. Guo, and Z. Lai, “Face recognition by sparse discriminant analysis via joint L2,1-norm minimization,” Pattern Recognit., vol. 47, no. 7, pp. 2447-2453, 2014.
S. J. Raudys and A. K. Jain, “Small Sample Size Effects in Statistical Pattern Recognition: Recommendations for Practitioners,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 13, no. 3, pp. 252-264, 1991.
F. Nie, H. Huang, X. Cai, and C. Ding, “Efficient and robust feature selection via joint L2;1-norms minimization,” in Advances in Neural Information Processing Systems 23: 24th Annual Conference on Neural Information Processing Systems 2010, NIPS 2010, 2010.
Z. Zhang, L. Shao, Y. Xu, L. Liu, and J. Yang, “Marginal Representation Learning with Graph Structure Self-Adaptation,” IEEE Trans. Neural Networks Learn. Syst., vol. 29, no. 10, pp. 4645-4659, 2018.
Y. Zhou and S. Sun, “Manifold partition discriminant analysis,” IEEE Trans. Cybern., vol. 47, no. 4, pp. 830-840, 2017.
A. S. Georghiades, P. N. Belhumeur, and D. Kriegman, “From few to many: Illumination cone models for face recognition under variable lighting and pose,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 23, no. 6, pp. 643-660, Jun. 2001.
A. Martinez and R. Benavente, “The AR face database,” CVC, New Delhi, India, Tech. Rep. 24, 1998.
[^1]: This work was supported in part by the National Key R$\&$D Program of China under Grant 2017YFC0803700, and in part by the National Natural Science Foundation of China under Grant U1636220 and Grant 61876183. *[(Corresponding author: Wensheng Zhang.)]{}*
[^2]: L. Hu is with the School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing 210094, China (e-mail: hlc\[email protected]).
[^3]: W. Zhang is with the Research Center of Precision Sensing and Control, Institute of Automation, Chinese Academy of Sciences, Beijing, 100190, China, and also with the University of Chinese Academy of Sciences, Beijing, 101408, China (e-mail: [email protected]).
[^4]: Manuscript received April 19, 2005; revised August 26, 2015.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We use a mathematical framework that we introduced in a previous paper to study geometrical and quantum mechanical aspects of a Hall system with finite size and general boundary conditions. Geometrical structures control possibly the integral or fractionnal quantization of the Hall conductivity depending on the value of $NB/2\pi$ ($N$ is the number of charge carriers and $B$ is the magnetic field). When $NB/2\pi$ is irrationnal, we show that monovalued wave functions can be constructed only on the graph of a free group with two generators. When $NB/2\pi$ is rationnal, the relevant space becomes a puncturated Riemann surface. We finally discuss our results from a phenomenological viewpoint.'
author:
- 'F. Chandelier$^a$, Y. Georgelin$^a$, T. Masson$^b$, J.-C. Wallet$^a$'
title: Quantum Hall Conductivity in a Landau Type Model with a Realistic Geometry II
---
$^a$ Groupe de Physique Théorique,\
Institut de Physique Nucléaire\
F-91406 Orsay Cedex, France\
$^b$ Laboratoire de Physique Théorique (UMR 8627)\
Bât 210, Université Paris-Sud Orsay\
F-91405 Orsay Cedex
LPT-Orsay 03-101\
IPNO-DR-03-11
Introduction
============
It is known that the integral quantization of the Hall conductivity can be understood as resulting from topological features underlying the Quantum Hall systems [@PranGirv:90; @Ston:92]. The first convincing argument in this direction was given by Laughlin for a system with cylindrical geometry [@Laug:81] and was refined in subsequent works considering[^1] the case of non-interacting charges carriers (electrons) in a periodic potential [@ThouKohmNighNijs:82; @Kohm:85], then further exented to take into account Quantum Hall systems with disorder and/or interacting electrons [@NiuThou:84; @NiuThouWu:85; @AvroSeil:85; @Thou:83]. It was realized that the Kubo formula giving rise to the Hall conductivity can be expressed in term of the integral of the first Chern class of a certain line vector bundle, stemming from the fact that in the quantum mechanical description of these systems two parameters submitted to some periodic conditions can be singled out or introduced. Attempts to explain the observed fractionnal quantization of the Hall conductivity have also appeared in which basically the above (topological) scheme is further combined with additional (physical) plausible assumptions, for instance the existence of a degeneracy of the ground state of the system together with the existence of a gap above it and/or the validity of averaging the Hall conductivity over gauge parameters [@NiuThou:84; @NiuThouWu:85; @AvroSeil:85; @Thou:83].
In most instance, the above topological approach does not permit one to really predict the value for the Hall conductivity starting from basic ingredients pertaining to the quantum mechanocal description of the systems. This feature may reflect the fact that the line vector bundle which shows up at an essential step of the approach does not (in most cases) retain a sufficient amount of the physical properties shared by the real experiments, such as the geometry of the experimental sample, the number of charges carriers $N$, the applied magnetic field $B$…Note that many of the quantum mechanical models introduced so far in the topological approach for the Quantum Hall Effect somehow rely on assumptions made on some of the characteristics of the experiment[^2], such as for instance torus [@WenNiu:90] or annular geometry [@Halp:82] for the two dimensional sample, which prove convenient to carry on the whole analysis but are sometime questionnable regarding the actual experimental situation. Ultimately, it would be desirable to incorporate into a single framework all the geometrical and physical constraints of the real experiments, control the validity of the hypothesis currently made on the applicability of the Kubo formula (in particular prove the existence of a gap in the Hamiltonian spectrum), study the possible degeneracy of the ground state. This would permit one in particular to determine clearly whether the observed fractionnal quantization of the Hall conductivity can be explainedm as the integral quantization, in terms of topological properties underlying the Hall systems or relies on non topological (additionnal) features.
In a previous paper [@CGMW:03], we considered a Hall system whose geometry, already introduced in [@AvroSeil:85; @NiuThou:87], mimics the one for the real experiments, namely a rectangular sample whose edges are connected by wires (see fig. \[fig-general\] of section \[previousresults\]). This system was assumed to obey general boundary conditions compatible with current and charge conservation and which may account for the respective quantum and classical nature of the electrons inside the sample and those circulating in the wires. We have introduced and discussed a mathematical framework allowing us to treat properly some quantum mechanical aspects specific to the present system with finite size sample and general boundary conditions. These latter in particular force the representation of the considered abstract operator algebra to be reducible. The relevant representation are found to be indexed by two real numbers (defined modulo $2\pi$) forming some “reciprocical” space that permits one to obtain from the Hamiltonian the velocity operators appearing in the Kubo formula. The represenation space has been shown, in turn, to be (continuously) decomposable into representaions of a unitary symmetry which connects translations acting *at the same time* on the reciprocical space and on the “direct” space built from the physical sample. In [@CGMW:03], we have found that when $NB/2\pi$ takes any integer value, noticeable geometrical structures emerge both in the direct and the reciprocical space: the wave functoins, extended to the whole $\gR^2$-plan, can be interpreted as sections of a line vector bundle over a torus[^3] living in the direct space which is rigidely linked through the above unitary symmetry to a line vector bundle over a torus living in the reciprocical space whose (integral of) first Chern class appears in the Kubo formula. Thanks to this correspondance, this latter formula can be explicitely calculated and we found that the Hall conductivity can take integer and fractionnal values.
When $NB/2\pi$ differs from integer values, the situation becomes far more complicated. It appears that the natural problem to be solved is to determine at least what geometrical structure (if any), that could be used to calculate physical quantities, would survive when $NB/2\pi$ is not integer. This is the purpose of the present paper. The paper is organized as follows. In section \[previousresults\], we summarize the main results obtained in [@CGMW:03] and collect the relevant material that will be used in the analysis. The section \[generalcase\] is devoted to a characterization of the space on which monovaluated wave functions can be construted for arbitrary (not integer) $NB/2\pi$. When $NB/2\pi$ is irrational, we find that the space is a tree corresponding to the graph of a free group with two generators and can ve identified with the universal covering space of $\gR^2 \setminus \gZ^2$. In section \[rationnalcase\], we consider the case where $NB/2\pi$ takes any irreducible rational value $\ell/k$. Using the van Kampen theorem, we find that the relevant space on which monovaluated wave functions exist is a Riemann surface with $k^2$ punctures and genus $g_k=1 + k^2(k-1)/2$. The complete proof is rather involved and is presented in the appendix. We further discuss under which circumstances the analysis presented in [@CGMW:03] for $NB/2\pi$ integer can be extended to rationnal values of $NB/2\pi$. When the extension can be performed, we find that the Hall conductivity still takes integer or fractionnal values. In section \[physicalcomments\], we discuss from a phenomenological viewpoint a situation when $NB/2\pi$ differs slightly from a rationnal value. The combinations of the fundamental boundary relations for the wave functions with some plausible physical assumptions, in particular the one stating that the geometrical structures exhibited in the pure rationnal case (for which the computation of the Hall conductivity is possible) would survive to the introduction of a sufficiently weak disorder, suggests the existence of an infinite set of domains in the disorder-filling factor plan whose architecture is reminiscent to the global phase diagram from the Quantum Hall Effect proposed in [@CGMW:02; @CGMW:02b]. Finally we summarize the results and conclude.
$N$-body Landau type Hamiltonian on a square {#previousresults}
============================================
In this section, we summarize the results obtained in our previous work [@CGMW:03], hereafter refered as (I), and collect some usefull properties of the corresponding mathematical framework that will be usefull in the sequel. In (I), we considered a spinless multiparticle Landau type Hamiltonian on a finite two-dimensional domain. The Hamiltonian can be expressed in the symmetric gauge[^4] as $$\label{totalhamiltonian}
H = \sum_{i=1}^N \frac{1}{2 m}\left( (p_{x_i} + \frac{1}{2} B y_i)^2 + (p_{y_i}
- \frac{1}{2} B x_i)^2\right) + \sum_{1\leq i<j \leq N} V(x_i - x_j, y_i - y_j)$$ where $(x_i, y_i)$ (resp. $(p_{x_i}, p_{y_i})$) denote the coordinates (resp. momentum components) of the particle $i$ with mass $m$, $B$ is the external magnetic field applied to the system and $V$ represents a two body interaction potential. The corresponding finite domain was assumed in (I) to be $[0,1]^2 = [0,1] \times [0,1]$, that is, a unit square on the $\gR^2$ plane. Furthermore, the boundary conditions we imposed on this quantum mechanical system were motivated by the geometry of the Quantum Hall Effect, as depicted on fig. \[fig-general\], where *classical* currents circulate in wires connecting apposite edges of the square while no charge accumulation on any edge is possible. As shown in (I), by using the conservation of the currents circulating in the wires together with the absence of charge accumulation on the edges, the computation of the transverse (Hall) conductivity can be performed in extenso from the Kubo formula provided a quantization condition on the product of the number of charge carriers $N$ by the external magnetic field $B$ is satisfied. Namely, one must have $NB=2 \pi \ell$ where $\ell$ is any integer.
![Schematic representation of the global geometry of the QHE.[]{data-label="fig-general"}](figures-0.eps)
Keeping in mind that the physical quantities we are interested in (such as the Hall conductivity) depend only on the collective motion of the $N$ charge carriers, we introduce the following variables
$$\label{collective}
x = \frac{1}{N} \sum_{i=1}^N x_i, \ y = \frac{1}{N} \sum_{i=1}^N y_i, \ p_x = \sum_{i=1}^N p_{x_i}, \ p_y = \sum_{i=1}^N p_{y_i}$$
$$\label{internal}
\tilde{x}_i = x_i - x, \quad
\tilde{y}_i = y_i - y, \quad
\tilde{p}_{x_i} = p_{x_i} - \frac{1}{N} p_x, \quad
\tilde{p}_{y_i} = p_{y_i} - \frac{1}{N} p_y$$
for $i=1,\dots,N$, where (\[collective\]) (resp. (\[internal\])) define the collective (resp. “internal”) variables and (\[internal\]) satisfy the relations $$\sum_{i=1}^N\tilde{x}_i = 0,\quad
\sum_{i=1}^N\tilde{y}_i = 0,\quad
\sum_{i=1}^N\tilde{p}_{x_i} = 0,\quad
\sum_{i=1}^N\tilde{p}_{y_i} = 0.$$
In these variables, the Hamiltonian (\[totalhamiltonian\]) can be splitted into two parts as
$$\begin{aligned}
H &= H_0 + H_I\\
H_0 &= \frac{1}{2 N m}\left( (p_{x} + \frac{1}{2} N B y)^2 + (p_{y} - \frac{1}{2} N B x)^2\right)\end{aligned}$$
and $H_I$ depends only on the “internal” variables. The Hilbert space on which this Hamiltonian acts is itself decomposed into $\cH = \cH_0 \otimes \cH_I$. In the following, we will mainly deal with $H_0$ and $\cH_0$.
The Hilbert space $\cH_0$ looks like $L^2([0,1]^2)$ (for the variables $x$ and $y$) but this identification is not correct as shown in (I). Due to the classical nature of the currents in the wires, the representation of the algebra generated by the operators $$P_x = -i\frac{\partial\hfill}{\partial x} + \frac{1}{2} N B y, \
P_y = -i\frac{\partial\hfill}{\partial y} - \frac{1}{2} N B x, \
H_0 = \frac{1}{2 N m} ( P_x^2 + P_y^2)
\stepcounter{equation}
\label{pxpyh0}
\tag{{\thesection.\arabic{equation}}a,b,c}$$ is reducible. Indeed, in order to define the operators $P_x$ and $P_y$, we have to specify their domain in the Hilbert space. Taking into account the various physical and mathematical constraints, in particular the hypothesis about the geometry of the system depicted on fig. \[fig-general\], the self-adjointness of the operators, we introduce the following domain $\cD_{\gamma, \eta}$: $$\begin{aligned}
\cD_{\gamma, \eta} = \{ \phi \in L^2([0,1]^2) /\ &
\phi \text{ absolutely continuous},\\
&\frac{\partial \phi}{\partial x}, \frac{\partial \phi}{\partial y} \in L^2([0,1]^2),\\
&\phi(1,y) = e^{i\gamma + \frac{i}{2} NB y} \phi(0,y),\\
&\phi(x,1) = e^{i\eta - \frac{i}{2} NB x} \phi(x,0),\\
&\phi(0,0) = \phi(1,0) =\phi(0,1) =\phi(1,1) = 0
\}.\end{aligned}$$ where $\gamma$ and $\eta$ are two real parameters in the interval $[0, 2\pi[$ which label the possible representations of the above algebra. Notice that in the following we will work with any real values for $\gamma$ and $\eta$, knowing that two values differing by $2\pi$ label the same representation. The arbitrariness in these parameters force us to consider them *all at once*. Then, as discussed in (I), the Hilbert space $\cH_0$ is the direct sum of the Hilbert spaces associated to each individual values of the couple $(\gamma, \eta)$. Because each representation indexed by $(\gamma, \eta)$ takes place in the Hilbert space $L^2([0,1]^2)$, $\cH_0$ is the direct hilbertian integral over $(\gamma, \eta)\in [0, 2\pi[^2$ of $L^2([0,1]^2)$.
We denote by $(x,y,\gamma, \eta) \mapsto \psi(x,y,\gamma, \eta)$ a general wave function in $\cH_0$ and $\cD$ the common domain of definition of $P_x$ and $P_y$ in $\cH_0$. As shown in (I), any function $\psi \in \cD$ satisfies the boundary conditions
\[boundarypsi12\] $$\begin{aligned}
\label{boundarypsi1}
\psi(1,y,\gamma,\eta) &= e^{i \gamma + \frac{i}{2} NB y} \psi(0,y,\gamma,\eta) \\
\label{boundarypsi2}
\psi(x,1,\gamma,\eta) &= e^{i \eta - \frac{i}{2} NBx} \psi(x,0,\gamma,\eta)\end{aligned}$$
These boundary conditions will play a central role in the analysis presented in the next sections, and we will refer to them as the *fundamental boundary relations*. From now on, the $(x,y)$ space will be called the direct space while the $(\gamma, \eta)$ space will be called the reciprocical space. Notice the this terminilogy is somehow inherited from solid state physics since in the present analysis, the variables $(\gamma, \eta)$ look very much like conjugate variables to $(x,y)$.
Let us now recall the situation in the special case where $NB=2\pi \ell$, where $\ell$ is an integer, hereafter called the integer case. In this situation, the wave functions $\psi \in \cD$ can be extended to the whole plane $(x,y) \in \gR^2$ by the relations $$\begin{aligned}
\label{boundarypsitorus1}
\psi(x+1,y,\gamma,\eta) &= e^{i \gamma + \frac{i}{2} NB y} \psi(x,y,\gamma,\eta) \\
\label{boundarypsitorus2}
\psi(x,y+1,\gamma,\eta) &= e^{i \eta - \frac{i}{2} NB x} \psi(x,y,\gamma,\eta)\ .\end{aligned}$$ Notice that this extension requires explicitely that $NB=2\pi \ell$ is verified, otherwise, an extra phase factor would appear in the wave function when one follows the path $(x,y) \rightarrow (x+1,y) \rightarrow (x+1,y+1) \rightarrow (x,y+1) \rightarrow (x,y)$. Now, for fixed values $(\gamma, \eta$), the extended functions $(x,y) \mapsto \psi(x,y,\gamma,\eta)$ can be interpreted as sections of a line vector bundle over a torus in the variables $(x,y)$ (of periodicity $1$ in $x$ and $y$). This identification was a crucial point in our explicit computation of the Hall conductivity performed in (I). Indeed, our computation made use of the Kubo formula in term of the Chern number of a line vector bundle over the reciprocical space $(\gamma, \eta)$ [@Kubo:57; @AvroSeilSimo:83] (this space being a torus of periodicity $2\pi$ in our case). Thanks to some symmetries in our system which combine some translations on the two torus *at the same time*, it was possible to caracterise this last line vector bundle over the torus in the variables $(\gamma, \eta)$ using the structure of the line vector bundle over the torus in the variables $(x,y)$. Notice also that in this case, the wave functions $\psi$ can have any values at the point $(0,0)$ and need not to vanish, contrary to the general case. We refer to (I) for details of the computation.
Characterization of the $(x,y)$-direct space for arbitrary $NB$ {#generalcase}
===============================================================
In this section, we study some features of the direct space when $NB$ takes an arbitrary value. As we will see, the situation is far more complicated than the “integer” case $NB = 2 \pi \ell$. Our starting point are the fundamental boundary relations (\[boundarypsi1\]) and (\[boundarypsi2\]) which caracterize the domain $\cD \subset \cH_0$.
As in the integer case, we want to extend the wave functions on the whole plane $(x,y) \in \gR^2$. Unfortunately, this is not possible as we will see. Let $\psi \in \cD$, a wave function. First, using the boundary relation (\[boundarypsi1\]), we extend $\psi$ in the $x$ direction by defining ($H$ stands for “horizontal”) $$\label{definitionPsiH}
\psi_H(x+p, y, \gamma, \eta) = e^{i p \gamma + \frac{i}{2} p N B y} \psi(x, y, \gamma, \eta)$$ for any integer $p$ and any $x \in [0,1[$. The function $\psi_H$ is then defined for $(x,y) \in \gR \times [0,1]$, and satisfies to the boundary relation in $y$ $$\label{boundarypsiHy}
\psi_H(x+p, 1, \gamma, \eta) = e^{i \eta + i p NB - \frac{i}{2} N B (x+p)} \psi_H(x+p, 0, \gamma, \eta)$$ This relation is somehow similar to the fundamental boundary relation (\[boundarypsi2\]) when $x$ is replaced by $x+p$, up to the presence of the extra phase $p NB$.
Now, using the boundary relation (\[boundarypsi2\]), $\psi$ can be extended in the $y$-direction by defining ($V$ stands for “vertical”) $$\label{definitionPsiV}
\psi_V(x, y+q, \gamma, \eta) = e^{i q \eta - \frac{i}{2} q N B x} \psi(x, y, \gamma, \eta)$$ for any integer $q$ and any $y \in [0,1[$. The function $\psi_V$ is then defined for $(x,y) \in [0,1] \times \gR$ and satisfies $$\label{boundarypsiVx}
\psi_V(1, y+q, \gamma, \eta) = e^{i \gamma - i q NB + \frac{i}{2} N B (y+q)} \psi_V(0, y+q, \gamma, \eta)$$ Because $\psi_H$ and $\psi_V$ coincide on the square $(x,y) \in [0,1] \times [0,1]$ with $\psi$ (recall that $\psi$ vanishes on the four corners of this square), the function $\psi$ has been extended to the domain $(x,y) \in \gR \times [0,1] \cup [0,1] \times \gR$.
Let us consider now $\psi_H$ on the square $(x,y) \in [1,2] \times [0,1]$. Using the boundary relation (\[boundarypsiHy\]), we can extend it to $(x,y)\in [1,2] \times \gR$ through $$\label{definitionPsiHV}
\psi_{HV}(x, y+q, \gamma, \eta) = e^{i q \eta + i q NB - \frac{i}{2} q N B x} \psi_H(x, y, \gamma, \eta)$$ for any integer $q$, any $y \in [0,1[$ and any $x \in [1,2]$. Performing the same procedure, with $\psi_V$ combined with the relation (\[boundarypsiVx\]), we define $\psi_{VH}$ on $(x,y) \in \gR \times [1,2]$ by $$\label{definitionPsiVH}
\psi_{VH}(x+p, y, \gamma, \eta) = e^{i p \gamma - i p NB + \frac{i}{2} p N B y} \psi_V(x, y, \gamma, \eta)$$ The two functions $\psi_{HV}$ and $\psi_{VH}$ are defined on the common domain $(x,y) \in [1,2] \times [1,2]$. Unfortunately, they do not necessary coincide there! Indeed, for $(x,y) \in [0,1] \times [0,1]$, a straighforward computation using the above definitions yields
$$\begin{aligned}
\psi_{HV}(x+1, y+1, \gamma, \eta) &= e^{i \eta + i NB - \frac{i}{2} NB (x+1)} e^{i \gamma + \frac{i}{2} NB y} \psi(x, y, \gamma, \eta) \\
\psi_{VH}(x+1, y+1, \gamma, \eta) &= e^{i \gamma - i NB + \frac{i}{2} NB (y+1)} e^{i \eta - \frac{i}{2} NB x} \psi(x, y, \gamma, \eta)\end{aligned}$$
Thus, we have $$\label{phase1}
\psi_{HV}(x+1, y+1, \gamma, \eta) = e^{i NB} \psi_{VH}(x+1, y+1, \gamma, \eta)$$ When $NB$ is not a multiple of $2\pi$, these two functions differ from each other by an extra phase factor. This therefore leads to *two* different extensions of $\psi$ on the square $[1,2] \times [1,2]$. Notice that the choice of the square $[1,2] \times [1,2]$ is arbitrary in this conclusion. Finally, a straighforward computation shows that $\psi_{HV}$ and $\psi_{VH}$ satisfy the boundary relations
$$\begin{aligned}
\label{boundarypsiHVx}
\psi_{HV}(2, y+q, \gamma, \eta) &= e^{i \gamma - i q NB + \frac{i}{2} N B (y+q)} \psi_{HV}(1, y+q, \gamma, \eta)\\
\label{boundarypsiVHy}
\psi_{VH}(x+p, 2, \gamma, \eta) &= e^{i \eta + i p NB - \frac{i}{2} N B (x+p)} \psi_{VH}(x+p, 1, \gamma, \eta)\end{aligned}$$
which are equivalent to (\[boundarypsiVx\]) and (\[boundarypsiHy\]) respectively.
Let us now generalize this construction. Let $\chi_H$ be a given extension of $\psi$ on a domain $(x,y) \in \gR \times [n,n+1]$ for an integer $n$, with some boundary relations of the form (\[boundarypsiHy\]) relating $y=n$ to $y=n+1$. It is easy to define a new extension $\chi_{HV}$ on a domain of the form $[m, m+1] \times \gR$ using the common subspace $[m, m+1] \times [n,n+1]$ and a relation generalizing (\[definitionPsiHV\]). In a similar way, it is possible to extend a function $\chi_V$ defined on a domain $[m, m+1] \times \gR$, with boundary relations in the $x$ direction similar to (\[boundarypsiVx\]), into a function $\chi_{VH}$ defined on a domain $\gR \times [n,n+1]$, using some natural generalisation of (\[definitionPsiVH\]).
Then, starting from the “fundamental” square $[0,1] \times [0,1]$, it is possible to define an extension of $\psi$ on any square $[m,m+1] \times [n,n+1]$ in the plane. Unfortunately, the value of this extension *depends* on the path used to go from the fundamental square to the final square. For instance, on fig. \[fig-path\], two extensions in two steps are possible from the fundamental square (in black) to the target square (in grey): a first extension $\psi_{HV}$, for which the first step in an horizontal extension ($\psi_H$) and the second step is vertical, and a second extension, $\psi_{VH}$, for which the first step is vertical ($\psi_V$) and the second horizontal. On the grey square, there is a phase difference between $\psi_{HV}$ and $\psi_{VH}$. Some elementary computation shows that this phase difference is of the form $$A \times NB$$ where $A$ is the area (evaluated in term of the number of squares) limited by the interior of the two paths joining the black square to the grey square used to define the two extensions $\psi_{HV}$ and $\psi_{VH}$. In (\[phase1\]), this area is just $1$. Note that if the grey square is located at $[m,m+1] \times [n,n+1]$ in $\gR^2$, one has $A = m\times n$.
![The “fundamental” domain is the black square. The wave function $\psi$ can be extended on the four stripes as $\psi_{H}$, $\psi_{V}$, $\psi_{HV}$ and $\psi_{VH}$, but on the grey square, $\psi_{HV}$ and $\psi_{VH}$ do not necessary coincide. The phase difference between them is of the form $A \times NB$ where $A$ is the area (counted in number of squares) enclosed by the stripes.[]{data-label="fig-path"}](figures-1.eps)
This “area rule” can be easily generalized to any path used to extend $\psi$ from the fundamental square to any target square. This rule will play a very important role in the following.
Since the extensions are multivalued, the space on which we want to introduce them cannot be $\gR^2$. The correct space on which an unique extension of $\psi$ can be defined is represented in fig. \[fig-tree\]. In this (infinite) tree representation, the squares have been shrunk to vertices, and the links connecting these vertices indicate that two squares are adjacent.
![This tree, denoted by $\cT$, represents the direct space one can use to define a monovalued extension of the wave function $\psi \in \cD$.[]{data-label="fig-tree"}](figures-2.eps)
This tree is constructed from the fundamental square (with removed corners), representented by the vertex at the center, by connecting four squares (with removed corners) at its four edges, and then by iterating this proceduce on any new introduced square, with the important rule that none of these squares are identified with any other, even if they have the same location in $\gR^2$. In the representation of fig. \[fig-tree\], at each step, a scale factor has been applied to avoid some overlapping. If one tries to represent this construction with squares and without any scale factor, one need at least a 3-dimensionnal representation, in which each square in $\gR^2$ is replaced by a stacking of an infinite number of copies. Even the fundamental square is replaced by an infinite number of copies. Mathematically, this tree corresponds to the graph of a two generators free group. One of these generators acts by translation in one direction, and the other as a translation in the orthogonal direction. This space is also the universal covering space of $\gR^2 \setminus \gZ^2$, the space $\gR^2$ where all the points with integer coordinates has been removed. We will denote by $\cT$ this covering space. Then we have a projection map $p : \cT \rightarrow \gR^2 \setminus \gZ^2$, and the inverse image of any square (without corners) $p^{-1}([m,m+1]\times [n,n+1])$ is an infinite number of squares, which we will call the “covering squares” of $[m,m+1]\times [n,n+1]$. This last caracterization makes $\cT$ the right space to extend the wave functions $\psi \in \cD$ into a wave function $\widetilde{\psi}$ on $\cT$. Indeed, for any $\psi \in \cD$, we have introduced a procedure to extend $\psi$ on any square in $\gR^2 \setminus \gZ^2$. This extension depends on the path used to build $\psi$ step by step starting from the fundamental square and going to the final square. Because in the universal covering space of $\gR^2 \setminus \gZ^2$ two such paths never meet, we can extend $\psi$ without ambiguities into a monovaluated wave function $\widetilde{\psi}$ on $\cT$. Notice that each path connecting two squares can be decomposed into a ordered family of elementary translation on $\cT$ (which goes from one square to one of its adjacent squares). These elementary translations are representation of the two generators of a free group. This explain why $\cT$ is related to such a group.
When $NB$ is of the form $2 \pi r$ for an *irrationnal* $r$ (which we call the “irrationnal case” in the following), the area rule tells us that the phase difference for the values of the extended wave function on two squares covering a square in $\gR^2 \setminus \gZ^2$ is never a multiple of $2 \pi$. This means that the values of $\widetilde{\psi}$ on the covering squares in any stack are always different (except if $\widetilde{\psi}$ vanishes there). On the contrary, when $NB$ is of the form $2 \pi \frac{\ell}{k}$, for any rationnal $\frac{\ell}{k}$ (the “rationnal case”), the phase difference can be a multiple of $2 \pi$ if the area is a multiple of $k$. It is desirable to identify two squares when the values of the extension of $\psi$ coincide on them. This is the purpose of the next section.
The rationnal case {#rationnalcase}
==================
In this section, we assume that $NB=2 \pi \frac{\ell}{k}$ where $\frac{\ell}{k}$ is an irreducible rationnal number. When $NB=2 \pi \frac{\ell}{k}$, it appears that the relevant space on which a monovalued extension of the wave function $\psi$ is defined is smaller than $\cT$. This space is obtained from $\cT$ by identifying some covering squares by using a simple rule that will be explained in a while. This space is a compact Riemann surface on which some points are removed (punctures), and depends only on the integer $k$. We turn now on the complete characterisation of these surfaces.
Recall that the main motivation for introducing $\cT$ was to obtain a monovaluated extension $\widetilde{\psi}$ of $\psi \in \cD$. As already noticed in section \[generalcase\], the space $\cT$ is well suited to deal with the irrational case and cannot be reduced to a smaller one. In the present case $NB=2 \pi \frac{\ell}{k}$, the situation is somehow different because the values of the extension of $\psi$ on differents covering squares over the same square in $\gR^2 \setminus \gZ^2$ can be equal. For instance, the values of the functions $\psi_{HV}$ and $\psi_{VH}$ in fig. \[fig-path\] coincide when $mn$ is a multiple of $k$.
We identify two covering squares over the same square in $\gR^2 \setminus \gZ^2$ when the values of $\widetilde{\psi}$ are equal on them (for any $\psi \in \cD$). Then, using the area rule, it is easy to show that the number of covering squares after this identification has been performed is exactly $k$. We denote by $\cT_k$ the corresponding space which is a $k$-fold covering of $\gR^2 \setminus \gZ^2$, and is no longer simply connected.
Let us illustrate the situation in the simplest case $k=2$, as shown in the left graphic of fig. \[fig-k2\]. There, four squares in $\gR^2 \setminus \gZ^2$ are considered. Each one has two covering squares, represented by the four stacks of two squares. A link connecting two squares means that these two squares are adjacents. Notice that starting with the top-upper-left square, and following the links, we have to go round *twice* in order to come back, which illustrates the area rule. Obviously, $\cT_2$ is composed of an infinity of such stacks of two squares, and the links which connect these squares.
Now, on $\cT_k$, it is easy to show that the extended wave function $\widetilde{\psi}$ satisfy the boundary relations
$$\begin{aligned}
\label{boundarytildepsix}
\widetilde{\psi}(x+k, y, \gamma, \eta) &= e^{i k \gamma + \frac{i}{2} 2 \pi \ell y} \widetilde{\psi}(x, y, \gamma, \eta)\\
\label{boundarytildepsiy}
\widetilde{\psi}(x, y+k, \gamma, \eta) &= e^{i k \eta - \frac{i}{2} 2 \pi \ell x} \widetilde{\psi}(x, y, \gamma, \eta)\end{aligned}$$
because the phases “$p NB$” or “$q NB$” (see for instance formula (\[boundarypsiVHy\]) and (\[boundarypsiHVx\])) are exactly $2\pi \ell$ when $p=q=k$ and then can be eliminated. With these boundary relations, it is possible to interpret $\widetilde{\psi}$ as a section of a line vector bundle over a manifold that we describe now. These boundary relations reveal a periodicity in the variable $x$ and $y$ of length $k$. This means that $\cT_k$ can be reduced by identifying the extreme edges of any stripe of length $k$ ($k$ adjacent squares in the same direction). We denote by $\cS_k$ the manifold obtained from $\cT_k$ by these identifications.
For instance, for $k=2$, this means that the left edge of the upper-left square is identified with the right edge of the upper-right square (the two squares being on top of their stack). Performing this identification for any square, one obtains the middle graphic of fig. \[fig-k2\], where the squares have been shrunk to some points, and the curved links represent these last identifications. The last graphic on the left represents an unfold version, which is topologically equivalent to $\cS_2$.
![Case $k=2$: a $2\times 2$ extract of $\cT_2$ on the left, $\cS_2$ in the middle and on the right.[]{data-label="fig-k2"}](figures-3.eps)
For $k=3$, it is still possible to represent graphically these steps, as done in fig. \[fig-k3\]: a $3\times 3$ extract of $\cT_3$ on the left and $\cS_3$ on the right. Notice that on $\cT_3$, if one follows some links from square to square, one need to enclose an area which must be a multiple of 3 to come back at the beginning. This is an illustration of the area rule.
![Case $k=3$: a $3\times 3$ extract of $\cT_3$ on the left, $\cS_3$ on the right.[]{data-label="fig-k3"}](figures-4.eps)
As it can be suspected on fig. \[fig-k3\], when $k$ takes higher values, it is very difficult to represent $\cS_k$. Nevertheless, it is possible to completely caracterize the topology of $\cS_k$ for any $k$.
Let us give some general properties of $\cS_k$ that can be obtained directly from its construction. First $\cT$ is the universal covering of $\cS_k$, and $\cT_k$ is an intermediaire covering. The two dimensionnal manifold (an orientable finite type Riemann surface [@FarKra:80]) $\cS_k$ is constructed by glueing together $k^3$ “elementary” squares, the final surface does not have boundary, but it has $k^2$ punctures. These punctures come from the removed points of $\gR^2 \setminus \gZ^2$. Beside the $k$-fold covering of $\cT_k$, each point in $\gZ^2$ is removed only once because in $\cT_k$ one can circulate once around such a point by following the links connecting the squares for which this point is a corner. This can be verified on fig. \[fig-k2\] or fig. \[fig-k3\]. This is just a consequence of the area rule.
For $k=1$, this construction can be done very easily, and $\cS_1$ is just the ordinary torus with one puncture. This case corresponds to the study presented in (I) when $NB=2 \pi \ell$. In that study, the puncture was removable (because $k=1$), and the entire torus was our base space for the line vector bundle for which the wave function was a section. For $k=2$, the glueing of the $8$ elementary squares is also possible, and gives rise to the Riemann surface of genus $3$ (a sphere $\gS^2$ with $3$ handles) with $4$ punctures. The case $k=3$ is also tractable, and produces the Riemann surface of genus $10$ with $9$ punctures.
We give in Appendix \[pi1\] a proof of the following result:
*$\cS_k$ is a Riemann surface of genus $g_k=\frac{1}{2}(2 + k^2(k-1))$ with $k^2$ punctures.*
Let us comment the results obtained in this section. When $NB=2\pi \frac{\ell}{k}$, the wave functions have a natural interpretation in terms of sections of a line vector bundle over a Riemann surface with $k^2$ puncture $\cS_k$. When $k=1$, which is that case studied in (I), $\cS_1$ is a torus with one puncture and the explicit computation of the transverse conductivity $\sigma_{xy}$ using a geometrical interpretation of the Kubo formula was possible. The extension to the general situation where $\cS_k$, $k>1$, comes into play, becomes extremely more complicated even for $k=2$, so that a direct computation based on the Kubo formula would be untractable. However, this technical difficulty can be partially overcome as we will see in a while.
As explained above, the Riemann surface $\cS_k$ is constructed by gluing $k^3$ copies of the “fundamental square” which is nothing but the real physical sample involved in the experiment submitted to the external magnetic field. However, $\cS_k$ is just a mathematical tool which indicates that the relevant wave functions, as sections of some line vector bundle, must satisfy some phase coherence on a distance larger that the length of the real physical sample (because $\cS_k$ is constructed from $k$ copies of the fundamental square in the $x$ and $y$ directions). This last observation will be examined more closely in the next section.
The total magnetic flux through the fundamental square is $NB=2\pi \frac{\ell}{k}$ (as seen by the center of mass of the system). This means that $\cS_k$ is submitted a flux $\Phi = NB\times k^3= 2 \pi \ell k^2$. If $\cS_k$ were closed (without punctures), this outgoing flux could only be produced by a monopole inside the Riemann surface. But thanks to the presence of $k^2$ punctures, this flux can penetrate into the surface, each puncture carrying a penetrating flux of $2 \pi \ell k^2/k^2= 2 \pi \ell$.
Due to the complexity of $\cS_k$, the direct computation of the transverse conductivity from a Kubo formula cannot be performed so that one has to evaluate it with an other method. Using simple arguments and natural physical hypothesis, it is possible to rely the computation of $\sigma_{xy}$ to the result we have obtained in (I). Indeed, the relation $NB=2\pi \frac{\ell}{k}$ can be written as $kNB=2 \pi \ell$. This can be interpreted as stemming from a system with $kN$ charge carriers in the magnetic field $B$. The Hamiltonian for these $kN$ particules is the direct sum of $k$ copies of the Hamiltonian (\[totalhamiltonian\]). In this new Hamiltonian, the charge carriers in differents copies do not interact. This means that the interaction potential is not exactly the same as the one used in (\[totalhamiltonian\]). Nevertheless, in the computation of $\sigma_{xy}$ performed in (I), the *explicit* expression for the potential was needed when calculating $\sigma_{xy}$: only some gap properties were assumed in order ensure the validity of the Kubo formula. Let us now make the natural physical hypothesis that the new potential has the properties needed for the Kubo formula to be used. Then, because the relation $kNB=2 \pi \ell$ holds, the use of formula (45) of (I), yields $$\sigma_{xy}(kN,B)=2 \frac{(kN)^2}{2\pi \ell}$$ Now, the conductivity for $kN$ charge carriers is $k$ times the conductivity of $N$ charge carriers: $\sigma_{xy}(kN,B)=k \sigma_{xy}(N,B)$. Then one has $$\label{sigmaxy}
\sigma_{xy}(N,B)=2 \frac{(kN)^2}{2\pi \ell}\frac{1}{k} = 2 \frac{N^2 k}{2\pi \ell} = 2\frac{N}{B}$$ As expected, provided the physical hypothesis made on the potential are valid, the result coincides with the one obtained in (I).
Physical discussion and conclusion {#physicalcomments}
==================================
In the previous section, we have shown that for all the rational values of $NB/2 \pi$, says $NB/2\pi = \ell/k$, the transverse conductivity $\sigma$ takes integers or fractionnal values while the wave functions in $\cD$ are rigidily linked with the topological and geometrical structures. The purpose of this section is to discuss heuristically what additionnal physical information could be extracted from these mathematical structures when $NB/2 \pi$ differs from a rational value, taking into account some plausible physical assumptions.
One of the main features of the real Quantum Hall experiments is the occurence of plateaux in the transverse conductivity when $B$ varies while no plateau can show up within the system described by fig. \[fig-general\] and (\[totalhamiltonian\]) (plus boundary conditions) that we have analysed up to now, since $\sigma$ is linear in $B$ for all the rational values of $NB/2 \pi$. Consider the case of integer QHE. Then, it is believed that charge carriers are localized by disorder effects (which are basically induced by the presence of impurities in the Hall sample). This can be shown in Landau type models on $\gR^2$ in which the localized states modify the Hamiltonian spectrum. At least in a weak disorder regime, for which the typical energy of the disorder potential is small compared to the cyclotron energy, each Landau level is broadened by disorder giving rise to a Landau band associated with an extended state with energy centered around the corresponding Landau energy. Now, when $B$ varies in such a way that the system stays in such a band, the number of states involved in the charge transport remains constant so that the resulting conductivity is constant.
A rigourous extension of the present analysis to take into account the disorder effects through the introduction of a suitable random potential in (\[totalhamiltonian\]) is a task that goes beyond the scope of the present paper. In the following, we will not perform a mathematical analysis but instead we will propose a heuristic scenario which might correspond to the situation with non zero disorder, introducing for this purpose some plausible assumptions. The main point underlying our discussion has in fact been implicitely suggested in [@NiuThouWu:85]. In fact, these autors gave convincing arguments suggesting that in the presence of disorder[^5] the system retains (a sufficient amount of its) underlying geometrical structure so that the transverse conductivity remains quantized. Keeping this in mind, let us assume that the following two hypothesis are valid:
i)
: The geometrical structures exhibited in the previous sections for rational values of $NB/2 \pi$ are not altered when a weak disorder potential is incorporated so that disorder effects would manifest themselves in the boundary conditions for the wave functions as phase changes compensating deviations of $NB/2 \pi$ from its rational values.
ii)
: The wave function $\widehat{\psi}(x,y,\gamma, \eta)$ for non zero disorder are related to those at zero disorder through a phase change $\xi(x,y)$ as $$\label{phase}
\widehat{\psi}(x,y,\gamma, \eta) = e^{i \xi(x,y)} \psi(x,y, \gamma, \eta)$$ which is suggested by scattering theory.
Now, in a weak disorder regime, the phase change $e^{i \xi(x,y)}$ in (\[phase\]) can be expected to be close to unity,[^6] $e^{i \xi(x,y)}\simeq 1$, so that $\xi(x,y)$ can be parametrized by $\xi(x,y) = 2 \pi r + \varepsilon(x,y)$ where the integer $r$ depends on the location $(x,y)$ and $\varepsilon(x,y)$ is bounded: $|\varepsilon(x,y)| \leq \delta/2$. In the spirit of hypothesis i), $\delta$ represents the upper value of the disorder contribution for which the geometrical structures exhibited in the previous sections are preserved giving rise to a quantized transverse conductivity. Then using (\[phase\]), the boundary conditions (\[boundarypsitorus1\])–(\[boundarypsitorus2\]) on the wave functions relevant for the case $NB/2 \pi$ integer must be given by:
$$\begin{aligned}
\label{boundarypsihattorus1}
\widehat{\psi}(x+1,y,\gamma,\eta) &= e^{i \gamma + \frac{i}{2} NB y} e^{i(\varepsilon(x+1,y) - \varepsilon(x,y))} \widehat{\psi}(x,y,\gamma,\eta) \\
\label{boundarypsihattorus2}
\widehat{\psi}(x,y+1,\gamma,\eta) &= e^{i \eta - \frac{i}{2} NB x} e^{i(\varepsilon(x,y+1) - \varepsilon(x,y))} \widehat{\psi}(x,y,\gamma,\eta)\ .\end{aligned}$$
Then, $\widehat{\psi}$ can be extended in a unique way on the whole $\gR^2$ plane provided $$e^{i(NB + \varepsilon(x,y+1) - \varepsilon(x+1,y))} = 1$$ or $$\label{extensioncondition}
NB + \varepsilon(x,y+1) - \varepsilon(x+1,y)) = 2\pi \ell$$ Since one has $|\varepsilon(x,y)| \leq \delta/2$, (\[extensioncondition\]) gives rise to the range of variations for $NB/2\pi$ around an integer value for which, according to hypothesis i), the transverse conductivity is quantized; namely, one obtains $$\label{inequalityinteger}
|NB - 2\pi \ell| \leq \delta$$ From a similar analysis applied to the boundary conditions relevant for the rational case, it can be easily realized that (\[inequalityinteger\]) becomes $$\label{inequalityfraction}
\left|NB - 2\pi \frac{\ell}{k}\right| \leq \delta$$ In view of (\[inequalityfraction\]), the main assumptions i), if finally correct, leads to the physical conclusions that the transverse conductivity stays fixed to its quantized value (\[sigmaxy\]) for $B$ varying in the range $\left[ -2 \pi \frac{\ell \delta}{N}, 2 \pi \frac{\ell \delta}{N} \right]$ for fixed $N$, which clearly can be viewed as a plateau.
The use of formula (\[inequalityinteger\]) permits one to obtain some insight into the global behavior of the physical system studied in this paper when $B$ is varied and the contribution stemming from the disorder comes into play. Recall that $\nu=\frac{N}{N_\phi}$ where $N_\phi$ is the number of magnetic fluxes through the sample and that in our units, $\frac{1}{2\pi} = e^2/h$ so that $N_\phi = B/2\pi$. Let us assume that the number of electrons $N$ is kept fixed to some value, says $N=N_0$.[^7] Then, from the analysis of section \[rationnalcase\], we know that when $B$ reaches the value $B^\ast_{\ell/k} = \frac{2\pi}{N_0}\frac{\ell}{k}$, the transverse conductivity must be $\sigma^\ast_{\ell/k} = \frac{e^2}{h} \left(2 N_0^2 \frac{k}{\ell}\right) = \left( \frac{e^2}{h}\right) s_0$ for integer or fractional $s_0$. Now, according to the analysis presented at the beginning of this section, the occurence of disorder will not modify the value for the transverse conductivity, $\sigma = \sigma^\ast_{\ell/k}$, provided $$\label{inequality}
\frac{2}{s_0} - \Delta \leq \frac{1}{\nu} \leq \frac{2}{s_0} + \Delta$$ still holds, where $\Delta = \frac{e^2}{h} (\delta / N_0^2)$ and the expression for $\sigma^\ast_{\ell/k}$ together with the definition of the filling factor $\nu$ have been used. Then, one concludes from (\[inequality\]) that each value for $s_0\in \gQ$ corresponds to a domain in the $\left( \frac{1}{\nu}, \Delta \right)$ plane whose associated state has transverse conductivity equal to $\sigma^\ast = (e^2/h) s_0$. Such a domain would therefore represent a phase in the $\left( \frac{1}{\nu}, \Delta \right)$ plane which could be fully characterized from the properties of its corresponding wave function and energy derived from the Hamiltonian (\[totalhamiltonian\]) enlarged with a disorder potential. In particular, this would permits one to compare the respective energies of states corresponding to the (partially) overlapping stemming from (\[inequality\]) in the $\left( \frac{1}{\nu}, \Delta \right)$ plane in order to determine which state is energically favored with respect to the other ones, therefore removing the spurious physically meaningless overlaps. Unfortunately, the full corresponding Schrödinger problem cannot be explicitely solved in the present case.
Besides, we further note that the value of the transverse conductivity we have found is twice the value which is expected in realistic Quantum Hall experiments. We have carefully checked that this factor 2 does appear when calculating *in extenso* the conductivity through the Kubo formula[^8] for the system depicted on fig. \[fig-general\] involving spinless charged particles and submitted to the boundary conditions (\[boundarypsi12\]). Stricto sensu, realistic Quantum Hall systems involve spin carrying electrons in a strong magnetic field and are therefore somehow different than the idealized system considered here for which the magnitude of the magnetic field seen by the spinless charge carriers is not specified. It is concevable that the extension of the present study to the case of charge carriers with spin may lead, in some strong magnetic field limit, to some mechanism and/or degeneracy of the ground state that triggers the compensation of the above mentionned factor 2. If we assume that this finally happens, then eq. (\[inequality\]) now becomes $$\label{inequality2}
\frac{1}{s_0} - \Delta \leq \frac{1}{\nu} \leq \frac{1}{s_0} + \Delta$$ where $\Delta$ has been defined above and is associated with the disorder. Consider now the following odd denominator states $s_0 = 1/3,\ s_0=2/5,\ s_0=2/7$ for which the corresponding energies $E_{s_0}$ have been evaluated to be such that $E_{1/3} < E_{2/7} < E_{2/5}$ [@LHDM]. This, combined with (\[inequality2\]) suggests the phase architecture in the $\left( \frac{1}{\nu}, \Delta \right)$ plane depicted on fig. \[fig-3phases\]. In this figure, we have used the plausible assumption that when two domains overlap, the indexing stable state corresponds to the one with lowest energy. It can be checked that successive actions of the Landau shift operator $\nu \mapsto \nu +1$ and of the flux attachment operator $\frac{1}{\nu} \mapsto \frac{1}{\nu} + 2$ generate a phase diagram similar to the one proposed in [@CGMW:02].
![Phase architecture in the $\left( \frac{1}{\nu}, \Delta \right)$ plane for the 3 states $s_0 = 1/3,\ s_0=2/5,\ s_0=2/7$.[]{data-label="fig-3phases"}](figures-6.eps)
Within the mathematical formalism introduced in (I), salient geometrical structures appear, depending on the values reached by the ratio $NB/2\pi$. These structures control possibly the (integral or fractionnal) quantization of the Hall conductivity. The simplest situation corresponds to integer values for $NB/2\pi$ and gives rise, as shown in (I), to a Hall conductivity taking integer or fractionnal values. In the present paper, we have analysed some of the main (new) geometrical features underlying the Hall system described in section \[previousresults\] in the more complicated situation where $NB/2\pi$ differs from an integer, paying special attention to the geometrical characterization of the space on which monovaluated wave functions can be defined. When $NB/2\pi$ takes a rational value $\ell/k$, this space is a Riemann surface that we have completely determined in term of the number of punctures and genus, both related to the denominator of the rational value. Provided the potential satisfies an additional condition, the explicit computation of the Kubo formula performed in (I) when $NB/2\pi$ is integer can be extended to rational values for $NB/2\pi$ and gives rise again to Hall conductivity taking integer or fractionnal values. When $NB/2\pi$ is irrationnal, we have found that the above space can be related to the graph of a free group with two generators. In that latter case however, the (explicit) computation of the Kubo formula cannot be performed. Finally, we have discussed from a phenomenological viewpoint what might be expected for a Hall system with $NB/2\pi$ differing slightly from a rational value.
We presented some physical arguments suggesting that interesting physical features (to be extracted) should be encoded in our formalism. This suggests to examine more closely the irrationnal case and in particular to determine how the geometrical structures that occur in the rationnal case evolve when $NB/2\pi$ differs from rationnal values. Presumably, those “deformations” should be better treated within an algebraic framework. We notice that an indication in this direction are already encoded in the unitary symmetries that we have exhibited in (I). Recall that these latter connect rigidely (structures in) the direct space to (structures in) the reciprocical space. Besides, the two operators generated the unitary symmetries bear some similarity with the irrationnal rotation algebra, the $C^\ast$-algebra introduced in [@Rief:81; @Conn:86]. As indicated in (I) and in the present paper, in the direct space quantization of $NB/2\pi$ relies on the existence of some geometrical fiber bundle, while in the reciprocical space, the quantization of the Hall conductance stemms from some geometrical counterpart fiber bundle that is induced by the unitary symmetries. Now, from an algebraic viewpoint, this is strongly reminiscent of $K$-theory in direct space connected to $K$-theory in reciprocical space. But a natural framework to deal with both $K$-theories *at the same time* is provided by the $KK$-theory of Kasparov (for a review see e.g. [@Skan:91]). The corresponding implications are presently under study.
Topological characterization of $\cS_k$ {#pi1}
=======================================
In this appendix, we characterize the Riemann surface with puncture $\cS_k$ introduced in section \[rationnalcase\]. The only missing information about this surface is its genus: from its construction, one knows that it is orientable, without boundary and with $k^2$ punctures. A elegant way to calculate its genus is to compute its first homotopy group $\pi_1(\cS_k)$. Indeed, using standard techniques in algebraic topology and the well known theorem of van Kampen, one can easily rely the genus $g$, the number of punctures $t$ and the number of generators $n$ of the group $\pi_1(\cS_k)$ by the formula $$n = 2g+t-1$$ In order to get $g$, we only need to compute the number of generators of $\pi_1(\cS_k)$. This can be done using the van Kampen theorem. Recall that this theorem permits one to compute the first homotopy group of a space $X$ using a decomposition $X = U \cup V$, where $U, V$ and $U \cap V$ are open, nonempty and arcwise connected, by the relation $$\pi_1(X) \simeq \pi_1(U) \star_{\pi_1(U\cap V)} \pi_1(V)$$ where $G_1 \star_H G_2$ is the free group amalgation of the group $G_1$ by $G_2$ over a group $H$ when there exists two morphisms $\phi_i : H \rightarrow G_i$. We refer to [@Bred:95] for the complete definition. We will make use of this formula when $\pi_1(U\cap V)$ is trivial, which means that the product is a free product. This means that the number of generators of $\pi_1(X)$ is the sum of the number of generators of $\pi_1(U)$ and $\pi_1(V)$.
The idea to compute $\pi_1(\cS_k)$ is to reconstruct the space $\cS_k$ by adding, step by step, the $k^3$ squares from which it is composed. The first space we take is a circle, composed of the $4k(k-1)$ peripherical squares of the $k^2$ stacks. As illustrated on fig. \[fig-k2\] and fig. \[fig-k3\], the links connecting these squares form a global circle. This property is easily demonstrated using the area rule. Let us now decompose these $k^2$ stacks of square into successive shells from outside to inside, as illustrated on fig. \[fig-computation\]. The shell labeled by $s=0$ is the one just described, composed of $4k(k-1)$ squares. The next one, labeled by $s=1$, is composed of $4(k-2-1)$ stacks, which gives $4k(k-2-1)$ squares. The shell labeled $s$ is composed of $4(k-2s-1)$ stacks, and $4k(k-2s-1)$ squares. The last internal shell depends on the parity of $k$: for $k$ odd it constains $1$ stack of $k$ squares, and for $k$ even, $4$ stacks. In the odd case, we will not consider the last shell as the one containing only $1$ stack, because it plays a different role in the following computation, but the one just before, containing $8$ stacks, which corresponds to $s=\frac{k-3}{2}$. In the even case, the last shell corresponds to $s=\frac{k-2}{2}$. This means also that we have to consider $k>3$ in the following analysis.
![General situation for generic $k$, with emphase for the last (inner) shell which depends on the parity of $k$.[]{data-label="fig-computation"}](figures-5.eps)
Let us now explain the reconstruction method. In the first step, we take $U$ to be the circle composed of the $4k(k-1)$ peripherical squares. First, we wish to add a square in the corner of the shell $s=1$. This square is connected to the shell $s=0$ by two links, as symbolised on fig. \[fig-computation\]. Notice that these two links do not connect the square to the same level in the stacks of the shell $s=0$. It is possible to close these links along the shell $s=0$, and then the space $V$ we take is just topologically a circle. When the four squares in the corner of the shell $s=1$ has been added, one can add the other squares one by one exactly by the same procedure. At each step the space $U$ is growing, and the space $V$ is topologically equivalent to a circle. We can repeat this procedure for all the shells. One has also to add the external links connected to the shell $s=0$, which are also topologically some circles.
Let us compute the number of generators of the first homotopy group from this construction using the van Kampen theorem. At each step, the space $V$ is a circle, which means that $\pi_1(V)$ has one generator, and the intersection $U\cap V$ is contractible, which means that $\pi_1(U\cap V)$ is trivial. Then at each step, the free group amalgation consists to add exactly one generator to the homotopy group. In order to get the number of generators of $\pi_1(\cS_k)$, one has to compute the number of circles one has to add to reconstruct $\cS_k$ from the original circle.
The shell $s=0$.
: This shell is a circle, it contributes to $1$ generator. This is the first space $U$.
The external links.
: Each corner of the shell $s=0$ is connected to exactly two other corners: they contribute to $4k$ generators. Each other square is connected to exactly one other square: they contribute to $2k(k-2)$ generators.
The generic shell $s$.
: There are $4k$ corners which contribute each to $1$ generator. For the other squares, one has to connect then to the shell $s-1$ and to the corners. A carefull analysis shows that there are $4k(k-2s-1)$ circles to add to connect them.
Let us treat separetely the odd and even cases for the last shell.
The odd case $k=2p+1$.
: One has to add one stack of $k$ squares to the shell $s=\frac{k-3}{2}=p-1$. Each square in this stack is connected to $4$ squares of the previous shell (see for instance the left graph of fig. \[fig-k3\]), but one need only to add $3$ circles to connect such a square. So this shell contributes to $3k$ generators.
The even case $k=2p$.
: One has to connect $4$ stacks of $k$ squares. These squares are corners, so they contribute to $4k$ generators to connect them to the previous shell. These squares are linked together, this adds $4k$ circles. Notice that this last contribution is the same as the contribution of a generic shell with $s=\frac{k-2}{2}$: $4k(k-2\frac{k-2}{2}-1) = 4k$. This means that this contribution can be added to the generic one.
Let us collect all these contributions in the two cases:
The odd case $k=2p+1$.
: $$1 + 4k + 2k(k-2) + \sum_{s=1}^{p-1} \left( 4k + 4k(k-2s-1) \right) + 3k = 1 + k^3$$
The even case $k=2p$.
: $$1 + 4k + 2k(k-2) + \sum_{s=1}^{p-1} \left( 4k + 4k(k-2s-1) \right) = 1 + k^3$$
The result is independant of the parity of $k$.[^9] Moreover, this computation supposes $k>3$, but the number of generators of $\pi_1(\cS_k)$ is also $1+k^3$ when $k=1,2,3$ as can be directly verified. So, the main result is that
*the number of generators of $\pi_1(\cS_k)$ is $1+k^3$.*
Knowing that $(\cS_k)$ has $k^2$ punctures, this implies that
*the number of handles of $\cS_k$ is $g_k=\frac{1}{2}(2 + k^2(k-1))$.*
[99]{}
R.E. Prange and S.M. Girvin eds., “The Quantum Hall Effect,” Springer-Verlag, 1990, and references therein.
M. Stone ed., “Quantum Hall Effect,” World Scientific, 1992, and references therein.
R.B. Laughlin, [*Phys. Rev. B*]{} **23** (1981), 5632.
D.J. Thouless, M. Kohmoto, M.P. Nightingale and M. den Nijs, [*Phys. Rev. Lett.*]{} **49** (1982), 405.
M. Kohmoto, [*Annals of Physics*]{} **160** (1985), 343.
Q. Niu and D.J. Thouless, [*J. Phys.*]{} **A17** (1984), 2453.
Q. Niu, D.J. Thouless and Y.-S. Wu, [*Phys. Rev. B*]{} **31** (1985), 3372.
J.E. Avron and R. Seiler, [*Phys. Rev. Lett.*]{} **54** (1985), 259.
D.J. Thouless, [*Phys. Rev. Lett.*]{} **57** (1983), 57.
X.G. Wen and Q. Niu, [*Phys. Rev. B*]{} **41** (1990), 9377.
B.I. Halperin, [*Phys. Rev. B*]{} **25** (1982), 2185.
F. Chandelier, Y. Georgelin, T. Masson, J.-C. Wallet [*Ann. Phys.*]{}, **305** (2003), 60.
Q. Niu and D.J. Thouless, [*Phys. Rev. B*]{} **35** (1987), 2188.
F. Chandelier, Y. Georgelin, T. Masson, J.-C. Wallet [*Phys. Lett.*]{}, **A301** (2002), 451.
see also Y. Georgelin, T. Masson, J.-C. Wallet [*J. Phys.*]{}, **A 30** (1997), 5065; Y. Georgelin, T. Masson, J.-C. Wallet [*J. Phys.*]{}, **A 33** (2000), 39; B.P. Dolan [*J. Phys.*]{}, **A 32** (1999), L243; B.P. Dolan [*Nucl. Phys.*]{}, **B 554** (1999), 487.
R. Kubo, [*J. Phys. Soc. Japan*]{} **12** (1957), 570; R. Kubo, [*Rep. Progr. Phys.*]{}, **29** (1966), 255.
J. Avron, R. Seiler and B. Simon, [*Phys. Rev. Lett.*]{} **51** (1983), 51; B. Simon, [*Phys. Rev. Lett.*]{} **51** (1983), 2167.
H. M. Farkas, I. Kra, “Riemann Surfaces”, Springer-Verlag, 1980.
see e.g. R.B. Laughlin in “The Quantum Hall Effect”, R.E. Prange, S.M. Girvin eds., Springer-Verlag (1990); see also B.I. Halperin, [*Phys. Rev. Lett*]{} **52** (1984), 1583; A.H. Mc Donalds and D.B. Murray, [*Phys. Rev. B*]{} **32** (1985), 2707.
M.A. Rieffel, [*Pacific J. Math.*]{}, **93**, (1981), 415.
A. Connes, [*Publ. Math. IHES*]{}, **62**, (1986), 257.
G. Skandalis, [*Exp. Math.*]{}, **9**, (1991), 193.
G. E. Bredon, “Topology and Geometry”, Springer-Verlag, 1995.
[^1]: assuming the existence of a spectral gap
[^2]: Some models consider both the sample and the threads as quantum mechanical objects which is questionnable from an experimental viewpoint.
[^3]: This torus in the direct space has no physical meaning and cannot be related in some way to the physical sample.
[^4]: As shown in (I), the final results do not depend on this gauge choice.
[^5]: at least for weak disorder
[^6]: see e.g. Prange in [@Ston:92] and ref. therein
[^7]: In the Quantum Hall experiment, this would be obtained by fixing the gate voltage controlling the electron density.
[^8]: Usual assumptions on the validity of the Kubo formula are assumed here.
[^9]: This may mean that our demonstration is not the most elegant one!
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present a multi-wavelength investigation of a large-scale physical system containing the W33 complex. The extended system ($\sim$50 pc $\times$ 37 pc) is selected based on the distribution of molecular gas at \[29.6, 60.2\] km s$^{-1}$ and of 88 ATLASGAL 870 $\mu$m dust clumps at d $\sim$2.6 kpc. The extended system/molecular cloud traced in the maps of $^{13}$CO and C$^{18}$O emission contains several H[ii]{} regions excited by OB stars (age $\sim$0.3–1.0 Myr) and a thermally supercritical filament (“fs1", length $\sim$17 pc). The filament, devoid of the ionized gas, shows dust temperature (T$_{d}$) of $\sim$19 K, while the H[ii]{} regions are depicted with T$_{d}$ of $\sim$21–29 K. It suggests the existence of two distinct environments in the cloud. The distribution of Class I young stellar objects (mean age $\sim$0.44 Myr) traces the early stage of star formation (SF) toward the cloud. At least three velocity components (around 35, 45, and 53 km s$^{-1}$) are investigated toward the system. The analysis of $^{13}$CO and C$^{18}$O reveals the spatial and velocity connections of cloud components around 35 and 53 km s$^{-1}$. The observed positions of previously known sources, W33 Main, W33 A and O4-7I stars, are found toward a complementary distribution of these two cloud components. The filament “fs1" and a previously known object W33 B are seen toward the overlapping areas of the clouds, where ongoing SF activity is evident. A scenario concerning the converging/colliding flows from two different velocity components appears to explain well the observed signposts of SF activities in the system.'
date:
title: Uncovering distinct environments in an extended physical system around the W33 complex
---
=1
\[firstpage\]
dust, extinction – HII regions – ISM: clouds – ISM: individual object (W33) – stars: formation – stars: pre–main sequence
Introduction {#sec:intro}
============
Over the past two decades, due to the availability of various large-scale multi-wavelength surveys, the research concerning the understanding of the formation processes of massive OB stars ($\gtrsim$ 8 M$_{\odot}$) and young stellar clusters has received immense interest, and is being debated [@zinnecker07; @tan14]. In this context, Galactic star-forming sites hosting mid-infrared (MIR) bubbles/shells associated with H[ii]{} regions and/or infrared dark clouds (IRDCs)/elongated filamentary features are thought to be prime targets for exploring the highlighted research problem [e.g., @churchwell06; @churchwell07; @andre10; @andre14]. The exploration of the existing multi-wavelength surveys allows to examine the physical conditions and the kinematics of embedded structures on large-scale areas of these sites, which hold crucial clues about the birth processes of clusters of young stellar objects (YSOs) and massive stars. In this connection, the target of this paper is a massive star-forming site W33, which is located in the inner Galaxy. Previous studies have reported that W33 hosts several massive OB stars and star-forming clumps, and filaments [@westerhout58; @immer13; @immer14; @maud17; @kohno18 and references therein].
Based on the radio line and continuum observations [@downes80; @haschick83; @lockman89; @anderson14; @anderson15] and the VLT/SINFONI near-infrared (NIR) spectroscopic observations [@messineo15], several ionized clumps/H[ii]{} regions and massive OB stars have been investigated in the W33 complex [see Figure 1 in @kohno18]. In the W33 complex, @messineo15 reported that the population of massive stars (including O4-7 type stars) formed during $\sim$2–4 Myr ago. At least six dust continuum clumps at 870 $\mu$m were detected toward the W33 complex within an area of 15 pc $\times$ 15 pc [e.g., @contreras13], which were referred to as W33 Main, W33 A, W33 B, W33 Main1, W33 A1, and W33 B1 [see Figure 1 in @kohno18]. @kohno18 identified two CO bipolar outflows toward W33 Main and W33 A, and the dynamical timescales of those outflows were computed to be $\sim$3 $\times$ 10$^{4}$ yr. @immer13 detected water masers toward W33 A (G012.90$-$0.24, G012.90$-$0.26) and W33 Main (G012.81$-$0.19) in a radial velocity (V$_\mathrm{lsr}$) range of 33–38 km s$^{-1}$, while the water masers in W33 B (G012.68$-$0.18) were observed in a velocity range of 57–63 km s$^{-1}$. Note that the radial velocity information has been found different for W33 B compared to W33 A and W33 Main. These authors suggested the kinematic distance of $\sim$3.7 kpc (corresponding to V$_\mathrm{lsr}$ of 36 km s$^{-1}$). They also reported the parallactic distance of the water masers in the massive star forming complex W33 (containing G012.68$-$0.18, G012.81$-$0.19, G012.90$-$0.24, G012.90$-$0.26) to be $\sim$2.4 kpc.
It was suggested that the distance estimated using the trigonometric parallax observations of maser sources is more reliable than the kinematic distance measurements [e.g., @reid09; @sato10; @immer13]. Hence, despite different V$_\mathrm{lsr}$ values, it has been considered that the sites (i.e., W33 A (G012.90$-$0.24, G012.90$-$0.26), W33 B (G012.68$-$0.18), and W33 Main (G012.81$-$0.19)) are located at the similar distance, and are part of the W33 complex [e.g., @immer13; @messineo15; @kohno18]. Based on the detection of methanol masers, @kohno18 also pointed out the presence of massive YSOs toward the sub-regions W33 Main, W33 A, and W33 B. Using the high-resolution molecular line data, inner environments of at least two sub-regions W33 A and W33 Main have been reported in the literature [@galvan10; @jiang15; @maud17]. To explain star formation (SF) activity in the direction of W33 A, @galvan10 proposed a triggered SF scenario by filamentary convergent gas flows from two different velocity components. In the direction of the W33 complex, @kohno18 reported three velocity components at 35, 45, and 58 km s$^{-1}$, and found the signatures of the collision between two clouds at 35 and 58 km s$^{-1}$. They assumed these clouds at a distance of $\sim$2.4 kpc. Considering this argument, a collision scenario was proposed to explain the massive SF in the W33 complex [see @kohno18 for more details]. @messineo15 also mentioned the possibility of sequential SF and feedback in W33. However, such study is not extensively carried out in W33. To our knowledge, earlier observational works toward W33 are restricted up to an area of 15 pc $\times$ 15 pc [see Figure 1 in @kohno18], and its large-scale environment (i.e., $>$ 35 pc $\times$ 35 pc) is yet to be examined.
In this context, Figure \[fig1\]a displays a large-scale area ($\sim$1$\degr$.1 $\times$ 0$\degr$.815) containing the W33 complex using the ATLASGAL 870 $\mu$m dust continuum map. Figure \[fig1\]a also indicates an area ($\sim$0$\degr$.4 $\times$ 0$\degr$.4) by a broken box, which was studied by @kohno18. The locations of W33 A (G012.90$-$0.24, G012.90$-$0.26), W33 B (G012.68$-$0.18), and W33 Main (G012.81$-$0.19) are also marked by filled circles in Figure \[fig1\]a. In Figure \[fig1\]a, the positions of 88 ATLASGAL clumps at 870 $\mu$m [from @urquhart18] are also marked by diamonds. @urquhart18 obtained velocities toward the ATLASGAL clumps and also tried to compute the distance information for the clumps. All these ATLASGAL clumps are depicted in a velocity range of \[30, 56\] km s$^{-1}$, and are located at a distance of $\sim$2.6 kpc [@urquhart18], which are in agreement with the previously published results [e.g., @immer13; @kohno18]. In this paper, we have adopted a distance of $\sim$2.6 kpc to the entire selected area around the W33 complex for all the analysis.
One can also notice that the ATLASGAL survey has allowed us to trace the boundary of an extended physical system hosting the W33 complex (area $\sim$50 pc $\times$ 37 pc; distance $\sim$2.6 kpc; see Figure \[fig1\]a). However, we do not find any study related to unearth the physical conditions and the kinematics of embedded structures on large-scale areas. In this paper, using a multi-wavelength approach, we aim to explore the ongoing physical processes in the extended physical system hosting W33. In particular, this observational work focuses to understand the formation mechanisms of massive stars and clusters of YSOs. In this connection, the study of molecular gas toward our selected extended system has been carefully carried out using the FOREST Unbiased Galactic plane Imaging survey with the Nobeyama 45-m telescope [FUGIN; @umemoto17] $^{13}$CO(J = 1–0) and C$^{18}$O (J = 1–0) line data.
The present paper is organized as follows. The information of the adopted data sets is given in Section \[sec:obser\]. In Section \[sec:data\], we present the observational findings derived using the multi-wavelength surveys. We discuss the possible SF processes in Section \[sec:disc\]. Finally, the main conclusions of this paper are presented in Section \[sec:conc\].
Data sets and analysis {#sec:obser}
======================
With the aid of the spatial distribution of 88 ATLASGAL clumps at a distance of $\sim$2.6 kpc, the selected target area ($\sim$1$\degr$.1 $\times$ 0$\degr$.815 (or $\sim$50 pc $\times$ 37 pc); centered at $l$ = 12$\degr$.946; $b$ = $-$0$\degr$.192) around W33 is presented in Figure \[fig1\]a. Table \[ftab1\] gives a summary of various multi-wavelength survey data adopted in this paper. In the selected wide-scale area, these surveys offer to examine the distribution of H$_{2}$ column densities, dust temperatures, dust clumps, ionized gas, YSOs, and gas kinematics as well as the embedded structures/morphologies.
In order to study the gas distribution toward the selected target area, the present paper employs the molecular $^{13}$CO (J=1–0) and C$^{18}$O (J=1–0) line data from the FUGIN survey [@umemoto17]. In the survey, the multi-beam receiver, FOur-beam REceiver System on the 45-m Telescope [FOREST; @minamidani16; @nakajima19], was used for the observations. The FUGIN molecular line data are calibrated in main beam temperature ($T_\mathrm{mb}$) [see @umemoto17 for more details]. The typical rms noise level ($T_\mathrm{mb}$) is $\sim$0.7 K for both the $^{13}$CO and C$^{18}$O lines [@umemoto17]. To improve sensitivities, we smoothed the FUGIN data cube with a Gaussian function having a half power beam width of 35$''$.
Survey band/line(s) Resolution ($\arcsec$) Reference
---------------------------------------------------------------------------------- -------------------------------- ------------------------ ------------- --
NRAO VLA Sky Survey (NVSS) 21 cm $\sim$46 @condon98
FUGIN survey $^{13}$CO, C$^{18}$O (J = 1–0) $\sim$20 @umemoto17
APEX Telescope Large Area Survey of the Galaxy (ATLASGAL) 870 $\mu$m $\sim$19.2 @schuller09
[*Herschel*]{} Infrared Galactic Plane Survey (Hi-GAL) 70–500 $\mu$m $\sim$5.8–37 @molinari10
[*Spitzer*]{} MIPS Inner Galactic Plane Survey (MIPSGAL) 24 $\mu$m $\sim$6 @carey05
[*Spitzer*]{} Galactic Legacy Infrared Mid-Plane Survey Extraordinaire (GLIMPSE) 3.6–8.0 $\mu$m $\sim$2 @benjamin03
Results {#sec:data}
=======
Wide-scale view around W33: environment and H[ii]{} regions {#subsec:radio}
-----------------------------------------------------------
Figure \[fig1\]a exhibits the spatial distribution of 88 ATLASGAL clumps (having d $\sim$2.6 kpc and V$_\mathrm{lsr}$ range $\sim$\[30, 56\] km s$^{-1}$) overlaid on the ATLASGAL continuum map at 870 $\mu$m, illustrating an extended physical system as mentioned earlier. Figure \[fig1\]b shows the overlay of the NVSS radio continuum emission contours on the MIPSGAL 24 $\mu$m image, revealing the locations of H[ii]{} regions in our selected system. The NVSS 1.4 GHz radio continuum data suggest the presence of the ionized gas, while the warm dust emission is traced in the 24 $\mu$m image. Additionally, one can also find the absorption features against the Galactic background in the 24 $\mu$m image, revealing the IRDCs (see arrows in Figure \[fig1\]b). The locations of previously identified at least three O4-7(super)-giant stars [see IDs \#7, 8, and 23; from @messineo15] are marked in Figures \[fig1\]a and \[fig1\]b (see filled stars). With the application of the [*clumpfind*]{} IDL program [@williams94] in the NVSS 1.4 GHz radio continuum map, we have selected 11 ionized clumps in our target area. Figure \[fig1\]c shows the boundary of each NVSS clump against the locations of earlier reported O4-7(super)-giant stars (see filled stars). A broken box is indicated in Figures \[fig1\]a, \[fig1\]b and \[fig1\]c, which shows the area investigated by @kohno18. The ionized clumps (see IDs \#c10, \#c8, and \#c1–c7) are found toward the MIR bubbles [i.e., N10 and N11; @churchwell06; @gama16], G013.210$-$0.144 [@white05], and the W33 complex (W33 A (G012.90$-$0.24, G012.90$-$0.26), W33 B (G012.68$-$0.18), and W33 Main (G012.81$-$0.19); see Figure \[fig1\]c), where we also find extended warm dust emission at 24 $\mu$m (see Figure \[fig1\]b). Two massive O4-7(super)-giant stars (i.e., \#7 and \#8) are seen toward the NVSS clump \#c3, while a massive O4-7(super)-giant star \#23 is found in the direction of the NVSS clump \#c4 (see Figure \[fig1\]c).
Following the work of @dewangan17, we have estimated the number of Lyman continuum photons [$N_\mathrm{uv}$; see @panagia73; @matsakis76] and the dynamical age [$t_\mathrm{dyn}$; see @dyson80] of each ionized clump (see Table \[tab2\]). Previously, @kohno18 adopted an initial particle number density of the ambient neutral gas (i.e., $n_{0}$ = 10$^{4}$ cm$^{-3}$) in the estimations of the formation timescale of the H[ii]{} regions G012.745$-$00.153 and G012.820$-$00.238, which was considered as a reasonable value for the W33 complex. Following the previous work of @kohno18, in this paper, we compute the ages of the ionized clumps for a value of $n_{0}$ = 10$^{4}$ cm$^{-3}$. The calculation uses the isothermal sound velocity in the ionized gas [$c_{s}$ = 11 km s$^{-1}$; @bisbas09], the radius of the H[ii]{} region ($R_{HII}$), $n_{0}$, and $N_\mathrm{uv}$. Dynamical ages of the ionized clumps vary between $\sim$0.35 – 1.0 Myr (see Table \[tab2\]). The bubble N10 (see ID \#c10 in Table \[tab2\]) is found to be excited by an O9.5V–O9V star (age $\sim$0.65 Myr), which is located away from the W33 complex. At least seven NVSS clumps (see IDs \#c1–c7) are located toward the W33 complex (area $\sim$15 pc $\times$ 15 pc; see a broken box in Figure \[fig1\]c), which contains OB stars (age $\sim$0.35 – 0.85 Myr). More description of the analysis can be found in @dewangan17. The implication of the ages of these ionized clumps is discussed in Section \[sec:disc\].
Dust temperature and column density maps {#sec:hermap}
----------------------------------------
In order to study the embedded structures in our selected target area, we have obtained the [*Herschel*]{} temperature and column density ($N(\mathrm H_2)$) maps from the site. The spatial resolution of these maps is $\sim$12$''$. Using the Bayesian [*PPMAP*]{} procedure operated on the [*Herschel*]{} data at wavelengths of 70, 160, 250, 350 and 500 $\mu$m [@marsh15; @marsh17], the [*Herschel*]{} temperature and column density maps were produced for the [*EU-funded ViaLactea project*]{} [@molinari10b].
In Figures \[fig2\]a and \[fig2\]b, we display the [*Herschel*]{} temperature and column density ($N(\mathrm H_2)$) maps of our selected target area, respectively. In both the [*Herschel*]{} maps, a broken box shows the area studied by @kohno18. The [*Herschel*]{} temperature map shows the existence of embedded structures with T$_\mathrm{d}$ $\sim$17–19 K in the system, and is overlaid with a temperature contour of 18.6 K, allowing us to depict filamentary structures (see labels fs1 and fs2 in Figure \[fig2\]a). The [*Herschel*]{} temperature map also exhibits extended features with T$_\mathrm{d}$ $\sim$21–29 K in the direction of the previously known H[ii]{} regions (see W33 A (G012.90$-$0.24, G012.90$-$0.26), W33 B (G012.68$-$0.18), W33 Main (G012.81$-$0.19), MIR bubbles, and G013.210$-$0.144).
In Figure \[fig2\]b, the column density contour (in white) is also overlaid on the $N(\mathrm H_2)$ map with a level of 3.35 $\times$ 10$^{22}$ cm$^{-2}$, indicating the presence of materials with high column densities. Using this $N(\mathrm H_2)$ contour level, we trace the boundary of the filament “fs1" (length $\sim$17 pc), which is also highlighted by a broken black curve in Figure \[fig2\]b. However, an elongated morphology of filament “fs2", as traced in the [*Herschel*]{} temperature map, is not seen by this $N(\mathrm H_2)$ contour level. One of the parts of the filament “fs2" is prominently depicted by this $N(\mathrm H_2)$ contour level (see a broken white circle). In the direction of the IRDCs highlighted in the image at 24 $\mu$m, both the [*Herschel*]{} maps indicate the presence of the embedded filaments. With the knowledge of T$_\mathrm{d}$ toward the filamentary structures and the H[ii]{} regions, two distinct environments are evident in the selected physical system.
Based on the analysis of the [*Herschel*]{} column density and temperature maps, we examine the stability of the elongated filament “fs1" (mass $\sim$51000 M$_{\odot}$; length $\sim$17 pc; T$_\mathrm{d}$ $\sim$19 K). To compute the mass of the filament/clump, we employed the equation, $M_{area} = \mu_{H_2} m_H Area_{pix} \Sigma N(H_2)$, where $\mu_{H_2}$ is the mean molecular weight per hydrogen molecule (i.e., 2.8), $Area_{pix}$ is the area subtended by one pixel (i.e., 6$''$/pixel), and $\Sigma N(\mathrm H_2)$ is the total column density [see also @dewangan17]. Following the works of @dewangan18, we have estimated the line mass or mass per unit length (i.e., $M_{\rm line,obs}$ $\sim$3000 M$_{\odot}$ pc$^{-1}$) and the critical line mass ($M_{\rm line,crit}$) of the filament “fs1". Due to unknown inclination angle of the filament, we consider the observed line mass as an upper limit. The expression of $M_{\rm line,crit}$ is given by $\sim$16 M$_{\odot}$ pc$^{-1}$ $\times$ (T$_{gas}$/10 K) for a gas filament, assuming that the filament is an infinitely extended, self-gravitating and isothermal cylinder without magnetic support [e.g. @ostriker64; @inutsuka97; @andre14]. The value of $M_{\rm line,obs}$ can be compared against the critical line mass $M_{\rm line,crit}$ of 16–32 M$_{\odot}$ pc$^{-1}$ at T = 10–20 K, suggesting that the filament “fs1" is thermally supercritical. In general, the thermally supercritical filaments are believed to be unstable to radial gravitational collapse and fragmentation [e.g., @andre10].
Star Formation Activities {#sec:sf}
-------------------------
In this section, to examine the ongoing SF activities, we have identified Class I YSOs in our selected target area around the W33 complex. In this context, the color-color plot (\[4.5\]$-$\[5.8\] vs \[3.6\]$-$\[4.5\]) is employed (not shown here), and the infrared color conditions (i.e., \[4.5\]$-$\[5.8\] $\ge$ 0.7 mag and \[3.6\]$-$\[4.5\] $\ge$ 0.7 mag) are adopted to depict Class I YSOs [see @hartmann05; @getman07; @dewangan17b; @dewangan18b; @dewangan18x for more details]. Photometric magnitudes of point-like sources at 3.6–5.8 $\mu$m were downloaded from the [*Spitzer*]{} GLIMPSE-I Spring’ 07 highly reliable catalog. In this work, we considered only sources with a photometric error of less than 0.2 mag in each [*Spitzer*]{} band.
Figure \[fig3\]a shows the overlay of the positions of 901 Class I YSOs on the ATLASGAL 870 $\mu$m map (see blue circles). In Figures \[fig3\]b and \[fig3\]c, we present surface density contours (in blue) of Class I YSOs overlaid on the ATLASGAL 870 $\mu$m dust continuum map and the NVSS 1.4 GHz radio continuum map, respectively. The surface density contours are shown with the levels of 1.2, 2, 3.5, and 6.5 YSOs/pc$^{2}$, where 1$\sigma$ = 1.1 YSOs pc$^{-2}$. A spatial correlation between the dust continuum emission and the Class I YSOs is also clearly seen in Figure \[fig3\]b. A majority of Class I YSOs [mean age $\sim$0.44 Myr; @evans09] are seen toward the [*Herschel*]{} filaments (i.e.,“fs1" and “fs2") and the H[ii]{} regions, indicating the presence of the early stage of SF in the selected physical system. The surface density analysis is performed using the nearest-neighbour (NN) method [see @casertano85; @gutermuth09; @bressert10; @dewangan17 for more details]. Following the similar procedures as reported in @dewangan17, we produce the surface density map of all the selected Class I YSOs using a 15$''$ grid and 6 NN at a distance of 2.6 kpc.
Kinematics of molecular gas {#sec:coem}
---------------------------
In Figure \[fig4\]a, we display the positions of 88 ATLASGAL clumps (having d $\sim$2.6 kpc and V$_\mathrm{lsr}$ range $\sim$\[30, 56\] km s$^{-1}$). Figure \[fig4\]b presents the distribution of V$_\mathrm{lsr}$ of 88 clumps against the Galactic longitude. In the direction of [*l*]{} = 13$\degr$.08–13$\degr$.4, this plot suggests the presence of at least two clouds (at \[30, 41\] km s$^{-1}$ (or around 35 km s$^{-1}$) and \[48, 56\] km s$^{-1}$ (or around 53 km s$^{-1}$)) containing the dust clumps. Furthermore, we also find two clumps with V$_{lsr}$ of $\sim$45 km s$^{-1}$ toward [*l*]{} $\approx$ 13$\degr$.1. A majority of the ATLASGAL dust clumps are associated with the gas at \[30, 41\] km s$^{-1}$. In the direction of our selected system, our analysis of the clumps indicates the presence of at least three velocity components. Previously, three cloud components at 35, 45, and 58 km s$^{-1}$ were also reported toward the W33 complex [see @kohno18]. In this section, we have analyzed the FUGIN molecular line data to further explore various cloud components. In Figure \[fig4\]c, we display the observed $^{13}$CO(J =1$-$0) and C$^{18}$O(J =1$-$0) spectra in the direction of an area highlighted by a solid box in Figure \[fig4\]a), revealing two noticeable velocity peaks (i.e., 35 and 53 km s$^{-1}$). These profiles are produced by averaging the selected area.
In the direction of the selected target area, Figures \[fig5\]a and \[fig5\]b display the integrated $^{13}$CO (J=1–0) and C$^{18}$O (J=1–0) intensity maps, respectively. In each intensity map, the molecular gas is integrated over a velocity range of \[29.6, 60.2\] km s$^{-1}$. Both the maps display an extended morphology of the molecular cloud associated with the selected target area, appearing like a giant molecular cloud (GMC). .
Figure \[fig6\] presents the integrated $^{13}$CO velocity channel maps (at intervals of 1.3 km s$^{-1}$). The distribution of $^{13}$CO gas in the channel maps suggests the existence of different velocity components in the direction of the selected physical system (see panels at \[36.82, 38.12\], \[44.62, 45.92\], and \[53.72, 55.02\] km s$^{-1}$). In Figures \[fig7\]a and \[fig7\]b, we show the longitude-velocity maps of $^{13}$CO and C$^{18}$O, respectively. Both the position-velocity maps depict three velocity components around 35, 45, and 53 km s$^{-1}$ (see arrows in Figures \[fig7\]a and \[fig7\]b). Similar result is also derived using the distribution of the radial velocities of the ATLASGAL clumps (see Figure \[fig4\]b). In the velocity space, the cloud components around 35 and 53 km s$^{-1}$ appear to be connected by a low-intensity intermediate velocity emission around 45 km s$^{-1}$ (see red broken curves toward [*l*]{} = 13$\degr$.1–13$\degr$.2 and [*l*]{} = 12$\degr$.7–12$\degr$.8 in Figure \[fig7\]a). Based on the observed different velocity components, Figures \[fig8\]a, \[fig8\]c, and \[fig8\]e show the spatial distribution of $^{13}$CO gas associated with the clouds at \[29.6, 43.3\], \[44, 46.5\], and \[47.2, 60.2\] km s$^{-1}$, respectively. In Figures \[fig8\]b, \[fig8\]d, and \[fig8\]f, we present the distribution of C$^{18}$O gas associated with the clouds at \[29.6, 43.3\], \[44, 46.5\], and \[47.2, 60.2\] km s$^{-1}$, respectively. In each panel of Figure \[fig8\], the positions of W33 A, W33 B, and W33 Main are also highlighted by open circles, while the positions of three O4-7I stars [see IDs \#7, 8, and 23 in @messineo15] are shown by filled stars (in cyan). The sites W33 A and W33 Main are seen with the highest intensities in the $^{13}$CO and C$^{18}$O maps at \[29.6, 43.3\] km s$^{-1}$, while the corresponding areas are seen with low intensities in the $^{13}$CO and C$^{18}$O maps at \[47.2, 60.2\] km s$^{-1}$. One can find a very weak $^{13}$CO and C$^{18}$O emission toward W33 B in the cloud at \[29.6, 43.3\] km s$^{-1}$, while the strong molecular emission is traced toward W33 B in the cloud at \[47.2, 60.2\] km s$^{-1}$. In the direction of the filamentary structures and the W33 complex, some areas with compact $^{13}$CO and C$^{18}$O emission are also seen in map at \[44, 46.5\] km s$^{-1}$, which is an intermediate velocity range between other two clouds.
In Figure \[fig9\]a, we show a color-composite image of our selected area with the $^{13}$CO maps at \[29.6, 43.3\] and \[47.2, 60.2\] km s$^{-1}$ in red and green, respectively. The filamentary feature “fs1" observed in the [*Herschel*]{} column density map and the positions of the ATLASGAL clumps are also highlighted in Figure \[fig9\]a. Figure \[fig9\]b displays the overlay of the surface density of Class I YSOs on the color-composite image produced using the $^{13}$CO maps at \[29.6, 43.3\] and \[47.2, 60.2\] km s$^{-1}$. A majority of the Class I YSOs are found toward the common zones of the two cloud components (see areas toward the filamentary structures and the W33 complex). In Figure \[fig9\]c, we present the overlay of the NVSS 1.4 GHz radio continuum contours on the color-composite image produced using the C$^{18}$O maps at \[29.6, 43.3\] and \[47.2, 60.2\] km s$^{-1}$. Both the color-composite images allow us to infer the overlapping zones of the two clouds at \[47.2, 60.2\] and \[29.6, 43.3\] km s$^{-1}$, where the early phase of SF, dust clumps as well as H[ii]{} regions are evident. Furthermore, in the color-composite images, a spatial fit of the intensity-depression region and the intensity-enhancement region is observed toward the W33 complex (see also Figure \[fig10\]). In the direction of the W33 complex, an intensity-enhancement is found in the molecular intensity maps at \[29.6, 43.3\] km s$^{-1}$ (see Figures \[fig10\]a and \[fig10\]b), while an intensity-depression region or cavity-like feature is observed in the molecular intensity maps at \[47.2, 60.2\] km s$^{-1}$ (see Figures \[fig10\]c and \[fig10\]d). The locations of W33 A (G012.90$-$0.24, G012.90$-$0.26), W33 Main (G012.81$-$0.19), and three O4-7I stars [see IDs \#7, 8, and 23 in @messineo15] are seen toward the areas of the spatial fit of the cavity and the intensity-enhancement region. However, the filament “fs1" and W33 B are found toward the common zones of the two clouds.
Overall, the study of the molecular line data shows the spatial and velocity connections of two cloud components in the direction of the selected physical system. We have discussed these results in more details in Section \[sec:disc\].
Discussion {#sec:disc}
==========
Physical environment around W33
-------------------------------
As highlighted earlier, previously published results were mainly focused to the H[ii]{} regions associated with the W33 complex (area $\sim$15 pc $\times$ 15 pc; see a broken box in Figures \[fig1\]a and \[fig1\]b). The large-scale environment around the W33 complex is not yet studied and explored. Such study can enable us to obtain new insights in the target site and to infer different physical mechanism(s) than earlier studies on small-scale environment. The present paper deals with an extended physical system ($\sim$50 pc $\times$ 37 pc), which is investigated by the distribution of 88 ATLASGAL clumps located at a distance of $\sim$2.6 kpc. It is also confirmed by the distribution of molecular gas at \[29.6, 60.2\] km s$^{-1}$, and at least three velocity components (around 35, 45, and 53 km s$^{-1}$) are identified in the direction of our selected physical system (see Section \[sec:coem\]). The physical system is found to host the embedded filaments (i.e., fs1 and fs2) and several H[ii]{} regions. The filaments are not associated with any radio continuum emission, and exhibit T$_\mathrm{d}$ of $\sim$17–19 K. In our selected system, a relatively warm dust emission (T$_\mathrm{d}$ $\sim$ 21–29 K) is also found toward the H[ii]{} regions excited by OB-type stars (see Table \[tab2\]). These results together show the existence of two distinct environments in the selected physical system, which differ substantially in their dust temperatures (see Section \[sec:hermap\]). The elongated filament (“fs1"; length $\sim$17 pc; T$_\mathrm{d}$ $\sim$19 K) is characterized as a thermally supercritical filament, and appears to be on the verge of collapse (see Section \[sec:hermap\]). Based on the distribution of Class I YSOs [mean age $\sim$0.44 Myr; @evans09], the early phase of SF activities is investigated toward the filaments and the H[ii]{} regions (age $\sim$0.35 – 1.0 Myr) in the system (see Section \[sec:sf\]).
Star Formation Scenarios
------------------------
Considering the presence of multiple velocity components and massive O-type stars, one can explore the applicability of the triggered SF scenarios concerning the expansion of H[ii]{} regions [e.g., @elmegreen77; @bertoldi89; @whitworth94; @lefloch94; @elmegreen98; @deharveng05; @dale07; @bisbas15; @walch15; @kim18; @haid19] and the colliding/converging flows from two velocity components [@ballesteros99; @heitsch08; @vazqez07] or the collision of two clouds [@elmegreen98]. Massive OB-stars are the important sources of mechanical and radiative energy, and can impact their surroundings through their intense energetic feedback (i.e., stellar wind, ionized emission, and radiation pressure). However, the stellar feedback cannot explain the existence of multiple velocity components in the selected physical system [see also @kohno18].
Concerning the “cloud cloud collision" (CCC) process, @habe92 studied numerical simulations of a head-on collision of two non-identical clouds. The simulations develop gravitationally unstable cores/clumps at the interface of the clouds due to the effect of their compression. In the CCC process, young stellar clusters and massive stars can be formed at the intersection of two molecular clouds or the shock-compressed interface layer [e.g., @habe92; @anathpindika10; @inoue13; @takahira14; @haworth15a; @haworth15b; @torii17; @bisbas17 and references therein]. In the CCC site, one may observationally find the spatial and velocity connections of two molecular clouds [e.g., @torii17; @dewangan17x; @dewangan17xx; @dewangan17b; @dewangan18b; @dewangan19a]. In the position-velocity map, one may also trace bridging features between two molecular clouds, which enable us to infer the association of the two clouds [e.g., @haworth15a; @haworth15b]. It may also hint the presence of a compressed layer of gas due to two colliding clouds/flows [e.g., @haworth15a; @haworth15b; @torii17]. Furthermore, one may also observe a complementary spatial distribution of two clouds in the CCC site [e.g., @torii17; @fukui18; @dewangan18x; @dewangan19a], which is related to the spatial fit of “Key/intensity-enhancement" and “Cavity/Keyhole/intensity-depression" features.
In the direction of the W33 complex (area $\sim$15 pc $\times$ 15 pc; see a broken box in Figures \[fig1\]a and \[fig1\]b), @kohno18 proposed the CCC model to explain the formation of massive stars. They found complementary distributions of two clouds (at 35 and 58 km s$^{-1}$) around the W33 complex. These authors also computed the timescale of the collision between these two clouds (i.e., 0.7–1.0 Myr), which was found to be older than the dynamical timescales of the outflows in W33 Main and W33 A [i.e., $\sim$10$^{4}$ yr; @kohno18]. They concluded that the sources in the W33 complex (i.e., W33 Main, W33 A, W33 B, W33 Main1, W33 A1, and W33 B1) were formed in the dense layer produced via the collision event. One can note that previously, no attempt was made to examine the wide-area around the W33 complex, which has been carefully carried out in the present paper.
In the present work, we have detected three velocity components around 35, 45, and 53 km s$^{-1}$ toward our selected physical system. @kohno18 reported the cloud at 58 km s$^{-1}$ as the third velocity component using the NANTEN2 $^{12}$CO (J = 1–0) emission. It appears that the cloud at 53 km s$^{-1}$ reported in this paper is a part of the cloud component at 58 km s$^{-1}$ identified by @kohno18 (see Figure 2b in their paper). In the direction of the W33 complex and the filamentary structures, in the velocity space, the two clouds around 35 and 53 km s$^{-1}$ are linked by lower intensity intermediate velocity emission, which may trace bridge-like features around 45 km s$^{-1}$ (see broken curves around [*l*]{} = 12$\degr$.8 and [*l*]{} =13$\degr$.1 in Figure \[fig7\]a). An analysis of the spatial distribution of molecular gas associated with two clouds around 35 and 53 km s$^{-1}$ shows the overlapping areas toward the filament “fs1" and W33 B (see arrows in Figure \[fig9\]a). These overlapping areas are ablaze with SF activities including massive stars (see Figures \[fig9\]b and \[fig9\]c), where we also find materials with high column densities.
Additionally, a complementary distribution of two clouds (around 35 and 53 km s$^{-1}$) is evident toward the W33 complex (see Figures \[fig10\]a & \[fig10\]c, and \[fig10\]b & \[fig10\]d), where the objects W33 Main, W33 A, and three massive O-type stars (i.e., \#7, \#8, and \#23) are located. As mentioned earlier, the complementary distribution of clouds was also reported by @kohno18. The key-like or intensity-enhancement feature is traced in the cloud at \[29.6, 43.3\] km s$^{-1}$ (or around 35 km s$^{-1}$), while the key-hole or cavity-like feature is evident in the cloud at \[47.2, 60.2\] km s$^{-1}$ (or around 53 km s$^{-1}$). However, we do not find any signature of complementary distribution of two clouds in the direction of the filament “fs1" and W33 B, but these objects are seen toward the common zones of the two clouds as discussed above. Overall, these results suggest the interaction of the two cloud components around 35 and 53 km s$^{-1}$ toward our selected physical system, which is associated with the early phase of SF activities and hosts H[ii]{} regions powered by massive OB stars. Note that our findings can be considered as an extension of the published work of @kohno18, but for a larger region or area (i.e., $\sim$50 pc $\times$ 37 pc). They reported the timescale of the collision to be 0.7–1.0 Myr.
In Figure \[fig10\]e, we display the overlay of the Scutum and Norma arms [from @reid16] on the longitude-velocity map of $^{13}$CO. The radial velocities of 88 ATLASGAL clumps are also shown in Figure \[fig10\]e. Figure \[fig10\]e also provides the information of the near side of the Norma and Scutum arms as well as the far side of the Scutum arm. Previously, based on the parallax measurements of maser sources, it was reported that the W33 complex is located toward the Scutum arm [see Figure 3 in @sato14]. Considering the distribution of the molecular gas and the ATLASGAL clumps in the velocity space, one may examine the interaction or collision of the gas clouds associated with the Scutum arm.
Earlier, using high resolution continuum and line data from the Submillimeter Array (SMA) and the Very Large Array (VLA) facilities, @galvan10 studied the SF activity toward W33 A, and interprated observed results in W33 A through a triggered SF scenario by filamentary convergent gas flows from two different velocity components [see also @inoue18]. In this scenario, one can expect the origin of molecular clouds via the convergence of streams of neutral gas. With time, the merging/converging/collision of the molecular clouds, filaments of molecular gas is expected, which produces the birth of cores and stars [@hunter86; @ballesteros99; @vazqez07; @heitsch08; @inoue18].
Previously, in the case of Orion Nebula Cluster [@fukui18] and RCW 38 [@fukui16], it was suggested that low-mass cluster members were formed before the onset of the collision event. However, in our selected target site, Class I YSOs (i.e., age $\sim$0.44 Myr) and H[ii]{} regions (age $\sim$0.35 – 0.85 Myr) appear relatively younger than the collision timescale (i.e., 0.7–1.0 Myr) as reported by @kohno18. Additionally, it was also pointed out that colliding gas flows may also lead to the birth of low-mass stars in regions of compressed gas and dust [@hunter86].
Taking into account the detection of Class I YSOs (i.e., age $\sim$0.44 Myr) and H[ii]{} regions (age $\sim$0.35 – 0.85 Myr) toward the common zones of the two cloud components around 35 and 53 km s$^{-1}$, the SF history seems to be explained by a triggered SF scenario by converging/colliding flows from two velocity components in our selected system. One can also keep in mind that the site G013.210$-$0.144 and the MIR bubbles (N10 and N11) appear away from the overlapping areas of the clouds. Hence, SF activities toward these objects are unlikely induced by the converging flows.
Summary and Conclusions {#sec:conc}
=======================
In this paper, to understand SF mechanisms, we have investigated a large-scale physical system hosting massive star-forming site W33 using a multi-wavelength approach. The present work is benefited with the analysis of the existing large-scale survey data at different wavelengths. The important results of this work are presented below.\
$\bullet$ A spatial spread of 88 ATLASGAL 870 $\mu$m dust clumps at d $\sim$2.6 kpc and the distribution of molecular gas at \[29.6, 60.2\] km s$^{-1}$ are utilized to select the extended physical system ($\sim$50 pc $\times$ 37 pc).\
$\bullet$ The information concerning the location and the radial velocity indicates our selected physical system in the direction of the Scutum and Norma arms.\
$\bullet$ Two distinct environments are investigated in the physical system, which are embedded filaments (with T$_{d}$ $\sim$17–19 K) and H[ii]{} regions powered by OB stars (with T$_{d}$ $\sim$ 21–29 K; age $\sim$0.3–1.0 Myr). The filaments are not associated with any ionized emission.\
$\bullet$ An elongated filament (“fs1"; length $\sim$17 pc) is found to be a thermally supercritical filament.\
$\bullet$ Based on the distribution of Class I YSOs [mean age $\sim$0.44 Myr; @evans09], the early stage of SF activities is observed toward the filaments and H[ii]{} regions in the cloud.\
$\bullet$ The FUGIN $^{13}$CO and C$^{18}$O line data trace three velocity components (around 35, 45, and 53 km s$^{-1}$) in the direction of the system.\
$\bullet$ The analysis of the molecular line data confirms the spatial and velocity connections of cloud components around 35 and 53 km s$^{-1}$.\
$\bullet$ The positions of W33 Main, W33 A and O4-7I stars are found toward a complementary distribution of cloud components around 35 and 53 km s$^{-1}$.\
$\bullet$ The spatial distribution of the clouds around 35 and 53 km s$^{-1}$ reveals overlapping zones toward the filament “fs1" and an object W33 B, where a majority of the ATLASGAL clumps and noticeable Class I objects are depicted.\
\
We conclude that the observed SF signposts seem to be explained well by a triggered SF scenario by converging/colliding flows from two velocity components around 35 and 53 km s$^{-1}$ in the physical system hosting two distinct environments (i.e., H[ii]{} regions and filamentary structures).
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank the anonymous reviewer for several useful comments and suggestions. The research work at Physical Research Laboratory is funded by the Department of Space, Government of India. This work is based \[in part\] on observations made with the [*Spitzer*]{} Space Telescope, which is operated by the Jet Propulsion Laboratory, California Institute of Technology under a contract with NASA. This publication makes use of data from FUGIN, FOREST Unbiased Galactic plane Imaging survey with the Nobeyama 45-m telescope, a legacy project in the Nobeyama 45-m radio telescope. TB is supported by the National Key Research and Development Program of China through grant 2017YFA0402702. TB also acknowledges support from the China Postdoctoral Science Foundation through grant 2018M631241. DKO acknowledges the support of the Department of Atomic Energy, Government of India, under project No. 12-R&D-TFR-5.02-0200.
{width="9.2cm"}
{width="13.4cm"}
{width="10.2cm"}
{width="13.2cm"}
{width="13.2cm"}
{width="17.0cm"}
{width="16.2cm"}
{width="17.2cm"}
{width="10.2cm"}
{width="19.2cm"} {width="7.2cm"}
----- ---------- ---------- ------------------ ----------- ----------------------- ---------------------------------- --------------- -- -- -- -- -- -- --
ID [*l*]{} [*b*]{} $R_\mathrm{HII}$ $S{_\nu}$ $\log{N_\mathrm{uv}}$ $t_\mathrm{dyn}$ (Myr) Spectral Type
(degree) (degree) (pc) (Jy) (s$^{-1}$) for $n_{0}$ = 10$^{4}$ cm$^{-3}$ (V)
c1 12.731 $-$0.130 1.31 688.81 47.56 0.79 B0.5-B0
c2 12.686 $-$0.188 1.08 361.55 47.28 0.66 B0.5-B0
c3 12.781 $-$0.167 1.39 1477.56 47.89 0.72 B0-O9.5
c4 12.811 $-$0.205 1.53 11370.76 48.78 0.51 O7.5-O7
c5 12.869 $-$0.234 1.34 653.54 47.54 0.83 B0.5-B0
c6 12.915 $-$0.284 0.81 361.55 47.28 0.39 B0.5-B0
c7 12.998 $-$0.201 1.12 163.63 46.94 0.86 B0.5-B0
c8 13.215 $-$0.147 0.87 888.69 47.67 0.36 B0-O9.5
c9 13.090 $-$0.072 1.16 92.10 46.69 1.05 B0.5-B0
c10 13.190 0.045 1.43 2698.41 48.16 0.65 O9.5-O9
c11 13.236 0.074 0.80 76.43 46.61 0.57 B0.5-B0
----- ---------- ---------- ------------------ ----------- ----------------------- ---------------------------------- --------------- -- -- -- -- -- -- --
[99]{} Anathpindika S. V., 2010, MNRAS, 405, 1431
Anderson L. D., Bania T. M., Balser D. S., Cunningham V., Wenger T. V., Johnstone B. M., Armentrout W. P., 2014, ApJS, 212, 1
Anderson L. D., Armentrout W. P., Johnstone B. M., Bania T. M., Balser D. S., Wenger T. V., Cunningham V., 2015, ApJS, 221, 26
Andr[é]{} P. et al., 2010, A&A, 518, L102
Andr[é]{} P., Di Francesco J., Ward-Thompson D., Inutsuka S. -I., Pudritz R. E., Pineda J. E., 2014, in Protostars and Planets VI, ed. H. Beuther et al. (Tucson, AZ; Univ. Arizona Press), 27
Ballesteros-Paredes J., Hartmann L., V[á]{}zquez-Semadeni E., 1999, ApJ, 527, 285
Benjamin R. A. et al., 2003, PASP, 115, 953
Bertoldi F., 1989, ApJ, 346, 735
Bisbas T. G., Wünsch R., Whitworth A. P., Hubber D. A., 2009, A&A, 497, 649
Bisbas T. G. et al., 2015, MNRAS, 453, 1324
Bisbas T. G., Tanaka K. E. I., Tan J. C., Wu B., Nakamura F., 2017, ApJ, 850, 23
Bressert E. et al., 2010, MNRAS, 409, 54
Carey S. J. et al., 2005, BAAS, 37, 1252
Casertano S., Hut P., 1985, ApJ, 298, 80
Churchwell E. et al., 2006, ApJ, 649, 759
Churchwell E. et al., 2007, ApJ, 670, 428
Condon J. J., Cotton W. D., Greisen E. W., Yin Q. F., Perley R. A., Taylor G. B., Broderick J. J., 1998, AJ, 115, 1693
Contreras Y. et al., 2013, A&A, 549, 45
Dale J. E., Clark P. C., Bonnell I. A., 2007, MNRAS, 377, 535
Deharveng L., Zavagno A., Caplan J., 2005, A&A, 433, 565
Dewangan L. K., 2017, ApJ, 837, 44
Dewangan L. K., Ojha D. K., 2017, ApJ, 849, 65
Dewangan L. K., Ojha D. K., Zinchenko I., Janardhan P., Luna, A., 2017a, ApJ, 834, 22
Dewangan L. K., Ojha D. K., Zinchenko I., 2017b, ApJ, 851, 140
Dewangan L. K., Ojha D. K., Zinchenko I., Baug T., 2018a, ApJ, 861, 19
Dewangan L. K., Baug T., Ojha D. K., Zinchenko I., Luna A., 2018b, ApJ, 864, 54
Dewangan L. K., Dhanya J. S., Ojha D. K., Zinchenko, I., 2018c, ApJ, 866, 20
Dewangan L. K., Sano H., Enokiya R., Tachihara K., Fukui Y., Ojha, D. K., 2019, ApJ, 878, 26
Downes D., Wilson T. L., Bieging J., Wink J., 1980, A&A, 40, 379
Dyson J. E., Williams D. A., 1980, Physics of the interstellar medium (New York, Halsted Press, p. 204)
Elmegreen B. G., Lada C. J., 1977, ApJ, 214, 725
Elmegreen B. G., 1998, in ASP Conf. Ser. 148, Origins, ed. C. E. Woodward, J. M. Shull, & H. A. Thronson, Jr. (San Francisco, CA: ASP), 150
Evans N. J. et al., 2009, ApJS, 181, 321
Fukui Y. et al., 2016, ApJ, 820, 26
Fukui Y. et al., 2018, ApJ, 859, 166
Getman K. V., Feigelson E. D., Garmire G., Broos P., Wang J., 2007, ApJ, 654, 316
Galván-Madrid R., Zhang Q., Keto E., Ho P. T. P., Zapata L. A., Rodr[í]{}guez L. F., Pineda J. E., V[á]{}zquez-Semadeni E., 2010, ApJ, 725, 17
Gama D. R. G., Lepine J. R. D., Mendoza E., Wu Y., Yuan J., 2016, ApJ, 830, 57
Gutermuth R. A., Megeath S. T., Myers P. C., Allen L. E., Pipher J. L., Fazio G. G., 2009, ApJS, 184, 18
Haschick A. D., Ho P. T. P., 1983, ApJ, 267, 638
Habe A., Ohta K., 1992, PASJ, 44, 203
Haid S., Walch S., Seifried D., Wünsch R., Dinnbier F., Naab T., 2019, MNRAS, 482, 4062
Hartmann L., Megeath S. T., Allen L., Luhman K., Calvet N., D’Alessio P., Franco-Hernandez R., Fazio G. G., 2005, ApJ, 629, 881
Haworth T. J. et al., 2015a, MNRAS, 450, 10
Haworth T. J., Shima K., Tasker E. J., Fukui Y., Torii K., Dale J. E., Takahira K., Habe A., 2015b, MNRAS, 454, 1634
Heitsch F., Hartmann L. W., Slyz A. D., Devriendt J. E.G., Burkert A., 2008, ApJ, 674, 316
Hunter J. H., Jr., Sandford M. T., II, Whitaker R. W., Klein R. I., 1986, ApJ, 305, 309
Inoue T., Fukui Y., 2013, ApJL, 774, 31
Inoue T., Hennebelle P., Fukui Y., Matsumoto T., Iwasaki K., Inutsuka S., 2018, PASJ, 70, S53
Immer K., Reid M. J., Menten K. M., Brunthaler A., Dame T. M., 2013, A&A, 553, A117
Immer K., Galván-Madrid R., Konig C., Liu H. B., Menten K. M., 2014, A&A, 572, A63
Inutsuka S., Miyama S. M., 1997, ApJ, 480, 681
Jiang X. J., Liu H. B., Zhang Q., Wang J., Zhang Zhi-Yu, Li J., Gao Y., Gu, Q., 2015, ApJ, 808, 114
Kim Jeong-Gyu, Kim Woong-Tae, Ostriker E. C.. 2018, ApJ, 859, 68
Kohno M. et al., 2018, PASJ, 70, S50
Lefloch B., Lazareff B., 1994, A&A, 289, 559
Lockman F. J., 1989, ApJS, 71, 469
Marsh K. A., Whitworth A. P., Lomax O., 2015, MNRAS, 454, 4282
Marsh K. A. et al., 2017, MNRAS, 471, 2730
Matsakis D. N., Evans N. J., II, Sato T., Zuckerman B., 1976, AJ, 81, 172
Maud L. T., Hoare M. G., Galván-Madrid R., Zhang Q., de Wit W. J., Keto E., Johnston K. G., Pineda J. E., 2017, MNRAS, 467, 120
Messineo M. et al., 2015, ApJ, 805, 110
Minamidani T. et al., 2016, Development of the new multi-beam100 GHz band SIS receiver FOREST for the Nobeyama 45-mTelescope. p. 99141Z, doi:10.1117/12.2232137
Molinari S. et al., 2010a, A&A, 518, L100
Molinari S. et al., 2010b, PASP, 122, 314
Nakajima T., Inoue H., Fujii Y., Miyazawa C., Iwashita H., Sakai T., Noguchi T., Mizuno A, 2019, PASJ, 71, 17
Ostriker J., 1964, ApJ, 140, 1056
Panagia N., 1973, AJ, 78, 929
Reid M. J., Dame T. M., Menten K. M., Brunthaler A., 2016, ApJ, 823, 77
Reid M. J. et al., 2009, ApJ, 700, 137
Sato M., Hirota T., Reid M. J., Honma, M., Kobayashi H., Iwadate K., Miyaji T., Shibata K. M., 2010, PASJ, 62, 287
Sato M. et al., 2014, ApJ, 793, 72
Schuller F. et al., 2009, A&A, 504, 415
Takahira K., Tasker E. J., Habe A., 2014, ApJ, 792, 63
Tan J. C., Beltrán M. T., Caselli P., Fontani F., Fuente A., Krumholz M. R., McKee C. F., Stolte A., 2014, in Protostars and Planets VI, ed. H. Beuther et al. (Tucson, AZ: Univ. Arizona Press), 149
Torii K. et al., 2017, ApJ, 835, 142
Urquhart J. S. et al., 2018, MNRAS, 437, 1059
Umemoto T. et al., 2017, PASJ, 69, 78
V[á]{}zquez-Semadeni E., G[ó]{}mez G. C., Jappsen A. K., Ballesteros-Paredes J., Gonz[á]{}lez R. F., Klessen R. S., 2007, ApJ, 657, 870
Walch S., Whitworth A. P., Bisbas T. G., Hubber D. A., Wünsch R., 2015, MNRAS, 452, 2794
Westerhout G., 1958, Bull. Astron. Inst. Netherlands, 14, 215
Whitworth A. P., Bhattal A. S., Chapman S. J., Disney M. J., Turner J. A., 1994, MNRAS, 268, 291
White R. L., Becker R. H., Hellfand D. J., 2005, AJ, 130, 586
Williams J. P., de Geus E. J., Blitz L., 1994, ApJ, 428, 693
Zinnecker H., Yorke H. W., 2007, ARA&A, 45, 481
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- Kate Ponto
bibliography:
- 'trace.bib'
date: |
\
The author was supported by a National Science Foundation postdoctoral fellowship.
title: Equivariant Fixed Point Theory
---
|
{
"pile_set_name": "ArXiv"
}
|
Nonlinear stratified media, in particular, nonlinear periodic media have drawn considerable attention over the past few decades due to their tremendous application potentials[@sdg-review]. In recent years there is a renewed interest because of fundamental issues involving $\mathcal{PT}$ symmetry and the associated spectral singularities [@pt-nonlinear1; @pt-nonlinear2; @pt-gap-1; @pt-gap-2; @pt-singl]. In the context of illumination from one side, particular attention was given to nonlinear periodic structures where the incident light frequency was chosen in the stop gap. It was shown that nonlinearity induced total transmission through the structure can be effected when the field distribution inside the layered medium corresponds to soliton-like spatial patterns[@chen-mills-prl; @chen-mills-prb]. A great deal of work has been carried out on such gap solitons both with continuous and discrete periodicities [@sdg-review; @sdg-imperial; @sipe-review; @molemed]. There are also reports on such gap solitons in $\mathcal{PT}$ symmetric systems [@pt-gap-1; @pt-gap-2]. Exact solutions for gap solitons for $s$-polarized light is now known [@chen-mills-prb]. Simultaneously approximate methods like nonlinear characterstic matrix method (NCMM) were also developed to capture the same physics [@ncmt; @sdg-review]. These methods were extensions of the linear characteristic matrix theory based on slowly varying envelope approximation (SVEA) [@marburg].
From a different angle there have been efforts to integrate the notions of recently invented coherent perfect absorption (CPA) [@cpa1; @cpa-longhi; @sdg-opex; @sdg-ol12] with nonlinearity [@ncpa2; @ncc; @ncpa-knr-sdg]. It was shown that nonlinearity can play a vital role in restoring CPA in off-resonant systems [@ncpa-knr-sdg; @ncc]. In the context of recent advancements in the theory and experiments on CPA, it is an interesting problem to integrate the notions of CPA with gap solitons. One can ask the question if it is possible to channel all the incident plane wave energy into the gap solitons. Clearly this would mean CPA like null-scattering leading to an energy efficient system. In this letter we investigate a symmetric periodic structure with alternating Kerr nonlinear and linear layers. Like in CPA the system is illuminated from both left and right ends [@sdg-opex]. We exploit the symmetry of the structure to calculate the incident intensities under which there is little or no outgoing waves from the structure. We also study the super scattering (SS) states which correspond to constructive interference as opposed to CPA. Depending on the parameters of the system, the null-scattering states correspond to one-, two-, and more-, soliton-like intensity distributions in the nonlinear layered medium. Calculations are performed using the NCMM and confirmed by exact numerical integration of the differential equations. We present details of each of the methods. The agreement of the approximate and the exact theory validates the NCMM applied to such nonlinear systems.
Consider the periodic layered medium shown in Fig. \[fig:fig1\]. Let the number of periods be $N$ resulting in $2N+1$ layers for the total structure. We assume the central layer to be Kerr nonlinear and extreme layers to be linear with lower refractive index. The dielectric function of the Kerr nonlinear slab is given by $$\bar{\epsilon}_a=\epsilon_a+\alpha|E|^2,
\label{eq1}$$ where $\alpha$ is the nonlinearity constant. Let the structure be illuminated at an angle, $\theta$ from both the sides by monochromatic $s$-polarized waves of wavelength $\lambda$. Moving to dimensionless variables as $$z\rightarrow k_0z \hspace{0.1cm}(k_0=\omega/c),\hspace{0.25cm} E\rightarrow \sqrt{\alpha} E, \label{eq2}$$ one can write down the Maxwell’s equations for tangential components of electric and magnetic fields in any $j^{th}$ layer as $$\begin{aligned}
\label{eq3}
\frac{dE_{jy}}{dz}&=&-iH_{jx},\\
\frac{dH_{jx}}{dz} &=&\begin{cases} \label{eq4}
-i[(\epsilon_j-p_{x}^2)+|E_{jy}|^2]E_{jy} \hspace{0.15cm} \text{for even $j$,}\\ \\
-i[(\epsilon_j-p_{x}^2)]E_{jy} \hspace{0.15cm} \text{for odd $j$},
\end{cases} \end{aligned}$$ where $p_x=\sqrt{\epsilon_{i}}\sin{\theta}$. Note that $E_{jy},~H_{jx}$ are in general complex and one has four coupled nonlinear equations for the real and imaginary parts of the fields. We have numerically integrated this set with proper boundary conditions to obtain the exact results (see below).
In what follows we present the outline of the approximate NCMM which has been used extensively in the past [@sdg-imperial; @sdg-review; @localization]. The propagation through each nonlinear layer is obtained in terms of the approximate (under SVEA) solution for the electric field as [@marburg] $$E_{jy}=A_{j+}e^{ip_{jz+}(z-z_j)}+A_{j-}e^{-ip_{jz-}(z-z_j)}, \label{eq5}$$ with $z_j \leq z \leq z_{j+1},~z_0=0$ and $p_{jz+}~(p_{jz-})$ denotes the normalized (to $k_0$) $z$ component of propagation constant for forward (backward) propagating waves with amplitudes $A_{j+}~(A_{j-})$. $p_{jz\pm}$ are given by the following $$p_{jz\pm}=\begin{cases}
\sqrt{\epsilon_j-p_x^2+(|A_{j\pm}|+2|A_{j\mp}|)} \hspace{0.2cm} \text{for even $j$}, \\ \\
\sqrt{\epsilon_j-p_x^2}\hspace{0.2cm} \text{for odd $j$}.
\end{cases}
\label{eq6}$$ We choose $\theta$ in range $$\epsilon_a-p_x^2>0, \hspace{0.5cm} \epsilon_b-p_x^2<0,\label{eq7}$$ in order to ensure that the waves are propagating (evanescent) in nonlinear (linear) layers. Thus the structure shown in Fig. \[fig:fig1\] represents a system of $N$ coupled nonlinear wave guides, sandwiched between two identical high index prisms representing a resonant tunneling geometry [@res-tunnel; @res-tunnel-sdg; @yeh]. Thus transmission through the structure can be realized when the resonant modes (the so called supermodes) are supported by the coupled guides [@yeh; @sdg-imperial]. For the illumination angle in the stop gap, total transmission of the nonlinear system for one sided illumination is known to lead to gap solitons [@sdg-imperial; @chen-mills-prb; @sipe-review; @chen-mills-prl; @sdg-review].
The column vectors with the tangential field components $E_{jy}$, $-H_{jx}$ as components, can be related at two different planes $z=z_j$ and $z=z_{j+1}$ as $$\begin{aligned}
\begin{pmatrix}
E_{jy} \\ -H_{jx} \end{pmatrix}_{z=z_{j}} &=&{{M}_j}
\begin{pmatrix}E_{jy} \\-H_{jx} \end{pmatrix}_{z=z_{j+1}}, \label{eq8}\end{aligned}$$ where $M_j$ denotes the nonlinear (linear) characteristic matrix [@sdg-review] when $j$ is even (odd). For CPA like excitation scheme (with both sided illumination) we exploit the symmetry (as in [@ncpa-knr-sdg]) of the structure allowing only the symmetric ($A_{0+}=A_{0-}=A_{0}$) and the antisymmetric ($A_{0+}=-A_{0-}=-A_{0}$) solutions. This results in $p_{0z+}=p_{0z-}=p_{z0}$. By treating $A_0$ as a parameter one can calculate the incident amplitude $A_{i+}(A_{f-})$ and the scattered amplitudes $A_{i-}(A_{f+})$ on the left (right) sides outside the layered medium. For example, starting from $z_0=0$ in the central layer one can calculate the amplitudes in the nonlinear layer labeled by $j=-2$ as $$\begin{aligned}
\begin{pmatrix}
A_{-2+} \\ A_{-2-} \end{pmatrix} = \begin{pmatrix} 1 & 1 \\ p_{-2z+} & -p_{-2z-} \end{pmatrix}^{-1}M_1(d_b)\times \nonumber \\ \times M_0(d_a/2)\begin{pmatrix} 1 & 1 \\ p_{0z} & -p_{0z} \end{pmatrix}\begin{pmatrix}A_0 \\ \pm A_0 \end{pmatrix}, \label{eq9}\end{aligned}$$ where $M_1(d_a/2)$ and $M_2(d_b)$ denote the characteristic matrices for nonlinear and linear layers of widths $d_a/2$, $d_b$, respectively. Since one doesn’t know $A_{-2\pm}$ *apriori*, $p_{-2\pm}$ are also not known. $p_{-2\pm}$ and hence $A_{-2\pm}$ are then evaluated by taking $|\cdots|^2$ on both the sides of Eq. (\[eq9\]) yielding the corresponding intensity relations and then using fixed point iteration for the coupled nonlinear equations [@sdg-imperial]. Continuing the same procedure for all the layers, for example, till the left end we have $$\begin{aligned}
\begin{pmatrix}
A_{i+} \\ A_{i-} \end{pmatrix} = \begin{pmatrix} 1 & 1 \\ p_{iz} & -p_{iz} \end{pmatrix}^{-1} M_N(d_b) M_{N-1}(d_a) \cdots \nonumber \\ \cdots M_1(d_b)M_0(d_a/2)\begin{pmatrix} 1 & 1 \\ p_{0z} & -p_{0z} \end{pmatrix}\begin{pmatrix}A_0 \\ \pm A_0 \end{pmatrix}, \label{eq10}\end{aligned}$$ where $p_{iz}$ denotes the $z$ component of the propagation constant for the incident (scattered) wave of amplitude $A_{i+}$ ($A_{i-}$) on the left side of the structure. One can also propagate towards right and evaluate the scattered ($A_{f+}$) and the incident ($A_{f-}$) wave amplitudes. As a consequence of inherent symmetry we find that for symmetric (antisymmetric) solution $A_{i+}=A_{f-}=A_{in}$ ($A_{i+}=-A_{f-}=-A_{in}$) and $A_{i-}=A_{f+}=A_{t}$ ($A_{i-}=-A_{f+}=-A_{t}$)[@ncpa-knr-sdg]. We can thus define the normalized scattering intensity as $S=|A_{i-}/A_{i+}|^2=|A_{f+}/A_{f-}|^2=|A_{t}/A_{in}|^2$ as a quantitative measure of CPA.
We now outline the symmetry considerations for the exact numerical integration. With the origin as defined in Fig. \[fig:fig1\], the supported solutions are either symmetric or antisymmetric with $$\begin{aligned}
E_y(-z)&=E_y(z),\hspace{0.5cm}H_x(-z)=-H_x(z),&\\ \label{eq11}
E_y(-z)&=-E_y(z),\hspace{0.5cm}H_x(-z)=H_x(z),& \label{eq12}\end{aligned}$$ respectively. Thus at $z=0$ in the central layer we have $H_x=0$ ($E_y=0$) for the symmetric (antisymmetric) solution. The other non vanishing component $E_y$ ($H_x$) at $z=0$ is treated as a parameter for symmetric (antisymmetric) solution. The initial conditions at $z=0$ for the tangential components of the electric and magnetic fields are thus known and Eqs. (\[eq3\])-(\[eq4\]) can be numerically integrated till the left/right end ensuring the continuity of the fields at every interface. Since the prisms are assumed to be linear the incident and scattered intensities can be computed in a straight forward manner from the knowledge of tangential fields.
We now present our numerical results. Most of the results presented are exact, obtained by numerical integration of Eqs. (\[eq3\])-(\[eq4\]). Direct numerical integration is used instead of the known exact solutions for Kerr nonlinear layers for $s$-polarized light[@chen-mills-prb] with future goals in mind. The technique used here is valid for any form of nonlinearity including the saturation type, which is attracting a lot of attention in the context of regularization of spectral singularities [@liu-sdg-gsa]. We also included some NCMM results to compare with exact ones to check the validity of the SVEA, which is often used in many other nonlinear systems. Throughout our calculations we have taken $\epsilon_a=4.0$, $\epsilon_b=2.25+i~0.5\times10^{-3}$, $N=19$, $\epsilon_i=6.145$, $d_a=0.2\lambda$, $d_b=0.365\lambda$ and $\theta$ is varied in between $41.0^{\circ}$ and $45.0^{\circ}$, where the condition given by Eq. (\[eq7\]) is satisfied.
We first study the linear properties. Let the amplitude reflection and transmission coefficients with unidirectional illumination be denoted by $r$ and $t$, respectively. The results for the corresponding intensity reflection $R=|r|^2$ and intensity transmission $T=|t|^2$ as functions of the angle of incidence are shown in Fig. \[fig:fig2\](a). The sharp resonances in Fig. \[fig:fig2\](a) in a small range of angles are due to resonant tunneling of electromagnetic radiation. The number of such resonances coincides with the number of the guides as expected. At every resonance we have the characteristic phase jump of $\pi$ (not shown here) between reflected and transmitted light ($\Delta\phi=|\phi_r-\phi_t|$). Thus if one satisfies $|r|=|t|$ for unidirectional illumination near any of these resonances then one can have CPA for two sided illumination geometry [@sdg-opex]. We studied the linear CPA for bidirectional illumination of the structure for both symmetric and antisymmetric solutions. The CPA-like dips in scattering are shown in Fig. \[fig:fig2\](b) for given losses and system parameters. We label these dips by integers $0,~2,~4$ ($1,~3,~5$) for symmetric (antisymmetric) solutions from the extreme right end. A prominent CPA-like dip can be seen in Fig. \[fig:fig2\](b) for the resonance labeled by $2$ as $|r|\approx|t|$ at that particular angle (see Fig. \[fig:fig2\](a)) and the phase condition is met by characteristic phase jump at the resonance.
Having discussed the linear properties, we investigate the nonlinearity induced changes when the operating point (angle of incidence) is chosen in the stop gap (marked by a cross in Fig. \[fig:fig2\], i.e., at $\theta=44.81^{\circ}$). Note that the stop gap boundary occurs at $\theta=44.76^\circ$. An increase in incident intensities leads to an increase (for $\alpha>0$) in the optical widths of the nonlinear layers and hence to bent resonances [@ncmt]. This is shown in Fig. \[fig:fig3\], where the scattered intensity is plotted against the incident intensity for both the symmetric (Fig. \[fig:fig3\](a)) and the antisymmetric (Fig. \[fig:fig3\](b)) solutions, respectively. The CPA-like dips are labeled following their linear counterparts in Fig. \[fig:fig2\](b). The symmetric (antisymmetric) dips $0,~2,~4$ ($1,~3,~5$) occur at $U_i=1.877\times10^{-4},~6.12\times10^{-5},~1.47\times10^{-4}$ ($U_i=5.10\times10^{-5},~9.56\times10^{-5},~2.004\times10^{-4}$). Since, the nonlinearity drives the on-resonant system away from the CPA resonance [@ncpa-knr-sdg], it is difficult to have CPA at both low and high powers for the same system under identical conditions. Till now our attention was on destructive interference and the resulting CPA under optimal conditions. In contrast, one can have constructive interference and super scattering, leading to maximal scattered intensities. Such nonlinear SS states are shown in the insets of Fig. \[fig:fig3\](a) and Fig. \[fig:fig3\](b). The lowest order symmetric (antisymmetric) SS states occur at $U_i=4.569\times10^{-3},~6.976\times10^{-3},~1.755\times10^{-2}$ ($U_i=4.872\times10^{-3},~1.136\times10^{-2},~2.555\times10^{-2}$).
Finally we study the normalized intensity $|E|^2$ distribution in the total structure corresponding to the nonlinear CPA-like and SS states by both exact and approximate methods. The typical field profiles corresponding to nonlinear CPA are shown in Fig. \[fig:fig4\], with left (right) column corresponding to the symmetric (antisymmetric) solutions. The dashed (solid) curves represent the field profiles as calculated by NCMM (exact integration). One can see the one-, two-, etc, soliton-like intensity profile (see Fig. \[fig:fig4\]) for CPA-like cases. For example, the envelope of the distribution in Fig. \[fig:fig4\](a) can be fitted with $A/\cosh^2(\beta z)$ to the exact (approximate) result with $\beta=0.0895$ and $A=0.0252~(0.0301)$ (see Fig. \[fig:fig4\](a) dotted curves). As compared to the CPA states, the SS ones are characterized by an intensity profile with fields localized away from the center (see Fig. \[fig:fig5\](a), (b) for SS states labeled by $0^{\prime}$ and $1^{\prime}$ in Fig. \[fig:fig3\]). As one can see, both the methods differ very little validating the SVEA in this context with NCMM estimates for $|E|^2$ to be slightly higher than those of the exact results.
. \[fig:fig5\]
In conclusion, we have studied a Kerr nonlinear periodic layered medium irradiated by $s$-polarized plane waves from both the sides and presented the exact numerical results for CPA. The angle of illumination was chosen inside the stop gap of the linear counterpart. We have shown that CPA in such structures correspond to one- two- etc. soliton-like distributions of the field inside the structure. These results are also obtained by making use of the NCMM offering a true test of the SVEA. The good correspondence of the exact and the approximate results is indicative of the validity of the SVEA and similar approximations of nonlinear optics so far as there are no evanescent waves in the Kerr-nonlinear media. We have also studied the field distributions corresponding to the cases of nonlinear super-scattering (maximal scattering). The excitation of the nonlinear gap solitons with CPA mediated near-null scattering may have interesting application in switching and integrated optical devices.
[99]{} S. Dutta Gupta, *Nonlinear optics of stratified media*, (Progress in Optics, E.Wolf, ed.(Elsevier Science, 1998), Vol. **38**, p.1.) Z. H. Musslimani, K. G. Makris, R. El-Ganainy, D. N. Christodoulides, **100**, 030402 (2008).
Z. Lin, H. Ramezani, T. Eichelkraut, T. Kottos, H. Cao, and D. N. Christodoulides, **106**, 213901 (2011).
H. Wang, W. He, L. Zheng , X. Zhu , H Li, and Y. He, J. Phys. B: At. Mol. Opt. Phys. **45** 245401 (2012).
F. K. Abdullaev, V. A. Brazhnyi, and M. Salerno, **88**, 043829 (2013). A. Mostafazadeh, **110**, 260402 (2013).
W. Chen and D. L. Mills, **58**, 160 (1987).
W. Chen and D. L. Mills, **36**, 6269 (1987).
S. Dutta Gupta, **6**, 1927 (1989).
C. M. de Sterke and J. E. Sipe, *Gap solitons*, (Progress in Optics, E. Wolf, ed. (Elsevier, 1994) Vol. **34**.)
H. Sakaguchi and B. A. Malomed **81**, 013624 (2010). S. Dutta Gupta and G. S. Agarwal, **4**, 691 (1987). J. H. Marburger and F. S. Felber, **17**, 335 (1978). W. Wan, Y. D. Chong, L. Ge, H. Noh, A. D. Stone and H. Cao, Science **331**, 889-892 (2011).
S. Longhi, **82**, 031801(R) (2010).
S. Dutta-Gupta, O. J. F. Martin, S. Dutta Gupta, G. S. Agarwal, **20**, 1330 (2012).
S. Dutta-Gupta, R. Deshmukh, A. V. Gopal, O. J. F. Martin, and S. Dutta Gupta, **37**, 4452 (2012).
S. Longhi, **107**, 033901 (2011).
K. N. Reddy, A. V. Gopal, S. Dutta Gupta, **38**, 2517 (2013).
K. N. Reddy and S. Dutta Gupta, **38**, 5252 (2013).
S. Dutta Gupta, D. S. Ray, Phys. Rev. B **41**, 8048 (1990).
I. R. Hooper, T. W. Preist, and J. R. Sambles, **97**, 053902 (2006).
D. Golla, S. Deb, and S. Dutta Gupta, Eur. Phys. J. Appl. Phys. **53**, 10702 (2011).
P. Yeh, *Optical Waves in Layered Media* (Wiley, New York, 1988).
X. Liu, S. Dutta Gupta, and G. S. Agarwal, **89**, 013824 (2014).
[99]{} S. Dutta Gupta, *Nonlinear optics of stratified media*, (Progress in Optics, E.Wolf, ed.(Elsevier Science, 1998), Vol. **38**, p.1.) Z. H. Musslimani, K. G. Makris, R. El-Ganainy, D. N. Christodoulides, “Optical solitons in PT periodic potentials,” **100**, 030402 (2008).
Z. Lin, H. Ramezani, T. Eichelkraut, T. Kottos, H. Cao, and D. N. Christodoulides, “Unidirectional invisibility induced by PT-symmetric periodic structures,” **106**, 213901 (2011).
H. Wang, W. He, L. Zheng , X. Zhu , H Li, and Y. He, “Defect gap solitons in real linear periodic optical lattices with parity-time symmetric nonlinear potentials,” J. Phys. B: At. Mol. Opt. Phys. **45** 245401 (2012).
F. K. Abdullaev, V. A. Brazhnyi, and M. Salerno, “Scattering of gap solitons by PT-symmetric defects,” **88**, 043829 (2013). A. Mostafazadeh, “Nonlinear spectral singularities for confined nonlinearities,” **110**, 260402 (2013)
W. Chen and D. L. Mills,“Gap solitons and optical response of the superlattice,” **58**, 160 (1987).
W. Chen and D. L. Mills, “Optical response of nonlinear multilayer structures: Bilayers and superlattices ,” **36**, 6269 (1987).
S. Dutta Gupta, “Optical multistability and soliton-like intensity distribution in a nonlinear superlattice for oblique incidence,” **6**, 1927 (1989).
C. M. de Sterke and J. E. Sipe, *Gap solitons*, (Progress in Optics, E. Wolf, ed. (Elsevier, 1994) Vol. **34**.)
H. Sakaguchi and B. A. Malomed, “Solitons in combined linear and nonlinear lattice potentials,” **81**, 013624 (2010). S. Dutta Gupta and G. S. Agarwal, “Dispersive bistability in coupled nonlinear Fabry-Perot resonators,” **4**, 691 (1987). J. H. Marburger and F. S. Felber, “Theory of a lossless nonlinear Fabry-Perot interferometer,” **17**, 335 (1978). W. Wan, Y. D. Chong, L. Ge, H. Noh, A. D. Stone and H. Cao, “Time-reversed lasing and interferometric control of absorption,” Science **331**, 889-892 (2011).
S. Longhi, “PT-symmetric laser absorber,” **82**, 031801(R) (2010).
S. Dutta-Gupta, O. J. F. Martin, S. Dutta Gupta, G. S. Agarwal, “Controllable coherent perfect absorption in a composite film,” **20**, 1330 (2012).
S. Dutta-Gupta, R. Deshmukh, A. V. Gopal, O. J. F. Martin, and S. Dutta Gupta, “Coherent perfect absorption mediated anomalous reflection and refraction,” **37**, 4452 (2012).
S. Longhi, “Time-Reversed Optical Parametric Oscillation,” **107**, 033901 (2011).
K. N. Reddy, A. V. Gopal, S. Dutta Gupta, “Nonlinearity induced critical coupling,” **38**, 2517 (2013).
K. N. Reddy and S. Dutta Gupta, “Light-controlled perfect absorption of light,” **38**, 5252 (2013).
S. Dutta Gupta, D. S. Ray, “Localization problem in optics: Nonlinear quasiperoidic media,” Phys. Rev. B **41**, 8048 (1990).
I. R. Hooper, T. W. Preist, and J. R. Sambles, “Making tunnel barriers (including metals) transparent,” **97**, 053902 (2006).
D. Golla, S. Deb, and S. Dutta Gupta, “The role of absorption and dispersion in resonant tunneling through a negative index medium,” Eur. Phys. J. Appl. Phys. **53**, 10702 (2011).
P. Yeh, *Optical Waves in Layered Media* (Wiley, New York, 1988).
X. Liu, S. Dutta Gupta, and G. S. Agarwal, “Regularization of the spectral singularity in PT-symmetric systems by all-order nonlinearities: Nonreciprocity and optical isolation,” **89**, 013824 (2014).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This paper deals with model predictive control problems for large scale dynamical systems with cyclic symmetry. Based on the properties of block circulant matrices, we introduce a complex-valued coordinate transformation that block diagonalizes and truncates the original finite-horizon optimal control problem. Using this coordinate transformation, we develop a modified alternating direction method of multipliers (ADMM) algorithm for general constrained quadratic programs with block circulant blocks. We test our modified algorithm in two different simulated examples and show that our coordinate transformation significantly increases the computation speed.'
author:
- 'Idris Kempf${^*}$, Paul J. Goulart and Stephen Duncan[^1]'
bibliography:
- 'IEEEabbrv.bib'
- 'bib\_circ\_admm.bib'
title: '**ADMM for Block Circulant Model Predictive Control**'
---
Model Predictive Control (MPC), Alternating Direction of Multipliers Method (ADMM), Block Circulant Systems, Quadratic Program
Introduction
============
The advantages of model predictive control (MPC) for constraint handling and feedforward disturbance modelling are widely recognised. However, its applicability is limited by the requirement to solve optimization problems in real-time to compute the control law. This constraint has inhibited the application of MPC to large-scale and high speed applications. While some approaches for accelerating the computing speed have focused on implementing optimization routines on specialised high-performance hardware [@FPGAMPC], other approaches have exploited the particular symmetric structure encountered in some classes of large-scale problems [@SYMMETRICMPC]. In this paper, we address systems with cyclic symmetry resulting in a block circulant structure. These systems can be interpreted as the symmetric interconnection of many subsystems, where each subsystem interacts in an identical way with its neighbors [@ANDREADISTRIBUTED]. Circulant systems can be found in a variety of applications, including vehicle formation control [@PHDTRAFFIC; @CIRCBALLOON], cross-directional control [@PAPERMACHINES], particle accelerator control [@PAROBUST] and in the approximation of partial differential equations [@PDE]. The mathematical properties of these systems have already been exploited in controller design [@HCIRC; @DISTRCIRC], stability analysis [@STABILITYCIRC] and subspace identification [@CIRCSYSID]. In this paper, the properties that a constrained quadratic program (CQP) inherits from a block circulant MPC problem are investigated. The main results of the paper show how exploiting the properties of the resulting CQP can reduce the computational cost when it is solved using the alternating direction of multipliers method [@ADMMBOYD].
This paper is structured as follows. In Section \[sec:problemstatement\], the linear model predictive control (MPC) problem and the alternating direction of multipliers method (ADMM) – an algorithm which is particularly suitable for solving the latter optimization problem – are introduced. Since this paper is concerned with the analysis of systems with block circulant symmetry, we introduce the notion of block circulant matrices in Section \[sec:circulantdecomposition\]. Furthermore, the block circulant MPC problem is formally defined and necessary conditions for its decomposition are stated. In Section \[sec:blockcirculantadmm\], we define a CQP with block circulant blocks and show how MPC problems with block circulant data can be written in this form. The block circulant decomposition is then applied to the CQP and a modified ADMM algorithm is then formulated for this problem. In Section \[sec:simulations\], we compare the performance of the original and modified ADMM algorithms. For the sake of comparison, both algorithms have been implemented in Matlab and tested on two illustrative examples.
*Notation and Definitions* The set of real and complex numbers is denoted by $\R$ and $\C$, respectively. Let $\otimes$ denote the Kronecker product and $\oplus$ denote the direct sum (i.e. the block diagonal concatenation) of two matrices. Let $\I_n$ represent the identity matrix in $\R^{n\times n}$. For a scalar, vector or matrix $a$, let $\bar{a}$ denote its complex conjugate; Let $Re(a)$ and $Im(a)$ denote its real and imaginary part, respectively; Let $a^H$ denote its Hermitian transpose. Let $\diag\lbrace a_1,\dots,a_n\rbrace$ denote a diagonal matrix with diagonal elements $a_1,\dots,a_n$.
Problem Statement {#sec:problemstatement}
=================
Model Predictive Control {#sec:mpc}
------------------------
Given a discrete-time linear dynamical system and an initial condition $x(t)$ at time $t$, a standard MPC scheme computes a control law by predicting the future evolution of the system and minimizing a quadratic objective function over some planning horizon $T$. This can achieved via repeated solution of the following quadratic program (QP):
$$\begin{aligned}
\min &\sum_{k=0}^{T-1} x_k^\Tr Q x_k+u_k^\Tr R u_k + x_N^\Tr P x_N\label{eq:mpcA}\\
s.t.\,\,\, &x_{k+1} = Ax_k+Bu_k, \quad x_0=x(t)\label{eq:mpcB}\\
&y_k = Cx_k+Du_k\label{eq:mpcC}\\
&\ubar{y} \leq y_k \leq \bar{y}\label{eq:mpcD}\end{aligned}$$
\[eq:mpc\]
for $k=0,\dots,T-1$ and outputting the first input vector $u_0$ of the optimal control law. The constraints on the states $x_k\in\R^{n_x}$ and the inputs $u_k\in\R^{n_u}$ are lumped into the variable $y_k\in\R^{n_y}$. The prediction horizon is $T$ and the dynamics of the system are described by . The stability of the state is guaranteed if $P=P^\Tr\succ 0$ is obtained from the discrete-time algebraic Riccati equation (DARE), $$\begin{aligned}
\label{eq:dare}
A^\Tr P A - A^\Tr P B \left(B^\Tr P B +R\right)^{-1}B^\Tr P A + Q =P.\end{aligned}$$ The problem has a unique solution if $R\succ 0$, $Q\succeq 0$ and the pairs $(A,B)$ and $(A,Q^{\frac{1}{2}})$ are controllable and observable, respectively [@MPCBOOK Chapter 12]. Throughout the paper, we will assume that $P$ is obtained from and that admits a unique solution.
By eliminating the state variables $(x_1, \dots, x_N)$ and defining $z \eqdef (u_0,\dots,u_{T-1})^\Tr$ and $v\eqdef(y_0,\dots,y_{T-1})^\Tr$, can be reformulated as
\[eq:mpcqp\] $$\begin{aligned}
\min \,\,\,&\frac{1}{2}z^\Tr J z +q^\Tr z\label{eq:mpcqpA}\\
s.t.\,\,\, &Kz-v=0\label{eq:mpcqpB}\\
&\ubar{v} \leq v \leq \bar{v},\label{eq:mpcqpC}\end{aligned}$$
where we do not make the dependency of $(\bar v, \ubar{v}, q)$ on $x_0$ explicit for simplicity of notation. The matrices $(J,K)$ and vectors $(\ubar{v},\bar{v},q)$ in are defined as
$$\begin{aligned}
J &\eqdef G^\Tr \left( (\I_T\otimes Q)\oplus P \right)G+(\I_T\otimes R)\\
K &\eqdef \left[ \I_T\otimes C \,\vert\, 0 \right] G+(\I_T\otimes D)\\
\ubar{v} &\eqdef (\B{1}_T\otimes\ubar{y}) - \left[\I_T\otimes C \,\vert\, 0 \right]Hx_0\\
\bar{v} &\eqdef (\B{1}_T\otimes\bar{y}) - \left[\I_T\otimes C \,\vert\, 0 \right]Hx_0\\
q &\eqdef G^\Tr H x_0\label{eq:mpcqpmatricesE}\end{aligned}$$
\[eq:mpcqpmatrices\]
where $\B{1}_T$ is a vector of ones of length $T$ and $G$ and $H$ arise from elimination of the equality constraints in , i.e. from setting $X=(x_0,\dots,x_T)^\Tr$ and writing as $X=Gz+Hx_0$, $$\begin{aligned}
\label{eq:dynamics}
G =
\begin{bmatrix}
0 & \dots \\
B & & \\
AB & B & & \\
\vdots & & \ddots\\
A^{T-1}B & A^{T-2}B & \dots & B
\end{bmatrix},\quad
H=
\begin{bmatrix}
\I_{n_x} \\ A \\ \vdots \\ A^T
\end{bmatrix}.\end{aligned}$$ Note that $J\succ 0$ because $R\succ 0$ by assumption.
ADMM Algorithm {#sec:admm}
--------------
We consider application of the alternating direction of multipliers method (ADMM) to the solution of , and will follow the specific ADMM formulation presented in [@ADMMBOYD] throughout. The method is summarized in Algorithm \[alg:admm\]. The augmented Lagrangian for can be written as $$\begin{aligned}
\label{eq:lang}
\begin{split}
L(z,v,\gamma) = &\frac{1}{2}z^\Tr J z + q^\Tr z+\frac{\rho}{2}\lVert Kz-v\rVert_2^2\\&+\gamma^\Tr(Kz-v) + \mathcal{I}_{[\ubar{v},\bar{v}]}(v),
\end{split}\end{aligned}$$ where $\mathcal{I}_{[\ubar{v},\bar{v}]}$ is the indicator function for the set $\set{v}{\ubar{v}\le v \le \bar{v}}$ and the penalty parameter $\rho>0$ and the dual variables $\gamma$ are associated with the constraint . ADMM solves by repeatedly minimizing w.r.t. $z$ and $v$ and updating the dual variables $\gamma$ using an approximate gradient ascent method. Even though the assumptions in section \[sec:mpc\] guarantee the convergence of Algorithm \[alg:admm\], it is common practice to limit it to a maximum number of iterations $i_{max}$.
State $x(t)$ Input $u(t)$ Set $x_0=x(t)$ and $v^0,\gamma^0=0$; compute $\ubar{v},\bar{v}$ and $q$ Update $z^i$ using (SP1) Update $v^i$ using (SP2) Update $\gamma^i$ using (SP3) $u(t)=(z_1,\dots,z_{n_u})^\Tr$
After initialization[^2], Algorithm \[alg:admm\] first minimizes w.r.t. to $z$, which, after completing the square, is equivalent to $$\begin{aligned}
\label{eq:minz}
z^i =\argmin_z \frac{1}{2}z^\Tr J z + q^\Tr z +\frac{\rho}{2}\lVert Kz-v^{i-1} +\rho^{-1}\gamma^{i-1}\rVert_2^2,\end{aligned}$$ with iteration index $i=1,\dots,i_{max}$. Since is an unconstrained QP, its derivative can be set to zero and the resulting linear system can then be solved from $$\begin{aligned}
\label{eq:sp1}
\left( J+\rho K^\Tr K \right) z^i = K^\Tr (\rho v^{i-1}-\gamma^{i-1}) -q.\tag{SP1}\end{aligned}$$ The linear system always admits a solution because $J+\rho K^\Tr K\succ 0$ under the assumptions from section \[sec:mpc\].
With $z^i$ obtained, Algorithm \[alg:admm\] then minimizes w.r.t. $v$ by solving $$\begin{aligned}
\label{eq:sp2}
v^i = &\argmin_{\ubar{v} \leq v \leq \bar{v}} \twonorm{Kz^i -v +\rho^{-1}\gamma^{i-1}}^2.\end{aligned}$$ The solution to can be written as $$\begin{aligned}
\label{eq:sp2_sat}
v^i = \sat_{[\ubar{v},\bar{v}]}\left\lbrace Kz^i+\rho^{-1}\gamma^{i-1}\right\rbrace,\tag{SP2}\end{aligned}$$ where the saturation function limits its argument to $\ubar{v}$ and $\bar{v}$.
Finally, algorithm \[alg:admm\] updates the dual variable $\gamma$ according to $$\begin{aligned}
\label{eq:sp3}
\gamma^i = \gamma^{i-1} +\rho(Kz^i-v^i). \nonumber\tag{SP3}\end{aligned}$$ Subproblems - are repeated until some convergence criterion or the maximum number of iterations is reached. Proofs and other variants of the ADMM can be found in [@ADMMBOYD; @OSQPMAIN].
Circulant Decomposition {#sec:circulantdecomposition}
=======================
Preliminaries {#sec:circ_matrices}
-------------
\[def:circ\_matrices\] Let $\inR{\Cn}{n}{n}$ denote the set of real invertible *circulant matrices*, i.e. all invertible $n\times n$ matrices in the form $$\begin{aligned}
C = \CircC,\end{aligned}$$ where each row is a cyclic shift of the previous row and each $c_i \in \R$.
Circulant matrices have a number of very useful basic properties [@CIRCBOOK Chapter 3], including
$$\begin{aligned}
C\in \Cn, \alpha \in \R &\longleftrightarrow \alpha C \in \Cn,\label{eq:prop1}\\
C\in \Cn &\longleftrightarrow C^T \in \Cn,\label{eq:prop2}\\
C\in \Cn &\longleftrightarrow C^{-1} \in \Cn,\label{eq:prop3}\\
A,C\in \Cn &\longleftrightarrow AC \in \Cn,\label{eq:prop4}\\
A,C\in \Cn &\longleftrightarrow A+C \in \Cn\label{eq:prop5}.\end{aligned}$$
\[eq:propcirc\]
\[def:fourier\_matrix\] Let $\inC{F_n}{n}{n}$ denote the *Fourier matrix*, defined as $$\begin{aligned}
F_n &= \frac{1}{\sqrt{n}}\begin{bmatrix} w_0 & w_1 & \dots & w_{n-1}\end{bmatrix}\label{eq:fourier_matrix}\end{aligned}$$ where the vectors $w_j = \begin{pmatrix}1&\rho_j & \rho_j^2 & \dots & \rho_j^{n-1} \end{pmatrix}^\Tr$ are mutually orthogonal and $\rho_j = e^{i\frac{2\pi}{n}j}$ are complex roots of unity.
Because the vectors $w_i$ composing the matrix $F_n$ in are orthogonal and $\Vert w_i \Vert_2 = \sqrt{n}$, the matrix $F_n$ is orthogonal and unitary, i.e. $F_nF_n^H=F_n^HF_n=\I_n$. The matrix is called the Fourier matrix since the Fourier coefficients of the discrete Fourier transformation of a vector $x \in \R^n$ can be obtained from the product $F_n x_n$ (or more efficiently using a fast Fourier transformation [@ROSE]).
Perhaps the most remarkable property of circulant matrices is that every circulant matrix of order $n$ is diagonalized by the same Fourier matrix $F_n$.
\[thm:diag\_cn\] For $\inR{C}{n}{n}$, it holds that $F_n^HCF_n$ is diagonal iff $C\in\Cn$. The diagonal elements $\lambda_0,\dots,\lambda_{n-1}$ of $F_n^HCF_n$ are $$\begin{aligned}
\lambda_j = c_o + c_1\rho_j+\dots+c_{n-1}\rho_j^{n-1},\,\,j=0,\dots,n-1.\end{aligned}$$
From the structure of $F_n$, the following corollary can be established on the structure of the eigenvalues $\lambda_0,\dots,\lambda_{n-1}$:
\[thm:pattern\_cn\] For $C\in\Cn$ the diagonal elements $\lambda_0,\dots,\lambda_{n-1}$ of $F_n^HCF_n$ have the following pattern of complex conjugates. If $n$ is odd, then $\lambda_0$ is real while $(\lambda_1,\dots,\lambda_{\frac{n-1}{2}}) = (\bar{\lambda}_{n-1},\dots,\bar{\lambda}_{\frac{n+1}{2}})$. If $n$ is even, then $\lambda_0$ and $\lambda_\frac{n}{2}$ are real while $(\lambda_1,\dots,\lambda_{\frac{n}{2}-1})=(\bar{\lambda}_{n-1},\dots,\bar{\lambda}_{\frac{n}{2}+1})$. If $C=C^\Tr$, then $\lambda_0,\dots,\lambda_{n-1}$ are real.
Note that the same pattern of complex conjugates applies to the elements of $F_n x$ for any $x \in \R^n$.
Let $\BCnpmC \subseteq \R^{np\times nm}$ denote the set of real block circulant matrices in the form $$\begin{aligned}
B = \CircB\end{aligned}$$ where the scalars $c_i$ from Definition \[def:circ\_matrices\] have been replaced by blocks $\inR{b_i}{p}{m}$.
Properties analogous to can be established for block circulant matrices [@CIRCBOOK Chapter 5.6], i.e.
$$\begin{aligned}
B\in \BCnpmC, \alpha \in \R &\longleftrightarrow \alpha B \in \BCnpmC,\label{eq:prop1_bc}\\
B\in \BCnpmC &\longleftrightarrow B^T \in \mathcal{BC}(n,m,p),\label{eq:prop2_bc}\\
B\in \mathcal{BC}(n,m,m) &\longleftrightarrow B^{-1} \in \mathcal{BC}(n,m,m),\label{eq:prop3_bc}\\
A,B\in \BCnpmC &\longleftrightarrow A+B \in \BCnpmC,\label{eq:prop4_bc}\\
A\in \BCnpmC,&\,\,\,B \in \mathcal{BC}(n,m,r)\nonumber \\ &\longleftrightarrow AB \in \mathcal{BC}(n,p,r).\label{eq:prop5_bc}\end{aligned}$$
\[eq:propbcirc\]
For it must hold that $\det B>0$. In [@CIRCSYSID], Theorem \[thm:diag\_cn\] was extended to the block circulant case and is described by the following corollary:
\[thm:diag\_bc\] For $\inR{B}{np}{nm}$, it holds that $(F_n\otimes\I_p)^HB(F_n\otimes\I_m)$ is block diagonal iff $B\in\BCnpmC$. The blocks $\inC{\nu_j}{p}{m}$ of the block diagonalized matrix are obtained as $$\begin{aligned}
\nu_j = b_o + b_1\rho_j+\dots+b_{n-1}\rho_j^{n-1},\,\,j=0,\dots,n-1.\label{eq:nu_j}\end{aligned}$$ The pattern of complex conjugates described in Corollary \[thm:pattern\_cn\] also holds for the blocks $\nu_j$.
Using the shuffle permutation matrix $\Pi_n^m$ from [@ROSE], we can rewrite $F_n\otimes\I_m$ as $$\begin{aligned}
\label{eq:PsiPerm}
F_n\otimes\I_m = (\Pi_n^m)^\Tr (I_m \otimes F_n) \Pi_n^m\end{aligned}$$ and see, that given a vector $x \in \R^{nm}$, $(F_n\otimes\I_m) x$ can be computed by permuting the vector, applying a sequence of $m$ fast Fourier transformations and applying the inverse permutation. This entails that $(F_n\otimes\I_m) x$ is of complexity $\co{mn\log n}$ compared to $\co{m^2n^2}$ for general matrix-vector multiplication. Note that the pattern of complex conjugates also holds for the $n$ blocks of the vector $(F_n\otimes\I_m)x$.
Block Circulant Decomposition {#sec:bcirc_decomp}
-----------------------------
We are concerned in particular with dynamic linear systems where the matrices present in are block circulant. More formally, a block circulant MPC problem is defined as follows.
\[def:blockcircmpc\] Consider with $x_k\in\R^{nn_x},u_k\in\R^{nn_u}$ and $y_k\in\R^{pnn_y}$, where $p$ is a positive integer, and partition matrices $C$ and $D$ as $C=[C_1,\dots,C_p]^\Tr$ and $D=[D_1,\dots,D_p]^\Tr$, respectively. We say that is a *block circulant MPC problem* of order $n$ if the following conditions holds: $$\begin{aligned}
A,Q &\in \mathcal{BC}(n,n_x,n_x), \qquad &&B \in \mathcal{BC}(n,n_x,n_u),\nonumber\\
R &\in \mathcal{BC}(n,n_u,n_u), \qquad &&C_i \in \mathcal{BC}(n,n_y,n_x),\nonumber\\
D_i &\in \mathcal{BC}(n,n_y,n_u).\nonumber\end{aligned}$$ for $i=1,\dots,p$.
Any model satisfying the above conditions can be interpreted as a periodic interconnection of $n$ identical subsystems with identical constraints and objective function penalties. Introducing the integer $p$ allows for multiple constraint sets, e.g. to restrict the state $x_k$ and the input $u_k$ separately.
\[thm:blockdiagmpc\] The matrices $$\begin{aligned}
\psi_x &\eqdef F_n\otimes\I_{n_x},\label{eq:psix}\\ \psi_u &\eqdef F_n\otimes\I_{n_u},\\\psi_y & \eqdef \I_p \otimes (F_n \otimes \I_{n_y}),\end{aligned}$$ decompose the dynamics - of a block circulant MPC problem into $$\begin{aligned}
\label{eq:decompABCD}
&\psixy^H\ABCDthin\psixu = \begin{bmatrix}
\hat{A} & \hat{B} \\ \hat{C} & \hat{D}
\end{bmatrix},\end{aligned}$$ where $$\begin{aligned}
\hat{A}&=\diag\lbrace\hat{a}_1,\dots,\hat{a}_n\rbrace,\qquad
\hat{B}=\diag\lbrace\hat{b}_1,\dots,\hat{b}_n\rbrace,\nonumber\\
\hat{C}&=\begin{bmatrix}\diag\lbrace\hat{c}_1^1,\dots,\hat{c}_n^1\rbrace\\ \vdots\\ \diag\lbrace\hat{c}_1^p,\dots,\hat{c}_n^p\rbrace\end{bmatrix},\,\,\,
\hat{D}=\begin{bmatrix}\diag\lbrace\hat{d}_1^1,\dots,\hat{d}_n^1)\\ \vdots\\ \diag\lbrace\hat{d}_1^p,\dots,\hat{d}_n^p\rbrace\end{bmatrix}\nonumber,\end{aligned}$$ the objective function matrices into $$\begin{aligned}
\label{eq:decompQR}
&\psixu^H\QRthin\psixu = \begin{bmatrix}
\hat{Q} & 0\\
0 & \hat{R}\end{bmatrix},\end{aligned}$$ where $$\begin{aligned}
\hat{Q}&=\diag\lbrace\hat{q}_1,\dots,\hat{q}_n\rbrace,\qquad
\hat{R}=\diag\lbrace\hat{r}_1,\dots,\hat{r}_n\rbrace,\nonumber\end{aligned}$$ and the terminal cost matrix $P$ into $$\begin{aligned}
\label{eq:decompP}
\hat{P} = \psi_x^H P \psi_x = \diag\lbrace\hat{p}_1,\dots,\hat{p}_n\rbrace.\end{aligned}$$
According to Definition \[def:blockcircmpc\], and are a direct consequence of Corollary \[thm:diag\_bc\]. The condition is proven through Lemma \[thm:dare\].
\[thm:dare\] The terminal cost matrix $P$ solves if and only if $\hat{P} = \psi_x^H P \psi_x$ solves with $A=\hat{A}$, $B=\hat{B}$, $Q=\hat{Q}$ and $R=\hat{R}$ for the decomposed system. Moreover, $\hat{P} = \diag\lbrace\hat{p}_1,\dots,\hat{p}_n\rbrace$.
Multiplying with $\psi_x^H$ from the left and $\psi_x$ from the right, inserting $\psi_x^H\psi_x=\I_{nn_x}$ and $\psi_u^H\psi_u=\I_{nn_u}$ where appropriate and substituting - yields $$\begin{aligned}
\hat{A}^H \psi^H_x P \psi_x \hat{A} - &\hat{A}^H \psi_x^H P \psi_x \hat{B} \left(\hat{B}^H \psi_x^H P \psi_x \hat{B}+\hat{R}\right)^{-1}&\nonumber\\
&\hat{B}^H \psi_x^H P \psi_x \hat{A} + \hat{Q} =\psi_x^H P \psi_x,\end{aligned}$$ where the properties of block circulant matrices - have been used. The transformation $\hat{P}=\psi^H P \psi_x$ therefore solves for the decomposed system. Starting with $$\begin{aligned}
\label{eq:decompdare}
\hat{A}^H \hat{P} \hat{A} - \hat{A}^H \hat{P} \hat{B} \left(\hat{B}^H \hat{P} \hat{B}+\hat{R}\right)^{-1}\hat{B}^H \hat{P} \hat{A} + \hat{Q} =\hat{P}\end{aligned}$$ and reversing substitutions - yields $P=\psi_x \hat{P} \psi_x^H$. Since all matrices in are block diagonal with $n$ blocks of size $m$, it can be solved for each for the blocks independently. It follows that $\hat{P}=\diag\lbrace\hat{p}_1,\dots,\hat{p}_n\rbrace$.
Theorem \[thm:blockdiagmpc\] states that the periodic interconnection of $n$ identical subsystems can be decomposed into $n$ independent systems, often referred to as modal subsystems. Corollary \[thm:diag\_bc\] is then applied to the decomposed block circulant MPC problem.
\[thm:truncation\] After the decomposition of Theorem \[thm:blockdiagmpc\] has been applied to a block circulant MPC problem of order $n$, it is sufficient to examine the first $n/2$ (for $n$ even) or the first $(n-1)/2$ (for $n$ odd) blocks of matrices $\hat{A},\hat{B},\hat{Q},\hat{R},\hat{P}$ and $\hat{C_i},\hat{D_i}$ for $i=1,\dots,p$.
Direct consequence of Corollary \[thm:diag\_bc\].
In case the original matrices are full, i.e. they have $n^2n_{x,y,u}^2$ nonzero (real) elements with $p=1$, Theorem \[thm:blockdiagmpc\] reduces the number of nonzero elements by $1/n$, resulting in $nn_{x,y,u}^2$ nonzero complex elements. Moreover, by exploiting the block-diagonal structure and pattern of complex conjugates, Corollary \[thm:truncation\] states that only half of the blocks are required for the ADMM algorithm. Similarly, when a vector is projected into the Fourier domain by setting $\hat{x} = \psi_x^H x$, only the first $nn_x/2$ elements must be examined.
Before advancing to the decomposition of , the truncation of Corollary \[thm:truncation\] and its counterpart operation, the augmentation, are formally defined.
\[def:truncaug\] Given $\inC{A=\diag\lbrace a_1,\dots,a_n\rbrace}{np}{nm}$, $x\in\C^{np}$ and $n$ even (odd), let $\trunc$ be the operator that extracts the first $\frac{n}{2}$ ($\frac{n-1}{2}$) blocks of $A$ and the first $\frac{n}{2}p$ rows of $x$. Conversely, let $\aug$ be the inverse operator which accepts a truncated vector $\p{x}=(x_1,\dots,x_{\frac{n}{2}})^\Tr\in\C^{\frac{n}{2}p}$ with $x_i\in\C^p$ and returns $x=(x_1,\dots,x_{\frac{n}{2}},\bar{x}_{\frac{n}{2}-1,\dots,\bar{x}_2})^\Tr\in\C^{np}$ for even $n$. For odd $n$, operator $\aug$ accepts a vector $\p{x}=(x_1,\dots,x_{\frac{n-1}{2}})^\Tr\in\C^{\frac{n-1}{2}p}$ with $x_i\in\C^p$ and returns $x=(x_1,\dots,x_{\frac{n-1}{2}},\bar{x}_{\frac{n-1}{2},\dots,\bar{x}_2})^\Tr\in\C^{np}$. Define $\aug$ in a similar way when invoked with a truncated block diagonal matrix $\p{A}=\diag\lbrace a_1,\dots,a_\frac{n}{2}\rbrace$. In addition, define $$\begin{aligned}
\trunc_N & \eqdef\I_N\otimes \trunc,\\
\aug_N & \eqdef \I_N\otimes \aug,\end{aligned}$$ i.e. the operators $\aug$ and $\trunc$ applied $N$ times.
Note that $\aug$ and $\trunc$ are linear operators. In addition, it holds that $\trunc\lbrace J z\rbrace=\trunc\lbrace J \rbrace \trunc \lbrace z \rbrace$ and $\aug\lbrace \p{J} \p{z}\rbrace=\aug\lbrace \p{J} \rbrace \aug \lbrace \p{z} \rbrace$ for block diagonal matrices $J,\p{J}$ and vectors $z,\p{z}$ of appropriate sizes.
Block Circulant ADMM Algorithm {#sec:blockcirculantadmm}
==============================
Constrained QP with Block Circulant Blocks {#sec:block_qp}
------------------------------------------
\[def:blockcircqp\] The following real valued constrained QP
$$\begin{aligned}
\min \,\,\,&\frac{1}{2}z^\Tr J z +q^\Tr z\\
s.t.\,\,\, &Kz-v=0\\
&\ubar{v} \leq v \leq \bar{v}.\label{eq:blockcircqp_ineq}\end{aligned}$$
\[eq:blockcircqp\]
is called a *constrained block circulant QP* of order $n$ if there exists a partitioning of vectors $z$ and $v$ into $N_z$ and $N_v$ segments of lengths $l_z^1,\dots,l_z^{N_z}$ and $l_v^1,\dots,l_v^{N_v}$, respectively, that partition matrices $J$ and $K$ as $$\begin{aligned}
J = \begin{bmatrix}
J_{11} & \dots & J_{1N_z}\\
\vdots & & \vdots\\
J_{N_z1} & \dots & J_{N_zN_z}
\end{bmatrix},
K = \begin{bmatrix}
K_{11} & \dots & K_{1N_z}\\
\vdots & & \vdots\\
K_{N_v1} & \dots & K_{N_vN_z}
\end{bmatrix},\end{aligned}$$ for which all blocks are block circulant matrices of order $n$, $$\begin{aligned}
J_{kj} \in \mathcal{BC}(n,l_z^k,l_z^j),\qquad
K_{wj} \in \mathcal{BC}(n,l_v^w,l_z^j),\end{aligned}$$ for $k,j=1,\dots,N_z$ and $w=1,\dots,N_v$.
The augmented Lagrangian for problem is obtained as $$\begin{aligned}
\label{eq:lagrangian}
\begin{split}
L(z,v,\gamma,\ubar{\lambda},&\bar{\lambda}) = \frac{1}{2}z^\Tr J z +q^\Tr z +\frac{\rho}{2} \twonorm{Kz-v}^{2}+\\
&\gamma^\Tr(Kz-v)+\ubar{\lambda}^\Tr(-v+\ubar{v})+\bar{\lambda}^\Tr(v-\bar{v}),
\end{split}\end{aligned}$$ where we have introduced dual variables $\ubar{\lambda}$ and $\bar{\lambda}$ associated with the inequality constraints . An optimal solution of is a saddle point of and must satisfy the following Karush-Kuhn-Tucker (KKT) conditions [@BOYDCONVEX],
$$\begin{aligned}
Kz^* - v^* = 0,\quad v^*-\bar{v} \leq 0,\quad -v^*+\ubar{v} \leq 0\label{eq:pf}\tag{PF}\\
\bar{\lambda}^* \geq 0,\quad \ubar{\lambda}^* \geq 0\label{eq:df}\tag{DF}\\
\bar{\lambda}^{*\Tr}(v^*-\bar{v}) = 0,\quad \ubar{\lambda}^{*\Tr}(\ubar{v}-v^*) = 0\label{eq:cs}\tag{CS}\\
\nabla L = \begin{bmatrix}\frac{J+J^\Tr}{2}+\rho K^\Tr K & -\rho K^\Tr\\ -\rho K & \rho\end{bmatrix}
\begin{pmatrix}z^* \\ v^*\end{pmatrix} +
\begin{bmatrix}K^\Tr\\ -\I\end{bmatrix}\gamma^*\nonumber\\
+\begin{bmatrix}0\\ \I\end{bmatrix}\bar{\lambda}^* +
\begin{bmatrix}0\\ -\I\end{bmatrix}\ubar{\lambda}^*+ \begin{pmatrix}q \\ 0\end{pmatrix}=0\label{eq:stat}\tag{ST}\end{aligned}$$
\[eq:KKToriginal\]
where stands for primal feasibility, for dual feasibility, complementary slackness and for stationarity. The following Corollary connects a block circulant MPC problem from Definition \[def:blockcircmpc\] to the constrained block circulant quadratic program (CBCQP):
\[thm:blocpmpisblockqp\] A block circulant MPC problem leads to a CBCQP with $N_z=T$, $N_v=Tp$ and $J_{kj} \in \mathcal{BC}(n,n_u,n_u)$ and $K_{wj} \in \mathcal{BC}(n,n_y,n_u)$.
This is a direct consequence of properties - and Definitions and .
As the matrices in Definition \[def:blockcircqp\] are composed of block circulant matrices, the results from Sections \[sec:circ\_matrices\] and \[sec:bcirc\_decomp\] suggest that there exists a coordinate transformation $(\tilde{z},\tilde{v})=(\psi_z^Hz, \psi_v^Hv)$ that block diagonalizes each block of $K$ and $J$. However, how the complex-valued transformation affects the minimization in is not immediately obvious. The following theorem answers this question:
\[thm:decompblockqp\] Given a CBCQP of order $n$, then the following CBCQP,
$$\begin{aligned}
\min_{\tilde{z}\in\Sc_z,\tilde{v}\in\Sc_v} \,\,\,&\frac{1}{2}\tilde{z}^H \tilde{J} \tilde{z} +\tilde{q}^H \tilde{z}\label{eq:tildeqpA}\\
s.t.\,\,\, &\tilde{K}\tilde{z}-\tilde{v}=0\label{eq:tildeqpB}\\
&\ubar{v} \leq \psi_v\tilde{v} \leq \bar{v}\label{eq:tildeqpC}.\end{aligned}$$
\[eq:blockcircqp\_decomp\]
where $\tilde{q}=\psi_z^H q,\,\,\tilde{J}=\psi_z^H J \psi_z,\,\,\tilde{K}=\psi_v^H K \psi_z,\,\,\psi_j = \diag \lbrace F_n\otimes\I_{l_j^1},\dots,F_n\otimes\I_{l_j^{N_j}} \rbrace$ for $j=\lbrace z,v\rbrace$ and the sets $\Sc_j$ restrict each of the segments of $\tilde{z}$ and $\tilde{v}$ to the pattern of complex conjugates from Corollary \[thm:diag\_bc\], is equivalent to in the sense that
$$\begin{gathered}
z^* = \psi_z \tilde{z}^*,\quad v^* = \psi_v \tilde{v}^*,\\
\gamma^* = \psi_v \tilde{\gamma}^*,\quad \bar{\lambda}^* = \tilde{\bar{\lambda}}^*,\quad \ubar{\lambda}^* = \tilde{\ubar{\lambda}}^*,\end{gathered}$$
\[eq:decomp\_equivalence\]
where $(z,v,\gamma,\bar\lambda,\ubar\lambda)^*$ and $(\tilde z,\tilde v,\tilde \gamma,\tilde{\bar\lambda},\tilde{\ubar\lambda})^*$ are primal and dual optimizers for and , respectively.
It is sufficient to show that using , the KKT conditions of yield . The augmented Lagrangian for is obtained as $$\begin{aligned}
\label{eq:lagrangian_decomp}
\begin{split}
\tilde{L}(&\tilde{z},\tilde{v},\tilde{\gamma},\tilde{\ubar{\lambda}},\tilde{\bar{\lambda}}) = \frac{1}{2}\tilde{z}^H \tilde{J} \tilde{z} +\tilde{q}^H \tilde{z} + \frac{\rho}{2} \twonorm{\tilde{K}\tilde{z}-\tilde{v}}^{2}\\
&+\tilde{\gamma}^H(\tilde{K}\tilde{z}-\tilde{v})+\tilde{\ubar{\lambda}}^\Tr(-\psi_v\tilde{v}+\ubar{v})+\tilde{\bar{\lambda}}^\Tr(\psi_v\tilde{v}-\bar{v})\\
&+\mathcal{I}_{\mathcal{S}_z}(\tilde{z})+\mathcal{I}_{\mathcal{S}_v}(\tilde{v}).
\end{split}\end{aligned}$$ Note that because $\tilde{z}\in\Sc_z$ and $\tilde{v}\in\Sc_v$, $\psi_v\tilde{v}$ is real-valued and $\tilde{K}\tilde{z}-\tilde{v}\in\Sc_v$. The former implies that $\tilde{\bar{\lambda}}$ and $\tilde{\ubar{\lambda}}$ are real-valued. The latter implies that $\tilde{\gamma}$ must have the same pattern of complex conjugates as $\tilde{v}$, which can be seen by formulating as two inequality constraints and evaluating the complementary slackness conditions. In addition, both observations entail that $\tilde{L}$ is a real-valued function. Conditions , and are recovered by substituting $\tilde{K}=\psi_v^H K \psi_z$ and . The partial derivatives of $\tilde{L}$ are calculated as $\frac{\partial \tilde{L}}{\partial (\cdot)} = \frac{\partial \tilde{L}}{\partial Re(\cdot)}+i\frac{\partial \tilde{L}}{\partial Im(\cdot)}$ [@ADAPTIVEFILTER p. 798]. The gradient $\nabla \tilde{L}$ evaluates to $$\begin{aligned}
\label{eq:gradienttilde}
\begin{split}
\begin{bmatrix}\frac{\tilde{J}+\tilde{J}^H}{2}+\rho \tilde{K}^H \tilde{K} & -\rho \tilde{K}^H\\ -\rho \tilde{K} & \rho\end{bmatrix}
\begin{pmatrix}\tilde{z}^* \\ \tilde{v}^*\end{pmatrix} +
\begin{bmatrix}\tilde{K}^H\\ -\I\end{bmatrix}\tilde{\gamma}^*\\
+\begin{bmatrix}0\\ \psi_v^H\end{bmatrix}\tilde{\bar{\lambda}}^* +
\begin{bmatrix}0\\ -\psi_v^H\end{bmatrix}\tilde{\ubar{\lambda}}^*+ \begin{pmatrix}\tilde{q} \\ 0\end{pmatrix}=0.
\end{split}\tag{$\tilde{\text{ST}}$}\end{aligned}$$ Pre-multiplying by $\diag\lbrace \psi_z,\psi_v\rbrace$ and inserting claim completes the proof.
Now that CBCQP has been block diagonalized, the following theorem uses the pattern of complex conjugates from Corollary \[thm:diag\_bc\] in order to truncate problem .
\[thm:truncblockqp\] Given a CBCQP of order $n$ which has been decomposed according to Theorem \[thm:decompblockqp\], then the following CBCQP,
$$\begin{aligned}
\min \,\,\,&\frac{1}{2}Re(\hat{z}^H \hat{J} \hat{z}) +Re(\hat{q}^H \hat{z})\label{eq:hatqpA}\\
s.t.\,\,\, &\aug_{N_v}\lbrace\hat{K}\hat{z}-\hat{v}\rbrace=0\label{eq:hatqpB}\\
&\ubar{v} \leq \psi_v\aug_{N_v}\lbrace\hat{v}\rbrace \leq \bar{v}\label{eq:hatqpC}.\end{aligned}$$
\[eq:blockcircqp\_trunc\]
where $\hat{q}=\trunc_{N_z}\tilde{q},\,\,\hat{J}=\trunc_{N_z}\tilde{J}$ and $\hat{K}=\trunc_{N_v}\tilde{K}$, is equivalent to in the sense that
$$\begin{gathered}
\hat{z}^* = \trunc_{N_z} \tilde{z}^*,\quad \hat{v}^* = \trunc_{N_v} \tilde{v}^*,\\
\hat{\gamma}^* = \tilde{\gamma}^*,\quad \hat{\bar{\lambda}}^* = \tilde{\bar{\lambda}}^*,\quad \hat{\ubar{\lambda}}^* = \tilde{\ubar{\lambda}}^*,\end{gathered}$$
\[eq:trunc\_equivalence\]
where $(\hat{z},\hat{v},\hat{\gamma},\hat{\bar\lambda},\hat{\ubar\lambda})^*$ and $(\tilde z,\tilde v,\tilde \gamma,\tilde{\bar\lambda},\tilde{\ubar\lambda})^*$ are primal and dual optimizers for and , respectively.
As for the proof of Theorem \[thm:decompblockqp\], one can show that problem and satisfy the same KKT conditions under claim . Note that it is necessary to augment the equality constraints in order to obtain a real-valued Lagrangian: $$\begin{aligned}
\label{eq:lagrangian_trunc}
\begin{split}
\hat{L}(\hat{z},\hat{v},&\hat{\gamma},\hat{\ubar{\lambda}},\hat{\bar{\lambda}}) =\frac{1}{2}Re(\hat{z}^H \hat{J} \hat{z}) +Re(\hat{q}^H \hat{z})\\
&+ \frac{\rho}{2} \twonorm{\hat{K}\hat{z}-\hat{v}}^{2}+\hat{\gamma}^H\aug_{N_v}\lbrace\hat{K}\hat{z}-\hat{v}\rbrace\\
&+\hat{\ubar{\lambda}}^\Tr(-\psi_v\aug_{N_v}\hat{v}+\ubar{v})+\hat{\bar{\lambda}}^\Tr(\psi_v\aug_{N_v}\hat{v}-\bar{v}).
\end{split}\end{aligned}$$ According to Definition \[def:truncaug\], $\aug_{N_v}\lbrace\hat{v}\rbrace\in\Sc_v$ and $\aug_{N_z}\lbrace\hat{z}\rbrace\in\Sc_z$. Conditions PF, DF and CS are recovered using $\tilde{K}=\aug_{N_v}\lbrace \hat{K}\rbrace$ and . The gradient of the augmented Lagrangian of the truncated problem is obtained as $$\begin{aligned}
\label{eq:gradienthat}
\begin{split}
\begin{bmatrix}\frac{\hat{J}+\hat{J}^H}{2}+\rho \hat{K}^H \hat{K} & -\rho \hat{K}^H\\ -\rho \hat{K} & \rho\end{bmatrix}
\begin{pmatrix}\hat{z}^* \\ \hat{v}^*\end{pmatrix} +
\begin{bmatrix}\hat{K}^H\\ -\I\end{bmatrix}\trunc_{N_v}\hat{\gamma}^*\\
+\begin{bmatrix}0\\ \trunc_{N_v}\lbrace \psi_v^H \hat{\bar{\lambda}}^* -\psi_v^H \hat{\ubar{\lambda}}^*\rbrace\end{bmatrix}+ \begin{pmatrix}\hat{q} \\ 0\end{pmatrix}=0.
\end{split}\tag{$\hat{\text{ST}}$}\end{aligned}$$ Augmenting both rows of yields and completes the proof.
ADMM for Block Circulant MPC {#sec:circ_admm}
----------------------------
The version of Algorithm \[alg:admm\] for a CBCQP, or equivalently a block circulant MPC problem, is outlined in Algorithm \[alg:circ\_admm\] and the individual steps are presented in the following paragraphs.
State $x(t)$ Input $u(t)$ Set $x_0^0=x(t)$ and $\hat{y}^0,\hat{\gamma}^0=0$; compute $\ubar{v},\bar{v}$ and $\hat{q}$ Update $\hat{z}^i$ using $(\hat{\text{SP1}})$ Update $\hat{v}^i$ using $(\hat{\text{SP2}})$ Update $\trunc_{N_v}\hat{\gamma}^i$ using $(\hat{\text{SP3}})$ $u(t)=\psi_u \aug\lbrace (\hat{z}_0,\dots,\hat{z}_{\frac{n-1}{2}n_u})^\Tr \rbrace$
Before entering the loop, Algorithm \[alg:circ\_admm\] requires the mapping of $q$ into the complex Fourier domain.
Next, the algorithm solves projected onto the Fourier domain, $$\begin{aligned}
\label{eq:sp1hat}
\left( \hat{J}+\rho \hat{K}^H\hat{K}\right) \hat{z}^i = \hat{K}^H (\rho \hat{v}^{i-1}-\trunc_{N_v}\hat{\gamma}^{i-1}) -\hat{q}.\tag{S\^{P}1}\end{aligned}$$ When subproblem is projected onto the Fourier domain as in , it is simplified in two ways. On one hand, the block diagonalized matrices have been reduced to a maximum of $Tn(\max\lbrace n_u,pn_y\rbrace)^2$ nonzero elements, where $T$ is the prediction horizon of the MPC problem. On the other, the pattern of complex conjugates allows vectors and matrices to be truncated as in Theorem \[thm:decompblockqp\].
The modified subproblem reads as $$\begin{aligned}
\label{eq:sp2hat_sat}
\hat{v}^i = \trunc_{N_v}\lbrace \psi_v^H\sat_{[\ubar{v},\bar{v}]}\left\lbrace \psi_v\aug_{N_v}\lbrace\hat{K}\hat{z}^i+\rho^{-1}\trunc_{N_v}\hat{\gamma}^{i-1}\rbrace\right\rbrace\rbrace.\nonumber\tag{S\^{P}2}\end{aligned}$$ The discrete Fourier transformations required in subproblem are the main drawbacks of algorithm and the problem sizes required to outperform algorithm are discussed in section \[sec:complexity\].
The decomposed and truncated dual variables are updated using $$\begin{aligned}
\label{eq:sp3hat}
\trunc_{N_v}\hat{\gamma}^i = \trunc_{N_v}\hat{\gamma}^{i-1} +\rho(\hat{K}\hat{z}^i-\hat{y}^i). \nonumber\tag{S\^{P}3}\end{aligned}$$ As mentioned in the proof of Theorem \[thm:truncblockqp\], $\hat{\gamma}\in\Sc_v$ and it is therefore sufficient to update the truncated dual variable. In practice, the term $\hat{K}\hat{z}^i$ can be cached during and reused in .
Finally, the convergence of Algorithm \[alg:circ\_admm\] is verified by $$\begin{aligned}
\label{eq:convergence}
\twonorm{\aug_{N_v}\lbrace\hat{v}^i-\hat{v}^{i-1}\rbrace}^2 < \epsilon,\\
\twonorm{\hat{\gamma}^i-\hat{\gamma}^{i-1}}^2 < \epsilon.\end{aligned}$$ To see that this criterion is equivalent to the one of Algorithm \[alg:admm\], note that $\twonorm{\aug_{N_v}\lbrace \hat{v}\rbrace}=\twonorm{\psi_v^Hv}=\twonorm{v}$. In practice, the augmentation is not required as we can relate $\twonorm{v}^2$ and $\twonorm{\hat{v}}^2$ by $$\begin{aligned}
2\twonorm{\hat{v}}^2-\sum_{k=1}^{N_v}(\hat{v}^k_1)^\Tr\hat{v}^k_1 = \twonorm{v}^2,\end{aligned}$$ where we assumed that $n$ is odd and where the sum over the segments of $\hat{v}$ extracts the first (real) block of each segment. Note that for the returned value $u(t)$ in the last step of Algorithm \[alg:circ\_admm\] it is assumed that $n$ is odd as well.
As shown in Theorems \[thm:decompblockqp\] and \[thm:truncblockqp\], Algorithms \[alg:admm\] and \[alg:circ\_admm\] produce an equivalent solution in the sense that the minimizers $z^*$ and $\hat{z}^*$ are related by $z^*=\psi_z\aug_{N_z}\lbrace\hat{z}^*\rbrace$.
Computational Complexity {#sec:complexity}
------------------------
In this section, the computational complexities of Algorithm \[alg:admm\] and \[alg:circ\_admm\] are compared. It is assumed that the matrices in Definition \[def:blockcircqp\] are full and that $\inR{J}{(Nn_zn)}{(Nn_zn)}$ and $\inR{K}{(Nn_vn)}{(Nn_zn)}$ are partitioned into segments of identical lengths $n_z$ and $n_v$, respectively, with $n_v=n_z$. Algorithm \[alg:circ\_admm\] makes use of discrete Fourier transformations during initialization, finalization and in subproblem . These can be performed using the identity and a sequence of fast Fourier transformations (FFT). For a vector of length $r$, the complexity of the FFT is $\co{r\log r}$.
For both algorithms, the main burden lies in solving the linear system in subproblems and . Depending on the structure of the linear system, it can be solved in numerous ways. For simplicity, it is assumed that the linear system is solved using a matrix inverse that is pre-computed offline. In that case, is of complexity $\co{(Nnn_z)^2}$, while the block diagonalization and truncation in reduces the complexity to $\co{4\frac{n}{2}(Nn_z)^2}$, where the factor $4$ accounts for the required complex arithmetic. Because the operations of the saturation function are negligible, accounts for a complexity of $\co{(Nnn_z)^2+Nnn_z}$. Note that the term $Kz^i$ is reused in . As the projection onto the boundaries $[\ubar{v},\bar{v}]$ must be carried out in the original domain, the drawbacks of the Fourier transformation become evident in . This results in a complexity of $\co{4\frac{n}{2}(Nn_z)^2+2Nn_zn\log n}$. Lastly, and are of complexities $\co{Nnn_z}$ and $\co{2N\frac{n}{2}n_z}$, respectively.
The total complexities of Algorithm \[alg:admm\] and \[alg:circ\_admm\] are summarized in Table \[tbl:complexity\]. The totals reveal that under the assumptions of this section, Algorithm \[alg:circ\_admm\] is cheaper than \[alg:admm\] if $n>2$ and $Nn_z>3$.
Algorithm \[alg:admm\] Algorithm \[alg:circ\_admm\]
------- ------------------------ --------------------------------
$\co{(Nnn_z)^2}$ $\co{2n(Nn_z)^2}$
$\co{(Nnn_z)^2}$ $\co{2n(Nn_z)^2+2Nn_zn\log n}$
$\co{Nnn_z}$ $\co{Nnn_z}$
Total $\co{Nnn_z(2Nnn_z+1)}$ $\co{Nnn_z(4Nn_z+1+2\log n)}$
: Computational Complexity for Algorithms \[alg:admm\] and \[alg:circ\_admm\][]{data-label="tbl:complexity"}
Simulations {#sec:simulations}
===========
Random Constrained QP with Block Circulant Blocks {#sec:random_qp}
-------------------------------------------------
We first consider a set of randomly generated constrained block circulant QPs to gauge performance of Algorithm \[alg:circ\_admm\]. Figure \[fig:decomposition\] shows the decomposed left-hand side of , $J+\rho K^\Tr K$, of a QP of order $n=4$ before applying the truncation. On one hand, it is evident how the transformation matrices $\psi_z$ and $\psi_v$ block diagonalize $J+\rho K^\Tr K$. On the other, the pattern of complex conjugate blocks motivating the truncation becomes apparent.
![Sparsity patterns of $J+\rho K^\Tr K$ (left) and $\tilde{J}+\rho \tilde{K}^H \tilde{K}$ (right) for $N_z=N_v=3$ and $l_z^j=l_v^j=4,\,j=1,2,3$. The colors are proportional to the magnitude of the matrix elements.[]{data-label="fig:decomposition"}](decomposition)
\[fig:resultsqp\]
[plots/results\_qp\_3\_logy]{} (45,-3) (-4,48) (-4,32) (-4,9.5)
\[fig:resultsqp\]
[results\_qp\_3\_logy\_large]{} (45,-3) (-2.5,66.5) (-2.5,42.5) (-2.5,13)
Figure \[fig:resultsqp\] shows the execution times of Algorithms \[alg:admm\] and \[alg:circ\_admm\] for random block circulant QPs of increasing order $n$ with $N=1$ and $l_z=l_v=10$. Even though the analysis of Section \[sec:complexity\] accounted for the required Fourier transformations, Figure \[fig:resultsqp\] reveals that for small $n$ the additional operations required in such as truncating, permuting and augmenting the vectors are not negligible. For larger $n$ these side effects lose their significance and the superiority of Algorithm \[alg:circ\_admm\] becomes evident.
Ring of Masses {#sec:oscillators}
--------------
Consider $n$ identical masses arranged in a ring of radius $R$ and connected through springs and dampers. At equilibrium the masses are uniformly spaced around the ring at angles $\phi_j=\frac{2\pi}{n}$ for $j=1,\dots,n$. The dynamics of a small deviation $\Delta\phi_j$ of mass $j$ from the equilibrium angle $\phi_j$ are obtained as $$\begin{aligned}
\label{eq:oscillator}
\begin{split}
\Delta\ddot{\phi}_j = &\frac{k}{m}(\Delta\phi_{j+1}+\Delta\phi_{j-1}-2\Delta\phi_{j})\\&+ \frac{d}{m}(\Delta\dot{\phi}_{j+1}+\Delta\dot{\phi}_{j-1}-2\Delta\dot{\phi}_{j})+\frac{1}{m}T_j,
\end{split}\end{aligned}$$ where all indices are modulo $n$, $T_j$ is a controllable torque acting on each mass, $m$ is the mass and $k$ and $d$ are the spring and damper coefficients, respectively. Discretizing the dynamics w.r.t. time and setting $x_k=(\Delta\phi_1^k,\Delta\dot{\phi}^k_1,\dots,\Delta\phi_n^k,\Delta\dot{\phi}^k_n)^\Tr$ yields a discrete-time block circulant dynamical system. By assigning $u_k=(T_1^k,\dots,T_n^k)^\Tr$, a block circulant MPC problem of order $n$ with $n_u=1$ and $n_x=2$ can be defined. For the following experiments, matrices $Q$ and $R$ were chosen to be identity matrices, the prediction horizon $T$ was set to $10$ and the states $x_k$ and the inputs $u_k$ were constrained separately, i.e. $C_1 = \I_{nn_x}$, $D_1=0$ and $C_2=0$, $D_2=\I_{nn_u}$. According to Corollary \[thm:blocpmpisblockqp\], the latter MPC problem defines a CBCQP.
\[fig:results\_oscillator\]
[plots/results\_oscillator\_4\_logy]{} (45,-3) (-4,48) (-4,32) (-4,10)
\[fig:results\_oscillator\]
[results\_oscillator\_logy\_large]{} (45,-3) (-2.5,66.5) (-2.5,42.5) (-2.5,13)
Figure \[fig:results\_oscillator\] compares the performance of Algorithm \[alg:admm\] and \[alg:circ\_admm\] solving the MPC problem with random initial conditions for an increasing number of masses $n$. Each problem was solved with an average number of 25 ADMM iterations. As for the example of Section \[sec:random\_qp\], Algorithm \[alg:circ\_admm\] performs worse than Algorithm \[alg:admm\] for small $n$ when the drawbacks of the Fourier transformation in and other side effects outweigh the computational gains in . The benefits of the Fourier transformation become evident for larger $n$.
Conclusions
===========
This paper demonstrated how to exploit the particular structure of a MPC problem for block circulant systems. Based on the properties of block circulant matrices, a block circulant MPC problem was defined and connected to a general constrained QP with block circulant blocks. A transformation was derived which block diagonalizes any constrained QP with block circulant blocks and allows to truncate the transformed vectors. A modified ADMM algorithm for the transformed and truncated system was developed. The modified ADMM algorithm was tested using a series of random constrained QPs with block-circulant problem data and using an academic example of a block-circulant MPC problem. In both cases, the evaluation of the results revealed that the modified ADMM algorithm performs significantly better for increasing problem sizes.
[^1]: $^*$Corresponding author: . All authors are with the Department of Engineering Science, University of Oxford, Oxford, UK. This research is supported by the Engineering and Physical Sciences Research Council (EPSRC) with a Diamond CASE studentship.
[^2]: Note that we assume that ADMM is *cold-started* in Algorithm \[alg:admm\] at each time step for simplicity, but in practice one would warm start the variables $(v^0,\gamma^0)$ from a previous solution.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The prisoner’s dilemma (PD) is a game-theoretic model studied in a wide array of fields to understand the emergence of cooperation between rational self-interested agents. In this work, we formulate a spatial iterated PD as a discrete-event dynamical system where agents play the game in each time-step and analyse it algebraically using Krohn-Rhodes algebraic automata theory using a computational implementation of the holonomy decomposition of transformation semigroups. In each iteration all players adopt the most profitable strategy in their immediate neighbourhood. Perturbations resetting the strategy of a given player provide additional generating events for the dynamics. Our initial study shows that the algebraic structure, including how natural subsystems comprising permutation groups acting on the spatial distributions of strategies, arise in certain parameter regimes for the pay-off matrix, and are absent for other parameter regimes. Differences in the number of group levels in the holonomy decomposition (an upper bound for Krohn-Rhodes complexity) are revealed as more pools of reversibility appear when the temptation to defect is at an intermediate level. Algebraic structure uncovered by this analysis can be interpreted to shed light on the dynamics of the spatial iterated PD.'
author:
- 'Isaiah Farahbakhsh [^1]'
- 'Chrystopher L. Nehaniv [^2] [^3]'
title: 'Spatial Iterated Prisoner’s Dilemma as a Transformation Semigroup'
---
Introduction {#sec:1}
============
Krohn-Rhodes (KR) theory offers powerful tools for understanding discrete-event dynamical systems (e.g. [@nehaniv2015symmetry]). This theory decomposes any system whose dynamics can be represented as a transformation semigroup into a cascade of permutation-group layers and identity-reset (flip-flop) layers using the wreath product [@krohn1965algebraic]. This yields a “coarse-to-fine graining” of both the system’s state and its dynamical transformations. The decomposition process can uncover subsystems represented by permutation groups which we call [*pools of reversibility*]{} or [*natural subsystems*]{} (see below). Algebraic structure uncovered by this analysis can be interpreted to shed light on the dynamics and complexity of many broad classes of discrete-event dynamical systems, including models found in the field of game theory.
The Prisoner’s Dilemma (PD) is an extensively studied game which explores the problem of individual vs. collective profit in a simple 2-strategy model. The model is usually presented describing a situation where two partners-in-crime are imprisoned and unable to communicate. The prosecutors lack evidence and can only convict each prisoner for a lesser charge, so they offer the prisoners a deal. This deal comes as a dilemma to the prisoners as they need to choose between remaining silent or betraying their partner which would grant them a sentence lighter than that of the lesser charge, only if their partner remains silent. These two options can be represented as strategies in a game where remaining silent is referred to as cooperation and betraying the partner-in-crime is referred to as defection. This game can be applied to any situation in which there is a temptation for individuals to defect, however the net benefit of all parties is maximized if all individuals cooperate. It has been used to study the emergence of cooperation in a wide array of models in the fields of ecology, economics and psychology [@clark2001sequential; @weitz2016oscillating; @wong2005dynamic].
In the PD, the payoff matrix for a given player is usually represented by: $$% \setlength\extrarowheight{2pt}
\renewcommand{\arraystretch}{1.45}
\setlength\tabcolsep{3pt}
\begin{tabular}{cc|c|c|}
& \multicolumn{1}{c}{} & \multicolumn{2}{c}{Player $2$}\\
& \multicolumn{1}{c}{} & \multicolumn{1}{c}{$D$} & \multicolumn{1}{c}{$C$} \\\cline{3-4}
\multirow{2}*{Player $1\:$} & $D\:$ & $(a,a)$ & $(b,c)$ \\\cline{3-4}
& $C\:$ & $(c,b)$ & $(d,d)$ \\\cline{3-4}
\end{tabular}$$ where every cell corresponds to each player choosing a strategy of either defect $(D)$ or cooperate $(C)$. The first and second elements of the tuple within each cell represent the payoff of players 1 and 2 respectively. To represent the dilemma, the payoffs are formulated with $b>d>a>c$ and to have the net payoff maximized for two cooperators, the system is further restricted such that $2d>b+c$. A common payoff matrix satisfying these conditions is: $$% \setlength\extrarowheight{2pt}
\setlength\tabcolsep{3pt}
\renewcommand{\arraystretch}{1.45}
\begin{tabular}{cc|c|c|}
& \multicolumn{1}{c}{} & \multicolumn{2}{c}{Player $2$}\\
& \multicolumn{1}{c}{} & \multicolumn{1}{c}{$D$} & \multicolumn{1}{c}{$C$} \\\cline{3-4}
\multirow{2}*{Player $1\:$} & $D\:$ & $(1,1)$ & $(b,0)$ \\\cline{3-4}
& $C\:$ & $(0,b)$ & $(3,3)$ \\\cline{3-4}
\end{tabular}
\label{eq:1}$$ where $b>3$ is a parameter referred to as the temptation to defect. When simulated as a two-player game, the players’ strategies will always converge to defection since it is the Nash equilibrium [@rubinstein1986finite]. However when iterated on a spatial structure with local interactions, more complex behaviour arises, including the persistence of cooperation due to the spatial clustering of alike strategies [@nowak1992evolutionary; @nowak1994spatial].
Spatial Algebraic Model {#sec:2}
=======================
For the model presented in this paper, the PD is iterated on rectangular lattice with periodic boundary conditions where each cell represents a player with one of two strategies; defection represented by ‘0’ and cooperation represented by ‘1’. A small $2\times 3$ lattice is used here due to current resource constraints of the computational algebraic analysis, but is illustrative of the general phenomena that arise.
The state space, $$X_{bin}=\{000000,000001,000010,000011,...,111110,111111\}$$ is made of 64 6-bit binary strings, where the $i$^th^ bit from the left represents the strategy of cell $i$ (Figure \[lattice\]). For a more notationally compact representation, this state set can also be written in decimal form with each state being the decimal integer equivalent of the binary string, $$X=\{0,1,2,3,...,62,63\}.$$
During each synchronous playing of the game, $t$ (which we call a time step), each cell plays the PD with their von Neumann neighbours and gains a net payoff over all games using the payoff matrix (\[eq:1\]). Note that since the system is a $2\times 3$ lattice, each cell has 3 von Neumann neighbours to avoid double-counting existing neighbours with the periodic boundaries. After playing against each other, each cell updates its strategy to match that of their neighbour with maximal payoff, only if the maximal payoff is greater than their own. If two neighbouring cells with different strategies have the same maximal payoff, then cooperation is chosen.
To allow for more complexity, the model was formulated such that certain cells can have their strategy perturbed outside of $t$-dependent strategy evolution. We call these cells “open”. If cell $i$ is open, there are two locally constant mappings associated with that cell; $d_i$ and $c_i$. These correspond to mapping the strategy of cell $i$ to defection or cooperation respectively, regardless of the change in payoff. These mappings on the set of open cells (denoted $O$) make a set of locally constant mappings, resetting cell $i$’s strategy to either $d$ or $c$ but leaving others’ unchanged. $$T'_O=\{d_i,c_i\}_{i\in O}.$$ The set of generators for the semigroup transformations is then given by $$T_O=T'_O\bigcup\{t\}.$$ Words made from elements of $T_O$ define mappings on the set of states by applying each transformation in order from left to right. The set of transformations generated from $T_O$ comprise a semigroup denoted by $$S_O=\langle T_O \rangle,$$ and $S_O$ acting on $X$ gives us the transformation semigroup $(X,S_O)$. As $T'_O$ is a set of locally constant mappings, it does not depend on the parameter $b$, however $t$ does and its $b$-dependence was explored using a *python* script which also generated the semigroup mappings. In this analysis the strict inequality $b>3=d$ was relaxed so that $b\geq 3$. Note that for $b=3$, the system still favours defection since although mutual cooperation has become a weak Nash equilibrium, mutual defection is still the only strict Nash equilibrium, meaning no player can change their strategy without suffering a loss in payoff. The mappings generated by the *python* script were then read into *GAP* [@linton2007gap] and the transformation semigroup was analyzed using the *SgpDec* package [@egri2014sgpdec] to carry out a holonomy decomposition [@eilenberg76; @nehaniv2015symmetry; @egri2015computational]. This yields a KR decomposition of the spatial PD model’s dynamics $(X,S_O)$ by identifying [*natural subsystems*]{}, i.e., nontrivial permutation groups whose state set is an image $X\cdot s$ of the state set $X$ under some semigroup element $s\in S_O$ and whose permutations are the restrictions of those members of $S_O$ which permute this set. Such an image set can be covered by the union of smaller image sets and singletons, which in turn must also be permuted by these transformations. The permutation group induced on the maximal covering sets of a natural subsystem by these sets is a [*holonomy group*]{}. See [@eilenberg76; @nehaniv2015symmetry; @egri2015computational] for details. In the next section, we will be referencing [*subduction*]{}, a generalized inclusion relation defined on the collection of images together with $X$ and the singletons. For subsets $P,Q \subseteq X$, we say $P$ [*subducts*]{} $Q$ if $P \subseteq Q\cdot s$ for some $s \in S$ or $s$ the identity mapping. Mutual subduction implies isomorphism of holonomy groups, so equivalent locally reversible dynamics in the hierarchical decomposition can be compressed [@egri2015computational], giving insight into complexity of a dynamical system $(X,S)$. In the analysis and diagrams below, subduction corresponds to subset inclusion.
Complexity Regimes {#sec:3}
==================
The investigation of the iterated PD’s $b$-dependence revealed four different regimes characterized by unique sets of transformations by $t$ (Table \[regimetab\]). The complexity of each regime was explored using the Krohn-Rhodes (KR) definition of semigroup complexity [@krohn1968complexity]: KR complexity is formulated such that the complexity of a transformation semigroup $(X,S)$ is equal to the smallest number of non-trivial groups needed for a wreath product decomposition of $(X,S)$. Therefore an upper bound for the KR complexity is the number of levels with non-trivial groups in the holonomy decomposition. For the remainder of this paper, upper bounds will be used when referring to KR complexity.
[p[1.7cm]{}p[3.5cm]{}]{} **Regime** & **Parameter Range**\
A & $b>4.0$\
B & $b=4.0$\
C & $3.0<b<4.0$\
D & $b=3.0$\
Regime A
--------
Beginning with regime A $(b>4.0)$, the system has a temptation to defect so large, that $t^2$ acting on any state containing at least one defector will bring that state to ‘000000’ (state 0), which we will call pure defection. (As $t$ maps the pure cooperation state ‘111111’ (state 63) to itself and no other states map to 63 by words generated by $t$, this state is left out of the subduction chains shown in Figures \[subA\] and \[subB\].) The defection attractor dynamics can be visualized from subduction chain for $(X\setminus\{63\},\langle t \rangle)$ (Figure \[subA\]). We can choose to only examine the mappings induced by words generated by $t$ when comparing regimes since the semigroup generated by $T'_O$ is unchanged by the parameter $b$. As $t^2$ only maps to pure defection and the rest of the mappings in $S_O$ are locally constant maps, there are no pools of reversibility and few levels in the holonomy decomposition, yielding a relatively trivial system.
Regime B
--------
Regime B $(b=4.0)$ can be seen as a critical point where the system changes from regime A to C. The main difference between regimes A and B is that mixed strategy equilibria under transformation $t$ appear in regime B. These equilibria fall under two spatial configurations up to isomorphism: “3-in-a-row’’ and “L-shape’’, shown in Figure \[mixeq\]. Similar to regime A, this regime does not have non-trivial groups in the holonomy decomposition giving both regimes a KR complexity of 0.
--------------------------------------------------- -------------------------------------------------------
\[2mm\] (a) “3-in-a-row” strategy configuration \(b) “L-shape” strategy configuration represented
represented by $\{21,42\}=[21]_{\cong} \subset X$ by $\{23, 29, 43, 46, 53, 58\}=[23]_{\cong}\subset X$
--------------------------------------------------- -------------------------------------------------------
Regime C
--------
In regime C $(3.0<b<4.0)$, the temptation to defect is at an intermediate level which now allows certain states to map to ones of higher cooperation with $t$. From the subduction chain (Figure \[subC\]) one can see that the decreased temptation to defect removes one class of mixed strategy equilibria states, $[21]_{\cong}$. In this state, the defector’s net payoff is $b+2$ as it receives $b$ for playing against one adjacent cooperator as well as 2 for playing against two adjacent defectors. Since $b<4$ in this regime, $b+2<6$, the total payoff for cooperators and this state will now map to pure cooperation when acted on by $t$.
This regime is drastically different from regimes A and B as there are now cyclic groups in the holonomy decomposition. For these intermediate temptations to defect, the system has pools of reversibility where dynamical cycles may recur, unlike the previous regimes where any non-trivial mappings induced by words in $S_O$ would bring the system to a state in which the previous state is inaccessible by the same transformation. This reversibility is entirely $t$-dependent since any words made of locally constant maps which act non-trivially on a state are by definition irreversible. Only when the temptation to defect is low enough such that an action by $t$ can bring the system to a new state of equivalent or higher cooperation will mappings induced by words from $S_O$ form non-trivial groups.
Additionally, the amount and distribution of open cells now play a significant role in the system complexity. In general, KR complexity increases with the number of open cells; yet for a given number of open cells, their distribution plays a significant role (Table \[opencells\]). Note the open cell configuration corresponding to $T'_{123}=\{d_1,d_2,d_3,c_1,c_2,c_3\}$ has its upper bound of KR complexity reduced from 6 to 4. This is because the configuration $T'_{1234}$ has KR complexity 4 and $T'_{123}$ is a subsemigroup of $T'_{1234}$. It follows naturally that $(X,S_{1234})$ emulates $(X,S_{123})$ and from the KR complexity axioms [@krohn1968complexity], for transformation semigroups $(Y,T)$ and $(X,S)$, if $(X,S)$ can emulate $(Y,T)$ then the complexity of $(Y,T)$ must be less than or equal to that of $(X,S)$.
Below are the orbits for a representation of the holonomy group $(3,C_2)$ found in the holonomy decomposition (Figure \[d2c1t\]a). This is one of two $C_2$ permutator groups in regime C with $O=\{1,2\}$. The generator of this permutation group is $d_2c_1t$ which represents mapping cell 2 to defection, cell 1 to cooperation and then letting one time step, $t$ pass. Since the holonomy group $(3,C_2)$ shows the group action on a set of 3 subsets permuted by permutations of 5 underlying states, the exact mechanism leading to this reversibility is not immediately clear. We can gain a better understanding of the dynamics of this cyclic group by examining the orbits of the transformation $d_2c_1t$ on specific states in these sets as shown in the natural subsystem (Figure \[d2c1t\]b).
----------------------------------------- -----------------------------------------------------
\(a) Orbits of holonomy group $(3,C_2)$ \(b) Natural subsystem for transformation $d_2c_1t$
for transformation $d_2c_1t$
----------------------------------------- -----------------------------------------------------
Most of these orbits act in an expected manner since for the two right-most orbits, the locally constant mappings do not change the state and as the states are mixed strategy equilibria for regime C, action by $t$ does not change the state. In the left-most orbit, the behaviour is also expected since $d_2c_1$ effectively turns a pure defection state into one with a single cooperator which will receive the lowest payoff of its defecting neighbours, thus reverting back to defection with $t$.
In the orbit second from the left in Fig. \[d2c1t\]b, the behaviour is much more interesting as the same transformation that removes four cooperators from the system, also leads it into a state of pure cooperation. For state ‘001010’, $d_2c_1$ acts as simply $c_1$ since cell 2 is already a defector. This maps the system to state ‘101010’. As shown above, this state now maps to one of pure cooperation with $t$. At the state of pure cooperation, $d_2c_1$ now acts as $d_2$ mapping the state to ‘101111’.
[>m[26mm]{}>m[9mm]{}>m[26mm]{}>m[9mm]{}>m[26mm]{}>m[9mm]{}]{} & $\xrightarrow{\: d_2c_1 \:}$ & & $\xrightarrow{\quad t \quad}$ &\
001010 & & 101010 & & 111111\
& $\xrightarrow{\: d_2c_1 \:}$ & & $\xrightarrow{\quad t \quad}$ &\
111111 & & 101111 & & 001010\
In this state, the single defector benefits from being surrounded by cooperators receiving the highest net payoff as $b>3$ and with $t$, all of its neighbours switch to the defector strategy resulting in the state ‘001010’ (See Figure \[fig7\]). These pools of reversibility can offer some insight regarding the spatial configuration of strategies which lead to the persistence of cooperation. Additionally, in regime C, we see the only symmetric group $(3,S_3)$ in the holonomy decomposition (Figure \[S3\]). Here the two group generators are given by $c_3td_1c_2tc_5c_2tc_2d_3$ and $c_3tc_2c_5td_1c_2tc_2d_3$ and although these words are long and hard to interpret, the possibility of appearance of such non-abelian group dynamics is not an obvious result of simple iterated PD.
Regime D
--------
In regime D $(b=3.0)$, which can be interpreted as a weak PD, the temptation to defect is very low and consequently the incentive to cooperate is highest. Due to this push towards cooperation, there are less pools of reversibility than in regime C and the highest upper bound for KR complexity is 2. From the subduction chain (Figure \[subD\]), we see that all equilibria are mapped to by a single time step $t$ and a new class of equilibria emerge. (Also present are the “L-shape" equilibria with two defectors in a row we encountered above.) This new class represents a single defector, which in all previous regimes had been beneficial to the lone defector. In this regime, the temptation to defect is low enough that the system has become immune to invasion by a single defector.
Conclusion
==========
Representing the iterated PD as a transformation semigroup allows the holonomy decomposition to reveal qualitative differences between distinct payoff-dependent regimes. When the temptation to defect is below a threshold, the KR complexity becomes non-zero and pools of reversibility form. The number of open cells also positively influences the KR complexity, however their spatial distribution plays an equally important role. With greater computational power, it would be interesting to further explore this system with a larger number of players as well as different topologies to see how the results presented in this paper compare to larger and more complex spatial configurations. With this information, one could explore how the KR complexity varies with both spatial size and configuration, as well as with the temptation to defect. Additionally, it could lead to insights resulting in algebraic theorems for more general iterated PD systems.
All code used to generate semigroup mappings, analysis and figures is available at [](https://gist.github.com/cello-kabob/eb0a56bbcd598bf613d91b8e773ff9cf)
[^1]: University of Waterloo, 200 University Ave W, Waterloo, ON N2L 3G1, Canada email: [email protected]
[^2]: Algebraic Intelligence & Computation Laboratory University of Waterloo, 200 University Ave W, Waterloo, ON N2L 3G1, Canada email: [email protected]
[^3]: This work was supported by the Natural Sciences and Engineering Research Council of Canada (NSERC), funding ref. RGPIN-2019-04669, and the University of Waterloo
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this article, we study the geometric invariant theory (GIT) compactification of quintic threefolds. We study singularities, which arise in non-stable quintic threefolds, thus giving a partial description of the stable locus. We also give an explicit description of the boundary components and stratification of the GIT compactification.'
address: |
Department of Mathematics & Computer Science\
High Point University\
High Point, NC, 27262\
USA
author:
- Chirag Lakhani
bibliography:
- 'bibliography.bib'
title: The GIT Compactification of Quintic Threefolds
---
Introduction {#sec:intro}
============
Quintic threefolds are some of the simplest examples of Calabi-Yau varieties. Physicists have given Calabi-Yau varieties a great deal of attention, in the last 30 years, because they give the right geometric conditions for some superstring compactifications [@wittenstrominger]. In mirror symmetry, in particular, the Kahler moduli space and complex structure moduli space of Calabi-Yau varieties are important objects of study. The purpose of this paper is to describe the complex structure moduli space of quintic threefolds using geometric invariant theory (GIT).
GIT constructions of moduli spaces of projective varieties are automatically projective, therefore have a natural compactification. The Hilbert-Mumford criterion provides a numerical tool, which is useful when constructing moduli spaces using GIT. Despite having this tool, it is still difficult to construct moduli spaces in dimension 2 or higher. There are a some cases where such moduli spaces have been constructed, such as degree 2 and degree 4 $K3$ surfaces by Shah [@shahdeg2; @shahdeg4], cubic threefolds by Allcock and Yokoyama [@allcock; @yokoyama], and cubic fourfolds by Laza [@laza].
A quintic threefold is the zero set of a homogeneous degree $5$ form $f \in {\mathbb{C}[x_0,x_1,x_2,x_3,x_4]}$. The space $H^0({\mathbb{P}}^4,{\mathcal{O}}_{{\mathbb{P}}^4}(1)) \cong {\mathbb{C}}^5$ represents the set of degree $1$ forms in ${\mathbb{P}}^4$. The parameter space of quintic threefolds is then represented by $V:={\mathbb{P}}(Sym^5({\mathbb{C}}^5))$, which is the projectivization of the space of coefficients of quintic forms $f$. ${\mathbb{C}}^5$ and $Sym^5({\mathbb{C}}^5)$ both have natural $SL(5,{\mathbb{C}})$-actions. Two threefolds are equivalent if one form can be transformed into another by an $SL(5,{\mathbb{C}})$-action. In order to construct the moduli space using GIT, the stable and semistable quintic threefolds must be identified. A quintic threefold $X$ is semistable if there is a $SL(5,{\mathbb{C}})$-invariant function on ${\mathbb{P}}(V)$ where $X$ does not vanish. A semistable quintic threefold $X$ is stable if its $SL(5,{\mathbb{C}})$-orbit, in the space of semistable quintic threefolds, is closed and the isotropy group of $X$ is finite. The space of semistable quintic threefolds is denoted ${\mathbb{P}}(V)^{ss}$ and the space of stable quintic threefolds is denoted ${\mathbb{P}}(V)^{s}$. The $SL(5,{\mathbb{C}})$-orbits of threefolds in ${\mathbb{P}}(V)^{s}$ are closed, so the quotient ${\mathbb{P}}(V)^{s} {/\!/}SL(5,{\mathbb{C}})$ forms an orbit space. The addition of semistable quintic threefolds compactifies the moduli space by making it a projective variety i.e. ${\mathbb{P}}(V)^{ss} {/\!/}SL(5,{\mathbb{C}})$ is projective. Two semistable quintic threefolds $X,Y \in {\mathbb{P}}(V)^{ss} \setminus {\mathbb{P}}(V)^{s}$ map to the same point in ${\mathbb{P}}(V)^{ss} {/\!/}SL(5,{\mathbb{C}})$ if their closures satisfy $$\overline{SL(5,{\mathbb{C}})\cdot X} \cap \overline{SL(5,{\mathbb{C}})\cdot Y} \neq \emptyset.
\label{eq:orbitclosure}$$
This establishes an orbit-closure relationship for semistable quintic threefolds where $X \sim Y$ if they satisfy the property \[eq:orbitclosure\]. Furthermore, all threefolds in the same orbit-closure equivalence class map to the same point in ${\mathbb{P}}(V)^{ss} {/\!/}SL(5,{\mathbb{C}})$. Every orbit-closure equivalence class in ${\mathbb{P}}(V)^{ss} \setminus {\mathbb{P}}(V)^{s}$ has a unique closed orbit representative called the *minimal orbit*. The boundary components of ${\mathbb{P}}(V)^{ss} {/\!/}SL(5,{\mathbb{C}}) \setminus {\mathbb{P}}(V)^{s} {/\!/}SL(5,{\mathbb{C}})$ are represented precisely by these minimal orbits.
We will follow the terminology in GIT [@mumford]. Unstable will mean not semistable, non-stable will mean not stable, and strictly semistable will mean semistable but not stable.
The main results of this paper are a description of the non-stable quintic threefolds in terms of singularities, partial description of the stable locus, and a complete description of boundary components and stratification of the GIT compactification using minimal orbits. The first main result of the paper is given in section \[sec:geointermaxsemi\] which describes the non-stable quintic threefolds in terms of singularities.
\[theorem:non-stable\] A quintic threefold $X$ is non-stable if and only if one of the following properties holds:
1. $X$ contains a double plane;
2. $X$ is a reducible variety, where a hyperplane is one of the components;
3. $X$ contains a triple line;
4. $X$ contains a quadruple point;
5. $X$ contains a triple point $p$ with the following properties:
1. the tangent cone of $p$ is the union of a double plane and another hyperplane;
2. the line connecting a point in the double plane with the triple point has intersection multiplicity 5 with $X$;
6. $X$ has a double line $L$ where every point $p \in L$ has the following properties:
1. the tangent cone of each point $p \in L$ is a double plane $P_p$;
2. each point $p \in L$ has the same double plane tangent cone i.e. $P_p=P$ for some double plane $P$;
3. the line connecting the point on the tangent cone $P_p$ and a point $p \in L$ has intersection multiplicity 4 with $X$;
7. $X$ contains a triple point $p$ and a plane $P$ with the following properties:
1. the tangent cone of $p$ contains a triple plane of $P$;
2. the singular locus of $X$, when restricted to $P$, is the intersection of two quartic curves $q_1$ and $q_2$;
3. the point $p$ is a quadruple point of $q_1$ and $q_2$.
A partial description of the stable locus is given in section \[sec:stablelocus\]. The analysis of singularities of the non-stable quintic threefolds gives a partial description of the singularities that occur in the stable locus. In particular, all smooth quintic threefolds and quintic threefold with at worst $A \hyph D \hyph E$ singularities will be GIT stable. Following the approach of Laza [@laza], the minimal orbits can be explicitly described using Luna’s criterion[@luna; @vinberg]. In order to find the minimal orbits, the non-stable quintic threefolds degenerate into families of quintic threefolds given by equations \[eq:MO-A\]-\[eq:MO-D\], which are denoted the *first level of minimal orbits*. Luna’s criterion determines which members of these families represent closed orbits, this is done in section \[sec:firstlevelminimalorbits\]. Certain hypersurfaces in these families are unstable and therefore represent unstable quintic threefolds or they degenerate, even further, into a family of quintic threefolds given by equations \[eq:MO2-I\]-\[eq:MO2-X\]. The families represented by equations \[eq:MO2-I\]-\[eq:MO2-X\] are called the *second level of minimal orbits*. The second level of minimal orbits represent how the boundary strata of the components in equations \[eq:MO-A\]-\[eq:MO-D\] intersect. Applying Luna’s criterion to the second level of minimal orbits will determine which quintic threefolds are closed orbits, unstable orbits, and which hypersurfaces degenerate even further. It will be shown in section \[sec:secondlevelminimalorbits\] that non-closed orbits in the second level of minimal orbits will eventually degenerate to the hypersurface $x_0x_1x_2x_3x_4$, which is the hypersurface with normal crossings singularities. This completely determines the boundary structure and stratification of the GIT compactification of the moduli space of quintic threefolds.
Section \[sec:maxnon-stable\] is devoted to the combinatorics and geometrical study of non-stable quintic threefolds. Using the Hilbert-Mumford criterion, a combinatorial study of the monomials to be included in maximal non-stable families of quintic threefolds will be done. Each maximal non-stable family has a destabilizing one-parameter subgroup (1-PS) $\lambda$ which gives rise to a “bad flag” that picks out the worst singularities in the family. This is used to prove Theorem \[theorem:non-stable\]. The last part of section \[sec:maxnon-stable\] gives a partial description of the stable locus. Section \[sec:minimalorbit\] introduces Luna’s criterion and then applies it to study the closed orbits in equations \[eq:MO-A\]-\[eq:MO-D\]. Section \[sec:boundary\] studies the closed orbits and further degenerations of equations \[eq:MO2-I\]-\[eq:MO2-X\], thereby giving a complete description of the boundary stratification of the GIT compactification of quintic threefolds.
Acknowledgements {#acknowledgements .unnumbered}
----------------
The author would like to Amassa Fauntleroy and Radu Laza for their advice and support in writing this paper. The author would also like to thank Ryan Therkelsen for introducing the author to Stembridge’s poset Maple package.
Maximal Nonstable Families {#sec:maxnon-stable}
==========================
The Hilbert-Mumford criterion is an important tool when establishing stability and semistability in GIT.
Following the convention in [@laza], a *normalized 1-PS $\lambda$* is a map $\lambda: {\mathbb{C}}^* \to T \subseteq SL(5,{\mathbb{C}})$, where $T$ is the the standard maximal torus $T$ of $SL(5,{\mathbb{C}})$, with the additional property that $\lambda(t) = diag(t^{a_0},t^{a_1},t^{a_2},t^{a_3},t^{a_4})$ satisfy $a_0 \geq a_1 \geq a_2 \geq a_3 \geq a_4$ and $a_0+a_1+a_2+a_3+a_4=0$.
For a quintic form $f \in {\mathbb{C}[x_0,x_1,x_2,x_3,x_4]}$ and a normalized 1-PS $\lambda = \langle a_0,a_1,a_2,a_3,a_4 \rangle$, the numerical function is defined as follows, $$\mu(f,\lambda):=max \{ a_0 i_0+a_1i_1+a_2i_2+a_3 i_3+a_4i_4 | f=\sum c_{i_0i_1i_2i_3i_4}x_0^{i_0}x_1^{i_1}x_2^{i_2}x_3^{i_3}x_4^{i_4}, c_{i_0i_1i_2i_3i_4} \neq 0 \}.$$
The Hilbert-Mumford criterion states *a quintic form f is stable (semistable) if and only if for every 1-PS $\lambda$ the numerical function $\mu(f,\lambda) > 0 ( \geq 0)$*. It can be restated so that a quintic form $f$ is non-stable (unstable) if there exists a 1-PS $\lambda$ where $\mu(f,\lambda) \leq 0$ $( < 0)$. If quintic forms are analyzed up to coordinate transformation ($SL(5,{\mathbb{C}})$-action), then the $G$-equivariance of the numerical function[@standardmonomial], $$\label{eq:gequivariant}
\mu(x,\lambda) = \mu(gx, g \lambda g^{-1}),$$ restricts to check checking the criterion for only normalized 1-PS $\lambda$ in the standard maximal torus $T$ of $SL(5,{\mathbb{C}})$.
In the remainder of the paper the a monomial $x_0^{i_0}x_1^{i_1}x_2^{i_2}x_3^{i_3}x_4^{i_4}$ will also be written as a vector denoted $[i_0,i_1,i_2,i_3,i_4]$. Also, the following combinatorial procedure described below is based on similar combinatorial techniques described in Mukai [@mukai] and applied by Laza [@laza] in the case of cubic fourfolds.
The normalized 1-PS $\lambda$ induce a partial order on the set of quintic monomials $[i_0,i_1,i_2,i_3,i_4]$ given by $$[i_0,i_1,i_2,i_3,i_4] \geq [j_0,j_1,j_2,j_3,j_4] \Longleftrightarrow
\begin{cases}
\mu([i_0,i_1,i_2,i_3,i_4],\lambda) \geq \mu([j_0,j_1,j_2,j_3,j_4],\lambda) \\
\mbox{for all normalized 1-PS } \lambda
\end{cases}$$ The following lemma is useful in creating an algorithm to determine the poset structure of quintic monomials.
\[c.f. [@mukai] p.225\] For two monomials $[i_0,i_1,i_2,i_3,i_4]$ and $[j_0,j_1,j_2,j_3,j_4]$ $$[i_0,i_1,i_2,i_3,i_4] \geq [j_0,j_1,j_2,j_3,j_4] \Longleftrightarrow
\begin{cases}
i_0 \geq j_0 \\
i_0+i_1 \geq j_0+j_1 \\
i_0+i_1+i_2 \geq j_0+j_1+j_2 \\
i_0+i_1+i_2+i_3 \geq j_0+j_1+j_2+j_3
\end{cases}$$
This criterion is useful because one can directly check whether two monomials are related in the poset by checking the subsequent inequalities. Using Maple[@maple], the above criterion can be used to find all partial order relationships between monomials. Stembridge’s poset package for Maple [@stembridgemaple] is used to find the minimal covering relationships for these monomials and thereby creating the poset for quintic monomials. The code for this entire procedure is given in appendix \[appendix:posetcode\]. The figure for this poset structure is given in figure \[fig:calabiyau\] at the end of the paper.
Combinatorics of Non-Stable Families
------------------------------------
The poset structure on quintic monomials greatly simplifies the Hilbert-Mumford criterion analysis on quintic polynomials $f$. For a fixed normalized 1-PS $\lambda$ and a monomial $[i_0,i_1,i_2,i_3,i_4]$, $\mu([i_0,i_1,i_2,i_3,i_4], \lambda)$ does not change when any monomials below it in the poset are added to $[i_0,i_1,i_2,i_3,i_4]$. For a fixed normalized 1-PS $\lambda$, ${M_{\leq 0}(\lambda)}$ (${M_{<0}(\lambda)}$) represent the set of monomials in the poset where $\mu \leq 0$ ($\mu <0$). By the Hilbert-Mumford criterion, the monomials of every non-stable quintic polynomial $f$, up to coordinate transformation, belong to a family of the form ${M_{\leq 0}(\lambda)}$. The *maximal non-stable families*, denoted $SSk$, are the largest possible families of the form ${M_{\leq 0}(\lambda)}$. The corresponding $\lambda$ of ${M_{\leq 0}(\lambda)}$ is called the family’s *destabilizing 1-PS*. From the poset structure, there will be a finite number of maximal non-stable families $SSk$. Any non-stable quintic form $f$, up to coordinate transformation; will be long to one of these families.
The procedure for determining the set of maximal non-stable families would be to start from the top monomial and work down the poset until one finds a monomial $[i_0,i_1,i_2,i_3,i_4]$ which has a normalized 1-PS $\lambda$ where $\mu([i_0,i_1,i_2,i_3,i_4], \lambda) \leq 0$. By restricting to normalized $\lambda$ one can use linear programming to determine whether such a $\lambda$ exists for a particular monomial. The linear programming script is given in appendix \[appendix:linearprogram\]. Using this procedure it is determined that the top most monomials which have $\mu \leq 0$ are \[3,0,0,2,0\], \[4,0,0,0,1\], \[2,0,3,0,0\], and \[1,4,0,0,0\], the poset of monomials below these top monomials are denoted SS1, SS2, SS3, and SS4. The figures for these posets are given in figures \[fig:SS1\], \[fig:SS2\], \[fig:SS3\], and \[fig:SS4\] at the end of the paper. The destabilizing 1-PS $\lambda$ is given in table 1. Other maximal non-stable families are found by finding the top-most monomials in the families $SS1 \hyph SS4$ which have a common destabalizing 1-PS $\lambda$. There are three other such families denoted $SS5 \hyph SS7$. These three families have multiple maximal monomials. The figures for these posets are given in figures \[fig:SS5a\], \[fig:SS5b\], \[fig:SS6a\], \[fig:SS6b\], \[fig:SS7a\] and \[fig:SS7b\] at the end of the paper.
The combinatorial procedure above determines the maximal families ${M_{\leq 0}(\lambda)}$ where $\mu \leq 0$. By the Hilbert-Mumford criterion, every semistable or unstable hypersurface will transform, via a coordinate transformation, into one of maximal non-stable families ${M_{\leq 0}(\lambda)}$.
$q_{a,b}(x_m,x_n \parallel x_o,x_p)$ represents polynomials which are a linear combination of degree $a$ monomials in $x_m$ and $x_n$ multiplied by a degree $b$ monomials in $x_o$ and $x_p$
$X$ is non-stable if and only if it belongs, via a coordinate transformation, to a hypersurface in families found in table 1.
For any non-stable $f$, the $SL(5,{\mathbb{C}})$-orbit will not necessarily be closed. Using the destabilizing 1-PS $\lambda$, the closure $\overline{\lambda f}$ = $f_0$ is a quintic form invariant with respect to $\lambda$. The forms $f$ and $f_0$ will map to the same point in the GIT quotient. If the orbit of $f_0$ is closed then it is a minimal orbit. $$\begin{aligned}
\label{eq:MO-A}
MO \hyph A:& q_{2,3}(x_0,x_1 \parallel x_2 , x_3, x_4), \\
\label{eq:MO-B}
MO \hyph B:& q_5(x_1,x_2,x_3)+x_0x_4q_3(x_1,x_2,x_3)+x_0^2x_4^2q_1(x_1,x_2,x_3), \\
\label{eq:MO-C}
MO \hyph C:& q_{1,4}(x_0,x_1 \parallel x_2,x_3) + x_4q_{2,2}(x_0,x_1 \parallel x_2,x_3) + x_4^2q_3(x_0,x_1), \\
\label{eq:MO-D}
MO \hyph D:& x_0q_4(x_1,x_2,x_3,x_4).\end{aligned}$$
In Section \[sec:minimalorbit\], it will be shown that a generic member of one of the families $MO \hyph A - MO \hyph D$ will represent a minimal orbit. Certain hypersurfaces in $MO \hyph A - MO \hyph D$ will degenerate further into a member of one of the families $MO2 \hyph I - MO2 \hyph X$.
$$\begin{aligned}
\label{eq:MO2-I}
MO2 \hyph I:& x_0^2 x_2x_4^2 + x_0x_1 x_2x_3x_4 + x_1^2 x_2x_3^2 , \\
\label{eq:MO2-II}
MO2 \hyph II:& x_0^2 x_3x_4^2 + x_0x_1 ( x_3^3 + x_2x_3x_4 ) + x_1^2 x_2^2x_3 , \\
\label{eq:MO2-III}
MO2 \hyph III:& x_0^2 ( x_2x_4^2 + x_3^2x_4 ) + x_0x_1 ( x_3^3 + x_2x_3x_4 ) + x_1^2 ( x_2x_3^2 + x_2^2x_4 ), \\
\label{eq:MO2-IV}
MO2 \hyph IV:& x_0x_1x_4 q_2(x_2,x_3) , \\
\label{eq:MO2-V}
MO2 \hyph V:& x_0x_1 x_2x_3x_4 , \\
\label{eq:MO2-VI}
MO2 \hyph VI:& x_1x_2^3x_3 + x_1^2x_2x_3^2 + x_0 x_1 x_2x_3x_4 , \\
\label{eq:MO2-VII}
MO2 \hyph VII:& x_4x_0^2x_3^2 + x_4x_0x_1x_2x_3 + x_4x_1^2x_2^2 , \\
\label{eq:MO2-VIII}
MO2 \hyph VIII:& x_0x_1q_3(x_2,x_3,x_4), \\
\label{eq:MO2-IX}
MO2 \hyph IX:& x_0q_{2,2}(x_1,x_2 \parallel x_3,x_4) , \\
\label{eq:MO2-X}
MO2 \hyph X:& x_0 (q_4(x_2,x_3) + x_1x_4q_2(x_2,x_3) + x_1^2x_4^2 ).
\end{aligned}$$
Bad Flags
---------
The maximal non-stable families will be characterized in terms of singularities found on a generic member of one of these families. A destabilizing 1-PS $\lambda$ has an associated “bad flag” of the vector spaces $H^0({\mathbb{P}}^4,{\mathcal{O}}_{{\mathbb{P}}^4}(1)) \cong {\mathbb{C}}^5$. A general principal, given by Mumford [@mumford], states that these “bad flags” pick out the singularities which cause the family to become semistable or unstable.
Using the approach given by Laza [@laza] it can be shown that a 1-PS $\lambda: {\mathbb{C}}^* \to T$ gives a weight decomposition of $H^0({\mathbb{P}}^4,{\mathcal{O}}_{{\mathbb{P}}^4}(1)) = \oplus_{i=0}^{5} W_i$ based on the eigenvalues of $\lambda$ acting on $H^0({\mathbb{P}}^4,{\mathcal{O}}_{{\mathbb{P}}^4}(1))$.
For a 1-PS $\lambda = \langle a,b,c,d,e \rangle$ let $m_i$ be a subset of $\left\{ a,b,c,d,e \right\}$ which have the same weights and let $n_i$ be the weight. $$W_{m_i} := \bigoplus_{\text{$i$ where $W_i$ has eigenvalue $n_i$}} W_i$$
The standard flag is
$$\label{eq:standardflagvariety} \begin{split} \emptyset \subseteq F_1 = ( x_1=x_2&=x_3=x_4=0 ) \subseteq F_2 = ( x_2=x_3=x_4=0 ) \subseteq \\ & F_3 = ( x_3=x_4=0 ) \subseteq F_4 = (x_4=0) \subseteq {\mathbb{P}}^4. \end{split}$$
Given a 1-PS $\lambda = \langle a,b,c,d,e \rangle$ let $m_1$, $m_2 \ldots$, $m_s$ represent the collection of common weights of $\lambda$. Let $m_i$ be ordered by increasing value of weights (i.e. $m_1$ has lowest weight). The *associated flag for $\lambda$* is
$$\label{eq:1psflagvariety} F_{\lambda}:\ \ \emptyset \subseteq F_{m_{s}}:= \bigoplus_{i=1}^{s} W_{m_i} \subset F_{m_{s-1}}:= \bigoplus_{i=1}^{s-1} W_{m_i} \subset \ldots \subseteq F_{m_{1}}:= W_{m_1} \subseteq {\mathbb{P}}^4.$$
This is a subflag of the standard flag (\[eq:1psflagvariety\]).
For the maximal destabilizing families $SS1 \hyph SS7$ the associated “bad flags” are
Geometric Interpretation of Maximal Semistable Families {#sec:geointermaxsemi}
-------------------------------------------------------
In order to determine the singularities which occur on threefolds in families $SS1 \hyph SS7$, we intersect the general form of the equation with its associated destabilizing flag. This will give some description of the types of singularities, which occur in each family. A precise description of each such family is given in the propositions below. Some of the singularity analysis is based on describing the tangent cone and intersection multiplicities of the tangent cone at singular points, a detailed introduction of these topics is given in Beltrametti et al. ([@beltrametti] ch.5)
A hypersurface $X$ is of type SS1 if and only if $X$ contains a double plane.
Let $X$ be of type $S1$ then it is equivalent, via a coordinate transformation, to the hypersurface $$\begin{split}
q_{3,2}(x_0,x_1,x_2 \parallel x_3,x_4) + q_{2.3}(x_0,x_1,x_2 \parallel x_3,x_4) \\ + q_{1,4}(x_0,x_1,x_2 \parallel x_3,x_4)+q_5(x_3,x_4).
\end{split}
\label{SS1}$$ This hypersurface contains the ideal $\langle x_3,x_4 \rangle^2$ which is a double plane in ${\mathbb{P}}^4$.
Let $X$ be a hypersurface which contains a double plane. By a coordinate transformation we can assume the double plane is $\langle x_3,x_4 \rangle^2$. The most general equation which contains the ideal $\langle x_3,x_4 \rangle^2$ is (\[SS1\]).
A hypersurface $X$ is of type SS2 if and only if $X$ is a reducible variety, where a hyperplane is one of the components. In particular, the singularity is the intersection of the hyperplane with the other component, which is generically a degree 4 surface.
Let $X$ be of type $S2$ then it is equivalent, via a coordinate transformation, to the hypersurface
$$\begin{split}
x_4q_4(x_0,x_1,x_2,x_3,x_4)
\end{split}
\label{SS2}$$
This hypersurface has the hyperplane $\langle x_4 \rangle$ as a component.
Let $X$ be a reducible hypersurface where a hyperplane is a component. The polynomial $f \in {\mathbb{C}}[x_0,x_1,x_2,x_3,x_4]$ defining $X$ can be factored into $f = gh$, where $h$ is a degree 1 polynomial. By a coordinate transformation we can map the hyperplane defining $h$ to $x_4$. Without loss of generality $f=x_4h$. Since since $f$ is of degree 5 then by neccesity $h$ is of degree 4 therefore $f$ is of the form (\[SS2\]).
A hypersurface $X$ is of type SS3 if and only if $X$ contains a triple line.
Let $X$ be of type $SS3$ then it is equivalent, via a coordinate transformation, to the hypersurface $$\begin{split}
q_{2,3}(x_0,x_1 \parallel x_2,x_3,x_4)+q_{1,4}(x_0,x_1 \parallel x_2,x_3,x_4)+q_5(x_2,x_3,x_4)
\end{split}
\label{SS3}$$ This hypersurface contains the ideal $\langle x_2,x_3,x_4 \rangle^3$ which is a triple line in ${\mathbb{P}}^4$.
Let $X$ be a hypersurface which contains a triple line. By a coordinate transformation, we can assume the triple line is $\langle x_2,x_3,x_4 \rangle^3$. The most general equation which contains the ideal $\langle x_2,x_3,x_4 \rangle^3$ is (\[SS3\]).
A hypersurface $X$ is of type $SS4$ if and only if $X$ contains a quadruple point.
Let $X$ be of type $SS4$ then it is equivalent, via a coordinate transformation, to the hypersurface $$\begin{split}
x_0q_4(x_1,x_2,x_3,x_4)+q_5(x_1,x_2,x_3,x_4)
\end{split}
\label{SS4}$$ This hypersurface contains the ideal $\langle x_1,x_2,x_3,x_4 \rangle^4$ which is a quadruple point in ${\mathbb{P}}^4$.
Let $X$ be a hypersurface which contains a quadruple point. By a coordinate transformation we can assume the quadruple point is $\langle x_1,x_2,x_3,x_4 \rangle^4$. The most general equation which contains the ideal $\langle x_1,x_2,x_3,x_4 \rangle^4$ is (\[SS4\]).
A hypersurface $X$ is of type $SS5$ if and only if $X$ has a triple point $p$ with the following properties:
- the tangent cone of $p$ is the union of a double plane and another hyperplane;
- the line connecting a point in the double plane with the triple point has intersection multiplicty 5 with the hypersurface.
Let $X$ be of type $SS5$ then it is equivalent, via a coordinate transformation, to the hypersurface $$\begin{split}
x_0^2 \bigg( x_4^2q_1(x_1,x_2,x_3,x_4) \bigg) + x_0x_4q_3(x_1,x_2,x_3,x_4) + q_5(x_1,x_2,x_3,x_4)
\end{split}
\label{SS5}$$ This hypersurface contains the triple point $\langle x_1,x_2,x_3,x_4 \rangle^3$. The tangent cone is the hypersurface defined by
$$\begin{split}
x_4^2q_1(x_1,x_2,x_3,x_4)
\end{split}$$
which is the union of a double hyperplane $\langle x_4 \rangle^2$ and another general hyperplane $q_1(x_1,x_2,x_3,x_4)$. The points whose lines passing through the triple point which have intersection multiplicity 5 with the hypersurface, is the locus of $\langle x_4^2q_1(x_1,x_2,x_3,x_4) \rangle$ and $\langle x_4q_3(x_1,x_2,x_3,x_4) \rangle$. Since $x_4$ is a component of both terms then a line emanating from the hyperplane $\langle x_4 \rangle$ to the triple point will have multiplicity 5.
Let $X$ be a hypersurface which contains a triple point. By a coordinate transformation we can assume the triple point is $\langle x_1,x_2,x_3,x_4 \rangle^3$. The most general equation which contains the ideal $\langle x_1,x_2,x_3,x_4 \rangle^3$ is
$$\begin{split}
x_0^2\bigg( q_3(x_1,x_2,x_3,x_4) \bigg) + x_0 \bigg( q_4(x_1,x_2,x_3,x_4) \bigg) + q_5(x_1,x_2,x_3,x_4).
\end{split}$$
If the tangent cone is the union of a double plane and another hyperplane then
$$\begin{split}
q_3(x_1,x_2,x_3,x_4)=f^2g
\end{split}$$
where $f$ and $g$ are linear forms. By a coordinate transformation which keeps the triple point fixed we can map the hyperplane $f$ to $x_4$. So without loss of generality
$$\begin{split}
q_3(x_1,x_2,x_3,x_4)=x_4^2g.
\end{split}$$
If a general line from the hyperplane $\langle x_4 \rangle$ to the triple point has multiplicity 5 then
$$\begin{split}
x_4=q_4=0.
\end{split}$$
This occurs only if $q_4$ has $x_4$ as a component so
$$\begin{split}
q_4=x_4q_3
\end{split}$$
which is precisely of the form (\[SS5\]).
A hypersurface $X$ is of type $SS6$ if and only if $X$ has a double line $L$ where every point $p \in L$ has the following properties:
- the tangent cone of each point $p \in L$ is a double plane $P_p$;
- each point $p \in L$ has the same double plane tangent cone i.e. $P_p=P$ for some double plane $P$;
- the line connecting the point on the tangent cone $P_p$ and a point $p \in L$ has intersection multiplicity 4 with the hypersurface.
Let $X$ be of type $SS6$ then it is equivalent, via a coordinate transformation, to the hypersurface $$\begin{split}
\bigg( q_{3} (x_0,x_1)x_4^2 \bigg) + \bigg( x_4 q_{2,2}(x_0,x_1 \parallel x_2,x_3,x_4) \bigg) \\ + q_{1,4}(x_0,x_1 \parallel x_2,x_3,x_4) + q_5(x_2,x_3,x_4)
\end{split}
\label{SS6}$$ This hypersurface contains the double line $\langle x_2,x_3,x_4 \rangle^2$. For any point $[\lambda: \nu: 0:0:0]$ of the double line the tangent cone is the same double plane given by $\langle x_4 \rangle^2$. The points which have intersection multiplicity 4 with the double line are the locus of $\langle x_4 \rangle^2 $ and $\langle x_4 q_{2,2}(\lambda, \nu \parallel x_2,x_3,x_4) \rangle$. Since $x_4$ is a component of both terms then the line emanating from the hyperplane $\langle x_4 \rangle$ to any point of the double line will have multiplicity 4.
Let $X$ be a hypersurface which contains a double line. By a coordinate transformation we can assume the double line is $\langle x_2,x_3,x_4 \rangle^2$. The most general equation which contains the ideal $\langle x_2,x_3,x_4 \rangle^2$ is
$$\begin{split}
q_{3,2}(x_0,x_1 \parallel x_2,x_3,x_4) + q_{2,3}(x_0,x_1 \parallel x_2,x_3,x_4) + q_{1,4}(x_0,x_1 \parallel x_2,x_3,x_4) + q_5(x_2,x_3,x_4).
\end{split}$$
If the tangent cone at every point on the double line is the same double plane then
$$\begin{split}
q_{3,2}(x_0,x_1 \parallel x_2,x_3,x_4) = q_3(x_0,x_1)f(x_2,x_3,x_4)^2
\end{split}$$
where $f$ is a linear form. By a coordinate transformation, which keeps the double line fixed, the hyperplane $f$ is mapped to $x_4^2$. So without loss of generality,
$$\begin{split}
q_{3,2}(x_0,x_1 \parallel x_2,x_3,x_4)=q_{3} (x_0,x_1)x_4^2.
\end{split}$$
If the line going from the hyperplane $\langle x_4 \rangle$ to any point of the double line has multiplicity 4 then
$$\begin{split}
x_4=q_{2,3}(\lambda,\nu \parallel x_2,x_3,x_4)=0.
\end{split}$$
This occurs only if $q_{2,3}(x_0,x_1 \parallel x_2,x_3,x_4)$ has $x_4$ as a component so
$$\begin{split}
q_{2,3}=x_4q_{2,2}(x_0,x_1 \parallel x_2,x_3,x_4)
\end{split}$$
which is precisely of the form (\[SS6\]).
A hypersurface $X$ is of type $SS7$ if and only if $X$ contains a triple point $p$ and a plane $P$, where $p \in P$ has the following properties:
- the tangent cone of $p$ contains a triple plane of $P$;
- the singular locus of $X$, when restricted to $P$, is the intersection of two quartic curves $q_1$ and $q_2$;
- the point $p$ is a quadruple point of $q_1$ and $q_2$.
Let $X$ be of type $SS7$ then it is equivalent, via a coordinate transformation, to the hypersurface $$\begin{split}
&x^{2}_0q_3(x_3,x_4) +x_0 \bigg( q_{2,2}(x_1,x_2 \parallel x_3,x_4)+q_{1,3}(x_1,x_2 \parallel x_3,x_4)+q_4(x_3,x_4) \bigg) \\ + &\bigg( q_{4,1}(x_1,x_2 \parallel x_3,x_4)+q_{3,2}(x_1,x_2 \parallel x_3,x_4) +q_{2,3}(x_1,x_2 \parallel x_3,x_4)\\ + &q_{1,4}(x_1,x_2 \parallel x_3,x_4) +q_5(x_3,x_4) \bigg)
\end{split}
\label{SS7}$$ This hypersurface contains the triple point $p$ given by the ideal $\langle x_1,x_2,x_3,x_4 \rangle^3$ and a plane $P$ given by $\langle x_3,x_4 \rangle$. The tangent cone is the hypersurface defined by $q_3(x_3,x_4)$ which which contains the triple plane $\langle x_3,x_4 \rangle^3$ of $P$. When the differential of $X$ is restricted to the plane $\langle x_3,x_4 \rangle$ the only non-trivial contribution comes from the term
$$\begin{split}
q_{4,1}(x_1,x_2 \parallel x_3,x_4) = q_4(x_1,x_2)x_3+ \tilde{q_4}(x_1,x_2)x_4.
\end{split}$$
The differential, when restricted to the plane, is zero when
$$\begin{split}
q_4(x_1,x_2)=\tilde{q_4}(x_1,x_2)=0.
\end{split}$$
Therefore, the plane contains two quartic curves $q_4(x_1,x_2)$ and $\tilde{q_4}(x_1,x_2)$ which contain $p$ as the quadruple point.
Let $X$ be a hypersurface which contains a triple point $p$ and a plane $P$, where $p \in P$. By a coordinate transformation we can assume the triple point is $\langle x_1,x_2,x_3,x_4 \rangle^3$ and the plane is $\langle x_3,x_4 \rangle$. The most general equation which contains the ideal $\langle x_1,x_2,x_3,x_4 \rangle^3$ and $\langle x_3,x_4 \rangle$ is
$$\begin{split}
x^{2}_0 & \bigg( q_{2,1}(x_1,x_2 \parallel x_3,x_4) +q_{1,2}(x_1,x_2 \parallel x_3,x_4)+ q_3(x_3,x_4) \bigg) \\ + & x_0 \bigg( q_{3,1}(x_1,x_2 \parallel x_3,x_4) + q_{2,2}(x_1,x_2 \parallel x_3,x_4)+q_{1,3}(x_1,x_2 \parallel x_3,x_4)+q_4(x_3,x_4) \bigg) \\ + & \bigg( q_{4,1}(x_1,x_2 \parallel x_3,x_4)+q_{3,2}(x_1,x_2 \parallel x_3,x_4) +q_{2,3}(x_1,x_2 \parallel x_3,x_4)\\ +&q_{1,4}(x_1,x_2 \parallel x_3,x_4)+q_5(x_3,x_4) \bigg)
\end{split}
\label{SS7converse}$$
If the tangent cone contains the triple plane of $P$ then it contains the ideal $\langle x_3,x_4 \rangle^3$. Then the coefficients of the $x_0^2$ term of (\[SS7converse\]) contains only the $q_3(x_3,x_4)$ term. The differential of (\[SS7converse\]), when restricted to the plane $\langle x_3,x_4 \rangle$, contains the equations of the form $x_0q_3(x_1,x_2)+q_4(x_1,x_2)$ and $x_0\tilde{q_3}(x_1,x_2)+\tilde{q_4}(x_1,x_2)$. If the singular locus of $X$ in the plane is the intersection of two quartic curves then
$$\begin{split}
x_0q_3(x_1,x_2)+q_4(x_1,x_2) = x_0 \tilde{q_3}(x_1,x_2)+ \tilde{q_4}(x_1,x_2)=0.
\end{split}$$
So $x_0q_3(x_1,x_2)+q_4(x_1,x_2)$ and $x_0\tilde{q_3}(x_1,x_2)+\tilde{q_4}(x_1,x_2)$ are the quartic curves. If $p$ is a quadruple point of both quartic curves then $q_3$ and $\tilde{q_3}$ are 0, so $X$ is of the form ($\ref{SS7}$).
This completes the proof of Theorem \[theorem:non-stable\].
Stable Locus {#sec:stablelocus}
------------
The classification of singularities of non-stable quintic threefolds can be used to give a partial description of the singularities which occur in the stable locus. The stable locus represents all of the closed orbits in the moduli space. Ideally, the stable locus would only include smooth hypersurfaces and the boundary would include hypersurfaces with singularities. Even in the case of cubic threefolds and cubic fourfolds, this is not the case as shown in[@laza; @allcock]. As the degree and dimension of hypersurfaces increases more singularity types will be included in the stable locus. In [@mumford] there is a general proposition which states that a smooth hypersurface will always be stable.
A smooth hypersurface $F$ in ${\mathbb{P}}^n$ with degree $\geq 2$ is a stable hypersurface.
A complete classification of all possible singularities in the stable locus of the moduli space of quintic threefolds has not yet been found. Using the results of the previous section a partial list of singularities can be determined.
If $X$ is a quintic threefold with at worst a double point then it is stable.
Suppose $X$ is not stable, then it is non-stable. Therefore, it belongs to one of the families SS1 - SS7, but $X$ does not satisfy the singularity criteria for any of these families. Hence, it is stable.
If $X$ is a quintic threefold with at worst a triple point whose tangent cone is an irreducible cubic surface and $X$ does not contain a plane then it is stable.
Suppose $X$ is not stable, then it is non-stable. Therefore, it belongs to one of the families SS1 - SS7. The only families, which have at worst a triple point are families SS5 and SS7. Since the tangent cone of $X$ is irreducible then it is not in SS5. Since $X$ does not contain a plane it is not in SS7, therefore it belong to neither family. Hence, it is stable.
If $X$ is a quintic threefold with at worst a double line whose tangent cone at each point on the line is irreducible then it is stable.
Suppose $X$ is not stable, then it is non-stable. So it belongs to one of the families SS1 - SS7. SS6 is the only family, which has at worst a double line as a singularity. Since the tangent cone of $X$ at each point is irreducible then it is not in SS6. Hence, it is stable.
These four classes of hypersurfaces give the most generic classes of hypersurfaces which are stable. There are also quintic threefolds which have singularities with degenerate tangent cones that do not fit into one of the classes SS1-SS7, but a complete classification is still unknown.
Minimal Orbits {#sec:minimalorbit}
==============
Luna’s Criterion {#sec:lunascriterion}
----------------
The ability to degenerate the large families $SS1 - SS7$ into much smaller invariant families $MO\hyph A - MO \hyph D$ makes the problem of finding minimal orbits much more tractable. Generically, a hypersurface in one of the families $MO\hyph A - MO \hyph D$ will be closed and thus minimal. To explicitly determine which elements in $MO\hyph A - MO\hyph D$ are closed and which elements further degenerate one can use Luna’s criterion. This approach was used by Laza [@laza] in the case of cubic fourfolds.
Luna’s criterion is used when there is an affine $G$-variety $Y$ and a point $y \in Y$ which has a non-finite stabilizer $H \subseteq G$. If $Y^H$ is the set of points in $Y$ which are $H$-invariant and $N_G(H)$ is the normalizer of $H$ in $G$ then there is a natural action of $N_G(H)$ on $Y^H$. Luna’s criterion reduces the problem of determining whether $Gy$ is closed in $Y$ to whether $N_G(H)$ is closed in $Y^H$.
[@vinberg; @luna]
Let $Y$ be an affine variety with a $G$-action and $y \in Y$ a point stabilized by a subgroup $H \subseteq G$. Then the orbit $Gy$ is closed in $Y$ if and only if the orbit $N_G(H)y$ is closed in $Y^H$.
In the case where $H$ is reductive and connected, $N_G(H) = H \cdot Z_G(H)$ where $Z_G(H)$ is the centralizer of $H$ in $G$. Since $N_G(H)$ acts on $Y^H$ we can quotient out by $H$. Thus, we can study the action of $Z_G(H)$, instead of $N_G(H)$, on $Y^H$.
The case of quintic threefolds consists of an $SL(5,{\mathbb{C}})$-action on the projective variety ${\mathbb{P}}(V)$. $V$ is the linearization of the $SL(5,{\mathbb{C}})$-action on ${\mathbb{P}}(V)$. The closed orbits of points in the linearization $V$ correspond to closed orbits of points in ${\mathbb{P}}(V)$. This correspondence between a projective variety and its linearization allows us to apply Luna’s criterion to $V$. Given a point from one of the families $MO\hyph A- MO \hyph D$, the stabilizer subgroup is the invariant 1-PS i.e. $H=\lambda$. The following lemma reduces the problem of finding minimal orbits to finding stable points in the $Z_G(H)$-action on $V^H$.
Let $v\in V$ be a point with stabilizer $H$ i.e. $v \in V^H$. If $v\in V$ is stable with respect to the $Z_G(H)$-action on the $H$-invariant space $V^H$ then the orbit $Gv$ is closed.
Let $v\in V^H$ be stable with respect to the $Z_G(H)$-action on $V^H$. By the definition of stable point, the orbit $Z_G(H)v$ is closed. By Luna’s criterion, $Gv$ is closed.
First Level of Minimal Orbits {#sec:firstlevelminimalorbits}
-----------------------------
The centralizer groups for the families $MO\hyph A - MO\hyph D$ are given in table 3. If $v \in V$ from $MO\hyph A- MO \hyph D$ is $Z_G(H)$-stable in $V^H$, then it is a minimal orbit. If that point $v$ is non-stable, with respect to the $Z_G(H)$-action, then it is either unstable or there is a destablizing 1-PS $\lambda$. This destabilizing 1-PS further degenerates $v$ into a smaller family, with a different stabilizer $H'$. The same process is then repeated for the smaller families. The Hilbert-Mumford criterion, with respect to the $Z_G(H)$-action on $V^H$, can be applied to determine precisely the stable, semistable, and unstable points.
### Minimal Orbit A {#sec:minorbita}
In the case of family $MO \hyph A$, the centralizer $Z_G(H)={\mathbb{C}}^2 \times SL(3,{\mathbb{C}}) \subseteq SL(5,{\mathbb{C}})$ acts on polynomials in $MO\hyph A$ which are of the form $q_{2,3}(x_0,x_1 \parallel x_2,x_3,x_4)$. The minimal orbit can be written as
$$\label{eq:moadecomp} x_0^2s_3(x_2,x_3,x_4) + x_0x_1t_3(x_2,x_3,x_4) + x_1^2u_3(x_2,x_3,x_4),$$
where $s_3(x_2,x_3,x_4)$, $t_3(x_2,x_3,x_4)$, $u_3(x_2,x_3,x_4)$ are degree $3$ polynomials in the variables $x_2$, $x_3$, and $x_4$. The polynomials which represent closed orbits in this family are stable with respect to the $SL(2,{\mathbb{C}}) \times SL(3,{\mathbb{C}})$ action on (\[eq:moadecomp\]). The polynomials that further degenerate are semistable with respect to the $SL(2,{\mathbb{C}}) \times SL(3,{\mathbb{C}})$ action. The unstable points are unstable with respect to the $SL(2,{\mathbb{C}}) \times SL(3,{\mathbb{C}})$ action. The set of semistable and unstable points can be found by modifying the techniques in section \[sec:maxnon-stable\] which involved classifying $G$-orbits by using (\[eq:gequivariant\]) to find which polynomials which satisfy the Hilbert-Mumford criterion for the set of normalized 1-PS $\lambda$. A normalized 1-PS in $SL(2,{\mathbb{C}}) \times SL(3,{\mathbb{C}})$ is of the following form:
$$\label{moaweylchamber}
\begin{pmatrix}
t & 0 & 0 & \cdots & 0 \\
0 & t^{-1} & 0 & \cdots & 0 \\
0 & 0 & t^a & 0 & 0 \\
\vdots & \vdots & 0 & t^b & 0 \\
0 & 0 & 0 & 0 & t^c
\end{pmatrix}$$
where $a+b+c=0$ and $a \geq b \geq c$. The normalization restriction of the $SL(3,{\mathbb{C}})$ block gives an ordering of the degree $3$ monomials in the variables $x_2,x_3,x_4$. The weights of $x_0^2$, $x_0x_1$, and $x_2^2$ are $2$,$0$, and $-2$ respectively. The set of maximal non-stable polynomials $F$ are those where $\mu(F,\lambda) \leq 0$ and the maximal unstable families have $\mu(F, \lambda) < 0$. A general polynomial in (\[eq:moadecomp\]) will be semistable if $s_3$, $t_3$, and $u_3$ have at most weights $-2$, $0$, and $2$, with respect to the action (\[moaweylchamber\]). These weights are neccesary in order balance the weights arising from $x_0^2$, $x_0x_1$, $x_1^2$ so that $\mu$ is less than zero. Similary, the weights for a polynomial in (\[eq:moadecomp\]) the weights for $s_3$, $t_3$, and $u_3$ are at most $-3$, $-1$, and $1$ for the polynomial to be unstable. The calculation below gives the set of semistable and unstable families. A similar method can be used for all other minimal orbits in order to explicitly calculate the set of semistable and unstable families.
\[fig:SS7a\]
From the GIT analysis of $MO \hyph A$, the semistable hypersurfaces and their corresponding destabilizing 1-PS $\lambda$ are given below.
1. $ x_0^2( q_3(x_3,x_4)+q_1(x_2,x_3)x_4^2+x_4^3) + x_0x_1 ( q_3(x_3,x_4) + x_2x_3x_4 + q_1(x_2,x_3)x_4^2+x_3x_4^2+x_3^2x_4+x_4^3 ) + x_1^2 ( x_2q_2(x_3,x_4) + q_3(x_3,x_4) )$
1. Destabilizing 1-PS: $\langle 1, -1, 4, -1, -3 \rangle$
2. $x_0^2 ( q_3(x_3,x_4) ) + x_0x_1 ( x_2q_2(x_3,x_4) + q_3(x_3,x_4) ) + x_1^2 ( x_2q_2(x_3,x_4) + q_3(x_3,x_4) )$
1. Destabilizing 1-PS: $\langle 1, -1, 2, -1, -1 \rangle$
3. $x_0^2 ( q_1(x_2,x_3)x_4^2+x_4^3 ) + x_0x_1 ( q_2(x_2,x_3)x_4 + q_1(x_2,x_3)x_4^2+ x_4^3 ) + x_1^2 ( q_2(x_2,x_3)x_4 + q_1(x_2,x_3)x_4^2+ x_4^3 )$
1. Destabilizing 1-PS: $\langle 1, -1, 1, 1, -2 \rangle$
4. $x_0^2 ( x_3x_4^2+x_4^3 ) + x_0x_1 ( q_3(x_3,x_4) + x_2x_3x_4 + q_1(x_2,x_3)x_4^2+x_3x_4^2+x_3^2x_4+x_4^3 ) + x_1^2 ( x_2^2q_1(x_3,x_4) + x_2q_2(x_3,x_4) + q_3(x_3,x_4) )$
1. Destabilizing 1-PS: $\langle 1, -1, 1, 0, -1 \rangle$
5. $x_0^2 ( q_1(x_2,x_3)x_4^2+x_3^2x_4+x_4^3 ) + x_0x_1 ( q_3(x_3,x_4) + x_2x_3x_4 + q_1(x_2,x_3)x_4^2+x_3x_4^2+x_3^2x_4+x_4^3 ) + x_1^2 ( x_2q_2(x_3,x_4) + q_3(x_3,x_4) + q_2(x_2,x_3)x_4 + x_4^3 )$
1. Destabilizing 1-PS: $\langle 1, -1, 2, 0, -2 \rangle$
From the poset analysis, the set of unstable families are given below.
1. $x_0^2(q_3(x_3,x_4)+x_2x_4^2) + x_0x_1(q_3(x_3,x_4)+x_2x_4^2) + x_1^2(q_3(x_3,x_4)+x_2x_3x_4+x_2x_4^2)$
2. $x_0^2(q_3(x_3,x_4)) + x_0x_1(q_3(x_3,x_4)) + x_1^2(q_3(x_3,x_4)+x_2q_2(x_3,x_4))$
3. $x_0^2(x_2x_4^2+x_3^2x_4+x_3x_4^2+x_4^3) + x_0x_1(x_2x_4^2+x_3^2x_4+x_3x_4^2+x_4^3) + x_1^2(q_3(x_3,x_4)+x_2x_3x_4+x_2x_4^2)$
4. $x_0^2(x_2x_4^2+x_3x_4^2+x_4^3) + x_0x_1(x_2x_4^2+x_3^2x_4+x_3x_4^2+x_4^3) + x_1^2(q_2(x_2,x_3)x_4+q_1(x_2,x_3)x_4^2+x_4^3)$
5. $x_0^2(x_4^3) + x_0x_1(q_1(x_2,x_3)x_4^2+x_3^2x_4+x_4^3) + x_1^2(q_3(x_3,x_4)+x_2^2x_4+x_2x_3x_4+x_2x_4^2)$
6. $x_0^2(x_3x_4^2+x_4^3) + x_0x_1(q_1(x_2,x_3)x_4^2+x_3^2x_4+x_4^3) + x_1^2(q_3(x_3,x_4)+x_2x_3x_4+x_2x_4^2)$
7. $x_0^2(x_3x_4^2+x_4^3) + x_0x_1(q_1(x_2,x_3)x_4^2+x_3^2x_4+x_4^3) + x_1^2(q_2(x_2,x_3)x_4+q_1(x_2,x_3)x_4^2+x_4^3)$
8. $x_0^2(x_3^2x_4+x_3x_4^2+x_4^3) + x_0x_1(q_1(x_2,x_3)x_4^2+x_3^2x_4+x_4^3) + x_1^2(x_2q_2(x_3,x_4)+q_3(x_3,x_4))$
9. $x_0^2(x_3x_4^2+x_4^3) + x_0x_1(q_1(x_2,x_3)x_4^2+x_3^2x_4+x_4^3) + x_1^2(x_2q_2(x_3,x_4)+q_3(x_3,x_4))$
10. $x_0^2(x_4^3) + x_0x_1(q_1(x_2,x_3)x_4^2+x_3^2x_4+x_4^3) + x_1^2(x_2q_2(x_3,x_4)+q_3(x_3,x_4))$
11. $x_0^2(x_3x_4^2+x_4^3) + x_0x_1(q_1(x_2,x_3)x_4^2+x_3^2x_4+x_4^3) + x_1^2(x_2x_3x_4+x_2x_4^2+q_3(x_3,x_4))$
12. $x_0^2(x_3x_4^2+x_4^3) + x_0x_1(q_1(x_2,x_3)x_4^2+x_3^2x_4+x_4^3) + x_1^2(q_2(x_2,x_3)x_4+q_1(x_2,x_3)x_4^2+x_4^3)$
13. $x_0^2(x_4^3) + x_0x_1(q_1(x_2,x_3)x_4^2+x_3^2x_4+x_4^3) + x_1^2(q_2(x_2,x_3)x_4+q_1(x_2,x_3)x_4^2+x_4^3+x_3^3)$
14. $x_0^2(x_4^3+x_3^2x_4) + x_0x_1(x_4^3+x_3^2x_4) + x_1^2(x_2q_2(x_3,x_4)+q_3(x_3,x_4))$
15. $x_0^2(x_4^3) + x_0x_1(x_4^3+x_3^2x_4) + x_1^2(x_2q_2(x_3,x_4)+q_3(x_3,x_4)+x_2^3x_4)$
16. $x_0^2(x_3^2x_4+x_4^3) + x_0x_1(x_4^3+x_3^2x_4) + x_1^2(q_3(x_3,x_4)+x_2x_3x_4+x_2x_4^2)$
17. $x_0^2(x_3^2x_4+x_4^3) + x_0x_1(x_4^3+x_3^2x_4) + x_1^2(q_2(x_2,x_3)x_4+q_1(x_2,x_3)x_4+x_4^3)$
18. $x_0^2(x_4^3) + x_0x_1(x_4^3+x_3^2x_4) + x_1^2(q_3(x_3,x_4)+x_2x_3x_4+x_2x_4^2+x_3^3)$
Let $X$, up to a coordinate transformation, be of the form $q_{2,3}(x_0,x_1 \parallel x_2,x_3,x_4)$.
1. If $X$ belongs to one of the families US1-A - US18-A then X is unstable.
2. If $X$ is of type SS1-A then the orbit is not closed and it degenerates to MO2-I.
3. If $X$ is of type SS2-A then the orbit is not closed and it degenerates to MO2-IV.
4. If $X$ is of type SS3-A then the orbit is not closed and it degenerates to MO2-IV.
5. If $X$ is of type SS4-A then the orbit is not closed and it degenerates to MO2-I.
6. If $X$ is of type SS5-A then the orbit is not closed and it degenerates to M02-IV.
Otherwise, $X$ is a closed orbit.
### Minimal Orbit B {#sec:minorbitb}
In the case of $MO \hyph B$, the centralizer $Z_G(H)={\mathbb{C}}^* \times SL(3,{\mathbb{C}}) \times {\mathbb{C}}^* \subseteq SL(5,{\mathbb{C}})$ acts on polynomials in $MO\hyph B$ which are of the form $q_5(x_1,x_2,x_3)+x_0x_4q_3(x_1,x_2,x_3)+x_0^2x_4^2q_1(x_1,x_2,x_3)$. The action on $Z_G(H)$ on $x_0x_4$ and $x_0^2x_4^2$ will have weights zero. By modifying the procedure in the $MO \hyph A$ case, one can obtain the maximal semistable and unstable families with respect to the $Z_G(H)$ action. The poset analysis shows that the maximal semistable families are the following families below.
1. $ ( x_1q_4(x_2,x_3) + q_5(x_2,x_3) ) + x_0x_4 ( x_1q_2(x_2,x_3) + q_3(x_2,x_3) ) + x_0^2x_4^2 ( q_1(x_2,x_3) )$
1. Destabilizing 1-PS: $\langle 1, 2, -1, -1, -1 \rangle$
2. $ ( x_1q_4(x_2,x_3) + q_5(x_2,x_3) + q_2(x_1,x_2)x_3^2 ) + x_0x_4 ( q_3(x_2,x_3) + x_1x_2x_3 + q_1(x_1,x_2)x_3^2 ) + x_0^2x_4^2 ( q_1(x_2,x_3) )$
1. Destabilizing 1-PS: $\langle 1, 3, -1, -2, -1 \rangle$
3. $ ( q_3(x_1,x_2)x_3^2 + q_2(x_1,x_2)x_3^3 + q_1(x_1,x_2)x_3^4 + x_3^5 ) + x_0x_4 ( q_2(x_1,x_2)x_3 + q_1(x_1,x_2)x_3^2 + x_3^3 ) + x_0^2x_4^2 ( x_3 )$
1. Destabilizing 1-PS: $\langle 1, 1, 1, -2, -1 \rangle$
4. $ ( q_5(x_2,x_3) + x_1x_2^3x_3 + x_1x_2^2x_3^2 + x_1x_2x_3^3 + x_1x_3^4 + x_1^2x_2x_3^2 + q_2(x_1,x_2)x_3^3 ) + x_0x_4 ( q_3(x_2,x_3) + x_1x_2x_3 + q_1(x_1,x_2)x_3^2 ) + x_0^2x_4^2 ( q_1(x_2,x_3) )$
1. Destabilizing 1-PS: $\langle 1, 1, 0, -1, -1 \rangle$
From the poset analysis, the set of unstable families are given below.
1. $(x_1q_4(x_2,x_3)+x_1^2x_3)+x_0x_4(q_3(x_2,x_3)+x_1x_3^2)+x_0^2x_4^2(x_2+x_3)$
2. $(q_5(x_2,x_3)+x_1x_2^3x_3+x_1x_2^2x_3^2+x_1^2x_3^3+x_1x_2x_3^3+x_1x_3^4)+x_0x_4(q_1(x_1,x_2)x_3^2+x_2^2x_3+x_3^3)+x_0^2x_4^2(x_3)$
3. $(x_1^2x_2x_3^2+x_1x_2^2x_3^2+x_2^4x_3+x_2^3x_3^2+x_2^2x_3^3+q_2(x_1,x_2)x_3^3+q_1(x_1,x_2)x_3^4+x_3^5)+x_0x_4(q_1(x_1,x_2)x_3^2+x_2^2x_3+x_3^3)+x_0^2x_4^2(x_3)$
4. $(q_3(x_1,x_2)x_3^2+q_2(x_1,x_2)x_3^3+q_1(x_1,x_2)x_3^4+x_4^5)+x_0x_4(q_1(x_1,x_2)x_3^2+x_2^2x_3+x_3^3)+x_0^2x_4^2(x_3)$
Let $X$, up to a coordinate transformation, be of the form $q_5(x_1,x_2,x_3)+x_0q_3(x_1,x_2,x_3)x_4+x_0^2q_1(x_1,x_2,x_3)x_4^2$.
1. If $X$ belongs to one of the families US1-B - US4-B then $X$ is unstable.
2. If $X$ is of type SS1-B then the orbit is not closed and it degenerates to MO2-IV.
3. If $X$ is of type SS2-B then the orbit is not closed and it degenerates to MO2-V.
4. If $X$ is of type SS3-B then the orbit is not closed and it degenerates to MO2-IV.
5. If $X$ is of type SS4-B then the orbit is not closed and it degenerates to MO2-VI.
Otherwise, $X$ is a closed orbit.
### Minimal Orbit C {#sec:minorbitc}
In the case of $MO \hyph C$, the centralizer $Z_G(H)= SL(2,{\mathbb{C}}) \times SL(2,{\mathbb{C}}) \times {\mathbb{C}}^* \subseteq SL(5,{\mathbb{C}})$ acts on polynomials in $MO\hyph C$ which are of the form $q_{1,4}(x_0,x_1 \parallel x_2,x_3) + x_4q_{2,2}(x_0,x_1 \parallel x_2,x_3) + x_4^2q_3(x_0,x_1)$. By modifying the procedure in the $MO \hyph A$, one can obtain the maximal semistable and unstable families with respect to the $Z_G(H)$ action. The poset analysis shows that the maximal semistable families are the following families below.
1. $ ( x_0(x_2x_3^3 + x_3^4) + x_1( x_2^2x_3^2+x_2x_3^3 + x_3^4) ) + x_4^2 ( x_0x_1^2 + x_1^3 ) + x_4 ( x_0^2 (x_3^2) + x_0x_1 ( x_2x_3 + x_3^2) + x_1^2(x_2^2+x_2x_3 + x_3^2) )$
1. Destabilizing 1-PS: $\langle 1, -1, 1, -1, 0 \rangle$
From the poset analysis, the set of unstable families are given below.
1. $x_0(x_2x_3^3+x_3^4)+x_1(x_2^2x_3^2+x_2x_3^3+x_3^3)+x_4^2(x_0x_1^2+x_1^3)+x_4x_0x_1(x_3^2)+x_4x_1^2(x_2x_3+x_3^2)$
Let $X$, up to a coordinate transformation, be of the form $q_{1,4}(x_0,x_1 \parallel x_2,x_3) + q_3(x_0,x_1)x_4^2+ q_{2,2}(x_0,x_1 \parallel x_2,x_3)x_4 $.
1. If $X$ belongs to the family US1-C then $X$ is unstable.
2. If $X$ is of type SS1-C then the orbit is not closed and it degenerates to MO2-VII.
Otherwise, $X$ is a closed orbit.
### Minimal Orbit D {#sec:minorbitd}
In the case of $MO \hyph D$, the centralizer $Z_G(H)= {\mathbb{C}}^* \times SL(4,{\mathbb{C}}) \subseteq SL(5,{\mathbb{C}})$ acts on polynomials in $MO\hyph D$ which are of the form $x_0q_4(x_1,x_2,x_3,x_4)$. By modifying the procedure in the $MO \hyph A$ case, one can obtain the maximal semistable and unstable families with respect to the $Z_G(H)$ action. The poset analysis shows that the maximal semistable families are the following families below.
1. $ x_0 ( x_1q_3(x_2,x_3,x_4) + q_4(x_2,x_3,x_4) )$
1. Destabilizing 1-PS: $\langle 0, 3, -1, -1, -1 \rangle$
2. $ x_0 ( q_3(x_1,x_2,x_3)x_4 + q_2(x_1,x_2,x_3)x_4^2 + q_1(x_1,x_2,x_3)x_4^3 + x_4^4 )$
1. Destabilizing 1-PS: $\langle 0, 1, 1, 1, -3 \rangle$
3. $ x_0 ( q_{2,2}(x_1,x_2 \parallel x_3,x_4) + q_{1,3}(x_1,x_2 \parallel x_3,x_4) + q_4(x_3,x_4) )$
1. Destabilizing 1-PS: $\langle 0, 1, 1, -1, -1 \rangle$
4. $ x_0 ( q_4(x_2,x_3,x_4) + x_1q_2(x_2,x_3)x_4 + q_2(x_1,x_2,x_3)x_4^2 + q_1(x_1,x_2,x_3)x_4^3 ) $
1. Destabilizing 1-PS: $\langle 0, 1, 0, 0, -1 \rangle$
From the poset analysis, the set of unstable families are given below.
1. $q_4(x_2,x_3,x_4) + q_{1,3}(x_1,x_2 \parallel x_3,x_4)+q_4(x_3,x_4)+x_1x_2x_4^2$
2. $x_2^2x_3x_4+ q_{1,3}(x_1,x_2 \parallel x_3,x_4)+q_4(x_3,x_4)+q_2(x_1,x_2,x_3)x_4^2+q_1(x_1,x_2,x_3)x_4^3$
3. $x_2q_3(x_3,x_4)+q_4(x_3,x_4)+q_3(x_2,x_3)x_4+q_2(x_2,x_3)x_4^2+q_1(x_2,x_3)x_4^3+q_1(x_1,x_2)x_3^2x_4+q_2(x_1,x_2,x_3)x_4^2+q_1(x_1,x_2,x_3)x_4^3$
Let $X$, up to a coordinate transformation, be of the form $x_0q_4(x_1,x_2,x_3,x_4)$.
1. If $X$ belongs to one of the families US1-D - US3-D then $X$ is unstable.
2. If $X$ is of type SS1-D then the orbit is not closed and it degenerates to MO2-VII.
3. If $X$ is of type SS2-D then the orbit is not closed and it degenerates to MO2-VIII.
4. If $X$ is of type SS3-D then the orbit is not closed and it degenerates to MO2-IX.
5. If $X$ is of type SS4-D then the orbit is not closed and it degenerates to MO2-X.
Otherwise, $X$ is a closed orbit.
Boundary Structure {#sec:boundary}
==================
From the propositions in sections \[sec:minorbita\], \[sec:minorbitb\], \[sec:minorbitc\], and \[sec:minorbitd\] it was shown that certain hypersurfaces in the families $MO \hyph A - MO \hyph D$ degenerate into even smaller families. In the GIT compactification these smaller families are the boundary strata where the boundaries components $MO \hyph A - MO \hyph D$ interesect. Again, Luna’s criterion it can be determined which members of these families represent closed orbit as well as which families degenerate further. The centralizers for the families in the second level of minimal orbits are given in table 4.
\[table:secondlevel\]
[ | c | c | c | ]{} *Family* & *Invariant 1-PS Subgroup (H)* & *Centralizer of H ($Z_G(H))$*\
MO2-I & $\langle 6,0,2,-3,-5 \rangle$ & Maximal Torus $T$\
\
MO2-II & $\langle 4,2,-1,-2,-3 \rangle$ & Maximal Torus $T$\
\
MO2-III & $\langle 4,2,0,-2,-4 \rangle$ & Maximal Torus $T$\
\
MO2-IV & $\langle 4,2,-1,-1,-5 \rangle$ & ${\mathbb{C}}^{*2} \times SL(2,{\mathbb{C}}) \times {\mathbb{C}}^*$\
\
MO2-V & $\langle 5,3,-1,-2,-7 \rangle$ & Maximal Torus $T$\
\
MO2-VI & $\langle 2,1,0,-1,-2 \rangle$ & Maximal Torus $T$\
\
MO2-VII & $\langle 5,3,0,-2,-6 \rangle$ & Maximal Torus $T$\
\
MO2-VIII & $\langle 4,2,-2,-2,-2 \rangle$ & ${\mathbb{C}}^{*2} \times SL(3,{\mathbb{C}})$\
\
MO2-IX & $\langle 4,0,0,-2,-2 \rangle$ & ${\mathbb{C}}^{*} \times SL(2,{\mathbb{C}}) \times SL(2, {\mathbb{C}})$\
\
MO2-X & $\langle 4,0,-1,-1,-2 \rangle$ & ${\mathbb{C}}^{*2} \times SL(2,{\mathbb{C}}) \times {\mathbb{C}}^* $\
\
Second Level of Minimal Orbits {#sec:secondlevelminimalorbits}
------------------------------
### MO2-I
In the case of family $MO2-I$, the maximal torus $T$ acts on polynomials of the form $x_0^2 (x_2x_4^2) + x_0x_1(x_2x_3x_4) + x_1^2(x_2x_3^2)$. The weights of a 1-PS $\lambda=\langle a_0, a_1, a_2, a_3, a_4 \rangle$ of the maximal torus $T$ acting on the monomials are subject to the constraint $a_0+a_1+a_2+a_3+a_4=0$. In a GIT semistable family all of the monomials, subject to the constraint, will have weight at most $0$. Similarly, all GIT unstable families will have weight at most $-1$. The set of semistable families are given below.
1. $x_0^2 ( x_2x_4^2 ) + x_0x_1 ( x_2x_3x_4 ) $
1. Destabilizing 1-PS: $\langle 1,1,-1,0,-1 \rangle$
2. $ x_0x_1 ( x_2x_3x_4 ) + x_1^2 ( x_2x_3^2 )$
1. Destabilizing 1-PS: $\langle 1,-1,0,0,0 \rangle$
The set of unstable families are given below.
1. $x_0^2x_2x_4^2$
2. $x_1^2x_2x_3^2$
Let $X$, up to a coordinate transformation, be of the form $ x_0^2 ( x_2x_4^2 ) + x_0x_1 ( x_2x_3x_4 ) + x_1^2 ( x_2x_3^2 )$.
1. If $X$ belongs to one of the families US1-I - US2-I then $X$ is unstable.
2. If $X$ is of type SS1-I then the orbit is not closed and it degenerates to MO2-V.
3. If $X$ is of type SS2-I then the orbit is not closed and it degenerates to MO2-V.
Otherwise, $X$ is a closed orbit.
### MO2-II
In the case of family $MO2-II$, the maximal torus $T$ acts on polynomials of the form $ x_0^2 ( x_3x_4^2 ) + x_0x_1 ( x_3^3 + x_2x_3x_4 ) + x_1^2 ( x_2^2x_3 )$. Following the same procedure as in MO2-I, the set of semistable families are given below.
1. $x_0^2 ( x_3x_4^2 ) + x_0x_1 ( x_3^3 + x_2x_3x_4 ) $
1. Destabilizing 1-PS: $\langle 1,1,0,-1,-1 \rangle$
2. $x_0x_1 ( x_3^3 + x_2x_3x_4 ) + x_1^2 ( x_2^2x_3 )$
1. **1-PS** $\langle 1,0,0,-1,0 \rangle$
3. $ x_0x_1 ( x_2x_3x_4 ) + x_1^2 ( x_2^2x_3 )$
1. Destabilizing 1-PS: $\langle 1,-1,0,0,0 \rangle$
4. $x_0^2 ( x_3x_4^2 ) + x_0x_1 ( x_2x_3x_4 ) $
1. Destabilizing 1-PS: $\langle 1,1,0,-1,-1 \rangle$
5. $ x_0x_1 ( x_3^3 + x_2x_3x_4 ) $
1. Destabilizing 1-PS: $\langle 1,-2,0,0,1 \rangle$
6. $ x_0x_1 ( x_2x_3x_4 )$
1. Destabilizing 1-PS: No 1-PS
The set of unstable families are given below.
### Unstable Family
1. $x_0^2x_3x_4^2+x_0x_1x_3^2$
2. $x_0x_1x_3^2+x_1^2x_2^2x_3$
3. $x_0x_1x_3^2$
4. $x_0^2x_3x_4^2$
5. $x_1^2x_2^2x_3$
Let $X$, up to a coordinate transformation, be of the form $ x_0^2 ( x_3x_4^2 ) + x_0x_1 ( x_3^3 + x_2x_3x_4 ) + x_1^2 ( x_2^2x_3 )$ .
1. If $X$ belongs to one of the families US1-II - US5-II then $X$ is unstable.
2. If $X$ is of type SS1-II then the orbit is not closed and it degenerates to MO2-V.
3. If $X$ is of type SS2-II then the orbit is not closed and it degenerates to MO2-V.
4. If $X$ is of type SS3-II then the orbit is not closed and it degenerates to MO2-V.
5. If $X$ is of type SS4-III then the orbit is not closed and it degenerates to MO2-V.
6. If $X$ is of type SS5-II then the orbit is not closed and it degenerates to MO2-V.
7. If $X$ is of type SS6-II then the orbit is not closed and it degenerates to MO2-V.
Otherwise, $X$ is a closed orbit.
### MO2-III
In the case of family $MO2-III$, the maximal torus $T$ acts on polynomials of the form $ x_0^2 ( x_2x_4^2 + x_3^2x_4 ) + x_0x_1 ( x_3^3 + x_2x_3x_4 ) + x_1^2 ( x_2x_3^2 + x_2^2x_4 )$. Following the same procedure as in MO2-I, the set of semistable families are given below.
1. $ x_0^2 ( x_2x_4^2 + x_3^2x_4 ) + x_0x_1 ( x_3^3 + x_2x_3x_4 ) $
1. Destabilizing 1-PS: $\langle 1,3,3,-4,-3 \rangle$
2. $ x_0x_1 ( x_3^3 + x_2x_3x_4 ) + x_1^2 ( x_2x_3^2 + x_2^2x_4 )$
1. Destabilizing 1-PS: $\langle 1,-8,1,2,4 \rangle$
3. $ x_0^2 ( x_2x_4^2 + x_3^2x_4 ) + x_0x_1 ( x_2x_3x_4 ) $
1. Destabilizing 1-PS: $\langle 1,2,0,0,-3 \rangle$
4. $ x_0x_1 ( x_2x_3x_4 ) + x_1^2 ( x_2x_3^2 + x_2^2x_4 )$
1. Destabilizing 1-PS: $\langle 1,-1,0,0,0 \rangle$
5. $ x_0^2 ( x_3^2x_4 ) + x_0x_1 ( x_3^3 + x_2x_3x_4 ) + x_1^2 ( x_2x_3^2 )$
1. Destabilizing 1-PS: $\langle 1,1,2,-7,3 \rangle$
6. $ x_0^2 ( x_3^2x_4 ) + x_0x_1 ( x_2x_3x_4 ) + x_1^2 ( x_2x_3^2 )$
1. Destabilizing 1-PS: $\langle 1,0,0,-2,1 \rangle$
7. $ x_0^2 ( x_2x_4^2 ) + x_0x_1 ( x_2x_3x_4 ) + x_1^2 ( x_2^2x_4 )$
1. Destabilizing 1-PS: $\langle 1,0,-1,1,-1 \rangle$
8. $ x_0^2 ( x_2x_4^2 ) + x_0x_1 ( x_3^3 + x_2x_3x_4 ) $
1. Destabilizing 1-PS: $\langle 1,1,1,-1,-2 \rangle$
9. $ x_0^2 ( x_3^2x_4 ) + x_0x_1 ( x_3^3 + x_2x_3x_4 ) $
1. Destabilizing 1-PS: $\langle 1,0,0,-2,1 \rangle$
10. $ x_0x_1 ( x_3^3 + x_2x_3x_4 ) + x_1^2 ( x_2^2x_4 )$
1. Destabilizing 1-PS: $\langle 1,-2,0,0,1 \rangle$
11. $ x_0x_1 ( x_3^3 + x_2x_3x_4 ) + x_1^2 ( x_2x_3^2 )$
1. Destabilizing 1-PS: $\langle 1,-2,0,0,1 \rangle$
12. $ x_0x_1 ( x_2x_3x_4 ) + x_1^2 ( x_2^2x_4 )$
1. Destabilizing 1-PS: $\langle 1,0,0,0,-1 \rangle$
13. $ x_0x_1 ( x_2x_3x_4 ) + x_1^2 ( x_2x_3^2 )$
1. **1-PS** $\langle 1,-1,0,0,0 \rangle$
14. $ x_0^2 ( x_2x_4^2 ) + x_0x_1 ( x_2x_3x_4 ) $
1. Destabilizing 1-PS: $\langle 1,1,-1,0,-1 \rangle$
15. $ x_0^2 ( x_3^2x_4 ) + x_0x_1 ( x_2x_3x_4 ) $
1. Destabilizing 1-PS: $\langle 1,2,0,0,-3 \rangle$
16. $ x_0x_1 (x_3^3 + x_2x_3x_4 ) $
1. Destabilizing 1-PS: $\langle 1,-2,0,0,1 \rangle$
17. $ x_0x_1 ( x_2x_3x_4 )$
1. Destabilizing 1-PS: No 1-PS
The set of unstable families are given below.
1. $ x_0^2 \bigg( x_2x_4^2 + x_3^2x_4 \bigg) + x_0x_1 \bigg( x_3^3 \bigg) $
2. $ x_0x_1 \bigg( x_3^3 + \bigg) + x_1^2 \bigg( x_2x_3^2 + x_2^2x_4 \bigg)$
3. $ x_0^2 \bigg( x_3^2x_4 \bigg) + x_0x_1 \bigg( x_3^3 \bigg) + x_1^2 \bigg( x_2x_3^2 \bigg)$
4. $ x_0^2 \bigg( x_2x_4^2 \bigg) + x_0x_1 \bigg( x_3^3 \bigg) + x_1^2 \bigg( x_2^2x_4 \bigg)$
5. $ x_0^2 \bigg( x_2x_4^2 + x_3^2x_4 \bigg) $
6. $ x_1^2 \bigg( x_2x_3^2 + x_2^2x_4 \bigg)$
7. $ x_0^2 \bigg( x_3^2x_4 \bigg) + x_1^2 \bigg( x_2x_3^2 \bigg)$
8. $ x_0^2 \bigg( x_2x_4^2 \bigg) + x_1^2 \bigg( x_2^2x_4 \bigg)$
9. $ x_0^2 \bigg( x_2x_4^2 \bigg) + x_0x_1 \bigg( x_3^3 \bigg) $
10. $ x_0x_1 \bigg( x_3^3 + \bigg) + x_1^2 \bigg( x_2^2x_4 \bigg)$
11. $ x_0^2 \bigg( x_3^2x_4 \bigg) + x_0x_1 \bigg( x_3^3 \bigg) $
12. $ x_0x_1 \bigg( x_3^3 + \bigg) + x_1^2 \bigg( x_2x_3^2 \bigg)$
13. $ x_0^2 \bigg( x_2x_4^2 \bigg) $
14. $ x_0^2 \bigg( x_3^2x_4 \bigg) $
15. $ x_0x_1 \bigg( x_3^3 \bigg) $
16. $ x_1^2 \bigg( x_2x_3^2 \bigg)$
17. $x_1^2 \bigg( x_2^2x_4 \bigg)$
Let $X$, up to a coordinate transformation, be of the form $ x_0^2 ( x_2x_4^2 + x_3^2x_4 ) + x_0x_1 ( x_3^3 + x_2x_3x_4 ) + x_1^2 ( x_2x_3^2 + x_2^2x_4 )$ .
1. If $X$ belongs to one of the families US1-III - US17-III then $X$ is unstable.
2. If $X$ is of type SS1-III then the orbit is not closed and it degenerates to MO2-V.
3. If $X$ is of type SS2-III then the orbit is not closed and it degenerates to MO2-V.
4. If $X$ is of type SS3-III then the orbit is not closed and it degenerates to MO2-V.
5. If $X$ is of type SS4-III then the orbit is not closed and it degenerates to MO2-V.
6. If $X$ is of type SS5-III then the orbit is not closed and it degenerates to MO2-V.
7. If $X$ is of type SS6-III then the orbit is not closed and it degenerates to MO2-V.
8. If $X$ is of type SS7-III then the orbit is not closed and it degenerates to MO2-V.
9. If $X$ is of type SS8-III then the orbit is not closed and it degenerates to MO2-V.
10. If $X$ is of type SS9-III then the orbit is not closed and it degenerates to MO2-V.
11. If $X$ is of type SS10-III then the orbit is not closed and it degenerates to MO2-V.
12. If $X$ is of type SS11-III then the orbit is not closed and it degenerates to MO2-V.
13. If $X$ is of type SS12-III then the orbit is not closed and it degenerates to MO2-V.
14. If $X$ is of type SS13-III then the orbit is not closed and it degenerates to MO2-V.
15. If $X$ is of type SS14-III then the orbit is not closed and it degenerates to MO2-V.
16. If $X$ is of type SS15-III then the orbit is not closed and it degenerates to MO2-V.
17. If $X$ is of type SS16-III then the orbit is not closed and it degenerates to MO2-V.
18. If $X$ is of type SS17-III then the orbit is not closed and it degenerates to MO2-V.
Otherwise, $X$ is a closed orbit.
### MO2-IV
In the case of family $MO2-IV$, the centralizer ${\mathbb{C}}^{*2} \times SL(2,{\mathbb{C}}) \times {\mathbb{C}}^*$ on polynomials of the form $ x_0x_4 ( x_1q_2(x_2,x_3) ) $. By following a similar procedure as in the $MO \hyph A$ case one can obtain the maximal semistable and unstable families with respect to the $Z_G(H)$ action. The poset analysis shows that the maximal semistable families are the following families below.
1. $ x_0x_4 ( x_1x_2x_3 + x_1x_3^2 ) $
1. Destabilizing 1-PS: $\langle 1,0,1,0,-2 \rangle$
The set of unstable families are given below.
1. $x_0x_4x_1x_3^2$
Let $X$, up to a coordinate transformation, be of the form $ x_0x_4 ( x_1q_2(x_2,x_3) ) $ .
1. If $X$ belongs to one of the families US1-IV then $X$ is unstable.
2. If $X$ is of type SS1-IV then the orbit is not closed and it degenerates to MO2-V.
Otherwise, $X$ is a closed orbit.
### MO2-V
In the case of family $MO2-V$, the maximal torus $T$ acts on polynomials of the form $ x_0x_4 ( x_1x_2x_3 )$. Following the same procedure as in MO2-I, the set of semistable families are given below.
1. $ x_0x_4 x_2x_3x_1 $
1. **1-PS** No 1-PS
The set of unstable families are given below.
1. NONE
Let $X$, up to a coordinate transformation, be of the form $ x_0x_4 x_1x_2x_3 $ then it is a closed orbit.
### MO2-VI
In the case of family $MO2-VI$, the maximal torus $T$ acts on polynomials of the form $ ( x_1x_2^3x_3 + x_1^2x_2x_3^2) + x_0x_4 ( x_1x_2x_3 )$ . Following the same procedure as in MO2-I, the set of semistable families are given below.
1. $ ( x_1x_2^3x_3 + x_1^2x_2x_3^2 ) + x_0x_4 ( x_1x_2x_3 ) $
1. Destabilizing 1-PS: $\langle 1,0,-1,0,0 \rangle$
2. $ ( x_1x_2^3x_3 ) + x_0x_4 ( x_1x_2x_3 ) $
1. Destabilizing 1-PS: $\langle 1,0,-1,0,0 \rangle$
3. $ ( x_1^2x_2x_3^2 ) + x_0x_4 ( x_1x_2x_3 ) $
1. Destabilizing 1-PS: $\langle 1,0,-1,0,0 \rangle$
4. $x_0x_4 ( x_1x_2x_3 ) $
1. Destabilizing 1-PS: No 1-PS
The set of unstable families are given below.
1. $( x_1x_2^3x_3 + x_1^2x_2x_3^2 )$
2. $( x_1x_2^3x_3 )$
3. $( x_1^2x_2x_3^2 )$
Let $X$, up to a coordinate transformation, be of the form $ ( x_1x_2^3x_3 + x_1^2x_2x_3^2 ) + x_0x_4 ( x_1x_2x_3 ) $ .
1. If $X$ belongs to one of the families US1-VI - US3-VI then $X$ is unstable.
2. If $X$ is of type SS1-VI then the orbit is not closed and it degenerates to MO2-V.
3. If $X$ is of type SS2-VI then the orbit is not closed and it degenerates to MO2-V.
4. If $X$ is of type SS3-VI then the orbit is not closed and it degenerates to MO2-V.
5. If $X$ is of type SS4-VI then the orbit is not closed and it degenerates to MO2-V.
Otherwise, $X$ is a closed orbit.
### MO2-VII
In the case of family $MO2-VII$, the maximal torus $T$ acts on polynomials of the form $ x_4x_0^2x_3^2 + x_4x_0x_1x_2x_3 + x_4x_1^2x_2^2 $. Following the same procedure as in MO2-I, the set of semistable families are given below.
1. $ x_4x_0^2x_3^2 + x_4x_0x_1x_2x_3 $
1. Destabilizing 1-PS: $\langle 1,2,0,0,0 \rangle$
2. $ x_4x_0x_1x_2x_3 + x_4x_1^2x_2^2 $
1. Destabilizing 1-PS: $\langle 1,0,0,0,-1 \rangle$
3. $ x_4x_0x_1x_2x_3 $
1. Destabilizing 1-PS: No 1-PS
The set of unstable families are given below.
1. $x_4x_1^2x_2^2$
2. $x_4x_0^2x_3^2$
Let $X$, up to a coordinate transformation, be of the form $x_4x_0^2x_3^2 + x_4x_0x_1x_2x_3 + x_4x_1^2x_2^2 $ .
1. If $X$ belongs to one of the families US1-VII - US2-VII then $X$ is unstable.
2. If $X$ is of type SS1-VII then the orbit is not closed and it degenerates to MO2-V.
3. If $X$ is of type SS2-VII then the orbit is not closed and it degenerates to MO2-V.
4. If $X$ is of type SS3-VII then the orbit is not closed and it degenerates to MO2-V.
Otherwise, $X$ is a closed orbit.
### MO2-VIII
In the case of family $MO2-VIII$, the centralizer ${\mathbb{C}}^{*2} \times SL(3,{\mathbb{C}})$ on polynomials of the form $x_0x_1q_3(x_2,x_3,x_4) $. By following a similar procedure as in the $MO \hyph A$ case one can obtain the maximal semistable and unstable families with respect to the $Z_G(H)$ action. The poset analysis shows that the maximal semistable families are the following families below.
1. $ x_0x_1( x_2q_2(x_3,x_4) + q_3(x_3,x_4)) $
1. Destabilizing 1-PS: $\langle 1,-1,2,-1,-1 \rangle$
2. $ x_0x_1( q_2(x_2,x_3)x_4 + q_1(x_2,x_3)x_4^2 + x_4^3 ) $
1. Destabilizing 1-PS: $\langle 1,-1,1,1,-2 \rangle$
The set of unstable families are given below.
1. $x_0x_1 \bigg( q_3(x_3,x_4)+x_2x_4^2 \bigg)$
Let $X$, up to a coordinate transformation, be of the form $ x_0x_1q_3(x_2,x_3,x_4) $ .
1. If $X$ belongs to one of the families US1-VIII then $X$ is unstable.
2. If $X$ is of type SS1-VIII then the orbit is not closed and it degenerates to MO2-IV.
3. If $X$ is of type SS2-VIII then the orbit is not closed and it degenerates to MO2-IV.
Otherwise, $X$ is a closed orbit.
### MO2-IX
In the case of family $MO2-IX$, the centralizer ${\mathbb{C}}^{*} \times SL(2,{\mathbb{C}}) \times SL(2, {\mathbb{C}})$ on polynomials of the form $ x_0q_{2,2}(x_1,x_2 \parallel x_3,x_4) $. By following a similar procedure as in the $MO \hyph A$ case one can obtain the maximal semistable and unstable families with respect to the $Z_G(H)$ action. The poset analysis shows that the maximal semistable families are the following families below.
1. $x_0 ( x_1x_2 + x_2^2 \parallel x_3x_4 + x_4^2 )$
1. Destabilizing 1-PS: $\langle 0,1,-1,1,-1 \rangle$
The set of unstable families are given below.
1. $x_0 \bigg( x_2^2 \parallel x_3x_4 + x_4^2 \bigg)$
2. $x_0 \bigg( x_1x_2 + x_2^2 \parallel x_4^2 \bigg)$
3. $x_0 \bigg( x_2^2 \parallel x_4^2 \bigg)$
Let $X$, up to a coordinate transformation, be of the form $ x_0q_{2,2}(x_1,x_2 \parallel x_3,x_4)$.
1. If $X$ belongs to one of the families US1-IX - US3-IX then $X$ is unstable.
2. If $X$ is of type SS1-IX then the orbit is not closed and it degenerates to MO2-V.
Otherwise, $X$ is a closed orbit.
### MO2-X
In the case of family $MO2-X$, the centralizer ${\mathbb{C}}^{*2} \times SL(2,{\mathbb{C}}) \times {\mathbb{C}}^* $ on polynomials of the form $ x_0 ( q_4(x_2,x_3) + x_1q_2(x_2,x_3)x_4 + x_1^2x_4^2 ) $ . By following a similar procedure as in the $MO \hyph A$ case one can obtain the maximal semistable and unstable families with respect to the $Z_G(H)$ action. The poset analysis shows that the maximal semistable families are the following families below.
1. $ x_0 ( q_4(x_2,x_3) + x_1 ( x_2x_3 + x_3^2 ) x_4 + x_1^2x_4^2 ) $
1. Destabilizing 1-PS: $\langle a,b,1,-1,c \rangle$
2. $ x_0 ( q_4(x_2,x_3) + x_1 (x_2x_3 + x_2^2 ) x_4 + x_1^2x_4^2 ) $
1. Destabilizing 1-PS: $\langle a,b,1,-1,c \rangle$
The set of unstable families are given below.
1. $ x_0 \bigg( x_2^2x_3^2+x_2x_3^3+x_3^4 + x_1x_3^2x_4 + x_1^2x_4^2 \bigg) $
2. $ x_0 \bigg( x_2x_3^3+x_3^4 + x_1(x_2x_3+x_3^2)x_4 + x_1^2x_4^2 \bigg) $
Let $X$, up to a coordinate transformation, be of the form $ x_0 ( q_4(x_2,x_3) + x_1q_2(x_2,x_3)x_4 + x_1^2x_4^2 ) $ .
1. If $X$ belongs to one of the families US1-X - US2-X then $X$ is unstable.
2. If $X$ is of type SS1-X then the orbit is not closed and it degenerates to MO2-V.
3. If $X$ is of type SS2-X then the orbit is not closed and it degenerates to MO2-V.
Otherwise, $X$ is a closed orbit.
Boundary Stratification
-----------------------
From the results above, the most degenerate point in the GIT compactification is the normal crossing singularities hypersurface $x_0x_1x_2x_3x_4$. The following chart show how the various degenerations occur on the boundary of the GIT compactification. The non-closed orbits of families $MO\hyph A - MO \hyph D$ degenerate further into the families $MO2 \hyph I - MO2 \hyph X$, which all eventually degenerate to the family of $x_0x_1x_2x_3x_4$.
Code for Poset Structure {#appendix:posetcode}
========================
Sample Linear Programming Calculation {#appendix:linearprogram}
=====================================
![Poset structure of quintic monomials[]{data-label="fig:calabiyau"}](calabiyau.eps)
![Poset structure of family SS1 []{data-label="fig:SS1"}](family30020.eps)
![Poset structure of family SS2 []{data-label="fig:SS2"}](family40001.eps)
![Poset structure of family SS3 []{data-label="fig:SS3"}](family20300.eps)
![Poset structure of family SS4 []{data-label="fig:SS4"}](family14000.eps)
![Poset structure of family SS5 []{data-label="fig:SS5b"}](familyss5-2.eps)
![Poset structure of family SS6 []{data-label="fig:SS6b"}](familyss6-3.eps)
![Poset structure of family SS7 []{data-label="fig:SS7b"}](familyss7-2.eps)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Let $H$ be a graph with $\Delta(H) \leq 2$, and let $G$ be obtained from $H$ by gluing in vertex-disjoint copies of $K_4$. We prove that if $H$ contains at most one odd cycle of length exceeding $3$, or if $H$ contains at most $3$ triangles, then $\chi(G) \leq 4$. This proves the *Strong Coloring Conjecture* for such graphs $H$. For graphs $H$ with $\Delta=2$ that are not covered by our theorem, we prove an approximation result towards the conjecture.'
address:
- 'Department of Mathematics and Statistics, Auburn University, Auburn, AL 36849'
- 'Department of Mathematics and Statistics, Auburn University, Auburn, AL 36849'
author:
- Jessica McDonald
- 'Gregory J. Puleo'
bibliography:
- 'biblio.bib'
title: 'Strong coloring $2$-regular graphs: cycle restrictions and partial colorings'
---
Introduction
============
In this paper all graphs are assumed to be simple, unless explicitly stated otherwise. Here we are primarily concerned the with the chromatic number of a graph obtained by gluing copies of $K_4$ onto a 2-regular graph. To contextualize this, consider the following general question.
\[ques:Kt\] Suppose that $G$ is obtained from a 2-regular graph $H$ by gluing in some number of vertex disjoint cliques of size $t$ (i.e. choosing vertex-disjoint sets of size $t$ and adding edges so that each of these sets induce $K_t$). Is $\chi(G)\leq t$?
When $H$ is a cycle and $t=3$, Question \[ques:Kt\] is the famous “cycle plus triangles problem” popularized by Erdős and resolved affirmatively by Fleischner and Stiebitz [@cycletri] (see [@cycletri] for more on the history of this particular problem). The result for $t=3$ does not hold for all 2-regular $H$; in particular a $C_4$ component in $H$ can allow $K_4$ to be created after gluing in triangles, and Fleischner and Stiebitz [@fleischner-remarks] found an infinite family of counterexamples without such $C_4$ components as well, answering a further question of Erdős [@erdos-fav].
Haxell [@haxell-strong] has answered Question \[ques:Kt\] affirmatively whenever $t\geq 5$ (in fact she proved something stronger, as we shall discuss shortly). Question \[ques:Kt\] remains open for $t=4$ however. There is an easy affirmative argument for $t=4$ when $H$ consists only of cycles of length 3 or 4, and the problem can also be resolved positively when $H$ has girth at least 4 (see Pei [@Pei]). In the present paper we step into the intermediate ground, where $H$ has both triangles and longer odd cycles, and prove the following result.
\[thm:main-cycles\] Let $H$ be a graph with $\Delta(H) \leq 2$, and let $G$ be obtained from $H$ by gluing in vertex-disjoint copies of $K_4$. If $H$ contains at most one odd cycle of length exceeding $3$, or if $H$ contains at most $3$ triangles, then $\chi(G) \leq 4$.
We also prove the following approximation result for graphs $H$ not dealt with by the above theorem.
\[thm:main-fraction\] Let $H$ be a graph with $\Delta(H) \leq 2$, and let $G$ be obtained from $H$ by gluing in vertex-disjoint copies of $K_4$. There is a set of vertices $Z$ with ${\left\lvert{Z}\right\rvert} \leq {\left\lvert{V(G)}\right\rvert}/22$ such that $\chi(G-Z) \leq 4$.
Theorem \[thm:main-cycles\] actually proves the *Strong Coloring Conjecture* for graphs $H$ of the sort we describe in the theorem. Given an $n$-vertex graph $H$ where $n$ is divisible by $t$, we say that $H$ is *strongly $t$-colorable* if, for any partition of $V(H)$ into parts of size $t$, $H$ has a $t$-coloring where each color class is a transversal of the partition (i.e. where each color class contains exactly one vertex from each part of the partition). In the case where $n$ is not divisible by $t$, we say that $H$ is *strongly t-colorable* if $H'$ is strongly $t$-colorable, where $H'$ is obtained from $H$ by adding $t\lceil\tfrac{n}{t}\rceil-n$ isolated vertices (the minimum amount to ensure divisibility by $t$). The notion of strong coloring was introduced independently by Alon [@Al] and Fellows [@Fel] about thirty years ago.
In the definition of strong coloring, instead of requesting that each color class is a transversal of the partition, we can equivalently ask for a copy of $K_t$ to be glued to each part of the partition, and then ask for the resulting graph to be $t$-colorable. Hence, given an $n$-vertex graph $H$ with $t\mathrel{|}n$, Question \[ques:Kt\] is exactly asking whether or not $H$ is strongly $t$-colorable. However, if $t\mathrel{\not|} n$, an affirmative answer to Question \[ques:Kt\] may not imply strong $t$-colorability. In particular, Fleischner and Stiebitz’s [@cycletri] result implies that cycles with length divisible by 3 are strongly 3-colorable, but it is not true that all cycles have this property (since by adding $K_3$’s to $C_4$ plus two necessary isolates we can create $K_4$). On the other hand, since we only require $\Delta(H)\leq 2$ in Theorem \[thm:main-cycles\], we get the following as an immediate corollary.
\[cor:strong\] Let $H$ be a graph with $\Delta(H) \leq 2$ which either contains at most one odd cycle of length exceeding $3$, or contains at most $3$ triangles. Then $H$ is strongly 4-colorable.
It is not obvious that a strongly $t$-colorable graph is necessarily strongly $(t+1)$-colorable, but in fact this can be shown using a short argument due to Fellows [@Fel]. Given this, it makes sense to define the *strong chromatic number* of $H$, $s\chi(H)$, as the minimum $t$ such that $H$ is strongly $t$-colorable. Note that for any graph $H$, $s\chi(H)\geq \Delta+1,$ since if a clique was added to the neighborhood of a $\Delta$-vertex, a copy of $K_{\Delta+1}$ would be created in the new graph, and obviously that is not $\Delta$-colorable. The previously-alluded to result by Haxell [@haxell-strong] says that for any graph $H$, $s\chi(H)\leq 3\Delta-1$. When $\Delta=2$ this says that $s\chi(H)\leq 5$ (hence answering Question \[ques:Kt\] affirmatively for $t\geq5$). Fleischner and Stiebitz [@fleischner-remarks] have given, for each $\Delta$, an example of a $\Delta$-regular graph $H$ for which $s\chi(H) \geq 2\Delta$, and hence the following conjecture would be best possible if it is true. (Attribution for this conjecture is somewhat tricky – according to [@ABS], it may have first appeared explicitly in a 2007 paper by Aharoni, Berger and Ziv [@ABZ] after being “folklore” for a while, although it could also be considered implicit in the 2004 paper of Haxell [@haxell-strong].)
\[conj:scc\] For any graph $H$, $s\chi(H)\leq 2\Delta(H)$.
The Strong Coloring Conjecture is trivial for $\Delta=1$, where it asks essentially for the union of two matchings to be bipartite. The conjecture remains open for all $\Delta \geq 2$, but several partial results are known for general $\Delta$. Haxell [@haxell-improved] proved that $s\chi(H) \leq (\tfrac{11}{4}+\varepsilon)\Delta(H)$ when $\Delta(H)$ is sufficiently large, improving her general upper bound of $s\chi(H) \leq 3\Delta(H)-1$. Aharoni, Berger and Ziv [@ABZ] proved a fractional-coloring version of the conjecture. Lo and Sanhueza–Matamala [@LS] proved the following asymptotic version of the conjecture: for any constant $c > 0$ and any graph $H$ with $\Delta(H) \geq c{\left\lvert{V(H)}\right\rvert}$, one can obtain an upper bound of the form $s\chi(H) \leq (2+o(1))\Delta(H)$. (Here the $o(1)$ term may depend on the constant $c$.)
In contrast to many of the above results, which weaken the conclusion of Conjecture \[conj:scc\] in some form or another, our Corollary \[cor:strong\] proves the exact conclusion of Conjecture \[conj:scc\] in several new cases. Corollary \[cor:strong\] improves previous work by Fleischner and Stiebitz [@fleischner-remarks] who (separately from their cycle + triangles solution) verified the conjecture for all cycles $H$. Our result also strengthens the previously discussed results of Pei [@Pei], who verified Conjecture \[conj:scc\] when $H$ consists only of cycles of length 3 or 4, or has girth at least 4.
If $H$ is a graph and $V_1, \ldots, V_n$ are disjoint subsets of $V(H)$, an *independent set of representatives* of $(V_1, \ldots, V_n)$ is an independent set $R$ containing exactly one vertex from each set $V_i$. Hence in the definition of strongly $t$-colorable, we may replace “$H$ has a $t$-coloring where each color class is a transversal of the partition” with “$H$ has $t$ disjoint ISRs of the partition”. The following theorem of Haxell [@haxell-maxdeg] (proved for a general $\Delta$ but stated here for the case $\Delta=2$) guarantees the existence of one ISR (where Conjecture \[conj:scc\] asks for four).
\[thm:haxell-maxdeg\] If $H$ is a graph with $\Delta(H)\leq 2$ and $V_1, \ldots, V_n$ are disjoint subsets of $V(H)$ with each ${\left\lvert{V_i}\right\rvert} \geq 4$, then $(V_1, \ldots, V_n)$ has an ISR.
We shall use Theorem \[thm:haxell-maxdeg\] in our proof of Theorem \[thm:main-fraction\]. We shall also prove the following extension of Theorem \[thm:haxell-maxdeg\], which serves as another approximation towards Conjecture \[cor:strong\] when $\Delta=2$.
\[thm:main-2isr\] If $H$ is a graph with $\Delta(H)\leq 2$ and $V_1, \ldots, V_n$ are disjoint subsets of $V(H)$ with each ${\left\lvert{V_i}\right\rvert} = 4$, then $(V_1, \ldots, V_n)$ has two disjoint ISRs.
It it worth noting that when $\Delta(H)\geq 3$ (and each ${\left\lvert{V_i}\right\rvert} \geq 2\Delta(G)$) Aharoni, Berger and Sprüssel [@ABS] had already proved the conclusion of Theorem \[thm:main-2isr\] (as a consequence of a more general result about matroids). It was also observed by Haxell (personal communication) that the $\Delta(H) = 2$ case follows quickly from a strengthened version of Theorem \[thm:haxell-maxdeg\]. Our approach has the advantage of being elementary and algorithmic, in contrast to the topological tools needed for the strengthening of Theorem \[thm:haxell-maxdeg\]. See [@ABZ] for more on various known results for ISRs, including the strengthened versions of Theorem \[thm:haxell-maxdeg\].
The rest of the paper is organized as follows. In Section \[sec:long-odd\] we show how to $4$-color a large fraction of the vertices when $H$ has few long odd cycles. In Section \[sec:isr-lems\] we prove two lemmas about ISRs that are needed for Section \[sec:few-tris\], in which we show how to $4$-color a large fraction of the vertices when $H$ has few triangles. Taking extreme cases of the results of Section \[sec:long-odd\] and Section \[sec:few-tris\] gives Theorem \[thm:main-cycles\]; combining the results gives Theorem \[thm:main-fraction\], which we also show in Section \[sec:few-tris\]. In Section \[sec:two-isrs\], we prove Theorem \[thm:main-2isr\].
Graphs with few long odd cycles {#sec:long-odd}
===============================
Our goal in this section is to prove the following theorem.
\[thm:longodd-910\] Let $H$ be a graph with $\Delta(H) \leq 2$, and let $G$ be a graph obtained from $H$ by gluing in vertex-disjoint copies of $K_4$. Let ${\mathcal{C}}$ be the set of odd cycles in $H$ with length exceeding $3$, and let $V({\mathcal{C}})$ be the set of vertices contained in these cycles. There is a set of vertices $Z {\subseteq}V({\mathcal{C}})$ with ${\left\lvert{Z}\right\rvert} \leq {\left\lvert{{\mathcal{C}}}\right\rvert}/2$ such that $G-Z$ is $4$-colorable and $Z$ contains at most one vertex from each cycle of ${\mathcal{C}}$. In particular, ${\left\lvert{V(G) - Z}\right\rvert} \geq (9/10){\left\lvert{V(G)}\right\rvert}$.
Before proving Theorem \[thm:longodd-910\], we observe that it implies one case of Theorem \[thm:main-cycles\].
\[cor:onelongoddcycle\] Let $G$ be a graph with $\Delta(H) \leq 2$, and let $G$ be a graph obtained from $H$ by gluing in vertex-disjoint copies of $K_4$. If $H$ has at most one odd cycle of length exceeding $3$, then $\chi(G) \leq 4$.
We will need the following lemma about equitable coloring in our proof of Theorem \[thm:longodd-910\].
\[lem:equitable\] If $G$ is a graph and $k > \Delta(G)$, then $G$ has a proper $k$-coloring with color classes $A_1, \ldots, A_k$ such that ${\left\lvert{A_i - A_j}\right\rvert} \leq 1$ for all $i,j$.
We start by showing that it suffices to prove the result for problem instances that satisfy the following additional assumptions:
(1) The vertices of each added copy of $K_4$ form an independent set in $H$,
(2) $H$ is $2$-regular, and
(3) The added copies of $K_4$ partition $V(H)$.
Let $H$ be any graph with $\Delta(H) \leq 2$ and let $Y_1, \ldots, Y_q$ be the vertex sets of added copies of $K_4$. We will modify $H$ and $Y_1, \ldots, Y_q$ to guarantee each of the properties (1)–(3) in turn, taking care not to invalidate earlier properties when establishing later ones, and taking care not to increase the size of ${\mathcal{C}}$. At each step, we will observe that a $4$-coloring of the modified graph implies $4$-colorability of the original $G$.
(1) Let $H^1$ consist of the graph $H$ where all edges within each $Y_j$ have been deleted, and let $G^1$ consist of $H^1$ with copies of $K_4$ glued in on the vertex sets $Y_1, \ldots, Y_j$. The only edges deleted in passing from $H$ to $H^1$ are restored when we glue in the $K_4$’s, so $G^1 = G$. Clearly, deleting edges cannot increase the size of ${\mathcal{C}}$.
(2) Enlarge $H^1$ to a larger graph $H^2$ by adding a set of new vertices $S$ and adding edges, as necessary, between $V(H^1)$ and $S$ and within $S$, to ensure that all vertices have degree $2$ (taking care not to create any new odd cycles). Let $G^2$ consist of $H^2$ with copies of $K_4$ glued in on the sets $Y_1, \ldots, Y_q$. We see that $G^2$ contains $G^1$ as an induced subgraph (since no new edges are added within $V(H^1)$), hence $4$-colorability of $G^2$ implies $4$-colorability of $G^1$. Furthermore, since all $Y_j \subseteq V(H_1)$, we see that Property (1) still holds. Since all new cycles created in this manner are even cycles, ${\left\lvert{{\mathcal{C}}}\right\rvert}$ has not increased in this step.
(3) Let $J$ be the subgraph of $H^2$ induced by the vertices not covered by $Y_1, \ldots, Y_q$. Let $J'$ be the disjoint union of $J$ and $t$ copies of $K_3$, where $t$ is chosen so that $J'$ has at least $12$ vertices and ${\left\lvert{V(J')}\right\rvert}$ is divisible by $4$. (Since $3$ and $4$ are coprime, such a $t$ can always be found.)
Let $k = {\left\lvert{V(J')}\right\rvert}/4$. By our choice of $t$, we see that $k$ is an integer with $k \geq 3 > \Delta(J')$. Hence, by Lemma \[lem:equitable\], we see that $J'$ has a $k$-coloring with color classes $A_1, \ldots, A_k$ such that ${\left\lvert{A_i - A_j}\right\rvert} \leq 1$. Since ${\left\lvert{V(J')}\right\rvert}$ is divisible by $k$, this implies that all ${\left\lvert{A_i}\right\rvert} = {\left\lvert{V(J')}\right\rvert}/k = 4$. These color classes are independent sets that we will be able to glue new copies of $K_4$ onto.
Let $H^3$ be the disjoint union of $H^2$ and $t$ copies of $K_3$ (with the latter having the same vertex set as those we added in passing from $J$ to $J'$). Let $Y'_1, \ldots, Y'_{q'}$ consist of the original sets $Y_1, \ldots, Y_q$ together with the color classes $A_1, \ldots, A_k$ from the coloring of $J'$. Let $G^3 = H^3$ with copies of $K_4$ glued in on the sets $Y'_1, \ldots, Y'_{q'}$. As $G^3$ contains $G^2$ as an induced subgraph, $4$-colorability of $G^3$ implies $4$-colorability of $G^2$ (and hence $G$).
Now $Y'_1, \ldots, Y'_{q'}$ partition $V(H^3)$, and by our construction, each $Y'_j$ is an independent set in $H^3$. In passing from $H^2$ to $H^3$ we have maintained $2$-regularity, so Properties (1) and (2) still hold. As we have only added triangles in passing from $H^2$ to $H^3$, we also see that we have not increased ${\left\lvert{{\mathcal{C}}}\right\rvert}$.
For the remainder of this proof, we will assume that Properties (1)–(3) hold. We will view the edges of $G$ (and its subgraphs) as being colored red and blue: all edges coming from the original graph $H$ will be colored red and all edges coming from the added copies of $K_4$ will be colored blue. (Obviously this is not a proper edge-coloring, as in $G$ every vertex is incident to two red edges and three blue edges.)
Let ${\mathcal{C}}_0$ be the set of all odd cycles in $H$ (including triangles), and let $V({\mathcal{C}}_0)$ be the set of vertices in these odd cycles. Observe that every component of $H- V({\mathcal{C}}_0)$ is an even cycle, hence $H$ has a matching (of red edges) that saturates $V(H) - V({\mathcal{C}}_0)$. Fix such a matching $M_0$.
Let $T$ be a transversal of $V({\mathcal{C}}_0)$ (i.e. $T$ is a set containing exactly one vertex from each cycle in ${\mathcal{C}}_0$). For $C \in {\mathcal{C}}_0$, we write $T(C)$ for the vertex of $C$ contained in $T$, and for a transversal vertex $t$, we write $C(t)$ for the cycle of ${\mathcal{C}}_0$ containing $t$. Observe that for any transversal $T$, there is a unique matching that extends $M_0$ and saturates $V(G) - T$. Let $M(T)$ denote this unique matching. (While the base matching $M_0$ is arbitrary, we use the same choice of $M_0$ for all transversals $T$ when defining $M(T)$.)
Let $J$ be an arbitrary perfect matching of the blue edges in $G$, and for any transversal $T$, let $B(T) = M(T) \cup J$, considered as a subgraph of $G$. (As before, while the choice of $J$ is arbitrary, we use the same $J$ for all $T$.) Inheriting the edge coloring from $G$, we observe that every vertex $v \in B(T)$ is incident to exactly one blue edge and either exactly one red edge (if $v \notin T$) or to zero red edges (if $v \in T$). In particular, $B(T)$ is bipartite, and its components consist of some number of even cycles together with ${\left\lvert{{\mathcal{C}}}\right\rvert}/2$ paths whose endpoints are vertices in $T$.
Now, among all the possible transversals $T$, we will choose an “optimal” transversal $T^*$. Our selection proceeds in two stages. First, among all transversals $T$, choose $T_1$ to minimize the sum of the lengths of all path components in $B(T_1)$. Call any transversal achieving this minimum a *semi-optimal* transversal.
Let $T_1$ be a semi-optimal transversal, let $t_1 \in T_1$, and let $Q$ be the component of $B(T_1)$ containing $t_1$. There is at least one vertex in the odd $H$-cycle $C(t_1)$ that is not contained in $Q$.
Let $C_1 = C(t_1)$, and suppose to the contrary that every vertex of $C_1$ lies in $Q$. Observe that $Q$ is a path with $t_1$ as one endpoint and another vertex of $T_1$ as the other endpoint. Let $t_2$ be the other endpoint of $Q$, so that $Q$ is a $(t_1,t_2)$-path.
Let $t'$ be the last vertex of $C_1$ along the $(t_1,t_2)$-path $Q$. Observe that since $Q$ contains every vertex of $C_1$, we have $t' \neq t_1$. Since $t'$ is the last vertex of $C_1$ along this path, the $(t',t_2)$-subpath does not contain any edges from $C_1$.
Let $T'$ be the transversal obtained from $t_1$ by replacing $t_1$ with $t'$. We claim that the path-components of $B(T')$ have a smaller sum of lengths than the path-components of $B(T_1)$, contradicting the choice of $B(T_1)$.
First observe that the only edges that lie in $B(T')$ but not $B(T_1)$, or vice versa, are edges from the cycle $C_1$, since this is the only cycle whose transversal-representative has changed. Since every vertex of $C_1$ lies in $Q$, no vertex outside $Q$ has gained or lost any incident edges in passing from $B(T_1)$ to $B(T')$. In particular, for every $t \in T - \{t_1, t_2\}$, the path-component containing $t$ is identical in $B(T')$ and $B(T_1)$.
Since the $(t',t_2)$-subpath of $Q$ did not use any edges of $C_1$, we see that it is still present in $B(T')$. However, the length of the $(t',t_2)$-path component in $B(T')$ is strictly less than the length of the $(t_1,t_2)$-path component in $B(T_1)$, with all other path components having the same length in both graphs. This contradicts the choice of $T_1$ and completes the proof of the claim.
Next, we refine our choice among the semi-optimal transversals. To define our optimality criterion, first fix a cyclic orientation of each cycle in ${\mathcal{C}}_0$. For each $v \in V({\mathcal{C}}_0)$, let $v^+$ be the successor of $v$ in this cyclic orientation, and for $v,w$ in the same cycle, let $d^+(v,w)$ be the “directed $H$-distance” from $v$ to $w$ along the directed $H$-edges.
For a transversal $T$ and a vertex $t \in T$, let $Q$ denote the component of $B(T)$ containing $t$ (note that $Q$ is a path). Say $t \in T$ is *bad* if $t^+ \in Q$. The *cost* of a bad vertex $t \in T$ is the minimum directed $H$-distance $d^+(t,v)$ for $v \in C(t) - Q$, and the cost of a non-bad vertex $t \in T$ is $0$. (Note that for semioptimal $T$, the set $C(t) - Q$ is nonempty, hence the cost of a bad vertex is always finite.)
Define the cost of a semi-optimal transversal $T$ to be the sum of the costs of the vertices in $T$. Among all semi-optimal transversals, choose $T^*$ to have minimum cost.
\[claim:nobb\] Let $T_1$ be a semi-optimal transversal, let $t_1 \in T_1$, and let $Q$ be the component of $B(T_1)$ containing $t_1$. If $t_1^+ \in Q$, then the unique $(t_1,t_1^+)$-path in $Q$ starts with a blue edge and ends with a red edge.
Since $t_1$ has no incident red edges in $B(T_1)$, the only other possibility is that the $(t_1,t_1^+)$-path starts and ends with a blue edge, as shown in Figure \[fig:t1plus\](a). As $t_1^+$ is covered by $M(T_1)$ and therefore has an incident red edge in $B(T_1)$, we see that the vertex following $t_1^+$ in $Q$ is its other $H$-neighbor, namely $t_1^{++}$. We know that the other endpoint of $Q$ is another vertex of $T_1$; let $t_2 \in T_1$ be the other endpoint of $Q$.
at (1cm, .5cm) [(a)]{}; iin [1,...,5]{} [ (ai) at (2\*icm, 0cm) ; (bi) at (1cm + 2\*icm, 0cm) ; ]{} at (a1) ; at (a4) ; at (b3) ; at (b5) ;
iin [1,...,5]{} [ (ai) – (bi); ]{} iin [1,...,4]{} [ ; (bi) – (a); ]{} (a1) .. controls ++(45:1.5cm) and ++(135:1.5cm) .. (b3);
at (1cm, .5cm) [(b)]{}; iin [1,...,5]{} [ (ai) at (2\*icm, 0cm) ; (bi) at (1cm + 2\*icm, 0cm) ; ]{} at (a1) ; at (a4) ; at (b3) ; at (b5) ;
iin [1,...,5]{} [ (ai) – (bi); ]{} iin [1,...,4]{} [ ; (bi) – (a); ]{} (a1) .. controls ++(45:1.5cm) and ++(135:1.5cm) .. (a4);
Let $t' = t_1^{++}$ and let $T'$ be the transversal obtained from $T_1$ by replacing $t_1$ with $t'$. As shown in Figure \[fig:rotate-matching\] (see also Figure \[fig:t1plus\](a)), the only effect of this replacement on the matching $M(T_1)$ is to replace the matching-edge $t_1^+t_1^{++}$ with the matching-edge $t_1t_1^+$. This splits the component $Q$ into a cycle containing $t_1$ and $t_1^+$ and a shorter $(t_1^{++},t_2)$-path and has no effect on the other components of $B(T_1)$. Since the sum of the lengths of the path components in $B(T')$ is shorter than the sum for $B(T_1)$, this contradicts the choice of $T_1$ as semi-optimal.
$T^*$ has no bad vertices.
Suppose to the contrary that $t_1 \in T^*$ is a bad vertex, and let $Q$ be the component of $B(T^*)$ containing $t_1$. We know that $Q$ is a path whose other endpoint is another vertex of $T^*$; let $t_2 \in T^*$ be the other endpoint of $Q$. Let $v$ be the vertex of $C(t_1)-Q$ minimizing $d^+(t_1, v)$, so that $d^+(t_1, v)$ is the cost of the bad vertex $t_1$.
Since $t_1$ is bad, we have $t_1^+ \in Q$. By Claim \[claim:nobb\], the unique $(t_1, t_1^+)$-path in $Q$ starts with a blue edge and ends with a red edge, as shown in Figure \[fig:t1plus\](b). In particular, the vertex preceding $t_1^+$ in this path is the other $H$-neighbor of $t_1^+$, namely $t_1^{++}$. Thus, $d^+(t_1, v) \geq 3$.
Let $t' = t_1^{++}$ and let $T'$ be the transversal obtained from $T^*$ by replacing $t_1$ with $t'$. As before, the only effect of this replacement on the matching $M(T^*)$ is to replace the matching-edge $t_1^+t_1^{++}$ with the matching-edge $t_1t_1^+$; see Figure \[fig:rotate-matching\]. In particular, this replacement does not alter the length of the path-component containing $t$ for any $t \in T^* - \{t_1,
t_2\}$. Furthermore, after removing the edge $t^+_1t'_1$ and adding the edge $t^+_1t_1$, we see that $B(T')$ has a $(t',t_2)$-path using exactly the vertices of $Q$, obtained by starting at $t'$, traversing $Q$ backwards until $t_1$, taking the new edge $t_1t^+_1$, and then completing the rest of the path from $t^+_1$ to $t_2$ (see Figure \[fig:t1plus\](b)). Hence, the sum of the lengths of the path-components in $B(T')$ is the same as the sum of the lengths in $B(T^*)$, that is, $T'$ is also semi-optimal.
iin [1,...,7]{} [ ; (ai) at (: 1cm) ; ]{} at (a1) ; at (a7) ; at (a6) ; (a2) – (a3); (a4) – (a5); (a6) – (a7); (a1) – (a7); (a2) – (a1); (a4) – (a3); (a6) – (a5); [(a1) circle (5pt)]{};
iin [1,...,7]{} [ ; (ai) at (: 1cm) ; ]{} at (a1) ; at (a7) ; at (a6) ; (a2) – (a3); (a4) – (a5); (a1) – (a7); (a7) – (a6); (a2) – (a1); (a4) – (a3); (a6) – (a5); [(a6) circle (5pt)]{};
at (0cm, 0cm) [$\Longrightarrow$]{};
Observe that since we replaced $t_1$ with $t' = t_1^{++}$ and since $d^+(t_1, v) \geq 3$, we have $d^+(t', v) \leq d^+(t_1, v) - 2$. Thus, the cost of $t'$ is strictly less than the cost of $t_1$ (regardless of whether $t'$ is bad). Furthermore, we have not altered any components of $B(T^*)$ except for the component $Q$, so every other vertex of $T'$ has the same cost it did in $T^*$. It follows that $T'$ is a semi-optimal transversal having lower cost than $T^*$, contradicting the choice of $T^*$.
Now we use the optimal transversal $T^*$ to produce the desired coloring. First we will randomly produce a $2$-coloring of $T$ using the colors black and white, then we will use the black-and-white coloring to $4$-color most of $G$. Each component of $B(T^*)$ is bipartite. Obtain a random black-and-white coloring of $B(T^*)$ by randomly choosing one of the two possible black-and-white colorings for each component independently and with equal probability. Let $\phi$ be the resulting coloring.
For any vertex $t \in T^*$, ${\mathbb{P}}[\phi(t) = \phi(t^+)] = 1/2$.
By the previous claim, $t$ is not a bad vertex, so $t$ and $t^+$ are in different components of $B(T^*)$. As the colorings on these components are chosen independently, this implies that ${\mathbb{P}}[\phi(t) = \phi(t^+)] = 1/2$.
Now we restrict our attention to the vertices of $T^*$ that lie in ${\mathcal{C}}$, disregarding the vertices of $T^*$ that lie in triangles. Say that a vertex $t \in T^* \cap V({\mathcal{C}})$ is *unhappy* if $\phi(t) = \phi(t^+)$. By the claim and by linearity of expectation, the expected number of unhappy $T^*$-vertices in a random coloring is at most ${\left\lvert{T^* \cap V({\mathcal{C}})}\right\rvert}/2 = {\left\lvert{{\mathcal{C}}}\right\rvert}/2$. Hence, there is a coloring $\phi^*$ with at most ${\left\lvert{{\mathcal{C}}}\right\rvert}/2$ unhappy vertices in $T^* \cap V({\mathcal{C}})$.
Let $Z$ be the set of unhappy vertices for $\phi^*$. Evidently, ${\left\lvert{Z}\right\rvert} \leq {\left\lvert{{\mathcal{C}}}\right\rvert}/2$, with $Z$ containing at most one vertex from each cycle in ${\mathcal{C}}$. We claim that $G-Z$ is $4$-colorable. Let $W_1$ be the set of vertices colored black in $G-Z$, let $W_2$ be the set of vertices colored white in $G-Z$, and let $G_i$ be the induced subgraph $G[W_i]$ for $i\in\{1, 2\}$.
Let $i \in \{1,2\}$. Every vertex of $G_i$ is incident, within $G_i$, to at most one red edge and one blue edge.
First suppose that some $w \in V(G_i)$ were incident with two blue edges within $G_i$, say $wx$ and $wy$. So, $w, x, y$ all received the same color under $\phi^*$. Since $B(T^*)$ is properly 2-colored by $\phi^*$, neither of $wx$ or $wy$ is in $B(T^*)$. Since $w, x, y$ must all be part of the same (blue) $K_4$, and $B(T^*)$ contains a perfect (blue) matching of this $K_4$, we know that $xy$ is in $B(T^*)$. But this is a contradiction, since $x$ and $y$ receive the same color under $\phi^*$.
Next suppose that some $w \in V(G_i)$ were incident with two red edges within $G_i$, say $wx$ and $wy$. If $w \notin T^*$, then one of the edges $wx$ and $wy$ must have appeared in $B(T^*)$, hence its endpoints must have received opposite colors in $B(T^*)$, contradicting that edge lying within $G_i$. However, if $w \in T^*$ we must be slightly more subtle.
If $w \in T^* \cap V({\mathcal{C}})$, then since $w \notin Z$, we see that $w$ is not unhappy under $\phi^*$. Hence, $\phi^*(w) \neq \phi^*(w^+)$, so the red edge $ww^+$ is not contained in $G_i$, meaning $w$ is not incident to two red edges in $G_i$. On the other hand, if $w \in T^* - V({\mathcal{C}})$, then $w$ lies in some triangle $wxy$ of $H$, so that $M(T^*)$ must contain the edge $xy$. Hence $x$ and $y$, the two neighbors of $w$ along red edges, receive opposite colors in the proper $2$-coloring of $B(T^*)$, meaning that $w$ is incident to exactly one red edge in $G_i$, contradiction.
The above claim tells us that each $G_i$ inherits a proper 2-edge-coloring in red and blue, and hence is bipartite. Since $V(G_1) \cup V(G_2) = V(G) - Z$, it follows that $\chi(G-Z) \leq 4$. Since $|Z|\leq \tfrac{|{\mathcal{C}}|}{2} \leq \tfrac{(|V(G)|/5)}{2}$, observe that $|V(G)-Z|\geq \tfrac{9}{10} |V(G)|$.
Two lemmas about ISRs {#sec:isr-lems}
=====================
A *total dominating set* in a graph $G$ is a set of vertices $X$ such that every vertex in $G$ is adjacent to a vertex in $X$. (In particular, every vertex of $X$ must also have a neighbor in $X$.) The *total domination number* of $G$, written ${\bar{\gamma}}(G)$, is the size of a smallest total dominating set; if $G$ has isolated vertices, then by convention we take ${\bar{\gamma}}(G) = \infty$.
Given a graph $H$ and disjoint subsets $V_1, \ldots, V_n {\subseteq}V(H)$, for each $S {\subseteq}[n]$ we define a subgraph $H_S$ by taking the subgraph induced by the vertex set $\bigcup_{i \in S}V_i$ and deleting all edges inside of each set $V_i$.
Using the above definitions, we can now state the following result of Haxell [@haxell-dom].
\[thm:haxell\] Let $H$ be a graph and let $V_1, \ldots, V_n$ be disjoint subsets of $V(H)$. If, for all $S {\subseteq}[n]$, we have ${\bar{\gamma}}(H_S) \geq 2{\left\lvert{S}\right\rvert} - 1$, then $(V_1, \ldots, V_n)$ has an ISR.
Theorem \[thm:haxell\] was originally stated in terms of hypergraphs (see also [@king-haxell] for a formulation not in terms of hypergraphs), and we have stated it above in a slightly modified but equivalent formulation.
The first of the two lemmas we will prove in this section is a deficiency version of Theorem \[thm:haxell\]: using weaker bounds on the size of total dominating sets, we can still obtain a “large” partial ISR. In particular, our proof will show that Theorem \[thm:haxell\] is “self-strengthening”, i.e., that the following can be obtained as a Corollary to Theorem \[thm:haxell\] itself.
\[lem:haxell-defic\] Let $H$ be a graph, let $V_1, \ldots, V_n$ be disjoint subsets of $V(H)$, and let $k$ be a nonnegative integer. If, for all $S \subseteq [n]$, we have ${\bar{\gamma}}(H_S) \geq 2{\left\lvert{S}\right\rvert} - 1 - 2k$, then $(V_1, \ldots, V_n)$ has a partial ISR of size at least $n-k$.
Let $H'$ be the disjoint union of $H$ with $k$ copies of $K_n$, and let $V'_1, \ldots, V'_n$ be obtained from $V_1, \ldots, V_n$ by defining $V'_i$ to be $V_i$ together with one vertex from each copy of $K_n$, chosen so that $V'_1, \ldots, V'_n$ are disjoint.
Observe that if $(V'_1, \ldots, V'_n)$ has an ISR, then at most $k$ of the vertices from the ISR are from the added copies of $K_n$; the remaining $n-k$ vertices yield a partial ISR of $(V_1, \ldots, V_n)$ of size at least $n-k$, as desired. Thus, it suffices to show that $(V'_1, \ldots, V'_n)$ has an ISR.
To do this, we apply Theorem \[thm:haxell\]. Let $S$ be any subset of $[n]$. We will show that ${\bar{\gamma}}(H'_S) \geq 2{\left\lvert{S}\right\rvert}-1$. If ${\left\lvert{S}\right\rvert} \leq 1$ then there is nothing to show, so assume that ${\left\lvert{S}\right\rvert} \geq 2$.
Let $H'_1$ be the subgraph of $H'_S$ induced by the original vertices of $H$ and let $H'_2$ be the subgraph of $H'_S$ induced by vertices from the added copies of $K_n$. Since each $V_i$ has exactly one vertex from each added $K_n$, we see that $H'_2$ is isomorphic to $k$ copies of $K_{{\left\lvert{S}\right\rvert}}$.
As there are no edges joining $H'_1$ and $H'_2$, clearly ${\bar{\gamma}}(H') = {\bar{\gamma}}(H'_1) + {\bar{\gamma}}(H'_2)$. By hypothesis, ${\bar{\gamma}}(H'_1) \geq 2{\left\lvert{S}\right\rvert}-1-2k$, and since ${\bar{\gamma}}(K_t) = 2$ for all $t \geq 2$, we have ${\bar{\gamma}}(H'_2) = 2k$. It follows that ${\bar{\gamma}}(H') \geq 2{\left\lvert{S}\right\rvert}-1$, and by Theorem \[thm:haxell\], it follows that $(V'_1, \ldots, V'_n)$ has an ISR.
The second lemma we will show in this section lets us “combine” ISRs for two different families of disjoint sets, under suitable conditions. In order to state it, we require the following technical definition.
Let ${\mathcal{X}}$ and ${\mathcal{Y}}$ be two collections of vertex-disjoint subsets of $G$. (That is, all sets in ${\mathcal{X}}$ are pairwise vertex-disjoint, and all sets in ${\mathcal{Y}}$ are pairwise vertex-disjoint, but we make no disjointness requirements between sets in ${\mathcal{X}}$ and sets in ${\mathcal{Y}}$.) The pair $({\mathcal{X}}, {\mathcal{Y}})$ is *admissible* for $G$ if, for each edge $e \in E(G)$, one of the following holds:
- There is some $X \in {\mathcal{X}}$ with both endpoints of $e$ in $X$,
- There is some $Y \in {\mathcal{Y}}$ with both endpoints of $e$ in $Y$,
- Both endpoints of $e$ are missing from all $X \in {\mathcal{X}}$, or
- Both endpoints of $e$ are missing from all $Y \in {\mathcal{Y}}$.
\[lem:combine-isr\] Let $G$ be a graph, and let $({\mathcal{X}}, {\mathcal{Y}})$ be an admissible pair of collections of vertex-disjoint subsets of $G$. If ${\mathcal{X}}$ has an ISR $R_{{\mathcal{X}}}$ in $G$ and ${\mathcal{Y}}$ has an ISR $R_{{\mathcal{Y}}}$ in $G$, then $G$ has an independent set $R \subseteq R_{{\mathcal{X}}} \cup R_{{\mathcal{Y}}}$ that is a transversal of both ${\mathcal{X}}$ and of ${\mathcal{Y}}$.
Initially, let $R = R_{{\mathcal{X}}} \cup R_{{\mathcal{Y}}}$. (Note that it is possible that some vertices may lie in $R_{{\mathcal{X}}} \cap R_{{\mathcal{Y}}}$.) The set $R$ clearly hits every $X_i$ and $Y_j$, but as there may be edges between $R_{{\mathcal{X}}}$ and $R_{{\mathcal{Y}}}$, the set $R$ may not be independent. We next describe an algorithm for iteratively deleting vertices from $R$ in order to obtain an independent subset of $R$ that still hits every $X_i$ and every $Y_j$. (Note that some vertices of $R$ may lie in $R_{{\mathcal{X}}} \cap R_{{\mathcal{Y}}}$; such vertices are automatically isolated vertices in $R$, and we will not need to worry about those vertices.)
To describe the algorithm, it will help to classify the edges between $R_{{\mathcal{X}}}$ and $R_{{\mathcal{Y}}}$. If $uv$ is an edge of $G$ with $u \in R_{{\mathcal{X}}}$ and $v \in R_{{\mathcal{Y}}}$, we say that $uv$ is an *${\mathcal{X}}$-edge* if $\{u,v\} \subseteq X_i$ for some $i$, and that $uv$ is a *${\mathcal{Y}}$-edge* if $\{u,v\} \subseteq Y_j$ for some $j$. Admissibility of the pair $({\mathcal{X}}, {\mathcal{Y}})$ implies that any edge joining a vertex of $R_{{\mathcal{X}}}$ with a vertex of $R_{{\mathcal{Y}}}$ must be an ${\mathcal{X}}$-edge or a ${\mathcal{Y}}$-edge, since such edges intersect both a set in ${\mathcal{X}}$ and a set in ${\mathcal{Y}}$. It may be possible for an edge to be both an ${\mathcal{X}}$-edge and a ${\mathcal{Y}}$-edge.
\[claim:edgedeg\] Every vertex of $R_{{\mathcal{X}}}$ is incident to at most one ${\mathcal{Y}}$-edge, and every vertex of $R_{{\mathcal{Y}}}$ is incident to at most one ${\mathcal{X}}$-edge.
Suppose that $u \in R_{{\mathcal{X}}}$ and $uv_1$, $uv_2$ are two different ${\mathcal{Y}}$-edges incident to $u$. It follows that $\{u, v_1\} \subseteq Y_1$ and $\{u, v_2\} \subseteq Y_2$ for some sets $Y_1$ and $Y_2$ in ${\mathcal{Y}}$, and since the sets $Y \in {\mathcal{Y}}$ are pairwise vertex-disjoint, this implies that $Y_1 = Y_2$. Hence, $v_1$ and $v_2$ lie in the same set $Y$. Since $R_{{\mathcal{Y}}}$ is an ISR of ${\mathcal{Y}}$ and $\{v_1, v_2\} \subseteq R_{{\mathcal{Y}}}$, this is a contradiction.
The same argument, interchanging the roles of ${\mathcal{X}}$ and ${\mathcal{Y}}$, proves the claim about ${\mathcal{X}}$-edges.
Now consider the following algorithm, starting with $R = R_{{\mathcal{X}}} \cup R_{{\mathcal{Y}}}$. Say a vertex $v$ is *dangerous* for $R$ if it has degree $1$ in $G[R]$, and either $v \in R_{{\mathcal{X}}}$ and the incident edge is not an ${\mathcal{X}}$-edge, or $v \in R_{{\mathcal{Y}}}$ and the incident edge is not a ${\mathcal{Y}}$-edge. (Thus, if $v \in R_{{\mathcal{X}}}$ is dangerous, then its incident edge is a ${\mathcal{Y}}$-edge, and vice versa; the awkward negative wording is intended to exclude the possibility that the incident edge may be *both* an ${\mathcal{X}}$-edge and a ${\mathcal{Y}}$-edge.)
Note that vertices which were not initially dangerous may become dangerous as their neighbors are deleted, while vertices which were initially dangerous become non-dangerous if their neighbor is deleted.
- While $R$ has a dangerous vertex:
- Let $v$ be a vertex that is dangerous for $R$, and let $w$ be its unique neighbor in $R$.
- Delete the vertex $w$ from $R$.
- Once $R$ has no dangerous vertices remaining, delete every vertex of $R_{{\mathcal{Y}}}$ that has positive degree in $R$.
This algorithm clearly terminates, and the resulting set $R$ is clearly independent. It remains to show that $R$ hits every set $X \in {\mathcal{X}}$ and every set $Y \in {\mathcal{Y}}$.
First consider any set $X \in {\mathcal{X}}$, and let $w$ be the representative of $X$ in the set $R_{{\mathcal{X}}}$. If $w$ is still in $R$, then clearly $R \cap X \neq \emptyset$. Otherwise, $w$ was deleted from $R$, which only occurs when $w$ is the neighbor of some dangerous vertex $v$ (since if $w\in R_{{\mathcal{Y}}}$ as well, it is originally isolated in $R$, and hence never deleted in our algorithm). Once $w$ is deleted, the vertex $v$ has no neighbors in $R$, and the algorithm therefore never deletes $v$ in the rest of its execution. Hence $v \in R$ at the end of the algorithm. Since $vw$ was a dangerous edge and $v \in R_{{\mathcal{Y}}}$, the edge $vw$ was an ${\mathcal{X}}$-edge. This means that $\{v,w\} \subseteq X'$ for some $X' \in {\mathcal{X}}$, and since the sets $X \in {\mathcal{X}}$ are pairwise disjoint, this forces $X = X'$, so $v \in X$. Hence $R \cap X \neq \emptyset$.
Next consider any set $Y \in {\mathcal{Y}}$, and let $w$ be the representative of $Y$ in the set $R_{{\mathcal{Y}}}$. As before, if $w \in R$ then we are done. Otherwise, $w$ was deleted from $R$. If $w$ was deleted from $R$ because of some dangerous vertex $v \in R_{{\mathcal{X}}}$, then by the same argument as before, we have $v \in R$ at the end of the algorithm, and since $vw$ was a ${\mathcal{Y}}$-edge, we have $v \in Y$, so that $R \cap Y \neq \emptyset$.
Otherwise, $w$ was deleted from $R$ in the last step, when no dangerous vertices remained in $R$. This implies that either $w$ had degree at least $2$ when it was deleted, or $w$ had degree $1$ and the incident edge was a ${\mathcal{Y}}$-edge (since if the incident edge were not a ${\mathcal{Y}}$-edge then $w$ itself would be dangerous). Since, by Claim \[claim:edgedeg\], the vertex $w$ is incident to at most one ${\mathcal{X}}$-edge, in both cases $w$ was incident to a ${\mathcal{Y}}$-edge in $R$ at the time of its deletion. Let $vw$ be a ${\mathcal{Y}}$-edge incident to $w$. Since $w$ was deleted in the last step, no subsequent step could have deleted $v$, so $v \in R$ at the end. Furthermore, since $vw$ is a ${\mathcal{Y}}$-edge, we have $v \in Y$. Hence $R \cap Y \neq \emptyset$.
Thus, after executing the algorithm, $R$ is an independent set in $G$ that intersects every set $X \in {\mathcal{X}}$ and every set $Y \in {\mathcal{Y}}$.
Graphs with few triangles {#sec:few-tris}
=========================
Our first goal in this section is to prove the following theorem.
\[thm:triangle-11\] Let $H$ be a graph with maximum degree at most $2$, and let $G$ be a graph obtained from $H$ by gluing in vertex-disjoint copies of $K_4$. Let $\mathcal{{\mathcal{T}}}$ be the set of triangles in $H$. There is a set of vertices $Z$ with ${\left\lvert{Z}\right\rvert} \leq {\left\lvert{{\mathcal{T}}}\right\rvert}/4$, containing at most one vertex from each cycle in $H$, such that $\chi(G-Z) \leq 4$. In particular, ${\left\lvert{V(G)-Z}\right\rvert} \geq 11{\left\lvert{V(G)}\right\rvert}/12$.
Before proving Theorem \[thm:triangle-11\], we observe that it immediately implies the following corollary, which, taken together with Corollary \[cor:onelongoddcycle\], proves Theorem \[thm:main-cycles\].
\[cor:3triangles\] Let $H$ be a graph with $\Delta(H) \leq 2$, and let $G$ be a graph obtained from $H$ by gluing in vertex-disjoint copies of $K_4$. If $H$ has at most three triangles, then $\chi(G) \leq 4$.
Given such a graph $G$, we can apply Theorem \[thm:triangle-11\] with ${\left\lvert{{\mathcal{T}}}\right\rvert} \leq 3$ to obtain a set of vertices $Z$ with ${\left\lvert{Z}\right\rvert} \leq 3/4$ such that $G-Z$ is $4$-colorable. As ${\left\lvert{Z}\right\rvert}$ is an integer, we have $Z = \emptyset$ and so $G$ is $4$-colorable.
If $H$ has no triangles (instead of few triangles as in Corollary \[cor:3triangles\]), then it has girth at least 4. As mentioned in the introduction, Pei [@Pei] has a proof that after gluing in $K_4$’s to such an $H$, we get a graph $G$ that is 4-colorable. The main idea of Pei’s proof is to find an independent set $R$ hitting each added $K_4$ and each cycle, and to observe that $G-R$ can then be viewed as a subgraph of a “cycle-plus-triangles” graph. Since every such graph is $3$-colorable (by the celebrated result of Fleischner and Stiebitz [@cycletri]), using a fourth color on the set $S$ gives the desired $4$-coloring of $G$. We will adapt this idea in order to $4$-color “most” of the vertices of $G$ in the case where $H$ has few triangles.
Letting $\mathcal{T}$ denote the set of triangles in $H$ (where $\Delta(H)\leq 2$), we observe that Pei’s result easily yields a partial result itself. Deleting one edge from each triangle yields a graph $H'$ with girth at least 4, and after $4$-coloring the resulting “glued graph” $G'$, we must, at worst, uncolor one vertex from each triangle (an endpoint of a deleted edge) in order to obtain a proper partial $4$-coloring (of at least $|V(G)-\mathcal{T}|\geq \tfrac{2}{3}|V(G)|$ vertices). We improve this to a partial coloring of $\tfrac{11}{12}$ of the vertices by showing that only ${\left\lvert{{\mathcal{T}}}\right\rvert}/4$ vertices must be deleted, rather than ${\left\lvert{{\mathcal{T}}}\right\rvert}$ vertices as in the simple argument.
Let $X_1, \ldots, X_p$ be the vertex sets of the cycles of $H$ and let $Y_1, \ldots, Y_q$ be the vertex sets of the added copies of $K_4$. Let ${\mathcal{X}}= (X_1, \ldots, X_p)$ and let ${\mathcal{Y}}= (Y_1, \ldots, Y_q)$.
Note that ${\mathcal{Y}}$ has an ISR in $G$ by Theorem \[thm:haxell-maxdeg\]. Our general strategy will be to first apply Lemma \[lem:haxell-defic\] to find a “large” subfamily ${\mathcal{X}}' \subseteq {\mathcal{X}}$ that admits an ISR. We will then be able to apply Lemma \[lem:combine-isr\] to get a $G$-independent set $R$ which is a transversal of both ${\mathcal{X}}'$ and ${\mathcal{Y}}$. Finally, we will use this $R$ (in a similar way to the set $R$ described above in Pei’s proof) to define our desired 4-coloring.
\[claim:p-k\] ${\mathcal{X}}$ has a partial ISR with $p-k$ vertices, where $k=\lfloor \tfrac{\mathcal{T}}{4}\rfloor$.
Let $S$ be any subset of $[p]$. In order for Lemma \[lem:haxell-defic\] to yield our desired result, We must show that ${\bar{\gamma}}(H_s) \geq 2{\left\lvert{S}\right\rvert}-2k-1$. Since all edges induced by the sets $X_i$ are removed from $H_S$, the only edges remaining in $H_S$ are edges that were added by the copies of $K_4$. Hence, every component of $H_S$ has at most $4$ vertices. Since ${\bar{\gamma}}(H_0) \geq 2$ for every graph $H_0$, it follows that $${\bar{\gamma}}(H_S) \geq {\left\lvert{V(H_S)}\right\rvert}/2.$$ On the other hand, since every set $X_i$ is the vertex set of either a triangle or a cycle of length at least $4$, we have $${\left\lvert{V(H_S)}\right\rvert} \geq 4{\left\lvert{S}\right\rvert} - {\left\lvert{{\mathcal{T}}}\right\rvert}.$$ Combining these inequalities yields $${\bar{\gamma}}(H_S) \geq 2{\left\lvert{S}\right\rvert} - \tfrac{{\left\lvert{{\mathcal{T}}}\right\rvert}}{2}= 2{\left\lvert{S}\right\rvert} - 2\left(\tfrac{{\left\lvert{{\mathcal{T}}}\right\rvert}}{4}\right)\geq 2{\left\lvert{S}\right\rvert} - \left(2{\left\lfloor{\tfrac{{\left\lvert{{\mathcal{T}}}\right\rvert}}{4}}\right\rfloor} + \tfrac{3}{4}\right) = 2{\left\lvert{S}\right\rvert} - 2k - \tfrac{3}{2}.$$ Since ${\bar{\gamma}}(H_S)$ is an integer, this gives our desired bound.
Let ${\mathcal{X}}'$ be the subfamily of ${\mathcal{X}}$ consisting of the sets containing a vertex from the partial ISR found in Claim \[claim:p-k\]. Then ${\mathcal{X}}'$ has an ISR in $G$.
\[claim:admissible\] There is an $G$-independent set $R$ that is a transversal of both ${\mathcal{X}}'$ and of ${\mathcal{Y}}$.
Since ${\mathcal{X}}'$ and ${\mathcal{Y}}$ both have an ISR in $G$, we get our desired result via Lemma \[lem:combine-isr\] provided that $({\mathcal{X}}', {\mathcal{Y}})$ is admissible for $G$. To this end, observe that every edge $e \in E(G)$ falls into one of the following categories:
- $e$ is induced by some cycle of ${\mathcal{X}}- {\mathcal{X}}'$, hence both endpoints of $e$ are missing from all $X \in {\mathcal{X}}$, or
- $e$ is induced by some cycle $X \in {\mathcal{X}}'$, hence both endpoints of $e$ are in $X$, or
- $e$ is an added edge from some $K_4$, hence both endpoints of $e$ are in $Y$ for some $Y \in {\mathcal{Y}}$.
It follows that $({\mathcal{X}}', {\mathcal{Y}})$ is admissible for $G$.
Let $F$ be a set consisting of one edge from each cycle of $H$ not contained in ${\mathcal{X}}'$ (so $|F|=k$). Let $J = H-R-F$ and observe that $J$ is a graph of maximum degree at most $2$ with no cycles. By adding edges between the endpoints of path components in $J$, we obtain a Hamiltonian cycle $J'$ on the same vertex set. For each $j \in [q]$, let $Y'_j = Y_j - R$; we have ${\left\lvert{Y'_j}\right\rvert} = 3$ for all $j$, since $R$ intersects each $Y_j$ in exactly one vertex. Let $J^*$ be the graph obtained from $J'$ by gluing in a triangle on each $Y_j'$. By Fleischner and Stiebitz’s [@cycletri] cycle + triangles result, we get that $J^*$ is $3$-colorable. As $G-F-R$ is a subgraph of $J^*$, it follows that $G-F-R$ is $3$-colorable. Using a fourth color on the independent set $R$ yields a $4$-coloring of $G-F$. Let $Z$ be a vertex set consisting of one endpoint of each monochromatic edge in $F$. Now $G-Z$ is properly $4$-colored, and we have $|Z|\leq k\leq \tfrac{|\mathcal{T}|}{4}$. Since $|\mathcal{T}|\leq \tfrac{|V(G)|}{3}$, this implies ${\left\lvert{V(G)-Z}\right\rvert} \leq 11{\left\lvert{V(G)}\right\rvert}/12$. Furthermore, since $F$ has at most one edge from each cycle, the vertex set $Z$ has at most one vertex from each cycle, as desired.
As stated, Theorem \[thm:triangle-11\] gives us no control over which cycles contain uncolored vertices, in contrast to Theorem \[thm:longodd-910\] which guarantees that the uncolored vertices are contained in the long odd cycles of $H$. In fact, by refining the statement of Lemma \[lem:haxell-defic\] to only add “dummy vertices” to the sets $X_i$ obtained from triangles, one can guarantee that the set $R$ hits all long odd cycles, and that all uncolored vertices lie in triangles of $H$. Proving this formally would require more technical conditions in the hypothesis of Lemma \[lem:haxell-defic\], so in the interest of clarity we have opted to only formally prove the simpler formulation.
Theorem \[thm:longodd-910\] is weakest when most vertices of $H$ lie in long odd cycles; Theorem \[thm:triangle-11\] is weakest when most vertices of $H$ lie in triangles. These worst-case scenarios cannot happen simultaneously; combining these bounds gives a stronger overall bound on the number of vertices in a $4$-colorable subgraph, namely, Theorem \[thm:main-fraction\]
Let ${n_t}$ and ${n_{\ell}}$ denote the number of vertices of $H$ that lie in triangles and in odd cycles of length exceeding $3$, respectively, and let ${n_0}$ denote the number of other vertices in $H$, so that $|V(H)|={\left\lvert{V(G)}\right\rvert}= {n_0}+{n_t}+ {n_{\ell}}$. Let ${n_4}$ denote the number of vertices of $G$ in a largest $4$-colorable induced subgraph.
Theorem \[thm:longodd-910\] says that ${n_4}\geq |V(G)|-\tfrac{|\mathcal{C}|}{2}$, where $\mathcal{C}$ is the set of all odd cycles in $H$ of length exceeding 3. Since $|\mathcal{C}|\leq \tfrac{{n_{\ell}}}{5}$, we get that $${n_4}\geq {n_0}+ {n_t}+ \frac{9}{10}{n_{\ell}}.$$ Similarly, Theorem \[thm:triangle-11\] says that ${n_4}\geq |V(G)|-\tfrac{|\mathcal{T}|}{4}$, where $\mathcal{T}$ is the set of all triangles in $H$. Since $|\mathcal{T}|=\tfrac{{n_t}}{3}$, we get that $${n_4}\geq {n_0}+ \frac{11}{12}{n_t}+ {n_{\ell}}.$$ Given any $\lambda \in [0,1]$, we can take a convex combination of the above two inequalities (with $\lambda$ times the first and $(1-\lambda)$ times the second), to get $${n_4}\geq {n_0}+ \left(\lambda + \frac{11}{12}(1-\lambda)\right){n_t}+
\left(\frac{9}{10}\lambda + (1-\lambda)\right){n_{\ell}}.$$ Setting $\lambda = 5/11$ equalizes the coefficients of ${n_t}$ and ${n_{\ell}}$, yielding the bound $${n_4}\geq {n_0}+ \frac{21}{22}{n_t}+ \frac{21}{22}{n_{\ell}}\geq \frac{21}{22}{\left\lvert{V(G)}\right\rvert}.$$
Finding Two ISRs {#sec:two-isrs}
================
By embedding $H$ in a larger graph we may assume that $V_1, V_2, \ldots, V_n$ partition $V(H)$.
Let $G_1 = H$, let $G_2$ consist of the edges of a $C_4$ on each part $V_i$, and let $G_3$ consist of the edges of a $K_4$ on each part $V_i$. We regard the edges in $G_1$ as colored red and the edges in $G_3$ as colored blue (as in the proof of Theorem \[thm:longodd-910\]); we shall also regard the edges of $G_2$ as colored green. Let $G$ be the multigraph $G = G_1 \cup G_2 \cup G_3$. (While the edge multiplicity is not relevant to the coloring problem, it simplifies things to be able to view a pair of vertices as possibly joined by edges of multiple colors.)
Our main idea is similar to the proof of Theorem \[thm:longodd-910\]: *ideally*, we would like to partition $V(G)$ into two sets $W_1$ and $W_2$ such that $G[W_1]$ and $G[W_2]$ each have maximum degree $1$ in each of red and blue. If we could do this, it would imply that $G[W_1]$ and $G[W_2]$ were both bipartite, so that $\chi(G) \leq 4$. While we are unable to achieve this goal for both classes, we will be able to ensure that $G[W_1]$ has a stronger condition: namely that it has maximum degree $1$ in each of red and blue and that it contains exactly two vertices from each $V_i$. This implies that not only is $G[W_1]$ bipartite, but that its partite sets (each of which must contain exactly one vertex from each $V_i$ since $G$ has a blue $K_4$ induced on each $V_i$) are our desired pair of ISRs.
There exists a partition $V(G)$ into two sets $W_1$ and $W_2$, with $|W_1|=|W_2|$, such that $G_i[W_i]$ has maximum degree $1$ for each $i\in\{1,2\}$ (that is, $G[W_1]$ has maximum degree $1$ with respect to red edges and $G[W_2]$ has maximum degree $1$ with respect to green edges).
A lemma of Haxell, Szabó, and Tardos (Lemma 2.6 of [@HST]) tells us that we can find a partition of $V(G)$ into two sets $W_1$ and $W_2$ such that $G_i[W_i]$ to has maximum degree $1$ for each $i\in\{1,2\}$ ($G[W_1]$ has maximum degree $1$ with respect to red edges and $G[W_2]$ has maximum degree $1$ with respect to green edges). We will show that by slightly modifying their proof, we can get their conclusion with ${\left\lvert{W_1}\right\rvert} = {\left\lvert{W_2}\right\rvert}$.
We form our partition using the following algorithm, which is adapted from [@HST] with one minor tweak (which we shall point out shortly). Fix an orientation of $G_1$ and of $G_2$ with maximum outdegree $1$ in each orientation. For each vertex $v \in V(G)$, let $v^{1+}$ and $v^{2+}$ denote the successor of $v$ in $G_1$ and $G_2$, respectvely, whenever these successors exist. Likewise, we write $v^{1-}$ and $v^{2-}$ for the predecessors of $v$ in $G_1$ and $G_2$ when these predecessors exist.
Initially, set $W_1 = W_2 = \emptyset$. We start by adding an arbitrarily-chosen vertex to $W_2$, and after this and each subsequent added vertex, we choose the next vertex to add as follows:
- For $i=1,2$, if the just-added vertex $v$ was added to $W_i$:
(i) If $v^{i+}$ exists and is not yet placed, then add $v^{i+}$ to $W_{3-i}$. (Now $v^{i+}$ is the just-added vertex for the next step.)
(ii) Otherwise, if $v^{i-}$ exists and is not yet placed, then add $v^{i-}$ to $W_{3-i}$. (Now $v^{i-}$ is the just-added vertex.)
(iii) Otherwise, if there is any unplaced vertex $w$, then add $w$ to $W_{3-i}$. (Now $w$ is the just-added vertex.)
(iv) Otherwise, terminate.
Our algorithm differs from the algorithm of [@HST] only in that our algorithm *always* alternates between placing a vertex into $W_1$ or into $W_2$, while the algorithm of [@HST] always places the next vertex into $W_1$ when it makes an arbitrary choice in Case (iii). Since $V_1, V_2, \ldots, V_n$ is a partition of $V(H)=V(G)$ where each part has size 4, we know that ${\left\lvert{V(G)}\right\rvert}$ is even, so this alternation guarantees ${\left\lvert{W_1}\right\rvert} = {\left\lvert{W_2}\right\rvert} = {\left\lvert{V(G)}\right\rvert}/2$ at the end.
The proof of Lemma 2.6 of [@HST] immediately implies that $\Delta(G_i[W_i]) \leq 1$ for each $i$. (As their proof never specifically uses the choice of $W_1$ in Case (iii), but rather only uses the choices made in Case (i) and Case (ii), it goes through without modification for this version of the algorithm.)
$G[W_1]$ has maximum degree $1$ in each of red and blue and it contains exactly two vertices from each $V_i$.
Taking $W_1, W_2$ as in the previous claim, we must additionally show that:
(1) $G[W_1]$ has maximum degree 1 with respect to blue edges, and
(2) $W_1$ contains exactly two vertices from each $V_i$.
If $W_1$ has least three vertices from some $V_i$, then this would force $G[W_1]$ to have maximum degree at least 2 with respect to blue edges. Hence, it suffices only to prove (1). To this end, suppose on the contrary that some vertex $v \in W_1$ is incident to two blue edges within $G[W_1]$. Then $W_1$ contains at least $3$ vertices of the corresponding copy of $K_4$, and $W_2$ contains at most $1$ vertex of that copy of $K_4$. Since ${\left\lvert{W_1}\right\rvert} = {\left\lvert{W_2}\right\rvert} = {\left\lvert{V(G)}\right\rvert}/2$, this forces $W_2$ to contain at least $3$ vertices of some other copy of $K_4$. However this would force $G[W_2]$ to have a vertex with green degree at least two, contradicting our choice of $W_2$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This work presents accurate [*ab initio*]{} investigations of various spectroscopic properties of a few Li-like ions in presence of a plasma environment within the Debye screening potential. The coupled-cluster theory in the relativistic framework has been employed to compute ionization potentials, excitation energies, electric dipole oscillator strengths, and electric quadrupole transition probabilities of Li-like C$^{3+}$, N$^{4+}$, and O$^{5+}$ ions. The unretarded Breit interaction has been implemented to increase the accuracy of the calculations. The effects of ion density and temperature on the ionization potentials, excitation energies, electric dipole oscillator strengths, and electric quadrupole transition probabilities have been investigated in the plasma environment. It is found that the plasma screening leads to a sharp decrease in the ionization potential as the screening strength increases. With increasing strength, the oscillator strengths associated with 2$s~^{2}S_{1/2}$$\rightarrow$2$p~^{2}P_{1/2, 3/2}$ transitions increase, whereas the transition probabilities associated with 3$d$ $^2D_{3/2, 5/2}$$\rightarrow$2$s$ $^2S_{1/2}$ transitions decrease.'
author:
- Pradip Kumar Mondal
- Narendra Nath Dutta
- Gopal Dixit
- Sonjoy Majumder
title: 'Effect of screening on spectroscopic properties of Li-like ions in plasma environment'
---
Introduction
============
With the advent of novel x-ray sources based on free-electron lasers (FELs) [@emma2; @ishikawa2012], laser plasmas [@rousse2001], and high-harmonic generations [@mckinnie2010; @popmintchev2012], it is possible to achieve extreme conditions in matter such as high energy density and high temperature using ultraintense, ultrashort, and tunable pulses, and hence it is possible to create matter in plasma form [@vinko2012; @ciricosta2012]. There have been many experimental [@murnane1991ultrafast; @riley1992plasma; @mourou1992development; @rogers1994astrophysical; @nazir1996x; @workman1997application; @nantel1998pressure; @woolsey1998competing; @saemann1999isochoric; @vinko2012; @ciricosta2012] and theoretical [@rouse1971screening; @gupta1982density; @seidel1995energy; @ray1998magnetic; @jung1999orientation; @ray2000influence; @pang2002analytic; @saha2002energy; @okutsu2005electronic; @saha2005spherically; @saha2006isoelectronic; @li2008influence; @bhattacharyya2008effect; @saha20092pnp; @paul2009hydrogen; @qi2009generalized; @sil2009spectra; @gao2011plasma; @xie2012energy] endeavors to explain and understand the effect of plasma environment on the spectroscopic properties of atoms and ions. In the situation, when atoms or ions embedded in plasma, the interaction between the nucleus and the bound electrons is screened by the surrounding ions and fast electrons. The modified interaction gives rise to phenomena such as pressure ionization and continuum lowering and affects the spectroscopic properties of atoms and ions [@griem1974spectral; @jaskolski1996confined]. Recent advanced experiment, carried out using FEL and electron beam ion trap (EBIT), provides an unexpected low oscillator strength of electric dipole ($E1$) transition of Fe$^{16+}$ and raises the concern about the quality of the atomic wave functions used to model such spectral properties [@bernitt2012]. Therefore, treating the effect of plasma environment in atoms and ions along with an accurate treatment of electron-electron correlation and relativistic effects are nontrivial. The ratio of Coulomb energy to thermal energy determine the strength of coupling ($\Gamma$) in plasma. The low density and high temperature situation corresponds to weakly coupled plasma ($\Gamma < 1$), where the screening of the nuclear Coulomb interaction by free electrons in the plasma is guided by the Debye model [@ichimaru1982strongly; @murillo1998dense].
Lithium and lithium-like ions in plasma are a few of the most abundant ionic species for specific temperature and density attainable in the laboratory [@kawachi1995]. Various spectroscopic properties of Li-like ions have significant importance in astrophysics due to evidence of high abundances of these ions in different astronomical systems like active galactic nuclei, x-ray binaries, quasars, and hot plasmas [@nahar2000; @gorczyca2006]. For such small-sized atoms, allowed and forbidden transitions with sufficient intensity are used as diagnostic tools of tokamak plasmas [@suckewer1981; @godefroid1984; @das1998]. It is well known that the ionized form of carbon, nitrogen, and oxygen and their various transition lines are important for the chromosphere region of the solar atmosphere [@fontenla2007], and to understand the dynamics and nature of the stellar and interstellar medium [@barstow2010; @ferrero2011; @prochaska2011; @fox2011; @de2012]. There is long literature on the applications of the isolated resonance lines having wavelength 1548.19 Å and 1550.77 Å for C$^{3+}$, 1238.82 Å and 1242.80 Å for N$^{4+}$ and 1031.91 Å and 1037.61 Å for O$^{5+}$ [@peach1988; @nist2012]. The astronomical observed lines are expected to be affected by the plasma atmosphere at the origin and therefore important for plasma diagnostic purpose. Forbidden transition lines, i.e., electric quadrupole ($E2$) and magnetic dipole ($M1$) transition lines provide very crucial parameters for estimations of density and internal temperature measurements at low density hot plasmas [@burgess1998; @allende2002; @chen2010]. Also, the transition rates of forbidden transitions provide accurate dielectric recombination rates for these ions [@andersen1992; @qu1999; @laming2003].
Several theoretical methods have been used to model the effect of plasma environment on the spectroscopic properties for one-electron [@saha2002energy; @ray2000influence; @bhattacharyya2008effect] and many-electron [@saha2006isoelectronic; @li2008influence; @das2011; @das2012; @chaudhuri2012] systems. Due to the screening effect, lowering of the ionization potential is demonstrated by Stewart and Pyatt [@stewart1966lowering]. The Debye plasma screening on lighter atoms or ions have been studied over last decade using different many-body approaches [@bielinska2004relativistic; @saha2006isoelectronic; @li2008influence; @bhattacharyya2008effect] and showed enough avenues of improvement. Recent works of correlation exhaustive Dirac-Coulomb based coupled -cluster calculations on He-, Li-, Be- and Na-like ions [@das2011; @das2012; @chaudhuri2012] are examples of this. The authors of these works emphasized the importance of relativistic correction on the plasma screening by using more accurate many-body theories within a relativistic framework [@das2011; @das2012; @chaudhuri2012].
In the present article, we analyze the influence of the plasma screening on the Li-like C$^{3+}$, N$^{4+}$, and O$^{5+}$ ions using the Debye model potential. The ionization potentials, excitation energies, oscillator strengths of $E1$ transitions, and transition rates of $E2$ transitions are estimated for these ions in the isolated (free) condition as well as within the plasma environment. Here, we have used the Fock-space coupled-cluster (FSCC) method within the Dirac-Coulomb-Breit (DCB) Hamiltonian to consider the relativistic effect on these spectral properties. Recently, Dutta $\it {et~al.}$ have implemented the unretarded Breit interaction in an all-order approach using the coupled-cluster (CC) theory and demonstrate the effect of electron-electron correlation and unretarded Breit interaction on the boron isoelectronic sequence [@dutta2012]. It is well known that the dynamical electron correlation, relaxation effect, and Breit interaction are important for moderately charged ions, which are considered here in an accurate way. This paper is structured as follows. Section II discusses brief theory and formalism of the relativistic FSCC approach with the Debye screening potential. Section III presents results and discussions on several spectral properties of Li-like ions and the effect of plasma screening on these properties. Conclusions and future outlooks are presented in Sec. IV.
Theory
======
In order to consider the effect of plasma environment on the spectroscopic properties, the Dirac-Coulomb (DC) Hamiltonian with unretarded Breit interaction for an $N$-electron atomic system can be written as $$H=\sum_{i=1}^{N}\left(c\overrightarrow{\alpha_i}\cdot\overrightarrow{p_i}+\left(
\beta_i-1\right) c^2 +V^{D}_{\textrm{eff}}(r_i)+
\sum_{j<i}\left(\frac{1}{r_{ij}}-\frac{\overrightarrow{\alpha_i}\cdot\overrightarrow{\alpha_j}}{r_{ij}}\right)\right),$$ with all the standard notations often used. Here, $V^{D}_{\textrm{eff}}(r_i)$ is the effective potential of the nucleus on the $i$-th electron due to the presence of plasma environment. The Debye-Hückel potential is considered to examine the effect of screening of nuclear Coulomb potential due to the presence of ions and free electrons in plasmas [@akhiezer1975plasma; @ichimaru1982strongly]. In case of weakly interacting plasma medium, the effective potential experienced by the $i$-th electron is given as $$V^{D}_{\textrm{eff}}(r_i)= -\frac{Ze^{-\mu r_i}}{r_i},$$ where $Z$ is the nuclear charge and $\mu$ is the Debye screening parameter, which is related to the ion density $n_{\textrm{ion}}$ and plasma temperature $T$ through the following relation: $$\mu = \left[ \frac{4\pi (1+Z)n_{\textrm{ion}}}{k_{B}T}
\right]^{\frac{1}{2}},$$ where $k_{B}$ is the Boltzmann constant. Therefore, a given value of $ \mu $ represents a range of plasma conditions with different ion densities and temperatures. The inverse of the Debye screening parameter is called the Debye screening length, i.e., $\lambda_{D} =
\mu^{-1}$. The pure Coulomb nuclear attraction corresponds to the zero screening situation ($\mu = 0$).
The wave functions, ionization potentials (IPs) of the ground, and the different excited states for the considered $N$-electron atomic system are obtained using the FSCC method with single, double, and partial triple excitations within the relativistic framework. The basic formalism of the FSCC method was developed several decades before [@lindgren1985; @lindgren1987; @haque1984; @pal1987; @pal1988]. The relativistic version of the FSCC theory has been developed recently and successfully employed to obtain the various properties in different single valence atomic systems [@eliav1994; @isaev2004; @sur2006; @dixit2007PRA; @dixitcd; @dixitzn; @pal2007relativistic; @mani2010atomic]. Here, we provide key steps of this method. In the relativistic coupled-cluster (RCC) theory, the correlated wave function of a single valence atomic state with a valance electron in the “$v$"th orbital is written in the form, $$|\Psi_v\rangle=e^{T}\{1+S_v\}|\Phi_v\rangle$$ where, $|\Phi_v\rangle$ is the corresponding reference state generated at the Dirac-Fock (DF) level of the $N-1$ electron closed-shell system using Koopman’s theorem [@szabo1996]. $T$ represents all possible excitations from the core orbitals of the closed-shell system, and $S_v$ represents all possible valence and core-valence excitations of the single-valence system. The detail inclusion of the unretarded Breit interaction in this formalism is described by Dutta and Majumder [@dutta2012].
The transition matrix element for any operator $O$ in the framework of the RCC wave function can be expressed as $$\begin{aligned}
O_{fi}& = & \frac{\langle \Psi_f|O|\Psi_i\rangle} {\sqrt{{\langle
\Psi_f|\Psi_f\rangle}
{\langle \Psi_i|\Psi_i\rangle}}} \nonumber \\
&=& \frac{{\langle
\Phi_f|\{1+{S_f}^{\dag}\}{e^T}^{\dag}Oe^T\{1+S_i\}|\Phi_i\rangle}}
{\sqrt{{\langle
\Phi_f|\{1+{S_f}^{\dag}\}{e^T}^{\dag}e^T\{1+S_f\}|\Phi_f\rangle}
{\langle
\Phi_i|\{1+{S_i}^{\dag}\}{e^T}^{\dag}e^T\{1+S_i\}|\Phi_i\rangle}}}.\end{aligned}$$ The single-electron reduced matrix elements corresponding to $E1$, $E2$, and $M1$ transitions are discussed in Refs. [@grant1974; @johnson1995].
The transition probabilities (in s$^{-1}$) corresponding to $E1$, $E2$, and $M1$ channels from state [*k*]{} to [*i*]{} is given as $$A^{E1}_{k \rightarrow i} =
\frac{2.0261\times10^{18}}{\lambda^{3}g_{k}}S^{E1},$$ $$A^{E2}_{k \rightarrow i} = \frac{1.11995\times10^{18}}{\lambda^{5}
g_{k}}S^{E2}$$ and $$A^{M1}_{k \rightarrow i} = \frac{2.69735\times10^{13}}{\lambda^{3}
g_{k}}S^{M1},$$ where $S = {|{\langle \Psi_k|O|\Psi_i\rangle}|}^2$ is the transition strength of the operator $O$ (in a.u.), $\lambda$ (in Å) is the corresponding transition wavelength, and $g_{k} =
2j_k+1$ is the degeneracy of the $k$ state. The oscillator strength of $E1$ transition from state [*i*]{} to [*k*]{} is given as $$f^{E1}_{i \rightarrow k} = 1.4992 \times 10^{-16}
\frac{g_{k}}{g_{i}} \lambda^{2} A^{E1}_{k \rightarrow i}.$$
Results and Discussions
=======================
The precise DF orbital wave functions, which are the building block of accurate correlation calculations, are generated via the basis set expansion technique in the potential of two core electrons at $1s_{1/2}$ orbital. The radial part of these basis wave functions are considered to be Gaussian type [@chaudhuri2000] having two optimized exponential parameters $\alpha_0$ and $\beta$. The nuclei are considered as finite size with a Fermi-type charge distribution [@parpia1992]. For all the ions considered here, the number of basis wave functions at the DF levels for $s$-, $p$-, $d$-, and $f$-type symmetries are 30, 25, 20, and 20, respectively. In order to choose the optimized parameters for each ion, the energies and wave functions of the DF orbitals are compared with the same as obtained from the GRASP92 code, where the DF equations are solved numerically [@parpia2006]. These exponential parameters are chosen as 0.005825 and 2.73 for C$^{3+}$, 0.003265 and 2.73 for N$^{4+}$, and 0.00525 and 2.73 for O$^{5+}$. The number of DF orbitals for different symmetries used in the RCC calculations are based on the convergent criteria of core correlation energies with increasing number of orbitals. There are 12, 11, 10, and 10 active orbitals, which include all core orbitals, considered in the calculations for the $s$, $p$, $d$, and $f$ type symmetries, respectively. As an improvement of the atomic Hamiltonian beyond the Dirac-Coulomb limit, the Breit interaction in it’s unretarded approximation has been included for more accurate relativistic descriptions of the wave functions [@dutta2012]. The $T$ amplitudes are solved first for the closed-shell systems and later the $S$ amplitudes corresponding to different single-valence states are solved for the open-shell systems using the RCC equations.
The quality of the wave functions for different eigen states is ensured by comparing the reduced matrix elements of the $E1$ transitions in length and velocity gauges [@johnson1995]. The calculated average deviations of these matrix elements between these two gauges at the CC levels have been found 1.11$\%$ for C$^{3+}$, 1.17$ \% $ for N$^{4+}$, and 1.58$ \% $ for O$^{5+}$, which indicate very good quality of the relativistic wave functions. The calculated values of the ground state IPs for C$^{3+}$, N$^{4+}$, and O$^{5+}$ are 520145, 789494 and 1113968 cm$^{-1}$, respectively; which are in excellent agreement with the NIST results: 520178, 789537 and 1114004 cm$^{-1}$, respectively. The average deviations of our calculated excitation energies (EEs) of different excited states with respect to the NIST results are estimated around 0.05$\%$, 0.06$\%$ and 0.06$\%$ for C$^{3+}$, N$^{4+}$, and O$^{5+}$, respectively [@nist2012].
The $E1$ oscillator strengths are presented in Table \[tab:results1\] along with the available NIST results [@nist2012] for comparison. The average deviations between our calculated results and the NIST values are about 1.6% for C$^{3+}$, 1.2% for N$^{4+}$, and 1.2% for O$^{5+}$. Our results for oscillator strength agree excellently with the recent FSCC results presented by Das [*et al.*]{} for C$^{3+}$ ion [@das2012]. The recently calculated results of Elabidi et al. using SUPERSTRUCTURE code [@elabidi2011] have an average deviation of about 3% with respect to the NIST results; though, in some cases deviations are around 25% [@elabidi2011]. Therefore, correlation exhaustive and relativistic calculations were wanted, and our calculations are motivated towards that. The oscillator strengths of 3$d$ $^{2}D_{3/2,5/2}$$\rightarrow$ 4$f$ ${^2}F_{5/2,7/2}$ transitions are estimated here as well.
------------------ --------------- ------------------ ------- ------- ------- ------- ------- -------
Lower Upper RCC NIST RCC NIST RCC NIST
2$s$ $^2S_{1/2}$ $\rightarrow$ 2$p$ $^2P_{1/2}$ 0.095 0.095 0.078 0.078 0.066 0.066
$\rightarrow$ 2$p$ $^2P_{3/2}$ 0.190 0.190 0.156 0.156 0.133 0.133
$\rightarrow$ 3$p$ $^2P_{1/2}$ 0.067 0.068 0.078 0.079 0.087 0.089
$\rightarrow$ 3$p$ $^2P_{3/2}$ 0.133 0.136 0.156 0.159 0.174 0.177
$\rightarrow$ 4$p$ $^2P_{1/2}$ 0.019 0.020 0.021 0.023 0.023 0.025
$\rightarrow$ 4$p$ $^2P_{3/2}$ 0.037 0.041 0.042 0.046 0.047 0.049
2$p$ $^2P_{1/2}$ $\rightarrow$ 3$d$ $^2D_{3/2}$ 0.645 0.646 0.651 0.652 0.654 0.657
$\rightarrow$ 4$d$ $^2D_{3/2}$ 0.122 0.123 0.122 0.122 0.120 0.123
2$p$ $^2P_{3/2}$ $\rightarrow$ 3$d$ $^2D_{3/2}$ 0.064 0.065 0.065 0.065 0.065 0.066
$\rightarrow$ 3$d$ $^2D_{5/2}$ 0.581 0.581 0.586 0.588 0.590 0.591
$\rightarrow$ 4$d$ $^2D_{3/2}$ 0.012 0.012 0.065 0.065 0.012 0.012
3$s$ $^2S_{1/2}$ $\rightarrow$ 3$p$ $^2P_{1/2}$ 0.161 0.160 0.131 0.131 0.112 0.111
$\rightarrow$ 3$p$ $^2P_{3/2}$ 0.323 0.320 0.267 0.263 0.226 0.224
$\rightarrow$ 4$p$ $^2P_{1/2}$ 0.063 0.068 0.077 0.082 0.092 0.092
$\rightarrow$ 4$p$ $^2P_{3/2}$ 0.126 0.136 0.153 0.164 0.182 0.185
3$p$ $^2P_{1/2}$ $\rightarrow$ 3$d$ $^2D_{3/2}$ 0.062 0.063 0.054 0.055 0.049 0.049
$\rightarrow$ 4$d$ $^2D_{3/2}$ 0.528 0.541 0.540 0.550 0.542 0.557
3$p$ $^2P_{3/2}$ $\rightarrow$ 3$d$ $^2D_{3/2}$ 0.006 0.006 0.005 0.005 0.005 0.005
$\rightarrow$ 3$d$ $^2D_{5/2}$ 0.055 0.056 0.048 0.049 0.043 0.044
$\rightarrow$ 4$d$ $^2D_{3/2}$ 0.053 0.054 0.054 0.055 0.054 0.056
$\rightarrow$ 4$d$ $^2D_{5/2}$ 0.476 0.486 0.486 0.495 0.489 0.501
3$d$ $^2D_{3/2}$ $\rightarrow$ 4$f$ $^2F_{5/2}$ 1.020 1.020 1.020 1.020 1.010
3$d$ $^2D_{5/2}$ $\rightarrow$ 4$f$ $^2F_{5/2}$ 0.049 0.049 0.048 0.049 0.048
$\rightarrow$ 4$f$ $^2F_{7/2}$ 0.974 0.975 0.967 0.974 0.966
------------------ --------------- ------------------ ------- ------- ------- ------- ------- -------
: Oscillator strengths of $E1$ transitions and their comparisons with the NIST results for C$^{3+}$, N$^{4+}$, and O$^{5+}$.[]{data-label="tab:results1"}
\
Table \[tab:results2\] presents the emission probabilities of $E2$ transitions having values of the order of 10$^4$ s$^{-1}$ or more using precisely calculated wavelengths for C$^{3+}$, N$^{4+}$, and O$^{5+}$. Sur and Chaudhuri have reported a few $E2$ transition probabilities for O$^{5+}$ using the RCC theory based on the DC Hamiltonian [@sur2007]. Their results differ by about 0.8% from our calculated values obtained by the same theory but based on the DCB Hamiltonian. All the transitions presented in the Table \[tab:results2\] fall in the ultraviolet region of electromagnetic spectrum. The wavelengths of these forbidden lines relative to those of allowed lines from the same ion make them very good candidates for line profile measurements and help to understand the excitation processes, like electron- and proton-impact excitations [@feldman1981; @dworetsky1998]. It is evident from the table that there are few strong $E2$ transitions with transition probabilities of the order of $10^{6}$ s$^{-1}$. These are the transitions between the ground states and 3$d$ ${^2}D_{3/2, 5/2}$ states for N$^{4+}$, and O$^{5+}$, and the transitions 3$p$ $^{2}P_{1/2}$ $\rightarrow$ 2$p$ $^{2}P_{3/2}$ and 4$f$ $^{2}F_{5/2}$ $\rightarrow$ 2$p$ $^{2}P_{1/2}$ for O$^{5+}$. Therefore, these forbidden transition lines having relatively higher probabilities can play in the determinations of density and internal temperature inside hot plasmas. Because the $M1$ transition probabilities are found to be quite low (of the order of $ 10^{-1} $ s$^{-1}$ or less), the lines associated with these transitions are hardly possible to detect and hence, are excluded from the consideration here.
------------------ --------------- ------------------ --------- ---------- ------- ------------ -------- ------------ -------- -------- --------
$A_{if}$ $ A_{if} $ $ A_{if} $
Upper Lower RCC NIST RCC NIST RCC NIST
3$d$ $^2D_{3/2}$ $\rightarrow$ 2$s$ $^2S_{1/2}$ 307.71 307.81 44.03 206.36 206.43 150.17 148.16 148.23 413.17
3$d$ $^2D_{5/2}$ $\rightarrow$ 2$s$ $^2S_{1/2}$ 307.70 307.79 44.04 206.35 206.43 150.23 148.15 148.21 413.41
4$d$ $^2D_{3/2}$ $\rightarrow$ 2$s$ $^2S_{1/2}$ 243.39 243.71 10.33 161.62 161.83 30.06 115.21 115.35 78.05
4$d$ $^2D_{5/2}$ $\rightarrow$ 2$s$ $^2S_{1/2}$ 243.39 243.71 12.30 161.61 161.83 30.09 115.21 115.35 78.14
3$p$ $^2P_{3/2}$ $\rightarrow$ 2$p$ $^2P_{1/2}$ 391.08 391.23 5.31 251.52 251.64 20.01 175.42 175.47 59.53
4$p$ $^2P_{3/2}$ $\rightarrow$ 2$p$ $^2P_{1/2}$ 290.56 290.83 1.91 186.81 187.01 7.19 130.32 130.37 22.87
4$f$ $^2F_{5/2}$ $\rightarrow$ 2$p$ $^2P_{1/2}$ 289.03 289.05 19.99 185.99 186.01 76.16 129.72 129.75 228.22
3$p$ $^2P_{1/2}$ $\rightarrow$ 2$p$ $^2P_{3/2}$ 391.32 391.45 10.61 251.76 251.86 40.01 175.65 175.68 118.97
3$p$ $^2P_{3/2}$ $\rightarrow$ 2$p$ $^2P_{3/2}$ 391.26 391.40 5.30 251.71 251.81 19.99 175.61 175.63 59.46
4$p$ $^2P_{1/2}$ $\rightarrow$ 2$p$ $^2P_{3/2}$ 290.68 290.93 3.82 186.92 187.11 14.35 130.43 130.47 45.62
4$p$ $^2P_{3/2}$ $\rightarrow$ 2$p$ $^2P_{3/2}$ 290.66 290.92 1.91 186.91 187.11 7.18 130.42 130.46 22.83
4$f$ $^2F_{5/2}$ $\rightarrow$ 2$p$ $^2P_{3/2}$ 289.13 289.14 5.71 186.09 186.11 21.76 129.82 129.84 65.20
4$d$ $^2D_{3/2}$ $\rightarrow$ 3$s$ $^2S_{1/2}$ 926.31 930.34 2.36 615.27 618.08 8.57 438.96 440.83 24.74
4$d$ $^2D_{5/2}$ $\rightarrow$ 3$s$ $^2S_{1/2}$ 926.27 930.30 2.36 615.23 618.05 8.57 438.92 440.79 24.74
4$p$ $^2P_{3/2}$ $\rightarrow$ 3$p$ $^2P_{1/2}$ 1130.07 1132.83 0.54 725.71 727.75 2.02 506.47 506.88 6.18
4$f$ $^2F_{5/2}$ $\rightarrow$ 3$p$ $^2P_{1/2}$ 1107.17 1106.38 2.22 713.41 712.79 8.29 497.54 497.61 24.40
4$p$ $^2P_{1/2}$ $\rightarrow$ 3$p$ $^2P_{3/2}$ 1130.75 1133.41 1.07 726.35 728.33 4.04 507.10 507.44 12.36
4$p$ $^2P_{3/2}$ $\rightarrow$ 3$p$ $^2P_{3/2}$ 1130.54 1133.24 0.54 726.16 728.16 2.02 506.91 507.28 6.18
4$f$ $^2F_{5/2}$ $\rightarrow$ 3$p$ $^2P_{3/2}$ 1107.62 1106.77 0.63 713.84 713.86 2.36 497.96 497.98 6.95
4$f$ $^2F_{7/2}$ $\rightarrow$ 3$p$ $^2P_{3/2}$ 1107.61 1106.77 2.85 713.82 713.15 10.64 497.94 497.96 31.29
4$d$ $^2D_{3/2}$ $\rightarrow$ 3$d$ $^2D_{3/2}$ 1164.01 1170.18 0.33 745.01 748.99 1.27 517.81 520.14 3.80
4$d$ $^2D_{5/2}$ $\rightarrow$ 3$d$ $^2D_{3/2}$ 1163.95 1170.13 0.095 744.95 748.94 0.36 517.75 520.08 1.08
4$d$ $^2D_{3/2}$ $\rightarrow$ 3$d$ $^2D_{5/2}$ 1164.13 1170.33 0.14 745.13 749.11 0.55 517.93 520.28 1.63
4$d$ $^2D_{5/2}$ $\rightarrow$ 3$d$ $^2D_{5/2}$ 1164.08 1170.27 0.38 745.08 749.06 1.41 517.88 520.22 4.34
------------------ --------------- ------------------ --------- ---------- ------- ------------ -------- ------------ -------- -------- --------
: Transition probabilities (in 10$^{4}$ s$^{-1}$) of $E2$ transitions along with corresponding transition wavelengths (in Å) for C$^{3+}$, N$^{4+}$, and O$^{5+}$.[]{data-label="tab:results2"}
\
After validating the quality and accuracy of our method and calculations for unscreened ions, i.e., with $\mu = 0$, several spectroscopic properties for Li-like ions are calculated for different values of Debye screening parameter $\mu$. In most of the scenario, the ions are presented in low density plasmas where all the spectroscopic properties are affected by the plasma atmosphere which are quantified by the Debye screening parameter $\mu$. We have chosen the values of $\mu$ ranging from 0 to 0.175 a.u. in interval of 0.025 a.u.. The ions become unstable after $\mu = 0.175$ a.u.. It is mentioned already in the “Theory" ection that $\mu$ is a function of ion density $n_{\textrm{ion}}$ and plasma temperature $T$. Therefore, the above ranges of $\mu$ mimic weakly coupled plasma; for example, $T = 10^{6}~\textrm{K}$ and $n_{\textrm{ion}} \sim 10^{22}~{\textrm{cm}^{-3}}$ correspond to $\mu = 0.15$ a.u. [@das2012; @sahoo2006]. This type of condition can be achieved in the laboratory plasmas for high temperature [@sahoo2006]. The relative variations in the IPs as a function of $\mu$ for C$^{3+}$, N$^{4+}$, and O$^{5+}$ ions are presented in Fig. \[fig1\]. For convenience, we define relative variation in any spectroscopic property of interest here, say $\textbf{O}$, as $$\textbf{Relative variation in O} = \frac{[\textbf{O}(\mu \neq 0 )
- \textbf{O}(\mu = 0)] \times 100}{\textbf{O}(\mu = 0)}.$$ It is evident from the Fig. \[fig1\] that as the ion density increases or temperature decreases, i.e., the screening strength $\mu$ increases, IPs decrease linearly and the systems become less and less stable. This is because, the screening of the nuclear charge increases with the increase of $\mu$ and hence, the attractive nuclear Coulomb potential at the valence electron decreases. This particular fact can be attributed as [*continuum lowering*]{} for the system surrounded in plasma environment. A similar trend was observed already for Li and Li-like ions in the presence of Debye plasma environment [@sahoo2006; @das2012]. In addition, this figure shows that the fall in IP decreases with increasing ionic charge, i.e., the IP of C$^{3+}$ decreases more rapidly than the IPs of N$^{4+}$, and O$^{5+}$ within a same interval of $\mu$. With increasing nuclear charge $Z$ of an isoelectronic sequence, the valence electron comes closer to the nucleus and hence can defend more effectively the screening of the plasma environment.
![(Color online) Relative variations of ionization potentials with Debye screening parameter $\mu$ for C$^{3+}$, N$^{4+}$, and O$^{5+}$.[]{data-label="fig1"}](fig1.eps){width="12cm"}
Figure \[fig2\] presents the relative variations in the excitation energies (EEs) of different low-lying excited states for $\mu = 0.075$ a.u. for the Li-like ions. For this value of $\mu$, the effect of screening on the EEs is found optimum. However, we have observed similar trends in the EEs for any other values of $\mu$ within the given range of weakly coupled plasma. It is evident from this figure that due to the screening, the ground-state transitions from the states 2$p$ ${^2}P_{1/2,3/2}$ are blue-shifted, and from the others are red-shifted. This figure further reflects that the shift in the EEs decreases as the nuclear charge increases. These shifts in the EEs can be attributed by the quantum confinement and electron screening in the presence of the plasma environment [@li2008influence]. It is well known that the quantum defect decreases as the orbital angular momentum quantum number increases, and for angular momentum quantum number equal and larger than two, the quantum defect is almost zero. Therefore, electronic states with higher angular momentum quantum number among the same principle quantum number experience relatively less effect of $\mu$. For example, EEs for 3$d$ $^{2}D_{3/2,5/2}$ are less perturbed than the states 3$s$ $^{2}S_{1/2}$ and 3$p$ $^{2}P_{1/2,3/2}$ due to screening. In addition, the effect of quantum confinement is the same for the fine structure states. The figure further confirms that the quantum confinement is more pronounced for higher excited states than relatively low-lying states [@li2008influence].
![(Color online) Relative variations of excitation energies of different excited states with Debye screening parameter $\mu = 0.075$ a.u. for C$^{3+}$, N$^{4+}$, and O$^{5+}$.[]{data-label="fig2"}](fig2.eps){width="12cm"}
The relative variations of oscillator strengths with respect to $\mu$ values for the most strong 2$s$ $^{2}S_{1/2}$ $\rightarrow$ 2$p$ $^{2}P_{1/2,3/2}$ transitions of C$^{3+}$, N$^{4+}$, and O$^{5+}$ ions are shown in Fig. \[fig2\]. The relative variations increase monotonically as $\mu$ increases and the effect is the same for both the transitions. For a particular value of $\mu$, the screening effect on the oscillator strengths decreases as nuclear charge increases. However, the transition energies corresponding to these transitions change significantly, which reflect in the total change in the oscillator strengths as observed in Fig. \[fig3\]. This present trend of the $E1$ oscillator strengths in Debye plasma has been reported in the recent past [@xie2012energy; @das2012]. Our present findings show that the lifetimes of 2p $^{2}P_{1/2,3/2}$ states decrease with increasing plasma strength $\mu$. The lifetimes of these states depend on the third power of wavelengths of the associated $E1$ transitions to the ground state. This enhances the screening effect on the lifetimes with respect to the oscillator strengths which depend inversely on the first power of the corresponding wavelengths.
![(Color online) Relative variations of oscillator strengths of $E1$ transitions with Debye screening parameter $\mu$ for C$^{3+}$, N$^{4+}$, and O$^{5+}$. Panels (a) and (b) correspond to ${2s}~^{2}S_{1/2} \rightarrow
{2p}~^{2}P_{1/2}$ and ${2s}~^{2}S_{1/2} \rightarrow
{2p}~^{2}P_{3/2}$ transitions, respectively.[]{data-label="fig3"}](fig3.eps){width="17cm"}
The transition probabilities ($A^{E2}$) of the $E2$ transition 3$d$ $^{2}D_{3/2}$$\rightarrow$ 2$s$ $^2S_{1/2}$ along with the corresponding transition wavelengths for different values of $\mu$ are presented in Table \[tab:results3\]. It is evident from the table that the $E2$ transition wavelengths are red shifted with increasing value of the screening strength. A similar trend is observed for the 3$d$ $^{2}D_{5/2}$$\rightarrow$ 2$s$ $^2S_{1/2}$ transition, which is not presented here. The relative variations in the transition probabilities for both these $E2$ transitions as a function of $\mu$ are depicted in Fig. \[fig4\]. Both these transitions are chosen due to their comparatively higher probabilities with respect to other transitions as evident from Table II. The systematic decrease of the transition probabilities with increasing $\mu$ is observed from the figure. One can also visualize that the relative decrease in $E2$ transition probabilities for a particular value of screening parameter is more for C$^{3+}$ compared to O$^{5+}$. If one compares the relative changes in $E1$ versus $E2$ transition amplitudes as a function of $\mu$, one finds that the latter is more influenced by $\mu$ than the former. This behavior can be explained from their radial dependence. The amplitude of $E1$ transition has $r$ dependence, where that of $E2$ transition has $r^2$ dependence. Due to this, $E2$ transition amplitude depends on further field region compared to $E1$ transition amplitude from the nucleus and hence, $E2$ transition amplitude is relatively more affected by the screening. At this point, it is important to emphasis that $E2$ transitions are the effective processes in the low density hot plasmas. Therefore, estimations of the influence of nuclear charge screening in the $E2$ transition probabilities along with their wavelengths are of great importance in plasma modeling.
------- ------------- ------------ ------------- ------------ ------------- -----------
$\mu$
$ \lambda $ $ A^{E2} $ $ \lambda $ $ A^{E2} $ $ \lambda $ $ A^{E2}$
0 307.71 44.03 206.36 150.17 148.16 413.17
0.025 308.22 43.68 206.57 149.47 148.26 411.92
0.05 309.69 42.70 207.18 147.46 148.56 408.29
0.075 312.10 41.14 208.17 144.26 149.04 402.46
0.1 315.44 39.06 209.53 139.95 149.70 394.56
0.125 319.75 36.48 211.27 134.62 150.54 384.72
0.15 325.13 33.42 213.40 128.27 151.56 373.04
0.175 331.71 29.87 215.94 121.00 152.76 359.60
------- ------------- ------------ ------------- ------------ ------------- -----------
: Effect of $\mu$ (in a.u.) on $A^{E2}_{{3d}~^{2}D_{3/2} \rightarrow {2s}~^{2}S_{1/2}}$ (in 10$^{4}$ s$^{-1}$) and corresponding transition wavelengths (in Å) for C$^{3+}$, N$^{4+}$, and O$^{5+}$.[]{data-label="tab:results3"}
\
![(Color online) Relative variations of transition probabilities of $E2$ transitions with Debye screening parameter $\mu$ for C$^{3+}$, N$^{4+}$, and O$^{5+}$. Panels (a) and (b) correspond to ${3d}~^{2}D_{3/2} \rightarrow
{2s}~^{2}S_{1/2}$, and (b) ${3d}~^{2}D_{5/2}
\rightarrow {2s}~^{2}S_{1/2}$ transitions, respectively.[]{data-label="fig4"}](fig4.eps){width="15cm"}
Conclusion
==========
We have investigated the influence of Debye screening of nuclear charges due to the presence of free electrons and ions in plasma medium on the ionization potentials, excitation energies, $E1$ oscillator strengths, and $E2$ transition probabilities of C$^{3+}$, N$^{4+}$, and O$^{5+}$. Especially, the study on $E2$ transitions can be considered a useful tool to model low density and high temperature plasmas. The transition wavelengths are mainly affected by this screening, which characterizes the screening effects on the other associated spectroscopic properties like the oscillator strengths and transition probabilities. Due to the high abundances of these Li-like ions in various astrophysical systems, we hope our investigations will be useful to the astrophysicist in the near future.
Pradip Kumar Mondal and Narendra Nath Dutta recognize financial help from the Council of Scientific and Industrial Research (CSIR), India.
[10]{}
P. Emma, R. Akre, J. Arthur, R. Bionta, C. Bostedt, J. Bozek, A. Brachmann, P. Bucksbaum, R. Coffee, F. J. Decker, Y. Ding, D. Dowell, S. Edstrom, A. Fisher, J. Frisch, S. Gilevich, J. Hastings, G. Hays, P. Hering, Z. Huang, R. Iverson, H. Loos, M. Messerschmidt, A. Miahnahri, S. Moeller, H. D. Nuhn, G. Pile, D. Ratner, J. Rzepiela, D. Schultz, T. Smith, P. Stefan, H. Tompkins, J. Turner, J. Welch, W. White, J. Wu, G. Yocky, and J. Galayda, Nature Photonics [**4**]{}, 641 (2010).
T. Ishikawa, H. Aoyagi, T. Asaka, Y. Asano, N. Azumi, T. Bizen, H. Ego, K. Fukami, T. Fukui, Y. Furukawa, S. Goto, H. Hanaki, T. Hara, T. Hasegawa, T. Hatsui, A. Higashiya, T. Hirono, N. Hosoda, M. Ishii, T. Inagaki, Y. Inubushi, T. Itoga, Y. Joti, M. Kago, T. Kameshima, H. Kimura, Y. Kirihara, A. Kiyomichi, T. Kobayashi, C. Kondo, T. Kudo, H. Maesaka, X. M. Marechal, S. Masuda, T.and Matsubara, T. Matsumoto, T. Matsushita, S. Matsui, M. Nagasono, N. Nariyama, H. Ohashi, T. Ohata, T. Ohshima, S. Ono, Y. Otake, C. Saji, T. Sakurai, T. Sato, K. Sawada, T. Seike, K. Shirasawa, T. Sugimoto, S. Suzuki, S. Takahashi, H. Takebe, K. Takeshita, K. Tamasaku, H. Tanaka, R. Tanaka, T. Tanaka, T. Togashi, K. Togawa, A. Tokuhisa, H. Tomizawa, K. Tono, S. K. Wu, M. Yabashi, M. Yamaga, A. Yamashita, K. Yanagida, C. Zhang, T. Shintake, H. Kitamura, and N. Kumagai, Nature Photonics [**6**]{}, 540 (2012).
A. Rousse, C. Rischel, and J. C. Gauthier, Rev. Mod. Phys. [**73**]{}, 17 (2001).
I. McKinnie and H. Kapteyn, Nature Photonics [**4**]{}, 149 (2010).
T. Popmintchev, M. C. Chen, D. Popmintchev, P. Arpin, S. Brown, S. Ali[š]{}auskas, G. Andriukaitis, T. Bal[č]{}iunas, O. D. M[ü]{}cke, A. Pugzlys, A. Baltuska, B. Shim, S. E. Schrauth, A. Gaeta, C. Hernandez-Garcia, L. Plaja, A. Becker, A. Jaron-Becker, M. M. Murnane, and H. C. Kapteyn, Science [**336**]{}, 1287 (2012).
S. M. Vinko, O. Ciricosta, B. I. Cho, K. Engelhorn, H. K. Chung, C. R. D. Brown, T. Burian, J. Chalupsk[y]{}, R. W. Falcone, C. Graves, V. Hajkova, A. Higginbotham, L. Juha, J. Krzywinski, H. J. Lee, M. Messerschmidt, Y. Murphy, C. D. Ping, A. Scherz, W. Schlotter, S. Toleikis, J. J. Turner, L. Vysin, T. Wang, B. Wu, U. Zastrau, D. Zhu, R. W. Lee, P. A. Heimann, B. Nagler, and J. S. Wark, Nature [**482**]{}, 59 (2012).
O. Ciricosta, S. M. Vinko, H. K. Chung, B. I. Cho, C. R. D. Brown, T. Burian, J. Chalupsk[y]{}, K. Engelhorn, R. W. Falcone, C. Graves, V. Hajkova, A. Higginbotham, L. Juha, J. Krzywinski, H. J. Lee, M. Messerschmidt, C. D. Murphy, Y. Ping, D. S. Rackstraw, A. Scherz, W. Schlotter, S. Toleikis, J. J. Turner, L. Vysin, T. Wang, B. Wu, U. Zastrau, D. Zhu, R. W. Lee, P. Heimann, B. Nagler, and J. S. Wark, Phys. Rev. Lett. [**109**]{}, 65002 (2012).
M. Murnane, H. Kapteyn, M. Rosen, and R. Falcone, Science [**251**]{}, 531 (1991).
D. Riley, L. A. Gizzi, F. Y. Khattak, A. J. Mackinnon, S. M. Viana, and O. Willi, Physical Review Letters [**69**]{}, 3739 (1992).
G. Mourou and D. Umstadter, Physics of Fluids B [**4**]{}, 2315 (1992).
F. J. Rogers and C. A. Iglesias, Science [**263**]{}, 50 (1994).
K. Nazir, S. J. Rose, A. Djaoui, G. J. Tallents, M. G. Holden, P. A. Norreys, P. Fews, J. Zhang, and F. Failles, Applied Physics Letters [**69**]{}, 3686 (1996).
J. Workman, M. Nantel, A. Maksimchuk, and D. Umstadter, Applied Physics Letters [**70**]{}, 312 (1997).
M. Nantel, G. Ma, S. Gu, C. Y. Cote, J. Itatani, and D. Umstadter, Physical Review Letters [**80**]{}, 4442 (1998).
N. C. Woolsey, B. A. Hammel, C. J. Keane, C. A. Back, J. C. Moreno, J. K. Nash, A. Calisti, C. Mosse, R. Stamm, B. Talin, A. Asfaw, L. S. Klein, and R. W. Lee, Physical Review E [**57**]{}, 4650 (1998).
A. Saemann, K. Eidmann, I. E. Golovkin, R. C. Mancini, E. Andersson, E. F[ö]{}rster, and K. Witte, Physical Review Letters [**82**]{}, 4843 (1999).
C. A. Rouse, Physical Review A [**4**]{}, 90 (1971).
U. Gupta and A. K. Rajagopal, Physics Reports [**87**]{}, 259 (1982).
J. Seidel, S. Arndt, and W. Kraeft, Physical Review E [**52**]{}, 5387 (1995).
D. Ray and P. K. Mukherjee, Journal of Physics B [**31**]{}, 3479 (1998).
Y. D. Jung, European Physical Journal D [**7**]{}, 249 (1999).
D. Ray, Physical Review E [**62**]{}, 4126 (2000).
J. Pang, G. Han, Z. Wu, and S. Li, Journal of Physics B [**35**]{}, 2117 (2002).
B. Saha, P. K. Mukherjee, and G. H. F. Diercksen, Astronomy and Astrophysics [**396**]{}, 337 (2002).
H. Okutsu, T. Sako, K. Yamanouchi, and G. Diercksen, Journal of Physics B [**38**]{}, 917 (2005).
B. Saha, P. K. Mukherjee, D. Bielinska-Waz, and J. Karwowski, Journal of Quantitative Spectroscopy and Radiative Transfer [**92**]{}, 1 (2005).
B. Saha and S. Fritzsche, Physical Review E [**73**]{}, 036405 (2006).
Y. Li, J. Wu, Y. Hou, and J. Yuan, Journal of Physics B [**41**]{}, 145002 (2008).
S. Bhattacharyya, A. N. Sil, S. Fritzsche, and P. K. Mukherjee, European Physical Journal D [**46**]{}, 1 (2008).
J. K. Saha, S. Bhattacharyya, T. K. Mukherjee, and P. K. Mukherjee, Journal of Physics B [**42**]{}, 245701 (2009).
S. Paul and Y. K. Ho, Physics of Plasmas [**16**]{}, 063302 (2009).
Y. Y. Qi, Y. Wu, J. G. Wang, and Y. Z. Qu, Physics of Plasmas [**16**]{}, 023502 (2009).
A. N. Sil, J. Anton, S. Fritzsche, P. K. Mukherjee, and B. Fricke, European Physical Journal D [**55**]{}, 645 (2009).
C. Gao, J. Zeng, and J. Yuan, High Energy Density Physics [**7**]{}, 54 (2011).
L. Y. Xie, J. G. Wang, R. K. Janev, Y. Z. Qu, and C. Z. Dong, European Physical Journal D [**66**]{}, 125 (2012).
H. R. Griem, , New York, Academic Press, 1974.
W. Jask[ó]{}lski, Physics Reports [**271**]{}, 1 (1996).
S. Bernitt, G. V. Brown, J. K. Rudolph, R. Steinbr�gge, A. Graf, M. Leutenegger, S. W. Epp, S. Eberle, K. Kubicek, V. M�ckel, M. C. Simon, E. Tr�bert, E. W. Magee, C. Beilmann, N. Hell, S. Schippers, A. M�ller, S. M. Kahn, A. Surzhykov, Z. Harman, C. H. Keitel, J. Clementson, F. S. Porter, W. Schlotter, J. J. Turner, J. Ullrich, P. Beiersdorfer, and J. R. Crespo L�pez-Urrutia, Nature [**492**]{}, 225 (2012).
S. Ichimaru, Reviews of Modern Physics [**54**]{}, 1017 (1982).
M. S. Murillo and J. C. Weisheit, Physics Reports [**302**]{}, 1 (1998).
T. Kawachi, T. Fujimoto, and G. Csanak, Physical Review E [**51**]{}, 1428 (1995).
S. N. Nahar, A. K. Pradhan, and H. L. Zhang, The Astrophysical Journal Supplement Series [**131**]{}, 375 (2000).
T. W. Gorczyca, I. Dumitriu, M. F. Haso[ğ]{}lu, K. T. Korista, N. R. Badnell, D. W. Savin, and S. T. Manson, The Astrophysical Journal Letters [**638**]{}, L121 (2006).
S. Suckewer, H. P. Eubank, R. J. Goldston, J. McEnerney, N. R. Sauthoff, and H. H. Towner, Nuclear Fusion [**21**]{}, 1301 (1981).
M. Godefroid and C. F. Fischer, Journal of Physics B [**17**]{}, 681 (1984).
G. C. Das and J. Sarma, Physics of Plasmas [**5**]{}, 3918 (1998).
J. M. Fontenla, K. S. Balasubramaniam, and J. Harder, The Astrophysical Journal [**667**]{}, 1243 (2007).
M. A. Barstow, D. D. Boyce, B. Y. Welsh, R. Lallement, J. K. Barstow, A. E. Forbes, and S. Preval, The Astrophysical Journal [**723**]{}, 1762 (2010).
R. F. Ferrero, C. M. Dur[á]{}n, J. L. Halbwachs, and A. M. C. Cubeiro, The Astronomical Journal [**143**]{}, 28 (2011).
J. X. Prochaska, B. Weiner, H. W. Chen, J. Mulchaey, and K. Cooksey, The Astrophysical Journal [**740**]{}, 91 (2011).
A. J. Fox, The Astrophysical Journal [**730**]{}, 58 (2011).
M. A. de Avillez and D. Breitschwerdt, Astronomy and Astrophysics [**539**]{}, L1 (2012).
G. Peach, H. E. Saraph, and M. J. Seaton, Journal of Physics B [**21**]{}, 3669 (1988).
Y. Ralchenko, A. E. Kramida, J. Reader, and N. A. Team, National Institute of Standards and Technology, Gaithersburg, MD (2012).
A. Burgess, M. C. Chidichimo, and J. A. Tully, Astronomy and Astrophysics Supplement Series [**131**]{}, 145 (1998).
C. Allende Prieto, D. L. Lambert, and M. Asplund, The Astrophysical Journal [**573**]{}, L137 (2002).
C. Y. Chen, K. Wang, M. Huang, Y. S. Wang, and Y. M. Zou, Journal of Quantitative Spectroscopy and Radiative Transfer [ **111**]{}, 843 (2010).
L. H. Andersen, G. Y. Pan, H. T. Schmidt, N. R. Badnell, and M. S. Pindzola, Physical Review A [**45**]{}, 7868 (1992).
L. Qu, Z. Wang, and B. Li, The European Physical Journal D [**5**]{}, 173 (1999).
J. M. Laming and U. Feldman, The Astrophysical Journal [**591**]{}, 1257 (2003).
M. Das, R. K. Chaudhuri, S. Chattopadhyay, U. S. Mahapatra, and P. K. Mukherjee, Journal of Physics B [**44**]{}, 165701 (2011).
M. Das, M. Das, R. K. Chaudhuri, and S. Chattopadhyay, Physical Review A [**85**]{}, 042506 (2012).
R. K. Chaudhuri, S. Chattopadhyay, and U. Sinha Mahapatra, Physics of Plasmas [**19**]{}, 2701 (2012).
J. C. Stewart and K. D. Pyatt Jr, The Astrophysical Journal [**144**]{}, 1203 (1966).
D. Bieli[ń]{}ska-Wa�[ż]{}, J. Karwowski, B. Saha, and P. K. Mukherjee, Physical Review E [**69**]{}, 016404 (2004).
N. N. Dutta and S. Majumder, Physical Review A [**85**]{}, 032512 (2012).
A. I. Akhiezer, I. A. Akhiezer, R. A. Polovin, A. G. Sitenko, and K. N. Stepanov, , volume 1, Oxford: Pergamon, 1975.
I. Lindgren and J. Morrison, , Berlin Springer-Verlag, 1985.
I. Lindgren and D. Mukherjee, Physics Reports [**151**]{}, 93 (1987).
M. A. Haque and D. Mukherjee, Journal of Chemical Physics [**80**]{}, 5058 (1984).
S. Pal, M. Rittby, R. J. Bartlett, D. Sinha, and D. Mukherjee, Chemical Physics Letters [**137**]{}, 273 (1987).
S. Pal, M. Rittby, R. J. Bartlett, D. Sinha, and D. Mukherjee, Journal of Chemical Physics [**88**]{}, 4357 (1988).
E. Eliav, U. Kaldor, and Y. Ishikawa, Physical Review A [**50**]{}, 1121 (1994).
T. A. Isaev, A. N. Petrov, N. S. Mosyagin, A. V. Titov, E. Eliav, and U. Kaldor, Physical Review A [**69**]{}, 30501 (2004).
C. Sur, K. V. P. Latha, B. K. Sahoo, R. K. Chaudhuri, B. P. Das, and D. Mukherjee, Physical Review Letters [**96**]{}, 193001 (2006).
G. Dixit, B. K. Sahoo, R. K. Chaudhuri, and S. Majumder, Physical Review A [**76**]{}, 042505 (2007).
G. Dixit, H. S. Nataraj, B. K. Sahoo, R. K. Chaudhuri, and S. Majumder, Physical Review A [**77**]{}, 012718 (2008).
G. Dixit, H. S. Nataraj, B. K. Sahoo, R. K. Chaudhuri, and S. Majumder, Journal of Physics B [**41**]{}, 025001 (2008).
R. Pal, M. S. Safronova, W. R. Johnson, A. Derevianko, and S. G. Porsev, Physical Review A [**75**]{}, 042515 (2007).
B. K. Mani and D. Angom, Physical Review A [**81**]{}, 042514 (2010).
A. Szabo and N. S. Ostlund, , Dover Pubns, 1996.
I. P. Grant, Journal of Physics B [**7**]{}, 1458 (1974).
W. R. Johnson, D. R. Plante, and J. Sapirstein, Advances In Atomic, Molecular, and Optical Physics , 255 (1995).
R. K. Chaudhuri, P. K. Panda, H. Merlitz, B. P. Das, U. S. Mahapatra, and D. Mukherjee, Journal of Physics B [**33**]{}, 5129 (2000).
F. A. Parpia and A. K. Mohanty, Physical Review A [**46**]{}, 3735 (1992).
F. A. Parpia, C. F. Fischer, and I. P. Grant, Computer Physics Communications [**175**]{}, 745 (2006).
H. Elabidi, S. Sahal-Br[é]{}chot, M. S. Dimitrijevi[ć]{}, and N. Ben Nessib, Monthly Notices of the Royal Astronomical Society [**417**]{}, 2624 (2011).
C. Sur and R. K. Chaudhuri, Journal of Physics B [**40**]{}, 4307 (2007).
U. Feldman, Physica Scripta [**24**]{}, 681 (1981).
M. M. Dworetsky, C. M. Jomaron, and C. A. Smith, Astronomy and Astrophysics [**333**]{}, 665 (1998).
S. Sahoo and Y. K. Ho, Physics of Plasmas [**13**]{}, 063301 (2006).
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
-
-
bibliography:
- 'LuNG.bib'
title: Synthetic Lung Nodule 3D Image Generation Using Autoencoders
---
Lung nodules, CT scan, machine learning, 3D image, image generation, autoencoder
Introduction {#sec:introduction}
============
Motivation {#sec:motivation}
==========
To improve early detection and reduce lung cancer mortality rates, the research community needs to improve lung nodule detection even given low resolution images and a small number of sample images for training. The images have low resolution because low radiation dosages allow for screening to be performed more frequently to aid in early detection, but the low radiation dosage limits the spatial resolution of the image. The number of training samples is small due in part to patient privacy concerns but is also related to the rate at which new medical technology is being created which generates a need for new training data on the new technology. Our primary goal is to create 3D voxel images that are within the broad class of legal nodule shapes that may be generated from a CT scan.
With the goal of creating improved images for training, we evaluate nodules generated from our trained network using the same software that analyzes the CT scans for lung nodules. Given the availability of ’accepted’ new nodules, we test augmenting the training set with these nodules to improve generality of the network. The feedback process we explore includes a nodule reconnection step (to insure final nodules are fully connected in 3D space) followed by a pass through the analyzer which will prune the generated set to keep 3D nodule feature means close to the original limited training set. The need to avoid overfitting the network for a small set of example images, as well as learning a 3D image category by examples, guided many of the network architecture decisions presented below.
While one goal of our work is to demonstrate the possibility to create a family of images which have *computed characteristics* (e.g., elongation, volume) that fit within a particular distribution range; another goal is to generate novel images that are similar to an observed input nodule image. Hence, in addition to creating a generator network, we shall create a feature network that can receive a seed image as input and produce as outputs values for the generator that reproduce the seed image. The goal of generating images related to a given input image motivates our inclusion of the reconnection algorithm. Other generative networks will prune illegal outputs as part of their use model [@Cummins17], but we wanted to provide more guarantee of valid images when exploring the feature space near a given sample input. The goal of finding latent feature values for existing images leads naturally to an autoencoder architecture for our neural network.
Generative adversarial networks (GANs) [@li_sig17] and variational autoencoders (VAEs) [@Doersch16] are two sensible approaches to generate synthetic images from a training set and could intuitively be applied to our problem. However, traditional GANs do not provide a direct mapping from source images into the generator input feature space [@GANGuide], which limits the ability to generate images similar to a specific input sample or require possibly heavy filtering of “invalid” images produced by the network. In contrast, using an autoencoder-based approach as we develop below allows to better explore the feature space near the seed images. A system that combines the training of an autoencoder with a discriminator network such as GRASS [@li_sig17] would allow some of the benefits of GAN to be explored relative to our goals. However, our primary goal is not to create images that match the distribution of a training set as determined by a loss function. As we show in section \[sec:metrics\], our goal can be summarized as creating novel images that are within a category acceptable to an automated nodule analyzer. As such, we strive to generate images that are not identical to the source images but fit within a broad category learned by the network.
A similar line of reasoning can be applied to VAEs relative to our goals. Variational autoencoders map the distribution of the input to a generator network to allow for exploration of images within a distributional space. In our work, we tightly constrain our latent feature space so that our training images map into the space but the space itself may not match the seed distribution exactly to aid in the production of novel images. Like GANs, there are ways to incorporate VAEs into our framework, and to some extent our proposed approach is a form of variational autoencoder, although with clear differences in both the training and evaluation loop, as developed below. Our work demonstrates one sensible approach for a full end-to-end system to create synthetic 3D images that can effectively cover the feature space of 3D lung nodules reconstructed via compressive sensing.
The LuNG System {#sec:contribution}
===============
The LuNG system is based on a neural network trained to produce realistic 3D lung nodules from a small set of seed examples to help improve automated cancer screening. To provide a broader range of legal images, guided training is used in which each nodule is modified to create 15 additional training samples. We call the initial nodule set, of which we were provided 51 samples, the ’seed’ nodules. The ’base’ nodules include 15 modified samples per seed nodule for a total of 816 samples. The base nodules are used to train an autoencoder neural network with 3 latent feature neurons in the bottleneck layer. The output of the autoencoder goes through a reconnection algorithm to increase the likelihood that viable fully connected nodules are being generated. A nodule analyzer program then extracts relevant 3D features from the nodules and prunes away nodules outside the range of interesting feature values. We use the ALNSB [@Shen15] nodule analyzer and classifier code for LuNG. The accepted nodules are the final output of LuNG for use in classifier training or medical evaluation. Given this set of generated images which have been accepted by the analyzer, we explore adding them to the autoencoder training set to improve the generality of the generator. We explore having a single generator network or 2 networks that exchange training samples.
Input images {#sec:inputimg}
------------
The input dataset comes from the Automatic Lung Screening Benchmark (ALNSB) [@Shen15], produced by the NSF Center for Domain-Specific Computing [@cdsc-web]. This pipeline, shown in figure \[fig:pipe\], is targeting the automatic reconstruction of 3D CT scans obtained with reduced (low-dose) radiation, i.e., reducing the number of samples taken by the machine. Compressive sensing is used to reconstruct the initial 3D image of a lung, including all artifacts such as airways, etc. A series of image processing steps are performed to isolate all nodules that could lie along the tissue. Then, each of the 3D candidate nodules is analyzed to obtain a series of domain-specific metrics, such as elongation, volume, surface area, etc. An initial filtering is done based on static criteria (e.g., volume greater than $4 \text{mm}^3$) to prune nodules that are not suspicious for potential cancerous origin. Eventually, the remaining nodules are fed to a trained SVM-based classifier, which classifies the nodules as potentially cancerous or not. The end goal of this process is to trigger a high-resolution scan for only the regions containing suspicious nodules, while the patient is still on the table.
This processing pipeline is extremely specific regarding both the method used to obtain and reconstruct the image, via compressive sensing, and the filtering imposed by radiologists regarding the nodule metrics and their values about potentially cancerous nodules. In this paper, we operate with a single-patient data as input, that is, a single 3D lung scan. About 2000 nodules are extracted from this single scan, out of which only 51 are true candidates for the classifier. Our objective is to create a family of nodules that are also acceptable inputs to the classifier (i.e., which have not been dismissed early on based on simple thresholds on the nodule metrics), starting from these 51 images. We believe this represents a worst-case scenario where the lack of input images is not even sufficient to adequately train a simple classifier and is therefore a sensible scenario to demonstrate our approach to generating images within a specific acceptable feature distribution.
These 51 seed images represent the general nodule shape that we wish to generate new nodules from. Based on the largest of these 51 images, we set our input image size to 25$\times$28$\times$28mm, which also aligns with other nodule studies [@Li08]. The voxel size from the image processing pipeline is 1.25$\times$0.7$\times$0.7mm, so our input nodules are 20$\times$40$\times$40 voxels. This results in an input to the autoencoder with 32,000 voxel values which can range from 0 to 1.
![Medical image processing pipeline [@cdsc-web][]{data-label="fig:pipe"}](imagepipe.png){width="45.00000%"}
![Six of the 51 seed nodules showing the middle 8 out of 20 2D slices[]{data-label="fig:raw"}](raw.png){width="40.00000%"}
Figure \[fig:raw\] shows 6 of the 51 seed images from the CT scan. Each of the images is centered in the 20$\times$40$\times$40 training size. One of our nodules was slightly too wide and 21 out of 1290 total voxels were clipped; all other nodules fit within the training size. From an original set of 51 images, 816 are generated: 8 copies of each nodule are the 8 possible reflections in X,Y, and Z of the original; and 8 copies are the X,Y, and Z reflections of the original shifted by 0.5 pixels in X and Y. The reflections are still representative of legal nodule shapes to the analyzer, so it improves the generality of the autoencoder to have them included. The 0.5-pixel shift also aids generalization of the network by training it to tolerate fuzzy edges and less precise pixel values. We do not do any resizing of the images as we found through early testing that utilizing the full voxel data resulted in better generated images than resizing the input and output of the autoencoder.
Our initial 51 seed images include 2 that are classified as suspicious nodules. These 2 seed images become 32 images in our base training set, but still provide us with a limited example of potentially cancerous nodules. A primary goal of the LuNG system is to create a wider variety of images for use in classification based on learning a nodule feature space from the full set of 51 input images.
Autoencoder network {#sec:autoencoder}
-------------------
Figure \[fig:autoencoder\] shows the autoencoder structure as well as the feature and generator networks that are derived from it. All internal layers use tanh for non-linearity, which results in a range of -1 to 1 for our latent feature space. The final layer of the autoencoder uses a sigmoid function to keep the output within the 0 to 1 range that we are targeting for voxel values.
We experimented with various sizes for our network and various methods of providing image feedback from the analyzer with results shown in section \[sec:results\]. The network shown in figure \[fig:autoencoder\] had the best overall $Score$.
![Autoencoder and derived feature/generator networks for nodules[]{data-label="fig:autoencoder"}](autoencoder.png){width="45.00000%"}
Our autoencoder is trained initially with the 816 images in our base set. We use Adam [@Kingma14] for stochastic optimization to minimize the mean squared error of the generated 32,000 voxel 3D images. After creating a well-trained autoencoder, the network can be split into feature and generator networks. The feature network can be used to map actual nodules into a latent feature space so that novel images similar to the actual input nodule can be created using the generator network. If stepping is done between the latent feature values of nodule suspected as cancerous and another suspected to be non-cancerous, a skilled neurologist could identify the shape at which the original suspicious nodule would not be considered suspicious to help train and improve an automated classifier. The generator network can also be used to generate fully random images for improving a classifier. For our random generation experiments we use uniform values from -1 to 1 as inputs for the 3 latent feature dimensions. We explore the reasons and benefits of this random distribution in section \[sec:features\].
The autoencoder structure which yielded the best results is not symmetric in that there are fewer layers before the bottleneck layer than after. Like the seminal work by Hinton and Salakhutdinov [@Hinton06], we explored various autoencoder sizes for our problem, but we added in an exploration of non-symmetric autoencoders. We found during hyperparameter testing that a 2-layer feature network (encoder) performed better than a 1-layer or 3-layer network. We suspect that a single layer for the feature network was not optimal due to limiting the feature encoding of the input images to linear combinations of principle components [@Haykin09]. We suspect that 3 layers for our feature network was less optimal than 2 layers due to overfitting the model to our limited training set. Given our goal of generating novel nodule shapes, overfitting is a particular concern and we address this using a network scoring metric discussed in section \[sec:metrics\].
Reconnection algorithm {#sec:reconnect}
----------------------
The autoencoder was trained on single component nodules in that all the ’on’ voxels for the nodule were connected in a single 3D shape. The variation produced by trained generator networks did not always result in a single component, and it is common for generative networks that have a technical constraint to discard output which fails to meet the requirements [@Cummins17]. However, for the use case of exploring the feature space near a known image, we chose to add a reconnection algorithm to our output nodules to minimize illegal outputs. This algorithm insures that for any input to the generative network, a fully-connected nodule is generated.
When the generator network creates an image, a single fully-connected component is usually generated and the reconnection algorithm does not need to be invoked. In the case where multiple components are detected, the algorithm will search through all empty voxels and set a small number of them to connect the components into a single nodule.
Metrics for nodule analyzer acceptance and results scoring {#sec:metrics}
----------------------------------------------------------
The nodule analyzer and classifier computes twelve 3D feature values for each nodule (features such as 3D volume, surface-to-volume ratio, and other data useful for classification). Our statistical approach to this data is related to Mahalanobis distances [@Mahalanobis36], hence we compute the mean and standard deviations on these 12 features for the 51 seed nodules. Random nodules from the generator are fed into the classifier code and accepted to produce similar feature values. This accepted set of images is the most useful image set for further analysis or use in classifier training.
#### Metrics for analyzer acceptance of the images {#metrics-for-analyzer-acceptance-of-the-images .unnumbered}
Using the mean and standard deviation values we create a distance metric $d$ based on concepts similar to the Mahalanobis distance. Given $S$ is the set of 51 seed nodules and $i$ is the index for one of 12 features, $\mu_{Si}$ is the mean value of feature $i$ and $\sigma_{Si}$ is the standard deviation. Given Y is the set of output nodules from LuNG, the running mean for feature $i$ of the nodules being analyzed is $\bar{Y_i}$. Given feature $i$ of a nodule $y$ is $y_i$ then if either (${y_i \geq \mu_{Si}}$ and $\bar{Y_i} \leq \mu_{Si}$) or ($y_i \leq \mu_{Si}$ and $\bar{Y_i} \geq \mu_{Si}$), then the nodule is accepted as it helps $\bar{Y_i}$ trend towards $\mu_{Si}$. In cases where the nodule’s $y_i$ moves $\bar{Y_i}$ away from $\mu_{Si}$, we compute a weighted distance $d$ from $\mu_{Si}$ in multiples of ${\sigma_{Si}}$ using:
$$d = |\frac{y_i + 3*\bar{Y_i} - 4*\mu_{Si}}{\sigma_{Si}}|$$
We compute the probability of keeping a nodule $y$ as $P_{keep}$ which drops as $d$ increases: [$$P_{keep} = \begin{cases}
0.7+\frac{0.9}{d} & \text{if } y_i > \mu_{Si} \text{ and } \bar{Y_i} > \mu_{Si} \text { and } d > 3\\
0.7+\frac{0.9}{d} & \text{if } y_i < \mu_{Si} \text{ and } \bar{Y_i} < \mu_{Si} \text { and } d > 3\\
1 & \text{otherwise} \\
\end{cases}$$ ]{} The specific numerical values used for computing $d$ and $P_{keep}$ were chosen to maximize the number of the original dataset which are accepted by this process while limiting the deviation from the seed features allowed by the generator. When using this process on a random sample from the 816 base nodules, 95% were accepted. Acceptance results for nodules generated by a trained network are provided in section \[sec:results\].
#### Metrics for scoring the accepted image set {#metrics-for-scoring-the-accepted-image-set .unnumbered}
The composite score that we use to evaluate networks for LUNG is comprised of 4 metrics used to combine key goals for our work. We compute the percentage of nodule images randomly generated by the generator that are accepted by the analyzer. For assessing the variation of output images relative to the seed images, we compute a feature distance ${FtDist}$ based on the 12 3D image features used in the analyzer. To track how well the distribution of output images matches the seed image variation, we compute a ${FtM\!M\!S\!E}$ based on the image feature means. The ability of the network to reproduce a given seed image is tracked with the mean squared error of the image output voxels, as is typical for autoencoder image training.
Our metric of variation, ${FtDist}$, is the average distance over all accepted images to the closest seed image in the 12-dimensional analyzer feature space and is scaled in a way similar to Mahalanobis distances. As ${FtDist}$ increases, the network is generating images that are less similar to specific samples in the seed images, hence it is a metric we want to increase with LuNG. Given an accepted set of $n$ images $Y$ and a set of 51 seed images $S$, and given $y_i$ denotes the value of feature $i$ for an image and $\sigma_{Si}$ denotes the standard deviation of feature $i$ within $S$:
$$FtDist = 1/n\sum_{y \in Y}\min_{s \in S}\sqrt{\sum_{i=1}^{12}(\frac{y_i - s_i}{\sigma_{Si}})^2}$$
$FtM\!M\!S\!E$ tracks how closely LuNG is generating images that are within the same analyzer feature distribution as the seed images. It is the difference between the means of the images in $Y$ and $S$ for the 12 3D features. As ${FtM\!M\!S\!E}$ increases, the network is generating images that are increasingly outside the seed image distribution, hence we want smaller values for LuNG. Given $\mu_{Si}$ is the mean of feature $i$ in the set of seed images and $\mu_{Yi}$ is the mean of feature $i$ in the final set of accepted images:
$$FtM\!M\!S\!E = 1/12\sum_{i=1}^{12}(\frac{\mu_{Yi} - \mu_{Si}}{\sigma_{Si}})^2$$
$Score$ is our composite network scoring metric used to compare different networks, hyperparameters, feedback options, and reconnection options. In addition to $FtDist$ and $FtM\!M\!S\!E$, we use $AC$, which is the fraction of generated images which the analyzer accepted, and $M\!S\!E$ which is the mean squared error that results when the autoencoder is used to regenerate the 51 seed nodule images.
$$Score = \frac{FtDist-1}{(FtM\!M\!S\!E+0.1)*(M\!S\!E+0.1)*(1-AC)}$$
$Score$ increases with ${FtDist}$ and ${AC}$ and decreases with ${FtM\!M\!S\!E}$ and $M\!S\!E$. The constants in the equation are based on qualitative assessments of network results; for example, using ${M\!S\!E+0.1}$ means that $M\!S\!E$ values below 0.1 don’t override the contribution of other components and aligns with the qualitative statement that an $M\!S\!E$ of 0.1 yielded visually acceptable images in comparison with the seed images.
Results using $Score$ to evaluate networks and LuNG interface features are discussed further is section \[sec:results\]. Our use of $Score$ to evaluate the entire nodule generation process rates the quality of the random input distribution, the generator network, the reconnection algorithm, the analyzer acceptance, and the interaction of these components into a system. Our use of the analyzer acceptance rate is similar in some functional respects to the discriminator network in a GAN as both techniques are used to identify network outputs as being inside or outside an acceptable distribution.
Updating the training set {#sec:training}
-------------------------
![Interaction between trained autoencoder and nodule analyzer. The images from figure \[fig:raw\] are always part of the training set to the autoencoder. The reconnected images after the network can be seen in figure \[fig:step\]. The analyzer accepted output of LuNG can be seen in figure \[fig:final\].[]{data-label="fig:analyzer"}](analyzer.png){width="45.00000%"}
After a trained generator network produces images which are reconnected and validated by the nodule analyzer, a new training set may optionally be created for the next round of training iterations of the autoencoder. Figure \[fig:analyzer\] diagrams the data loop between the autoencoder and the nodule analyzer. We explored various approaches for augmenting the training set, but ultimately found that our best results came from proper autoencoder sizing and training with only the 816 base images created by adding guided training examples to the original 51 seed images. Although the image feedback into the training set did not improve the LuNG system, we include the description and results to illustrate the drawbacks of the approach.
We were motivated to explore augmentation approaches because we wanted to learn if adding accepted nodules to the training set could improve network $Score$ results by improving $FtDist$. One of the approaches used images from one trained network to augment a second trained network, similar to the multi-adversarial networks discussed in [@Durugkar16]. The intent of this approach is to improve the breadth of the feature space used to represent all legal nodules by importing images that were not generated by the network being trained.
We analyze the feedback approach we considered the best in detail in section \[sec:results\]. In this approach, we train the network for 50,000 iterations then generate 400 output images (less than half the size of the base set). We pass those images through the analyzer to insure they are considered legal and chose a single random reflection of each image to add to the training set. Because we are feeding in the image reflection, we did not find added value in having 2 networks trade generated images - the network itself was generating images whose reflections were novel to both its own feature network as input and its generator network as output. This training feedback is only done for 2 rounds of 25,000 training iterations and then a final 50,000 training iterations train only on the 816 base images.
The augmentation experiments explored whether having some training image variation helps fill the latent feature space with more variation on legal images. The intent of testing multiple approaches was to learn if an analyzer feedback behavior can be found that improves the criteria LuNG is trying to achieve: novel images accepted by the analyzer.
Experimental Results {#sec:results}
====================
Using the $Score$ metric and its components introduced in section \[sec:metrics\], we evaluated various network sizes, the analyzer feedback approaches discussed in section \[sec:training\], and the reconnection algorithm discussed in section \[sec:reconnect\]. Our naming for networks is given with underscores separating neuron counts between the fully connected layers, so 32\_3\_64\_1024 is a network with 4 hidden layers that have 32, 3, 64, and 1024 neurons respectively. As seen by referencing tables \[tab:params\] and \[tab:params2\], depending on the metric which is rated as most important, different architectures would be recommended.
Results for reconnection, feedback options, and depth {#sec:optresults}
-----------------------------------------------------
Table \[tab:params\] shows training results for some of the feedback options discussed in section \[sec:training\]. The networks in this table did not use the reconnection algorithm discussed in section \[sec:reconnect\], so ’illegal’ inputs to the analyzer could occur, reducing the acceptance rate. The MSE column shows 1000 times the mean squared error per voxel for the autoencoder when given the 51 original seed nodules as inputs and targets. The “AC%” column shows what percentage of 400 images randomly generated by the generator network were accepted by the analyzer. The “FtDist” column shows the average minimum distance in analyzer feature space from any image to a seed image. The “FtMMSE” column shows the average mean squared error of all 12 analyzer features between the images and the 51 seed images. “No reflections” is one of our feedback options referring to using the accepted images from the analyzer directly in the training set. “multiple” feedback refers to using 2 autoencoders and having each autoencoder use the accepted images that were output by the other. Using this table and other early results, we observed that the network with no analyzer feedback had overall good metrics, although the FtDist column indicating the novelty of images generated was lower than we would prefer, so we weighed ${FtDist}$ heavier in our final scoring of networks as we explored network sizing.
[ |p[3.5cm]{}||p[0.6cm]{}|p[0.6cm]{}|p[0.7cm]{}|p[1.0cm]{}| ]{}\
Parameters & AC% & MSE & FtDist & FtMMSE\
16\_4\_64\_256\_1024 & 54 & 0.03 & 1.75 & 0.07\
No Feedback & & & &\
16\_4\_64\_256\_1024 & 44 & 0.03 & 2.23 & 0.21\
FB: no reflections & & & &\
16\_4\_64\_256\_1024 & 36 & 0.05 & 2.43 & 0.26\
FB: no reflections, multiple & & & &\
16\_4\_64\_256\_1024 & 29 & 0.04 & 2.68 & 0.45\
FB: 4 reflections, multiple & & & &\
\[tab:params\]
Table \[tab:params2\] shows experiments in which the reconnection algorithm was used. When using the reconnection algorithm, the analyzer always has a full set of 400 images to consider for acceptance, leading to higher acceptance rates. This table includes data on the number of raw generator output images which were clean when generated (one fully connected component) and the number that were inverted (white background with black nodule shape). The fact that deeper generation networks sometimes resulted in inverted output images is an indication that they have too many degrees of freedom and contributed to the decision to limit the depth of our autoencoder. The “1 reflection” feedback label refers to having a single reflected copy of each accepted image used to train the autoencoder for 2 of the 6 rounds. This “1 reflection” feedback was our most promising approach as described in section \[sec:training\].
[ |p[3.9cm]{}||p[0.45cm]{}|p[0.4cm]{}|p[0.55cm]{}|p[0.5cm]{}|p[0.5cm]{}| ]{}\
Parameters & AC% & MSE & FtDist & Clean & Invert\
64\_4\_64\_1024, No Feedback & 85 & 0.08 & 1.78 & 109 & 0\
64\_4\_64\_1024, 1 reflection & 64 & 0.06 & 3.13 & 63 & 0\
64\_4\_64\_256\_1024, No Feedback & 80 & 0.02 & 1.96 & 117 & 2\
64\_4\_64\_256\_1024, 1 reflection & 61 & 0.03 & 4.11 & 77 & 6.5\
\[tab:params2\]
From the results in these 2 tables and other similar experiments, we concluded that the approach in section \[sec:training\], which used analyzer feedback for 2 of the 6 training rounds, had the best general results of the 4 feedback approaches considered. Also, the approach in section \[sec:reconnect\], which will reconnect and repair generator network outputs, yielded 3D images preferable to the legal subset left when the algorithm was not applied. The results of these explorations informed the final constants that we used to create the $Score$ metric for rating networks as described in section \[sec:metrics\].
Results for tuning network sizes {#sec:sizresults}
--------------------------------
We analyzed neuron counts as well as total number of network layers using $Score$. We tested multiple values for the neuron counts in each layer and figure \[fig:scorefb\] shows the results for testing the dimensions in the latent feature space. As can be seen, from the networks tested, the network which yielded the highest score of 176 was 32\_3\_64\_1024, which is the network used to generate the nodule images shown in section \[sec:imagerslt\].
Our final network can train on our data in an acceptable amount of time. Even though our experiments gathered significant intermediate data to allow for image feedback during training, the final 32\_3\_64\_1024 network can be trained in approximately 2 hours. Our system for training has 6 Intel Xeon E5-1650 CPUs at 3.6GHz and an Nvidia GeForce GTX 1060 6GB GPU. Those 2 hours break down as: 10 minutes for creation of 816 base images from 51 seed images, 80 minutes to train for 150,000 epochs on the images, 20 minutes to generate and connect 400 nodules, and 10 minutes to run the analyzer on the nodules. Code tuning would be able to improve the image processing parts of that time, but the training was done using PyTorch [@Pytorch] on the GPU and is already highly optimized. When generating images for practical use, we would recommend training multiple networks and using the results from the network that achieved the highest score.
![$Score$ comparisons between networks that use 816 base images with no analyzer feedback for 150,000 training iterations versus networks that trained for 25,000 iterations on the base images, then added 302 generated nodules to train for 25,000 iterations, then added a different 199 generated nodules to train for 25,000 iterations, and then finished with 75,000 training iterations with no feedback.[]{data-label="fig:scorefb"}](scorefb.png){width="45.00000%"}
Figure \[fig:metrics\] shows the components of $Score$ for the final parameter analysis we did on the network. Note that the MSE metric (mean squared error of the network on training set) continues to decrease with larger networks, but $Score$ is optimal with 3 bottleneck latent feature neurons. Our intuition is that limiting our network to 3 bottleneck neurons results in most of the available degrees of freedom being required for proper image encoding. As such, using a $-1$ to $1$ uniform random distribution as the prior distribution for our generative network creates a variety of acceptable images. The $Score$ metric helps us to tune the system such that we do not require VAE techniques to constrain our random image generation process, although such techniques may be a valuable path for future research.
![There are 4 components used to compute the network score. The component values are scaled as shown so that they can all be plotted on the same scale. []{data-label="fig:metrics"}](metrics.png){width="45.00000%"}
Latent feature results {#sec:features}
----------------------
To visualize the weaknesses of too many bottleneck neurons in our autoencoder network, we plot the 51 seed nodule positions in the latent feature space in figure \[fig:bottle\]. To save space, we only present 2 of the 3 neurons for the final trained network, and we compare their positions to 2 of the 8 neurons from a trained network with 8 bottleneck neurons.
For the network with 3 bottleneck neurons, the plot shows that the 51 seed nodules are relatively well distributed in the 4 quadrants of the plot and the full range of both neurons is used to represent all the input images. For the network with 8 bottleneck neurons, most of the seed nodules map to the upper left quadrant in the plot and the full range of the 2 neurons is not used. This is a symptom of having a network with more degrees of freedom than needed to represent the nodule training space. Our intuition is that this contributes to the high ${FtM\!M\!S\!E}$ measurements shown in figure \[fig:metrics\] for a network with 8 bottleneck neurons. The figures also show how a random -1 to 1 uniform range for our generator will likely result in a higher acceptance rate for images generated with 3 bottleneck neurons versus 8 bottleneck neurons.
![Distribution of images in feature space after training with no feedback[]{data-label="fig:bottle"}](neurons.png){width="50.00000%"}
Image results {#sec:imagerslt}
-------------
The images shown in this section are from a network with 4 hidden layers with 32, 3, 64, and 1024 neurons. The network was trained without analyzer feedback and the output is processed to guarantee fully connected nodules.
Figure \[fig:step\] shows the quality of 3D nodules our network can produce with 6 steps through the 3D bottleneck neuron latent feature space starting at the 2nd nodule from figure \[fig:raw\] and ending at the 4th nodule. First, note that the learned images for the start and end nodules are very similar to the 2nd and 4th input images, validating the MSE data that the network is correctly learning the seed nodules. The 4 internal step images have some relation to the start and end images but depending on the distance between the 2 nodules in latent feature space a variety of shapes may be involved in the steps.
![6 steps through 3D latent feature space between original nodules 2 and 4 from figure \[fig:raw\][]{data-label="fig:step"}](step.png){width="40.00000%"}
Figure \[fig:final\] shows 6 images generated by randomly selecting 3 values between -1 and 1 for the latent feature inputs to the generator network and then being processed by the analyzer to determine acceptance. When using the network to randomly generate nodules (for classification by a trained specialist or training automated classifiers), this is an example of quality final results.
![6 images generated using uniform distribution from 3D feature space after passing nodule analyzer[]{data-label="fig:final"}](final.png){width="40.00000%"}
Image quality for training {#sec:imagerqual}
--------------------------
Figure \[fig:all\] shows the 12 features that are used by the nodule analyzer and demonstrates another key success of our full approach. Characteristics like volume, surface area, and other values are used. We normalized the mean and standard deviation of each feature to 1.0 and the figure shows that the mean and standard deviation of the generated nodules for all 12 features stays relatively close to 1 for our proposed network with no analyzer image feedback. However, when feedback is used, one can see that the nodule features which have some deviation from the mean get amplified even though the analyzer tries to accept nodules in a way that maintains the same mean. For example, “surface area$^3$/volume$^2$” is a measure of the compactness of a shape; the generated images from the network with no feedback tended to have higher surface area to volume than the seed images, and when these images were used for further training the generated images had a mean that was about 2.6 times higher than the seed images and a much higher standard deviation.
![Computed 3D features from nodule analyzer. Our proposed method avoids the deviations shown by the grey and yellow bars.[]{data-label="fig:all"}](features.png){width="50.00000%"}
![Classifier distances to positive and negative centroids of SVM for 1000 analyzer accepted network generated samples. Nodules closer to positive than negative centroid after support vectors are applied are more likely cancerous.[]{data-label="fig:posneg"}](SVM.png){width="45.00000%"}
The ALNSB classifier we interact with uses a support vector machine to map images onto a 2D space representing distances to positive (suspicious) or negative (non-cancerous) centroids. Figure \[fig:posneg\] shows the positive and negative centroid distances for the seed data and 1000 samples of analyzer accepted generated data. Nodules that have a smaller distance to the positive centroid than to the negative centroid are classified as likely cancerous. The general distribution of generated images fits the shape of the seed images rather well, and there is a collection of nodules being generated near the decision line between cancerous and non-cancerous, allowing for improved training based on operator classification of the nodules. Even though the original seed dataset only included 2 nodules eventually classified as potentially cancerous, our approach can use shape data from all 51 nodules to create novel images that can be useful for improving an automated screening system. Note that these centroid distances themselves are not part of the 12 features that are used to filter the nodules, so this figure validates our approach to create images usable for further automated classification work.
Related Work {#sec:related}
============
Improving automated CT lung nodule classification techniques and 3D image generation are areas that are receiving significant research attention.
Recently, Valente et al. provided a good overview of the requirements for Computer Aided Detection systems in medical radiology and they survey the status of recent approaches [@Valente16]. Our aim is to provide a tool which can be used to improve the results of such systems by both decreasing the false positive rate and increasing the true positive rate of classifiers through the use of an increase in nodules for training and analysis. Their survey paper discusses in detail preprocessing, segmentation, and nodule detection steps similar to those used in the ALNSB nodule analyzer/classifier which we used in LuNG.
Li et. al provide a thorough overview of recent approaches to 3D shape generation in their paper “GRASS: Generative Recursive Autoencoders for Shape Structures” [@li_sig17]. While we do not explore the design of an autoencoder with convolutional and deconvolutional layers, the same image generation quality metrics that we teach could be used to evaluate such designs. Tradeoffs between low reproduction error rates and overfitting would have to be considered when setting the network depth and feature map counts in the convolutional layers.
Durugkar et al. describe the challenges of training GANs well and discuss the advantages of multiple generative networks trained with multiple adversaries to improve the quality of images generated [@Durugkar16]. LuNG explored using multiple networks during image feedback experiments. Larsen et al. [@Larsen15] teach a system which combines a GAN with an autoencoder which could be a basis for future work introducing GAN methodologies into the LuNG system by preserving our goal of generating shapes similar to existing seed shapes.
Conclusion {#sec:conclusion}
==========
To produce quality image classifiers, machine learning requires a large set of training images. This poses a challenge for application areas where large training sets are rare, such as for new medical techniques using computer-aided diagnosis of cancerous lung nodules.
In this work we developed LuNG, a lung nodule image generator, allowing us to augment the training dataset of image classifiers with meaningful (yet computer-generated) lung nodule images. Specifically, we have developed an autoencoder-based system that learns to produce 3D images with features that resemble the original training set. LuNG was developed using PyTorch and is fully implemented and automated. We have shown that the 3D nodules generated by this process visually and numerically align well with the general image space presented by the limited set of seed images.
#### Acknowledgment {#acknowledgment .unnumbered}
This work was supported in part by the U.S. National Science Foundation award CCF-1750399.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'F.O. Bunnin [ [^1] ]{}'
- 'A. Shenvi [^2] [ ]{}'
- 'J.Q. Smith [^3] [ ]{}'
bibliography:
- 'bdn.bib'
title: Network Modelling of Criminal Collaborations with Dynamic Bayesian Steady Evolutions
---
**Keywords:** chain event graphs, dynamic non-Gaussian models, Markov processes, probabilistic graphical models, terrorist networks, statistical network models, decision support systems
Acknowledgements {#acknowledgements .unnumbered}
================
This work was funded by The Alan Turing Institute Defence and Security Project G027. Jim Q. Smith was supported by the Alan Turing Institute and funded by the EPSRC \[grant number EP/K03 9628/1\].
[^1]: The Alan Turing Institute, British Library, 96 Euston Road, London NW1 2DB
[^2]: Centre for Complexity Science, University of Warwick, Coventry, CV4 7AL
[^3]: Department of Statistics, University of Warwick, Coventry, CV4 7AL
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'J.L. Goveas and P.D. Olmsted'
title: A Minimal Model for Vorticity and Gradient Banding in Complex Fluids
---
Introduction
============
There is a significant body of experimental evidence documenting the existence of sharp, stable interfaces separating two or more phases or “bands”, in shear flow in complex fluids. This phenomena has been reported in various types of surfactant solutions [@Pine; @Bonn], polymers [@Eiser], liquid crystals [@neutron] and colloidal suspensions [@Zcolloids]. There appears to be a compelling generality between these “phase transitions” in different complex fluids:
- The onset of banding or phase separation manifests itself as a discontinuity in the “flow curve” of the system. The flow curve is the unique relationship between the measured shear stress and the applied shear rate (or vice versa) at steady-state. An experimental flow curve typically contains segments that correspond to homogeneous flow, as well as segments corresponding to inhomogeneous flow. The individual homogeneous bands which make up the inhomogeneous state each have their own homogeneous flow curve, which we shall refer to as a “constitutive curve”. The inhomogeneous flow curve then represents the response of the system, averaged over different spatial regions that occupy different homogeneous flow branches, in proportions to maintain the externally controlled shear stress or shear rate. \[Henceforth we will use the terms stress and shear stress interchangeably, unless otherwise specified.\]
- The transition only occurs above a unique and reproducible critical stress or shear rate.
- The flow curve can be qualitatively different depending on whether the average stress or the average shear rate in the system is held fixed. \[In a typical rheological experiment this is achieved by controlling the torque or angular velocity respectively.\] For intermediate stresses or shear rates, the flow curve usually has multiple branches which are not equally accessible under both stress and shear rate control. For weak and strong flows, the flow curve is single-valued, and the same locus of points is traced out under stress or shear rate control.
- The flow-induced bands have different shear rates or shear stresses, and are generally also distinguished by some combination of different degrees of order and different microstructures.
- The interfaces between the bands may be aligned in the direction of the flow gradient or the flow vorticity. Each banding orientation has its own rheological signature. In shear-thinning systems, for example, a stress plateau in the flow curve usually indicates gradient banding, while extrema in the stress (as a function of shear rate) usually indicate vorticity banding. One of us [@Peter] has constructed possible flow curves based on the banding orientation and the character of phase coexistence (shear-thinning versus shear-thickening).
Gradient banding has been unambiguously observed in solutions of wormlike micelles. In strain-controlled experiments on shear-thinning solutions, a stress plateau coincides with shear-banding in the gradient direction [@Callaghan]. In shear-thickening solutions [@Pine], a gel-like phase can be induced by flow. Under shear rate control the induced phase fills the system at steady-state, resulting in a discontinuous stress jump in the flow curve. Under stress control phase coexistence between solution and gel is observed, the gel fraction being an increasing function of stress. In the corresponding flow curve the shear rate shows a minimum and maximum.
Vorticity banding has been reported in dense colloidal suspensions [@Zcolloids] and surfactant solutions of multilamellar vesicles [@Bonn]. When the shear rate is held fixed, the flow curve shows a maximum and minimum in the stress. Under controlled stress, there is a jump up in shear rate upon increasing stress, and a jump down in shear rate upon decreasing stress. The same qualitative curves have also been observed in surfactant hexagonal phases [@LRamos], although in that case vorticity banding has not yet been explicitly verified. Such behavior is analogous to that of the shear-thickening wormlike micelles, if the roles of stress and shear rate are interchanged. In bcc cubic crystals of triblock copolymers [@Eiser], Eiser *et al.* observe two stress plateaus in the flow curve under controlled shear rate. X-ray diffraction shows that each plateau corresponds to different orientations (relative to the flow direction) of dense planes in the crystal.
In steady-state there can be no acceleration, so the total stress must be divergence free. In planar shear flow, this implies that the shear rate in the vorticity direction and the shear stress in the gradient direction are uniform. Vorticity banding thus corresponds to a scenario where bands share a common shear rate but can have different shear stresses (see Figure \[banding figure\]). Similarly, when bands lie in the gradient direction the stress is uniform across the bands and the shear rate can vary. Most experiments where banding has been observed have been carried out in the curved geometries of cone-and-plate or Couette rheometers. The gaps in these rheometers are usually very thin, and in this limit the flow is approximately a planar shear flow. \[We also note that we consider flows in the low Reynolds number limit.\]
The microscopic mechanisms causing the transitions in all these complex fluids are likely to be highly system specific, and govern the critical shear rate or shear stress, the structure of the flow-induced phases, and the detailed shape of the flow curves. At a macroscopic level, however, there appears to be a high degree of universality between systems. As we have discussed, different complex fluids can produce qualitatively similar flow curves. By simply analysing the shape of these flow curves, we have extracted information about the banding orientation of the transition [@Peter], as well as the stability of the system [@Goveas; @and; @Pine].
The obvious analog to this way of thinking is the well-known Landau-Ginzburg theory of equilibrium phase transitions. A free energy functional consisting of a double-well local free energy and a square gradient term reproduces all the phenomenology of a phase transition in the region of the critical temperature, for many different systems. However, a microscopic theory is required to calculate the Landau coefficients. In this work, we use a multivalued reaction diffusion scheme to construct a general phenomenological theory to describe phase transitions in flow; in the spirit of Landau-Ginzburg theory, such a model could in principle be derived from microscopic theories.
The most difficult step in devising such a nonequilibrium theory is determining the important variables. Typically there are three significant quantities: i) a species concentration, which is a conserved scalar, ii) the momentum density, which is a conserved vector and iii) the stress, which is a non-conserved tensor. The momentum density is described by the Navier-Stokes equation, and its current is the stress. In addition non-conserved “microstructural” order parameters exist which contribute to the stress. These may be scalars such as chain length in wormlike micelle solutions, or tensors such as molecule orientation. Since all of these variables have different relative relaxation times, one must distinguish between slow variables, which require their own equations of motion, and fast variables, which relax quickly to a steady-state value.
The choice of slow variables affects the structure of the equations of motion and the couplings between them, and therefore the dynamics of the system. In models of hydrodynamic instabilities for example, the momentum is considered to be a slow variable. A phenomenological stress constitutive equation is often used: if the stress is taken to be a fast variable, this relation is simply an algebraic function of the rate of strain tensor, such as the Newtonian relation for simple fluids; if the stress is taken to be a slow variable, this relation takes the form of a differential equation, such as the Upper Convected/Oldroyd-B Maxwell model for polymer melts. The hallmark of complex fluid rheology, however, is the coupling between the velocity and/or the stress to the microstructure of the fluid. In microscopic theories, generally an equation of motion is not written for the total stress, but for another slow variable which makes an important contribution to it, such as the director in nematic liquid crystals or the second moment of the configuration tensor in polymer melts.
Schmitt, Marques and Lequeux [@SchmittMarquesLequeux] have classified flow instabilities in complex fluids as “mechanical” or “spinodal” instabilities, using a model where concentration and momentum are the slow variables. If a perturbation to the shear rate first makes the system go unstable, the instability is mechanical, while it is spinodal if the concentration becomes unstable first. Note that “instability” as it is used here refers to a linear instability. Any instability, linear or nonlinear (we return to this issue at the end of the paper), can lead to a macroscopically shear-banded state that resolves the instability.
Shear-banding associated with momentum instabilities have been analyzed in detail at a high (macroscopic) level, using the phenomenological Johnson-Segalman model [@Spenley; @Yuan; @Olmsted; @JS]. Here a non-conserved “polymer” stress tensor, playing the role of the slow variable, is included in the total momentum density, resulting in a multi-valued stress constitutive relation. This model produces gradient banding and a flow curve with a stress plateau, and is considered a reasonable mimic of shear-thinning wormlike micelles. Microscopically derived theories for wormlike micelles [@Cates] and nematic liquid crystalline melts [@Olmsted; @LC] yield a non-monotonic relation similar to the Johnson-Segalman model, but with the benefit of a molecular interpretation.
An alternative caricature to these models has been developed in phenomenological theories for shear-thickening. Originally, Ajdari [@Ajdari] proposed an equation of motion for the position of an interface that separates high and low viscosity phases under shear. By coupling this equation with conservation laws and a Newtonian stress constitutive equation for the micellar solution, a non-monotonic flow curve was produced. Goveas and Pine adopted this approach to describe shear-thickening wormlike micelles and were able to successfully reproduce much of the experimental phenomenology. The flow curve was then used to explain the differences in stress versus shear rate control, based on a linear stability analysis of the interfacial height equation. In this case, the momentum and micellar solution stress were taken to be fast variables, while an equation of motion was written for a scalar variable, which is the macroscopic manifestation of changes in the fluid microstructure.
However, the formulation of Goveas and Pine did not contain any mechanism for the formation of the shear-induced state, so that the *existence* of the new phase was simply postulated by the presence of an interface. In this paper we present a generic phenomenological model that naturally admits a flow-induced phase and incorporates spatial gradients so that an interface structure and its stability can be determined. The model consists of an equation of motion for the volume fraction of a reacting species, *i.e.* a scalar non-conserved order parameter representing microstructural change in a complex fluid. There are fast stress variables associated with the reactant and product species, which contribute additively to the total stress. We have continuously evolved the model from a homogeneous to a phase-separated state, and examined how thinning and thickening flow curves, as well as the size of the gradient terms, affect phase transitions, and in particular the banding orientation (vorticity versus gradient banding). Most significantly, we are able to probe the nonlinear dynamic behavior of the system.
Minimal model
=============
Our phenomenological theory consists of a general reaction-diffusion scheme. The reaction terms represent the creation and destruction of a variable under flow, and are analogous to the local free energy terms in a Landau-Ginzburg theory. This variable may embody a species concentration, or a structural parameter such as aggregate size or molecule orientation. While this scheme is meant to be quite general and is a vehicle for capturing the general physics for many complex fluids, a reaction diffusion scheme has a literal basis for wormlike micelles and onion solutions. In the wormlike micelle case, such “reaction” terms correspond to the constant breaking and recombination of the “living” polymers; while in onion solutions, the reaction terms might correspond to the formation of onions. The steady-state onion size scales as the inverse square root of the shear rate [@Roux] and is a reversible function of the shear rate; *i.e.* the size is independent of whether smaller onions are created by increasing the shear rate applied to larger onions, or larger onions are created by decreasing the shear rate applied to smaller onions. This indicates that onion combination and fracture processes compete to attain steady state, and these processes have different dependences on shear rate.
In the same way that a double well potential signals the possibility of equilibrium phase coexistence, a multivalued reaction term can allow for flow-induced banding. The diffusion terms are the analog of the non-local terms in the free energy and provide gradients which can support inhomogeneities and describe interfaces between states. However, unlike in equilibrium, where a global minimization principle applies, the diffusion terms are necessary for determining the conditions for phase coexistence in flow [@nonlocal; @stress].
Consider a system which is one-phase at equilibrium, and consists solely of a species, $A$. Planar shear flow is then applied to this system: the coordinate system is shown in Figure \[banding figure\], where $x,y$ and $z$ denote the flow, gradient and vorticity directions respectively. We consider only variations in $y$ and $z$ in this work. Suppose that a new phase, $B$, can be induced by flow, such that at a given shear rate (or stress) a dynamic equilibrium between $A$ and $B$ is established. We write this schematically as: $$A \rightleftharpoons B.$$ Let us define an order parameter, $\phi_{B}\equiv \phi$, which corresponds to the volume fraction of the $B$-species (although allowing $\phi$ to correspond to a structural variable is also viable). The system is constrained to have constant density such that $$\label{density constraint}
\phi_{A}+\phi_{B}=1.$$ Notice that $\phi_{A}$ and $\phi_{B}$ are non-conserved variables, although the total density is conserved. We write the following equation of motion for $\phi$ as $$\label{deq}
\frac{d \phi}{dt}= R(\phi, \dot{\gamma})+D \nabla^{2} \phi,$$ where $R(\phi, \dot{\gamma})$ represents the forward and backward “reactions” which create and destroy the new phase, $\dot{\gamma}(y,z)=dv_x/dy$ is the local shear rate (where $v_{x}$ is the component of the velocity in the flow direction), and $D$ is an effective diffusion coefficient (taken to be a constant). The [*homogeneous*]{} steady-state solutions to equation (\[deq\]), where $\nabla^{2}
\phi=0$, are given by $$\label{reaction eq}
R(\phi,\dot{\gamma})=0.$$
Constitutive curves {#flow curve section}
===================
To compute flow curves for the system, stress constitutive equations for the different components must also be specified. The simplest possible scheme involves additive Newtonian relations for each species $$\label{stress}
\begin{split}
\sigma&=\sigma_{A}+\sigma_{B} \\
\sigma_{\alpha}&=\eta_{\alpha} \phi_{\alpha} \dot{\gamma},
\end{split}$$ where $\sigma$ is the total shear stress, and $\sigma_{\alpha}$ and $\eta_{\alpha}$ are the shear stress and viscosity, respectively, of species $\alpha=A,B$. Stress and composition are thus effectively coupled in the system. Applying the density constraint, equation (\[density constraint\]), gives an expression for the constitutive curve of the system, $$\label{constitutive stress}
\sigma=[\phi(c-1)+1] \dot{\gamma},$$ where $\phi=\phi(\dot{\gamma})$ is the solution to equation (\[reaction eq\]), $c=\eta_{B}/\eta_{A}$ is the ratio of the viscosities of the two components, and we have set $\eta_{A}=1$ for simplicity.
=3.3in
Notice that a multivalued $R(\phi,\dot{\gamma})$ produces multiple homogeneous steady-states. Choosing $$\label{reaction term}
R(\phi,\dot{\gamma})=|\dot{\gamma}| \phi_{A} \phi_{B}^{2}- k \phi_{B}$$ yields the curve shown in Figure \[mfig\], for an imposed $\dot{\gamma}$. \[$k$ represents a rate constant for a backward reaction, which has dimensions of inverse time and is henceforth set to unity.\] Notice that the forward reaction term in equation (\[reaction term\]) has a linear shear rate dependence, so we must take its absolute value from symmetry considerations. In microscopic theories for flow-induced reactions in wormlike micelles [@Cates; @and; @Turner] and polymers [@Glenn], such an [ *effective*]{} reaction rate also has a linear or non-analytic form, resulting from the projection of tensorial degrees of freedom onto a scalar order parameter. For a given local shear rate, the reaction scheme of equations (\[density constraint\], \[reaction eq\], \[reaction term\]) yields the following homogeneous steady-states:
\[shear steady states\] $$\begin{aligned}
\phi_{1}&=0\\
\phi_{2}& =\frac{1}{2}- \frac{1}{2}
\sqrt{1-\frac{4}{\dot{\gamma}}}\\
\phi_{3}&=\frac{1}{2}+\frac{1}{2} \sqrt{1-\frac{4}{\dot{\gamma}}} .\end{aligned}$$
Performing a linear stability analysis on equation (\[deq\]), at fixed $\dot{\gamma}$, shows that $\phi_{1}$ and $\phi_{3}$ are stable fixed points for the system, while $\phi_{2}$ is an unstable fixed point. This is evident from Figure \[mfig\], where the middle root has positive slope $dR/d\phi>0$.
![Flow curves (from equations \[reaction eq\], \[constitutive stress\] and \[reaction term\]) corresponding to the minimal model. At low shear rates, the flow curve is single-valued (branch $1$). Above a certain shear rate (or stress), an additional unstable branch, $2$, and a stable branch, $3$, exist. These branches correspond to the homogeneous roots $\{\phi_1,\phi_2, \phi_3 \}$ from equations (\[shear steady states\]) for imposed shear rate, and $\{ \phi_1,\phi_2', \phi_3' \}$ from equations (\[phi stress\]) for imposed stress. In the former case, branches $2$ and $3$ are separated by $P$, while they are separated by $Q$ in the latter. Thus, the line segment between $P$ and $Q$ is stable under controlled shear rate, but unstable under controlled stress. (a) Shear-thickening flow curve for $c>1$, illustrating the controlled shear rate case. At a fixed shear rate, $\Gamma$, the system can choose between homogeneous states on branches $1$ and $3$, or gradient band between these branches at stress $\sigma^{*}$. (b) Shear-thinning flow curve for $c<1$, illustrating the constrolled stress case. At fixed stress $\Sigma$, the system can vorticity band at shear rate $\dot{\gamma}^{*}$ between $1$ and $3$, or choose between homogeneous states on these same branches.[]{data-label="minimal flow curves"}](./cartoong2.eps)
Substituting the homogeneous roots from equation (\[shear steady states\]) into equation (\[constitutive stress\]) produces three branches of the constitutive curve (labelled as $1, 2$ and $3$ respectively in Figure \[minimal flow curves\]). Notice that below a certain shear rate, $\dot{\gamma}=4$, only the $\phi_{1}$ root is real and the reaction curve is single-valued. This is marked as point $P$ on Figure \[minimal flow curves\], and has coordinates $$\left\{\dot{\gamma}_{P},\sigma_P\right\} = \left\{4,
2(c+1) \right\}.$$ Physically, this means that only species $A$ exists at low (uniform) shear rates. From equation (\[constitutive stress\]), we can see that this one-phase system ($\phi_{A}=1$) is Newtonian, and the corresponding flow curve has a slope of unity. The slope of the stable flow-induced branch $3$ depends on the value of the parameter $c$. For $c<1$, a transition from branch $1$ to branch $3$ is shear-thinning, while for $c>1$ such a transition is shear-thickening. The locus of flow induced roots $(\phi_2,\phi_3)$ exhibits a minimum in the shear stress as a function of shear rate, which is denoted as point $Q$ in Figure \[minimal flow curves\]: $$\label{stressmin}
\left\{\dot{\gamma}_{Q},\sigma_Q\right\}=
\left\{\frac{(1+\sqrt{c})^2}{\sqrt{c}}, (1+\sqrt{c})^2 \right\}.$$
For a given value of the local stress, the homogeneous steady-states are given by
\[phi stress\] $$\begin{aligned}
\phi_{1}& =0 \\
\phi_{2'}& =\frac{ (\sigma-c+1)-
\sqrt{(\sigma-c+1)^{2} - 4 \sigma}} {2 \sigma}\\
\phi_{3'}& =\frac{ (\sigma-c+1)+ \sqrt{(\sigma-c+1)^{2} - 4 \sigma}}
{2 \sigma}.\end{aligned}$$
Note that these roots are found by recasting equation (\[reaction term\]) in terms of the stress, by using equation (\[constitutive stress\]). This procedure is not equivalent to substituting equation (\[constitutive stress\]) into equation (\[shear steady states\]). This is because while the loci of homogeneous states is the same under fixed local stress or shear rate, the [*stability*]{} of these steady-states is not; *i.e.* the portion of the constitutive curve between $P$ and $Q$ is unstable under fixed local stress, but stable under fixed local shear rate. In Figure \[minimal flow curves\], point $Q$ marks the stress above which the constitutive curve is multivalued for controlled stress, while $P$ marks the strain rate above which the constitutive curve is multivalued for controlled shear rate.
In an experiment, however, only the [*average*]{} stress and shear rate can be controlled. If the average shear rate is held fixed at $\Gamma$, for example, the system can choose between various options (illustrated in Figure \[minimal flow curves\]a):
- A homogeneous low stress state, $\phi_{1}$.
- A homogeneous high stress state, $\phi_{3'}$.
- A mixture of states (i) and (ii), where the interfaces between phases lies in the vorticity direction (vorticity banding). Note that (i) and (ii) cannot coexist with each other in the $y$-direction, since the stress must be homogeneous in the gradient direction.
- A mixture of high shear rate phase, $\phi_{3'}$ and a low shear rate phase, $\phi_{1}$. Here the system attains an intermediate stress, $\sigma^{*}$, and the relative proportions of the two phases are set such that the average shear rate is maintained at $\Gamma$. Since the bands have the same stress, but different shear rates, this scenario corresponds to gradient banding.
While it appears from Figure \[minimal flow curves\] that there is a multiplicity of stresses $\sigma^*$ at which the system can gradient band, in fact the system selects a particular stress (see next Section).
If instead, the average stress is fixed at $\Sigma$ (illustrated in Figure \[minimal flow curves\]b), the system can choose a high or low shear rate homogeneous phase, or it can gradient band between these. Alternatively, it can band in the vorticity direction between high and low stress states, by adopting a shear rate $\gamma^{*}$. There is also a selected shear rate for vorticity banding. [*The essential question is: which of the many possible states available to it does the system actually choose and why?*]{}
Calculating the banding stress and shear rate
=============================================
The inclusion of gradient terms in equation (\[deq\]) causes stress selection for gradient banding, and shear rate selection for vorticity banding [@nonlocal; @stress]. The selected stress and shear rate are determined by mathematically connecting two different homogeneous stable states to form an inhomogeneous profile. To find the banding shear rate $\dot{\gamma}^{*}$ at which vorticity banding can occur, equation (\[deq\]) is integrated across the domain at steady-state. A banding solution (homogeneous phases separated by interfaces) is, by definition, one which has no gradients in $\phi$ at the boundaries. We obtain the following condition: $$\label{banding eq}
\int_{\phi_{1}(\dot{\gamma}^{*})}^{\phi_{3}(\dot{\gamma}^{*})} d \phi
\, R[\phi, \dot{\gamma}^{*}]=0.$$ where $\phi_{1}$ and $\phi_{3}$ are given by equations (\[shear steady states\]). Defining a new function, $F[\phi,\dot{\gamma}]=\int_{0}^{\phi} d \,
\phi' R[\phi', \dot{\gamma}]$, we can rewrite equation (\[banding eq\]) as $$\label{bandq}
F[\phi_{1}(\dot{\gamma}^{*}),\dot{\gamma}^*]=F[\phi_{3}
(\dot{\gamma}^{*}),\dot{\gamma}^*],$$ which is analogous to the common tangent construction from equilibrium thermodynamics. If this were an equilibrium system, $F$ would be identified as the free energy. Since this is a dynamic system however, $F$ cannot be given the same physical interpretation, so that the analogy is purely formal. Using equation (\[shear steady states\]) in equation (\[bandq\]) gives $$\dot{\gamma}^{*}=4.5.$$
To calculate the banding stress $\sigma^{*}$ at which gradient banding can occur, equation (\[banding eq\]) must be recast in terms of the shear stress using the stress constitutive relation, equation (\[constitutive stress\]) to obtain a relation $\dot{\gamma}(\sigma,\phi)$. The banding stress in our minimal model is only a function of $c$ and is given by the solution of the following equation: $$F[\phi_{1'}(\sigma^{*}),\sigma^*]=F[\phi_{3'}(\sigma^{*}),\sigma^*],$$ where $F[\phi,\sigma^*]=\int_{0}^{\phi} d \, \phi'
R[\phi', \dot{\gamma}(\sigma,\phi')]$ yields $$\begin{aligned}
F[\phi(\sigma^{*}), \sigma^{*}]&=\frac{\sigma^{*}}{(c-1)^{4}} \left
[ -\tfrac13\tilde{\phi}^{3} +
\tfrac12(c+2)\tilde{\phi}^{2} + \tfrac32 c\right.\label{eq:Fbig}\\
&\left.\phantom{\tfrac13} -(2c+1) \tilde{\phi} +
\tfrac13 + c \ln
\tilde{\phi} \right ] - \tfrac12\phi^{2}(\sigma^*), \nonumber\\
\tilde{\phi}&=\phi(\sigma^{*})(c-1)+1,\end{aligned}$$ and $\phi_{1'}$ and $\phi_{3'}$ are given by equations (\[phi stress\]). The selected stress $\sigma^*$ and shear rate $\dot{\gamma}^*$ are shown in Figure \[minimal flow curves\] for $c=0.3, 2.7$, and given in Table \[tab:bandingvalues\].
$c$ $\sigma^{\ast}$ $\phi_3(\sigma^{\ast})$ $\dot{\gamma}(\phi_3^{\ast})$ $\sigma_P$ $\dot{\gamma}_Q$ $\sigma_Q$
------- ----------------- ------------------------- ------------------------------- ------------ ------------------ ------------
$0.3$ $2.815$ $0.810$ $6.504$ $2.6$ $4.374$ $2.395$
$0.6$ $3.606$ $0.732$ $5.100$ $3.2$ $4.066$ $3.149$
$1.2$ $4.910$ $0.642$ $4.352$ $2.4$ $4.008$ $4.391$
$2.7$ $7.625$ $0.529$ $4.014$ $5.4$ $4.252$ $6.986$
: Banding stress $\sigma^{\ast}$, points of instability $P$ and $Q$, and coexistence conditions for different values of $c$. In all cases $\dot{\gamma}_P=4$ and $\dot{\gamma}^*=4.5$., while the stress and shear rate on branch $\phi_1$ are related by $\sigma_1=\dot{\gamma}_1$.[]{data-label="tab:bandingvalues"}
Dynamical Selection of Steady-States
====================================
In the preceding Sections we have seen that certain homogeneous and banded states are available to the system, based on a [ *steady-state*]{} analysis. To determine which of these states is selected in practice, equation (\[deq\]) must be evolved in time; to make contact with experiments we can only impose constraints of fixed average shear rate or stress. In this work only one-dimensional calculations are performed, so that the equation of motion is solved either in the $y$ (gradient) or $z$ (vorticity) directions. If there are composition modulations in the $y$-direction, these can only cause modulations in the shear rate (gradient banding), since the shear stress must be uniform in $y$. If the average stress is controlled, there is only one “interesting” stress $\sigma^{*}$ at which the system can become inhomogeneous in the $y$ direction. Due to the numerical difficulty of fixing a precise stress in the system, we do not consider this case. By comparison, if the average shear rate is controlled, there is a wide range of shear rates for which we can investigate whether the system remains homogeneous or gradient bands. Similarly, when composition modulations in $z$ are allowed, we can only look for vorticity banding under controlled stress within this calculation. Gradient banding can occur only if the shear rate is set exactly at $\dot{\gamma}^{*}$, which we do not study here.
Controlled Average Shear Rate {#shear rate}
-----------------------------
We first consider the system under shear rate control, and only allow for spatial variations in $y$. Integrating the stress relation, equation (\[constitutive stress\]), across the domain (where the shear stress is independent of $y$) gives the local shear rate as a function of $V$, the velocity difference across the system. Then equation (\[deq\]) becomes $$\label{average shear rate}
\frac{d \phi}{dt}= \frac{V}{\int dy
\left\{1/\left[\phi\left(c-1\right) + 1\right]\right\}}
\frac{\phi^{2}(1-\phi)}{\phi(c-1)+1}
-\phi +D \frac{d^{2} \phi}{dy^{2}}.$$ This is an integro-differential equation, instead of the differential equation which yielded the analysis of Section \[flow curve section\]. Notice, however, that the same homogeneous steady-states are obtained. Equation (\[average shear rate\]) is solved using random initial conditions with $\phi$ uniformly chosen within the range $[0-1]$, and no flux boundary conditions, keeping $V$ at a fixed value. The domain size is normalized to unity, so that $V$ is synonymous with the average shear rate.
To solve equations (\[average shear rate\]), we use a fully implicit finite difference scheme, using a central difference approximation for first and second spatial derivatives, and a forward difference approximation for the time derivative. Nonlinear terms are linearized in time as follows: $$W[\phi(x,t+\Delta t)]=W[\phi(t)]+ [\phi(t+\Delta t)-\phi(t)]\frac{d
\,W[\phi,t]}{d \phi}.$$ The integral in equation (\[average shear rate\]) is evaluated explictly, *i.e.* at the previous time step. In general, $300$ spatial mesh points are used with a time-step of $1/10000$ [@mesh; @points].
For some initial conditions, the resulting steady-states are homogeneous, while for others, they are banded. This implies that there is a [*basin of attraction*]{} for attaining a banded state. Our results can be categorized according to the shape of the constitutive curve and the magnitude of the diffusion coefficient, and whether the system is shear-thinning or thickening. We find that decreasing the diffusion coefficient increases the basin of attraction of the banded state and destabilizes the homogeneous state. That is, the system is more likely to band for narrow interfaces. Intuitively, this makes sense since the banded state represents a mathematical connection between two homogeneous states: the wider the interface, the more difficult it is for gradients to be non-zero near the boundaries of the system.
For $c<1$, for $V<\dot{\gamma}_{P}$ the system usually chooses branch 1 of the constitutive curve over the banded state, even for small diffusion coefficients. Notice that banding is first allowed, in principle, when the imposed shear rate is larger than that of the low shear rate band. The “critical” shear rate, where the system actually first starts to band, is generally somewhat higher than this: it is the low shear rate limit of the stress plateau that would be measured experimentally. Below the critical shear rate the system [*always*]{} chooses branch 1, although the exact location of this point depends on the diffusion coefficient. \[Increasing the diffusion coefficient widens the interface, affecting where the interface first “touches" the wall [@Olmsted; @JS] and the ability of a banded system to satisfy the boundary conditions, as discussed above.\]
When $\dot{\gamma}_{P}< V < \dot{\gamma}_{Q}$, the system is more inclined to band rather than remain homogeneous, even for large diffusion coefficients (see Fig. \[fig:basinregions\]). The chosen homogeneous states always lie on branch 3 of the constitutive curve. Beyond $\dot{\gamma}_{Q}$, there is a crossover to where homogeneous states (on branch 3) are preferred over banded states. The exact crossover depends on the value of the diffusion coefficient. This shows that the system has made a transition under flow, so that for $c<1$ a shear-thinning transition is seen, as discussed in Section \[flow curve section\]. These findings are illustrated in Table 2, for 49 runs with different initial conditions with $c=0.3$. The “critical” shear rate in this case is around $V=3.3$. Figures \[zbandingb\] and \[zbandinga\] show similar behavior for $c=0.6$.
Gradient banding is never observed in our numerical experiments when $c>1$. Here, the chosen final steady state, above the critical shear rate, is always the stable high stress homogeneous state on branch 3, which makes this a shear-thickening transition.
------ ------- ---------- -------- ----------
Result
V D $\phi_1$ banded $\phi_3$
3.3 0.01 49 0 0
0.005 49 0 0
0.001 48 1 0
3.6 0.01 45 5 0
0.005 34 15 0
0.001 6 43 0
3.8 0.01 15 34 0
0.005 8 41 0
0.001 0 49 0
3.9 0.01 0 49 0
0.005 0 49 0
0.001 0 49 0
4.0 0.01 0 49 0
0.005 0 49 0
0.001 0 49 0
4.05 0.01 0 49 0
0.005 0 49 0
0.001 0 49 0
4.1 0.01 0 49 0
0.005 0 49 0
0.001 0 49 0
4.25 0.01 0 39 10
0.005 0 47 2
0.001 0 49 0
4.37 0.01 0 10 39
0.005 0 33 16
0.001 0 49 0
4.4 0.01 0 10 39
0.005 0 21 28
0.001 0 49 0
4.55 0.01 0 2 47
0.005 0 8 41
0.001 0 13 36
4.6 0.01 0 1 48
0.005 0 7 42
0.001 0 20 29
4.7 0.01 0 0 49
0.005 0 4 45
0.001 0 19 30
------ ------- ---------- -------- ----------
: Summary of results for $c=0.3$, for $49$ runs with random initial conditions and controlled mean shear rate $V$. In this case, $\dot{\gamma}_{P}$ corresponds to $V=4.0$, and $\dot{\gamma}_{Q}$ corresponds to $V=4.374$. Banding is first allowed at $V=2.815$.
\[tab:basins\]
The preceding results are for the case where the shear rate is held at a steady value, and might apply to a system where the mean shear rate is applied to an initially noisy system. However, most experiments are conducted by starting up the system from zero shear rate, and then discontinuously ramping the shear rate to higher values. We have tried to mimic this scenario, by bringing the system to steady-state for $V<\dot{\gamma}_{P}$, and then suddenly increasing $V$ to greater than $\dot{\gamma}_{P}$. In order to dislodge the system from branch 1 to the banded state, however, we need to add noise of amplitude order unity. Such a large amount of noise essentially obliterates any memory of the initial steady-state, suggesting that a nucleation event is required for an experimental system to band from start-up, as was found previously in Ref. [@Olmsted; @JS].
1.0truecm =3.0truein 1.0truecm =3.0truein
1.0truecm =3.0truein 1.0truecm =3.0truein
In general, the banded state consists of two bands, corresponding to branches 1 and 3 on the constitutive curve. For some initial conditions, multiple bands are found. We do not attach much significance to this result, because it is known that multiple interfaces are allowed for planar flow, and the number of allowed interfaces in such one-dimensional systems is known to increase as the diffusion coefficient decreases [@Grindrod; @book]. Britton and Callaghan have reported multiple gradient bands for wormlike micelles in Couette flow [@Callaghan]. However, it has been shown [@Olmsted; @JS] that simple constitutive relations (like the one derived here) do not permit multiple interfaces in Couette flow. This implies that the current model cannot describe these experimental observations.
Controlled Average Stress
-------------------------
Next we fix the average shear stress and solve the model in the $z$-direction, hence allowing for vorticity banding at different stresses. Following a procedure analogous to that of the previous Section, the observation that the shear rate is uniform in $z$ allows us to convert equation (\[deq\]) into the following integro-differential equation $$\label{average stress}
\frac{d \phi}{dt}= \frac{\Sigma \, \phi^{2}(1-\phi)}{\int dz
[\phi(c-1)+1]} -\phi +D \frac{d^{2} \phi}{dz^{2}},$$ where $\Sigma$ is the imposed average shear stress across the domain.
Note that interchanging the stress and shear rate variables in the phenomenological model would “reverse” all the results of Section \[shear rate\]. Thus, branch 3 would be thickening for $c<1$, and the system would be inclined to vorticity band under stress control, for $\sigma_{P}<\Sigma<\sigma_{Q}$. Similarly, the system would remain homogeneous for $c<1$. This reasoning implies that the shape of the flow curve sets the attractors for the system. If this is true, then when stress and shear rate are not interchanged, the system should be predisposed to vorticity band for $c>1$ when the stress is fixed between $\sigma_{P}$ and $\sigma_{Q}$, but should remain homogeneous for all stresses when $c<1$.
This is indeed what we find from numerical solution of equation (\[average stress\]). Table \[tab:basinsig\] shows results for various values of $\Sigma$ and $c=2.7$. Decreasing the diffusion coefficient destabilizes the homogeneous state, as in the controlled shear rate case.
---------- ------- ---------- -------- ----------
Result
$\Sigma$ D $\phi_1$ banded $\phi_3$
5.0 0.01 49 0 0
0.005 49 0 0
0.001 49 0 0
5.4 0.01 49 0 0
0.005 49 0 0
0.001 49 0 0
6.0 0.01 49 0 0
0.005 48 1 0
0.001 36 13 0
7.1 0.01 28 21 0
0.005 0 49 0
0.001 0 49 0
7.2 0.01 0 4 45
0.005 0 40 9
0.001 0 49 0
7.3 0.01 0 0 49
0.005 0 9 40
0.001 0 43 6
7.4 0.01 0 0 49
0.005 0 2 47
0.001 0 23 26
7.5 0.01 0 0 49
0.005 0 0 49
0.001 0 13 36
8.5 0.01 0 0 49
0.005 0 0 49
0.001 0 0 49
---------- ------- ---------- -------- ----------
: Summary of results for $c=2.7$, for $49$ runs with random initial conditions and controlled stress $\Sigma$. In this case, $\sigma_{P}=
5.4$ and $\sigma_{Q}=6.896$.
\[tab:basinsig\]
We sometimes observe multiple bands under stress control, as with the fixed shear rate cases. Bonn *et al.* [@Bonn] and Chen *et al.* have seen multiple bands in the vorticity direction in Couette flow which is probably due to a combination of the slow coarsening expected in one-dimensional systems [@OneD] and multiple allowed interfaces [@Grindrod; @book]. The stress is non-uniform and monotonic in the flow gradient direction of a cylindrical Couette device, which implies a single stable interface. The cylindrical geometry does not, however, impose such an inhomogeneity along the vorticity direction.
Conclusions
===========
We have shown that a simple phenomenological reaction-diffusion scheme can produce a flow-induced phase transition, as a consequence of a multi-valued reaction term. The model consists of an equation of motion for a non-conserved composition variable, while the stresses induced in the reactants and products are assumed to be fast variables. The character of the model depends on a single parameter $c$, that controls whether or not the transition is shear-thinning or shear-thickening. Above a critical shear rate (or shear stress), the system may band or remain homogeneous. The steady-states that are selected from random initial conditions depend on the shape of the constitutive curves and the magnitude of the diffusion coefficient:
1. *Imposed shear rates:* For $c<1$ (shear-thinning transition), the system chooses a low stress homogeneous state at low shear rates. Above a critical shear rate, gradient banding tends to occur for imposed shear rates around the region of the constitutive curve with negative slope $d\sigma/d\dot{\gamma}<0$ (see Figure \[minimal flow curves\]a). At shear rates higher than this, the system is predisposed towards the high stress homogeneous state. For $c>1$ (shear-thickening transition), the system always chooses this homogeneous state above the critical shear rate and gradient banding is never observed.
2. *Imposed stress:* For $c>1$ (shear-thickening transition), the system chooses a low shear rate homogeneous at low stresses. Above a critical stress, vorticity banding tends to occur for imposed stresses around the region of the constitutive curve with negative slope. For higher stresses, the system is predisposed towards the high shear rate homogeneous state. For $c<1$, the system always chooses this homogeneous state above the critical stress and vorticity banding is never observed.
3. In the regions of the flow curve where banding is observed, we find the apparent basin of attraction for banding increases upon decreasing the value of the diffusion coefficient.
While banding is more pronounced in the vicinity of the flow curve with a negative slope (where the system is linearly unstable), it is also observed in regions of the flow curve with positive slope. In particular, the critical shear rate or stress (where banding is first initiated), lies in the latter section of the flow curve. Here, the system is nonlinearly unstable to perturbations. Such behavior has been seen in experiments on shear-thinning wormlike micelles [@grand], where the onset of banding occurs at a lower stress (and shear rate) if the system is given enough time to explore all fluctuations, as compared to where banding is induced upon rapidly varying the control parameters. Porte *et al.* [@PorteBerretHarden] have discussed various flow curves which can contain both linearly and nonlinearly unstable regions: the equilibrium analog of the former is the spinodal curve, and that of the latter is the metastable region, where an instability must be nucleated.
Our results are significant because they show that a minimal model can exhibit a rich phenomenology, and that the selection rules for phase coexistence are simple. To understand why the system chooses certain states over others in some regions of the flow curve, a nonlinear dynamics analysis of the model must be performed. We believe that our scheme represents a new class of reaction-diffusion equations, because the constraint of fixed average stress or shear rate turns the governing partial differential equations into integro-differential equations, which represents a general class of dynamical equations that, to our knowledge, has not been studied. This system exhibits fascinating and complex nonlinear dynamics, which we will discuss in a future publication.
Our current scheme is missing much physics: a complete model would involve coupled equations of motion for conserved variables (concentration of the various species) and non-conserved, tensorial variables (structural variables, stress). Also, we have assumed that the individual species obey Newtonian stress constitutive relations. Typically, these species are themselves complex fluids, and are either shear-thinning or exhibit a yield stress. In future work, by systematic exclusion of certain dynamic variables, we will be able to investigate the individual roles played by the stress, concentration etc., in order to determine which variables are essential to the problem formulation.
One of us [@Olmsted; @LC] has already considered a theory with stress, momentum and concentration variables in the context of rigid rod suspensions. Separate phase diagrams for shear-induced phase separation in both the vorticity and gradient directions were calculated, but the model was too prohibitively complicated to study which of these orientations would in fact be selected by the system. In this work, we have used a much simpler scheme to demonstrate the neccessary analysis (albeit within a one-dimensional model- see next paragraph) to unambiguously determine whether banding actually occurs in a system, as well as the banding orientation. While Schmitt *et al.* [@SchmittMarquesLequeux] have also presented quite a simple phenomenological model (including both concentration and momentum as dynamical variables), they did not go beyond a linear stability analysis. They also did not consider the case (as we have here) of a non-conserved variable initiating an instability in the system.
Our calculations have been carried out only for the case of planar flow. It has been shown for the Johnson-Segalman model [@Olmsted; @JS] that the nonuniformity of stress in a curved geometry has significant effects on banding. In addition, we have examined the issue of gradient versus vorticity banding using a one-dimensional model. Realistically, the model should be solved considering both vorticity and gradient directions simultaneously. The band orientation may be influenced by anisotropy in the diffusion coefficient. A convective term of the form $v \cdot{\nabla \phi}$ should also be included in the equation of motion. Such a term does not appear in a one-dimensional shear flow, but it can qualitatively affect transients in phase separation in two dimensions. Finally, noise has been incorporated into our model through the initial conditions. While Gaussian noise is present in the equation of motion through the diffusion term, in a driven system there may be other noise terms that should be added.
We stress that our phenomenological theory only aims to describe the general macroscopic physics of flow-induced phase transitions. Details concerning the underlying structural transformations can only be probed by more specific microscopic models.
[99]{} Boltenhagen, P., Hu, Y. T., Matthys, E. F. and Pine, D. J., [Europhys. Lett.]{}, [**38**]{}, 389 (1997).
Bonn, D., Meunier, J., Greffier, O., Alkahwaji, A. and Kellay, H., [Phys. Rev. E.]{}, [**58**]{}, 2115 (1998).
Eiser, E., Molino, F. and Porte, G., [Rheol. Acta]{}, [**39**]{}, 201 (2000); [Phys. Rev. E]{}, [**61**]{}, 6759 (2000).
Noirez, L. and Lapp, A., [Phys. Rev. Lett.]{}, [ **78**]{}, 70 (1997).
Chen, L.B., Zukoski, C. F., Ackerson, B. J., Hanley, H. J. M., Straty, G. C., Barker, J. and Glinka, C. J., [Phys. Rev. Lett.]{}, [**64**]{}, 688 (1992).
Olmsted, P. D., [Europhys. Lett.]{}, [**48**]{}, 339 (1999).
Britton, M. M. and Callaghan, P. T., [J. Rheol.]{}, [**41**]{}, 1365 (1997)
Ramos, L., Molino, F. and Porte, G., [Langmuir]{}, [ **16**]{}, 5846 (2000).
Goveas, J. L. and Pine, D. J., [Europhys. Lett.]{}, [**48**]{}, 706 (1999).
Schmitt, Marques and Lequeux, [Phys. Rev. E]{}, [**52**]{}, 4009 (1995).
Spenley, N. A., Yuan, X. F. and Cates, M. E., [J. Phys. II (France)]{}, [**6**]{}, 551 (1996).
Yuan, X. F., [Europhys. Lett.]{}, [**46**]{}, 542 (1999).
Olmsted, P. D., Radulescu, O. and Lu, C.-Y. D., [J. Rheol.]{}, [**44**]{}, 257 (2000).
Cates, M. E. i) [Macromolecules]{}, [**20**]{}, 2289 (1987); ii) [J. Phys. Chem.]{}, [**94**]{}, 371 (1990).
Ajdari, A., [Phys. Rev. E]{}, [**58**]{}, 6294 (1998).
Diat, O. and Roux, D., [J. Phys. II (France)]{}, [**3**]{}, 9 (1992).
Cates, M. E. and Turner, M. S., [Europhys. Lett.]{}, [**11**]{}, 681 (1990).
Fredrickson G. H., Leibler L, Macromolecules, [**29**]{}, 2674 (1996).
Lu, C.-Y. D, Olmsted, P. D. and Ball, R. C., [Phys. Rev. Lett.]{}, [**84**]{}, 642 (2000).
In the numerical solution of the integro- differential equation, we observe a small (on the order of a few tenths of a percent) variation with mesh-size of the steady-state selected stress of the banded states.
Grindrod, P., *The Theory and Applications of Reaction-Diffusion Equations: Patterns and Waves (Second Edition)* (Clarendon Press, Oxford 1996).
Alikakos,N., Bates, P. W., and Fusco, G., [J. Diff. Eqs.]{}, [**90**]{}, 81 (1991).
Grand, C., Arrault, J., and Cates, M. E., [J. Phys. II (France)]{}, [**7**]{}, 1071 (1997).
Porte, G., Berret, J. F. and Harden, J., [J. Phys. II (France)]{}, [**7**]{}, 459 (1997).
Olmsted, P. D. and Goldbart, P. M., [Phys. Rev. A]{}, [**41**]{} (1990) 4578; [**46**]{} (1992) 4966; Olmsted, P. D. and Lu, C.-Y. D., [Phys. Rev. E.]{}, [**56**]{}, 55 (1997); [**60**]{} (1999) 4397.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Returning to a classical question in Harmonic Analysis we strengthen an old result of Walter Rudin. We show that there exists a weakly almost periodic function on the group of integers ${\mathbb{Z}}$ which is not in the norm-closure of the algebra $B({\mathbb{Z}})$ of Fourier-Stieltjes transforms of measures on the dual group $\hat{{\mathbb{Z}}}={\mathbb{T}}$, and which is [**recurrent**]{}. We also show that there is a Polish monothetic group which is reflexively but not Hilbert representable.'
address:
- |
Department of Mathematics\
Tel Aviv University\
Tel Aviv\
Israel
- |
Institute of Mathematics\
Hebrew University of Jerusalem\
Jerusalem\
Israel
author:
- Eli Glasner and Benjamin Weiss
date: 'February 10, 2010'
title: On Hilbert dynamical systems
---
[^1]
Introduction {#introduction .unnumbered}
============
Walter Rudin [@Ru] was the first to show that the algebra $WAP({\mathbb{Z}})$, of weakly almost periodic functions on ${\mathbb{Z}}$, is strictly larger than the algebra $H({\mathbb{Z}})= {\overline}{B({\mathbb{Z}})}$, the norm closure in the Banach space $\ell_\infty({\mathbb{Z}})$ of the algebra $B({\mathbb{Z}})$ of Fourier-Stieltjes transforms of complex measures on the dual group $\hat{{\mathbb{Z}}} = {\mathbb{T}}$. Many other examples of functions in $WAP({\mathbb{Z}}) \setminus H({\mathbb{Z}})$ followed (see e.g. [@DR]).
As it turns out, in all of these examples the function in question is non-recurrent. We say that a function $f \in \ell_\infty({\mathbb{Z}})$ is [*recurrent*]{} if there is a sequence $n_k \nearrow \infty$ with $\lim_{k \to \infty}\sup_{j \in {\mathbb{Z}}} |f(j) - f(j +n_k)| =0$. Now in some sense the more interesting functions in $WAP({\mathbb{Z}})$ are the recurrent ones, and moreover they are prevalent; e.g. every ${\mathbb{Z}}$-subalgebra of $WAP({\mathbb{Z}})$ which is ${\mathbb{Z}}$-generated by a recurrent function consists entirely of recurrent functions (see Lemma \[Rec-WAP\] below). Also, a Fourier-Stieltjes transform $\hat{\mu}$ of a continuous measure $\mu$ on ${\mathbb{T}}$ is recurrent iff $\mu$ is Dirichlet [@HMP]. Thus the question whether there are [*recurrent*]{} functions in $WAP({\mathbb{Z}}) \setminus H({\mathbb{Z}})$ is a natural one.
Another related open question is the following. Is there a Polish monothetic group $P$ which can be represented as a group of isometries of a reflexive Banach space but is not representable as a group of unitary operators on a Hilbert space?
In this work we see how these questions are related and show that the answer to both is affirmative (Theorem \[WAPH-1\] and Corollary \[rf\]). We also show that if a Polish monothetic group $P$ is Hilbert-representable and $K \subset P$ is a compact subgroup, then the quotient group $P/K$ is also Hilbert-representable (Corollary \[AK-ext-cor\]).
Our proofs are based on the theory of topological dynamics and rely on a well known construction of Banaszczyk [@Ba]. We also use an idea of Megrelishvili [@Me-00] who showed that the topological groups $L_{2n}([0,1])$, for $n$ a positive integer, are reflexively but not Hilbert representable. For the sake of simplicity we state our results in the most basic setup, where the acting group is ${\mathbb{Z}}$ and the dynamical systems are usually assumed to be point-transitive.
We refer the reader to the following related recent works: Gao and Pestov [@Gao-P], Megrelishvili [@Me], Ferri and Galindo [@Ga-F], and Galindo [@Ga].
Some preliminaries from topological dynamics
============================================
A [*dynamical system*]{} (or sometimes just a [*system*]{}) is for us a pair $(X,T)$ where $X$ is a compact Hausdorff space and $T : X \to X$ is a self homeomorphism. With $(X,T)$ we associate an action of the group of integers ${\mathbb{Z}}$ via the map $n \mapsto T^n$. The [*orbit*]{} of a point $x \in X$ is the set ${\mathcal{O}}_T(x) =\{T^n x : n \in {\mathbb{Z}}\}$. The [*orbit closure*]{} of $x$ is the set ${\bar{\mathcal{O}}}_T(x) = {\overline}{{\mathcal{O}}_T(x)}$. The system $(X,T)$ is [*point-transitive*]{} if there is a point $x \in X$ with ${\bar{\mathcal{O}}}_T(x)= X$. Such a point is called a [*transitive point*]{} and the collection of transitive points (when it is not empty) is denoted by $X_{tr}$. For a point-transitive metric system $X_{tr}$ is a dense $G_{\delta}$ subset of $X$. We will mostly work in the category of [*pointed*]{} dynamical systems $(X,x_0,T)$, where the latter is a point-transitive dynamical system with a distinguished point $x_0 \in X_{tr}$. The restriction of $T$ to a closed invariant subset $Y \subset X$ in a dynamical system $(X,T)$ defines a dynamical system $(Y,T)$. Such a system is called a [*subsystem*]{} of $(X,T)$.
A continuous surjective map $\pi: (X,T) \to (Y,S)$ between two dynamical systems $(X,T)$ and $(Y,S)$ which intertwines the ${\mathbb{Z}}$-actions (i.e. $\pi(Tx) = S\pi(x)$ for every $x \in X$) is called a [*homomorphism*]{} of dynamical systems and we sometimes say that $Y$ is a [*factor*]{} of $X$ or that $X$ is an [*extension*]{} of $Y$. When dealing with pointed systems a homomorphism $\pi: (X,x_0,T) \to (Y,y_0,S)$ is further assumed to satisfy $\pi(x_0) = y_0$. An extension $\pi : (X,T) \to (Y,S)$ is called [*almost 1-1*]{} if there is a dense $G_{\delta}$ subset $X_0 \subset X$ with $\pi^{-1}(\pi(x)) =\{x\}$ for every $x \in X_0$. The extension $\pi: (X,x_0,T) \to (Y,y_0,S)$ is called a [*group-extension*]{} if there is a compact subgroup $K \subset {{\rm{Homeo\,}}}(X)$ such that each $k \in K$ is an automorphism of $(X,T)$ (i.e. $Tk = kT$ for every $k \in K$) and such that the quotient dynamical system $(X/K, T)$ is isomorphic to $(Y,S)$.
A point $x$ in a metric dynamical system $(X,T)$ is called [*recurrent*]{} if there is a sequence $\{n_k\} \subset {\mathbb{Z}}$ with $|n_k| \to \infty$ such that $\lim T^{n_k} x = x$. Note that in a point-transitive system if the set of isolated points is not empty then it coincides with the orbit of a transitive point. On the other hand, when there are no isolated points in a point transitive system then every point of $X_{tr}$ is recurrent. We call a point-transitive system with no isolated points a [ *recurrent-transitive system*]{}.
Let $\ell_\infty({\mathbb{Z}})$ be the Banach space (and $C^*$-algebra) of bounded complex valued functions on ${\mathbb{Z}}$ with the sup norm: $\|f\|_\infty= \sup_{n \in {\mathbb{Z}}} |f(n)|$. We write $S: \ell_\infty({\mathbb{Z}}) \to \ell_\infty({\mathbb{Z}})$ for the shift operator, where $Sf(n) = f(n+1),\ (n \in {\mathbb{Z}})$. An $S$-invariant, conjugacy invariant subalgebra of $\ell_\infty({\mathbb{Z}})$ containing the constant function ${\mathbf{1}}$ will be called a [*${\mathbb{Z}}$-algebra*]{}.
Given a pointed dynamical system $(X,x_0,T)$ we define a map $j_{x_0}: C(X) \to \ell_\infty({\mathbb{Z}})$ by $$j_{x_0}F(n) = F(T^n x_0), \quad F \in C(X),\ n \in {\mathbb{Z}}.$$ It is easy to see that $j_{x_0}$ is an isometry with $j_{x_0}\circ T
=S \circ j_{x_0}$. We denote its image in $\ell_\infty({\mathbb{Z}})$ by ${\mathcal{A}}(X,x_0)$. Clearly ${\mathcal{A}}(X,x_0)$ is a ${\mathbb{Z}}$-algebra.
Conversely, given a ${\mathbb{Z}}$-algebra ${\mathcal{A}}\subset \ell_\infty({\mathbb{Z}})$ we denote its Gelfand space (comprising the non-zero $C^*$-homomorphisms of ${\mathcal{A}}(X,x_0)$ into ${\mathbb{C}}$) by $X=|{\mathcal{A}}(X,x_0)|$. It is easy to see that the operator $S$ induces a homeomorphism $T : X \to X$ and that the resulting dynamical system $(X,T)$ is a point-transitive system. In fact, the point $x_0={{\rm{eva}}}_0 \in X$, which corresponds to the multiplicative complex valued homomorphism defined on ${\mathcal{A}}$ by evaluation at $0$, is a transitive point.
These operations are inverse to each other and we have $$(|{\mathcal{A}}(X,x_0)|,{{\rm{eva}}}_0,S) \cong (X,x_0,T).$$ A ${\mathbb{Z}}$-algebra ${\mathcal{A}}$ is [*cyclic*]{} if there is a function $f \in {\mathcal{A}}$ such that ${\mathcal{A}}= {\mathcal{A}}_f$, where the latter is the smallest ${\mathbb{Z}}$-algebra that contains $f$.
Given $f \in \ell_\infty({\mathbb{Z}})$ we can consider $f$ as an element of the compact metrizable space ${\Omega}=[-\|f\|_\infty,\|f\|_\infty]^{\mathbb{Z}}$. Again denote by $S : {\Omega}\to {\Omega}$ the homeomorphism defined by $Sg(n) = g(n+1),\ (g \in \ell_\infty,\ n \in {\mathbb{Z}})$. We let $X_f = {{\rm{cls\,}}}\{S^n f : n \in {\mathbb{Z}}\}$, where the closure is taken in ${\Omega}$ with respect to the product topology. It can be easily verified that ${\mathcal{A}}(X_f,f,S) ={\mathcal{A}}_f$. Note that as ${\mathcal{A}}_f$ is always separable, $X_f$ is metrizable.
If $\pi:(X,x_0,T)\to (Y,y_0,S)$ is a homomorphism of pointed transitive systems then the diagram $$\xymatrix
{
C(Y) \ar[d]_{\pi^*} \ar[r]^{j_{y_0}} & {\mathcal{A}}(Y,y_0)
\ar[d]^i \\
C(X) \ar[r]_{j_{x_0}} & {\mathcal{A}}(X,x_0)
}$$ commutes. Here $(\pi^* F)(x)=F(\pi x)$, for $F\in C(Y), x\in X$, and $i$ is the inclusion map. Conversely, if $B\subset C(X)$ is a closed conjugation-invariant, $T$-invariant subalgebra containing the constant functions, then the restriction map $$\pi: (|{\mathcal{A}}(X,x_0)|,{{{\rm{eva}}}}_0,T) \to
(|j_{x_0}(B)|,{{{\rm{eva}}}}_0,T)$$ is a pointed homomorphism of dynamical systems.
With every dynamical system $(X,T)$ we associate its [*enveloping semigroup*]{} $E=E(X,T) \subset X^X$. This is the pointwise closure of the set $\{T^n : n \in {\mathbb{Z}}\}$, as a subset of $X^X$. $E(X,T)$ is a compact [*right topological semigroup*]{}, i.e. for each $p \in E(X,T)$ right multiplication $q \mapsto qp,\ q \in E$ is continuous. The set $\{T^n : n \in {\mathbb{Z}}\}$ is contained in the center of $E$. In particular, for each $n \in {\mathbb{Z}}$ the map $p \mapsto T^n x,\ p \in E$ is a homeomorphism of $E$, so that via multiplication by $T$, $(E,T)$ is also a dynamical system.
A function $f \in \ell_\infty({\mathbb{Z}})$ is an [*almost periodic*]{} ([*weakly almost periodic*]{}) function if its orbit $\{S^n f: n \in {\mathbb{Z}}\}$ is norm precompact (weakly precompact) in the Banach space $\ell_\infty({\mathbb{Z}})$. We denote by $AP({\mathbb{Z}})$ and $WAP({\mathbb{Z}})$ the collections of almost periodic and weakly almost periodic functions in $\ell_\infty({\mathbb{Z}})$ respectively. These are ${\mathbb{Z}}$-algebras with $AP({\mathbb{Z}})
\subset WAP({\mathbb{Z}})$. A point-transitive dynamical system $(X,T)$ is called [*almost periodic*]{} iff ${\mathcal{A}}(X,x_0,T) \subset AP({\mathbb{Z}})$ iff for every $F$ in the Banach space $C(X)$ the orbit $\{F \circ T^n : n \in {\mathbb{Z}}\}$ is norm pre-compact. The system $(X,T)$ is [*weakly almost periodic*]{} iff ${\mathcal{A}}(X,x_0,T) \subset WAP({\mathbb{Z}})$ iff for every $F \in C(X)$ the orbit $\{F \circ T^n : n \in {\mathbb{Z}}\}$ is weakly pre-compact. It is well known that a point-transitive system $(X,T)$ is almost periodic iff it is equicontinuous and minimal. A theorem of Ellis and Nerurkar [@EN] based on a theorem of Grothendieck asserts that $(X,T)$ is weakly almost periodic iff its [*enveloping semigroup*]{} $E(X,T)$ consists of continuous maps. Another characterization of WAP systems (which again goes back to Grothedieck) is that $E(X,T)$ be a commutative semigroup. Using this observation it is also easy to deduce the following well known [*double limit criterion*]{} (stated here for a general topological group $G$, see e.g. [@Rupp]).
\[DLC\] For a topological group $G$, a bounded continuous function $f : G \to {\mathbb{C}}$ is WAP iff whenever $g_m, h_n$ are sequences in $G$ such that the double limits $$a=\lim_{n \to \infty}\lim_{m \to \infty} f(g_m h_n)
\quad {\text{and}}\quad
b=\lim_{m \to \infty}\lim_{n \to \infty} f(g_m h_n)$$ exist, then $a=b$.
For WAP ${\mathbb{Z}}$-systems we have $E(X,T) \cong (X,T)$, see e.g. [@Do] or [@G].
We summarize these results in the following:
\[WAP-thm\] Let $(X,T)$ be a point-transitive dynamical system and let $E=E(X,T)$ be its enveloping semigroup. The following conditions are equivalent.
1. The system $(X,T)$ is WAP.
2. $E$ is a semi-topological semigroup, i.e. both right and left multiplications are continuous.
3. $E$ is a commutative semigroup.
4. $E$ consists of continuous maps.
A point $x$ in a metric dynamical system $(X,T)$ is an [*equicontinuity point*]{} if for every ${\epsilon}>0$ there is a ${\delta}>0$ such that $d(x,x')< {\delta}{\Rightarrow}d(T^nx, T^nx')< {\epsilon}$ for all $n \in {\mathbb{Z}}$. A dynamical system with a residual set of equicontinuity points is called an [*almost equicontinuous*]{} system. A recurrent-transitive metric almost equicontinuous system $(X,T)$ is uniformly rigid and the the set of equicontinuity points in $X$ coincides with $X_{tr}$. Moreover, the map ${\Lambda}_X {\twoheadrightarrow}X_{tr},\ {\lambda}\mapsto {\lambda}x_0$, is a homeomorphism of the Polish group ${\Lambda}_X$ onto the dense $G_{\delta}$ subset $X_{tr} \subset X$ [@GW2].
By results of Glasner and Weiss [@GW] and Akin, Auslander and Berg [@AAB] it follows that if $(X,T)$ is a metric recurrent-transitive WAP dynamical system, then the system $(X,T)$ is [*uniformly rigid*]{}, i.e. there is a sequence of positive integers $n_k \to \infty$ such that the sequence $\{T^{n_k}: k \in {\mathbb{N}}\}$ tends uniformly to the identity map on $X$ (see [@GMa]). In a uniformly rigid system the uniform closure $${\Lambda}_X = {{\rm{cls\,}}}\{T^n: n \in {\mathbb{Z}}\} \subset {{\rm{Homeo\,}}}(X),$$ is a non-discrete Polish monothetic topological group. We then have $X_{tr} = {\Lambda}_X x_0$.
When $(X,T)$ is recurrent-transitive metric WAP we know much more: The system $(X,T)$ is [*hereditarily almost equicontinuous*]{}; i.e., every subsystem $Y \subset X$ is almost equicontinuous (see [@GW], [@AAB], [@GW2] and [@GM]).
Reflexive and Hilbert representability of groups and dynamical systems
======================================================================
1. We say that a Polish topological group $P$ is [*representable*]{} on a reflexive Banach space $V$ if there is a topological isomorphism of $P$ into the group of isometries ${{\rm{Iso\,}}}(V)$ of $V$ equipped with the strong operator topology. The group $P$ is [*reflexibly-representable*]{} it is representable on some reflexive Banach space.
2. The group $P$ is [*Hilbert-representable*]{} if there is a topological isomorphism of $P$ into the unitary group ${\mathcal{U}}({\mathcal{H}})$ of a Hilbert space ${\mathcal{H}}$ equipped with the strong operator topology (see Megrelishvili [@M1] and [@Me]).
3. As in Banaszczyk [@Ba] we say that $P$ is [*exotic*]{} if it does not admit any nontrivial continuous unitary representations. We say that $P$ is [*strongly exotic*]{} if it does not admit any nontrivial weakly continuous representations in Hilbert spaces.
<!-- -->
1. A metric dynamical system $(X,T)$ is called [*reflexively-representable*]{} if there is a reflexive Banach space $V$, a linear isometry $U \in Iso(V)$ and a weakly compact $U$-invariant subset $Z$ of $V$ such that the dynamical systems $(X,T)$ and $(Z,U)$ are isomorphic.
2. A metric dynamical system $(X,T)$ is called [*Hilbert-representable*]{} if there is a Hilbert space ${\mathcal{H}}$, a unitary operator $U \in {\mathcal{U}}({\mathcal{H}})$ and a weakly compact $U$-invariant subset $Z$ of ${\mathcal{H}}$ such that the dynamical systems $(X,T)$ and $(Z,U)$ are isomorphic.
3. Let $B({\mathbb{Z}})$ be the sub-algebra of $\ell_\infty({\mathbb{Z}})$ which consists of the Fourier-Steiltjes transforms of complex measures on the circle ${\mathbb{T}}$, i.e. $B({\mathbb{Z}}) = {\mathcal{F}}(M({\mathbb{T}}))$ where ${\mathcal{F}}: M({\mathbb{T}}) \to \ell_\infty({\mathbb{Z}})$ is the Fourier transform: $${\mathcal{F}}(\mu)(n) = \hat\mu(n) = \int_{{\mathbb{T}}} e^{int} \,d\mu(t).$$
4. Let $H({\mathbb{T}})={\overline}{B({\mathbb{Z}})}$ be the norm closure of $B({\mathbb{Z}})$ in the Banach space $\ell_\infty({\mathbb{Z}})$. Clearly $H({\mathbb{Z}})$ is a ${\mathbb{Z}}$-subalgebra of $\ell_\infty({\mathbb{Z}})$.
5. A point-transitive system $(X,x_0,T)$ is called [*Hilbert*]{} if ${\mathcal{A}}(X,x_0) \subset H({\mathbb{Z}})$. The elements of $H({\mathbb{Z}})$ are called [*Hilbert functions*]{}.
Suppose $(X,T)$ is reflexively-representable. Thus we assume that there is a reflexive Banach space $V$, a linear isometry $U \in {{\rm{Iso\,}}}(V)$ and that $X \subset V$ is a weakly compact $U$-invariant subset, with $T=U{\upharpoonright}X$. Let $x_0$ be a vector in $X \subset V$ and $\phi$ a vector in $V^*$. Set $F(x) =\langle x, \phi\rangle$ and $f(n) = F(T^nx_0)=\langle T^nx, \phi\rangle$ (such a function is called a [*matrix coefficient*]{} of the representation). Then, it is easy to see how the weak-compactness of the unit ball of $V^*$ implies that the $T$-orbit of $F$ in $C(X)$ is weakly precompact. Since ${\mathcal{A}}(X,x_0)\cong C(X)$ it follows that the $S$-orbit of $f$ in $\ell_\infty({\mathbb{Z}})$ is also weakly precompact, i.e. the function $f$ is in $WAP({\mathbb{Z}})$.
It turns out that the converse is also true. We have the following basic theorems of Shtern [@S] and Megrelishvili [@M1] concerning reflexive representability. We formulate these results in the context of a general topological group $G$, where the $C^*$-algebra $LUC(G)$ of bounded, complex valued, left uniformly continuos functions takes the place of $\ell_\infty({\mathbb{Z}})$. Thus e.g. $WAP(G)$ is the $G$-subalgebra of $LUC(G)$ comprising functions whose $G$-orbit is weakly pre-compact.
\[S-M\]
1. A topological group $G$ can be faithfully represented on a reflexive Banach space $V$ iff the $WAP(G)$ functions separate points and closed sets on $G$ [@S], [@M1].
2. Let $G$ be a topological group, then every $f \in WAP(G)$ is a matrix coefficient of a reflexive representation of $G$ [@M1].
3. A metrizable $G$-system $(X,G)$ is WAP iff it is reflexively-representable [@M1].
\[HR\] If a point-transitive metric system $(X,x_0,T)$ is Hilbert-representable then there is a function $F \in C(X)$ such that the corresponding $f \in {\mathcal{A}}(X,x_0)$, i.e. the function $f \in \ell_\infty({\mathbb{Z}})$ defined by $f(n) = F(T^nx_0)$, is positive definite and satisfies ${\mathcal{A}}(X,x_0) = {\mathcal{A}}_f$, where the latter is the ${\mathbb{Z}}$-subalgebra generated by $f$ in $\ell_\infty({\mathbb{Z}})$.
By assumption we can consider $X$ as a weakly compact subset of a separable Hilbert space ${\mathcal{H}}$ and $T= U{\upharpoonright}X$, where $U \in {\mathcal{U}}({\mathcal{H}})$, the unitary group of ${\mathcal{H}}$. With no loss in generality we also assume that $\|x_0\|=1$ and that ${\mathcal{H}}= Z_U(x_0)$, where the latter is the $U$-cyclic space generated by $x_0$. Set $F(x) = \langle x, x_0 \rangle$. Then $F\in C(X)$ and $$f(n) = F(U^nx_0) = \langle U^n x_0, x_0 \rangle$$ is indeed positive definite.
If $x \ne y$ are points in $X$ then, as ${\mathcal{H}}= Z_U(x_0)$, there is some $n \in {\mathbb{Z}}$ with $$F(U^{-n}x)=\langle x, U^{-n}x_0 \rangle
\ne \langle y, U^{-n}x_0 \rangle = F(U^{-n}y),$$ so that the sequence of functions $\{F\circ U^n\}_{n\in {\mathbb{Z}}}$ separates points on $X$, whence ${\mathcal{A}}(X,x_0) = {\mathcal{A}}_f$.
\[PD\] Suppose $f \in \ell_\infty({\mathbb{Z}})$ is positive definite.
1. The system $(X_f,f,S)$ is Hilbert-representable, i.e. there exists a Hilbert space ${\mathcal{H}}$, a unit vector $x_0 \in {\mathcal{H}}$ and a unitary operator $U \in {\mathcal{U}}({\mathcal{H}})$ such that ${\mathcal{H}}= Z_U(x_0)$, $X = w$-${{\rm{cls\,}}}\{U^nx_0 : n \in {\mathbb{Z}}\}$, and $(X_f,f,S) \cong (X,x_0,U)$.
2. Every element $g$ of $X_f= {\bar{\mathcal{O}}}(f) \subset [-\|f\|,\|f\|]^{\mathbb{Z}}$ has the form $g(j) = \langle U^j x_0, x \rangle$ for some $x \in X$. In particular $g \in B({\mathbb{Z}})$.
3. For $g \in X_f = {\bar{\mathcal{O}}}(f)$ we have: $g \in {\mathcal{O}}(f)$ iff there is $n \in {\mathbb{Z}}$ with $g(n) = \|g\|=\|f\|$, in which case $g = S^n f$. In particular for $g \in X_f$ we have $g(0) = \|g\|=\|f\| {\Rightarrow}g =f$.
4. $(X_f)_{tr} = \{g \in X_f : \|g\| = \|f\|\}$.
1. This is a consequence of Herglotz’ theorem. 2.Let $g \in X_f$. Then there exists a sequence $n_k \to \infty$ with $$g(j) = \lim_{k \to \infty} f(j + n_k) = \lim_{k \to \infty} S^{n_k} f(j) =
\lim_{k \to \infty} \langle U^j x_0, U^{n_k} x_0 \rangle
= \langle U^j x_0, x \rangle ,$$ where we assume, with no loss in generality, that $x = w$-$\lim_{k \to\infty}
U^{n_k}x_0$ exists. It follows that $g \in B({\mathbb{Z}})$.
3. Suppose $g \in X_f$ satisfies $g(0) = \|g\|=\|f\|$. As we have seen in part 1, there is an $x \in X$ with $g(j) = \langle U^j x_0, x \rangle$ for every $j\in {\mathbb{Z}}$ and our assumption reads: $$1= g(0) = \langle x , x_0 \rangle.$$ We conclude that $x = x_0$, hence $$g(j) = \langle U^j x_0, x_0 \rangle = f(j),$$ for all $j \in {\mathbb{Z}}$, i.e. $g = f$.
4. Let $g \in (X_f)_{tr}$. There is the a sequence $n_k$ with $\lim_{k \to \infty} S^{n_k} g(j) = \lim_{k \to \infty} g(j + n_k) = f(j)$ for every $j \in {\mathbb{Z}}$. In particular $\lim_{k \to \infty} g(n_k) = f(0) = \|f\|$. Thus $\|f \| \le \|g\| \le \|f\|$, hence $\|f \|= \|g\|$. Conversely, assuming $\|f \|= \|g\|$, we have $\lim_{k \to \infty} g(n_k) = f(0)=\|f\|$. With no loss of generality we can assume that $h = \lim_{k \to \infty} S^{n_k}g$ exists in $X_f$, so that $h(0) =\|f\|$. By part 3, $h = f$ and we conclude that $g \in (X_f)_{tr}$.
\[sh=pd\] A point-transitive system $(X,x_0,T)$ is Hilbert-representable iff there is a positive definite function $f \in \ell_\infty({\mathbb{Z}})$ such that $(X,x_0,T) \cong (X_f,f,S)$, where $X_f$ is the orbit closure of $f$ in $\ell_\infty({\mathbb{Z}})$ under the shift $S$ with respect to the pointwise convergence topology. In particular every Hilbert-representable system is Hilbert.
Combine Lemmas \[HR\] and \[PD\].
Every Hilbert system is WAP, i.e. $H({\mathbb{Z}}) \subset WAP({\mathbb{Z}})$.
It is well known (and easy to see) that $B({\mathbb{Z}}) \subset WAP({\mathbb{Z}})$. Since the algebra $WAP({\mathbb{Z}})\subset \ell_\infty({\mathbb{Z}})$ is closed we also have $H({\mathbb{Z}}) \subset WAP({\mathbb{Z}})$.
A structure theorem for Hilbert systems
=======================================
\[AK-ext\] Every metrizable recurrent-transitive Hilbert system $(Y,T)$ admits an almost 1-1 extension $(\tilde{Y},\tilde{T})$ which is a compact group-factor of a recurrent-transitive Hilbert-representable system $(X,U)$. Thus there exists a commutative diagram $$\label{AK}
\xymatrix
{
(X,U) \ar[dd]_{\pi}\ar[dr]^{{\sigma}} & \\
& (\tilde{Y},\tilde{T})\ar[dl]^{\rho}\\
(Y,T) &
}$$ with ${\sigma}$ a group-extension and $\rho$ an almost 1-1 extension.
1. Let $(Y,T)$ be a metrizable recurrent-transitive Hilbert system. Thus, picking a transitive point $y_0 \in Y$ we have ${\mathcal{A}}(Y,y_0) \subset H({\mathbb{Z}})$. Since every function in $B({\mathbb{Z}})$ is a linear combination of four positive definite functions, we can find a sequence $f_n$ of positive definite functions (with $f_1 \equiv {\mathbf{1}}$ and $f_n(0) = \|f_n\|_\infty = 1\ (n=1,2,\dots)$ such that ${\mathcal{A}}(\{f_n: n \in {\mathbb{N}}\})$, the closed translation and conjugation invariant algebra generated by the functions $f_n$, contains ${\mathcal{A}}(Y,y_0)$. For each $n$ there is a separable Hilbert space ${\mathcal{H}}_n$, a unit vector $x_n \in {\mathcal{H}}_n$, and a unitary operator $U_n \in {\mathcal{U}}({\mathcal{H}}_n)$ such that $f_n(k) = \langle U_n^k x_n,x_n \rangle \ (\forall k \in {\mathbb{Z}})$. Let ${\mathcal{H}}= \oplus_{n=1}^\infty {\mathcal{H}}_n$ (an $\ell_2$-sum), $x_0 = \oplus_{n=1}^\infty x_n \in {\mathcal{H}}$, and $U = \oplus_{n=1}^\infty U_n \in {\mathcal{U}}({\mathcal{H}})$. Set $X = {\text{weak-}}{{\rm{cls\,}}}\{U^k x_0 : k \in {\mathbb{Z}}\}$. Then it is easy to verify that the Hilbert-representable system $(X,U)$ admits $(Y,T)$ as a factor, say $\pi: (X,U) \to (Y,T)$.
2. Let ${\mathcal{X}}$ be the collection of all subsystems $(X',U)$ of $(X,U)$ with $\pi(X')= Y$. By Zorn’s lemma there is a minimal element $(Z,U)$ in ${\mathcal{X}}$. Clearly $(Z,U)$ is point-transitive Hilbert-representable system and, for convenience, we now rename it as $(X,U)$. Also, let us rename ${\mathcal{H}}=Z_U(x_0)$, the cyclic space generated by $\{U^n x_0 : n \in {\mathbb{Z}}\}$, where $x_0$ is now some transitive point in $X$.
3. Let $X_{tr} \subset X$ be the dense $G_{\delta}$ subset of the transitive points in $X$. Similarly let $Y_{tr} \subset X$ be the dense $G_{\delta}$ subset of the transitive points in $Y$. Clearly $\pi(X_{tr}) \subset Y_{tr}$ and we claim that these sets are equal. Suppose $y_1 \in Y_{tr}$. Let $x_1 \in X$ be some point with $\pi(x_1) = y_1$. By assumption $y_1$ is a transitive point and there is a sequence $T^{n_i}y_1 \to y_0$. We can assume that also $U^{n_i}x_1 \to x'_0$ for some point $x'_0 \in X$ with $\pi(x'_0) = y_0$. It follows that $\pi({\bar{\mathcal{O}}}_U(x'_0)) = Y$ and therefore, by minimality, ${\bar{\mathcal{O}}}_U(x'_0) =X$. Thus $x'_0$ is a transitive point and hence so is $x_1$ hence $y_1 = \pi(x_1) \not\in X_{tr}$. We conclude that indeed $\pi(X_{tr}) = Y_{tr}$.
4. Next recall that with every recurrent-transitive metrizable WAP system $(W,R)$ there is an associated non-discrete Polish monothetic group ${\Lambda}_W = {\text{unif-cls}}\{R^n: n \in {\mathbb{Z}}\} \subset {{\rm{Homeo\,}}}(W)$. Moreover, for every transitive point $w_0 \in W$, the map $g \mapsto g w_0$ is a homeomorphism from ${\Lambda}_W$ onto the set $W_{tr}$ of transitive points of $W$, see [@GW2]. Now using this observations and with notations as in the previous step, we note that the surjection $\pi: X_{tr} \to Y_{tr}$ defines a surjective Polish group homomorphism $\rho: {\Lambda}_X \to {\Lambda}_Y$. Moreover, we note that ${\rm ker}\,\rho = K$ is a compact subgroup of ${\Lambda}_X$.
5. Set $\tilde{Y} = X/K$ and let ${\sigma}: (X,U) \to (\tilde{Y}, \tilde{T})$ denote the corresponding group homomorphism. We now obtained the diagram (\[AK\]).
A Polish monothetic group $P$ which is reflexive but not Hilbert-representable
==============================================================================
We will use Banaszczyk’s construction of some families of Polish monothetic strongly exotic groups [@Ba]. For our purposes it suffices to consider one particular such example which we now proceed to describe. Let $E=\ell_4({\mathbb{N}})$ be the Banach space comprising the real valued sequences $u=\{x_j\}_{j=1}^\infty$ with $|u|:=(\sum_{j=1}^\infty x_j^4)^{1/4} < \infty$.
Let $\{e_n\}_{n=1}^\infty$ be the standard basis of unit vectors and choose a dense sequence $\{a_n\}_{n=1}^\infty \subset E$ with the property $a_n \in {{\rm{span}}}\{e_j\}_{j < n}$ for every $n \ge 1$. Let ${\Gamma}$ be the subgroup of $E$ generated by the sequence $\{e_n + a_n\}_{n=1}^\infty$. It can be easily checked that $|{\gamma}| \ge 1$ for every $0 \ne {\gamma}\in {\Gamma}$; and that, on the other hand, $E = {\Gamma}+ 2 B_1(0)$, where $B_1(0)$ denotes the unit ball in $E$. Set $P = E/{\Gamma}$ and equip $P$ with its quotient topology. It follows that $P$ is a Polish topological group. By Theorem 5.1 and Remark 5.2 in [@Ba] $P$ is strongly exotic and monothetic. In Remark 5.2 Banaszczyk leaves the proof of the fact that $P$ is monothetic as an exercise; for completeness we provide a proof of this fact at the end of this section, Proposition \[Mo\].
\[Ban\] The Polish monothetic group $P = \ell_4({\mathbb{N}})/{\Gamma}$ is reflexively-representable.
We will follow the ideas of Megrelishvili’s proof that the group $L_4([0,1])$ is reflexively-representable ([@Me-00], Lemmas 3.3 and 3.4). Note that Megrelishvili’s proof works almost verbatim to show that $\ell_4({\mathbb{N}})$ is reflexively-representable.
We first define an invariant metric on $P$ by $$d(u+{\Gamma},0) = \|u+{\Gamma}\| = \inf_{{\gamma}\in {\Gamma}} |u + {\gamma}|,\qquad u \in E.$$ Note that for $u \in E$ we have $$\label{=}
|u| \le \frac12
{\Rightarrow}\|u + {\Gamma}\| = |u|.$$ Next define the following continuous function on $P$: $$F(u +{\Gamma}) = \min\{\|u +{\Gamma}\|, \frac{1}{10}\}.$$ We will show that $F$ satisfies Grothendieck’s [*double limit condition*]{}, i.e. we have to show that given two sequences $\{u_m + {\Gamma}\}_{m=1}^\infty, \{v_n + {\Gamma}\}_{n=1}^\infty$ in $P$, if the limits $$\label{DL}
a=\lim_m\lim_n F(u_m + v_n + {\Gamma}) \quad {\text{and}} \quad
b=\lim_n\lim_m F(u_m + v_n + {\Gamma}),$$ exist, then necessarily $a =b$.
Of course when $a=b = \frac{1}{10}$ we are done; so we now assume that $a < \frac{1}{10}$. Then for some ${\delta}>0$ eventually $$a_m := \lim_{n \to \infty} F(u_m +v_n +{\Gamma}) =
\lim_{n \to \infty} \|u_m + v_n + {\Gamma}\| < \frac{1}{10}-2{\delta}.$$ Thus for some $m_0$ we have $a_m \le \frac{1}{10}-2{\delta}$ for every $m \ge m_0$; and for a fixed $m \ge m_0$ there is an $n(m)$ such that for $n \ge n(m)$ we have $\|u_m + v_n + {\Gamma}\| \le \frac{1}{10}-{\delta}$.
Now for $n \ge \max\{n(m), n(m_0)\}$ we have $$\|u_m + v_n + {\Gamma}\| \le \frac{1}{10}-{\delta}\quad {\text{and}}
\quad \|u_{m_0} + v_n + {\Gamma}\| \le \frac{1}{10}-{\delta},$$ hence $ \|u_m - u_{m_0} + {\Gamma}\| \le \frac{2}{10}$. Thus by an appropriate choice of representatives we can now assume that $\|u_m - u_{m_0} + {\Gamma}\| = |u_m - u_{m_0}| \le \frac{2}{10}$ for all $m\ge m_0$. By the same token, for $n \ge n(m_0)$ we can assume that also $\|u_{m_0} + v_n + {\Gamma}\| = |u_{m_0} + v_n| \le \frac{2}{10} -{\delta}$. This means that for sufficiently large $m$ and $n$ all the vectors $u_m$ and $-v_n$ lie in a ball of radius $\frac{2}{10}$ around $u_{m_0}$. By the triangle inequality and (\[=\]) $$\|u_m + v_n +{\Gamma}\| = |u_m + v_n| \le \frac{4}{10}-{\delta},$$ for all $m \ge m_0$ and $n \ge n(m_0)$. Now in $E$ we have: $$\begin{gathered}
|u_m + v_n|^4
= \sum_{i=1}^\infty u_m(i)^4 + 4u_m(i)^3v_n(i) + 6u_m(i)^2v^2_n(i)
+ 4u_m(i)v_n(i)^3 + v_n(i)^4.\end{gathered}$$ As all these vectors lie in a ball we can assume, by passing to subsequences, that $$\begin{aligned}
& \lim_{n \to \infty} v_n = v \quad {\text{weakly in}} \ \ell_4\\
& \lim_{n \to \infty}v^2_n = v^2 \quad {\text{weakly in}} \ \ell_2\\
& \lim_{n \to \infty}v^3_n = v^3 \quad {\text{weakly in}} \ \ell_{\frac{4}{3}},\end{aligned}$$ whence $$a_m = |u_m|^4 + 4\langle u_m^3,v \rangle + 6 \langle u_m^2,v^2 \rangle
+4 \langle u_m,v^3 \rangle + V^4,$$ where $V = \lim_{n \to \infty}|v_n|
$.
Again we can assume that also $$\begin{aligned}
& \lim_{m \to \infty} u_m = u \quad {\text{weakly in}} \ \ell_4\\
& \lim_{m \to \infty}u^2_n = u^2 \quad {\text{weakly in}} \ \ell_2\\
& \lim_{m \to \infty}u^3_n = u^3 \quad {\text{weakly in}} \ \ell_{\frac{4}{3}}.\end{aligned}$$ We now get $$a = \lim_{m\to\infty}a_m =
U^4 + 4\langle u^3,v \rangle + 6 \langle u^2,v^2 \rangle
+4 \langle u,v^3 \rangle + V^4,$$ where $U = \lim_{n \to \infty}|u_n|$.
Computing the double limit the other way we have similarly $$b =
U^4 + 4\langle u^3,v \rangle + 6 \langle u^2,v^2 \rangle
+4 \langle u,v^3 \rangle + V^4,$$ whence $a =b$, as required.
We have thus shown that the function $F$ satisfies the double limit condition. Now observe that, by the same computations, for any fixed $v \in E$ the function $F_v(u) = F(u - v)$ also satisfies the double limit property. By the double limit criterion \[DLC\] we conclude that the family $F_v, v \in E$, is contained in $WAP(P)$. As clearly the collection $\{F_v\}_{v \in E}$ generates the topology on $P$ it follows that this family separates points and closed sets on $P$ and we conclude from Theorem \[S-M\].(1), as in [@Me-00], that the group $P$ is indeed reflexively-representable.
\[WAPH-1\] There is a Polish monothetic group $P$ which admits a faithful representation on a reflexive Banach space but is not Hilbert-representable (in fact it is strongly exotic).
Let $P= \ell_4({\mathbb{N}})/{\Gamma}$ be the strongly exotic Polish monothetic group described by Banaszczyk [@Ba]. By Theorem \[Ban\] $P$ is reflexively-representable and, being strongly exotic, it is not Hilbert-representable.
\[Mo\] The group $P$ is monothetic.
Let $\pi_k: \ell_4({\mathbb{N}}) \to \ell^k_4:=E_k$ be the projection map onto the first $k$ coordinates. Let ${\Gamma}_k = \pi_k({\Gamma})$ and $P_k = E_k/{\Gamma}_k$. Note that for each $k \ge 1$ the group $P_k$ is a $k$-dimensional torus and by Kronecker’s theorem the vectors $z=(z_1,\dots,z_k) \in E_k$ such that $\{nz +{\Gamma}_k: n \in {\mathbb{Z}}\}$ is dense in $P_k$, form a dense subset of $E_k$. We will construct, by induction, a Cauchy sequence $\{z^{(k)}: k \in {\mathbb{N}}\}$ in $\ell_4({\mathbb{N}})$ whose limit $x =(x_1,x_2,\dots) \in \ell_4({\mathbb{N}})$ will have the property that $\{nx +{\Gamma}: n \in {\mathbb{Z}}\}$ is a dense subgroup of $P$.
Suppose we have already chosen $w^{(k)} =(z^{(k)}_1,z^{(k)}_2,\dots,z^{(k)}_k)
\in E_k$ with the property that for some positive integer $N_k$ $$\label{Nk}
\{nw^{(k)} +{\Gamma}_k: |n| \le N_k\},$$ is $\frac{1}{k}$-dense in $P_k$. We set $$z^{(k)}=(z^{(k)}_1,z^{(k)}_2,\dots,z^{(k)}_k,0,0,\dots)\in \ell_4({\mathbb{N}}).$$ There exists an ${\epsilon}_k >0$ such that the condition (\[Nk\]) (with $\pi_k(z)$ instead of $w^{(k)}$) is satisfied for every $z$ in an ${\epsilon}_k$-ball around $z^{(k)}$. As we have seen there is then a vector $w^{(k+1)} =(z^{(k+1)}_1,z^{(k+1)}_2,\dots,z^{(k+1)}_{k+1})
\in E_{k+1}$ and a positive integer $N_{k+1}$ such that the set $\{nw^{(k+1)} +{\Gamma}_{k+1}: |n| \le N_{k+1}\}$ is $\frac{1}{k+1}$-dense in $P_{k+1}$, and such that for $z^{(k+1)}=(z^{(k+1)}_1,z^{(k+1)}_2,\dots,z^{(k+1)}_k,0,0,\dots)\in \ell_4({\mathbb{N}})$ we have, $|z^{(k)} - z^{(k+1)}| < \min\{2^{-k},{\epsilon}_k\}$. This completes the inductive construction and clearly the unique limit point $x = \lim_{k \to \infty} z^{(k)}$ is the required one.
A WAP recurrent-transitive system which is not Hilbert
======================================================
\[WAPH-2\] The Banaszczyk group $P=\ell_4/{\Gamma}$ admits no nontrivial Hilbert system. Thus any nontrivial WAP action of $P$ provides an example of a recurrent-transitive WAP system which is not Hilbert.
1. Let $T\in P$ be a generator of a dense cyclic subgroup in $P$. Suppose that $P$ admits a nontrivial point-transitive Hilbert system $(Y,y_0,P)$ and consider the ${\mathbb{Z}}$-system $(Y,y_0,T)$. Since $P$ is exotic it is in particular MAP (= minimally almost periodic, i.e. it admits no nontrivial unitary finite dimensional representations) and it follows that $y_0$ is a non-periodic, recurrent point of $(Y,T)$.
Let $(X,x_0,U)$ with $x_0$ a unit vector in a Hilbert space ${\mathcal{H}}$, $U \in {\mathcal{U}}({\mathcal{H}})$ a unitary operator on ${\mathcal{H}}$, and $X = w{\text{-}}{{\rm{cls\,}}}\{U^nx_0 : n \in {\mathbb{Z}}\}$, be the extension provided by Step 5 of Theorem \[AK-ext\] (diagram (\[AK\]) above).
Let $G_X = {{\rm{cls\,}}}\{U^n : n \in {\mathbb{Z}}\}$, where the closure is taken in the strong operator topology on ${\mathcal{U}}({\mathcal{H}})$. It is not hard to see that there is a canonical topological isomorphism between the groups ${\Lambda}_X$ and $G_X$. There are also canonical topological isomorphisms ${\Lambda}_X/K \cong {\Lambda}_Y \cong {\Lambda}_{\tilde{Y}}$. Note that our assumption that $(Y,T)$ extends to a $P$-system means that the map $n \mapsto T^n$ extends to a continuous homomorphism $P \to {\Lambda}_Y$ with a dense image (see Step 4 of the proof of Theorem \[AK-ext\]).
By Proposition \[sh=pd\] we have an isomorphism $(X,U)\cong (X_f,S)$ where $f$ is the positive definite function $f(n) = \langle U^n x_0,x_0 \rangle$. Now let $({\Omega}, {\mathcal{F}}, \mu, R)$ be the Gauss dynamical system which corresponds to $f$; thus there is a ${\mathbb{Z}}$-sequence $\{\xi_n\}_{n \in {\mathbb{Z}}}$ of real valued random variables defined on ${\Omega}$ with joint Gauss distribution such that $\xi_n = \xi_0 \circ R^n$ with ${\mathbb{E}}(\xi_n\xi_m) = f(n -m)$, and such that the ${\sigma}$-algebra generated by the sequence $\{\xi_n\}_{n \in {\mathbb{Z}}}$ is ${\mathcal{F}}$. (For more details on Gauss dynamical systems see e.g. [@CFS] and [@LPT].) Let ${\mathcal{H}}_0 \subset L_2({\Omega},\mu)$ be the closed subspace spanned by $\{\xi_n\}_{n \in {\mathbb{Z}}}$ (the first Wiener chaos). Then ${\mathcal{H}}_0$ is isomorphic, as a Hilbert space, to ${\mathcal{H}}$ (which is the $U$-cyclic space spanned by $x_0$). In particular we can think of ${\Lambda}_X$ as a subgroup of ${\mathcal{U}}_0({\mathcal{H}}_0)$ where it commutes with $U_R$, the Koopman operator on $L_2({\Omega},\mu)$ which corresponds to $R$. It follows that ${\Lambda}_X$ can be realized a a group of measure preserving transformations (i.e. a subgroup of ${{\rm{Aut\,}}}({\Omega},\mu)$; see e.g. [@LPT]). In particular we can now view $K$ as a compact subgroup of ${{\rm{Aut\,}}}({\Omega},\mu)$ and then define the compact group-factor map $\pi_K : ({\Omega}, {\mathcal{F}}, \mu, R) \to ({\Omega}/K, {\mathcal{F}}/K, \mu_K, R_K)$.
Finally considering $L_2({\Omega}/K,\mu_K)$ we see that the Polish group ${\Lambda}_X/K$ is faithfully represented on this Hilbert space. Since we have a continuous homomorphism $P \to
{\Lambda}_Y \cong {\Lambda}_X/K$ with a dense image, this contradicts the fact that $P$ is exotic and our proof is complete.
\[rf\] There exist a recurrent-transitive WAP function $f \in WAP({\mathbb{Z}}) \setminus H({\mathbb{Z}})$.
Let $P$ be the Polish monothetic exotic Banaszczyk group. Let $T\in P$ be a generator of a dense cyclic subgroup in $P$. As we have seen in Theorem \[Ban\], the functions $\{F_v\}_{ v \in E}$ are in $WAP(P)$ and they separate points and closed sets on $P$. Therefore their restrictions $f_v= F_v {\upharpoonright}{\mathbb{Z}}$, to the dense subgroup $\{T^n : n \in {\mathbb{Z}}\} \subset P$, are in $WAP({\mathbb{Z}})$. As elements of $WAP(P)$ the functions $F_v$ are recurrent and hence so are the functions $f_v$,$v \in E$. Finally by Theorem \[WAPH-2\] each $f_v$ is an element of $WAP({\mathbb{Z}}) \setminus H({\mathbb{Z}})$.
\[AK-ext-cor\] If a Polish monothetic group $P$ is Hilbert-representable and if $K \subset P$ is a compact subgroup, then the quotient group $P/K$ is Hilbert-representable.
Follow the last steps of the proof of Theorem \[WAPH-2\] with $P$ taking the place of ${\Lambda}_X$.
Recurrent functions in $H({\mathbb{Z}}) \setminus B({\mathbb{Z}})$ {#R-H}
==================================================================
\[Rec-WAP\] For every recurrent $f \in WAP({\mathbb{Z}})$ the system $(X_f,S)$ is uniformly rigid and the ${\mathbb{Z}}$-algebra ${\mathcal{A}}_f$ consists entirely of recurrent functions.
The dynamical system $X_f=
{{\rm{cls\,}}}\{S^n f: n \in {\mathbb{Z}}\}$ (in the pointwise convergence topology) is WAP and therefore uniformly rigid. It then follows that the non-discrete Polish monothetic group ${\Lambda}_{X_f}= {{\rm{cls\,}}}\{S^n : n \in {\mathbb{Z}}\}$ (in the uniform convergence topology on ${{\rm{Homeo\,}}}(X_f)$) acts on $X_f$ as a group of automorphisms. In turn this implies that every function $g \in {\mathcal{A}}(X_f,f) = {\mathcal{A}}_f$ is a recurrent point: $g = \lim_{k \to \infty} S^{n_k}g$, for every sequence $n_k \nearrow\infty$ with $ S^{n_k}$ tending to the identity in ${\Lambda}_{X_f}$.
\[dense\] If ${\mathcal{A}}$ is an infinite-dimensional ${\mathbb{Z}}$-subalgebra of $H({\mathbb{Z}}) = {\overline}{B({\mathbb{Z}})}$ then ${\mathcal{A}}\setminus B({\mathbb{Z}})$ is a (norm) dense Borel subset of ${\mathcal{A}}$.
The Fourier transform ${\mathcal{F}}: M({\mathbb{T}}) {\twoheadrightarrow}B({\mathbb{Z}})$ is a 1-1 onto bounded linear map ($\|{\mathcal{F}}(\mu)\|_\infty \le \|\mu\|$). If ${\mathcal{A}}\subset {\mathcal{F}}(M({\mathbb{T}})) = B({\mathbb{Z}})$ then ${\mathcal{F}}^{-1}{\mathcal{A}}\subset M({\mathbb{T}})$ is an infinite-dimensional $C^*$-algebra, which is impossible (see e.g. Theorem 2.1 in [@DR]). Thus ${\mathcal{A}}\subsetneq B({\mathbb{Z}})$. Moreover, if the algebra ${\mathcal{A}}\cap B({\mathbb{Z}})$ contains a nonempty ${\mathcal{A}}$-norm-open subset then it must coincide with ${\mathcal{A}}$ which again is impossible. Thus ${\mathcal{A}}\setminus B({\mathbb{Z}})$ is norm-dense in ${\mathcal{A}}$.
Finally the algebra $B({\mathbb{Z}})$, being a 1-1 continuous image of $M({\mathbb{T}})$ (under ${\mathcal{F}}$), is a Borel subset of $H({\mathbb{Z}}) = {\overline}{B({\mathbb{Z}})}$, and hence so are the sets ${\mathcal{A}}\cap B({\mathbb{Z}})$ and ${\mathcal{A}}\setminus B({\mathbb{Z}})$.
In every infinite dimensional ${\mathbb{Z}}$-subalgebra ${\mathcal{A}}\subset H({\mathbb{Z}})$ the set ${\mathcal{A}}_{rec} \setminus B({\mathbb{Z}})$, where ${\mathcal{A}}_{rec}$ is the collection of recurrent functions in ${\mathcal{A}}$, is a (norm) dense subset of ${\mathcal{A}}_{rec}$. In particular there are recurrent functions in $H({\mathbb{Z}})$ which are not Fourier-Stieltjes transforms (i.e. are not in $B({\mathbb{Z}})$).
Combine Lemmas \[Rec-WAP\] and \[dense\].
A structure theorem for weakly almost periodic systems
======================================================
\[AK-ext-WAP\] Let $(X,T)$ be a metrizable recurrent-transitive WAP dynamical system and $\pi: (X,T) \to (Y,S)$ a factor. Then there is an almost 1-1 extension which is a compact group-factor of a recurrent-transitive subsystem $Z \subset X$. More explicitly there is a commutative diagram $$\label{di}
\xymatrix
{
(Z,T) \ar[dd]_{\pi}\ar[dr]^{{\sigma}} & \\
& (\tilde{Y},\tilde{T})\ar[dl]^{\rho}\\
(Y,S) &
}$$ with $Z \subset X$ a subsystem, ${\sigma}$ a group-extension and $\rho$ an almost 1-1 extension.
Just repeat the steps 2. to 5. of the proof of Theorem \[AK-ext\] above.
Some open problems
==================
\[abc\] In Theorem \[AK-ext\] we have shown that every metrizable recurrent-transitive Hilbert system admits an almost 1-1 extension which is a group-factor of a Hilbert-representable system. Can one get rid in this structure theorem of either one of these extensions or maybe of both? Thus, our question is whether a metrizable recurrent-transitive Hilbert system is always: (a) Hilbert-representable, or (b) a group-factor of a Hilbert-representable system, or (c) an almost 1-1 factor of a Hilbert-representable system. Question (a) can be reformulated as follows: Is a factor of a Hilbert-representable system also Hilbert-representable? This latter question is stated as Problem 996 in [@M-O].
Is there a metrizable recurrent-transitive Hilbert system $(X,T)$ with no nontrivial Hilbert-representable factors? In other words, is there a nontrivial ${\mathbb{Z}}$-algebra ${\mathcal{A}}\subset H({\mathbb{Z}})$ with ${\mathcal{A}}\cap B({\mathbb{Z}}) = {\mathbb{C}}$? Such a system will provide a counter-example to option (a) in Problem \[abc\].
[MMMMM]{}
E. Akin, J. Auslander, and K. Berg, [*When is a transitive map chaotic* ]{}, Convergence in Ergodic Theory and Probability, Walter de Gruyter & Co. 1996, pp. 25-40.
W. Banaszczyk, [*Additive subgroups of topological vector spaces*]{}, Lecture Notes in Math. 1466, Springer-Verlag, 1991.
I. P. Cornfeld, S. V. Fomin and Ya. G. Sinai, [*Ergodic theory*]{}, Springer-Verlag, New York, 1982.
T. Downarowicz, [*Weakly almost periodic flows and hidden eigenvalues*]{}, Topological dynamics and applications, Contemporary Mathematics [**215**]{}, a volume in honor of R. Ellis, 1998, pp. 101-120.
C. F. Dunkl and D. E. Ramirez, [*Topics in harmonic analysis*]{}, 1971, Appleton-Century-Crofts.
R. Ellis, [*Distal transformation groups*]{}, Pacific J. Math. [**8**]{}, (1957), 401-405.
R. Ellis and M. Nerurkar, [*Weakly almost periodic flows*]{}, Trans. Amer. Math. Soc. [**313**]{}, (1989), 103-119.
S. Ferri and J. Galindo, [*Embedding a topological group into its WAP-compactification*]{}, Studia Math. [**193**]{}, (2009), 99-108.
J. Galindo, [*On unitary representability of topological groups*]{}, Math. Z. [**263**]{}, (2009), 211-220.
S. Gao and V. Pestov, [*On a universality property of some abelian Polish groups*]{}, Fund. Math. [**179**]{}, (2003), 1-15.
E. Glasner, [*Ergodic Theory via joinings*]{}, Math. Surveys and Monographs, AMS, [**101**]{}, 2003.
E. Glasner and D. Maon, [*Rigidity in topological dynamics*]{}, Ergod. Th. Dynam. Sys. [**9**]{}, (1989), 309-320.
E. Glasner and M. Megrelishvili, *Linear representations of hereditarily non-sensitive dynamical systems*, Colloq. Math. [**104**]{}, (2006), no. 2, 223–283.
E. Glasner and B. Weiss, [*Sensitive dependence on initial conditions*]{}, Nonlinearity [**6**]{}, (1993), 1067-1075.
E. Glasner and B. Weiss, [*Locally equicontinuous dynamical systems*]{}, Collloq. Math., [**84/85**]{}, Part 2, (2000), 345-361.
B. Host, J.-F. Méla and F. Parreau, [*Analyse harmonique des mesures*]{}, Astérisque No. [**135-136**]{}, (1986), 261 pp.
M. Lemańczyk, F. Parreau and J.-P. Thouvenot, [*Gaussian automorphisms whose ergodic self-joinings are Gaussian*]{}, Fund. Math. [**164**]{}, (2000), 253–293.
M. Megrelishvili, [*Reflexively but not unitarily representable topological groups*]{}, Topology Proc. 25 (2000), 615-625, (2002).
M. Megrelishvili, [*Fragmentability and representations of flows*]{}, Topology Proceedings, [**27:2**]{} (2003), 497-544.
M. Megrelishvili, *Reflexively but not Hilbert representable compact flows and semitopological semigroups*, Colloq. Math., [**110**]{}, (2008), 383-407.
M. Megrelishvili, [*Topological transformation groups: selected topics*]{}, in [*Open problems in topology II*]{}, Editor: E. Pearl, Elsevier, 2007, pp. 421-435.
W. Rudin, [*Weak almost periodic functions and Fourier-Stieltjes transforms*]{}, Proc. Amer. Math. Soc., [**26**]{}, (1959), 215-220.
W. Ruppert, [*Compact semitopological semigroups*]{}, Springer Verlag, Lecture notes in Math., [**1079**]{}, 1984.
A. Shtern, [*Compact semitopological semigroups and reflexive representability*]{}, Russian J. of Math. Physics [**2**]{}, (1994), 131-132.
[^1]: [Both author’s work is supported by ISF grant \#1157/08 and the first author’s work is also supported by BSF grant \# 2006119]{}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'High spatial and spectral resolution observations of the atomic interstellar medium in nearby dwarf galaxies reveal evidence for warm and cold neutral gas, just like the phases in our own Galaxy. The cold or quiescent phase ($\approx$ 20% of the HI in the galaxies studied, except for LGS 3) seems to be associated with star formation activity— it may mark the regions where the conditions are right for star formation. These results help to explain the patterns of star formation activity which are seen in color-magnitude data for the dwarf irregulars.'
author:
- 'L. M. Young'
- 'L. van Zee'
- 'R. C. Dohm-Palmer'
- 'K. Y. Lo'
title: Star Formation and the ISM in Dwarf Galaxies
---
Motivation
==========
Color-magnitude diagrams of the stars within the nearest dwarf galaxies give detailed information about how many stars formed, where, and when. In general, star formation rates go up and down as time passes and the galaxies evolve. What is happening in the interstellar medium (ISM) to cause these fluctuations in the star formation rate? The issue is critical to an understanding of galaxy evolution. We attempt to answer this question by studying the properties of the ISM in a sample of dwarf galaxies with good color-magnitude information.
We also study the ISM in the nearby dwarfs because it may offer the closest analog to conditions which prevailed in the very early universe. Specifically, the nearby dwarfs have extremely low metallicities. Heavy metals in the gas phase and in dust are responsible for most of the heating and cooling of the interstellar medium, not to mention the formation of molecular gas. A better understanding of how the ISM works in the nearby dwarf galaxies may find broader application to galaxy evolution and very distant systems.
Sample and Observations
=======================
It has long been known that a careful analysis of HI line profiles can give insight into the temperature and density structure of the ISM (e.g. Radhakrishnan et al. 1972). We are applying these ideas to a sample of seven nearby ($\leq$ 3 Mpc) gas-rich dwarf galaxies. In order to study the HI line profiles without too much interference from rotational broadening in the finite HI beam, the galaxies must be low mass and/or close to face-on. The galaxies in our sample have HI masses about $10^7$ M$_\odot$, and dynamical masses are a few times larger than that. These galaxies have oxygen abundances in the range 4% to 10% of solar, which are approaching the lowest metallicities known for galaxies.
The HI line profiles are obtained from VLA observations with 15 (70 to 200 pc) and 1.3 km/s resolution. We also use ground-based optical broadband and H$\alpha$ imaging (e.g. van Zee 2000) to show regions of current and past star formation activity. Four of the galaxies have stellar color-magnitude information from HST, which allows us to trace the star formation activity as a function of time and place within the galaxy.
The Phase Structure of the Neutral ISM
======================================
Contrary to expectations, the HI line width as measured by a second moment of intensity [*decreases*]{} in regions of high column density and near star formation activity. This happens because the [*shape*]{} of the HI line profile changes significantly near regions of star formation. Throughout each galaxy studied, we find a broad HI component, well fitted by a Gaussian of dispersion 8–10 km/s, and containing about 80% of the HI in the galaxy. Its dispersion limits its kinetic temperature to be $< 10^4$ K. In some regions, near the locations of current star formation activity, there is superposed a narrower component of dispersion 3–5 km/s, with about 20% of the HI mass and kinetic temperature $<$ 1000 K. These temperature limits do not take into account possible line broadening from magnetic fields, turbulence, and so on. Figure 1 shows some of these types of spectra from the galaxy GR 8. More details about the analysis of line profiles can be found in Young & Lo (1996; 1997), where we discuss the galaxies Leo A, Sag DIG, and LGS 3.
![HI spectra from the dwarf irregular galaxy GR 8. The solid line is the data and the dashed line shows the fitted Gaussian components. Panel a is a spectrum which is well described by the broad HI component alone. Panel b requires a broad and a narrow HI component; panel c shows two components with velocities offset by about 5 km/s, possible evidence for kinetic energy input into the ISM by young massive stars. \[spectrafig\]](spectra.ps)
We identify the broader HI component as a warm (or more turbulent) neutral HI phase, and the narrower component as a cold (or more quiescent) HI phase. That’s because the properties of the dwarf irregular HI spectra are exactly analogous to those of the HI emission spectra in our own Galaxy (Kulkarni & Heiles 1988). In addition, the broad and narrow (warm and cold) HI features are also seen in an intermediate velocity HI cloud (Shaw et al. 1996) and the Magellanic Clouds (e.g. Dickey et al. 1994). The broad and narrow HI components should be seen in many other galaxies as well, if the sensitivity, spectral resolution, and no-rotational-broadening requirements can be met.
Because of the dependence of heating and cooling on heavy metals in the ISM, as mentioned above, it’s not obvious whether a stable multiphase medium similar to the one found in our Galaxy should exist in the dwarf irregulars. Young & Lo (1996, 1997) consider some theoretical models of the phase structure of the Milky Way ISM. They find that these Galactic models can be extended to conditions which are appropriate for the dwarf irregulars, and the models are consistent with observations if the thermal pressure in the dwarf irregulars is a bit lower than in our own Galaxy. These results give greater confidence that the Galactic ISM models may be extended to fit other systems as well— for example, high-redshift quasar absorption lines also indicate multiphase media (Churchill, this conference; Chengalur, this conference).
Star Formation and the ISM
==========================
What do the warm and cold phases of the [*atomic*]{} medium have to do with star formation? After all, the star formation is probably happening in [*molecular*]{} gas. But CO emission has not been detected in galaxies of this type (Taylor et al. 1998), presumably because of the low metallicity. We hypothesize that the cold, dense atomic phase is telling us where conditions are right for the formation of molecular gas and stars. Figure \[opticalfig\] shows that the narrow (cold) HI component is localized near regions of star formation activity in GR 8, whereas the broad (warm) HI component is spread throughout the galaxy.
The properties of the cold HI phase in the dwarf galaxies also explain some important results about how star formation happens in these galaxies. Dohm-Palmer et al. (1997, 1998) traced the star formation activity in Sextans A and in GR 8 as a function of both time and location. They find that star formation occurs in regions or complexes of size 100-200 pc; the activity lasts in one region for about 100 Myr and moves around through the galaxy in a way which is loosely organized at best. The narrow (cold) atomic component is found in patches of about the same size as the star formation regions (Figure \[opticalfig\]), so it may trace the raw material for star formation activity. The sizes and linewidths of the narrow HI clumps imply virial masses which are a few times larger than the HI masses alone; part of the difference may be caused by molecular gas.
We plan to test these ideas more fully by exploring the evolution of the ISM with time. For example, the cold HI phase may be a prerequisite for star formation but should be gradually destroyed as the young massive stars inject radiation and kinetic energy into the surrounding gas. We cannot watch the evolution of an individual region of a galaxy, but stellar color-magnitude diagrams for a sample of galaxies will place ages on many different regions. By finding a sequence of star formation regions of different ages, we may be able to watch the evolution of the ISM through the onset of star formation and the disruption of the raw material for star formation.
J. Salzer is gratefully acknowledged for providing the optical and H$\alpha$ images of GR 8.
Dickey, J. M., Mebold, U., Marx, M., Amy, S., Haynes, R. F., & Wilson, W. 1994, , 289, 357
Dohm-Palmer, R. C., Skillman, E. D., Gallagher, J. S., et al. 1998, , 116, 1227
Dohm-Palmer, R. C., Skillman, E. D., et al. 1997, , 114, 2527
Kulkarni, S., & Heiles, C. 1988, in Galactic and Extragalactic Radio Astronomy, ed. G. Verschuur & K. Kellermann (2d ed; New York: Springer), 95
Radhakrishnan, V., Murray, J. D., Lockhart, P., & Whittle, R. P. J. 1972, , 24, 15
Shaw, C. R., Bates, B., Kemp, S. N., Keenan, F. P., Daview, R. D., & Roger, R. S. 1996, , 473, 849
Taylor, C. L., Kobulnicky, H. A. and Skillman, E. D. 1998, , 116, 2746
van Zee, L. 2000, , 119, 2757
Young, L. M., & Lo, K. Y. 1996,
Young, L. M., & Lo, K. Y. 1997,
[**Question:**]{} (J. Rhoads) What are the dynamical times in these galaxies. ie., how long does it take a star to orbit the galaxy and forget where it was born?
[**Response:**]{} Dohm-Palmer et al. (1997) discuss this issue, and they argue that the remnant structures of star formation activity could easily last for 0.5 to 1 Gyr. One of the reasons for this longevity is that these small galaxies appear to be rotating in solid-body fashion, with no shear and no spiral density waves.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We investigate the coupling of a nanomechanical oscillator in the quantum regime with molecular (electric) dipoles. We find theoretically that the cantilever can produce single-mode squeezing of the center-of-mass motion of an isolated trapped molecule and two-mode squeezing of the phonons of an array of molecules. This work opens up the possibility of manipulating dipolar crystals, which have been recently proposed as quantum memory, and more generally, is indicative of the promise of nanoscale cantilevers for the quantum detection and control of atomic and molecular systems.'
author:
- 'S. Singh, M. Bhattacharya, O. Dutta and P. Meystre'
title: Coupling nanomechanical cantilevers to dipolar molecules
---
Recent experimental advances have brought macroscopic oscillators closer than ever before to operating in the quantum regime [@cohadon1999; @gigan2006; @kleckner2006; @arcizet2006; @schliesser2006; @corbitt2007]. Technically progress has been enabled by improvements in nanofabrication and non-equilibrium cooling. Foundational interest in this frontier lies in the fact that quantum mechanics has never been tested at such a macroscopic scale, particularly with respect to counter-intuitive effects such as superposition and entanglement. From a practical point of view it is important to explore the behavior of mechanical oscillators in the quantum regime since they serve as sensors whose precision is fundamentally restricted by quantum mechanics [@bocko1996; @rugar2004]. A broader perspective on the subject may be assumed by taking into account related successes in atomic physics, where laser cooling and trapping techniques have enabled impressive coherent control of microscopic systems. As part of an ongoing merger between atomic and condensed matter physics, it has become realistic to explore the interaction of cold atomic systems with quantum nano-mechanical oscillators. Examples include the coupling of cold ions to vibrating electrodes [@zoller2004], of a nanomechanical cantilever to a Bose-Einstein condensate [@haensch07], of an atomic vapor to an oscillating mirror [@genes2008], etc. As a first step in this direction, this Letter investigates the coupling of a laser-cooled nanomechanical cantilever to an ultracold lattice of polar molecules. Due to the anisotropic, long-range interaction between these molecules, ensembles of ultra-cold polar molecules are believed to have a rich phase diagram and are the subject of intense theoretical and experimental interest [@doyle2004]. The cantilever-molecule coupling is assumed to be enabled by a ferroelectric domain mounted on the former leading to a strong dipole-dipole interaction, and a strong polarizing field freezes out the rotational freedom of the dipoles.
To set the stage for the discussion we consider first the simple case of a single molecule, see setup A of Fig. \[fig:setup\], and demonstrate that its coupling to the cantilever leads to the parametric squeezing of its center-of-mass motion. These considerations are then generalized to the situation of a linear chain of electric dipoles (setup B). Such a self-arranged crystal has recently been proposed as memory for quantum information processing [@rabl2007]. We find that for an appropriate choice of cantilever frequency the phonons in the crystal can be two-mode squeezed, i.e. entangled [@merlin1997; @nori1997], hinting at the possibility of exploiting such a set-up for the coherent control of the quantum state of the dipolar lattice. This could for instance be achieved by using a learning algorithm to determine an appropriate time dependence of the cantilever frequency.
The Hamiltonian describing the coupling of the cantilever to a single molecule is $H = H_c + H_m + V_I $, where $$\label{eq:NMC}
H_c=\hbar \omega_{c} \left(a^{\dagger}a+\frac{1}{2}\right),$$ describes the quantized (single mode of) vibration of the cantilever of effective mass $m_c$ at frequency $\omega_c$, $a$ and $a^{\dagger}$ being bosonic creation and annihilation operators obeying the commutation rules $\left[a,a^{\dagger}\right]=1$. In terms of the displacement $x_c$ of the cantilever along the $y$-axis, we have $$x_c=\sqrt{\frac{\hbar}{2 \omega_c m_c}}(a+a^\dagger).$$ The Hamiltonian $$\label{eq:1mol} H_M=\hbar \omega_{t}
\left(b^{\dagger}b+\frac{1}{2}\right),$$ describes the center-of-mass motion of the trapped dipole of mass $m$, where $\omega_t$ is the trap frequency and $b, b^{\dagger}$ are bosonic annihilation and creation operators with $$x_m=\sqrt{\frac{\hbar}{2 \omega_t m}}(b+b^\dagger),$$ $x_m$ being the displacement of the molecule along the $x$-axis. Finally, the interaction between the molecule and the oscillator is given by $$\label{eq:1int}
V_I= \frac{d_{m}d_{c}}{4 \pi\epsilon_0 r^{3}}\left[1-\frac{3 (R+x_c)^2}
{r^{2}}\right],$$ where $d_c$ is the dipole moment of the ferroelectric domain attached to the tip of the nanomechanical cantilever and $R$ is its distance from the equilibrium position of the molecule. The distance between the cantilever and the molecule is $r=\left[
(R+x_c)^2+x_m^2\right]^{1/2}$. For $R \gg x_m,x_c$ the dipolar interaction can be approximated as: $$\label{eq:Vint1}
V_I \approx \frac{d_m d_c}{2 \pi \epsilon_0 R^6}(-R^3
+ 3x_cR^2+3x_m^2R- 15x_cx_m^2).$$ The typical trap level spacing is much larger than the thermal energy of the ultracold molecule, which justifies its zero-temperature description. We will be including thermal effects for the cantilever below.
The presence of the cantilever has two major effects on the molecule dynamics. First, it leads to a tightening of the trap for small distances $R$, resulting in a shifted trapping frequency $$\omega_{t}' = \left [\omega_t^2+\frac{3d_m d_c}{\pi \epsilon_0 m
R^5}\right ]^{1/2}.$$ The second, more interesting effect is parametric squeezing. In an interaction picture with respect to the free Hamiltonian $H_c+H_m$, taking $\omega_c = 2 \omega_{t}'$, performing the rotating-wave approximation, and further assuming that the cantilever motion can be treated classically, $(a \rightarrow \alpha)$, the interaction potential $V_I$ reduces to $$\label{eq:oms}
V_I= - \hbar C\left( b^2 + b^{\dagger 2}\right)$$ where $$C= \sqrt{\bar{N}}\frac{15d_m d_c}{4\pi\epsilon_{o} m \omega_{t}'R^{6}}\left(\frac{\hbar}{2m_c\omega_c}\right)^{1/2}.$$ Here $\bar{N}$ is the average number of quanta of excitation of the mechanical oscillator, $\bar{N}=k_BT_c/\hbar\omega_c$, where $k_B$ is the Boltzmann constant and $T_c$ the cantilever temperature. Eq. (\[eq:oms\]) is the standard squeezing Hamiltonian familiar from studies of the degenerate parametric amplifier in quantum optics, see e.g. Ref. [@zubairy1983].
Thermal noise can be introduced simply in the description of the system in the form of phase fluctuations in the cantilever field. These fluctuations are related to the cantilever damping rate, $D$ by the fluctuation-dissipation theorem. For times t such that $D < t^{-1}<2 C$, the variance in the dimensionless position quadrature for the molecule, $x_1= \frac{1}{2}(b+b^{\dagger})$, is then given by [@zubairy1983] $$(\Delta x_1)^2_t = \frac{1}{4}e^{-2u}+\frac{1}{8}e^{2u}Dt$$ where $u=2Ct$ is the squeezing parameter. Consider for example a nanomechanical cantilever with frequency $\omega_c = 4$MHz, effective mass $m_c = 10^{-16}$kg, and linewidth $D = 1$Hz. A ferroelectric domain with dipole moment $d_c = 2.1 \times 10^{-23}$C-m is attached to the cantilever and is placed at $R = 2\mu$m from a SrO molecule. These parameters give an oscillator frequency $\omega_t' = 2$MHz. We assume $\bar{N}=100$, yielding $C = 20.4$ Hz. Fig. \[fig:squeeze1mode\] shows the variance in $x_1$ as a function of the squeezing parameter for that example. We remark that the squeezing in single trapped ions [@meekhof96] and atoms [@raithel97] has been experimentally demonstrated, and similar measurement techniques techniques can possibly be implemented to detect squeezing in the present case.
We now extend these considerations to the case of a lattice of $N$ heteronuclear molecules contained in a harmonic trap $V_{t}$, as shown in Fig. 1.B. The trap is arranged so as to confine the sample weakly along $x$, tightly along $y$ and tightly or weakly along the $z$ direction depending on whether a one- or two-dimensional lattice crystal is desired. (We restrict our considerations to the case of a one-dimensional chain in the following.) A polarizing electric field is provided along $y$ so that all the dipoles align along that direction. The system can be described by the Hamiltonian $$\label{eq:classphon}
H_p = \displaystyle \sum_{i}^{N} \frac{p_{i}^{2}}{2m}
+\frac{d_{m}^{2}}{4\pi\epsilon_{o}} \displaystyle \sum_{i < j}^{N}
\frac{1}{\left|x_{i}-x_{j}\right|^3}+V_{t},$$ where $x_{i},p_{i}$ are the position and momentum, of the $i$th molecule and $V_{t}$ is the external trapping potential. The first term in Eq. (\[eq:classphon\]) corresponds to the kinetic energy of the dipoles, the second to their dipole-dipole interaction and the last term denotes the trap energy.
Due to their mutual repulsion along the $x$ direction the molecules self-organize into a linear lattice [@rabl2007]. For small oscillations of the molecules about their equilibrium positions Eq. (\[eq:classphon\]) can be expressed in terms of acoustic phonon modes of momentum $k$ and energy $\hbar \omega_{k}$, $$\label{eq:phonons}
H_p = \displaystyle \sum_{k} \hbar \omega_{k} \left(b^{\dagger}_{k}b_{k}
+\frac{1}{2}\right),$$ where $b_{k}, b^{\dagger}_{k}$ are the phonon creation and annihilation operators obeying the bosonic commutation rules $\left[b_{k},b_{k'}^{\dagger}\right]=\delta_{kk'}$, the phonon frequencies are given by $\omega_{k}=2\omega_{o}\left|\sin(kl/2)\right|$, where $\omega_{o}=d_{m}\left(3/2\pi\epsilon_{o}ml^{5}\right)^{1/2}$, and $l$ is the lattice spacing. We note that only terms harmonic in the $x_i$ have been retained in deriving Eq. (\[eq:phonons\]) from Eq. (\[eq:classphon\]). Higher, anharmonic terms represent phonon-phonon interactions and in particular determine the phonon lifetime in the crystal [@pathak65].
We consider ultracold molecules at a temperature $T$ such that $k_{B}T \ll \hbar \omega_{o},$ so that a $T=0$ description is appropriate as before for the molecules. The energy of the nanomechanical cantilever is again given by Eq. (\[eq:NMC\]), so that the interaction between the chain of molecules and the oscillator is $$\label{eq:int}
V_I= \displaystyle \sum_{i}\frac{d_{m}d_{c}}{r_{i}^{3}}\left[1-\frac{3 (R+x_c)^2}
{r_{i}^{2}}\right],$$ where $d_c$ is the dipole moment of the ferroelectric domain attached to the tip of the cantilever. Here $x_c$ is the displacement of the cantilever along the $y$ axis, $R$ is its distance from the center of the dipolar crystal and its distance from the $i$th molecule is given by $r_{i}=\left[
(R+x_c)^2+(il+x_{i})^2\right]^{1/2}$. Exploiting the hierarchy of length-scales, $x_{i} \ll l, Nl \ll R$, we expand Eq. (\[eq:int\]) to find that the oscillator produces a slight shift in the phonon frequency, $$\omega_{k}'\approx \omega_{k}+\frac{d_{m}d_{c}}{4\pi\epsilon_{o}m \omega_{k}^{2}R^{5}}.$$ and the coupling of the cantilever to the phonons is given by $$V_{I}=-\displaystyle \sum_{k} \hbar C_{k}' \left(a+a^{\dagger}\right)
\left(b_{k}b_{-k}+b_{k}^{\dagger}b_{-k}^{\dagger}+b_{k}^{\dagger}
b_{k}+b_{-k}b_{-k}^{\dagger}\right),$$ where $$\label{eq:ckp}
C_{k}'=-\frac{3d_m d_c}{2\pi\epsilon_{o} m \omega_{k}'R^{6}}\left(\frac{\hbar}{2m_c\omega_c}\right)^{1/2}.$$ We remark that for sufficiently small $R$ and/or sufficiently large $d_c$ the cantilever will couple to the individual dipoles rather than collectively to the acoustic phonons.
Following an approach that parallels the single-molecule description we work in an interaction picture with respect to the free Hamiltonian of the cantilever and the (frequency shifted) phonon mode. We further choose the cantilever frequency such that $\omega_c=2\omega_{k}'$, with $k=\pi/l$, implying that the cantilever couples primarily to excitations near the edge of the first Brillouin zone, where the density of phonon states is largest. We assume for simplicity that the motion of the nanomechanical cantilever can be described classically, $(a \rightarrow \alpha)$, a reasonable approximation since it is still challenging to cool these systems to their quantum regime. Performing the rotating wave approximation we then obtain the approximate interaction picture interaction Hamiltonian $$\label{eq:tms}
V_I= - \hbar C_{k}\left( b_{k}b_{-k}
+ b_{-k}^{\dagger}b_{k}^{\dagger}\right).$$ where $C_k =\sqrt{\bar{N}} C_k'$ and $\bar{N}$ is the average occupation number of the cantilever. This Hamiltonian is known from quantum optics to lead to the generation of two-mode squeezing between acoustic phonon modes of momenta $\pm k$ (within the bandwidth of the nanomechanical resonance), and hence their quantum entanglement.
In order to characterize that two-mode squeezing we introduce the two dimensionless quadratures as follows: $$\begin{aligned}
s_1= \frac{1}{\sqrt{2}}(b_k+b_{-k}+b_k^{\dagger}+b_{-k}^{\dagger})\\
s_2=\frac{1}{\sqrt{2}i}(b_k-b_k^{\dagger}-b_{-k}+b_{-k}^{\dagger})\end{aligned}$$ Taking into account the phase fluctuations in the cantilever motion resulting from thermal noise, the sum of variances in the two quadratures is then [@zubairy2006]: $$\begin{aligned}
&&\lefteqn{(\Delta s_1)^2+(\Delta s_2)^2= } \nonumber\\
& & \frac{e^{-\frac{Dt}{2}}}{C_{k0}}\{ D\sinh(C_{k0}t)+ 2C_{k0} \cosh (C_{k0}t)\} \nonumber \\
& & -\sum_{i,j,k,i\neq j,j\neq k} e^{\lambda_it}\frac{2 C_{k0}(\lambda_i+4D)}{(\lambda_i-\lambda_j)(\lambda_i-\lambda_k)},\end{aligned}$$ where $u$ is the squeezing parameter and is equal to $2C_{k0}t$, with $C_{k0} = \frac{1}{2}\sqrt{4C_k^2-D^2}$ and the $\lambda_i$’s are the roots of the cubic equation: $$\lambda^{3} + 5D\lambda^{2}+(4D^2-C_{k0}^2)\lambda - 2C_{k0}^2 D = 0$$ For example, let us consider a nanomechanical cantilever with frequency $\omega_c = 2$MHz, effective mass $m_c = 10^{-16}$kg, and linewidth $D = 1$Hz. We assume an average occupation number of 100. A ferroelectric domain with dipole moment $d_c = 2.1 \times 10^{-23}$C-m is attached to the cantilever and is placed at $R = 2\mu$m from a one dimensional SrO crystal. The crystalline phase of these dipolar molecules is formed with inter-molecular distances $l \approx 200$nm. These parameters give a phonon frequency $\omega_o = 4$MHz, and thus an interaction $C_k = 6.2$. The squeezing parameter $u$ is given by $2C_kt$. Fig. \[fig:squeeze2mode\] gives the sum of variances of $s_1$ and $s_2$ as a function of the squeezing parameter for this system.
The sum of the variances in Eq. 20 is actually a measure of the entanglement of the system [@zoller2000]. In our case, this inseparability criterion implies that the system is entangled if the sum of the variances is less than 2. We observe from Fig. 3 that as expected intuitively the introduction of phase fluctuations destroys the entanglement over time. As far as the experimental verification of this prediction is concerned we remark that squeezed states of phonons have been previously detected experimentally in solid-state systems [@merlin1997], and similar techniques can possibly be used to detect squeezed phonon modes in the present system.
As indicated earlier, our results assume that the motion of the cantilever is classical and dominated by thermal rather than quantum fluctuations. In addition to being a realistic description of the current experimental situation, this approximation enabled us to present the physics of the coupling between the cantilever and the dipolar molecules using simple analytical models. However, our results are expected to still hold at least qualitatively when treating the cantilever quantum-mechanically. Exact numerical solutions for one-mode and two-mode squeezing using a coherent pump with low $\bar{N}$ are available in literature [@kinsler93; @crouch88]. These results are consistent with our simple analytical treatment and point to the existence of squeezing for this system.
In conclusion, we studied the coupling of nanomechanical cantilevers to dipolar molecules. We found that for a single trapped molecule, the presence of the cantilever leads to tighter confinement and parametric squeezing. We also demonstrated squeezing and entanglement of the phonon modes of a linear chain of dipolar molecules. These results open up the way to extremely promising novel methods for the quantum manipulation and control of the state of ultracold dipolar systems, and are indicative of the general use of nanoscale cantilevers in the detection and control of atomic and molecular systems.
This work is supported in part by the US Office of Naval Research, by the National Science Foundation, and by the US Army Research Office.
[10]{}
P. F. Cohadon, A. Heidmann and M. Pinard, Phys. Rev. Lett. [**83**]{}, 3174 (1999).
S. Gigan *et al.*, Nature [**444**]{}, 67 (2006).
D. Kleckner and D. Bouwmeester, Nature [**444**]{}, 75 (2006).
O. Arcizet, P. -F. Cohadon, T. Briant, M. Pinard, and A. Heidmann, Nature [**444**]{}, 71 (2006).
A. Schliesser *et al.*, Phys. Rev. Lett. [**97**]{}, 243905 (2006).
T. Corbitt *et al.*, Phys. Rev. Lett. [**98**]{}, 150802 (2007).
M. Bocko and R. Onofrio, Rev. Mod. Phys. [**68**]{}, 755 (1996).
D. Rugar, R. Budakian, H. J. Mamin and B. W. Chui, Nature [**430**]{}, 329 (2004).
L. Tian and P. Zoller, Phys. Rev. Lett. [**93**]{}, 266403 (2004).
P. Treutlein *et. al*, Phys. Rev. Lett. [**99**]{}, 140403 (2007).
C. Genes, D. Vitali, P. Tombesi, arXiv:0801.2266 (2008).
J. M. Doyle, B. Friedrich, R. V. Krems, and F. Masnou-Seeuws, Eur. Phys. J. D [ **31**]{}, 149 (2004).
P. Rabl and P. Zoller, Phys. Rev. A [**76**]{}, 042308 (2007).
G. A . Garrett, A. G. Rojo, A. K. Sood, J. F. Whitaker and R. Merlin, Science [**275**]{}, 1638 (1997).
X. Hu and F. Nori, Phys. Rev. Lett. [**79**]{}, 4605 (1997).
K. Wodkiewicz and M. S. Zubairy, Phys. Rev. A [**27**]{}, 2003 (1983).
D.M. Meekhof, C. Monroe, B.E. King, W.M. Itano and D.J. Wineland, Phys. Rev. Lett. [**76**]{}, 1796 (1996).
G. Raithel, G. Birkl, W.D. Phillips, and S.L. Rolston, Phys. Rev. Lett. [**78**]{}, 2928 (1997).
K. Pathak, Phys. Rev. [**139**]{}, 1569 (1965).
K. Ahmed, H. Xiong and M.S. Zubairy, Opt. Comm. [**262**]{}, 129 (2006).
L-M. Duan, G. Giedke, J.I. Cirac and P. Zoller, Phys. Rev. Lett. [**84**]{}, 2722 (2000).
P. Kinsler, M. Fernee and P.D. Drummond, Phys. Rev. A [**48**]{}, 3310 (1993).
D. Crouch and S. Braunstein, Phys. Rev. A [**38**]{}, 4696 (1988).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
In this paper, we construct stationary classical solutions of the incompressible Euler equation approximating singular stationary solutions of this equation. This procedure is carried out by constructing solutions to the following elliptic problem $$\begin{cases}
-{\varepsilon}^2 \Delta u=\sum_{j=1}^m \chi_{{\Omega}_j}(u-q-\frac{\kappa_j}{2\pi}\ln\frac{1}{{\varepsilon}})_+^p, \quad & x\in\Omega, \\
u=0, \quad & x\in\partial\Omega,
\end{cases}$$ where $p>1$, $\Omega\subset\mathbb{R}^2$ is a bounded domain, ${\Omega}_i\subset\subset{\Omega}, i=1\cdots,m$ are suitable small domains such that ${\Omega}_i\bigcap{\Omega}_j $ is empty if $i\neq j$ and $q$ is a harmonic function.
We showed that if $\Omega$ is simply-connected smooth domain, then for any given stable critical point of Kirchhoff-Routh function $\mathcal{W}(z_1,\cdots,z_m)$ with the positive strength $\kappa_i>0$, there is a stationary classical solution approximating stationary $m$ points vortex solution of incompressible Euler equations with vorticity $\sum_{j=1}^m\kappa_i$.
Existence and asymptotic behavior of single point non-vanishing vortex solutions were studied by D. Smets and J. Van Schaftingen in [@SV].\
*AMS 2000 Subject Classifications: Primary $35\mathrm{J}60$; Secondary $35\mathrm{JB}05$; $35\mathrm{J}40$ Keywords: The Euler equation; Multiple non-vanishing vortices; Free boundary problem.*
address:
- 'Institute of Applied Mathematics, Chinese Academy of Science, Beijing 100190, P.R. China'
- 'Institute of Applied Mathematics, Chinese Academy of Science, Beijing 100190, P.R. China'
- 'Department of Mathematics, The Chinese University of Hong Kong, Shatin, N.T., Hong Kong '
author:
- Daomin Cao
- Zhongyuan Liu
- Juncheng Wei
title: Regularization of point vortices for the Euler equation in dimension two
---
Introduction and main results
=============================
The incompressible Euler equations $$\label{1.2}
\begin{cases}
\mathbf{v}_t+(\mathbf{v}\cdot\nabla)\mathbf{v}=-\nabla P,\\
\nabla\cdot\mathbf{v}=0,
\end{cases}$$ describe the evolution of the velocity $\mathbf{v}$ and the pressure $P$ in an incompressible flow. In $\mathbb{R}^2$, the vorticity of the flow is defined by $\omega=\nabla\times\mathbf{v}:=\partial_1
v_2-\partial_2 v_1$, which satisfies the equation $$\omega_t+\mathbf{v}\cdot\nabla\omega=0.$$
Suppose that $\omega$ is known, then the velocity $\mathbf{v}$ can be recovered by Biot-Savart law as following: $$\mathbf{v}=\omega\,*\frac{1}{2\pi}\frac{-x^\bot}{|x|^2},$$ where $x^\bot=(x_2,\,-x_1)$ if $x=(x_1,\,x_2)$. One special singular solutions of Euler equations is given by $\omega=\sum^m_{i=1}\kappa_i\delta_{x_i(t)}$, which is related $$\mathbf{v}=-\sum^m_{i=1}\frac{\kappa_i}{2\pi}\frac{(x-x_i(t))^\bot}{|x-x_i(t)|^2}.$$ and the positions of the vortices $x_i: \mathbb{R}\rightarrow
\mathbb{R}^2$ satisfy the following Kirchhoff law: $$\kappa_i\,\frac{dx_i}{dt}=(\nabla_{x_i}\mathcal{W})^\bot$$ where $\mathcal{W}$ is the so called Kirchhoff-Routh function defined by $$\mathcal{W}(x_1,\cdots, x_m)=\frac{1}{2}\sum_{i\neq
j}^m\frac{\kappa_i\kappa_j}{2\pi}\log\frac{1}{|x_i-x_j|}.$$
In simply-connected bounded domain $\Omega\subset \mathbb{R}^2$, similar singular solutions also exist. Suppose that the normal component of $\mathbf{v}$ vanishes on $\partial\Omega$, then the Kirchhoff-Routh function is
$$\label{W1}
\mathcal{W}(x_1,\cdots, x_m)=\frac{1}{2}\sum_{i\neq
j}^m{\kappa_i\kappa_j}G(x_i,\,x_j)+
\frac{1}{2}\sum_{i=1}^m{\kappa_i^2}H(x_i,\,x_i),$$
where $G$ is the Green function of $-\Delta$ on $\Omega$ with 0 Dirichlet boundary condition and $H$ is its regular part (the Robin function).
Let $v_n$ be the outward component of the velocity $\mathbf{v}$ on the boundary $\partial\Omega$, then we see that $\int_{\partial\Omega}v_n=0$ due to the fact that $\nabla\cdot\mathbf{v}=0$. Suppose that $\mathbf{v}_0$ is the unique harmonic field whose normal component on the boundary $\partial\Omega$ is $v_n$, then $\mathbf{v}_0$ satisfies
$$\label{v0}
\begin{cases}
\nabla\cdot\mathbf{v}_0=0,\,\,\text{in}\,\Omega,\\
\nabla\times\mathbf{v}_0=0,\,\,\text{in}\,\Omega,\\
n\cdot\mathbf{v}_0=v_n,\,\, \text{on}\,\partial\Omega.
\end{cases}$$
If $\Omega$ is simply-connected, then $\mathbf{v}_0$ can be written $\mathbf{v}_0=(\nabla\psi_0)^\bot$, where the stream function $\psi_0$ is determined up to a constant by $$\label{psi}
\begin{cases}
-\Delta \psi_0=0,\,\,\text{in}\,\Omega,\\
-\displaystyle\frac{\partial\psi_0}{\partial\tau}=v_n,\,\,
\text{on}\,\partial\Omega,
\end{cases}$$ where $\frac{\partial\psi_0}{\partial\tau}$ denotes the tangential derivative on $\partial\Omega$. The Kirchhoff-Routh function associated to the vortex dynamics becomes(see Lin [@Lin])
$$\label{KR}
\mathcal{W}(x_1,\cdots,x_m)=\frac{1}{2}\sum_{i\neq
j}^m\kappa_i\kappa _jG(x_i,x_j)+\frac{1}{2}\sum^{m}_{i=1}\kappa^2
_iH(x_i,x_i)+\sum^{m}_{i=1}\kappa_i\psi_0(x_i).$$
It is known that critical points of the Kirchhoff-Routh function $\mathcal{W}$ give rise to stationary vortex points solutions of the Euler equations. As for the existence of critical points of $\mathcal{W}$ given by , we refer to [@BT].
Roughly speaking, there are two methods to construct stationary solutions of the Euler equation, which are the vorticity method and the stream-function method. The vorticity method was first established by Arnold and Khesin [@AK] and further developed by Burton [@B] and Turkington [@T].
The stream-function method consists in observing that if $\psi$ satisfies $-\Delta \psi=f(\psi)$ for some function $f\in
C^1(\mathbb{R})$, then $\mathbf{v}=(\nabla\psi)^\bot$ and $P=F(\psi)-\frac{1 }{2}|\nabla\psi|^2$ is a stationary solution to the Euler equations, where $(\nabla\psi)^\bot:=(\frac{\partial\psi}{\partial
x_2},-\frac{\partial\psi}{\partial x_1}), F(t)=\int_0^tf(s)ds$. Moreover, the velocity $\mathbf{v}$ is irrotational on the set where $f(\psi)=0$.
Set $q=-\psi_0$ and $u=\psi-\psi_0$, then $u$ satisfies the following boundary value problem $$\label{1.3}
\begin{cases}
-\Delta u=f(u-q),\quad & x\in\Omega,\\
u=0,\quad &x\in\partial\Omega.
\end{cases}$$ In addition, if we suppose that $\inf_\Omega q>0$ and $f(t)=0,~t\leq0$, the vorticity set $\{x: f(\psi)>0\}$ is bounded away from the boundary.
The motivation to study is to justify the weak formulation for point vortex solutions of the incompressible Euler equations by approximating these solutions with classical solutions.
Marchioro and Pulvirenti [@MP] have approximated these solutions on finite time intervals by considering regularized initial data for the vorticity. On the other hand, the stationary point vortex solutions can also be approximated by stationary classical solutions. See e.g. [@BF1; @N; @SV; @T; @Y1] and the references therein. It is worth pointing out that the above approximations can just give explanation for the formulation to single point vortex solutions. In this paper, we will show that multi-point vortex solutions can be approximated by stationary classical solutions. There are many results for problem on the existence and asymptotic behavior of solutions under various assumptions. In [@Ba; @BF1; @FB; @N; @T], the constrained variation methods were used to find solutions for the equation $$\label{1.4}
\begin{cases}
-\Delta u=\lambda f(u-q),\quad & x\in\Omega,\\
u=0,\quad &x\in\partial\Omega,
\end{cases}$$ under the constraint $\int_\Omega F(u-q)=\mu$, where $\lambda>0$ is a Lagrange multiplier a priori unknown. On the other hand, in [@AS; @AY; @Ni; @Y1; @Y2], the solutions were obtained by using Mountain Pass Lemma for various nonlinearities. For the asymptotic behavior, Berger and Fraenkel [@BF1] began studying the asymptotic behavior for variable $\mu$ and $q$, but the lack of information about $\lambda$ is still an obstacle.
To avoid this obstacle, Yang [@Y1] studied the minimization of the functional $\frac{1}{2}\int_\Omega|\nabla
u|^2-\frac{1}{{\varepsilon}^2}\int_\Omega F(u-q)$ under the natural constraint $\int_\Omega|\nabla u|^2-\frac{1}{{\varepsilon}^2}\int_\Omega uf(u-q)=0$ and obtained the asymptotic behavior of the solutions $u_{\varepsilon}$ as ${\varepsilon}\rightarrow0$ for $\Omega=\mathbb{R}^2_+,~q(x)=Wx_1+d$, where $W,d>0$. That is, set $A_{\varepsilon}=\{x\in\mathbb{R}^2_+: f(u_{\varepsilon}-q)>0\},
\kappa_{\varepsilon}=\frac{1}{{\varepsilon}^2}\int_\Omega f(u_{\varepsilon}-q)$ and $x_{\varepsilon}\in
A_{\varepsilon}$, then $diam A_{\varepsilon}\rightarrow0,~
dist(x_{\varepsilon},\partial\mathbb{R}_+^2)\rightarrow0$ and $\frac{u_{\varepsilon}}{\kappa_{\varepsilon}}-G(x_{\varepsilon},\cdot)\rightarrow0$ in $W_{loc}^{1,r}(\mathbb{R}_+^2)$ for $r\in[1,2)$. Later on, similar results were obtained in [@LYY] for bounded domains with additional information that $q(x_{\varepsilon})\rightarrow\min_\Omega q$. However, it has been pointed out in [@SV] that the solutions obtained above corresponded to desingularization of point-vortex solutions with vanishing vorticity. To get non-vanishing vortex solutions, D. Smets and J. Van Schaftingen [@SV] investigated the following problem $$\label{1.1}
\begin{cases}
-{\varepsilon}^2 \Delta u=\left(u-q-\frac{\kappa}{2\pi}\ln\frac{1}{{\varepsilon}}\right)_+^{p}, & \text{in}\;{\Omega},\\
u=0, &\text{on}\; \partial{\Omega},
\end{cases}$$ and gave exact asymptotic behavior and expansion of the least energy solution by estimating the upper bounds on the energy. The solutions for in [@SV] were obtained by finding a minimizer of the corresponding functional in a suitable function space, which can only give approximation to a single point non-vanishing vortex. This method is hard to obtain multiple non-vanishing solutions.
In this paper, we approximate stationary vortex solutions of Euler equations with multiple non-vanishing vorticity by stationary classical solutions.
Our main result concerning is the following:
\[nth1\]
Suppose that $\Omega\subset \mathbb{R}^2$ is a bounded simply-connected smooth domain. Let $v_n:
\partial\Omega\rightarrow \mathbb{R}$ be such that $v_n\in
L^s(\partial\Omega)$ for some $s>1$ satisfying $\int_{\partial\Omega}v_n=0$. Let $\kappa_i>0,~i=1,\cdots,m$. Then, for any given stable critical point $(x_1^*,\cdots,x_m^*)$ of Kirchhoff-Routh function $\mathcal{W}(x_1,\cdots,x_m)$ defined by , there exists ${\varepsilon}_0>0$, such that for each ${\varepsilon}\in
(0,{\varepsilon}_0)$, problem has a stationary solution $\mathbf{v}_{\varepsilon}$ with outward boundary flux given by $v_n$, such that its vorticities $\omega_{\varepsilon}$ satisfying $$supp (\omega_{\varepsilon})\subset\cup_{i=1}^m
B(x_{i,\,{\varepsilon}},C{\varepsilon})~~\text{for}~~x_{i,\,{\varepsilon}}\in\Omega,
~~i=1,\cdots,m$$ and as ${\varepsilon}\rightarrow0$ $$\int_\Omega \omega_{\varepsilon}\rightarrow \sum_{i=1}^m\kappa_i,$$
$$(x_{1,\,{\varepsilon}},\cdots,x_{m,\,{\varepsilon}})\rightarrow (x_1^*,\cdots,x_m^* ).$$
The simplest case, corresponding to a single point vortex ( $m=1$ ) was studied by Smets and Van Schaftingen [@SV] by minimizing the corresponding energy functional. In their paper $\mathcal{W}(x_{1,\,{\varepsilon}})\rightarrow
sup_{x\in\Omega}\mathcal{W}(x)$. Even in the case $m=1$, our result extends theirs to general critical points (with additional assumption that the critical point is non-degenerate). The method used in [@SV] can not be applied to deal with general critical point cases. The method used here is constructive and is completely different from theirs.
In this case that $m=1$ suppose that $x_1$ is a strict local maximum(or minimum) point of Kirchhoff-Routh function $\mathcal{W}(x)$ defined by , statement of Theorem \[nth1\] still holds which can be proved similarly(see Remark \[re1.4\]). Thus we can obtain corresponding existence result in [@SV].
Theorem \[nth1\] is proved via considering the following problem $$\label{0}
\begin{cases}
-{\varepsilon}^2 \Delta u=\sum_{j=1}^m \chi_{{\Omega}_j}(u-q-\frac{\kappa_j}{2\pi}\ln\frac{1}{{\varepsilon}})_+^p, \quad & x\in\Omega, \\
u=0, \quad & x\in\partial\Omega,
\end{cases}$$ where $p>1$, $q\in C^2(\Omega)$, $\Omega\subset\mathbb{R}^2$ is a bounded domain, ${\Omega}_j\subset\Omega$ is a subdomain such that $x^*_j\in{\Omega}_j,\,j=1,\cdots,m$ and ${\Omega}_i\cap{\Omega}_j={\O}$ if $ i\neq
j$.
\[th3\]
Suppose $q\in C^2(\Omega)$. For $\kappa_i>0,~i=1,\cdots,m$. Then, for any given $C^1$-stable critical point $(x_1^*,\cdots,x_m^*)$ of Kirchhoff-Routh function $\mathcal{W}(x_1,\cdots,x_m)$ defined by , there exists ${\varepsilon}_0>0$, such that for each ${\varepsilon}\in (0,{\varepsilon}_0)$, has a solution $u_{\varepsilon}$, such that the set $\Omega_{{\varepsilon},i} =\{x:
u_{\varepsilon}(x)-\frac{\kappa_i}{2\pi}\,\ln\frac{1}{{\varepsilon}}-q(x)>0\}\subset\subset{\Omega}_i,
\,i=1,2,\cdots,m$ and each $\Omega_{{\varepsilon},\,i}$ shrinks to $x_i^*\in
\Omega$, as ${\varepsilon}\to 0$.
\[re1.4\] For the case $m=1$ suppose that $x_1$ is a strict local maximum(or minimum) point of Kirchhoff-Routh function $\mathcal{W}(x)$ defined by , statement of Theorem \[th3\] still holds which can be proved by making corresponding modification of the proof of Theorem \[th3\] in obtaining critical point of $K(z)$ defined by (see Propositions 2.3,2.5 and 2.6 [@Cao] for detailed arguments).
For domains which may not be simply-connected, we show in the following result that the topology of the domain plays an important role in the existence of solutions.
\[th1\]
Suppose that the homology of ${\Omega}$ is nontrivial. Then, for any positive integer $m$, there exists ${\varepsilon}_0>0$, such that for each ${\varepsilon}\in (0,{\varepsilon}_0)$, has a solution $u_{\varepsilon}$, such that the set $\Omega_{{\varepsilon},i} =\{x:
u_{\varepsilon}(x)-\frac{\kappa_i}{2\pi}\,\ln\frac{1}{{\varepsilon}}-q(x)>0\}\subset\subset{\Omega}_i,
\,i=1,2,\cdots,m$ and each $\Omega_{{\varepsilon},\,i}$ shrinks to a point $x_i^*\in \Omega$, as ${\varepsilon}\to 0$. Moreover $x_i^*\neq x_j^*$, if $
i\neq j$.
Since $m$ is arbitrary, from Theorem \[th1\], we can see that the number of solutions for is unbounded as ${\delta}\to 0$.
Not as in [@SV] where is investigated directly, we prove Theorem \[nth1\], Theorem \[th3\] and Theorem \[th1\] by considering an equivalent problem of instead. Let $ w=\frac{2\pi}{|\ln{\varepsilon}|}u$ and $\delta={\varepsilon}(\frac{2\pi}{ |\ln{\varepsilon}|})^{\frac{p-1}{2}}$, then becomes $$\label{1}
\begin{cases}
-{\delta}^2 \Delta w=\sum_{j=1}^m\chi_{{\Omega}_j}\left(w-\kappa_j-\frac{2\pi}{ |\ln{\varepsilon}|}q(x)\right)_+^{p}, & \text{in}\;{\Omega},\\
w=0, &\text{on}\; \partial{\Omega}.
\end{cases}$$
We will use a reduction argument to prove Theorem \[th3\] and Theorem \[th1\]. To this end, we need to construct an approximate solution for . For the problem studied in this paper, the corresponding “limit" problem in $\mathbb{R}^2$ has no bounded nontrivial solution. So, we will follow the method in [@CPY; @DY] to construct an approximate solution. Since there are two parameters ${\delta},~{\varepsilon}$ in problem , which causes some difficulty, we must take this influence into careful consideration and give delicate estimates in order to perform the reduction argument. For example we need to consider $(s_{1,{\delta}},\cdots,s_{m,{\delta}})$ and $(a_{1,{\delta}},\cdots,a_{m,{\delta}})$ together in Lemma \[l2.1\].
As a final remark, we point out that problem can be considered as a free boundary problem. Similar problems have been studied extensively. The reader can refer to [@CF; @CPY; @DY; @FW; @LP] for more results on this kind of problem.
This paper is organized as follows. In section 2, we construct the approximate solution for . We will carry out a reduction argument in section 3 and the main results will be proved in section 4. We put some basic estimates in the appendix.
Approximate solutions
======================
In the section, we will construct approximate solutions for .
Let $R>0$ be a large constant, such that for any $x\in {\Omega}$, ${\Omega}\subset\subset B_R(x)$. Consider the following problem:
$$\label{2.1}
\begin{cases}
-\delta^2\Delta w=( w-a)_+^{p},& \text{in}\; B_R(0),\\
w=0, &\text{on}\;\partial B_R(0),
\end{cases}$$
where $a>0$ is a constant. Then, has a unique solution $W_{{\delta},a}$, which can be written as
$$\label{2.2}
W_{{\delta},a}(x)=
\begin{cases}
a+{\delta}^{2/(p-1)}s_{\delta}^{-2/(p-1)}\phi\bigl(\frac{|x|}{s_{\delta}}\bigr), & |x|\le s_{\delta},\\
a\ln\frac {|x|} R/\ln \frac {s_{\delta}}R, & s_{\delta}\le |x|\le R,
\end{cases}$$
where $\phi(x)=\phi(|x|)$ is the unique solution of $$-\Delta \phi=\phi^{p},\quad\phi>0,~~\phi\in H_0^1\bigl(B_1(0)\bigr)$$ and $s_{\delta}\in (0,R)$ satisfies $${\delta}^{2/(p-1)}s_{\delta}^{-2/(p-1)}\phi^\prime(1)=\frac{a}{\ln(s_{\delta}/R)},$$ which implies $$\frac{s_{\delta}}{{\delta}|\ln{\delta}|^{(p-1)/2}}\rightarrow\left(\frac{|\phi^\prime(1)|}{a}\right)^{(p-1)/2}>0,\quad\text{as}~~{\delta}\rightarrow0.$$ Moreover, by Pohozaev identity, we can get that $$\int_{B_1(0)}\phi^{p+1}=\frac{\pi(p+1)
}{2}|\phi^\prime(1)|^2~~\text{and}~~\int_{B_1(0)}\phi^{p}=2\pi|\phi^\prime(1)|.$$
For any $z \in {\Omega}$, define $W_{{\delta},z,a}(x)=W_{{\delta},a}(x-z)$. Because $W_{{\delta},z,a}$ does not vanish on $\partial\Omega$, we need to make a projection. Let $PW_{{\delta},z,a}$ be the solution of
$$\begin{cases}
-{\delta}^2 \Delta w=( W_{{\delta},z,a}-a)_+^{p},& \text{in } \; {\Omega},\\
w=0, &\text{on}\; \partial {\Omega}.
\end{cases}$$ Then
$$\label{2.3}
PW_{{\delta},z,a}= W_{{\delta},z,a}-\frac a{\ln \frac{R}{s_{\delta}}} g(x,z),$$
where $g(x,z)$ satisfies
$$\begin{cases}
- \Delta g=0,& \text{in } \; {\Omega},\\
g=\ln\frac{R}{|x-z|}, &\text{on}\; \partial {\Omega}.
\end{cases}$$ It is easy to see that
$$g(x,z)=\ln R +2\pi h(x,z),$$ where $h(x,z)=-H(x,z)$.
We will construct solutions for of the form
$$\sum_{j=1}^m PW_{{\delta},z_{j},a_{{\delta},j}} +\omega_{\delta},$$ where $z_j\in\Omega$ for $j=1,\cdots,m$, $\omega_{\delta}$ is a perturbation term. To obtain a good estimate for $\omega_{\delta}$, we need to choose $a_{{\delta},j}$ properly.
By , we have
$$\label{2.4}
\begin{split}
&-{\delta}^2 \Delta \sum_{j=1}^m PW_{{\delta},z_j,a_{{\delta},j}}
-\sum_{j=1}^m\chi_{{\Omega}_j}\left(\sum_{i=1}^mPW_{{\delta},z_i,a_{{\delta},i}}-\kappa_j-\frac{2\pi q}{|\ln{\varepsilon}|}\right)_+^{p}\\
=& \sum_{j=1}^m \left(
W_{{\delta},z_j,a_{{\delta},j}}-a_{{\delta},j}\right)_+^{p}-
\sum_{j=1}^m\chi_{{\Omega}_j}\left(\sum_{i=1}^m
\left(W_{{\delta},z_i,a_{{\delta},i}}-\frac{a_{{\delta},i}}{\ln\frac R{s_{\delta}} }
g(y,z_i)\right) -\kappa_j-\frac{2\pi q}{|\ln{\varepsilon}|}\right)_+^{p}.
\end{split}$$
Denote $Z=(z_1,\cdots,z_m)\in \mathbb{R}^{2m}$. In this paper, we always assume that $z_j\in{\Omega}$ satisfies
$$\label{2.5}
d(z_j,\partial{\Omega})\ge \varrho>0,\quad |z_i-z_j|\ge \varrho^{\bar
L},\quad i, j=1,\cdots,m,\; i\ne j,$$
where $\varrho>0$ is a fixed small constant and $\bar L>0$ is a fixed large constant.
\[l2.1\] For ${\delta}>0$ small, there exist $(s_{{\delta},1}(Z),\cdots,s_{{\delta},m}(Z))$ and $(a_{{\delta},1}(Z),\cdots,a_{{\delta},m}(Z))$ satisfying the following system $$\label{2.6}
{\delta}^{2/(p-1)}s_i^{-2/(p-1)}\phi^\prime(1)=\frac{a_{i}}{\ln(s_i/R)},\qquad
i=1,\cdots,m$$ and $$\label{2.7}
a_{i}= \kappa_i + \frac{2\pi
q(z_i)}{|\ln{\varepsilon}|}+\frac{g(z_i,z_i)}{\ln\frac R {s_{i}} } a_{i}
-\sum_{j\ne i}^m \frac{\bar G(z_i,z_j)}{\ln\frac R{s_{j}} }a_{j}
,\qquad i=1,\cdots, m,$$ where for $i=1,\cdots,m$, $$\bar G(x,z_i)=\ln \frac R{|x-z_i|} -g(x,z_i).$$
We will show that system - has a solution $(s_1,\cdots,s_m, a_1,\cdots,a_m)$ in $\mathcal{D}\doteq[\frac{{\delta}}{|\ln{\delta}|},{\delta}|\ln{\delta}|\,]^m\times
\Pi_{i=1}^m[\frac{\kappa_i}{2},\,\frac{3\kappa_i}{2}]$. It is easy to see, for fixed $(s_1,\cdots,s_m)$ with $0<{\delta}<{\delta}^*$ small, that has a solution $(a_1,\cdots,a_m)$ depending on $(s_1,\cdots,s_m)$, such that $\frac{\kappa_i}{2}\leq
a_i\leq\frac{3\kappa_i}{2}$. For such $(a_1,\cdots,a_m)$ and for $i=1,\cdots, m$ define $$\theta_i(s_1,\cdots,s_m)=\frac{s_i^\frac{2}{p-1}}{\ln\frac{R}{s_i}}+\frac{\phi^\prime(1)}{a_i}{\delta}^\frac{2}{p-1},$$ then it is easy to verify that $$\left\{
\begin{array}{ll}
\theta_i(s_1,\cdots,s_m)>0,& s_i={\delta}|\ln{\delta}|, s_j\in [\frac{{\delta}}{|\ln{\delta}|},\,\,{\delta}|\ln{\delta}|\,] \,\,{\rm for}\,\, j=1,\cdots,m, j\neq i, \\
\theta_i(s_1,\cdots,s_m)<0,& s_i=\frac{{\delta}}{|\ln{\delta}|},\,\,s_j\in
[\frac{{\delta}}{|\ln{\delta}|},\,\,{\delta}|\ln{\delta}|\,] \,\,\,\,\,{\rm for}\,\,
j=1,\cdots,m, j\neq i.
\end{array}
\right.$$
By the Poincaré-Miranda Theorem in [@K; @M], we can get $(s_{{\delta},1},\cdots,s_{{\delta},m})$ such that $\theta_i(s_{{\delta},1},\cdots,s_{{\delta}, m})=0$. Therefore we have completed our proof of Lemma \[l2.1\].
For simplicity, for given $Z=(z_1,\cdots,z_m)$, in this paper, we will use $a_{{\delta},i}$,$s_{{\delta},i}$ instead of $a_{{\delta},i}(Z)$,$s_{{\delta},i}(Z)$.
\[remark2.2\] More precisely, we have the following relation $$\label{r2.2.1}
\frac{1}{\ln\frac{R}{s_{{\delta},i}}}=\frac{1}{\ln\frac{R}{{\varepsilon}}}
+O\left(\frac{\ln|\ln{\varepsilon}|}{|\ln{\varepsilon}|^2}\right),\, i=1,\cdots,m,$$
$$\label{r2.2.2}
a_{{\delta},i}=\kappa_i+\frac{2\pi
q(z_i)}{|\ln{\varepsilon}|}+\frac{g(z_i,z_i)}{\ln\frac{R}{{\varepsilon}}}-\sum_{j\neq
i}^m\frac {\bar G(z_i,z_j)}{\ln\frac{R}{{\varepsilon}}}+
O\Bigl(\frac{\ln|\ln{\varepsilon}|}{|\ln{\varepsilon}|^2}\Bigr),\, i=1,\cdots,m,$$
$$\label{2.10}
\frac{\partial a_{{\delta},i}}{\partial
z_{j,h}}=O\left(\frac1{|\ln{\varepsilon}|}\right),\quad~~\frac{\partial
s_{{\delta},i}}{\partial
z_{j,h}}=O\left(\frac{{\varepsilon}}{|\ln{\varepsilon}|}\right),\,i,j=1,\cdots,m,\,h=1,2.$$
Indeed, can be deduced from (see [@DY], for example). can been deduced from and . Differentiating both sides of and with respect to $z_{j,\,h}$ we can get a linear system of $\frac{\partial a_{{\delta},i}}{\partial z_{j,h}}$ and $\frac{\partial s_{{\delta},i}}{\partial z_{j,h}}$, which will deduces .
From now on we will always choose $(a_{{\delta},1},\cdots,a_{{\delta},m})$ and $(s_{{\delta},1},\cdots,s_{{\delta},m})$ such that and are satisfied. For $(a_{{\delta},1},\cdots,a_{{\delta},m})$ and $(s_{{\delta},1},\cdots,s_{{\delta},m})$ chosen in such a way define $$\label{2.8}
P_{{\delta}, Z,j}=
PW_{{\delta},z_j,a_{{\delta},j}},\,\,P_{{\delta},Z}=\sum_{j=1}^mP_{{\delta}, Z,j}.$$
Then, we find that for $x\in
B_{L s_{{\delta},i}}(z_i)$, where $L>0$ is any fixed constant,
$$\begin{split}
& P_{{\delta}, Z,i}(x)-\kappa_i-\frac{2\pi q(x)}{|\ln{\varepsilon}|}= W_{{\delta},z_i,
a_{{\delta},i}}(x) -\frac {a_{{\delta},i}}{\ln \frac{R}{s_{{\delta},i} }}
g(x,z_i)-\kappa_i-\frac{2\pi q(x)}{|\ln{\varepsilon}|}
\\
=& W_{{\delta},z_i, a_{{\delta},i}}(x)- \kappa_i-\frac {a_{{\delta},i}}{\ln
\frac{R}{s_{ {\delta},i}}}
g(z_i,z_i)-\frac {a_{{\delta},i}}{\ln \frac{R}{s_{{\delta},i}}}\Bigl(
\left\langle D g(z_i,z_i), x-z_i\right\rangle +O( |x-z_i|^2)\Bigr)\\
\quad&-\frac{2\pi q(z_i)}{|\ln{\varepsilon}|}-\frac{2\pi}{|\ln{\varepsilon}|}\left(\left\langle Dq(z_i),x-z_i\right\rangle+O(|x-z_i|^2)\right)\\
=& W_{{\delta},z_i, a_{{\delta},i}}(x)- \kappa_i-\frac{2\pi q(z_i)}{|\ln{\varepsilon}|}-\frac{2\pi}{|\ln{\varepsilon}|}\left\langle
Dq(z_i),x-z_i\right\rangle\\
\quad&-\frac {a_{{\delta},i}}{\ln \frac{R}{s_{ {\delta},i}}}
g(z_i,z_i)-\frac {a_{{\delta},i}}{\ln \frac{R}{s_{{\delta},i}}}
\left\langle D g(z_i,z_i), x-z_i\right\rangle +O\left(\frac{s_{{\delta},i}^2}{|\ln{\varepsilon}|}\right),
\end{split}$$ and for $j\ne i$ and $x\in B_{Ls_{{\delta},i}}(z_i)$, by
$$\begin{split}
& P_{{\delta}, Z,j}(x)=W_{{\delta},
z_j,a_{{\delta},j}}(x)-\frac{a_{{\delta},j}}{\ln\frac R{s_{{\delta},j}} }
g(x,z_j)=
\frac{a_{{\delta},j}}{\ln\frac R{s_{{\delta},j}}
} \bar G(x,z_j)\\
=&
\frac{a_{{\delta},j}}{\ln\frac R{s_{{\delta},j}}
} \bar G(z_i,z_j)+\frac{a_{{\delta},j}}{\ln\frac R{s_{{\delta},j}} }
\left\langle D \bar G(z_i,z_j),x-z_i\right\rangle +
O\Bigl( \frac{s_{{\delta},i}^2 }{|\ln{\varepsilon}|
} \Bigr).
\end{split}$$ So, by using , we obtain
$$\label{2.9}
\begin{split}
& P_{{\delta}, Z}(x)-\kappa_i-\frac{2\pi q(x)}{|\ln{\varepsilon}|}\\
=&
W_{{\delta},z_i, a_{{\delta},i}}(x)- a_{{\delta},i} -\frac{2\pi}{|\ln{\varepsilon}|}\left\langle Dq(z_i),x-z_i\right\rangle-\frac {a_{{\delta},i}}{\ln \frac{R}{s_{{\delta},i}}}
\left\langle D g(z_i,z_i), x-z_i\right\rangle \\
& +\sum_{j\ne i}^m \frac{a_{{\delta},j} }{\ln\frac R{s_{{\delta},j}} }
\left\langle D \bar G(z_i,z_j),x-z_i\right\rangle
+O\left( \frac{s_{{\delta},i}^2 }{|\ln{\varepsilon}| } \right),\quad x\in B_{L
s_{{\delta},i}}(z_i).
\end{split}$$
We end this section by giving the following formula which can be obtained by direct computation and will be used in the next two sections.
$$\label{2.11}
\begin{array}{ll}
\displaystyle\frac{\partial W_{{\delta},z_i,a_{{\delta},i}}(x)}{\partial z_{i,h}}&\\
=\left\{
\begin{array}{ll}
\displaystyle\frac{1}{{\delta}}\Bigl(\frac{a_{{\delta},i}}{|\phi^\prime(1)||\ln\frac{R}{s_{{\delta},i}}|}\Bigr)^{(p+1)/2}
\phi^\prime\bigl(\frac{|x-z_i|}{s_{{\delta},i}}\bigr)
\frac{z_{i,h}-x_h}{|x-z_i|}+ O\left(\frac{1}{|\ln{\varepsilon}|}\right),
~~x\in B_{s_{{\delta},i}}(z_i),\\
\,\\
\displaystyle-\frac{a_{{\delta},i}}{\ln\frac{R}{s_{{\delta},i}}}\frac{z_{i,h}-x_h}{|x-z_i|^2}+O\left(\frac{1}{|\ln{\varepsilon}|}\right),
\qquad\qquad\qquad\qquad\qquad\quad x\in \Omega\setminus B_{s_{{\delta},i}}(z_i).
\end{array}
\right.\\
\end{array}$$
the reduction
=============
Let
$$w(x)=
\begin{cases}
\phi(|x|), &|x|\le 1,\\
\phi^\prime(1)\ln |x|, & |x|>1.
\end{cases}$$ Then $w\in C^1(\mathbb{R}^2)$. Since $\phi^\prime(1)<0$ and $\ln
|x|$ is harmonic for $|x|>1$, we see that $w$ satisfies
$$\label{3.1}
-\Delta w= w_+^{p}, \quad \text{in}\; \mathbb{R}^2.$$
Moreover, since $w_+$ is Lip-continuous, by the Schauder estimate, $w\in C^{2,\alpha}$ for any $\alpha\in (0,1)$.
Consider the following problem:
$$\label{3.2}
-\Delta v- pw_+^{p-1} v=0,\quad v\in L^\infty(\mathbb{R}^2),$$
It is easy to see that $\frac{\partial w}{\partial x_i}$, $i=1,2,$ is a solution of . Moreover, from Dancer and Yan [@DY], we know that $w$ is also non-degenerate, in the sense that the kernel of the operator $Lv:=-\Delta v- pw_+^{p-1} v,~~v\in
D^{1,2}(\mathbb{R}^2)$ is spanned by $\bigl\{\frac{\partial
w}{\partial x_1}, \frac{\partial w}{\partial x_2}\bigr\}$.
Recall that $Z=(z_1,\cdots,z_m)$, and $z_j\in{\Omega}$ satisfies
$$\label{3.3}
d(z_j,\partial{\Omega})\ge \varrho>0,\quad |z_i-z_j|\ge \varrho^{\bar
L},\qquad i\ne j,$$
where $\varrho>0$ is a fixed small constant, and $\bar L>0$ is a large constant.
Let $P_{{\delta},Z,j}$ be the function defined in . Set
$$F_{{\delta},Z}=\left\{u: u\in L^p({\Omega}),\; \int_{\Omega}\frac{\partial
P_{{\delta},Z,j}}{\partial z_{j,h}}u =0, j=1,\cdots,m,\;h=1,2 \right\},$$ and
$$E_{{\delta},Z}=\left\{u:\;u\in W^{2,p}({\Omega})\cap H_0^1({\Omega}), \int_{\Omega}\Delta\left( \frac{\partial
P_{{\delta},Z,j}}{\partial z_{j,h}}\right)u =0, j=1,\cdots,m,\; h=1,2
\right\}.$$
For any $u\in L^p({\Omega})$, define $Q_{\delta}u$ as follows: $$Q_{\delta}u= u-\sum_{j=1}^m \sum_{h=1}^2
b_{j,h}\left(-{\delta}^2\Delta\Bigl(\frac{\partial P_{{\delta},Z,j}}{\partial
z_{j,h}}\Bigr)
\right),$$ where the constants $b_{j,h}$, $j=1,\cdots,m$, $h=1, 2$, satisfy
$$\label{3.4}
\sum_{j=1}^m \sum_{h=1}^2 b_{j,h}\left(-{\delta}^2\int_{{\Omega}}\Delta\Bigl(
\frac{\partial P_{{\delta},Z,j}}{\partial z_{j, h}}\Bigr) \frac{\partial
P_{{\delta},Z,i}}{\partial z_{i,\bar h}}\right)=\int_{{\Omega}} u
\frac{\partial P_{{\delta},Z,i}}{\partial z_{i,\bar h}}.$$
Since $\int_{\Omega}\frac{\partial
P_{{\delta},Z,j}}{\partial z_{j,h}} Q_{\delta}u=0$, the operator $Q_{\delta}$ can be regarded as a projection from $L^p(\Omega)$ to $F_{{\delta},Z}$. In order to show that we can solve to obtain $b_{j,h}$, we just need the following estimate ( by and ):
$$\label{3.5}
\begin{split}
&-{\delta}^2\int_{{\Omega}}\Delta\Bigl( \frac{\partial P_{{\delta},Z,j}}{\partial
z_{j, h}}\Bigr)
\frac{\partial P_{{\delta},Z,i}}{\partial z_{i,\bar h}}\\
=&
p\int_{{\Omega}}\bigl(W_{{\delta},z_j,a_{{\delta},j}}-a_{{\delta},j}\bigr)_+^{p-1}\left(\frac{\partial
W_{{\delta},z_{j},a_{{\delta},j}}}{\partial z_{j, h}}-\frac{\partial
a_{{\delta},j}}{\partial z_{j,h}}\right) \frac{\partial
P_{{\delta},Z,i}}{\partial z_{i,\bar h}}
\\
=& \delta_{ij h\bar h}\frac
c{|\ln{\varepsilon}|^{p+1}}+O\left(\frac{{\varepsilon}}{|\ln{\varepsilon}|^{p+1}}\right),
\end{split}$$
where $c>0$ is a constant, $\delta_{ijh \bar h}=1$, if $i=j$ and $h=\bar h$; otherwise, $\delta_{ijh \bar h}=0$.
Set
$$L_{\delta}u= -{\delta}^2 \Delta u-\sum_{j=1}^m p\chi_{{\Omega}_j}\left(
P_{{\delta},Z}-\kappa_j-\frac{2\pi q(x)}{|\ln{\varepsilon}|}\right)_+^{p-1}u.$$
We have the following lemma.
\[l31\] There are constants $\rho_0>0$ and ${\delta}_0>0$, such that for any ${\delta}\in (0,{\delta}_0]$, $Z$ satisfying , $ u\in E_{{\delta},Z}$ with $Q_{\delta}L_{\delta}u =0$ in $\Omega\setminus \cup_{j=1}^m
B_{L s_{{\delta},j}}
(z_j)$ for some $L>0$ large, then
$$\|Q_{\delta}L_{\delta}u\|_{L^p({\Omega})} \ge
\frac{\rho_0{\delta}^{\frac{2}{p}}}{|\ln{\delta}|^{\frac{(p-1)^2}{p}}}
\|u\|_{L^\infty({\Omega})}.$$
Set $s_{n,j}=s_{{\delta}_n,j}$. We will use $\|\cdot\|_p, \|\cdot\|_\infty$ to denote $\|\cdot\|_{L^p({\Omega})}$ and $ \|\cdot\|_{L^\infty({\Omega})}$ respectively.
We argue by contradiction. Suppose that there are ${\delta}_n\to 0$, $Z_n$ satisfying and $u_n\in E_{{\delta}_n,Z_n}$ with $Q_{{\delta}_n}L_{{\delta}_n} u_n =0$ in $\Omega\setminus \cup_{j=1}^m
B_{L s_{n,j}}
(z_{j,n})$, $\|u_n\|_\infty =1$, such that $$\|Q_{{\delta}_n} L_{{\delta}_n } u_n\|_{p} \le
\frac{1}{n}\frac{{\delta}_n^{\frac{2}{p}}}{|\ln{\delta}_n|^{\frac{(p-1)^2}{p}}}.$$
Firstly, we estimate $b_{j,h,n}$ in the following formula:
$$\label{3.6}
Q_{{\delta}_n} L_{{\delta}_n } u_n= L_{{\delta}_n } u_n-\sum_{j=1}^m \sum_{h=1}^2
b_{j,h,n} \left(-{\delta}_n^2\Delta\frac{\partial
P_{{\delta}_n,Z_n,j}}{\partial z_{j,h}}\right).$$
For each fixed $i$, multiplying by $
\frac{\partial
P_{{\delta}_n,Z_n,i}}{\partial z_{i,\bar h}}$, noting that
$$\int_{\Omega}\bigl( Q_{{\delta}_n} L_{{\delta}_n } u_n\bigr)
\frac{\partial
P_{{\delta}_n,Z_n,i}}{\partial z_{i,\bar h}}=0,$$ we obtain
$$\begin{split}
& \int_{\Omega}u_n\, L_{{\delta}_n} \left(\frac{\partial
P_{{\delta}_n,Z_n,i}}{\partial z_{i,\bar h}} \right)= \int_{\Omega}\bigl(
L_{{\delta}_n} u_n\bigr)\, \frac{\partial
P_{{\delta}_n,Z_n,i}}{\partial z_{i,\bar h}} \\
&=\sum_{j=1}^m
\sum_{\bar h=1}^2 b_{j,h,n} \int_{\Omega}\left(-{\delta}_n^2\Delta\frac{\partial P_{{\delta}_n,Z_n,j}}{\partial
z_{j,h}}\right)
\frac{\partial P_{{\delta}_n,Z_n,i}}{\partial z_{i,\bar h}}\\
\end{split}$$ Using and Lemma \[al1\], we obtain
$$\begin{split}
&\int_{\Omega}u_n\, L_{{\delta}_n} \left(\frac{\partial
P_{{\delta}_n,Z_n,i}}{\partial z_{i,\bar h}}\right)\\
&=\int_{\Omega}\left(-{\delta}_n^2\Delta \left(\frac{\partial
P_{{\delta}_n,Z_n,i}}{\partial z_{i,\bar
h}}\right)-\sum_{j=1}^mp\chi_{{\Omega}_j}\left(
P_{{\delta}_n,Z_n}-\kappa_j-\frac{2\pi
q(x)}{|\ln{\varepsilon}_n|}\right)_+^{p-1}\frac{\partial
P_{{\delta}_n,Z_n,i}}{\partial z_{i,\bar h}}\right)u_n\\
&=p\int_{\Omega}\left(W_{{\delta}_n,z_{i,n},a_{{\delta}_n,i}}-a_{{\delta}_n,i}\right)_+^{p-1}\left(\frac{\partial
W_{{\delta}_n,z_{i,n},a_{{\delta}_n,i}}}{\partial z_{i,\bar h}}-\frac{\partial
a_{{\delta}_n,i}}{\partial z_{i,\bar h}}\right)u_n\\
&\quad-\sum_{j=1}^mp\int_{\Omega_j}\left(W_{{\delta}_n,z_{j,n},a_{{\delta}_n,j}}-a_{{\delta}_n,j}+O\left(\frac{s_{n,j}}{|\ln{\varepsilon}_n|}\right)\right)_+^{p-1}\frac{\partial
P_{{\delta}_n,Z_n,i}}{\partial
z_{i,\bar h}}u_n\\
&=O\left(\frac{{\varepsilon}_n^2}{|\ln{\varepsilon}_n|^p}\right).
\end{split}$$
Using , we find that $$b_{i,h,n}=O\left({\varepsilon}_n^2|\ln{\varepsilon}_n|\right).$$ Therefore, $$\begin{split}
&\sum_{j=1}^m\sum_{h=1}^2b_{j,h,n}\left(-{\delta}_n^2\Delta\frac{\partial
P_{{\delta}_n,Z_n,j}}{\partial z_{j,h}}\right)\\
&=p\sum_{j=1}^m\sum_{h=1}^2b_{j,h,n}\left(W_{{\delta}_n,z_{j,n},a_{{\delta}_n,j}}-a_{{\delta}_n,j}\right)_+^{p-1}
\left(\frac{\partial W_{{\delta}_n,z_{j,n},a_{{\delta}_n,j}}}{\partial
z_{j,h}}-\frac{\partial a_{{\delta}_n,j}}{\partial z_{j,h}}\right)\\
&=O\left(\sum_{j=1}^m\sum_{h=1}^2 \frac{{\varepsilon}_n^{\frac{2}{p}-1}|b_{j,h,n}|}{|\ln{\varepsilon}_n|^{p}}\right)\\
&=O\left(\frac{{\varepsilon}_n^{\frac{2}{p}+1}}{|\ln{\varepsilon}_n|^{p-1}}\right)\quad
\text{in}~~L^p({\Omega}).
\end{split}$$
Thus, we obtain
$$L_{{\delta}_n}u_n = Q_{{\delta}_n}L_{{\delta}_n}u_n
+O\left(\frac{{\varepsilon}_n^{\frac{2}{p}+1}}{|\ln{\varepsilon}_n|^{p-1}}\right)
=O\left(\frac 1n\frac{{\delta}_n^{\frac{2}{p}}}{|\ln{\delta}_n|^{\frac{(p-1)^2}{p}}}\right).$$
For any fixed $i$, define
$$\tilde u_{i,n} (y)= u_n(s_{n,i} y+z_{i,n}).$$
Let $$\tilde L_n u= -\Delta u
-\sum_{l=1}^mp\frac{s_{n,i}^2}{{\delta}_n^2}\chi_{{\Omega}_l}\left(
P_{{\delta}_n,Z_n}(s_{n,i}y+z_{i,n})-\kappa_l-\frac{2\pi
q}{|\ln{\varepsilon}_n|}\right)_+^{p-1}u,$$ Then
$$s_{n,i}^{\frac{2}{p}}\times\frac{{\delta}_n^2}{s_{n,i}^{2}}\|\tilde L_n
\tilde u_{i,n}\|_p= \| L_{{\delta}_n} u_n\|_p.$$
Noting that $$\left(\frac{{\delta}_n}{s_{n,i}}\right)^2=O\left(\frac{1}{|\ln{\delta}_n|^{p-1}}\right),$$ we find that $$L_{{\delta}_n}u_n=o\left(\frac{{\delta}_n^{\frac{2}{p}}}{|\ln{\delta}_n|^{\frac{(p-1)^2}{p}}}\right).$$ As a result,
$$\tilde L_{n } \tilde u_{i,n}
=o(1),\quad \text{in}\; L^p(\Omega_n),$$ where $\Omega_n=\bigl\{y: s_{n,i} y+z_{i,n}\in\Omega\bigr\}$.
Since $\|\tilde u_{i,n}\|_\infty=1$, by the regularity theory of elliptic equations, we may assume that
$$\tilde u_{i,n}\to u_i, \quad \text{in}\; C_{loc}^1(\mathbb{R}^2).$$
It is easy to see that $$\begin{split}
&\sum_{l=1}^m\frac{s_{n,i}^2}{{\delta}_n^2}\chi_{{\Omega}_l}\left(
P_{{\delta}_n.Z_n}(s_{n,i}y+z_{i,n})-\kappa_l-\frac{2\pi q}{|\ln{\varepsilon}_n|}\right)_+^{p-1}\\
&=\frac{s_{n,i}^2}{{\delta}_n^2}\left(W_{{\delta}_n,z_{i,n},a_{{\delta}_n,i}}-a_{{\delta}_n,i}
+O\left(\frac{s_{n,i}}{|\ln{\varepsilon}_n|}\right)\right)_+^{p-1}+o(1)\\
&\rightarrow w_+^{p-1}.
\end{split}$$ Then, by Lemma \[al1\], we find that $u_i$ satisfies
$$-\Delta u_i-pw_+^{p-1} u_i= 0.$$ Now from the Proposition 3.1 in [@DY], we have
$$\label{3.7}
u_i= c_1 \frac{\partial w}{\partial x_1}+ c_2 \frac{\partial
w}{\partial x_2}.$$
Since
$$\int_{\Omega}\Delta \bigl(\frac{\partial P_{{\delta}_n,Z_n,i}}{\partial z_{i,
h}}\bigr) u_n =0,$$ we find that
$$\int_{\mathbb{R}^2}\phi_+^{p-1} \frac{\partial \phi}{\partial z_h}
u_i =0,$$ which, together with , gives $u_i=0$. Thus,
$$\tilde u_{i,n} \to 0,\quad \text{in}\; C^1(B_{L}(0)),$$ for any $L>0$, which implies that $u_n=o(1)$ on $\partial
B_{Ls_{n.i}}(z_{i,n})$.
By assumption,
$$Q_{{\delta}_n} L_{{\delta}_n} u_n = 0,\quad\text{in}\; {\Omega}\setminus
\cup_{i=1}^k B_{L s_{n,i}}(z_{i,n}).$$
On the other hand, by Lemma \[al1\], for $j=1,\cdots,m$, we have
$$\left(P_{{\delta}_n,Z_n} -\kappa_j-\frac{2\pi q(x)}{|\ln{\varepsilon}_n|} \right)_+
=0, \quad x\in {\Omega}_j\setminus B_{L s_{n,j}}(z_{j,n}).$$ Thus, we find that
$$-\Delta u_n =0,\quad \text{in}~{\Omega}\setminus \cup_{i=1}^m B_{L
s_{n,i}}(z_{i,n}).$$ However, $u_n=0$ on $\partial{\Omega}$ and $u_n=o(1)$ on $\partial
B_{Ls_{n,i}}(z_{i,n})$, $i=1,\cdots,m$. So we have
$$u_n=o(1).$$ This is a contradiction.
\[p32\]
$Q_{\delta}L_{\delta}$ is one to one and onto from $E_{{\delta},Z}$ to $F_{{\delta},Z}$.
Suppose that $Q_{\delta}L_{\delta}u=0$. Then, by Lemma \[l31\], $u=0$. Thus, $Q_{\delta}L_{\delta}$ is one to one.
Next, we prove that $Q_{\delta}L_{\delta}$ is an onto map from $E_{{\delta},Z}$ to $F_{{\delta},Z}$.
Denote
$$\tilde E= \Bigl\{ u: u\in H_0^1(\Omega), \; \int_{\Omega}
D\frac{\partial P_{{\delta},Z,j}}{\partial z_{j,h} } Du=0,\;
j=1,\cdots,m, h=1, 2\Bigr\}.$$ Note that $E_{{\delta},Z}=\tilde E\cap W^{2,p}(\Omega)$.
For any $\tilde h\in F_{{\delta},Z}$, by the Riesz representation theorem, there is a unique $u\in H_0^1({\Omega})$, such that
$$\label{3.8}
{\delta}^2\int_{\Omega}Du D\varphi =\int_{\Omega}\tilde h \varphi,\quad \forall\;
\varphi\in H_0^1({\Omega}).$$
On the other hand, from $\tilde h\in F_{{\delta},Z}$, we find that $u\in
\tilde E$. Moreover, by the $L^p$-estimate, we deduce that $u\in
W^{2,p}(\Omega)$. As a result, $u\in E_{{\delta},Z}$. Thus, we see that $Q_{\delta}(-{\delta}^2\Delta )=-{\delta}^2\Delta$ is an one to one and onto map from $ E_{{\delta},Z}$ to $F_{{\delta},Z}$. On the other hand, $Q_{\delta}L_{\delta}u=h$ is equivalent to
$$\label{3.9}
u=p{\delta}^{-2} (-Q_{\delta}\Delta )^{-1}\left[
Q_{\delta}\left(\sum_{j=1}^m\chi_{{\Omega}_j} \left(
P_{{\delta},Z}-\kappa_j-\frac{2\pi q(x)}{|\ln{\varepsilon}|}\right)_+^{p-1}
u\right)\right]+{\delta}^{-2} (-Q_{\delta}\Delta )^{-1} h,\quad u\in
E_{{\delta},Z}.$$
It is easy to check that ${\delta}^{-2} (-Q_{\delta}\Delta )^{-1}\left[
Q_{\delta}\left(\sum_{j=1}^m\chi_{{\Omega}_j} \left(
P_{{\delta},Z}-\kappa_j-\frac{2\pi q(x)}{|\ln{\varepsilon}|}\right)_+^{p-1}
u\right)\right]$ is a compact operator in $E_{{\delta},Z}$. By the Fredholm alternative, is solvable if and only if
$$u=p{\delta}^{-2} (-Q_{\delta}\Delta )^{-1}\left[
Q_{\delta}\left(\sum_{j=1}^m\chi_{{\Omega}_j} \left(
P_{{\delta},Z}-\kappa_j-\frac{2\pi q(x)}{|\ln{\varepsilon}|}\right)_+^{p-1}
u\right)\right]$$ has trivial solution, which is true since $Q_{\delta}L_{\delta}$ is a one to one map. Thus the result follows.
Now consider the equation
$$\label{3.10}
Q_{\delta}L_{\delta}\omega=
Q_{\delta}l_{\delta}+ Q_{\delta}R_{\delta}(\omega),$$
where
$$\label{3.11}
l_{\delta}=\sum_{j=1}^m \chi_{{\Omega}_j}\left( P_{{\delta},Z}
-\kappa_j-\frac{2\pi q(x)}{|\ln{\varepsilon}|} \right)_+^{p}-\sum_{j=1}^m
\left(W_{{\delta},z_j,a_{{\delta},j}}-a_{{\delta},j}\right)_+^{p},$$
and
$$\label{3.12}
\begin{split}
R_{\delta}(\omega)=& \sum_{j=1}^m \chi_{{\Omega}_j}\left(P_{{\delta},Z}
-\kappa_j+\omega-\frac{2\pi q(x)}{|\ln{\varepsilon}|} \right)_+^{p}
-\sum_{j=1}^m \chi_{{\Omega}_j}
\left(P_{{\delta},Z}
-\kappa_j-\frac{2\pi q(x)}{|\ln{\varepsilon}|}\right)_+^{p} \\
&-\sum_{j=1}^m \chi_{{\Omega}_j}p\left(P_{{\delta},Z} -\kappa_j-\frac{2\pi
q(x)}{|\ln{\varepsilon}|}\right)_+^{p-1}\omega.
\end{split}$$
Using Proposition \[p32\], we can rewrite as
$$\label{3.13}
\omega =G_{\delta}\omega =: (Q_{\delta}L_{\delta})^{-1} Q_{\delta}\bigl(
l_{\delta}+ R_{\delta}(\omega)\bigr).$$
The next Proposition enables us to reduce the problem of finding a solution for to a finite dimensional problem.
\[p33\]
There is an ${\delta}_0>0$, such that for any ${\delta}\in (0,{\delta}_0]$ and $Z$ satisfying , has a unique solution $\omega_{\delta}\in
E_{{\delta},Z}$, with
$$\|\omega_{\delta}\|_\infty =O\Bigl({\delta}|\ln{\delta}|^{\frac{p-1}{2}}\Bigr).$$
It follows from Lemma \[al1\] that if $L$ is large enough, ${\delta}$ is small then
$$\left(P_{{\delta},Z}
-\kappa_j-\frac{2\pi q(x)}{|\ln{\varepsilon}|} \right)_+=0, \quad x\in
{\Omega}_j\setminus B_{Ls_{{\delta},j}}(z_j),j=1,\cdots,m.$$
Let
$$M= E_{{\delta},Z}\cap\Bigl\{ \|\omega\|_\infty\le
{\delta}|\ln{\delta}|^{\frac{p-1}{2}}\Big\}.$$ Then $M$ is complete under $L^\infty$ norm and $G_{\delta}$ is a map from $ E_{{\delta},Z}$ to $ E_{{\delta},Z}$. We will show that $G_{\delta}$ is a contraction map from $M$ to $M$.
Step 1. $G_{\delta}$ is a map from $M$ to $M$.
For any $\omega\in M$, similar to Lemma \[al1\], it is easy to prove that for large $L>0$, ${\delta}$ small
$$\label{3.14}
\left(P_{{\delta},Z}+\omega
-\kappa_j-\frac{2\pi q(x)}{|\ln{\varepsilon}|}\right)_+=0, \quad \text{in}\;
{\Omega}_j\setminus B_{Ls_{{\delta},j}}(z_j).$$
Note also that for any $u\in L^\infty({\Omega})$,
$$Q_{\delta}u= u\quad \text{in}\; {\Omega}\setminus \cup_{j=1}^m
B_{Ls_{{\delta},j}}(z_j).$$ Therefore, using Lemma \[al1\], and , we find that for any $\omega\in M$,
$$Q_{\delta}l_{\delta}+ Q_{\delta}R_{\delta}(\omega)= l_{\delta}+ R_{\delta}(\omega)
=0, \quad \text{in}\; {\Omega}\setminus \cup_{j=1}^m
B_{Ls_{{\delta},j}}(z_j).$$ So, we can apply Lemma \[l31\] to obtain
$$\| (Q_{\delta}L_{\delta})^{-1} \bigl(
Q_{\delta}l_{\delta}+ Q_{\delta}R_{\delta}(\omega)\bigr)\|_\infty
\le \frac{C|\ln{\delta}|^{\frac{(p-1)^2}{p}}}{{\delta}^{\frac{2}{p}}} \| Q_{\delta}l_{\delta}+ Q_{\delta}R_{\delta}(\omega)\|_p.$$
Thus, for any $\omega\in M$, we have
$$\label{3.15}
\begin{split}
\| G_{\delta}(\omega)\|_\infty =& \| (Q_{\delta}L_{\delta})^{-1} Q_{\delta}\bigl(
l_{\delta}+ R_{\delta}(\omega)\bigr)\|_\infty\\
\le & \frac{C|\ln{\delta}|^{\frac{(p-1)^2}{p}}}{{\delta}^{\frac{2}{p}}}\|Q_{\delta}\bigl( l_{\delta}+ R_{\delta}(\omega)\bigr)\|_p.
\end{split}$$
It follows from – that the constant $b_{j,h}$, corresponding to $u\in L^\infty({\Omega})$, satisfies
$$|b_{j,h}| \le C |\ln{\delta}|^{p+1}\sum_{i,\,\bar h}
\int_{\Omega}\Bigl|\frac{\partial P_{{\delta},Z,i}}
{\partial z_{i,\bar h}}\Bigr| |u|.$$
Since
$$l_{\delta}+ R_{\delta}(\omega)=0, \quad \text{in}\; {\Omega}\setminus \cup_{j=1}^m
B_{Ls_{{\delta},j}}(z_j),$$ we find that the constant $b_{j,h}$, corresponding to $l_{\delta}+ R_{\delta}(\omega)$ satisfies
$$\begin{split}
|b_{j,h}|\le & C|\ln{\delta}|^{p+1} \sum_{i,\,\bar h} \left(\sum_{j=1}^m
\int_{B_{Ls_{{\delta},j} }(z_j)}
\Bigl|\frac{\partial P_{{\delta},Z,i}}
{\partial z_{i,\bar h}}\Bigr||l_{\delta}+ R_{\delta}(\omega)
|\right)\\
\le & C{\varepsilon}^{1-\frac{2}{p}}|\ln{\varepsilon}|^{p}\|l_{\delta}+ R_{\delta}(\omega) \|_p.
\end{split}$$ As a result,
$$\begin{split}
&\|Q_{\delta}(l_{\delta}+ R_{\delta}(\omega))\|_p \\
\le & \| l_{\delta}+ R_{\delta}(\omega)\|_p+C \sum_{j,\,h} |b_{j,h}| \left\|
-{\delta}^2\Delta\Bigl(\frac{\partial P_{{\delta},Z,j}}{\partial
z_{j,h}}\Bigr)
\right\|_p\\
\le & C \| l_{\delta}\|_p+ C\| R_{\delta}(\omega) \|_p.
\end{split}$$
On the other hand, from Lemma \[al1\] and , we can deduce
$$\begin{split}
\|l_{\delta}\|_p =&\left\|\sum_{j=1}^m\chi_{{\Omega}_j}\left(
P_{{\delta},Z}-\kappa_j-\frac{2\pi q(x)}{|\ln{\varepsilon}|}\right)_+^{p}-\sum_{j=1}^m\bigl(W_{{\delta},z_j,a_{{\delta},j}}-a_{{\delta},j}\bigr)_+^{p}\right\|_p\\
\le&\sum_{j=1}^m\frac{Cs_{{\delta},j}}{|\ln{\varepsilon}|}\Big\|\bigl(W_{{\delta},z_j,a_{{\delta},j}}-a_{{\delta},j}\bigr)_+^{p-1}\Big\|_p\\
=&O\left(\frac{{\delta}^{1+\frac{2}{p}}}{|\ln{\delta}|^{\frac
{p-1}2+\frac{1}p}}\right).
\end{split}$$
For the estimate of $\| R_{\delta}(\omega) \|_p$, we have $$\label{3.16}
\begin{split}
\| R_{\delta}(\omega) \|_p=&\bigg\|\sum_{j=1}^m\chi_{{\Omega}_j}\Big(
P_{{\delta},Z}-\kappa_j+\omega-\frac{2\pi
q(x)}{|\ln{\varepsilon}|}\Big)_+^{p}-\sum_{j=1}^m\chi_{{\Omega}_j}\Big(
P_{{\delta},Z}-\kappa_j-\frac{2\pi q(x)}{|\ln{\varepsilon}|}\Big)_+^{p}\\
&-\sum_{j=1}^mp\chi_{{\Omega}_j}\Big(
P_{{\delta},Z}-\kappa_j-\frac{2\pi q(x)}{|\ln{\varepsilon}|}\Big)_+^{p-1}\omega\bigg\|_p\\
\le&\sum_{j=1}^mC\|\omega\|_\infty^2\left\|\chi_{{\Omega}_j}\left(
P_{{\delta},Z}-\kappa_j-\frac{2\pi q(x)}{|\ln{\varepsilon}|}\right)_+^{p-2}\right\|_p\\
=&O\left(\frac{{\delta}^{\frac{2}{p}}\|\omega\|_\infty^2}{|\ln{\delta}|^{p-3+\frac{1}p}}\right).
\end{split}$$
Thus, we obtain
$$\label{3.17}
\begin{split}
\| G_{\delta}(\omega)\|_\infty \le &
\frac{C|\ln{\delta}|^{\frac{(p-1)^2}{p}}}{{\delta}^{\frac{2}{p}}}\Bigl(\|
l_{\delta}\|_p+\| R_{\delta}(\omega)\|_p
\Bigr)\\
\le & C|\ln{\delta}|^{\frac{(p-1)^2}{p}}\left(\frac{{\delta}}{|\ln{\delta}|^{\frac
{p-1}2+\frac{1}p}}+\frac{\|\omega\|_\infty^2}{|\ln{\delta}|^{p-3+\frac{1}p}}\right)\\
\le&{\delta}|\ln{\delta}|^{\frac{p-1}2}
\end{split}$$
Thus, $G_{\delta}$ is a map from $M$ to $M$.
Step 2. $G_{\delta}$ is a contraction map.
In fact, for any $\omega_i\in M$, $i=1,2 $, we have
$$G_{\delta}\omega_1-G_{\delta}\omega_2= (Q_{\delta}L_{\delta})^{-1} Q_{\delta}\bigl(
R_{\delta}(\omega_1)-R_{\delta}(\omega_2)\bigr).$$ Noting that
$$R_{\delta}(\omega_1)=R_{\delta}(\omega_2)=0,\quad\text{in}\; {\Omega}\setminus
\cup_{j=1}^m B_{Ls_{{\delta},j}}(z_j),$$ we can deduce as in Step 1 that $$\begin{split}
\|G_{\delta}\omega_1-G_{\delta}\omega_2\|_\infty\le& \frac{C|\ln{\delta}|^{\frac{(p-1)^2}{p}}}{{\delta}^{\frac{2}{p}}}\|R_{\delta}(\omega_1)-R_{\delta}(\omega_2)\|_p\\
\le&C|\ln{\delta}|^{p-1}\left(\frac{\|\omega_1\|_\infty}{|\ln{\delta}|^{p-2}}+\frac{\|\omega_2\|_\infty}{|\ln{\delta}|^{p-2}}\right)\|\omega_1-\omega_2\|_\infty\\
\le&C{\delta}|\ln{\delta}|^{\frac {p+1}2}\|\omega_1-\omega_2\|_\infty \le\frac
12 \|\omega_1-\omega_2\|_\infty.
\end{split}$$
Combining Step 1 and Step 2, we have proved that $G_{\delta}$ is a contraction map from $M$ to $M$. By the contraction mapping theorem, there is an unique $\omega_{\delta}\in M$, such that $\omega_{\delta}= G_{\delta}\omega_{\delta}$. Moreover, it follows from that
$$\|\omega_{\delta}\|_\infty\le {\delta}|\ln{\delta}|^{\frac {p-1}2}.$$
Proof of The main results
=========================
In this section, we will choose $Z$, such that $ \sum_{j=1}^m
P_{{\delta},Z,j}+\omega_{\delta}$, where $\omega_{\delta}$ is the map obtained in Proposition \[p33\], is a solution of .
Define $$I(u)=\frac{{\delta}^2}{2}\int_\Omega
|Du|^2-\sum_{j=1}^m\frac{1}{p+1}\int_\Omega\chi_{{\Omega}_j}\left(u-\kappa_j-\frac{2\pi
q(x)}{|\ln{\varepsilon}|}\right)_+^{p+1}$$ and $$\label{K}
K(Z)= I\left( P_{{\delta},Z} +\omega_{\delta}\right).$$ It is well known that if $Z$ is a critical point of $K(Z)$, then $\sum_{j=1}^m P_{{\delta},Z,j} +\omega_{\delta}$ is a solution of .
In the following, we will prove that $K(Z)$ has a critical point.
\[l42\]
We have
$$K(Z)= I\left(\sum_{j=1}^m P_{{\delta},Z,j}
\right)+O\left(
\frac{{\varepsilon}^3}{|\ln{\varepsilon}|^p}\right).$$
Recall that
$$P_{{\delta},Z}=\sum_{j=1}^m P_{{\delta},Z,j}.$$ We have
$$\begin{split}
K(Z)=& I\bigl( P_{{\delta},Z}\bigr) +\int_{\Omega}{\delta}^2
DP_{{\delta},Z}D\omega_{\delta}+\frac{{\delta}^2}{2}\int_\Omega |D\omega_{\delta}|^2\\
&-\frac1{p+1} \sum_{j=1}^m\int_{\Omega}\chi_{{\Omega}_j}\Biggl[ \biggl(
P_{{\delta},Z} +\omega_{\delta}-\kappa_j-\frac{2\pi q(x)}{|\ln{\varepsilon}|}
\biggr)_+^{p+1}- \biggl( P_{{\delta},Z}-\kappa_j-\frac{2\pi
q(x)}{|\ln{\varepsilon}|} \biggr)_+^{p+1}\Biggr].
\end{split}$$
Using Proposition \[p33\] and , we find
$$\begin{split}
& \int_{{\Omega}_j} \Biggl[ \biggl( P_{{\delta},Z}
+\omega_{\delta}-\kappa_j-\frac{2\pi q(x)}{|\ln{\varepsilon}|} \biggr)_+^{p+1}-
\biggl(
P_{{\delta},Z}-\kappa_j-\frac{2\pi q(x)}{|\ln{\varepsilon}|} \biggr)_+^{p+1}\Biggr]\\
= &\int_{ B_{Ls_{{\delta},j}}(z_j)} \Biggl[ \biggl( P_{{\delta},Z}
+\omega_{\delta}-\kappa_j-\frac{2\pi q(x)}{|\ln{\varepsilon}|} \biggr)_+^{p+1}-
\biggl(
P_{{\delta},Z}-\kappa_j-\frac{2\pi q(x)}{|\ln{\varepsilon}|} \biggr)_+^{p+1}\Biggr]\\
=&O\left(\frac{s_{{\delta},j}^2\|\omega_{\delta}\|_\infty}{|\ln{\varepsilon}|^{p}}\right)
= O\Bigl(
\frac{{\varepsilon}^3}{|\ln{\varepsilon}|^{p}}\Bigr).
\end{split}$$
On the other hand,
$$\begin{split}
&{\delta}^2 \int_{\Omega}DP_{{\delta},Z}D\omega_{\delta}=
\sum_{j=1}^m\int_{\Omega}\left(W_{{\delta},z_j,a_{{\delta},j}}-a_{{\delta},j}\right)_+^{p} \omega_{\delta}\\
=&\sum_{j=1}^m
\int_{\cup_{k=1}^m B_{s_{{\delta},k}}(z_k)}
(W_{{\delta},z_j,a_{{\delta},j}}-a_{{\delta},j})_+^{p} \omega_{\delta}\\
=& O\Bigl(
\frac{{\varepsilon}^3}{|\ln{\varepsilon}|^{p}}\Bigr).
\end{split}$$ Finally, we estimate ${\delta}^2\int_{\Omega}|D\omega_{\delta}|^2$.\
Note that $$\begin{split}
-{\delta}^2\Delta\omega_{\delta}=&\sum_{j=1}^m\chi_{{\Omega}_j}\left(P_{{\delta},Z}+\omega_{\delta}-\kappa_j-\frac{2\pi
q(x)}{|\ln{\varepsilon}|}\right)_+^{p}
-\sum_{j=1}^m\left(W_{{\delta},z_j,a_{{\delta},j}}-a_{{\delta},j}\right)_+^{p}\\
&+\sum_{j=1}^m\sum_{\bar h=1}^2b_{j,\bar
h}\left(-{\delta}^2\Delta\frac{\partial P_{{\delta},Z,j}}{\partial z_{j,\bar
h}}\right),
\end{split}$$ Hence, by , we have $$\begin{split}
{\delta}^2\int_{\Omega}|D\omega_{\delta}|^2=&\sum_{j=1}^m\int_{{\Omega}_j}\Biggl[\left(P_{{\delta},Z}+\omega_{\delta}-\kappa_j-\frac{2\pi
q(x)}{|\ln{\varepsilon}|}\right)_+^{p}
-\left(W_{{\delta},z_j,a_{{\delta},j}}-a_{{\delta},j}\right)_+^{p}\Biggr]\omega_{\delta}\\
&+\sum_{j=1}^m\sum_{\bar h=1}^2b_{j,\bar
h}\int_{\Omega}\left(-{\delta}^2\Delta\frac{\partial P_{{\delta},Z,j}}{\partial
z_{j,\bar h}}\right)\omega_{\delta}\\
=&p\sum_{j=1}^m\int_{{\Omega}_j}\left(W_{{\delta},z_j,a_{{\delta},j}}-a_{{\delta},j}\right)_+^{p-1}\left(\frac{s_{{\delta},j}}{|\ln{\varepsilon}|}+\omega_{\delta}\right)\omega_{\delta}+O\left(\sum_{j=1}^m\sum_{\bar h=1}^2\frac{{\varepsilon}|b_{j,\bar
h}|\|\omega_{\delta}\|_\infty}{|\ln{\varepsilon}|^{p}}\right)\\
=&O\left(\frac{{\varepsilon}^4}{|\ln{\varepsilon}|^{p-1}}\right).
\end{split}$$
So we can obtain that
$$K(Z)= I\left(\sum_{j=1}^m P_{{\delta},Z,j}
\right)+O\left(
\frac{{\varepsilon}^3}{|\ln{\varepsilon}|^{p}}\right).$$
\[l43\]
We have
$$\frac{\partial K(Z)}{\partial z_{i,h}}=\frac{\partial }{\partial
z_{i,h}}I\left(\sum_{j=1}^m P_{{\delta},Z,j} \right)+O\Bigl(
\frac{{\varepsilon}^3}{|\ln{\varepsilon}|^{p-1}}\Bigr).$$
First, we have
$$\label{4.4}
\begin{split}
&\frac{\partial K(Z)}{\partial z_{i,h}}=\left\langle I^\prime\Bigl(
P_{{\delta},Z} +\omega_{\delta}\Bigr),\frac{\partial P_{{\delta},Z}}{\partial
z_{i,h}}+ \frac{\partial
\omega_{\delta}}{\partial z_{i,h}}\right\rangle\\
=&\frac{\partial }{\partial z_{i,h}}I\Bigl(P_{{\delta},Z} \Bigr)
+\left\langle I^\prime\big( P_{{\delta},Z}+\omega_{\delta}\big),
\frac{\partial\omega_{\delta}}{\partial z_{i,h}}\right\rangle
\\
&-\sum_{j=1}^m \int_{\Omega_j}\Biggl[
\biggl(P_{{\delta},Z}+\omega_{\delta}-\kappa_j-\frac{2\pi
q(x)}{|\ln{\varepsilon}|}\biggr)_+^{p}-\biggl(P_{{\delta},Z}-\kappa_j-\frac{2\pi
q(x)}{|\ln{\varepsilon}|}\biggr)_+^{p}\Biggr]\frac{\partial
P_{{\delta},Z}}{\partial z_{i,h}}.
\end{split}$$
Since $\omega_{\delta}\in E_{{\delta},Z}$, we have $$\int_{\Omega}\left(W_{{\delta},z_j,a_{{\delta},j}}-a_{{\delta},j}\right)_+^{p-1}\left(\frac{\partial
W_{{\delta},z_j,a_{{\delta},j}}}{\partial z_{j,h}}-\frac{\partial
a_{{\delta},j}}{\partial z_{j,h}}\right)\omega_{\delta}=0.$$ Differentiating the above relation with respect to $z_{i, h}$, we can deduce
$$\begin{split}
\Bigg\langle & I^\prime\bigl(P_{{\delta},Z} +\omega_{\delta}\bigr),
\frac{\partial \omega_{\delta}}{\partial z_{i, h}}
\Bigg\rangle=\sum_{j=1}^m\sum_{\bar h=1}^2 b_{j,\bar h}
\int_{\Omega}\left(-{\delta}^2\Delta\frac{\partial P_{{\delta},Z,j}}{\partial
z_{j,\bar h}}\right) \frac{\partial \omega_{\delta}}{\partial z_{i, h}}\\
=&\sum_{j=1}^m\sum_{\bar h=1}^2 p b_{j,\bar h}
\int_{\Omega}\left(W_{{\delta},z_j,a_{{\delta},j}}-a_{{\delta},j}\right)_+^{p-1}\left(\frac{\partial
W_{{\delta},z_j,a_{{\delta},j}}}{\partial z_{j,\bar h}}-\frac{\partial
a_{{\delta},j}}{\partial z_{j,\bar h}}\right) \frac{\partial
\omega_{\delta}}{\partial z_{i, h}}\\
=&O\left(\sum_{j=1}^m\sum_{\bar h=1}^2 \frac{{\varepsilon}|b_{j,\bar
h}|}{|\ln{\varepsilon}|^{p}}\right)=O\left(\frac{{\varepsilon}^3}{|\ln{\varepsilon}|^{p-1}}\right).
\end{split}$$
On the other hand, using (for the definition of $R_{\delta}(\omega)$, see ), we obtain
$$\begin{split}
&\sum_{j=1}^m\int_{\Omega_j}\Biggl[
\biggl(P_{{\delta},Z}+\omega_{\delta}-\kappa_j-\frac{2\pi
q(x)}{|\ln{\varepsilon}|}\biggr)_+^{p}-\biggl(P_{{\delta},Z}-\kappa_j-\frac{2\pi
q(x)}{|\ln{\varepsilon}|}\biggr)_+^{p}\Biggr]\frac{\partial
P_{{\delta},Z,i}}{\partial z_{i,h}}\\
= & \sum_{j=1}^m\int_{\Omega_j}\Biggl[
\biggl(P_{{\delta},Z}+\omega_{\delta}-\kappa_j-\frac{2\pi q(x)}{|\ln{\varepsilon}|}\biggr)_+^{p}-\biggl(P_{{\delta},Z}-\kappa_j-\frac{2\pi q(x)}{|\ln{\varepsilon}|}\biggr)_+^{p}\\
&-p\biggl(P_{{\delta},Z}-\kappa_j-\frac{2\pi
q(x)}{\kappa|\ln{\varepsilon}|}\biggr)_+^{p-1}\omega_{\delta}\Biggr]\frac{\partial
P_{{\delta},Z,i}}{\partial z_{i,h}} +\sum_{j=1}^m
p\int_{\Omega_j}\Biggl[\biggl(P_{{\delta},Z}-\kappa_j-\frac{2\pi q(x)}{|\ln{\varepsilon}|}\biggr)_+^{p-1}\\
&-\bigl(W_{{\delta},z_j,a_{{\delta},j}}-a_{{\delta},j}\bigr)_+^{p-1}\Biggr]\frac{\partial
P_{{\delta},Z,i}}{\partial z_{i,h}}\omega_{\delta}+O\left(\frac{s_{{\delta},j}^2\|\omega_{\delta}\|_\infty}{|\ln{\varepsilon}|^{p}}\right)\\
=&\int_\Omega R_{\delta}(\omega_{\delta})\frac{\partial P_{{\delta},Z,i}}{\partial
z_{i,h}} +
\sum_{j=1}^mp\int_{\Omega_j}\Biggl[\biggl(P_{{\delta},Z}-\kappa_j-\frac{2\pi
q(x)}{|\ln{\varepsilon}|}\biggr)_+^{p-1}-\bigl(W_{{\delta},z_i,a_{{\delta},i}}-a_{{\delta},i}\bigr)_+^{p-1}\Biggr]\frac{\partial
P_{{\delta},Z,i}}{\partial
z_{i,h}}\omega_{\delta}\\
&+O\left(\frac{{\varepsilon}^3}{|\ln{\varepsilon}|^{p}}\right)\\
=& O\left(\frac{{\varepsilon}^3}{|\ln{\varepsilon}|^{p-1}}\right).
\end{split}$$ Thus, the estimate follows.
Define
$$c_{{\delta},1}=\frac{ C {\delta}^2
}{\ln \frac{R}{{\varepsilon}}
}-\frac{ \pi{\delta}^2\ln\frac1{\varrho}
}{|\ln\frac{R}{{\varepsilon}}|^2
},\quad
c_{{\delta},2}=\frac{(C+\eta) {\delta}^2
}{\ln\frac{R}{{\varepsilon}}
},$$ where $\eta>0$ is a small constant and $\varrho>0$ is a fixed small constant. Let
$$D=\bigl\{ Z=(z_1,\cdots,z_m): \; z_i\in\Omega_\varrho,
i=1,\cdots,m,\; |z_i-z_j|\ge \varrho^{\bar L}, i\ne j\bigr\},$$ where $\Omega_\varrho=\bigl\{y: y\in\Omega,\; d(y,\partial\Omega)\ge
\varrho\bigr\}$, and $\bar L>0$ is a large constant.
Denote $K^c=\bigl\{Z: Z\in D,\; K(Z)\le c\bigr\}$. Consider
$$\left\{
\begin{array}{ll}
\frac{d Z(t)}{dt}=- DK(Z(t)), \; t\ge 0,&\\ \vspace{0.05cm}\\
Z(0)\in K^{c_{{\delta},2}}.&
\end{array}
\right.$$
\[l44\] $Z(t)$ does not leave $D$ before it reaches $K^{c_{{\delta},1}}$.
Note that $$\label{4.5}
h(x,z)=\frac{1}{2\pi}\ln\frac{1}{|x-\bar{z}|}+o(1),\quad
\frac{\partial h(x,z)}{\partial
n}=-\frac{1}{2\pi|x-\bar{z}|}\left\langle\frac{x-\bar{z}}{|x-\bar{z}|},n\right\rangle+o(1),$$ if $z$ is close to $\partial\Omega$, where $n$ is the outward normal unit vector of $\partial\{x: x\in\Omega, d(x,
\partial\Omega)\leq d(z, \partial\Omega)\}$ and $\bar{z}$ is the reflection point of $z$ with respect to $\partial\Omega$.
Suppose that there is $t_0>0$, such that $Z(t_0)=:(z_1,\cdots,z_m)\in
\partial D$.
1. Suppose that there are $i, j\in\{1,2,\cdots,m\}$, such that $i\neq j$ and $|z_i-z_j|=\varrho^{\bar
L}$.
Since $d(\bar{z},\partial\Omega)\geq \varrho$ and $\bar{z}\notin\Omega$, using , we get $|h(z_j,z_j)|\le
C^\prime\ln\frac1{\varrho}$ for any $i, j$, where $C^\prime>0$. Thus, we have
$$\bar G(z_i,z_j)\ge
\ln\frac1{|z_i-z_j|}-C^\prime\ln\frac1{\varrho}\ge \bar
L\ln\frac1{\varrho}-C^\prime\ln\frac1{\varrho}.$$ Then, by Lemma \[l42\] and Proposition \[ap1\], we have $$K(Z)\le \frac{C {\delta}^2
}{\ln \frac{R}{{\varepsilon}}
}+ \frac{ mC^\prime {\delta}^2\ln\frac1\varrho
}{(\ln\frac{R}{{\varepsilon}})^2}-\frac{ \bar L {\delta}^2\ln\frac1\varrho\
}{(\ln\frac{R}{{\varepsilon}})^2}<c_{{\delta},1},$$ if $\bar L>0$ is large.\
2. Suppose that there is $i$, such that $z_i\in\partial\Omega_\varrho$.
Let $n$ be the outward unit normal of $\partial\Omega_\varrho$ at $z_i$. We have
$$\frac{\partial \bar G(z_j,z_i)}{\partial
n}=-\frac{1}{|z_j-z_i|}\left\langle
\frac{z_i-z_j}{|z_i-z_j|},n\right\rangle-\frac{\partial
g(z_j,z_i)}{\partial n},$$ where $n$ is the outward normal unit vector of $\partial\Omega_\varrho$ at $z_i$.
On the other hand, if $z_j\in\Omega_\varrho$, $j\ne i$, satisfies
$$\left\langle \frac{z_i-z_j}{|z_i-z_j|},n\right\rangle<0,$$ then,
$$\left\langle \frac{z_i-z_j}{|z_i-z_j|},n\right\rangle= O(|z_i-z_j|).$$ So, we obtain
$$\left\langle \frac{z_i-z_j}{|z_i-z_j|},n\right\rangle\ge
-C|z_i-z_j|, \quad\forall\; j\ne i.$$ As a result, by Lemma \[l43\] and Proposition \[ap2\], we have
$$\begin{split}
\frac{\partial K}{\partial n}\ge &
\frac{4\pi^2{\delta}^2\kappa_i}{|\ln{\varepsilon}||\ln\frac{R}{{\varepsilon}}|}\frac{\partial
q(z_i)}{\partial n}+
\frac{2\pi{\delta}^2\kappa_i^2}{(\ln\frac{R}{{\varepsilon}})^2}\frac{\partial
g(z_i,z_i)}{\partial n}\\
&+\sum_{j\ne i}^m \frac{2\pi{\delta}^2\kappa_i\kappa_j}{(\ln\frac{R}{{\varepsilon}})^2}\frac{\partial g(z_j,z_i)}{\partial
n}-\frac{C
{\delta}^2}{(\ln\frac{R}{{\varepsilon}})^2}.
\end{split}$$ On the other hand, we derive from
$$\frac{\partial g(z_i,z_i)}{\partial n}=\frac{1+o(1)}{2\varrho},$$ and
$$\frac{\partial g(z_j,z_i)}{\partial
n}=\frac{1+o(1)}{|\bar z_i-z_j|}\bigl\langle \frac{\bar z_i-z_j}{|\bar
z_i-z_j|},n\bigr\rangle,$$ where $\bar z_i$ is the reflection point of $z_i$ with respect to $\partial\Omega$.
It is easy to check that if $|z_j-z_i|\le M\varrho$, where $M>0$ is a fixed large constant, then
$$\left\langle \frac{\bar z_i-z_j}{|\bar
z_i-z_j|},n\right\rangle\ge 0.$$ So
$$\frac{\partial K}{\partial n}\ge \frac{2\pi
{\delta}^2}{(\ln\frac{R}{{\varepsilon}})^2}\Bigl(
\frac{\kappa_i^2+o(1)}{2\varrho}-\frac{\kappa_i\kappa_j+o(1)}{M\varrho}-C\Bigr)>0.$$
Therefore, the flow does not leave $D$.
We will prove that $K(Z)$ has a critical point in $K^{c_{{\delta},2}}\setminus K^{c_{{\delta},1}}$.
Suppose that $K(Z)$ has no critical point in $K^{c_{{\delta},2}}\setminus
K^{c_{{\delta},1}}$. Then from Lemma \[l44\] that $K^{c_{{\delta},1}}$ is a deformation retract of $K^{c_{{\delta},2}}$.
It is easy to see that $K^{c_{{\delta},2}}=D$ and
$$\bigl\{Z: Z\in D, |z_i-z_j|=\varrho^{\bar L}, \;\text{for some}\;
i\ne j\bigr\}\subset K^{c_{{\delta},1}}.$$
On the other hand, take $R$ large enough such that $\inf\limits_{\Omega}q\ge
-\sum_{i=1}^m\frac{(p-1)\kappa_i^2}{16m\pi\min_l\{\kappa_l\}}-\sum_{i=1}^m\frac{\kappa_i^2g(z_j,z_j)}{4m\pi\min_l\{\kappa_l\}}$, then $K(Z)\le c_{{\delta},1}$ implies that
$$-\sum_{j\ne i}^m \frac{\pi{\delta}^2\kappa_i\kappa_j \bar
G(z_j,z_i)}{(\ln\frac{R}{{\varepsilon}})^2}\le
-\frac{\pi{\delta}^2\ln\frac1{\varrho}}{|\ln{\varepsilon}||\ln\frac{R}{{\varepsilon}}|},$$ which implies that there are $i\ne j$, such that
$$\bar G(z_j,z_i)\ge c'\ln\frac1{\varrho},\quad\text{where}~~c'>0
~~\text{is a constant}.$$ So, there is a $\alpha>0$, independent of $\delta$, such that
$$|z_i-z_j|\le \varrho^\alpha.$$ Therefore,
$$\label{4.6}
\begin{split}
& \bigl\{Z: Z\in D, |z_i-z_j|=\varrho^{\bar L}, \;\text{for
some}\; i\ne
j\bigr\}\\
&\subset K^{c_{{\delta},1}}\subset \bigl\{Z: Z\in D, |z_i-z_j|\le
\varrho^\alpha, \;\text{for some}\; i\ne j\bigr\}.
\end{split}$$
Filling the hole $D^*=: \bigl\{Z: Z\in D, |z_i-z_j|=\varrho^{\bar
L},
\;\text{for some}\; i\ne
j\bigr\}$ in $D$, we obtain
$$\label{4.7}
\begin{split}
& \bigl\{Z: z_i\in\Omega_\varrho,
|z_i-z_j|\le \varrho^{\bar L}, \;\text{for some}\; i\ne
j\bigr\}\\
&\subset K^{c_{{\delta},1}}\cup D^* \subset \bigl\{Z:
z_i\in\Omega_\varrho,
|z_i-z_j|\le
\varrho^\alpha, \;\text{for some}\; i\ne j\bigr\}.
\end{split}$$
Since $K^{c_{{\delta},1}}$ is a deformation retract of $K^{c_{{\delta},2}}$, we find that $K^{c_{{\delta},1}}\cup D^*$ is a deformation retract of $K^{c_{{\delta},2}} \cup D^*$. On the other hand, $\bigl\{Z: z_i\in\Omega_\varrho, z_i=z_j, \;\text{for some}\; i\ne
j\bigr\}$ is a deformation retract of $\bigl\{Z:
z_i\in\Omega_\varrho,
|z_i-z_j|\le
\varrho^\alpha, \;\text{for some}\; i\ne j\bigr\}$ if $\varrho>0$ is small. Using , we see that
$$\bigl\{Z: z_i\in\Omega_\varrho, z_i=z_j, \;\text{for some}\; i\ne
j\bigr\}$$ is a deformation retract of $$\underbrace{\Omega_\varrho\times\cdots\times \Omega_\varrho}_m=
K^{c_{{\delta},2}} \cup D^*.$$ This is impossible if $\Omega$ has nontrivial homology.
Thus we get a solution $w_{\delta}$ for . Let $u_{\varepsilon}=\frac{|\ln{\varepsilon}|}{2\pi}w_{\delta},~{\delta}={\varepsilon}\left(\frac{|\ln{\varepsilon}|}{2\pi}\right)^{\frac{1-p}{2}}$, it is not difficult to check that $u_{\varepsilon}$ has all the properties listed in Theorem \[th1\] and thus the proof of Theorem \[th1\] is complete.
\[r44\] In the proof of Theorem \[th1\], what we actually need is that the following function $$\Phi(Z)=\sum_{i=1}^m4\pi^2\kappa_iq(z_i) +\sum_{i=1}^m
\pi\kappa_i^2g(z_i,z_i)-\sum_{j\neq
i}^m\pi\kappa_i\kappa_j\bar{G}(z_j,z_i)$$ as well as its small perturbation (in a suitable sense) has a critical point in $D$. Moreover, using the estimates as in Lemma \[l43\], it is easy to see that if $\sum_{j=1}^mP_{{\delta},Z_{{\delta}},j}(x)+\omega_{\delta}$ is a solution of , and $Z_{\delta}\rightarrow Z_0$ as ${\delta}\rightarrow0$, then $Z_0$ is a critical point of $\Phi(Z)$.
Note that the Kirchhoff–Routh function associated to the vortex dynamics is $$\mathcal{W}(x_1,\cdots,x_m)=\frac{1}{2}\sum_{i\neq
j}^m\kappa_i\kappa _jG(x_i,x_j)+\frac{1}{2}\sum^{m}_{i=1}\kappa^2
_iH(x_i,x_i)+\sum^{m}_{i=1}\kappa_i\psi_0(x_i).$$
Recall that $h(z_i,z_j)=-H(z_i,z_j)$, it is easy to check that $$\Phi(Z)=-4\pi^2\mathcal {W}(Z)+\sum_{i=1}^m\pi\kappa_i^2\ln R.$$ Hence, $\Phi(Z)$ and $\mathcal{W}(Z)$ possess the same critical points.
By Lemma \[l42\], \[l43\] and Proposition \[ap1\], \[ap2\], we have $$K(Z)= \frac{ C{\delta}^2
}{\ln\frac{R}{{\varepsilon}}
} +\sum_{i=1}^m\frac{\pi(p-1){\delta}^2\kappa_i^2}{4(\ln\frac{R}{{\varepsilon}})^2}+\frac{{\delta}^2}{|\ln{\varepsilon}|^2}\Phi(Z)
+ O\left(\frac{{\delta}^2\ln|\ln{\varepsilon}|}{|\ln{\varepsilon}|^3} \right)$$ and $$\frac{\partial K(Z)}{\partial
z_{i,h}}=\frac{{\delta}^2}{|\ln{\varepsilon}|^2}\frac{\partial \Phi(Z)}{\partial
z_{i,h}}+O\left(\frac{{\delta}^2\ln|\ln{\varepsilon}|}{|\ln{\varepsilon}|^3} \right).$$ Thus, stable critical point of Kirchhoff-Routh function $\mathcal{W}(Z)$ implies that $K(Z)$ has a critical point. So the result follows.
Now we are in the position to prove Theorem \[nth1\].
By Theorem \[th3\], we obtain that $u_{\varepsilon}$ is a solution to .
Set $$\mathbf{v}_{\varepsilon}=(\nabla
(u_{\varepsilon}-q))^\bot,~\omega_{\varepsilon}=\nabla\times\mathbf{v}_{\varepsilon},$$ $$P_{\varepsilon}=\sum_{j=1}^m\frac{1}{p+1}\chi_{{\Omega}_j}\left(u_{\varepsilon}-q-\frac{\kappa_j|\ln{\varepsilon}|}{2\pi}\right)_+^{p+1}-\frac{1}{2}|\nabla
(u_{\varepsilon}-q)|^2.$$ then $(\mathbf{v}_{\varepsilon}, P_{\varepsilon})$ forms a stationary solution for problem .
We now just need to verify $$\int_\Omega
\omega_{\varepsilon}\rightarrow \sum_{j=1}^m\kappa_j,
~~\text{as}~~{\varepsilon}\rightarrow0.$$
By direct calculations, we find that $$\begin{split}
\int_\Omega\omega_{\varepsilon}&=\sum_{j=1}^m\frac{1}{{\varepsilon}^2}\int_\Omega\chi_{{\Omega}_j}
\left(u_{\varepsilon}-q-\frac{\kappa_j|\ln{\varepsilon}|}{2\pi}\right)_+^p\\
&=\sum_{j=1}^m\frac{|\ln{\varepsilon}|^p}{(2\pi)^p {\varepsilon}^2}\int_{\Omega_j}
\left(w_{\delta}-\kappa_j-\frac{2\pi q}{|\ln{\varepsilon}|}\right)^p_+\\
&=\sum_{j=1}^m\frac{|\ln{\varepsilon}|^p}{(2\pi)^p{\varepsilon}^2}\int_{B_{Ls_{{\delta},j}(z_j)}}
\left(W_{{\delta},z_j,a_{{\delta},j}}-a_{{\delta},j}+O\Big(\frac{s_{{\delta},j}}{|\ln{\varepsilon}|}\Big)\right)^p_+\\
&=\sum_{j=1}^m\frac{ s_{{\delta},j}^2|\ln{\varepsilon}|^p}{(2\pi)^p{\varepsilon}^2}\left(\frac{{\delta}}{s_{{\delta},j}}\right)^{\frac{2p}{p-1}}\int_{B_1(0)}\phi^p+o(1)\\
&=\sum_{j=1}^m\frac{ a_{{\delta},j}
|\ln{\varepsilon}|}{\ln\frac{R}{s_{{\delta},j}}}+o(1)\\
&\rightarrow \sum_{j=1}^m\kappa_j, \quad \text{as}~~{\varepsilon}\rightarrow0.
\end{split}$$ Therefore, the result follows.
To regularize point vortices with equi-strength $\kappa$, we do not need $\chi_{{\Omega}_j}$, that is, we just need to consider the following problem $$\begin{cases}
-{\varepsilon}^2 \Delta u=(u-q-\frac{\kappa}{2\pi}\ln\frac{1}{{\varepsilon}})_+^p, \quad & x\in\Omega, \\
u=0, \quad & x\in\partial\Omega.
\end{cases}$$
[**Acknowledgements:**]{} D. Cao and Z. Liu were supported by the National Center for Mathematics and Interdisciplinary Sciences, CAS. D. Cao and J. Wei were also supported by CAS Croucher Joint Laboratories Funding Scheme.
Energy expansion
================
In this section we will give precise expansions of $I\left(\sum_{j=1}^m P_{{\delta},Z,j}\right)$ and $\frac{\partial
}{\partial z_{i,h}} I\left(\sum_{j=1}^m P_{{\delta},Z,j}\right)$, which have been used in section 4.
We always assume that
$$d(z_j,\partial{\Omega})\ge \varrho>0,~~~|z_i-z_j|\ge \varrho^{\bar
L},\quad i\ne j,$$ for some small $\varrho>0$ and large $\bar L>0$.
\[al1\] For $x\in{\Omega}_i,~i=1,\cdots,m$, we have $$P_{{\delta},Z}(x)>\kappa_i+\frac{2\pi q(x)}{|\ln{\varepsilon}|},\quad x\in
B_{s_{{\delta},i}(1-Ts_{{\delta},i})}(z_i),\;$$ where $T>0$ is a large constant; while
$$P_{{\delta},Z}(x)<\kappa_i+\frac{2\pi q(x)}{|\ln{\varepsilon}|},\quad
x\in{\Omega}_i\setminus B_{s_{{\delta},i}(1+s_{{\delta},i}^{\sigma})}(z_i),$$ where $\sigma>0$ is a small constant.
Suppose that $x\in B_{s_{{\delta},i}(1-Ts_{{\delta},i})}(z_i)$. It follows from and $\phi_1'(s)<0$ that $$\begin{split}
& P_{{\delta},Z}(x)-\kappa_i-\frac{2\pi q(x)}{|\ln{\varepsilon}|}= W_{{\delta},z_i,a_{{\delta},i}}(x)-a_{{\delta},i}+O\left(\frac{s_{{\delta},i}}{|\ln{\varepsilon}|}\right)\\
=&
\frac{a_{{\delta},i}}{|\phi'(1)||\ln\frac{R}{s_{{\delta},i}}|}\phi\Bigl(\frac{|x-z_i|}{s_{{\delta},i}}\Bigr)+O\Bigl(\frac{{\varepsilon}}{|\ln{\varepsilon}|}\Bigr)>0,
\end{split}$$ if $T>0$ is large. On the other hand, if $x\in{\Omega}_i\setminus
B_{s_{{\delta},i}^{\tilde\sigma}}(z_i)$, where $\tilde\sigma>\sigma>0$ is a fixed small constant, then $$\begin{split}
& P_{{\delta},Z}(x)-\kappa_i-\frac{2\pi q(x)}{|\ln{\varepsilon}|}= \sum_{j=1}^m
a_{{\delta},j}\ln\frac R{|x-z_j|}/\ln\frac R{s_{{\delta},j}} -\kappa_i
-\frac{2\pi q(x)}{|\ln{\varepsilon}|}+o(1)\\
\le & C\tilde \sigma -\kappa_i+o(1)<0.
\end{split}$$
Finally, if $x\in B_{s_{{\delta},i}^{\tilde\sigma}}(z_i)\setminus
B_{s_{{\delta},i}(1+Ts_{{\delta},i}^{\tilde \sigma})}(z_i)$ for some $i$, then $$\begin{split}
& P_{{\delta},Z}(x)-\kappa_i-\frac{2\pi q(x)}{|\ln{\varepsilon}|}= W_{{\delta},z_i,a_{{\delta},i}}(x)-a_{{\delta},i}+O\left(\frac{s_{{\delta},i}^{\tilde\sigma}}{\ln\frac{R}{s_{{\delta},i}}}\right)\\
=& a_{{\delta},i}\frac{\ln\frac
R{|x-z_i|}}{\ln\frac{R}{s_{{\delta},i}}}-a_{{\delta},i}+O\left(\frac{s_{{\delta},i}^{\tilde\sigma}}{\ln\frac{R}{s_{{\delta},i}}}\right)\\
\le & -a_{{\delta},i}\frac{\ln(1+Ts_{{\delta},i}^{\tilde\sigma})}{\ln\frac
{R}{s_{{\delta},i}}}+O\left(\frac{s_{{\delta},i}^{\tilde\sigma}}{\ln\frac{R}{s_{{\delta},i}}}\right)
<0,
\end{split}$$ if $T>0$ is large. Note that by the choice of $\tilde\sigma$, $B_{s_{{\delta},i}(1+s_{{\delta},i}^\sigma)}(z_i)\supset
B_{s_{{\delta},i}(1+Ts_{{\delta},i}^{\tilde\sigma})}(z_i)$ for small ${\delta}$. We therefore derive our conclusion.
\[ap1\]
We have
$$\begin{split}
I\left( P_{{\delta},Z}\right)=&\frac{ C{\delta}^2
}{\ln\frac{R}{{\varepsilon}}
} +\sum_{i=1}^m\frac{\pi(p-1){\delta}^2\kappa_i^2}{4(\ln\frac{R}{{\varepsilon}})^2}
+\sum_{i=1}^m\frac{4\pi^2{\delta}^2\kappa_iq(z_i)}{|\ln{\varepsilon}||\ln\frac{R}{{\varepsilon}}|}+
\sum_{i=1}^m \frac{\pi{\delta}^2\kappa_i^2 g(z_i,z_i)}{(\ln\frac{R}{\varepsilon})^2}
\\
&-\sum_{j\ne i}^m \frac{\pi{\delta}^2\kappa_i\kappa_j \bar G(z_j,z_i)}{{(\ln\frac{R}{\varepsilon})^2}}
+ O\left(\frac{{\delta}^2\ln|\ln{\varepsilon}|}{|\ln{\varepsilon}|^3} \right),
\end{split}$$ where $C$ is a positive constant.
Taking advantage of , we have
$$\begin{split}
&{\delta}^2\int_{\Omega}\big|D P_{{\delta},Z}\big|^2= \sum_{j=1}^m \sum_{i=1}^m
\int_{\Omega}\bigl(W_{{\delta},z_j,a_{{\delta},j} }-a_{{\delta},j}\bigr)_+^{p}
P_{{\delta},Z,i}\\
=& \sum_{j=1}^m \sum_{i=1}^m\int_{B_{s_{{\delta},j}}(z_j)}
\left(W_{{\delta},z_j,a_{{\delta},j}}-a_{{\delta},j}\right)_+^{p} \left(
W_{{\delta},z_i,a_{{\delta},i}}-\frac{a_{{\delta},i}}{\ln\frac R{s_{{\delta},i}}}
g(x,z_i)\right).
\end{split}$$
First, we estimate
$$\begin{split}
&\int_{B_{s_{{\delta},i}} (z_i)}
\left(W_{{\delta},z_i,a_{{\delta},i}}-a_{{\delta},i}\right)_+^{p}\left(
W_{{\delta},z_i,a_{{\delta},i}}-\frac{a_{{\delta},i}}{\ln\frac R{s_{{\delta},i}}}
g(x,z_i)
\right)\\
=&\int_{B_{s_{{\delta},i}} (z_i)}
\bigl(W_{{\delta},z_i,a_{{\delta},i}}-a_{{\delta},i}\bigr)^{p+1}+a_{{\delta},i}
\int_{B_{s_{{\delta},i}} (z_i)}
\bigl(W_{{\delta},z_i,a_{{\delta},i}}-a_{{\delta},i}\bigr)^{p}\\
&-\frac{a_{{\delta},i}}{\ln\frac R{s_{{\delta},i}}}\int_{B_{s_{{\delta},i}} (z_i)}
\bigl(W_{{\delta},z_i,a_{{\delta},i}}-a_{{\delta},i}\bigr)^{p}
g(x,z_i)\\
=&\Bigl(\frac{{\delta}}{s_{{\delta},i}}\Bigr)^{\frac{2(p+1)}{p-1}}s_{{\delta},i}^2\int_{B_1(0)}\phi^{p+1}
+a_{{\delta},i}\Bigl(\frac{{\delta}}{s_{{\delta},i}}\Bigr)^{\frac{2p}{p-1}}s_{{\delta},i}^2\int_{B_1(0)}\phi^{p}\\
&-\frac{a_{{\delta},i}}{\ln\frac
R{s_{{\delta},i}}}\Bigl(\frac{{\delta}}{s_{{\delta},i}}\Bigr)^{\frac{2p}{p-1}}g(z_i,z_i)s_{{\delta},i}^2\int_{B_1(0)}\phi^{p}+O\left(
\frac{s_{{\delta},i}^3}{|\ln{\varepsilon}|^{p+1}}\right)\\
=&\frac{\pi(p+1)}{2}\frac{{\delta}^2a_{{\delta},i}^2}{(\ln\frac{R}{s_{{\delta},i}})^2}+\frac{2\pi
{\delta}^2a_{{\delta},i}^2}{\ln\frac{R}{s_{{\delta},i}}}-\frac{2\pi{\delta}^2
a_{{\delta},i}^2}{(\ln\frac{R}{s_{{\delta},i}})^2}g(z_i,z_i)+O\left(
\frac{{\varepsilon}^3}{|\ln{\varepsilon}|^{p+1}}\right).
\end{split}$$
Next, for $j\ne i$,
$$\begin{split}
&\int_{B_{s_{{\delta},j}} (z_j)}
\bigl(W_{{\delta},z_j,a_{{\delta},j}}-a_{{\delta},j}\bigr)_+^{p} \left(
W_{{\delta},z_i,a_{{\delta},i}}-\frac{a_{{\delta},i}}{\ln\frac R{s_{{\delta},i}}}
g(x,z_i)
\right)\\
=&
\Bigl(\frac{{\delta}}{s_{{\delta},j}}\Bigr)^{\frac{2p}{p-1}}\frac{a_{{\delta},i}}{\ln\frac{R}{s_{{\delta},i}}}
\int_{B_{s_{{\delta},j}}(z_j)}\phi^{p}\Bigl(\frac{|x-z_j|}{s_{{\delta},j}}\Bigr)\bar
G(x,z_i)
\\
=&
\Bigl(\frac{{\delta}}{s_{{\delta},j}}\Bigr)^{\frac{2p}{p-1}}\frac{a_{{\delta},i}s_{{\delta},j}^2}{\ln\frac{R}{s_{{\delta},i}}}
\int_{B_{1}(0)}\phi^{p}(|x|)\bar G(z_j+s_{{\delta},j}x,z_i)
\\
=&\Bigl(\frac{{\delta}}{s_{{\delta},j}}\Bigr)^{\frac{2p}{p-1}}\frac{a_{{\delta},i}s_{{\delta},j}^2}{\ln\frac{R}{s_{{\delta},i}}}
\bar G(z_j,z_i)
\int_{B_{1}(0)}\phi^{p}+O\left(\frac{s_{{\delta},j}^3}{|\ln{\varepsilon}|^{p+1}}\right)\\
=&\frac{2\pi{\delta}^2
a_{{\delta},i}a_{{\delta},j}}{|\ln\frac{R}{s_{{\delta},i}}||\ln\frac{R}{s_{{\delta},j}}|}\bar{G}(z_i,z_j)+O\left(
\frac{{\varepsilon}^3}{|\ln{\varepsilon}|^{p+1}}\right).
\end{split}$$
By Lemma \[al1\] and ,
$$\begin{split}
&\sum_{j=1}^m\int_{{\Omega}_j} \left( P_{{\delta},Z} -\kappa_j-\frac{2\pi
q(x)}{|\ln{\varepsilon}|}\right)_+^{p+1}= \sum_{j=1}^m
\int_{B_{Ls_{{\delta},j}}(z_j)}\left( P_{{\delta},Z} -\kappa_j-\frac{2\pi
q(x)}{\kappa|\ln{\varepsilon}|}\right)_+^{p+1}
\\
=& \sum_{j=1}^m \int_{B_{Ls_{{\delta},j}}(z_j)}
\left(W_{{\delta},z_j,a_{{\delta},j}} -a_{{\delta},j} +O\bigg(\frac
{s_{{\delta},j}}{|\ln{\varepsilon}|}
\bigg)\right)_+^{p+1}\\
=&\sum_{j=1}^m\left(\frac{{\delta}}{s_{{\delta},j}}\right)^{\frac{2(p+1)}{p-1}}\int_{B_{s_{{\delta},j}}(z_j)}\phi^{p+1}\Bigl(\frac{|x-z_j|}{s_{{\delta},j}}\Bigr)
+O\left(\frac{s_{{\delta},j}^3}{|\ln{\varepsilon}|^{p+1}}\right)\\
=&\sum_{j=1}^m\left(\frac{{\delta}}{s_{{\delta},j}}\right)^{\frac{2(p+1)}{p-1}}s_{{\delta},j}^2\int_{B_{1}(0)}\phi^{p+1}
+O\left(\frac{s_{{\delta},j}^3}{|\ln{\varepsilon}|^{p+1}}\right)\\
=&\sum_{j=1}^m\frac{\pi(p+1)}{2}\frac{{\delta}^2a_{{\delta},j}^2}{(\ln\frac{R}{s_{{\delta},j}})^2}+O\left(
\frac{{\varepsilon}^3}{|\ln{\varepsilon}|^{p+1}}\right).
\end{split}$$ So, we have proved
$$\begin{split}
I\left(\sum_{i=1}^m P_{{\delta},Z,j}\right)=&\sum_{i=1}^m\left(\frac{\pi
(p+1)}{4} \frac{ {\delta}^2
a_{{\delta},i}^2}{|\ln \frac{R}{s_{{\delta},i}}|^2
} +\frac{\pi{\delta}^2 a^2_{{\delta},i}}{|\ln \frac{R}{s_{{\delta},i}}|}-\frac{\pi g(z_i,z_i){\delta}^2a_{{\delta},i}^2}{|\ln\frac{R}{s_{{\delta},i}}|^2}\right)
\\
&+\sum_{j\ne i}^m \frac{\pi\bar G(z_j,z_i){\delta}^2 a_{{\delta},i}
a_{{\delta},j}}{|\ln\frac{R}{s_{{\delta},i}}||\ln \frac{R}{s_{{\delta},j}}|}-\frac{\pi{\delta}^2}{2}\left(\sum_{j=1}^m \frac{a_{{\delta},j}^2}{|\ln \frac{R}{s_{{\delta},j}}|^2}\right)
+ O\left(\frac{{\varepsilon}^3}{|\ln{\varepsilon}|^{p+1}} \right).
\end{split}$$
Thus, the result follows from Remark \[remark2.2\].
\[ap2\]
We have
$$\begin{split}
\frac{\partial }{\partial z_{i,h}} I\left(
P_{{\delta},Z}\right)=&\frac{4\pi^2{\delta}^2\kappa_i}{|\ln{\varepsilon}||\ln\frac{R}{{\varepsilon}}|}\frac{\partial
q(z_i)}{\partial z_{i,h}}
+\frac{2\pi{\delta}^2\kappa_i^2}{(\ln\frac{R}{{\varepsilon}})^2} \frac{\partial
g(z_i,z_i)}{\partial z_{i,h}}
-\sum_{j\ne i}^m \frac{2\pi{\delta}^2\kappa_i\kappa_j}{(\ln\frac{R}{{\varepsilon}})^2}\frac{\partial \bar G(z_j,z_i)}{\partial z_{i,h}} \\
&+ O\left(\frac{{\delta}^2\ln|\ln{\varepsilon}|}{|\ln{\varepsilon}|^3} \right).
\end{split}$$
Direct computation yields that
$$\begin{split}
\frac{\partial }{\partial z_{i,h}} I\left( P_{{\delta},Z}\right)=&
{\delta}^2\int_{{\Omega}} DP_{{\delta},Z}D \frac{\partial P_{{\delta},Z}}{\partial
z_{i,h}}-\sum_{j=1}^m\int_{{\Omega}_j} \left( P_{{\delta},Z}-\kappa_j
-\frac{2\pi q(x)}{|\ln{\varepsilon}|}\right)_+^{p}\frac{\partial
P_{{\delta},Z}}{\partial z_{i,h}}
\\
=& \sum_{l=1}^m\sum_{j=1}^m \int_{B_{Ls_{{\delta},j}}(z_j)}
\left[ \left(W_{{\delta},z_j,a_{{\delta},j}}-a_{{\delta},j}\right)_+^{p}-
\left(
P_{{\delta},Z}-\kappa_j -\frac{2\pi
q(x)}{|\ln{\varepsilon}|}\right)_+^{p}\right]\frac{\partial
P_{{\delta},Z,l}}{\partial z_{i,h}} .
\end{split}$$
Using , Lemma \[al1\] and Remark \[remark2.2\], we find that
$$\begin{split}
&\int_{B_{Ls_{{\delta},i}}(z_i)}
\left[ \left(W_{{\delta},z_i,a_{{\delta},i}}-a_{{\delta},i}\right)_+^{p}-
\left(
P_{{\delta},Z}-\kappa_i-\frac{2\pi q(x)}{|\ln{\varepsilon}|} \right)_+^{p}\right]\frac{\partial P_{{\delta},Z,i}}{\partial z_{i,h}}\\
=&\int_{B_{s_{{\delta},i}(1+s_{{\delta},i}^{\sigma})}(z_i)}
\left[ \bigl(W_{{\delta},z_i,a_{{\delta},i}}-a_{{\delta},i}\bigr)_+^{p}-
\left(
P_{{\delta},Z}-\kappa_i-\frac{2\pi q(x)}{|\ln{\varepsilon}|} \right)_+^{p}\right]\frac{\partial P_{{\delta},Z,i}}{\partial z_{i,h}}\\
=
&p\int_{B_{s_{{\delta},i}}(z_i)}\bigl(W_{{\delta},z_i,a_{{\delta},i}}-a_{{\delta},i}\bigr)_+^{p-1}\bigg[
\frac{2\pi}{|\ln{\varepsilon}|}\bigl\langle D q(z_i), x-z_i\bigr\rangle+\frac{a_{{\delta},i}}{\ln\frac{R}{s_{{\delta},i}}}\bigl\langle D g(z_i,z_i), x-z_i\bigr\rangle \\
\quad&- \sum_{j\ne i}^m
\frac{a_{{\delta},j}}{\ln\frac{R}{s_{{\delta},j}}}\bigl\langle D \bar
G(z_i,z_j),x-z_i\bigr\rangle \bigg]\frac{\partial
P_{{\delta},Z,i}}{\partial z_{i,h}}+O\Bigl(\frac{{\varepsilon}^{2+\sigma}}{|\ln{\varepsilon}|^{p+1}}\Bigr)\\
=&-\frac{p{\delta}^2a_{{\delta},i}}{|\phi'(1)||\ln\frac{R}{s_{{\delta},i}}|}\bigg(\frac{2\pi}{|\ln{\varepsilon}|}\frac{\partial
q(z_i)}{\partial
z_{i,h}}+\frac{a_{{\delta},i}}{\ln\frac{R}{s_{{\delta},i}}}\frac{\partial
g(z_i,z_i)}{\partial z_{i,h}}-\sum_{j\neq
i}^m\frac{a_{{\delta},j}}{\ln\frac{R}{s_{{\delta},j}}}\frac{\partial \bar
G(z_i,z_j)}{\partial z_{i,h}}\bigg)\\
&\times \int_{B_1(0)}\phi^{p-1}(|x|)\phi^\prime(|x|)\frac{x_h^2}{|x|}
+O\Bigl(\frac{{\varepsilon}^{2+\sigma}}{|\ln{\varepsilon}|^{p+1}}\Bigr)\\
=&\frac{4\pi^2 {\delta}^2a_{{\delta},i}}{|\ln{\varepsilon}||\ln\frac{R}{s_{{\delta},i}}|}\frac{\partial q(z_i)}{\partial z_{i,h}}
+\frac{2\pi {\delta}^2a_{{\delta},i}^2}{(\ln\frac{R}{s_{{\delta},i}})^2}\frac{\partial g(z_i,z_i)}{\partial z_{i,h}}
-\sum_{j\neq i}\frac{2\pi {\delta}^2a_{{\delta},i}a_{{\delta},j}}{|\ln\frac{R}{s_{{\delta},j}}||\ln\frac{R}{s_{{\delta},i}}|}\frac{\partial \bar G(z_i,z_j)}{\partial z_{i,h}}
+O\Bigl(\frac{{\varepsilon}^{2+\sigma}}{|\ln{\varepsilon}|^{p+1}}\Bigr),
\end{split}$$ since
$$\int_{B_1(0)}\phi^{p-1}(|x|)\phi^\prime(|x|)\frac{x_h^2}{|x|}=
-\frac{2\pi}{p}|\phi^\prime(1)|.$$
On the other hand, for $l\ne i$, from , we have
$$\begin{split}
&\int_{B_{Ls_{{\delta},j}}(z_j)}
\left[ \bigl(W_{{\delta},z_j,a_{{\delta},j}}-a_{{\delta},j}\bigr)_+^{p}-
\left(
P_{{\delta},Z}-\kappa_j-\frac{2\pi q(x)}{|\ln{\varepsilon}|} \right)_+^{p}\right]\frac{\partial P_{{\delta},Z,l}}{\partial z_{i,h}}\\
= &\int_{B_{Ls_{{\delta},j}}(z_j)}\left[\left(W_{{\delta},z_j,a_{{\delta},j}}-
a_{{\delta},j}\right)^{p-1}\frac{s_{{\delta},j}}{|\ln{\varepsilon}|}\right]\times \frac{C}{\ln\frac{R}{s_{{\delta},l}}}\\
=& O\left(\frac{{\varepsilon}^3}{|\ln{\varepsilon}|^{p+1}}
\right).
\end{split}$$ Thus, the result follows.
[99]{}
A. Ambrosetti and M. Struwe, Existence of steady vortex rings in an ideal fluid, *Arch. Rational Mech. Anal.*, 108(1989), 97–109.
A. Ambrosetti and J. Yang, Asymptotic behaviour in planar vortex theory, *Atti Accad. Naz. Lincei Cl. Sci. Fis. Mat. Natur. Rend. Lincei (9) Mat. Appl.*, 1(1990), 285–291.
V.I. Arnold and B.A. Khesin, Topological methods in hydrodynamics, *Applied Mathematical Sciences, Vol. 125.* Springer, New York, 1998.
T.V. Badiani, Existence of steady symmetric vortex pairs on a planar domain with an obstacle, *Math. Proc. Cambridge Philos. Soc.*, 123(1998), 365–384.
T.Bartsch, A.Pistoia and T.Weth, N-vortex equilibria for ideal fluids in bounded planar domains and new nodal solutions of the sinh-Poisson and the Lane-Emden-Fowler equations, *Comm.Math.Phys.*, 297(2010), 653–686.
M.S. Berger and L.E. Fraenkel, Nonlinear desingularization in certain free-boundary problems, *Comm. Math. Phys.*, 77(1980), 149–172.
G.R. Burton, Vortex rings in a cylinder and rearrangements, *J. Diff. Equat.*, 70(1987), 333–348.
L. Caffarelli and A. Friedman, Asymptotic estimates for the plasma problem, *Duke Math. J.*, 47(1980), 705–742.
D.Cao and T.Küpper, On the existence of multi-peaked solutions to a semilinear Neumann problem, *Duke Math. J.*, 97(1999), 261–300.
D. Cao, S. Peng and S. Yan, Multiplicity of solutions for the plasma problem in two dimensions, *Adv. Math.*, 225(2010), 2741–2785.
E.N. Dancer and S. Yan, The Lazer-McKenna conjecture and a free boundary problem in two dimensions, *J. London Math. Soc.*, 78(2008), 639–662.
L.E. Fraenkel and M.S. Berger, A global theory of steady vortex rings in an ideal fluid, *Acta Math.*, 132(1974), 13–51.
F. Flucher and J. Wei, Asymptotic shape and location of small cores in elliptic free-boundary problems, *Math. Z.*, 228(1998), 638–703.
W. Kulpa, The Poincaré-Miranda theorem, *The Amer. Math. Monthly*, 104(1997), 545–550.
G. Li, S. Yan and J. Yang, An elliptic problem related to planar vortex pairs, *SIAM J. Math. Anal.*, 36(2005), 1444–1460.
Y. Li and S. Peng, Multiple solutions for an elliptic problem related to vortex pairs, *J. Diff. Equat.*, 250(2011), 3448–3472.
C.C.Lin, On the motion of vortices in two dimension – I. Existence of the Kirchhoff-Routh function, *Proc. Natl. Acad. Sci. USA*, 27(1941), 570–575.
C. Miranda, Un’osservazione su un teorema di Brouwer., *Boll. Un. Mat. Ital.*, 3(1940), 5–7.
C. Marchioro and M. Pulvirenti, Euler evolution for singular initial data and vortex theory, *Comm. Math. Phys.*, 91(1983), 563–572.
W.-M. Ni, On the existence of global vortex rings, *J. Anal. Math.*, 37(1980), 208–247.
J. Norbury, Steady planar vortex pairs in an ideal fluid, *Comm. Pure Appl. Math.*, 28(1975), 679–700.
D. Smets and J. Van Schaftingen, Desingulariation of vortices for the Euler equation, *Arch. Rational Mech. Anal.*, 198(2010), 869–925.
B. Turkington, On steady vortex flow in two dimensions. I, II, *Comm. Partial Diff. Equat.*, 8(1983), 999–1030, 1031–1071.
J. Yang, Existence and asymptotic behavior in planar vortex theory, *Math. Models Methods Appl. Sci.*, 1(1991), 461–475.
J. Yang, Global vortex rings and asymptotic behaviour, *Nonlinear Anal.*, 25(1995), 531–546.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We analyze excited baryon states using a holographic dual of QCD that is defined on the basis of an intersecting D4/D8-brane system. Studies of baryons in this model have been made by regarding them as a topological soliton of a gauge theory on a five-dimensional curved spacetime. However, this allows one to obtain only a certain class of baryons. We attempt to present a framework such that a whole set of excited baryons can be treated in a systematic way. This is achieved by employing the original idea of Witten, which states that a baryon is described by a system composed of $N_c$ open strings emanating from a baryon vertex. We argue that this system can be formulated by an ADHM-type matrix model of Hashimoto-Iizuka-Yi together with an infinite tower of the open string massive modes. Using this setup, we work out the spectra of excited baryons and compare them with the experimental data. In particular, we derive a formula of the nucleon Regge trajectory assuming that the excited nucleons lying the trajectory are characterized by the excitation of a single open string attached on the baryon vertex.'
author:
- Yasuhiro Hayashi
- Takahiro Ogino
- Tadakatsu Sakai
- Shigeki Sugimoto
title: '[Stringy excited baryons in holographic QCD]{}'
---
Introduction
============
Ever since the AdS/CFT correspondence was proposed by Maldacena (for a review, see [@adscft]), it has been recognized that it may provide us with a powerful tool for analyzing nonperturbative dynamics of nonabelian gauge theories. One of the most intensive applications of the AdS/CFT correspondence is to hadron physics of QCD. A key ingredient of hadron physics is how to understand spontaneous breaking of chiral symmetry. A holographic dual of QCD (in the top down approach) with manifest chiral symmetry was presented in [@SS1; @SS2] on the basis of an intersecting D4/D8-brane configuration. It was argued there that chiral symmetry breaking is realized as a smooth interpolation of D8 - anti-D8-brane (${\overline}{\rm D8}$) pairs in a curved background corresponding to D4-branes in type IIA supergravity. The associated Nambu-Goldstone mode (pion) is shown to arise from the 5 dimensional gauge field on the interpolated D8-branes. This model is formulated in large $N_c$ and large ’t Hooft coupling $\lambda$ regime with $N_c\gg N_f$, where $N_c$ and $N_f$ are the numbers of color and flavor, respectively, for the purpose of suppressing intricate stringy and quantum gravity effects. In spite of this approximation, predictions of this model matches well with various experimental data in the low energy hadron physics.
In particular, it has been shown that the meson effective theory is given by a 5 dimensional $U(N_f)$ gauge theory, and a tower of vector and axial-vector mesons including $\rho$ and $a_1$ mesons appear as the Kaluza-Klein (KK) modes of the 5 dimensional gauge field. Other mesons including higher spin mesons are interpreted as excited open string modes attached on the D8-branes. [@ISS] As they are described by an open string, nearly-linear Regge trajectories with mild nonlinear corrections are obtained quite naturally, and it has been argued that the predicted meson spectrum agrees at least qualitatively with what is observed in nature.
The holographic model is also used to study the baryon sector. This is performed by noting that a baryon can be realized as a topological soliton in the 5 dimensional gauge theory with a baryon number identified with a topological number. The original idea is due to Skyrme [@Skyrme] by adding a so-called Skyrme term to the chiral Lagrangian of the massless pion. In the holographic model, the soliton solution is given by an instanton solution with the instanton number regarded as the baryon number [@SS1]. The analysis of the moduli space quantum mechanics analogous to the work [@ANW] in the Skyrme model was performed in [@HSSY] and [@HSS] to obtain the baryon spectrum and the static properties, respectively,[^1] and again many of the results turned out to be consistent with the experimental data. However, one of the limitations in [@HSSY] is that it describes only a subclass of baryons with $I=J$ for $N_f=2$. Here, $J$ and $I$ denotes the spin and the isospin of a baryon, respectively. The reason for this limitation is clear: the moduli space approximation only takes into account the light degrees of freedom that correspond to the massless sector in the open string spectrum. We are led naturally to expect that incorporation of massive open string states enables us to obtain a larger class of baryons with $I\ne J$,[^2] as it was done in [@ISS] for the meson sector.
The purpose of this paper is to examine the holographic baryons following this line. To this end, we utilize the idea of Witten [@Witten:bv] that a holographic description of baryons is made by introducing a D-brane configuration, called a baryon vertex. In the present holographic model, we add a D4-brane that wraps around an ${{\mathbb S}}^4$ with $N_c$ units of RR-flux over it. It was found in [@Witten:bv; @Gross:1998gk] that the RR-flux forces $N_c$ open string to extend between the D4-brane and the D8-branes. The whole system is regarded as a holographic baryon. As a consistency check, the instanton solution is identical with the baryon vertex D4-brane in the context of the effective theory. The baryon states can be computed by working out a bound state of a many-body quantum mechanics that is defined from open strings attached on the baryon vertex. There are two types of open strings that should be taken into account. One of them is the 4-4 strings whose both end points are attached on the baryon vertex D4-brane and the other is the 4-8 strings that extend between the D4-brane and one of the D8-branes. As it was shown in [@Witten:1995gx; @Douglas:1995bn], the massless degrees of freedom that arise from these strings correspond to the instanton moduli space in the ADHM construction [@Atiyah:1978ri] and it is expected to be equivalent to the moduli space quantum mechanics in the soliton approach. This approach was proposed in [@HIY], in which a matrix quantum mechanics describing multiple baryon systems was derived. Our main idea is to incorporate the massive open string states into this quantum mechanics to describe heavier baryons. Solving the bound state problem in quantum mechanics is highly involved in general. In this present case, however, we argue that taking the large $N_c$ limit makes the problem tractable. This is because the string coupling is of ${{\mathcal O}}(1/N_c)$ so that interactions among open strings are mostly negligible in the large $N_c$ limit.
The fundamental degrees of freedom in the quantum mechanics are given by massless and an infinite tower of massive modes of open strings attached on the baryon vertex D4-brane. The mass spectrum can be worked out by quantizing the open strings in the curved background (\[metric\]), but this is technically difficult to achieve. As suggested in [@ISS], this problem gets simplified drastically by taking the limit $\lambda\gg 1$, where the spacetime curvature becomes negligible. Nontrivial curvature effects into the mass spectrum are incorporated perturbatively in $1/\lambda$ expansions. Using these results, the many-body quantum mechanics is formulated in a manner that is simple and powerful enough to study a wide range of holographic baryons quantitatively. As an application, we derive the mass formula of the nucleon and its excited states. We also discuss its implication to the nucleon Regge trajectory.
The organization of this paper is as follows. In section 2, after giving a brief review of the holographic model of QCD with emphasis on a baryon vertex, we compute the mass spectrum of the open strings attached on the baryon vertex and D8-branes. With this result, section 3 formulates a many-body quantum mechanics that enables one to compute the mass spectrum of baryons that are missing in [@HSSY]. In section 4, we compare the predictions of this model to experiments. We conclude this paper in section 5 with summary and some comments about future directions. Some technical formulas that are used in this paper are summarized in appendix A.
Holographic model of QCD and baryons {#sec:baryonstates}
====================================
Brief review of the model
-------------------------
The holographic model of QCD we work with is constructed from an intersecting D4/D8-brane system [@SS1; @SS2]. The $N_c$ D4-branes wrap around a circle on which SUSY breaking boundary condition is imposed, and yield gluons of gauge group $SU(N_c)$ on the worldsheet at low energy compared with the circle radius $1/{M_{\rm KK}}$. $N_f$ D8- and ${\overline}{\rm D8}$-branes are placed at the anti-podal points of the SUSY breaking circle. Quantization of D4-D8 and D4-${\overline}{\rm D8}$ strings gives left- and right-handed quarks in the fundamental representation of $SU(N_c)$, respectively This system has a manifest chiral $U(N_f)_L\times U(N_f)_R$ symmetry.
The holographic dual of this model is formulated by replacing the D4-branes with a solution of type IIA supergravity with a nontrivial dilaton $\phi$ [@Witten:d4]: $$\begin{aligned}
&ds^2=\frac{4}{27}\lambda l_s^2 \,d{\widetilde}{s}^2\ ,
{\nonumber}\\
&d{\widetilde}{s}^2=
K(r)^{1/2} \eta_{\mu\nu}dx^{\mu}
dx^{\nu}+ K(r)^{-5/6}dr^2 + K(r)^{-1/2}r^2d\theta^2
+\frac{9}{4}K(r)^{1/6}d\Omega_4^2\ ,
\label{metric}\end{aligned}$$ $$\begin{aligned}
e^\phi= \frac{\lambda^{3/2}}{3\sqrt{3}\pi N_c}K(r)^{1/4}\ ,
\label{dilaton}\end{aligned}$$ Here, $\mu,\nu=0,1,2,3$ denotes the indices of 4d Minkowski spacetime where QCD is defined. $d\Omega_4^2$ is the metric of a unit ${{\mathbb S}}^4$, and $K=1+r^2$.[^3] $\theta$ is the coordinate of the SUSY breaking circle. In addition, there exists $N_c$ units RR-4-form flux over the ${{\mathbb S}}^4$: $$\begin{aligned}
\frac{1}{2\pi}\int_{S^4}F_4=N_c\ .
\label{RR}\end{aligned}$$ It is useful to define $$\begin{aligned}
z=r\sin\theta \ ,~~ y=r\cos\theta \ .{\nonumber}\end{aligned}$$ The metric (\[metric\]) is defined in the decoupling limit, where the dependence on $l_s$, the string length, factorizes as a prefactor. As a consequence, the string theory on this background is independent of $l_s$. This allows one to set $$\begin{aligned}
\alpha'\equiv l_s^2=\frac{27}{4\lambda}
\label{lslam}\end{aligned}$$ in units of ${M_{\rm KK}}=1$ so that $ds^2=d{\widetilde}s^2$. (See [@ISS] for more details on this point.) It follows that the stringy excitation modes have mass of ${{\mathcal O}}(\lambda^{1/2})$ and may be neglected at low energies for $\lambda\gg 1$.
Assuming $N_c \gg N_f$,[^4] the D8-branes can be regarded as probes with no backreaction to the metric (\[metric\]) taken into account. It is shown [@SS1] that the D8- and ${\overline}{\rm D8}$-brane pairs interpolate with each other smoothly at $z=y=0$ and the resultant D8-brane worldvolume is specified by the embedding equation $y=0$. In this setup, the mesons are identified with the open strings attached on the D8-branes that can move along the $z$ direction.
In order to incorporate baryon degrees of freedom into the model, we introduce a baryon vertex [@Witten:bv], which is given by a single D4-brane wrapping around ${{\mathbb S}}^4$ at $z=y=0$. We refer to this D4-brane as a D4$_{\rm BV}$ in order to distinguish it from $N_c$ color D4-branes. The RR flux (\[RR\]) forces $N_c$ open strings to extend between the D8-branes and the baryon vertex. This configuration is identified with a single baryon. It is argued in [@SS1] that this brane system is realized as an instanton solution on the D8-brane worldvolume theory. By analyzing the moduli space quantum mechanics corresponding to this instanton solution, the paper [@HSSY; @HSS] has shown that aspects of the baryon dynamics are reproduced from this model both qualitatively and quantitatively. One of the limitations in this analysis, however, is that describing a baryon vertex as a classical solution of the $U(N_f)$ gauge theory on the D8-branes is valid only for low-lying baryons. This is because the $U(N_f)$ gauge theory is an effective theory of the D8-branes with only the massless degrees of freedom taken into account. In addition, the moduli space approximation only keeps light degrees of freedom in the fluctuations around the soliton solution. In fact, these are the main reasons why the analysis in [@HSSY] leads to only baryons with the spin $J$ and isospin $I$ equal to each other for the $N_f=2$ case. For the purpose of obtaining more general baryons, we thus have to consider stringy effects in the baryon vertex.
Quantization of open strings in a flat spacetime limit {#subsec:states}
------------------------------------------------------
It is highly difficult to make a full quantization of a string that propagates in the curved background (\[metric\]) in the presence of the RR flux (\[RR\]). In order to circumvent this problem, we follow [@ISS]. We first take the large $\lambda$ limit, where the curved background can be approximated with a 10 dimensional flat spacetime. Then, the baryon configuration reduces to the system with $N_f$ D8-branes and a D4$_{\rm BV}$-brane with $N_c$ open strings stretched between them in the flat background. For a technical reason, it is useful to formally T-dualize the system in the $y$ direction. The D8/D4$_{\rm BV}$-brane system gets mapped to a D9/D5$_{\rm BV}$-brane configuration shown below:
[|c|c|c|c|c|c|c|c|c|c|c|]{} & 0 & 1 & 2 & 3 & $z$ & ${\tilde{y}}$ & 6 & 7 & 8 & 9\
$N_f\times$D9 & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$\
D5$_{\rm BV}$ & $\bigcirc$ & & & & & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$ & $\bigcirc$\
The $123z$- and $6789$-directions are labeled by indices $M$ and $i$, respectively. The $6789$-directions span ${{\mathbb R}}^4$, which results from ${{\mathbb S}}^4$ that is decompactified for $\lambda\gg 1$. Quantization of a 9-5 and 5-5 string is performed most easily by using a light-cone quantization, where the light-cone coordinate is taken to be $x^0\pm {\tilde{y}}$. The manifest spacetime symmetry of the brane system is $SO(4)_{123z}\times SO(4)_{6789}$.
We first study the light-cone quantization of a 9-5 string. The equations of motion (EOM) of the worldsheet boson in the 6789-directions is solved in terms of Fourier expansions with an integer modding, while that in the 123$z$-directions in terms of those with a half-integer modding, because of the boundary conditions imposed on them. For the worldsheet fermions in the NS (R) sector, the solutions of EOM in the 6789-directions are written in terms of Fourier expansions with a half-integer (integer) modding, while those in the 123$z$-directions written in terms of those with an integer (half-integer) modding. It follows that the NS ground state is degenerate due to the fermion zero modes, belonging to a spinor representation of $SO(4)_{123z}$. The R ground state is denigrate too and belongs to a spinor representation of $SO(4)_{6789}$. We label an irreducible representation of $SO(4)_{123z}\simeq
(SU(2)_L\times SU(2)_R)/{{\mathbb Z}}_2$ by $(s_L,s_R)$, where $s_L$ and $s_R$ are the spin of $SU(2)_L$ and $SU(2)_R$, respectively. The (integer spin) representation of $SO(4)_{6789}$ is labeled by Young tableau as ${\textbf{1}}$, ${\ytableausetup{smalltableaux} \ydiagram{1}_{4}}$, ${\ytableausetup{smalltableaux} \ydiagram{1,1}_{6}}$, ${\ytableausetup{smalltableaux} \ydiagram{2}_{9}}$, etc.,where the subscripts denote the dimensions. Then, the low-lying 9-5 string states in the NS sector with the GSO projection imposed are summarized in Table \[table:single\].
-------------- ----------------------------------------------------------------- ------------------------------------------- --------------------------------------------------------------------------
$N_{95}=0$ $\ket{a}_{{\mathrm{NS}}}$ $(1/2,0)$ ${\textbf{1}}$
$N_{95}=1/2$ $\alpha_{-1/2}^M \ket{a}_{{\mathrm{NS}}}$ $(0,1/2) {\oplus}(1,1/2)$ ${\textbf{1}}$
$\psi_{-1/2}^i \ket{{\dot{a}}}_{{\mathrm{NS}}}$ $(0,1/2)$ ${\ytableausetup{smalltableaux} \ydiagram{1}_{4}}$
$N_{95}=1$ $\alpha_{-1}^i \ket{a}_{{\mathrm{NS}}}$ $(1/2,0)$ ${\ytableausetup{smalltableaux} \ydiagram{1}_{4}}$
$\psi_{-1}^M \ket{{\dot{a}}}_{{\mathrm{NS}}}$ $(1/2,0) {\oplus}(1/2,1)$ ${\textbf{1}}$
$\alpha_{-1/2}^M \alpha_{-1/2}^N \ket{a}_{{\mathrm{NS}}}$ $(1/2,0) {\oplus}(1/2,1) {\oplus}(3/2,1)$ ${\textbf{1}}$
$\alpha_{-1/2}^M \psi_{-1/2}^i \ket{{\dot{a}}}_{{\mathrm{NS}}}$ $(1/2,0) {\oplus}(1/2,1)$ ${\ytableausetup{smalltableaux} \ydiagram{1}_{4}}$
$\psi_{-1/2}^i \psi_{-1/2}^j \ket{a}_{{\mathrm{NS}}}$ $(1/2,0)$ $\raisebox{0.1cm}{${\ytableausetup{smalltableaux} \ydiagram{1,1}_{6}}$}$
-------------- ----------------------------------------------------------------- ------------------------------------------- --------------------------------------------------------------------------
Although it is not manifest in the light-cone quantization, the 6 dimensional Lorentz symmetry on the ${\mathrm{D5_{BV}}}$-brane worldvolume allows one to summarize the massive excitations into the irreducible representations of the little group $SO(5)_{{\tilde{y}}6789}$, which contains ${SO(4)_{6789}}$ as a subgroup. Table \[table:soyi\] is the list of the low-lying 9-5 string states in the NS sector in terms of $SO(5)_{{\tilde{y}}6789}$.
-------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
$N_{95}=0$ $(1/2,0)\,{\textbf{1}}$
$N_{95}=1/2$ $(0,1/2)\,{\ytableausetup{smalltableaux} \ydiagram{1}_{5}}{\oplus}(1,1/2)\,{\textbf{1}}$
$(1/2,0)\,{\ytableausetup{smalltableaux} \ydiagram{1,1}_{10}}{\oplus}(1/2,0)\,{\ytableausetup{smalltableaux} \ydiagram{1}_{5}}{\oplus}(1/2,1)\,{\ytableausetup{smalltableaux} \ydiagram{1}_{5}}$
${\oplus}\, (1/2,0)\,{\textbf{1}}{\oplus}(1/2,1)\,{\textbf{1}}{\oplus}(3/2,1)\,{\textbf{1}}$
-------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
We next study the mass spectrum of a 5-5 string using the light-cone quantization. The worldsheet bosons can be Fourier expanded with an integer modding for both $123z$- and $6789$-directions. The worldsheet fermions in the NS (R) sector can be Fourier expanded with a half-integer (integer) modding for $123z$- and $6789$-directions. The physical ground state in the NS sector is massless and given by $\psi_{-1/2}^M|0\rangle_{\rm NS}$ and $\psi_{-1/2}^i|0\rangle_{\rm NS}$. Here, $|0\rangle_{\rm NS}$ is tachyonic, being GSO-projected out. The first excited 5-5 string states in the NS sector that survive the GSO projection are given by acting on $|0\rangle_{\rm NS}$ with a set of the creation operators with the total excitation number equal to $3/2$. These have the mass squared $(3/2-1/2)/l_s^2=1/l_s^2$ and are listed in Table \[table:55ns\].
------------ ------------------------------------------------------------------- --------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
$N_{55}=0$ $\psi_{-1/2}^M \ket{0}_{{\mathrm{NS}}}$ $(1/2,1/2)$ ${\textbf{1}}$
$\psi_{-1/2}^i \ket{0}_{{\mathrm{NS}}}$ $(0,0)$ ${\ytableausetup{smalltableaux} \ydiagram{1}_{4}}$
$N_{55}=1$ $\psi_{-1/2}^M\psi_{-1/2}^N\psi_{-1/2}^L \ket{0}_{{\mathrm{NS}}}$ $(1/2,1/2)$ ${\textbf{1}}$
$\psi_{-1/2}^M\psi_{-1/2}^N\psi_{-1/2}^i \ket{0}_{{\mathrm{NS}}}$ $(1,0){\oplus}(0,1)$ ${\ytableausetup{smalltableaux} \ydiagram{1}_{4}}$
$\psi_{-1/2}^M\psi_{-1/2}^i\psi_{-1/2}^j \ket{0}_{{\mathrm{NS}}}$ $(1/2,1/2)$ ${\ytableausetup{smalltableaux} \ydiagram{1,1}_{6}}$
$\psi_{-1/2}^i\psi_{-1/2}^j\psi_{-1/2}^k \ket{0}_{{\mathrm{NS}}}$ $(0,0)$ $\ytableausetup{smalltableaux} \ydiagram{1,1,1}_{4}$
$\psi_{-3/2}^M \ket{0}_{{\mathrm{NS}}}$ $(1/2,1/2)$ ${\textbf{1}}$
$\psi_{-3/2}^i \ket{0}_{{\mathrm{NS}}}$ $(0,0)$ ${\ytableausetup{smalltableaux} \ydiagram{1}_{4}}$
$\alpha_{-1}^M \psi_{-1/2}^N \ket{0}_{{\mathrm{NS}}}$ $(0,0) {\oplus}(1,0) {\oplus}(0,1) {\oplus}(1,1)$ ${\textbf{1}}$
$\alpha_{-1}^M \psi_{-1/2}^i \ket{0}_{{\mathrm{NS}}}$ $(1/2,1/2)$ ${\ytableausetup{smalltableaux} \ydiagram{1}_{4}}$
$\alpha_{-1}^i \psi_{-1/2}^M \ket{0}_{{\mathrm{NS}}}$ $(1/2,1/2)$ ${\ytableausetup{smalltableaux} \ydiagram{1}_{4}}$
$\alpha_{-1}^i \psi_{-1/2}^j \ket{0}_{{\mathrm{NS}}}$ $(0,0)$ ${\ytableausetup{smalltableaux} \ydiagram{1}_{4}}{\oplus}\raisebox{0.1cm}{${\ytableausetup{smalltableaux} \ydiagram{1,1}_{6}}$} {\oplus}{\ytableausetup{smalltableaux} \ydiagram{2}_{9}}$
------------ ------------------------------------------------------------------- --------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
As in the 9-5 string states, any massive state of the 5-5 string is summarized into an irreducible representation of $SO(4)_{123z}\times SO(5)_{{\tilde{y}}6789}$. It is found that the first excited states with $N_{55}=1$ in Table \[table:55ns\] are rearranged as $$\begin{aligned}
{\left[ (1,0) {\oplus}(0,1) \right] }&\,\ytableausetup{smalltableaux} \ydiagram{1}_{5}\ ,\ \ \
(1/2,1/2)\,\ytableausetup{smalltableaux} \ydiagram{1,1}_{10},\ \ \
{(1/2,1/2)\,\ytableausetup{smalltableaux} \ydiagram{1}_{5}}\ ,\ \ \
(0,0)\,\ytableausetup{smalltableaux} \ydiagram{1,1,1}_{10}\ ,\ \ \
{\nonumber}\\
(0,0)&\,\ytableausetup{smalltableaux} \ydiagram{2}_{15}\ , \ \
{(1/2,1/2)\,{\textbf{1}}}\ ,\ \ \
{\left[ (0,0){\oplus}(1,1) \right] }\,{\textbf{1}}\ , \ \ \
\label{55:GSO1st}\end{aligned}$$ where the Young tableaux are those of $SO(5)_{{\tilde{y}}6789}$. In fact, these states are obtained as the decomposition of $~\ytableausetup{smalltableaux} \ydiagram{2}_{44} {\oplus}~
\ytableausetup{smalltableaux} \ydiagram{1,1,1}_{84}$ of $SO(9)$, which is the same as the first excited 9-9 string states considered in [@ISS].
Symmetries in the presence of a baryon vertex {#subsec:disc}
---------------------------------------------
It is discussed in [@ISS] that the D4/D8-brane system has discrete symmetries that are identified with those in massless QCD. Parity $P$ and charge conjugation $C$ are given by $$\begin{aligned}
P=I_{123z}\ ,~~~ C=I_{z89}\Omega\,(-1)^{F_L}\ ,\end{aligned}$$ respectively, where $I_{i_1i_2\cdots}$ is spacetime involution along the $i_1,i_2,\cdots$ directions, $\Omega$ is a worldsheet parity, and $F_L$ is a spacetime fermion number in the left-moving sector of a string worldsheet. A ${\mathrm{D4_{BV}}}$-brane placed at $x^1=x^2=x^3=y=z=0$[^5] is invariant under $P$, while it is mapped to a ${\overline}{\rm D4}_{\rm BV}$-brane under $C$. To see the latter, note that when the ${{\mathbb Z}}_2$ action generated by $C$ is gauged, a background has an O6-plane at $z=x^8=x^9=0$, and it is known that the ${\mathrm{D4_{BV}}}$-brane has to be paired with a ${\overline}{\rm D4}_{\rm BV}$-brane in the presence of the O6-plane.[@ISS2] This is consistent with the fact that the baryon is invariant under the parity, up to sign of the wavefunction, while it is mapped to an anti-baryon under the charge conjugation.
In order to see how $P$ acts on the NS ground state of the 9-5 string considered in section \[subsec:states\], it is useful to write the parity operator in a bosonized form. We note that the worldsheet fermions of a 9-5 string can be expressed using free worldsheet complex scalars $H^1$ and $H^2$ as $$\begin{aligned}
\psi^1 \pm i \psi^2 =e^{\pm i H^1} \ ,~~
\psi^3 \pm i \psi^z =e^{\pm i H^2} \ .{\nonumber}$$ Parity acts on the worldsheet fermions as $$\begin{aligned}
\psi^M \to - \psi^M \ ,{\nonumber}$$ which in turn induces the transformation of $H^1,H^2$ as $$\begin{aligned}
(H^1,H^2) \to (H^1 + (2n_1+1)\pi, H^2 + (2n_2+1)\pi)\ ,
\label{eq:PHp}\end{aligned}$$ with a choice of $n_1,n_2\in{{\mathbb Z}}$. The vertex operator corresponding to the NS ground state of a 9-5 string is given by $$\begin{aligned}
e^{i(s_1H^1+s_2H^2)} \ ,\end{aligned}$$ up to a ghost sector that is invariant under $P$, with $s_1=s_2=\pm 1/2$ for $\ket{a}_{{\mathrm{NS}}}$ and $s_1=-s_2=\pm 1/2$ for $\ket{\dot a}_{{\mathrm{NS}}}$. Therefore, the parity transformation (\[eq:PHp\]) acts as the chirality operator on the spinor representation of $SO(4)_{123z}$ up to a sign ambiguity. We choose $n_1$ and $n_2$ in (\[eq:PHp\]) such that $\ket{a}_{{\mathrm{NS}}}$ and $\ket{\dot a}_{{\mathrm{NS}}}$ are parity even and odd, respectively. With this convention, the parity of the proton and the neutron turn out to be even. This is consistent with the conventional choice of the parity in QCD, in which the parity of quarks are chosen to be even. For a ${\overline}{\rm D5}_{\rm BV}$-brane, which represents an anti-baryon, since the GSO projection is opposite, the parity of the the NS ground state is odd. This is again consistent with the fact that the anti-quarks have odd parity.
Then, the parity of the excited states can be computed by using the transformation laws of the creation operators that act on the ground state. Namely, $\psi_{-r}^M$ and $\alpha_{-r}^M$ with $M=1,2,3,z$ are parity odd and $\psi_{-r}^i$ and $\alpha_{-r}^i$ with $i=6,7,8,9$ are parity even operators.
In addition to these symmetries, the D4/D8-system admits a discrete symmetry that has no counterpart in QCD. That is called $\tau$-parity[^6] and defined as $$\begin{aligned}
P_\tau=I_{y9}\,(-1)^{F_L} \ .\end{aligned}$$ As discussed in [@ISS], both the quarks that originate from 4-8 and 4-$\bar{8}$ strings in the open string picture, and the gluons that originate from the 4-4 strings are even under the $\tau$-parity. This implies that all the states that can be interpreted as the genuine color singlet states of QCD have to be $\tau$-parity even as well. There are $\tau$-parity odd states in the spectrum of the bound states in our model. However, such states are artifacts of the model, which do not have counterparts in QCD, and we will not consider them in the following.
Assuming that the ${\mathrm{D4_{BV}}}$-brane is placed at $y=0$, one can show that the ${\mathrm{D4_{BV}}}$-brane is invariant under the $\tau$-parity $P_\tau$. To see this, we note that $I_{y9}$ maps the D4$_{\rm BV}$ to a $\overline{\rm D4}_{\rm BV}$ and $(-1)^{F_L}$ maps it back to a D4$_{\rm BV}$.
For the purpose of reading off $\tau$-parity of an open string state, it is useful to work in the T-dualized description used in section \[subsec:states\]. When the $y$-direction is T-dualized, $P_\tau$ is mapped to $$\begin{aligned}
\widetilde{P}_\tau=I_{9\tilde{y}} \ ,
\label{eq:tauparity}\end{aligned}$$ where $\tilde{y}$ is the T-dualized coordinate of $y$. This is simply a 180 degree rotation in the 9-${\tilde{y}}$ plane and it is easy to find the action of ${\widetilde}P_\tau$ from the representation of $SO(5)_{{\widetilde}6789}$ listed in Table \[table:soyi\] and (\[55:GSO1st\]).
In addition to the $\tau$-parity discussed above, we can also use the $SO(5)$ isometry of ${{\mathbb S}}^4$ in the background to single out the open string states that could be used to construct a baryon in QCD. It is easy to see that both quarks and gluons are invariant under this $SO(5)$, and hence the baryons in QCD have to be $SO(5)$ singlet. In the flat spacetime limit, the requirement of the $SO(5)$ invariance amounts to demanding the states to be $SO(4)_{6789}$-singlet and carry no momentum along the 6789-directions. In the T-dualized picture, we should also impose the condition that the momentum along ${\tilde{y}}$ is zero, since the original $y$ direction is not compactified and there is no winding mode along $y$. Therefore, among the open string states obtained in section \[subsec:states\], we only consider the states that are invariant under $SO(4)_{6789}$ and the $\tau$-parity ${\widetilde}{P}_\tau$, and carry no momentum along the ${\tilde{y}}6789$ directions.
Summary of the results
----------------------
We first derive the 9-5 string states that meet the conditions discussed in the last subsection. The requirement of the $SO(4)_{6789}$ invariance implies that the R-sector must be removed because all the states in the R-sector are $SO(4)_{6789}$-nonsinglet. It follows from the $\tau$-parity condition that among the $SO(4)_{6789}$-singlet NS states, only those with an even number of the spacetime index $\tilde{y}$ are allowed. The NS ground state satisfies these conditions. For the first excited states (those with $N_{95}=1/2$) listed in Table \[table:soyi\], only the state with $(s_L,s_R)=(1,1/2)$ is allowed. From the second excited states with $N_{95}=1$, we pick up $$\begin{aligned}
(1/2,0){\textbf{1}}{\oplus}(1/2,1){\textbf{1}}{\oplus}(3/2,1){\textbf{1}}\ . {\nonumber}$$ Finally, we set the momenta along the ${\tilde{y}}6789$ direction to zero, which is equivalent to omitting the dependence of the corresponding wavefunctions on ${\tilde{y}}$ and $x^{6,7,8,9}$. These results are summarized in Table \[table:95summary\].
-------------- ----------- ------------------------------ ----- -----
$N_{95}=0$ $(1/2,0)$ $1/2$ $+$
$N_{95}=1/2$ $(1,1/2)$ $3/2 {\oplus}1/2$ $-$ $1$
$N_{95}=1$ $(1/2,0)$ $1/2$ $+$ $2$
$(1/2,1)$ $3/2{\oplus}1/2$ $+$ $3$
$(3/2,1)$ $5/2{\oplus}3/2 {\oplus}1/2$ $+$ $4$
-------------- ----------- ------------------------------ ----- -----
In this table, we also listed the representation (spin) of $SU(2)_J$, which is related to the $SO(3)_{123}$ subgroup of $SO(4)_{123z}$ by $SU(2)_J/{{\mathbb Z}}_2\simeq SO(3)_{123}$. Note that $SO(4)_{123z}$ symmetry appears only in the flat spacetime limit and it is broken to $SO(3)_{123}$ due to the $z$-dependence of the background. The masses of these states in the flat spacetime limit are proportional to the excitation number $N_{95}$ as $$\begin{aligned}
m^2=\frac{N_{95}}{\alpha'}=\frac{4\lambda}{27}N_{95}\ ,~~~
(N_{95}=0,1/2,1,\cdots )\ ,
\label{95mass}\end{aligned}$$ where we have used the relation (\[lslam\]).
The quantum field corresponding to the 9-5 massless state is denoted by $\omega^I_a$, which reduces to a function of time $t$ only as discussed above. Here $a=1,2$ is the spin index for $SU(2)_J$ and $I=1,2,\cdots,N_f$ is index for the flavor $U(N_f)$ symmetry.
Next, we discuss the 5-5 string states. As in the 9-5 string case, all the R-states are non-singlet under $SO(4)_{6789}$ and thus ruled out. The NS massless states that satisfies all the conditions are given by $\psi_{-1/2}^M\ket{0}_{{\mathrm{NS}}}$ ($M=1,2,3,z$) only. The corresponding fields are denoted as $X^M$. Again, these fields reduce to the functions of $t$. Among the first excited states with $N_{55}=1$ listed in (\[55:GSO1st\]), the states listed below satisfy all the conditions $$\begin{aligned}
2\,(0,0) {\oplus}(1/2,1/2) {\oplus}(1,1) \ .\end{aligned}$$
------------ ------------- ------------------------- ----- -------
$N_{55}=0$ $(1/2,1/2)$ $1{\oplus}0$ $-$
$N_{55}=1$ $2\,(0,0)$ $0{\oplus}0$ $+$ $1,2$
$(1/2,1/2)$ $1 {\oplus}0$ $-$ $3$
$(1,1)$ $2 {\oplus}1 {\oplus}0$ $+$ $4$
------------ ------------- ------------------------- ----- -------
Note here that there are two $(0,0)$ states and one of them comes from $(0,0)\ytableausetup{smalltableaux}\,\ydiagram{2}_{15}$ in (\[55:GSO1st\]) with two ${\tilde{y}}$ indices. The masses are given by $$\begin{aligned}
m^2=\frac{N_{55}}{\alpha'}=\frac{4\lambda}{27}N_{55}\ ,~~~
(N_{55}=0,1,2,\cdots )\ .
\label{55mass}\end{aligned}$$ The results for the 5-5 strings are summarized in Table \[table:55summary\].
One baryon quantum mechanics {#sec:quant}
============================
In the previous section, we obtained the spectrum of the open strings attached on the baryon vertex ${\mathrm{D4_{BV}}}$-brane.[^7] Here, we write down the quantum mechanical ($0+1$ dimensional) action for these open string degrees of freedom. This action is a generalization of the quantum mechanical action obtained in a solitonic approach of the baryons in holographic QCD [@HSSY], which is related to that of the collective coordinates in the Skyrme model [@ANW], and the nuclear matrix model formulated in [@HIY], which is obtained by considering the ground states in the open string spectrum. The baryon states are obtained by quantizing this system. In this section, we give the general procedure to obtain baryon spectrum including the contributions from the excited open string states. The explicit construction of some of the low lying baryon states will be given in section \[sec:comparison\].
The action
----------
The action for the open string states attached on the baryon vertex ${\mathrm{D4_{BV}}}$-brane is written as $$\begin{aligned}
S=\int dt\,(L_0+L_m)\ ,\end{aligned}$$ where $L_0$ is the Lagrangian for the ground states while $L_m$ is the part that involves the excited states. $L_0$ is derived in [@HIY] as $$\begin{aligned}
L_0=
\frac{M_0}{2}\left[\dot X^2
+|D_0 w|^2
-V_{\rm ADHM}(w)
-V_0(X,w)\right]+N_c A_0\ ,\end{aligned}$$ where $w=(w_a^I)$ is a complex $N_f\times 2$ matrix variable with a spin ($SU(2)_J$) index $a=1,2$ and a flavor ($SU(N_f)_{\rm flavor}$) index $I=1,\cdots,N_f$, $X=(X^M)$ ($M=1,2,3,z$) is a real 4 component variable, and $A_0$ the $U(1)$ gauge field on the ${\mathrm{D4_{BV}}}$-brane. $w$ and $X$ corresponds to the ground state for 8-4 strings and 4-4 strings, respectively. The value of $X$ represents the position of the ${\mathrm{D4_{BV}}}$-brane in the 4 dimensional space parametrized by $(x^1,x^2,x^3,z)$. The dot denotes the time derivative as $\dot X\equiv \frac{d}{dt}X$ and $$\begin{aligned}
D_0 w\equiv\dot w-iA_0 w\equiv \frac{dw}{dt}-iA_0 w\ ,\end{aligned}$$ is the covariant derivative. The potential terms are given by $$\begin{aligned}
V_{\rm ADHM}(w)
&=&c \left({\mathrm{tr}}(\vec\tau\, w^\dag w)\right)^2
=c\left(
2|w^\dag w|^2-(|w|^2)^2
\right)
\ ,
\label{VADHM}
\\
V_0(X,w)&=&m_z^2 (X^z)^2+\gamma |w|^2
+\frac{v}{|w|^2}\ .
\label{V0}\end{aligned}$$ Here, $\vec\tau=(\tau^1,\tau^2,\tau^3)$ is the Pauli matrix and we have used the notation $|a|^2\equiv {\mathrm{tr}}(a^\dag a)=\sum_{b,I}(a^\dag)^b_I a_b^I$ for a complex matrix $a=(a_a^I)$. $M_0$, $c$, $m_z$, $\gamma$, $v$ are constants. $M_0$, $c$ and $m_z$ are related to the number of color $N_c$ and the ’t Hooft coupling $\lambda$ as [^8] $$\begin{aligned}
M_0=\frac{\lambda N_c}{27\pi} \ ,~~
c=\frac{\lambda^2}{3^6\pi^2}\ ,~~m_z^2=\frac{2}{3}\ .
\label{constants}\end{aligned}$$
The potential $V_{\rm ADHM}$ (\[VADHM\]) is obtained by integrating out the auxiliary fields in [@HIY]. The condition $V_{\rm ADHM}(w)=0$ is equivalent to the ADHM constraints for the ADHM construction of the self-dual instanton solution. The first term in $V_0$ (\[V0\]) represents the fact that the ${\mathrm{D4_{BV}}}$-brane is attracted to the origin in the $z$-direction due to the curved background. The second and third terms in $V_0$ (\[V0\]) are added rather phenomenologically. $\gamma$ is chosen to be $\gamma=1/6$ in [@HIY] so that the second term in (\[V0\]) recovers the corresponding term in the soliton approach [@HSSY]. The third term in (\[V0\]) was not present in [@HIY], but one could add it to have more flexibility. We treat $\gamma$ and $v$ as unspecified parameters for the moment.[^9]
$L_m$ is the Lagrangian with the excited states obtained in section \[sec:baryonstates\]. It can be written as $$\begin{aligned}
L_m=\frac{M_0}{2}
\left[\sum_j
\left(|D_0\Psi_j|^2-m_j^2|\Psi_j|^2
\right)+\sum_{k}\left(\dot\Phi_k^2-m_k^2\Phi_k^2
\right)+L_{\rm int}\right]\ ,
\label{Lm0}\end{aligned}$$ where $\Psi_j$ and $\Phi_k$ denote the fields corresponding to the excited states created by 8-4 strings and 4-4 strings, respectively. We call these fields “massive fields” in the following. The indices $j$ and $k$ label all the excited states and $m_j^2$ and $m_k^2$ are the mass squared of these states given in (\[95mass\]) and (\[55mass\]), which are of order $1/\alpha'\sim{{\mathcal O}}(\lambda)$. $\Psi_j$ are complex fields that couple with the $U(1)$ gauge field $A_0$ with the unit charge, while $\Phi_k$ are real fields, which are neutral under the $U(1)$ gauge symmetry. $L_{\rm int}$ gives the interaction terms for the massive fields that may also contain massless fields. We put the overall factor $M_0/2$ by convention so that all the fields have the dimension of length. Since the evaluation of the interaction terms including the massive states is beyond the scope of this paper, we assume that the contribution from $L_{\rm int}$ is small as far as the qualitative features of the baryon spectrum are concerned. In section \[subsec:Lint\], we argue that though most of the possible terms in $L_{\rm int}$ are suppressed in the large $N_c$ limit, there are some terms that could survive even in the large $N_c$ limit.
Gauss law constraint and Hamiltonian
------------------------------------
To quantize our system, we follow the approach developed recently in [@HMM]. We take the $A_0=0$ gauge and impose the EOM for $A_0$ (Gauss law constraint) as a physical state condition on the Hilbert space. The Gauss law constraint can be written as $$\begin{aligned}
q_w+\sum_j q_j=N_c\ ,
\label{Glaw}\end{aligned}$$ where $$\begin{aligned}
q_w\equiv \frac{M_0}{2}{\mathrm{tr}}(i(\dot w^\dag w-w^\dag\dot w))
\ ,~~~
q_j\equiv \frac{M_0}{2}
i(\dot \Psi_j^\dag \Psi_j-\Psi_j^\dag\dot\Psi_j)
\ .\end{aligned}$$ These $q_w$ and $q_j$ correspond to the charge associated to the phase rotation symmetries $w\rightarrow e^{i\alpha_w}w$ and $\Psi_j\rightarrow e^{i\alpha_j}\Psi_j$, respectively, which are approximate symmetries that exist when the interaction term $L_{\rm int}$ is neglected. The Gauss law constraint (\[Glaw\]) represents the fact that $N_c$ open strings have to be attached on the ${\mathrm{D4_{BV}}}$-brane and $q_j$ is interpreted as the number of the excited open strings associated with $\Psi_j$.[^10]
It is interesting to note that the Gauss law constraint (\[Glaw\]) implies that the spin of the baryon state is half-integer or integer for odd or even $N_c$, respectively.[^11] Indeed, the wavefunction for the baryon state satisfying the Gauss law constraint (\[Glaw\]) is of the form[^12] $$\begin{aligned}
\psi(X,w,w^\dag,\Psi_j,\Psi_{j}^\dag,\Phi_k)
=\underbrace{
w^{I_1}_{a_1}\cdots w^{I_{q_w}}_{a_{q_w}}\Psi_{j_1}\cdots
\Psi_{j_{N_c-q_w}}
}_{N_c}
{\widetilde}\psi(X, w^\dag w,\Psi_j^\dag\Psi_{j'},\Psi_j^\dag w,w^\dag\Psi_j,\Phi_k)
\ .
\label{wf}\end{aligned}$$ Here, ${\widetilde}\psi$ is a $U(1)$-invariant wavefunction that is written only through $U(1)$ invariants. Because 8-4 strings ($w$ and $\Psi_j$) and 4-4 strings ($X$ and $\Phi_k$) carry half-integer and integer spin, respectively, ${\widetilde}\psi$ can only have an integer spin and the spin of the state (\[wf\]) is $N_c/2 \mod~{{\mathbb Z}}$.
Omitting $L_{\rm int}$, the Hamiltonian in the $A_0=0$ gauge is given by $$\begin{aligned}
H=H_0+H_m\ ,\end{aligned}$$ with $$\begin{aligned}
H_0&=&\frac{1}{2M_0}(P_X^2+|P_w|^2)+\frac{M_0}{2}(V_{\rm ADHM}(w)+V_0(X,w))\ ,
\label{H0}
\\
H_m&=&\sum_j\left(\frac{1}{2M_0}|P_{\Psi_j}|^2+{\frac{1}{2}}M_0 m_j^2|\Psi_j|^2\right)
+\sum_k\left(P_{\Phi_k}^2+{\frac{1}{2}}M_0 m_k^2\Phi_k^2\right)\ ,
\label{Hm}\end{aligned}$$ where $P_X$, $P_w$, $P_{\Psi_j}$ and $P_{\Phi_k}$ are the momenta conjugate to $X$, $w$, $\Psi_j$ and $\Phi_k$, respectively. $H_m$ (\[Hm\]) is simply a collection of harmonic oscillators associated with the excited open string states obtained in section \[sec:baryonstates\]. The quantum mechanics for $H_0$ (\[H0\]) has been studied in [@HIY; @HMM], though the part with $w$ is treated in a different way in the following.
$H_0$ for $N_f=2$ {#Nf2}
-----------------
We are particularly interested in the cases with $N_f=2$, in which $w$ is a $2\times 2$ complex matrix and can be parametrized as $$\begin{aligned}
w=Y_0 1_2+i\vec Y\cdot\vec\tau\ ,~~~(Y=(Y_0,\vec Y)\in{{\mathbb C}}^4)\ ,\end{aligned}$$ where $1_2$ is the $2\times 2$ unit matrix. $Y$ transforms as the (complex) 4 dimensional vector representation of $SO(4)\simeq (SU(2)_I\times SU(2)_J)/{{\mathbb Z}}_2$, where $SU(2)_J$ and $SU(2)_I=SU(N_f)_{\rm flavor}$ with $N_f=2$ corresponds to the spin and isospin groups, respectively. The kinetic term for $w$ in (\[H0\]) is written as $$\begin{aligned}
\frac{1}{2M_0}|P_w|^2= -\frac{1}{4M_0}\Delta_Y\ ,\end{aligned}$$ where $\Delta_Y=4\frac{{\partial}^2}{{\partial}{\overline}Y_A{\partial}Y_A}$ is the Laplacian in ${{\mathbb C}}^4$.
Using the relations: $$\begin{aligned}
|w|^2= 2|Y|^2\ ,~~~
|w^\dag w|^2=4(|Y|^2)^2-2|Y^2|^2\ ,\end{aligned}$$ where $Y^2\equiv Y_0^2+\vec Y\cdot\vec Y$ and $|Y|^2\equiv |Y_0|^2+\vec Y^\dag\cdot\vec Y$, the ADHM potential can be written as $$\begin{aligned}
V_{\rm ADHM}(Y)
=4c\left((|Y|^2)^2-|Y^2|^2\right)\ .\end{aligned}$$ The minimum of this potential is parametrized by $$\begin{aligned}
Y=e^{i\theta}y\ ,~~~(\theta\in{{\mathbb R}}\ ,~y\in{{\mathbb R}}^4)\ .
\label{Ymin}\end{aligned}$$ Note that $y$ together with $X$ correspond to the collective coordinates of the one instanton configuration considered in [@HSSY]. More explicitly, $$\begin{aligned}
\rho\equiv \sqrt{y^2}\ ,~~~a\equiv y/\rho \end{aligned}$$ corresponds to the size and the $SU(2)$ orientation of the instanton solution, respectively.[^13] One way to include the components that are orthogonal to the directions along (\[Ymin\]) is to parametrize $Y$ as[^14] $$\begin{aligned}
Y=e^{i\theta}(y+i{\widetilde}y)\ ,~~~(\theta\in{{\mathbb R}}\ ,~y,{\widetilde}y\in{{\mathbb R}}^4)
\label{Y}\end{aligned}$$ with $$\begin{aligned}
{\widetilde}y=\beta_a i\Sigma^a a\ ,~~~
((\beta_a)= (\beta_1,\beta_2,\beta_3)\in{{\mathbb R}}^3)\ ,\end{aligned}$$ where $\Sigma^a$ ($a=1,2,3$) are the generators of $SU(2)_I$ acting on $y$, which are chosen to be pure imaginary anti-symmetric matrices. See Appendix \[SO4\] for the explicit forms. One can easily show $$\begin{aligned}
y\cdot{\widetilde}y=0\ ,~~~{\widetilde}y^2=\beta^2\ ,\end{aligned}$$ where $\beta^2=\beta_a\beta_a$ and the ADHM potential becomes $$\begin{aligned}
V_{\rm ADHM}(Y)=16c\, \rho^2\beta^2\ .
\label{VADHMrhobeta}\end{aligned}$$
Note that the parametrization (\[Y\]) has a redundancy induced by the ${{\mathbb Z}}_2$ transformation $$\begin{aligned}
\theta\rightarrow \theta+\pi\ ,~~y\rightarrow -y\ .
\label{Z2tr}\end{aligned}$$ When the wavefunction is written in terms of $\theta$, $y$ and $\beta_a$ instead of $Y$, we should impose the invariance of the wavefunction under this ${{\mathbb Z}}_2$ transformation.
In this paper, we consider the cases that $\beta$ takes small values so that $V_{\rm ADHM}$ does not generate additional mass term for $\rho$. One important observation is that the kinetic term of the Hamiltonian (\[H0\]) contains a term as $$\begin{aligned}
-\frac{1}{2M_0\rho^2}\frac{{\partial}^2}{{\partial}\theta^2}\ ,
\label{ddtheta}\end{aligned}$$ for $\beta^2\ll\rho^2$. (See (\[Laplacian\])) Since $q_w$ is the generator of the phase rotation of $Y$, we have the relation $$\begin{aligned}
q_w=-i{\frac{{\partial}}{{\partial}\theta}}
\label{qw}\end{aligned}$$ in the quantum mechanics. When we consider the cases with $\sum_j q_j\sim{{\mathcal O}}(1)$, $q_w$ has to be of ${{\mathcal O}}(N_c)$ because of the Gauss law constraint (\[Glaw\]). In such cases, the term (\[ddtheta\]) gives a potential of the form $$\begin{aligned}
-\frac{1}{2M_0\rho^2}\frac{{\partial}^2}{{\partial}\theta^2}
\sim \frac{N_c}{\lambda\rho^2}\ ,
\label{Nclamrho}\end{aligned}$$ up to a numerical factor in the large $N_c$ limit, which has an effect to push $\rho$ to have a larger value. Let $\rho_0$ be the value of $\rho$ that minimizes the effective potential given by adding this term to $V_0$ (\[V0\]). Assuming that the third term in (\[V0\]) is either negligible or of the same order as (\[Nclamrho\]), i.e. $v\sim {{\mathcal O}}(\lambda^{-2})$, we find $\rho_0^2\sim{{\mathcal O}}(\lambda^{-1})$, which is consistent with the results in [@Hong:2007kx; @HSSY]. We will shortly obtain an explicit expression for $\rho_0$ in the large $N_c$ limit (see (\[rho0\])), and show that it has an effect of generating a large mass term for $\beta_a$ in the next subsection.
Large $N_c$ limit {#LargeN}
-----------------
Now, let us figure out which terms in $H_0$ are important in the large $N_c$ limit. First we decompose $\rho$ as $\rho=\rho_0+{\delta\rho}$, and regard $M_0^{1/2}{\delta\rho}$, $M_0^{1/2}\beta_a$ and $a$ to be order 1 variables,[^15]which means $$\begin{aligned}
{\delta\rho}\sim\beta_a\sim{{\mathcal O}}(\lambda^{-1/2}N_c^{-1/2})\ ,~~~
{\frac{{\partial}}{{\partial}\rho}}\sim{\frac{{\partial}}{{\partial}\beta_a}}\sim
{{\mathcal O}}(\lambda^{1/2}N_c^{1/2})\ .\end{aligned}$$ Then, the leading (${{\mathcal O}}(\lambda N_c^2)$) and subleading (${{\mathcal O}}(\lambda N_c)$) terms in the Laplacian $\Delta_Y$ turn out to be $$\begin{aligned}
\Delta_Y\simeq
\left(\frac{1}{\rho^2}
+\frac{3\beta^2}{\rho_0^4}
\right){\frac{{\partial}{}^2}{{\partial}\theta^2}}
+{\frac{{\partial}{}^2}{{\partial}\rho^2}}+\left({\frac{{\partial}}{{\partial}\beta_a}}\right)^2
+{{\mathcal O}}(\lambda N_c^{1/2})\ .
\label{Laplacian}\end{aligned}$$ Keeping these terms, the Hamiltonian for $\rho$ and $\beta_a$ becomes $$\begin{aligned}
H_0|_{\rho,\beta} &\simeq&
\frac{1}{4M_0}\left[
\frac{q_w^2}{\rho^2}+\frac{3q_w^2}{\rho_0^4}\beta^2
-{\frac{{\partial}{}^2}{{\partial}\rho^2}}-\left({\frac{{\partial}}{{\partial}\beta_a}}\right)^2
\right]
+\frac{M_0}{2}\left[16c\,\rho^2\beta^2+2\gamma(\rho^2+\beta^2)
+\frac{v}{2(\rho^2+\beta^2)}
\right]
{\nonumber}\\
&\simeq&
2M_0\gamma\rho_0^2
-\frac{1}{4M_0}\left[
{\frac{{\partial}{}^2}{{\partial}\rho^2}}+\left({\frac{{\partial}}{{\partial}\beta_a}}\right)^2
\right]
+M_0\left(\omega_{{\delta\rho}}^2{\delta\rho}^2
+\omega_\beta^2\beta^2\right)\ ,
\label{H0rhobeta}\end{aligned}$$ where $$\begin{aligned}
\omega_{{\delta\rho}}^2=4\gamma\ ,~~~
\omega_\beta^2=8c\rho_0^2+\gamma-\frac{v}{4\rho_0^4}+\frac{3q_w^2}{4M_0^2\rho_0^4}\ .
\label{omega}\end{aligned}$$ Here we have imposed the condition that the $\rho_0$ minimizes the potential for $\rho$, which reads $$\begin{aligned}
\rho_0^2={\frac{1}{2}}\sqrt{\frac{1}{\gamma}\left(\frac{q_w^2}{M_0^2}+v\right)}
\ .
\label{rho0}\end{aligned}$$ The Hamiltonian (\[H0rhobeta\]) is a sum of the Harmonic oscillators for $\rho$ and $\beta_a$.
A few comments are in order: First, $\omega_{{\delta\rho}}^2$ coincides with $m_z^2$ in (\[constants\]) for $\gamma=1/6$ used in [@HIY], which is consistent with [@HSSY]. Second, the value of $\rho_0$ in (\[rho0\]) agrees with that in [@HIY] when $q_w=N_c$ and $v=0$. However, as pointed out in [@HIY], it is larger than the value in [@Hong:2007kx; @HSSY] by a factor of $5/4$. One can adjust the value of $v$ as $v=-\frac{N_c^2}{5M_0^2}$ to match with the value in [@Hong:2007kx; @HSSY]. Third, in the right side of $\omega_\beta^2$ in (\[omega\]), the first term $8c\rho_0^2$ is of order $\lambda$, while the other terms are of order 1. Recall that the masses of the excited open string states are $m^2\propto 1/\alpha'\sim{{\mathcal O}}(\lambda)$. This means that although $\beta_a$ arises as the ground states (the open string states with $N_{95}=0$), it acquires a large mass comparable to the massive excited states due to the ADHM potential (\[VADHM\]) together with the Gauss law constraint (\[Glaw\]).
Mass formula {#bmassf}
------------
As we have argued in section \[Nf2\], the Hamiltonian is reduced to a collection of harmonic oscillators in the large $N_c$ limit, which can be easily solved. Then, the masses of the baryons are obtained as $$\begin{aligned}
M= M_0^*
+m_z n_z+\omega_{{\delta\rho}}n_\rho+
\omega_\beta\sum_{a=1}^3 n_\beta^a
+\sum_{j}m_j (n^{\Psi}_j+n^{{\overline}\Psi}_j)+\sum_{k}m_k n^{\Phi}_k\ ,
\label{mass}\end{aligned}$$ where $n_z$, $n_\rho$, $n_\beta^a$, $n^{\Psi}_j$, $n^{{\overline}\Psi}_j$ and $n^{\Phi}_k$ are non-negative integers corresponding to the excitation levels of the harmonic oscillators associated with $X^z$, ${\delta\rho}$, $\beta_a$, $\Psi_j$, ${\overline}\Psi_j$ and $\Phi_k$, respectively. $m_z$, $\omega_{{\delta\rho}}$, $\omega_{\beta}$ are given in (\[constants\]) and (\[omega\]). $m_j$ and $m_k$ are the masses for the corresponding open string states given in (\[95mass\]) and (\[55mass\]), respectively. $M_0^*$ is a ($q_w$ dependent) constant whose classical value is $$\begin{aligned}
M^*_{0\,{\rm classical}}=(1+2\gamma\rho_0^2)M_0\ ,
\label{M0cl}\end{aligned}$$ where the first term $M_0$ comes from the tension of the ${\mathrm{D4_{BV}}}$-brane placed at $y=z=0$ and the second term $2\gamma\rho_0^2M_0$ is the first term in (\[H0rhobeta\]). It also contains the contributions from the zero point energies of all the fields in the system, including those neglected in section \[sec:baryonstates\]. Since there are infinitely many fields involved, it is not easy to evaluate it explicitly.[^16] For this reason, we leave $M_0^*$ as an unknown parameter and focus on the mass differences.
Note that the mass (\[mass\]) implicitly depends on the value of $q_w$ through the parameters $M_0^*$ and $\omega_\beta$. Because the Gauss law constraint (\[Glaw\]) implies that $q_w$ is related to $n_j^\Psi$ and $n_j^{{\overline}\Psi}$ by $$\begin{aligned}
q_w+\sum_j(n^{\Psi}_j-n^{{\overline}\Psi}_j)=N_c\ ,\end{aligned}$$ these parameters are state dependent.
As a consistency check, one can show a that the formula (\[mass\]) agrees with the leading order terms in the baryon mass formula obtained in [@HSSY] when $q_w=N_c$ and $n_\beta=n_j^{\Psi}=n_j^{{\overline}\Psi}=n_k^\Phi=0$. In fact, the baryon mass formula in [@HSSY] can be written as $$\begin{aligned}
M&=&
M_0+\sqrt{\frac{(\ell+1)^2}{6}+\frac{2}{15}N_c^2}
+\frac{2(n_\rho+n_z)+2}{\sqrt{6}}
\label{HSSYmass}
\\
&\simeq&
M_0+2M_0\gamma\rho_0^2
+\frac{(\ell+1)^2}{4M_0\rho_0^2}
+\omega_{{\delta\rho}} n_\rho +m_z n_z+{\frac{1}{2}}(\omega_{{\delta\rho}}+m_z)
+{{\mathcal O}}(N_c^{-3})\ ,
\label{HSSYmass2}\end{aligned}$$ where $\ell\in{{\mathbb Z}}_{\ge 0}$ is related to the spin $J$ and isospin $I$ as $I=J=\ell/2$. The $\ell$ dependence appears because the Laplacian in the $y$-space: $$\begin{aligned}
\Delta_y\equiv\left({\frac{{\partial}}{{\partial}y_A}}\right)^2
=\frac{1}{\rho^3}{\partial}_\rho(\rho^2{\partial}_\rho)+\frac{1}{\rho^2}
\Delta_{S^3}\ ,\end{aligned}$$ contains the Laplacian on $S^3$ parametrized by $a$, denoted by $\Delta_{S^3}$, whose eigenvalue is $-\ell(\ell+2)$. In (\[Laplacian\]), we have neglected this contribution, though it also appears in $\Delta_Y$ if we keep the ${{\mathcal O}}(N_c^0)$ term.
In [@HSSY], $\ell$ was chosen to be odd (or even) for odd (or even) $N_c$ by hand, so that the spin of the baryon obtained in the soliton approach is consistent with that in the quark model, as it is also the case for the Skyrme model with $N_f=2$. In our case, this condition is replaced with $\ell\equiv q_w~({\rm mod}~2 )$, which automatically follows from the fact that the eigenfunction of $\Delta_{S^3}$ is given by $$\begin{aligned}
T^{(\ell)}(a)\equiv
C^{A_1\cdots A_\ell}a_{A_1}\cdots a_{A_\ell}\ ,
\label{Tell}\end{aligned}$$ where $C^{A_1\cdots A_\ell}$ is a traceless symmetric tensor of rank $\ell$, and $\theta$ appears in the wavefunction as an overall factor $e^{iq_w\theta}$. As explained around (\[Z2tr\]), the wavefunction has to be invariant under the ${{\mathbb Z}}_2$ transformation (\[Z2tr\]), which implies $\ell\equiv q_w~({\rm mod}~2 )$.
Wavefunctions of the baryon states {#wavefunc}
----------------------------------
As discussed above, the Hamiltonian of the one baryon quantum mechanics is a collection of infinitely many harmonic oscillators in the large $N_c$ limit. The eigenfunction can be written as a product of a function of $X$, $a$, ${\delta\rho}$ and $\beta_a$, and that of $\Psi_j$, $\Psi_j^\dag$ and $\Phi_k$ as $$\begin{aligned}
\psi(X, a, {\delta\rho},\beta_a,\Psi_j,\Psi_j^\dag,\Phi_k )=
\psi_0(X, a, {\delta\rho},\beta_a)\,\psi_m(\Psi_j,\Psi_j^\dag,\Phi_k )\ .\end{aligned}$$ We call $\psi_0$ and $\psi_m$ to be wavefunctions for the massless and massive sectors, respectively.[^17]
The massless sector wavefunction $\psi_0$ can be written as $$\begin{aligned}
\psi_0= e^{i\vec p\cdot\vec X}
T^{(\ell)}(a)\psi_{n_z}(X^z)\psi_{n_\rho}({\delta\rho})\psi_{n_\beta}(\beta_a)\ ,\end{aligned}$$ where $e^{i\vec p\cdot \vec X}$ is the wavefunction for the plane wave with momentum $\vec p$, $T^{(\ell)}(a)$ is defined in (\[Tell\]), and $\psi_{n_z}$, $\psi_{n_\rho}$ and $\psi_{n_\beta}$ are the eigenfunctions of the harmonic oscillators for $X^z$, ${\delta\rho}$ and $\beta_a$ with the excitation numbers $n_z$, $n_\rho$ and $n^a_\beta$, respectively. We set $\vec p=0$ in the following for simplicity. We also use the bra-ket notation as $$\begin{aligned}
\ket{\psi_0}=\ket{\ell,n_z,n_\rho,n^a_\beta,q_w}\ .\end{aligned}$$ Here, $q_w$ is included in the notation to remember that the massless sector wavefunction also depends on $q_w$.
If $\psi_{n_\beta}$ is trivial, $\psi_0$ agrees with the large $N_c$ limit of the wavefunction obtained in [@HSSY]. As it was shown in [@HSSY], $T^{(\ell)}(a)$ has a degeneracy of $(\ell+1)^2$ that corresponds to the states in the representation of $I=J=\ell/2$. The mass formula (\[mass\]) appears to be independent of $\ell$, because the $\ell$ dependence is a subleading effect in the large $N_c$ limit. Upon taking finite $N_c$ effects into account, we expect that the energy is an increasing function of $\ell$ as it was the case in [@HSSY].[^18]
Note that since $X^z$ is parity odd, $\psi_{n_z}$ has parity $(-1)^{n_z}$. As mentioned in section \[LargeN\], $\omega_{{\delta\rho}}$ coincides with $m_z^2$ for $\gamma=1/6$ and hence the states with $(n_\rho,n_z)=(1,0)$ and $(n_\rho,n_z)=(0,1)$ are degenerate. This implies a degeneracy between parity even and odd states for those with $(n_\rho,n_z)\ne (0,0)$. This could be a hint toward an understanding of the parity doubling phenomenon in the excited baryons.[^19]
$\psi_{n_\beta}$ is a wavefunction for a 3 dimensional harmonic oscillator with respect to $\beta_a$ ($a=1,2,3$). The energy contribution in the mass formula (\[mass\]) for this part is $\omega_\beta n_\beta$ with $$\begin{aligned}
n_\beta\equiv \sum_{a=1}^3 n_\beta^a\ .\end{aligned}$$ The degeneracy is $$\begin{aligned}
{\frac{1}{2}}(n_\beta+1)(n_\beta+2)\ ,\end{aligned}$$ and the eigenspace for a given $n_\beta$ can be decomposed into a direct sum over the states with isospin $I=0,2,\cdots,n_\beta$ or $I=1,3,\cdots,n_\beta$ for even or odd $n_\beta$, respectively. For example, for the state with $\ell=1$ and $n_\beta=1$, the massless wavefunction $\psi_0$ has spin $1/2$ and isospin $1/2\otimes 1=3/2{\oplus}1/2$.
The wavefunction for the massive sector is given by the eigenfunctions of the harmonic oscillators associated with $\Psi_j$, $\Psi_j^\dag$ and $\Phi_k$, which is written in the bra-ket notation as $$\begin{aligned}
\ket{\psi_m}=\ket{n_j^\Psi,n_{j}^{{\overline}\Psi},n_k^\Phi}\ .\end{aligned}$$ In order to classify these states, we introduce a notation $$\begin{aligned}
{{\mathcal N}}={{\mathcal N}}_{84}+{{\mathcal N}}_{44}\ ,
\label{cN}\end{aligned}$$ which we call the level of a baryon, with $$\begin{aligned}
{{\mathcal N}}_{84}=\sum_{j}(n_j^\Psi+n_j^{{\overline}\Psi}) N_{84}^{(j)}\ ,~~~
{{\mathcal N}}_{44}=\sum_{k} n_k^\Phi N_{44}^{(k)}\ ,\end{aligned}$$ where $N_{84}^{(j)}$ and $N_{44}^{(k)}$ are the excitation numbers for $\Psi_j$ and $\Phi_k$ given in Table \[table:95summary\] and Table \[table:55summary\], respectively.[^20] It will become increasingly complicated to extract the spin and isospin for the states with larger ${{\mathcal N}}$. We will give some explicit examples of the baryon states in section \[sec:comparison\].
Comments on $L_{\rm int}$ {#subsec:Lint}
-------------------------
Here, we make some comments on $L_{\rm int}$ in (\[Lm0\]). First, we classify $L_{\rm int}$ depending on the order of the massive fields multiplied and assume that each term contains at least two massive fields so that the trivial configuration $\Psi_j=\Phi_k=0$ is a solution of the EOM for the massive fields. Note that the overall factor $M_0$ in the Lagrangian (\[Lm0\]) is proportional to $N_c$, which reflects the fact that the leading terms of the open string action are given by the string worldsheet of disk topology. As always, we neglect the loop corrections of string theory which are suppressed by $1/N_c$. Then, $L_{\rm int}$ is order 1 in the $1/N_c$ expansion with fixed $\lambda$. If one writes down the Lagrangian using canonically normalized massive fields $$\begin{aligned}
{\widetilde}\Psi_j\equiv \sqrt{M_0}\Psi_j\ ,~~~
{\widetilde}\Phi_k\equiv \sqrt{M_0}\Phi_k\ ,\end{aligned}$$ one finds that all the terms with more than two massive fields are suppressed in the large $N_c$ limit. Therefore, the terms in $L_{\rm int}$ that survive in the large $N_c$ limit are quadratic with respect to the massive fields. For the same reason, it should not contain $\dot w$, $\dot X$ or $X$. Then, the possible terms consistent with the $U(1)$ gauge symmetry are schematically written as $$\begin{aligned}
w^n (w^\dag)^n \Psi_j^\dag \Psi_{j'}\ ,~~~
w^n (w^\dag)^n \Phi_k^\dag \Phi_{k'}\ ,~~~
w^n (w^\dag)^{n+2} \Psi_j \Psi_{j'}\ ,~~~
w^n (w^\dag)^{n+1} \Psi_j \Phi_k\ ,
\label{wPP}\end{aligned}$$ with properly contracted indices and their complex conjugates. As we have seen in sections \[Nf2\] and \[LargeN\], $w$ is treated as order 1 variable, these terms may appear even in the large $N_c$ limit.
One might think that these terms are perhaps suppressed for large $\lambda$. However, unfortunately, the answer is no. Consider, for example, a term proportional to $|w|^{2n}|\Psi_j|^2\propto |Y|^{2n}|\Psi_j|^2$ for $N_f=2$. As we observed in section \[LargeN\], the leading term in $Y$ is $Y\sim \rho_0 a\sim {{\mathcal O}}(\lambda^{-1/2})$. Recall that all the fields have the dimension of length in our convention. To have the correct dimensions, there should be appropriate number of $\alpha'$ or ${M_{\rm KK}}$ in the coefficient of (\[wPP\]) to saturate the correct dimension of $L_{\rm int}$. A possible term is of the form $$\begin{aligned}
L_{\rm int}
\sim \alpha'^{-n-1}|Y|^{2n}|\Psi_j|^2
\sim \lambda |\Psi_j|^2\ ,
\label{Lint}\end{aligned}$$ which shifts the mass for $\Psi_j$ in the same order as the original mass term. This is the same mechanism as the mass generation of $\beta_a$ discussed in section \[LargeN\]. $L_{\rm int}$ may also induce mixing terms as well, and the diagonalization of the mass matrix may become very complicated. Because we do not know the explicit form of $L_{\rm int}$, we are not able to evaluate it explicitly and leave the detailed analysis including $L_{\rm int}$ for future research.
Comparison with experiments {#sec:comparison}
===========================
Regge trajectory {#Regge}
----------------
Here, we focus on the baryons listed in Table \[Nregge\], which are the lightest baryons with $I=1/2$ and $J^P=(n+1/2)^{(-)^n}$ ($n=0,1,\cdots,5$) found in the experiments.
baryons N N(1520) N(1680) N(2190) N(2220) N(2600)
------------- --------- ---------------- ---------------- ---------------- ---------------- ----------------
$J^P$ $1/2^+$ $3/2^-$ $5/2^+$ $7/2^-$ $9/2^+$ $11/2^-$
mass\[MeV\] 939 1510$\sim$1520 1680$\sim$1690 2140$\sim$2220 2250$\sim$2320 2550$\sim$2750
These baryons have been considered to be described by an excited (rotating) open string with a pair of quark and diquark attached on the two end points.[@Sharov:2013tga; @cobi][^21] An analogous object in our model is a ${\mathrm{D4_{BV}}}$-brane with $(N_c-1)$ 8-4 strings in the ground state and only one 8-4 string gets excited as $J$ increases. The aim of this subsection is to discuss whether our model gives us plausible predictions assuming that this is the correct interpretation. More explicitly, the lightest one in Table \[Nregge\], which is the nucleon (proton or neutron), is identified with $q_w=N_c$, $\ell=1$ and $n_\rho=n_z=n_j^\Psi=n_j^{{\overline}\Psi}=n_k^\Phi=0$.[^22] The excited nucleons with spin $J\ge 3/2$ in Table \[Nregge\] are interpreted as the highest spin state among those with $q_w=N_c-1$, $\ell=0$, $n_\rho=n_z=n_{j'}^{{\overline}\Psi}=n_k^\Phi=0$ and $n_{j'}^\Psi=\delta_{j'j}$ for some $j$. These states are most likely to be the lightest state among the highest spin states with isospin 1/2 for each level. Let us discuss if the quantum numbers and the masses of these states are consistent with the experimental data with this interpretation.
The states we consider are labeled uniquely by the level ${{\mathcal N}}$ introduced in (\[cN\]). Let $E_{{\mathcal N}}$ denote the baryon mass for a given ${{\mathcal N}}$. The nucleon corresponds to the case ${{\mathcal N}}=0$, which has $J^P=1/2^+$ and the mass given by $$\begin{aligned}
M_{{{\mathcal N}}=0}=M_0^\ast(q_w=N_c,\ell=1) \ .
\label{EN0}\end{aligned}$$ Here, $M_0^\ast$ is considered to be a function of $q_w$ and $\ell$ as argued in section \[bmassf\]. As it is technically hard to compute the quantum $M_0^\ast(q_w,\ell)$, we regard it as an unknown parameter.
For ${{\mathcal N}}\ge 1/2$, because $\ell=0$, the massless sector has vanishing spin and isospin. Then, the total spin of the excited baryons with ${{\mathcal N}}\ge 1/2$ is fixed by the massive sector. Let the excitation number of the excited 8-4 string be $N^{(j)}_{84}$, which is to be identified with the level ${{\mathcal N}}$ for the excited nucleons as seen before. For each ${{\mathcal N}}=1/2,1,3/2,2,\cdots$, the highest spin states are contained in the states of the form $$\begin{aligned}
\left(
\alpha_{-1/2}^{M_1}\cdots \alpha_{-1/2}^{M_{2{{\mathcal N}}}}
-(\mbox{trace parts})\right)\ket{a,I}_{{\mathrm{NS}}} \ ,
\label{highest59}\end{aligned}$$ which belongs to the spin $({{\mathcal N}},{{\mathcal N}})\otimes (1/2,0)$ representation of $SU(2)_L\times SU(2)_R$. Here, we have included the flavor index $I$ to show that it is an isospin $1/2$ state for $N_f=2$. Decomposing this under the vector-like subgroup $SU(2)_{J}\subset SU(2)_L\times SU(2)_R$, one finds that the highest spin is given by $J=2{{\mathcal N}}+1/2$. The parity of these excited nucleons are given by $P=(-)^{2{{\mathcal N}}}$, because the state (\[highest59\]) has parity $(-)^{2{{\mathcal N}}}$ and the massless sector is parity even for $n_z=0$. Therefore, the spin, isospin and parity for the excited nucleon states constructed above are consistent with those in Table \[Nregge\].
The baryon mass formula (\[mass\]) implies that the masses for these excited nucleons with $J\ge 3/2$ states are $$\begin{aligned}
M_{{{\mathcal N}}}=
M_0^\prime+\sqrt{\frac{{{\mathcal N}}}{\alpha'}}
=
M_0^\prime+\frac{1}{\sqrt{2\alpha^\prime}}
\sqrt{J-\frac{1}{2}} \ ,
\label{EN}\end{aligned}$$ where, $M_0^\prime\equiv M_0^\ast(q_w=N_c-1,\ell=0)$. This formula can be recast as a formula for spin $J$ as a function of mass $M$: $$\begin{aligned}
J=2\alpha^\prime(M-M_0^\prime)^2+\frac{1}{2} \ .
\label{JE}\end{aligned}$$
It has been observed that, when the spin $J$ is plotted as a function of the mass squared $M^2$, the excited nucleon states listed in Table \[Nregge\] lies on a linear trajectory that satisfies $$\begin{aligned}
J=\alpha_0+\alpha^\prime M^2 \ .
\label{linear}\end{aligned}$$ with $\alpha_0|_{\rm exp}\simeq -0.3$ and $\alpha^\prime|_{\rm exp}\simeq 0.9\,{\rm GeV}^{-2}$. Our formula (\[JE\]) is a nonlinear function with respect to $M^2$, and one would think it disagrees with the observation. However, choosing $$\begin{aligned}
\alpha'\simeq 0.6\,{\rm GeV}^{-2}\ ,~~~M_0'\simeq 0.5\,{\rm GeV}\ ,
\label{alpha_m0}\end{aligned}$$ we get a plot in Figure \[plot\], which shows that it can fit the data reasonably well.
Due to the nonlinear term in (\[JE\]), the trajectory in Figure \[plot\] is curved toward the left and the value of mass squared for $J=1/2$ becomes significantly smaller compared to that of the nucleons (proton or neutron). This is, however, not a problem of the formula (\[JE\]) as it is derived for the states with $J\ge 3/2$. Our expression for the nucleon mass is given in (\[EN0\]). Though we are not able to predict its value, this observation suggests that the difference between (\[EN0\]) and $M_0'$ $$\begin{aligned}
M_{{{\mathcal N}}=0}-M_0'=
M_0^*(q_w=N_c,\ell=1)-M_0^*(q_w=N_c-1,\ell=0) \ ,\end{aligned}$$ is positive, as we have expected.[^23]
We emphasize that the values (\[alpha\_m0\]) should not be considered to be an accurate estimate, because we have neglected all the $1/N_c$ and $1/\lambda$ corrections, as well as the possible contributions from the interaction term (\[Lint\]) for the massive fields. Nevertheless, let us here make a few comments on the value of $\alpha'$. In [@SS1; @SS2], the parameters ${M_{\rm KK}}$ and $\lambda$ was chosen to be $$\begin{aligned}
{M_{\rm KK}}\simeq 949\,{\rm MeV}\ ,~~\lambda\simeq 16.6 \ ,
\label{mkklambda}\end{aligned}$$ to fit the experimental values of the $\rho$-meson mass and the pion decay constant. If we use these values and the relation (\[lslam\]), we obtain $\alpha^\prime\simeq 0.452\,{\rm GeV}^{-2}$, which is a bit small compared with the value in (\[alpha\_m0\]). On the other hand, the value of $\alpha'$ evaluated from the Regge slope of the $\rho$-meson trajectory is $\alpha'|_{\rm exp}\simeq 0.88~{\rm GeV}^{-2}$. In [@HSS], the $\rho$-meson Regge behavior is analyzed theoretically using the same holographic model of QCD as in the present paper. It was argued there that the $\rho$-meson trajectory has some nonlinear corrections similar to that in (\[JE\]) and the value of $\alpha'$ that fits well with the experimental data turned out to be around $1.1~{\rm GeV}^{-2}$. The value of $\alpha'$ in (\[alpha\_m0\]) is close to neither of these values, though it is not too far from them. It is important to resolve this discrepancy by making more accurate estimate of $\alpha^\prime$.
Note that the slope $\alpha'$ of the linear Regge trajectory (\[linear\]) for the excited nucleons is very close to that of the $\rho$-mesons. This is one of the motivations to conjecture that both of them are described by open strings with some particles attached on the end points as investigated in [@Sharov:2013tga; @cobi]. Our description is similar to these models in that only one of $N_c$ strings attached on the baryon vertex gets excited while the rest remains to be in the ground state. This system may be approximated with a single open string by regarding the effect of the baryon vertex as a massive end point. However, a clear distinction to the models in [@Sharov:2013tga; @cobi] is that the mass of the end point in the present model is of ${{\mathcal O}}(N_c)$ and considered to be much heavier than the energy scale determined by the string tension. In fact, it is not difficult to verify that a rotating open string with a massive end point of mass $M_0$ has a classical energy $E$ that reduces in the heavy end point limit to $$\begin{aligned}
J=2\alpha^\prime(E-M_0)^2 \ ,
\label{JEcl}\end{aligned}$$ which agrees with (\[JE\]) up to an additive constant $1/2$ and the contributions from the zero point energy in $M_0'$.[^24] We note that the difference between the mass formula (\[JE\]) and the (\[JEcl\]) is due to quantum $1/N_c$ corrections.
More about excited baryon states
--------------------------------
In this subsection, we show some examples of low-lying excited baryons that are obtained in a manner explained in the previous sections. For simplicity, we set $(n_\rho,n_z)=(0,0)$ and $n_\beta^a=0$. The states in the massless and massive sectors are denoted by $|\ell,q_w\rangle$ and $|n_j^\Psi,n_j^{{\overline}\Psi},n_k^\Phi\rangle$, respectively, where only nonvanishing quantum numbers are indicated explicitly for notational simplicity.
We start from the sector ${{\mathcal N}}=1/2$. This sector is constructed only from the excitation of a single 8-4 string with $N_{84}^{(1)}=1/2$, because any 4-4 excited state has $N_{44}^{(k)}\ge 1$.[^25] The corresponding field $\Psi_1$ belongs to $(1,1/2)_-$ under $SU(2)_L\times SU(2)_R$ with the subscript denoting parity, and yields an harmonic oscillator with the angular frequency given by $m_1=\sqrt{1/(2\alpha')}$. The condition ${{\mathcal N}}_{84}=1/2$ is satisfied when $(n_1^\Psi,n_1^{{\overline}\Psi})=(1,0)$ or $(0,1)$. The excited states $|n_1^\Psi=1\rangle$ and $|n_1^{{\overline}\Psi}=1\rangle$ have the same energy eigenvalue of $H_m$ with $SU(2)_L\times SU(2)_R$ spin given by $(1,1/2)_-$. We consider only the former, because this leads to $q_w=N_c-1$ so that (\[M0cl\]) shows that the corresponding massless sector have less energy compared to that with $q_w=N_c+1$, which corresponds to $|n_1^{{\overline}\Psi}=1\rangle$. As $q_w$ is even, the massless sector is allowed to have $\ell=0,2,4,\cdots$. We first consider the case $\ell=0$, which yields the lightest state in the massless sector $\ket{\ell=0,q_w}$, which belongs to the trivial $SU(2)_L\times SU(2)_I$ representation and has even parity because $n_z=0$. Hence, the tensor product state of $\ket{\ell=0,q_w}$ with $|n_1^{\Psi}=1\rangle$ has $SU(2)_L\times SU(2)_R\times SU(2)_I$ spin given by $$\begin{aligned}
(1,1/2)^{1/2}_-\otimes(0,0)^0_+=(1,1/2)^{1/2}_- \ ,\end{aligned}$$ where the superscripts represent the isospin. The tensor product state of $\ket{\ell=2,q_w}$ with $|n_1^{\Psi}=1\rangle$ decomposes under $SU(2)_L\times SU(2)_R\times SU(2)_I$ as $$\begin{aligned}
(1,1/2)^{1/2}_-\otimes(1,0)^1_+=
\left[(2,1/2)\oplus(1,1/2)\oplus(0,1/2)\right]^{3/2}_-
\oplus
\left[(2,1/2)\oplus(1,1/2)\oplus(0,1/2)\right]^{1/2}_-
\ .{\nonumber}\end{aligned}$$ It is straightforward to decompose all these states in terms of $SU(2)_J\subset SU(2)_L\times SU(2)_R$. The results are summarized in Table \[cN1/2\].
product states $SU(2)_L\times SU(2)_R\times SU(2)_I$ $SU(2)_J\times SU(2)_I$
----------------------------------------------------- ---------------------------------------------------------------- ---------------------------------------------------
$|\ell=0,q_w=N_c-1\rangle\otimes|n_1^\Psi=1\rangle$ $(1,1/2)^{1/2}_-$ $(3/2)^{1/2}_-\oplus(1/2)^{1/2}_-$
$|\ell=2,q_w=N_c-1\rangle\otimes|n_1^\Psi=1\rangle$ $\left[(2,1/2)\oplus(1,1/2)\oplus(0,1/2)\right]^{3/2}_-$ $[(5/2)\oplus 2(3/2)\oplus 2(1/2)]^{3/2}_-$
$\oplus\left[(2,1/2)\oplus(1,1/2)\oplus(0,1/2)\right]^{1/2}_-$ $\oplus[(5/2)\oplus 2(3/2)\oplus 2(1/2)]^{1/2}_-$
Note that $(3/2)^{1/2}_-$ appearing at the first row is identified with N(1520) in the previous section.
We next turn to discussing the ${{\mathcal N}}=1$ states. This is possible only when $({{\mathcal N}}_{84},{{\mathcal N}}_{44})=(1,0)$ or $(0,1)$. The first condition is further divided into two cases: (i) $(n_1^\Psi,n_1^{{\overline}\Psi})=(2,0),(1,1),(0,2)$ and (ii) $(n_j^\Psi,n_j^{{\overline}\Psi})=(1,0),(0,1)$ for $j=2,3,4$. Note that the 8-4 massive states with $j=2,3,4$ are given by the three states with $N_{84}=1$ listed in Table \[table:95summary\]. Again, we focus on the lightest states in each case, implying that we pick up only $(n_1^\Psi,n_1^{{\overline}\Psi})=(2,0)$ and $(n_j^\Psi,n_j^{{\overline}\Psi})=(1,0)$ with $j=2,3,4$. The second condition $({{\mathcal N}}_{84},{{\mathcal N}}_{44})=(0,1)$ is solved by $n_k^{\Phi}=1$ with $k=1,2,3,4$ with the rest of the excitation numbers set to be zero. Note that the 4-4 string states labeled by $k=1,2,3,4$ are given by the excited states with $N_{44}=1$ shown in Table \[table:55summary\]. As no excitation is made by any 8-4 string mode, this case gives $q_w=N_c$.
Let us now work out the baryons states for the above three cases. For the first case, the state in the massive sector is given by $|n_1^\Psi=2\rangle$, which transforms under $SU(2)_L\times SU(2)_R\times SU(2)_I$ as $$\begin{aligned}
\left[(1,1/2)_-^{1/2}\otimes (1,1/2)_-^{1/2}\right]_{\rm symmetrized}
=\left[
(0,1)^1\oplus
(0,0)^0\oplus
(2,1)^1\oplus
(2,0)^0\oplus
(1,1)^0\oplus
(1,0)^1
\right]_+ \ .
\label{nn20}\end{aligned}$$ The massless sector for this case is characterized by $q_w=N_c-2={\rm odd}$. We are thus allowed to set $\ell=1$ as the lightest state, whose $SU(2)_L\times SU(2)_R\times SU(2)_I$ spin is given by $(1/2,0)_+^{1/2}$. By taking the tensor product of this state $\ket{\ell=1,q_w=N_c-2}$ with $|n_1^\Psi=2\rangle$, we find the baryon states listed below $$\begin{aligned}
& [
(1/2,0)\oplus
(1/2,1)\oplus
(3/2,0)\oplus
(3/2,1)\oplus
(5/2,1)
]_+^{3/2}
{\nonumber}\\
&\oplus
[
2(1/2,0)\oplus
2(1/2,1)\oplus
2(3/2,0)\oplus
2(3/2,1)\oplus
(5/2,0)\oplus
(5/2,1)
]_+^{1/2}\ .\end{aligned}$$
For the second case, we take the massive sector state to be $|n_j^\Psi=1\rangle$ with $j=2,3,4$. This corresponds to $q_w=N_c-1={\rm even}$. We can take $\ell=0,2,4\cdots$. The state $\ket{\ell=0,q_w}$ has a trivial spin so that the tensor product of this state with $|n_j^\Psi=1\rangle$ has the same spin as that of $|n_j^\Psi=1\rangle$. The massless sector with $\ell=2$ has $SU(2)_L\times SU(2)_R\times SU(2)_I$ spin given by $(1,0)^1$. The tensor product of this state with $|n_j^\Psi=1\rangle$ is easy to evaluate for each $j=2,3,4$.
Finally, the massive sector for the third case is characterized by the four states $|n_k^\Phi=1\rangle$ with $k=1,2,3,4$. As noted before, this corresponds to $q_w=N_c={\rm odd}$ so that odd $\ell$ is allowed. We pick up $\ell=1$, which is expected to give the lightest state among those with odd $\ell$, and take its tensor product with $|n^\Phi_k=1\rangle$. Note that any 4-4 string state has a vanishing isospin. The same computation is easy to perform for the next lightest state with $\ell=3$.
All the results are summarized in Table \[cN1\].
---------------------------------------------------------------------------------------------------
product states $SU(2)_L\times SU(2)_R\times SU(2)_I$
----------------------------------------------------- ---------------------------------------------
$|\ell=1,q_w=N_c-2\rangle\otimes|n_1^\Psi=2\rangle$ $
[
(1/2,0)\oplus
(1/2,1)\oplus
(3/2,0)\oplus
(3/2,1)\oplus
(5/2,1)
]_+^{3/2}
$
$
\oplus[
2(1/2,0)\oplus
2(1/2,1)\oplus
2(3/2,0)\oplus
2(3/2,1)\oplus
(5/2,0)\oplus
(5/2,1)
]_+^{1/2}
$
$|\ell=0,q_w=N_c-1\rangle\otimes|n_2^\Psi=1\rangle$ $
(1/2,0)^{1/2}_+
$
$|\ell=0,q_w=N_c-1\rangle\otimes|n_3^\Psi=1\rangle$ $
(1/2,1)_+^{1/2}
$
$|\ell=0,q_w=N_c-1\rangle\otimes|n_4^\Psi=1\rangle$ $
(3/2,1)_+^{1/2}
$
$|\ell=2,q_w=N_c-1\rangle\otimes|n_2^\Psi=1\rangle$ $
[(3/2,0)\oplus(1/2,0)]^{3/2}_+
\oplus
[(3/2,0)\oplus(1/2,0)]^{1/2}_+
$
$|\ell=2,q_w=N_c-1\rangle\otimes|n_3^\Psi=1\rangle$ $
[(3/2,1)\oplus(1/2,1)]^{3/2}_+
\oplus
[(3/2,1)\oplus(1/2,1)]^{1/2}_+
$
$|\ell=2,q_w=N_c-1\rangle\otimes|n_4^\Psi=1\rangle$ $
[(5/2,1)\oplus(3/2,1)\oplus(1/2,1)]^{3/2}_+
\oplus
[(5/2,1)\oplus(3/2,1)\oplus(1/2,1)]^{1/2}_+
$
$|\ell=1,q_w=N_c\rangle\otimes|n_1^\Phi=1\rangle$ $
(1/2,0)^{1/2}_+
$
$|\ell=1,q_w=N_c\rangle\otimes|n_2^\Phi=1\rangle$ $
(1/2,0)^{1/2}_+
$
$|\ell=1,q_w=N_c\rangle\otimes|n_3^\Phi=1\rangle$ $
\left[ (1,1/2)\oplus (0,1/2)\right]^{1/2}_-
$
$|\ell=1,q_w=N_c\rangle\otimes|n_4^\Phi=1\rangle$ $
[(3/2,1)\oplus(1/2,1)]^{1/2}_+
$
$|\ell=3,q_w=N_c\rangle\otimes|n_1^\Phi=1\rangle$ $
(3/2,0)^{3/2}_+
$
$|\ell=3,q_w=N_c\rangle\otimes|n_2^\Phi=1\rangle$ $
(3/2,0)^{3/2}_+
$
$|\ell=3,q_w=N_c\rangle\otimes|n_3^\Phi=1\rangle$ $
\left[
(2,1/2)
\oplus (1,1/2)
\right]^{3/2}_-
$
$|\ell=3,q_w=N_c\rangle\otimes|n_4^\Phi=1\rangle$ $
[(5/2,1)
\oplus(3/2,1)
\oplus(1/2,1)
]^{3/2}_+
$
---------------------------------------------------------------------------------------------------
Decomposing these states in terms of $SU(2)_J\subset SU(2)_L\times SU(2)_R$ is straightforward.
Now we discuss possible identifications of the states listed in Tables \[cN1/2\] and \[cN1\] with the baryons found in the experiments. Because we haven’t been able to derive the $\ell$ dependence in the baryon mass formula (\[mass\]), we have to rely on some qualitative arguments. Our guiding principles are as follows. First, we expect that the states with the same $\ell,q_w$ and ${{\mathcal N}}$ are nearly degenerate. Second, for a given $(\ell,q_w)$, the states with ${{\mathcal N}}=1$ are heavier than those with ${{\mathcal N}}=1/2$. Third, for a given ${{\mathcal N}}$, the mass is an increasing function of both $\ell$ and $q_w$ except for the state with $n_1^\Psi=2$ listed at the first row in Table \[cN1\], which is expected to be heavier than the others according to the baryon mass formula (\[mass\]). [^26]
The predictions for the low-lying excited baryons with $I=1/2$ are summarized in Table \[I1/2\], whose data are taken from Tables \[cN1/2\] and \[cN1\].
----------------------------------------------------------------------------------------------------------------------------------------------
level states $SU(2)_J\times SU(2)_I$
---------------------- ----------------------------------------------------------- -----------------------------------------------------------
${{\mathcal N}}=1/2$ $|\ell=0,q_w=N_c-1\rangle\otimes|n_1^\Psi=1\rangle$ $[{\color{blue} (3/2)}\oplus (1/2)]_-^{1/2}$
$|\ell=2,q_w=N_c-1\rangle\otimes|n_1^\Psi=1\rangle$ $[(5/2)\oplus 2(3/2)\oplus 2(1/2)]_-^{1/2}$
${{\mathcal N}}=1$ $|\ell=0,q_w=N_c-1\rangle\otimes|n_{2,3,4}^\Psi=1\rangle$ $[{\color{blue}(5/2)}\oplus 2(3/2)\oplus 3(1/2)]_+^{1/2}$
$|\ell=1,q_w=N_c\rangle\otimes|n_{1,2,3,4}^\Phi=1\rangle$ $
[(5/2)\oplus 2(3/2)\oplus 4(1/2)]_+^{1/2}
\oplus
[(3/2)\oplus 2(1/2)]_-^{1/2}
$
$|\ell=2,q_w=N_c-1\rangle\otimes|n_{2,3,4}^\Psi=1\rangle$ $
[(7/2)\oplus 3(5/2)\oplus
6(3/2)\oplus 5(1/2)]_+^{1/2}
$
----------------------------------------------------------------------------------------------------------------------------------------------
Here, we will not attempt to relate the states with $J=1/2$ in this table to those in the baryon summary table [@pdg], because these might be regarded as excited states with nonvanishing $n_\rho$, $n_z$ and $n_\beta$ without excitations in the massive sector.[^27] Note that the states with $(3/2)_-^{1/2}$ and $(5/2)_+^{1/2}$ in the first and third rows in Table \[I1/2\] are identified with N(1520) and N(1680), respectively, in section \[Regge\]. The $(5/2)_-^{1/2}$ state at ${{\mathcal N}}=1/2$ is expected to be the lightest state with this quantum number and hence it may be identified with N(1675), which is the lightest baryon with the same quantum number listed in the baryon summary table. Then, the $(3/2)_-^{1/2}$ states at the second row are expected to have mass nearly equal to N(1675). A natural candidate for one of them is N(1700).[^28]
As for the ${{\mathcal N}}=1$ states, we find that the $(3/2)_+^{1/2}$ states at the third row are expected to have mass nearly equal to N(1680). A natural candidate for one of them is N(1720).[^29] Since the fourth row has larger values of $\ell$ and $q_w$ compared with the third row, the $(5/2)_+^{1/2}$ state at the fourth row is expected to be heavier than N(1680) and N(1720). A natural candidate for it is N(1860), though this state is not established in the experiments. If this is the case, the $(3/2)_\pm^{1/2}$ states at the fourth row are expected to be nearly degenerate with N(1860). These states could be identified with N(1900) and N(1875). The baryon states at the fifth row contains a state with $(7/2)_+^{1/2}$. The only baryon with this quantum number listed in the baryon summary table is N(1990), though this is not considered to be established. Then, the $(5/2)_+^{1/2}$ and $(3/2)_+^{1/2}$ states at the fifth row in Table \[cN1/2\] could be identified with N(2000) and N(2040), respectively, which are again poorly established in the experiments.
Unfortunately, the identification we have made is not a clear one-to-one correspondence. There are more than one candidate states in the model for many of the baryons listed in the baryon summary table. In particular, the degeneracy of the states in Table \[I1/2\] doesn’t match the experimental data perfectly. Furthermore, as mentioned in the footnotes, some of the baryons may be identified with the states that are not listed in Table \[I1/2\]. Lack of the one-to-one correspondence would be in part because all the excited baryons we consider are unstable resonances (for finite $N_c$) and many of them, in particular heavier ones, are probably not easy to identify in the experiment. Furthermore, some of the states in Tables \[I1/2\] and $\ref{I3/2}$ could be the artifacts of the model. Although, as discussed in section \[sec:baryonstates\], we have imposed the invariance with respect to the $SO(5)$ symmetry and $\tau$-parity to get rid of the artifacts, we are not able to show that this is sufficient to exclude all of them. It is expected that incorporation of full $1/\lambda$ corrections into the baryon mass formula makes the artifacts of the model infinitely heavy in the ${M_{\rm KK}}\rightarrow\infty$ ($\lambda\rightarrow 0$) limit with $\Lambda_{\rm QCD}$ kept fixed. However, the extrapolation to the small $\lambda$ regime is a notoriously difficult problem in the holographic description, because we have to deal with all the stringy corrections in a highly curved spacetime. A similar observation was made also in [@ISS]. We leave it as an open problem to study a dictionary between the theoretical predictions and the experimental data in more detail.
We also examine the mass spectrum of $\Delta$ baryons with isospin $I=3/2$. The theoretical predictions for this case is summarized in Table \[I3/2\], whose data is taken from Tables \[cN1/2\] and \[cN1\].
--------------------------------------------------------------------------------------------------------------------------------
level states $SU(2)_J\times SU(2)_I$
---------------------- ----------------------------------------------------------- ---------------------------------------------
${{\mathcal N}}=1/2$ $|\ell=2,q_w=N_c-1\rangle\otimes|n_1^\Psi=1\rangle$ $[(5/2)\oplus 2(3/2)\oplus 2(1/2)]_-^{3/2}$
${{\mathcal N}}=1$ $|\ell=2,q_w=N_c-1\rangle\otimes|n_{2,3,4}^\Psi=1\rangle$ $[(7/2)\oplus 3(5/2)
\oplus 6(3/2)\oplus 5(1/2)]_+^{3/2}$
--------------------------------------------------------------------------------------------------------------------------------
It is natural to identify the $(5/2)_-^{3/2}$ and $(7/2)_+^{3/2}$ states at the first and second rows in Table \[I3/2\] with the lightest $\Delta$ baryons having the same quantum numbers listed in the baryon summary table, which are $\Delta(1930)$ and $\Delta(1950)$, respectively. This suggests that the $(5/2)_+^{3/2}$ states in the second row in Table \[cN1\] are nearly degenerate with $\Delta(1950)$. A good candidate to be identified with one of these states is $\Delta(1905)$. However, this identification is problematic: Although our formula (\[mass\]) suggests that the ${{\mathcal N}}=1$ states are significantly heavier than ${{\mathcal N}}=1/2$ states, $\Delta(1930)$ and $\Delta(1950)$ are nearly degenerate and $\Delta(1905)$ is even lighter than $\Delta(1930)$.
Conclusions
===========
In this paper, we have discussed stringy excited baryons using the holographic dual of QCD on the basis of an intersecting D4/D8-brane system. A key step to this end is to work on the whole system of a baryon vertex without describing it by a topological soliton on an effective 5 dimensional gauge theory. We formulated this system as a many-body quantum mechanics that is composed of the ADHM-type matrix model of Hashimoto-Iizuka-Yi [@HIY] and an infinite number of open string massive modes. This is done by relying on an approximation that is valid in the large $N_c$ and $\lambda$ regime. The resultant quantum mechanics provides us with a powerful framework for making a systematic analysis of excited baryons including those with $I\ne J$ that are difficult to obtain in the soliton picture.
By construction, it would be too ambitious that the theoretical predictions from the present model match the experimental data to good accuracy. Interestingly, we have seen that the present model reproduces a qualitative feature of the nucleon Regge trajectory. It has been argued that the stringy excited baryons to be identified with the excited nucleons are interpreted as a rotating open string with a massive end point. Such a picture of baryon Regge trajectories has been studied extensively so far in the literature. [@Sharov:2013tga; @cobi] It is worth emphasizing that the massive end point in this model is due to a D4$_{\rm BV}$, having mass of ${{\mathcal O}}(N_c)$. The Regge trajectory formula (\[JE\]) that we proposed in this paper is not given by a simple, linear relation between the spin and the mass squared because of the heavy end point.
We conclude this paper by making some comments about future directions. First, it is important to improve a theoretical accuracy of the model by incorporating the interacting terms in $L_{\rm int}$ that have been neglected for technical difficulties. It would be almost impossible to fix the mass terms of the mass fields $\Psi_j$ and $\Phi_k$ precisely, because an infinitely many higher-order terms could contribute to a single mass term as discussed in section \[subsec:Lint\]. Instead, what may be performed immediately is to take into account the effects of the mixing terms like $\Psi_j\Psi_j$ into the baryon mass formula. With these mixing terms, $q_j$ is not a conserved charge any more so that an exact diagonalization of $H_m$ in a manner consistent with the Gaussian constraint appears highly involved. It is interesting to compute a perturbative effect of the mixing terms into the mass formula.
One of the unsatisfactory points is that the values of the parameters $\gamma$ and $v$ in the potential (\[V0\]) are not determined from the first principle. Though it is possible to adjust them to fit the results in the soliton picture as in [@HSSY], a derivation within our framework is desired to make sure that all the parameters can be fixed, in principle, without any ambiguities. Compared with the soliton picture, the origin of the potential (\[V0\]) is expected to be due to the energy contribution from the $U(N_f)$ gauge field on the flavor D8-branes in the presence of a baryon vertex. It would be interesting to examine this in more detail.
Finally, it would be of great interest to apply the results in this paper to a more complicated system that are made out of multiple baryon and anti-baryon vertices. A typical example is given by a stringy realization of tetraquarks. It is nice to try to formulate a holographic model for tetraquarks following this paper and compare the theoretical predictions with experiments.
Acknowledgements {#acknowledgements .unnumbered}
================
We would like to thank K. Hashimoto, S. Hirano and J. Sonnenschein for useful discussions. The work of SS has been supported by JSPS KAKENHI (Grant-in-Aid for Scientific Research (C)) Grant Number JP16K05324 and (Grant-in-Aid for Scientific Research (B)) Grant Number JP19H01897.
$SO(4)\simeq SU(2)_I\times SU(2)_J$ {#SO4}
===================================
The generators of the Lie algebra of $SO(4)\simeq (SU(2)_I\times SU(2)_J)/{{\mathbb Z}}_2$ can be chosen as $$\begin{aligned}
i\Sigma^1&=&i\sigma_2\otimes\sigma_1
=\mat{,\sigma_1,-\sigma_1,}\ ,{\nonumber}\\
i\Sigma^2&=&-i\sigma_2\otimes\sigma_3
=\mat{,-\sigma_3,\sigma_3,}\ ,{\nonumber}\\
i\Sigma^3&=&i{\bf 1}_2\otimes\sigma_2
=\mat{i\sigma_2,,,i\sigma_2}\ ,\end{aligned}$$ $$\begin{aligned}
i{\widetilde}\Sigma^1&=&-i\sigma_1\otimes\sigma_2
=\mat{,-i\sigma_2,-i\sigma_2,}\ ,{\nonumber}\\
i{\widetilde}\Sigma^2&=&-i\sigma_2\otimes{\bf 1}_2
=\mat{,-{\bf 1}_2,{\bf 1}_2,}\ ,{\nonumber}\\
i{\widetilde}\Sigma^3&=&i\sigma_3\otimes\sigma_2
=\mat{i\sigma_2,,,-i\sigma_2}\ .\end{aligned}$$
$\{\Sigma^a\}$ and $\{{\widetilde}\Sigma^a\}$ satisfy the same algebra as the Pauli matrices; $$\begin{aligned}
\Sigma^a\Sigma^b=\delta^{ab}+i\epsilon^{abc}\Sigma^c\ ,~~~
{\widetilde}\Sigma^a{\widetilde}\Sigma^b=\delta^{ab}+i\epsilon^{abc}{\widetilde}\Sigma^c\ ,\end{aligned}$$ and they commute with each other; $$\begin{aligned}
\Sigma^a{\widetilde}\Sigma^b&=&{\widetilde}\Sigma^b\Sigma^a\ .\end{aligned}$$ $\{i\Sigma^a\}_{a=1,2,3}$ and $\{i{\widetilde}\Sigma^a\}_{a=1,2,3}$ are the generators of $SU(2)_I$ and $SU(2)_J$, respectively.
[99]{}
O. Aharony, S. S. Gubser, J. M. Maldacena, H. Ooguri and Y. Oz, “Large N field theories, string theory and gravity,” Phys. Rept. [**323**]{}, 183 (2000) \[hep-th/9905111\].
T. Sakai and S. Sugimoto, “Low energy hadron physics in holographic QCD,” Prog. Theor. Phys. [**113**]{}, 843 (2005) \[hep-th/0412141\]. T. Sakai and S. Sugimoto, “More on a holographic dual of QCD,” Prog. Theor. Phys. [**114**]{}, 1083 (2005) \[hep-th/0507073\]. T. Imoto, T. Sakai and S. Sugimoto, “Mesons as Open Strings in a Holographic Dual of QCD,” Prog. Theor. Phys. [**124**]{}, 263 (2010) \[arXiv:1005.0655 \[hep-th\]\]. T. H. R. Skyrme, “A Nonlinear field theory,” Proc. Roy. Soc. Lond. A [**260**]{}, 127 (1961). G. S. Adkins, C. R. Nappi and E. Witten, “Static Properties of Nucleons in the Skyrme Model,” Nucl. Phys. B [**228**]{}, 552 (1983). H. Hata, T. Sakai, S. Sugimoto and S. Yamato, “Baryons from instantons in holographic QCD,” Prog. Theor. Phys. [**117**]{}, 1157 (2007) \[hep-th/0701280 \[HEP-TH\]\]. K. Hashimoto, T. Sakai and S. Sugimoto, “Holographic Baryons: Static Properties and Form Factors from Gauge/String Duality,” Prog. Theor. Phys. [**120**]{} (2008) 1093 \[arXiv:0806.3122 \[hep-th\]\]. D. K. Hong, M. Rho, H. U. Yee and P. Yi, “Chiral Dynamics of Baryons from String Theory,” Phys. Rev. D [**76**]{} (2007) 061901 \[hep-th/0701276 \[HEP-TH\]\], D. K. Hong, M. Rho, H. U. Yee and P. Yi, “Dynamics of baryons from string theory and vector dominance,” JHEP [**0709**]{} (2007) 063 \[arXiv:0705.2632 \[hep-th\]\]. H. Hata, M. Murata and S. Yamato, “Chiral currents and static properties of nucleons in holographic QCD,” Phys. Rev. D [**78**]{} (2008) 086006 \[arXiv:0803.0180 \[hep-th\]\]. K. Y. Kim and I. Zahed, “Electromagnetic Baryon Form Factors from Holographic QCD,” JHEP [**0809**]{} (2008) 007 \[arXiv:0807.0033 \[hep-th\]\].
K. Hashimoto, Y. Matsuo and T. Morita, “Nuclear states and spectra in holographic QCD,” JHEP [**1912**]{}, 001 (2019) \[arXiv:1902.07444 \[hep-th\]\].
K. Hashimoto, N. Iizuka and P. Yi, “A Matrix Model for Baryons and Nuclear Forces,” JHEP [**1010**]{}, 003 (2010) \[arXiv:1003.4988 \[hep-th\]\].
E. Witten, “Baryons and branes in anti-de Sitter space,” JHEP [**9807**]{}, 006 (1998) \[hep-th/9805112\].
D. J. Gross and H. Ooguri, “Aspects of large N gauge theory dynamics as seen by string theory,” Phys. Rev. D [**58**]{} (1998) 106002 \[hep-th/9805129\]. E. Witten, “Small instantons in string theory,” Nucl. Phys. B [**460**]{}, 541 (1996) \[hep-th/9511030\]. M. R. Douglas, “Branes within branes,” NATO Sci. Ser. C [**520**]{}, 267 (1999) \[hep-th/9512077\]. M. F. Atiyah, N. J. Hitchin, V. G. Drinfeld and Y. I. Manin, “Construction of Instantons,” Phys. Lett. A [**65**]{} (1978) 185.
E. Witten, “Anti-de Sitter space, thermal phase transition, and confinement in gauge theories,” Adv. Theor. Math. Phys. [**2**]{}, 505 (1998) doi:10.4310/ATMP.1998.v2.n3.a3 \[hep-th/9803131\].
T. Imoto, T. Sakai and S. Sugimoto, “O(N(c)) and USp(N(c)) QCD from String Theory,” Prog. Theor. Phys. [**122**]{} (2010) 1433 \[arXiv:0907.2968 \[hep-th\]\]. R. C. Brower, S. D. Mathur and C. I. Tan, “Glueball spectrum for QCD from AdS supergravity duality,” Nucl. Phys. B [**587**]{}, 249 (2000) \[hep-th/0003115\].
K. Hashimoto and N. Iizuka, “Nucleon Statistics in Holographic QCD : Aharonov-Bohm Effect in a Matrix Model,” Phys. Rev. D [**82**]{} (2010) 105023 \[arXiv:1006.3612 \[hep-th\]\].
M. Tanabashi [*et al.*]{} \[Particle Data Group\], “Review of Particle Physics,” Phys. Rev. D [**98**]{}, no. 3, 030001 (2018).
S. S. Afonin, “Parity Doubling in Particle Physics,” Int. J. Mod. Phys. A [**22**]{} (2007) 4537 \[arXiv:0704.1639 \[hep-ph\]\].
G. S. Sharov, “String Models, Stability and Regge Trajectories for Hadron States,” arXiv:1305.3985 \[hep-ph\]. J. Sonnenschein and D. Weissman, “A rotating string model versus baryon spectra,” JHEP [**1502**]{} (2015) 147 \[arXiv:1408.0763 \[hep-ph\]\], J. Sonnenschein, “Holography Inspired Stringy Hadrons,” Prog. Part. Nucl. Phys. [**92**]{}, 1 (2017) \[arXiv:1602.00704 \[hep-th\]\], J. Sonnenschein and D. Weissman, “Quantizing the rotating string with massive endpoints,” JHEP [**1806**]{}, 148 (2018) \[arXiv:1801.00798 \[hep-th\]\]. G. S. Sharov, “Instability of classic rotational motion for three string baryon model,” hep-ph/0001154, G. S. Sharov, “Instability of the Y string baryon model within classical dynamics,” Phys. Atom. Nucl. [**65**]{} (2002) 906 \[Yad. Fiz. [**65**]{} (2002) 938\]. G. ’t Hooft, “Minimal strings for baryons,” hep-th/0408148, E. Santopinto, “An Interacting quark-diquark model of baryons,” Phys. Rev. C [**72**]{} (2005) 022201 \[hep-ph/0412319\], C. L. Gutierrez and M. De Sanctis, “A relativistic quark-diquark model for the nucleon,” Pramana [**72**]{} (2009) 451, J. Ferretti, A. Vassallo and E. Santopinto, “Relativistic quark-diquark model of baryons,” Phys. Rev. C [**83**]{} (2011) 065204, C. Gutierrez and M. De Sanctis, “A study of a relativistic quark-diquark model for the nucleon,” Eur. Phys. J. A [**50**]{} (2014) no.11, 169.
[^1]: See also [@Hong:2007kx; @HMY; @Kim:2008pw].
[^2]: For another approach to holographic baryons with $I\ne J$, see [@HMM], which is based on the study of a matrix model formulated in [@HIY].
[^3]: The radial coordinate $r$ is related to $U/U_{\rm KK}$ used in [@SS1; @SS2] by $(U/U_{\rm KK})^3=1+r^2$.
[^4]: For this, we mean that we consider $N_f$ to be of ${{\mathcal O}}(1)$ and only take into account the leading terms in the $1/N_c$ expansion.
[^5]: For a ${\mathrm{D4_{BV}}}$-brane wrapped on ${{\mathbb S}}^4$, it can be shown that $y=z=0$ is energetically favored and realized in the classical minimal energy configuration. It may be located anywhere in ${{\mathbb R}}^3\ni x^{1,2,3}$, because of the translational invariance. Here we just put it at $x^1=x^2=x^3=0$ to have a $P$ invariant configuration.
[^6]: The $\tau$-parity was originally introduced in [@BMT] in the context of glueball spectrum and then generalized to the system with quarks in [@ISS].
[^7]: In this and the following sections, we consider the original D8/${\mathrm{D4_{BV}}}$ system, rather than the T-dualized version (D9/${\mathrm{D5_{BV}}}$ system) considered in the previous section. Therefore, 9-5 and 5-5 strings in the previous section correspond to 8-4 and 4-4 strings, respectively.
[^8]: There is a mass parameter ${M_{\rm KK}}$ that gives the mass scale of the model. We mainly work in the ${M_{\rm KK}}=1$ unit. The ${M_{\rm KK}}$ dependence can be easily recovered by the dimensional analysis.
[^9]: One motivation to add these terms is to accommodate possible additional energy contributions from the gauge fields on the D8-branes. The second and third terms in (\[V0\]) mimic the $\rho$ dependent energy contributions from the gauge fields in [@HSSY]. Note that we should not trust this potential near $w=0$ when $v\ne 0$, since the third term in (\[V0\]) diverges at $w=0$. As we will see in sections \[Nf2\] and \[LargeN\], the wavefunctions of the baryon states that we are mostly interested in are peaked away from $w=0$ and we expect that it does not affect the main features of the analysis.
[^10]: $q_w$ and $q_j$ can be negative. The sign reflects the orientation of the fundamental string attached on the ${\mathrm{D4_{BV}}}$-brane.
[^11]: See [@Hashimoto:2010rb; @HMM] for the related discussions.
[^12]: Here, we discuss the cases with $0\le q_w\le N_c$ for simplicity. Other cases can also be discussed in a similar way.
[^13]: Using the relation $a^2=1$, one can show that $\mbox{\boldmath $a$}\equiv a_01_2+i\vec a\cdot\vec\tau$ is an element of $SU(2)$. This $\mbox{\boldmath $a$}$ is also related to the collective coordinate of Skyrmion for $N_f=2$.[@ANW]
[^14]: The notation $y$ and ${\widetilde}y$ in this section should not be confused with those in section \[subsec:states\].
[^15]: This is equivalent to writing down the Lagrangian in terms of the canonically normalized fields ${\widetilde}{\delta\rho}\equiv M_0^{1/2}{\delta\rho}$ and ${\widetilde}\beta_a\equiv M_0^{1/2}\beta_a$ and taking the large $N_c$ limit with these fields kept finite. On the other hand, $a$ satisfies $a^2=1$ by definition and hence we regard it as an order 1 variable. We also assume here that quantum numbers for the baryon state such as spin and isospin are all order 1, except for $q_w$ which is assumed to be of order $N_c$ as discussed around (\[qw\]).
[^16]: As pointed out in [@HSSY], a similar problem also appears in the soliton approach.
[^17]: Although $X^z$, ${\delta\rho}$ and $\beta_a$ have mass terms in the Hamiltonian (\[H0\]) and (\[H0rhobeta\]), we consider them to be in the massless sector, because these modes originate from the the massless open string states in the flat spacetime limit.
[^18]: If we set $N_c=3$ in the mass formula (\[HSSYmass\]) given in [@HSSY], the expansion as (\[HSSYmass2\]) is not justified for $\ell>1$. This suggests that the $\ell$ dependence is actually important to compare with the realistic QCD. (See [@HSSY] for further discussion.)
[^19]: See, e.g., [@Afonin:2007mj] for a review.
[^20]: $N_{84}$ and $N_{44}$ correspond to $N_{95}$ and $N_{55}$ in section \[sec:baryonstates\], respectively.
[^21]: For earlier and closely related works, see [@3strings]. See also [@diquark] for related works based on quark-diquark models.
[^22]: Here, we consider $N_c$ to be a large odd number. Recall that the condition $\ell\equiv q_w$ (mod 2) has to be satisfied. (See section \[bmassf\].)
[^23]: To get a rough estimate, one could try to evaluate it by assuming that $\ell$ dependence is small and the mass difference $\Delta M_0^*\equiv M_{{{\mathcal N}}=0}-M_0'$ is entirely determined by (\[M0cl\]). Then, one gets $\Delta M_0^*=2 M_0\gamma(\rho_0^2|_{q_w=N_c}-\rho_0^2|_{q_w=N_c-1})$. For $\gamma=1/6$ and $v=0$, using (\[rho0\]), we get $\Delta M_0^*={M_{\rm KK}}/\sqrt{6}$, where we have recovered the ${M_{\rm KK}}$ dependence by the dimensional analysis. Using the value of ${M_{\rm KK}}$ in (\[mkklambda\]), this is estimated as 387 MeV.
[^24]: For a systematic treatment of classical motion of rotating strings with massive end points, see the third paper in [@cobi].
[^25]: $N^{(j)}_{84}$ and $N^{(k)}_{44}$ are the excitation numbers for $\Psi_j$ and $\Phi_k$, respectively. $\Psi_j$ with $j=1,2,3,4$ and $\Phi_k$ with $k=1,2,3,4$ are listed in Table \[table:95summary\] and Table \[table:55summary\], respectively.
[^26]: Here, we have assumed that $M_0^*|_{q_w=N_c}-M_0^*|_{q_w=N_c-2}$ is smaller than $(\sqrt{2}-1)/\alpha'$, which can be justified for large $\lambda$.
[^27]: Part of such states were already discussed in [@HSSY].
[^28]: There are other possibilities for this identification. For example, $\ket{\ell=0,n_\rho=1,q_w=N_c-1}\otimes\ket{n_1^\Psi=1}$ and $\ket{\ell=3,n_z=1,n_\beta=1,q_w=N_c}$ also have $(3/2)_-^{1/2}$ components that could be identified with N(1700).
[^29]: As in the case of N(1700), $\ket{\ell=0,n_z=1,q_w=N_c-1}\otimes\ket{n_1^\Psi=1}$ and $\ket{\ell=3,n_\rho=1,n_\beta=1,q_w=N_c}$ also have $(3/2)_+^{1/2}$ components that could be identified with N(1720).
|
{
"pile_set_name": "ArXiv"
}
|
---
address:
- |
Department of Physics and Astronomy, SUNY at Stony Brook,\
E-mail: [email protected]
- '$^A$ Niels Bohr Institute, Denmark; $^B$ LANL, USA; $^C$ Columbia U., USA; $^D$ Nuclear Physics Laboratory of Nantes, France; $^E$ Hiroshima U., Japan; $^F$ CERN; $^G$ Rudjer Boscovic Institute, Croatia; $^H$ Texas A&M U., USA; $^I$ The Ohio State U., USA; $^J$ SUNY at Stony Brook, USA; $^K$ U. of Lund, Sweden; $^L$ BNL, USA. '
author:
- 'Mikhail L. KOPYTINE[^1] for The NA44 Collaboration'
- 'I.Bearden$^{a}$, H.B[Ø]{}ggild$^{a}$, J.Boissevain$^{b}$, L.Conin$^{d}$, J.Dodd$^{c}$, B.Erazmus$^{d}$, S.Esumi$^{e}$, C.W.Fabjan$^{f}$, D.Ferenc$^{g}$, D.E.Fields$^{b}$, A.Franz$^{f}$, J.J.Gaardh[Ø]{}je$^{a}$, A.G.Hansen$^{a}$, O.Hansen$^{a}$, D.Hardtke$^{i}$, H. van Hecke$^{b}$, E.B.Holzer$^{f}$, T.J.Humanic$^i$, P.Hummel$^f$, B.V.Jacak$^j$, R.Jayanti$^i$, K.Kaimi$^e$, M.Kaneta$^e$, T.Kohama$^e$, M.L.Kopytine$^j$, M.Leltchouk$^c$, A.Ljubicic, Jr$^g$, B. L[ö]{}rstad$^k$, N.Maeda$^e$, L.Martin$^d$, A.Medvedev$^c$, M.Murray$^h$, H.Ohnishi$^e$, G.Paic$^f$, S.U.Pandey$^i$, F.Piuz$^f$, J.Pluta$^d$, V.Polychronakos$^l$, M.Potekhin$^c$, G.Poulard$^f$, D.Reichhold$^i$, A.Sakaguchi$^e$, J.Schmidt-S[Ø]{}rensen$^k$, J.Simon-Gillo$^b$, W.Sondheim$^b$, T.Sugitate$^e$, J.P.Sullivan$^b$, Y.Sumi$^e$, W.J.Willis$^c$, K.L.Wolf$^h$, N.Xu$^b$, D.S.Zachary$^i$.'
title: Search for Critical Phenomena in Pb+Pb Collisions
---
An ultrarelativistic collision of heavy ions presents a phenomenon whose most interesting features are conditioned by the large multitude of degrees of freedom involved, and yet offer an opportunity for the fundamental physics of the strong interaction to manifest itself. The very notion of a phase transition in such collisions is inherently of a multiparticle nature. Truly multiparticle observables, defined on event-by-event basis – a few have been constructed so far – therefore attract attention. Recently published event-by-event analyses of the 158 GeV/A $Pb+Pb$ data either deal with a small number of events [@EMU15] or analyse properties of a large ensemble of events using a single observable ($p_T$) to compare different ensemble averages[@NA49_phi]. In the first case, accumulation of feature information from large data sets remains open. In the second case [@NA49_phi], one can not establish a scale independency in event textures by observing a logical consequence thereof [@CLT]. Furthermore, an ensemble average on a set of *post-freeze-out* events is not representative of the *pre-freeze-out* history of those events, due to the dramatic non-stationarity of the open system, with a consequent lack of ergodicity.
Here we concentrate on *texture*, or *local fluctuation* observables, where a single event determines its own correlation/fluctuation content, and the scale composition thereof manifests itself in the observables in a positive way. The idea to look at particle distributions in rapidity $y$ to search for critical behaviour was proposed [@Scalapino_Sugar; @Carruthers_Sarcevic] based upon a Ginzburg-Landau type of multihadron production theory [@Scalapino_Sugar], where a random hadronic field $\phi(y)$ plays the role of an order parameter in a hadronization transition. Enhanced large scale correlations of hadrons in $y$ at the phase transition would signal critical fluctuations in the order parameter. Stephanov and coworkers [@tricritical] indicated a second order QCD phase transition point which should exist under certain initial conditions, within the reach of today’s experiments.
In our work, a power spectrum analysis of event texture in pseudorapidity $\eta$ and azimuthal angle $\zeta$, based on a Discrete Wavelet Transformation (DWT)[@DWT], is performed on a number of large event ensembles sampled according to their multiplcity, thereby studying the impact parameter dependence of the observables. DWT quantifies contributions of different $\zeta$ and $\eta$ scales into the overall event’s texture, thus testing the possible large scale enhancement.
The SPS beam was collimated to a $1\times2$ mm profile. The NA44 Si pad array, installed 10 cm downstream from the target, in the magnetic field of the first dipole[@NA44ex], measured ionization energy loss of charged particles in its 512 300 $\mu$m thick Si pads. The silicon detector had inner radius 7.2 mm and outer radius 43 mm. The detector was split *radially* into 16 rings of equal $\eta$ coverage. Each ring was divided *azimuthally* into 32 sectors of equal angular coverage to form pads. The pads were read out by AMPLEX [@AMPLEX] chips, one chip per sector. $\delta$-electrons, produced by the $Pb$ beam traversing the target, were swept away to one side by the dipole magnetic field ($\le 1.6$ Tl). Only the $\delta$-electron-free side was used in this analysis. Channel pedestals had, on the average, $FWHM=0.48 <dE>$ of 1 MIP. In the texture analysis, every event was represented by a 2D array of the calibrated digitized amplitudes of the channels ( an *amplitude array*). Empty target runs were used to measure the background. Cross-talk in the detector was evaluated off-line. DWT formalizes the images of the $PbPb$ collision events in pseudorapidity $\eta$ and azimuthal angle $\zeta$ by expanding them into a set of functions orthogonal with respect to scale and position, and allows one to accumulate the texture information by averaging the power spectra of many events. The simplest DWT basis is the Haar one, built upon the scaling function $\phi(x) = 1$ for $0\le x<1$ and 0 otherwise. If the interaction vertex lies on the detector’s symmetry axis, every pad’s acceptance is a rectangle in the $(\zeta,\eta)$ space. Then, the Haar basis is the natural one, as its scaling function in two dimenstions (2D) $\Phi(\zeta,\eta) = \phi(\zeta)\phi(\eta)$ is just a pad’s acceptance (modulo units). We set up a 2D wavelet basis: $$\Psi^{\lambda}_{m,i,j}(\zeta,\eta) =
2^{m}\Psi^{\lambda}(2^{m}\zeta-i,2^{m}\eta-j)
\label{dilate_translate}$$ $\Phi_{m,i,j}(\zeta,\eta)$ is constructed from $\Phi(\zeta,\eta)$ similarly. Here, $m$ is an integer scale fineness index; $i$ and $j$ index the discrete positions of pad centers in $\zeta$ and $\eta$ ( $1 \le m \le 4$ and $1\le i,j \le 16$ because we use $16=2^4$ rings and 16 sectors ). Different values of $\lambda$ (denoted as $\zeta$, $\eta$, and $\zeta\eta$) distinguish, respectively, functions with azimuthal, pseudorapidity, and diagonal texture sensitivity: $$\Psi^\zeta=\psi(\zeta)\phi(\eta), \ \
\Psi^\eta=\phi(\zeta)\psi(\eta), \ \
\Psi^{\zeta\eta}=\psi(\zeta)\psi(\eta)$$ In the Haar basis, for any variable $x$ $$%% \psi(x) = \left\{ \begin{array}
%% {r@{\quad:\quad}l}
%% +1 & 0\le x<\frac{1}{2} \\ -1 & \frac{1}{2}\le x<1 \\ 0 & otherwise
%% \end{array} \right.
\psi(x) = \{ +1 \mbox{\ for\ } 0\le x<\frac{1}{2};
-1 \mbox{\ for\ } \frac{1}{2}\le x<1;
0 \mbox{\ otherwise}
\}$$ is the wavelet function. Then, $\Psi^\lambda_{m,i,j}$ with integer $m$, $i$, and $j$ are known [@DWT] to form an othonormal basis in $L^2({\mathbb{R}}^2)$. We adopt the existing [@DWT_power] 1D DWT power spectrum analysis technique and expand it to 2D. Track density in an individual event is $\rho(\zeta,\eta)$ and its *local* fluctuation in a given event is $ \sigma^2 \equiv \langle \rho - \bar{\rho},\rho - \bar{\rho}\rangle,$ where $\bar{\rho}$ is the average $\rho$ in the given event. Using completeness of the basis, we expand $$\rho - \bar{\rho} =
\langle \rho,\Psi^\lambda_{m,i,j}\rangle \Psi^\lambda_{m,i,j}
- \langle \bar{\rho},\Psi^\lambda_{m,i,j}\rangle \Psi^\lambda_{m,i,j}$$
Notice that $\bar{\rho}$, being constant within detector’s rectangular acceptance, is orthogonal to any $\Psi^\lambda_{m,i,j}$ with $m \ge 1$. Due to the orthonormality condition $\langle \Psi^\lambda_{m,i,j},\Psi^{\lambda'}_{m',i',j'}\rangle =
\delta_{\lambda,\lambda'}\delta_{m,m'}\delta_{i,i'}\delta_{j,j'}$, the $\rho - \bar{\rho}$ components for different scales do not form cross-terms in the $\sigma^2$ sum, and the sum contains no cross-terms between $\rho$ and $\bar{\rho}$ for the four observable scales. Instead of a $\langle \rho, \Phi_{m=5,i,j} \rangle$ set, the amplitude array – its closest experimentally achievable approximation – is used as the DWT input.
The Fourier images of 1D wavelet functions occupy a set of wave numbers whose characteristic broadness grows with scale fineness $m$ as $2^m$; $2^{2m}$ should be used in the 2D case. In 2D, we find it most informative to present the three modes of a power spectrum with different directions of sensitivity $P^{\zeta\eta}(m)$, $P^\zeta(m)$, $P^\eta(m)$ separately. We define *power spectrum* as $$P^\lambda(m) =
\frac{1}{2^{2m}}\sum_{i,j}\langle \rho,\Psi^\lambda_{m,i,j}\rangle^2 ,
\label{eq:P_m}$$ where the denominator gives the meaning of spectral *density* to the observable. So defined, the $P(m)$ of a random white noise field (for example) is independent of $m$ [@N_Wiener]. In the first approximation, the white noise example provides a base-line case for comparisons in search for non-trivial effects.
We used WAILI [@WAILI] software library to obtain the wavelet expansions.
=3.5cm
Figure \[compare\] shows the power spectra for one multiplicity range. The first striking feature is that the power spectra of physical events are indeed enhanced on the coarse scale. The task of the analysis is to quantify and, as much as possible, eliminate “trivial” and experiment-specific reasons for this enhancement.
=8.4cm
The average event, formed by summing amplitude arrays of the measured events within a multiplicity range, and dividing by the number of events, has a much reduced texture as fluctuations cancel. However it retains the texture associated with the $\,d^2N/\,d\eta\,d\zeta$, with the dead channels and the finiteness of the beam’s geometrical cross-section. A better way to get rid of the “trivial” texture is to use mixed events. The event mixing is done by taking different channels from different events. Therefore, the mixed events preserve the texture associated with the detector position offset, the inherent $\,dN/\,d\eta$ shape and the dead channels. In order to reproduce the electronics cross-talk effects in the mixed event sample, mixing is done sector-wise, i.e. the sectors constitute the subevents subjected to the event number scrambling. In other words, the mixed events preserve the texture coupled with the channels of the detector. This is *static* texture as it reproduces its pattern event after event; we are interested in *dynamic* texture. We reduce sources of static texture in the power spectra by empty target subtraction and by subtraction of the mixed events power spectra, thus obtaining the *texture correlation* $P^\lambda(m)_{true} - P^\lambda(m)_{mixed}$. Its multiplicity dependence is plotted on Figure \[multi\_dep\]. For comparison with models, a MC simulation (done with RQMD [@RQMD]) includes the known static texture effects and undergoes the same elimination procedure. This allows the effects irreducible by the subtraction methods to be taken into account in the comparison. One such example is the finite beam size, which has been shown by the MC studies to cause the RQMD points to rise with $\,dN/\,d\eta$.
The systematic errors were evaluated by removing the $Pb$ target and switching magnetic field polarity to expose the analyzed side of the detector to $\delta$-electrons, while minimizing nuclear interactions. All correlations (i.e. deviations of $P^\lambda(m)_{true}$ from $P^\lambda(m)_{mixed}$) in such events are considered to be systematic errors. Thus this component of the systematic error gets a sign, and the systematic errors are asymmetric. The other component (significant only on the coarsest scale) is the uncertainty of our knowledge of the beam’s geometrical cross-section.
This novel method of event-by-event analysis, applied to the SPS $PbPb$ data, does not reveal any evidence of critical phenomena. The authors thank N.Antoniou, I.Dremin, E.Shuryak, M.Stephanov, and T.Trainor for illuminating discussions. The NA44 Collaboration thanks the staff of the CERN PS-SPS accelerator complex for their excellent work, and the technical staff in the collaborating institutes for their valuable contributions. This work was supported by the Austrian Fonds zur F[ö]{}rderung der Wissenschaftlichen Forschung; the Science Research Council of Denmark; the Japanese Society for the Promotion of Science; the Ministry of Education, Science and Culture, Japan; the Science Research Council of Sweden; the US Department of Energy and the National Science Foundation.
[99]{} N.M.Astafeva, I.M.Dremin, K.A.Kotelnikov, Mod. Phys. Lett. [**A12**]{} 1185-1192 (1997); I.M.Dremin et al, hep-ph/0007060.
NA49 Collaboration, Phys. Lett. [**B459**]{}, 679-686 (1999)
namely, by using the Central Limit Theorem – see: T.A.Trainor, hep-ph/0001148.
D.J.Scalapino, R.L.Sugar, Phys. Rev. [**D8**]{}, No 7, 2284-2294 (1973)
P.Carruthers, I.Sarcevic, Phys. Lett. [**B189**]{}, No 4, 442-448 (1987)
M.Stephanov, K.Rajagopal, E.Shuryak, Phys. Rev. Lett. [**81**]{} 4816 (1998)
I.Daubechies, Ten Lectures on Wavelets, SIAM, 1992
NA44 Collaboration, Phys. Lett. [**B302**]{} (1993) 510.
E. Beuville *et al.*, NIM [**A288**]{}, 157-167 (1990)
J.Pando, L-Z.Fang, Phys. Rev. [**E57**]{}, 3593-3601 (1998); L-Z.Fang, atro-ph/9791228.
N. Wiener, Generalized Harmonic Analysis.
G. Uytterhoeven *et al.*, WAILI: Wavelets with Integer Lifting. TW Report 262, Department of Computer Science, Katholieke Universiteit Leuven, Belgium, July 1997.
H. Sorge, Phys. Rev. C [**52**]{} (1995) 3291; we use version 2.4 of RQMD.
[^1]:
|
{
"pile_set_name": "ArXiv"
}
|
---
address: 'Babeş-Bolyai University, Faculty of Mathematics and Computer Science, Str. Mihail Kogălniceanu 1, RO-400084 Cluj-Napoca, Romania'
author:
- Tiberiu Coconeţ and Andrei Marcus
title: Group graded basic Morita equivalences
---
We introduce group graded basic Morita equivalences between algebras determined by blocks of normal subgroups, and by using the extended Brauer quotient, we show that they induce graded basic Morita equivalences at local levels.
Introduction
============
Categorical equivalences between blocks of group algebras have been intensely studied over the last four decades, as they provide a structural explanation for various character correspondences which have been observed much earlier. It has turned out that in many cases, these are not mere equivalences between algebras – they are also compatible with the $p$-local structure of the blocks, encoded in terms of Brauer pairs or local pointed groups associated to subgroups of the defect groups, and their fusions. The motivation is related to the older idea of obtaining information about the block from information about blocks of local subgroups.
In the case of principal blocks, J. Rickard [@Ri] introduced the so-called splendid derived equivalences, which involve permutation bimodules with diagonal vertices and trivial sources, with the feature that they give rise to derived equivalences between principal blocks of centralizers of $p$-subgroups. Splendid equivalences have been generalized to arbitrary blocks by M. Linckelmann [@L1], [@L2] and M.E. Harris [@Ha]. Later, a far reaching generalization has been achieved by L. Puig [@Puig], who introduced basic equivalences, which also have a significant local structure, and induce equivalences between block of the centralizers of subgroups of the defect groups by employing the Brauer construction. Moreover, L. Puig and Y. Zhou [@Puig3], [@PZ2] proved that these local equivalences extend to equivalences between blocks of the normalizers of subgroups of the defect groups. It was pointed out by X. Hu [@Hu] that in fact, one gets in [@Puig3] graded Morita equivalences, via the construction introduced in [@Marcus].
In this paper, we investigate the local structure of graded Morita equivalences in a general setting. We fix a complete discrete valuation ring of characteristic zero with residue field $k$ of characteristic $p>0$, finite groups $G$ and $G'$ with normal subgroups $N$ and $N'$ respectively, such that the factor groups $G/N$ and $G'/N'$ are isomorphic. Let $\ddot G$ be the diagonal subgroup of $G\times G'$ with respect to the given isomorphism between $G/N$ and $G'/N'$. Let $b$ be a $G$-invariant block of $\mathcal{O}N$, and let $b'$ be a $G'$-invariant block of $\mathcal{O}N'$.
A $G/N$-graded bimodule $\ddot M$ inducing a Morita equivalence between $A=\mathcal{O}Gb$ and $A'=\mathcal{O}G'b'$ has an $1$-component which can be regarded as an indecomposable $\mathcal{O}\ddot{G}$-module, so it has a vertex $\ddot P$ and a source $\ddot N$. The discussion of the Morita equivalences in terms of $\ddot P$ and $\ddot N$ requires the introduction of a natural group graded structure on Puig’s $\mathcal{O}G$-interior Hecke algebra ${\operatorname{End}}_{\mathcal{O}(1\times G')}({\operatorname{Ind}}_{\ddot{P}}^{G\times G'}(\ddot{N}))$. We studied this graded structure in detail in [@CM]. In Section 3 we characterize in these terms the $G/N$-graded Morita equivalences between $\mathcal{O}Gb$ and $\mathcal{O}G'b'$, by linking, via the source module $\ddot N$, a defect pointed group $P_\gamma$ of $G_{\{b\}}$ on $\mathcal{O}Nb$ and a defect pointed group $P'_{\gamma'}$ of ${G'}_{\{b'\}}$ on $\mathcal{O}N'b'$. Our first main result, Theorem \[main\] below, relates the Morita equivalence between the block extensions $A$ and $A'$ induced by $\ddot M$ to a $G/N$-graded Morita equivalence between the source algebras $A_\gamma$ and $A'_{\gamma'}$.
In Section 4 we introduce the graded version of basic Morita equivalence between $\mathcal{O}Gb$ and $\mathcal{O}G'b'$, and we show in Corollary \[c:truncation\] that the truncation (that is, restriction of the grading to a subgroup of $\Gamma$) of a group graded basic Morita equivalence is again a group graded basic Morita equivalence.
The extended Brauer quotient of $\mathcal{O}G$ with respect to a $p$-subgroup $Q$ of $G$ was introduced in [@Puig3 Section 3], and it was generalized to $N$-interior $G$-algebras in [@CT]. In our situation, we show in Section 5 that the $G/N$-grading on $\mathcal{O}G$ induces a group grading on the extended Brauer quotient, given by a certain subgroup of the normalizer $N_G(Q)$ modulo the centralizer $C_N(Q)$.
This observation is used in our second main result, Theorem \[t:localequiv\] below, where we prove that a basic graded Morita equivalence between $\mathcal{O}Gb$ and $\mathcal{O}G'b'$ induces a group graded basic equivalence at local levels. This is a common generalization of [@Puig3 Theorem 1.4] and [@Marcus Corollary 3.9]. More precisely, [@Puig3 Theorem 1.4] is obtained in the case $G=N$, while in [@Marcus Corollary 3.9] we have only principal blocks $b$ and $b'$, with Sylow $p$-subgroups $P\le N$ and $P'\le N'$, and we do not consider the extended Brauer quotient.
One should mention that Puig also discussed in his book [@Puig] the other types of equivalences relevant to block theory – stable equivalences of Morita type and Rickard equivalences, and moreover, the results of [@Puig3] are extended in [@PZ2] to obtain local Rickard equivalences with a group graded structure. However, the comparison of the results on group graded equivalences presented in [@M Theorem 5.1.2] (Morita), [@M Theorem 5.2.5] (Rickard) and [@M Proposition 5.3.7] (stable) suggests that the assumption that the characteristic $p$ does not divide the order of the grading group $G/N$ is quite reasonable in the case of stable and Rickard equivalences.
We introduce our notations and basic assumptions in Section 2. We will freely use the language of pointed groups on $G$-algebras and their fusions, as presented in [@Puig], [@Puig1] and [@The]. We will not need to assume that $k$ is a splitting field for the group algebras in discussion, so certain field extensions of $k$ will occur below. We also refer to [@M] for results on group graded algebras and modules.
Preliminaries {#sect:prelim}
=============
The following notations and assumptions will be in force for the remaining of the paper.
\[s:GN\] Let $G$ and $G'$ be finite groups, let $\omega:G\to \Gamma$ and $\omega':G'\to \Gamma$ be group epimorphisms, and denote $N:=\mathrm{Ker} \omega$ and $N':=\mathrm{Ker} \omega'$, so that $$\Gamma\simeq G/N\simeq G'/N'.$$ Let $\ddot G$ be the inverse image in $G\times G'$ of the diagonal subgroup of $\Gamma\times \Gamma$, that is $$\ddot G=\{(g,g')\in G\times G'\mid \omega(g)=\omega'(g') \}.$$ Consider the diagonal subalgebra $$\Delta=\bigoplus_{(g,g')\in [\ddot G/N\times N']}\mathcal{O}Ng\otimes_{\mathcal{O}}\mathcal{O}N'g',$$ of the $\Gamma\times\Gamma$-graded algebra $\mathcal{O}G\otimes_{\mathcal{O}}\mathcal{O}G'$ (with respect to the maps $\omega$ and $\omega'$), so $\Delta$ is naturally isomorphic, as a $\Gamma$-graded algebra, to the group algebra $\mathcal{O}\ddot G$.
\[assumption\_on\_M\] Let $b$ be a block of $\mathcal{O}N$ and let $b'$ be a block of $\mathcal{O}N'$ such that $b$ is $G$-invariant and $b'$ is $G'$-invariant. Set $A:=\mathcal{O}Gb$ and $A':=\mathcal{O}G'b'.$ Then both $A$ and $A'$ are $\Gamma$-graded in an obvious way, and we denote by $A_1$ and $A_1'$ respectively the identity components.
We consider the $\Gamma$-graded $G$-acted $\mathcal{O}$-algebra structure on ${A},$ that is, we have $a^g\in A_{x^g},$ for all $x\in \Gamma$ and $a_x\in A_x$.
\[s:ddotM\] Let $M$ be an indecomposable $A_1\otimes A'_1$-module, so $M$ is an indecomposable $\mathcal{O}(N\times N')$-module associated with $b\otimes (b')^\circ$. We assume that $M$ extends to $\Delta$ (that is, the action of $N\times N$ on $M$ extends to $\ddot{G})$, and that $M$ restricted to both $\mathcal{O}(1\times N')$ and $\mathcal{O}(N\times 1)$ is projective. Set $$\ddot{M}:={\operatorname{Ind}}_{\ddot G}^{G\times G'}(M).$$ Then, by [@M Lemma 1.6.3], $\ddot{M}$ is a $\Gamma$-graded $(\mathcal{O}G,(\mathcal{O}G')^{\mathrm{op}})$-bimodule with $1$-component naturally isomorphic to $M$.
Since $M$ is an indecomposable $\mathcal{O}\ddot G$-module, we consider a vertex $\ddot{P}$ of $M$ in $\ddot G,$ and we choose an $\mathcal{O}\ddot{P}$-source $\ddot{N}$ of $M.$ Consider the $\mathcal{O}\ddot{P}$-interior algebra $$S:={\operatorname{End}}_{\mathcal{O}}(\ddot{N}).$$
It follows in particular that $M$ is a direct summand of ${\operatorname{Ind}}_{\ddot{P}}^{\ddot G}(\ddot{N})$ as $\mathcal{O}\ddot G$-modules, and $\ddot{M}$ is a direct summand of ${\operatorname{Ind}}_{\ddot{P}}^{G\times G'}(\ddot{N})$ as $\Gamma$-graded $\mathcal{O}(G\times G')$-modules, where note that the map $\ddot{P}\to G\times G'$ is not injective in general. Let $X$ be the Green correspondent of $M$, so $X$ is an indecomposable $\mathcal{O}N_{\ddot G}(\ddot{P})$-module with vertex $\ddot{P}$ such that $X\mid {\operatorname{Ind}}_{\ddot{P}}^{N_{\ddot G}(\ddot{P})}(\ddot{N}).$
Denote also by $\pi:G\times G'\to G$ the natural projection, by $\rho:\ddot{P}\to G,$ the restriction of $\pi$ to $\ddot{P}$ and by $\sigma:\ddot{P}\to P$ the surjective group homomorphism determined by $\pi$ and $\rho.$ Similarly, we set $\pi',$ $\rho'$ and $\sigma':\ddot{P}\to P'.$
For later use, we record the following lemma relating points on $A$ and points on $A_1$.
\[point\_relation\] For any subgroup $H$ of $G$ and for any point $\hat{\alpha}$ of $H$ on ${A}$ there exists a point $\alpha$ of $H$ on $A_1$ such that $\hat{i}\cdot j=j\cdot \hat{i}=\hat{i},$ for some $\hat{i}\in \hat{\alpha}$ and $j\in \alpha.$
Let $1_{{A}}=\sum j$ be an orthogonal primitive decomposition of the identity element of ${A}$ in $(A_1)^H.$ Let $\hat{i}'\in \hat{\alpha}.$ Since $\hat{i}'\cdot 1_{{A}}=1_{{A}}\cdot \hat{i}'=\hat{i}'$ it follows that there exists an idempotent $j$ such that $j\cdot \hat{i}'\neq 0.$ Now [@Puig1 Proposition 3.19] shows that there exists $\hat{a}\in {A}^*$ such that $j^{\hat{a}}\cdot \hat{i}'=\hat{i}'\cdot j^{\hat{a}}=\hat{i}'$ in ${A}^H,$ as $\hat{i}'$ is a primitive idempotent. The statement follows by taking $\hat{i}:=(\hat{i}')^{\hat{a}^{-1}}\in \hat{\alpha}$ and $\alpha $ the point containing $j.$
Graded Hecke interior algebras and Morita equivalences
======================================================
Let $\ddot{M}$ be an indecomposable $\Gamma$-graded $\mathcal{O}(G\times G')$-module with $1$-component $M$ having vertex $\ddot{P}$ and source $\ddot{N}$ as in \[s:ddotM\]. Consider the induced algebra $$\ddot{A}:={\operatorname{Ind}}_{\ddot{P}}^{G\times G'}({\operatorname{End}}_{\mathcal{O}}(\ddot{N}))\simeq {\operatorname{End}}_{\mathcal{O}}({\operatorname{Ind}}_{\ddot{P}}^{G\times G'}(\ddot{N})),$$ and set $$\hat{A}:=\ddot{A}^{1\times G'}\simeq {\operatorname{End}}_{\mathcal{O}(1\times G')}({\operatorname{Ind}}_{\ddot{P}}^{G\times G'}(\ddot{N})).$$ The algebra $\hat A$ is called in [@Puig Section 4] the [*Hecke $\mathcal{O}G$-interior algebra*]{} associated with $G'$, $\ddot{P}$ and ${\operatorname{End}}_{\mathcal{O}}(\ddot{N})$. In our situation, by [@CM Section 6], $\ddot{A}$ and $\hat{A}$ are naturally $\Gamma$-graded $\mathcal{O}(G\times G')$-interior respectively $\mathcal{O}G$-interior algebras, such that the structural map $$\mathcal{O}G\to \hat{A}$$ is a homomorphism of $\Gamma$-graded $\mathcal{O}G$-interior algebras.
\[s:isoms\] By [@Puig Theorem 4.4] there is an $\mathcal{O}G$-interior algebra isomorphism $$\label{1}
\hat{A}\simeq {\operatorname{Ind}}_{\rho}({\operatorname{End}}_{\mathcal{O}}(\ddot{N})\otimes_{\mathcal{O}}{\operatorname{Res}}_{\rho'}(\mathcal{O}G')),$$ which in our case, by [@CM Section 7], is an isomorphism of $\Gamma$-graded algebras. As a consequence, we obtain the isomorphism $$\label{2}{\operatorname{Ind}}_{\ddot{P}}^{P\times G'}({\operatorname{End}}_{\mathcal{O}}(\ddot{N}))^{1\times G'}\simeq {\operatorname{Ind}}_{\sigma}({\operatorname{End}}_{\mathcal{O}}(\ddot{N})\otimes_{\mathcal{O}}{\operatorname{Res}}_{\rho'}(\mathcal{O}G'))$$ of $\Gamma$-graded $\mathcal{O}P$-interior algebras. Note that constructions made in [@Puig 6.7] can be applied in our situation since $b'$ is $G'$-invariant, hence fron (\[1\]) we obtain the isomorphism $$\label{3} \hat{A}\cdot (b')^\circ\simeq {\operatorname{Ind}}_{\rho}({\operatorname{End}}_{\mathcal{O}}(\ddot{N})\otimes_{\mathcal{O}}{\operatorname{Res}}_{\rho'}(A'))$$ of $\Gamma$-graded $\mathcal{O}G$-interior algebras.
\[n:endom-alg\] Denote $\bar N_{G\times G'}(\ddot P)=N_{G\times G'}(\ddot P)/\ddot P$ and $\bar N_{\ddot G}(\ddot P)=N_{\ddot G}(\ddot P)/\ddot P$. Consider the $\bar N_{\ddot G}(\ddot P)$-graded endomorphism algebra $$E:={\operatorname{End}}_{\mathcal{O}N_{\ddot G}(\ddot{P})}({\operatorname{Ind}}_{\ddot{P}}^{N_{\ddot G}(\ddot{P})}(\ddot{N}))^{\mathrm{op}}$$ and the $\bar N_{G\times G'}(\ddot P)$-graded endomorphism algebra $$\ddot{E}:={\operatorname{End}}_{\mathcal{O}N_{G\times G'}(\ddot{P})}({\operatorname{Ind}}_{\ddot{P}}^{N_{G\times G'}(\ddot{P})}(\ddot{N}))^{\mathrm{op}}.$$ The quotient algebra $\ddot E/\mathrm{J}_{\mathrm{gr}}(\ddot E)$ modulo the graded Jacobson radical $\mathrm{J}_{\mathrm{gr}}(\ddot E)$ is a $\bar N_{G\times G'}(\ddot{P}_{\ddot{N}})$-graded crossed product denoted $\bar k_*\hat{\bar{N}}_{G\times G'}(\ddot{P}_{\ddot{N}})$, where $$\bar k={\operatorname{End}}_{\mathcal{O}{\ddot P}}({\ddot N})/\mathrm{J}({\operatorname{End}}_{\mathcal{O}{\ddot P}}({\ddot N}))$$ is an extension of $k$, while $E/\mathrm{J}_{\mathrm{gr}}(E)$ is naturally isomorphic to the $\bar N_{G\times G'}(\ddot{P}_{\ddot{N}})$-graded subalgebra $\bar k_*\hat{\bar{N}}_{\ddot G}(\ddot{P}_{\ddot{N}})$
\[n:Clifford-corr\] By the Clifford correspondence (see, for instance, [@M Section 2.3.B]), the indecomposable $\mathcal{O}N_{\ddot G}(\ddot{P})$-summand $X$ of ${\operatorname{Ind}}_{\ddot{P}}^{N_{\ddot G}(\ddot{P})}(\ddot{N})$ determines an indecomposable projective $E$-module, which, in turn, determines the indecomposable projective $\bar k_*\hat{\bar{N}}_{\ddot G}(\ddot{P}_{\ddot{N}})$-module $V$, similarly to [@Puig 6.6]. Since by [@M Theorem 2.3.10.d)] the Clifford correspondence commutes with induction, we get that $\ddot{M}$ corresponds to $${\operatorname{Ind}}_{N_{\ddot G}(\ddot{P})}^{N_{G\times G'}(\ddot{P})}(X),$$ which, again by Clifford correspondence, determines a projective $\bar k_*\hat{\bar{N}}_{G\times G'}(\ddot{P}_{\ddot{N}})$-module $\ddot{V}$ such that $$\ddot{V}\simeq {\operatorname{Ind}}_{\bar k_*{\bar{N}}_{\ddot{G}}(\ddot{P}_{\ddot{N}})}^{\bar{k}_*{\bar{N}}_{G\times G'}(\ddot{P}_{\ddot{N}})}(V).$$
\[gammahat\] Now let $\tilde{\ddot{V}}$ be an indecomposable direct summand of $\ddot{V}$ and let $\ddot{W}$ be an isotypic component of the restriction to $\bar k_*\hat{\bar{N}}_{P\times G'}(\ddot{P}_{\ddot{N}})$ of $\mathrm{\tilde{\ddot{V}}}.$ By [@Puig 6.8], $\ddot{W}$ determines a local point $\ddot{\hat{\gamma}}$ of $P$ on the induced algebra $${\operatorname{Ind}}_{\sigma}({\operatorname{End}}_{\mathcal{O}}(\ddot{N})\otimes_{\mathcal{O}}{\operatorname{Res}}_{\rho'}(\mathcal{O}G')),$$ occurring in \[s:isoms\] (\[2\]), and a local point $\hat{\gamma}'$ of $P'$ on $\mathcal{O}G'$, such that $\ddot{\hat{\gamma}}$ is a point of $P$ on $${\operatorname{Ind}}_{\sigma}({\operatorname{End}}_{\mathcal{O}}(\ddot{N})\otimes_{\mathcal{O}}{\operatorname{Res}}_{\rho'}(\mathcal{O}G')_{\hat{\gamma}'}).$$
Let $\tilde{\ddot{M}}$ be the indecomposable $\mathcal{O}(G\times G')$-summand of $\ddot{M}$ that corresponds to $\tilde{\ddot{V}}.$ Since $\ddot{M}$ is associated with $b\otimes (b')^o$, so is $\tilde{\ddot{M}},$ and then we obtain the $\mathcal{O}(G\times G')$-interior algebra embeddings $${\operatorname{End}}_{\mathcal{O}}(\tilde{\ddot{M}})\to {\operatorname{End}}_{\mathcal{O}}(\ddot{M})\to \ddot{A},$$ which induce the $\mathcal{O}G$-interior algebra embeddings $${\operatorname{End}}_{\mathcal{O}}(\tilde{\ddot{M}})^{1\times G'}\to {\operatorname{End}}_{\mathcal{O}}(\ddot{M})^{1\times G'}\to \hat{A}.$$ Now $\tilde{\ddot{M}}$ determines, as in [@Puig 6.6], a point $\hat{\alpha}$ of $G$ on ${\operatorname{End}}_{\mathcal{O}}(\ddot{M})^{1\times G'}$ and on $\hat{A}$ such that $b\cdot \hat{\alpha}=\hat{\alpha}.$ As in [@Puig 6.8], we obtain that $$(P\times G')_{\ddot{\hat{\gamma}}}\leq (G\times G')_{\hat{\alpha}},$$ and thus, by using the isomorphism \[s:isoms\] (\[3\]), we have that $P'_{\hat{\gamma}'}$ is a local point on $A'.$
Consider again the $\Gamma$-graded $\mathcal{O}G$-interior algebra $$\hat{C}:={\operatorname{Ind}}_{\rho}({\operatorname{End}}_{\mathcal{O}}(\ddot{N})\otimes_{\mathcal{O}}{\operatorname{Res}}_{\rho'}(A'))$$ occurring in \[s:isoms\] (\[3\]), whose $1$-component is $$\hat{C}_1=(\mathcal{O}\otimes_{\mathcal{O}(1\times N')} ({\operatorname{End}}_{\mathcal{O}}({\operatorname{Ind}}_{\ddot{P}}^{\ddot G}(\ddot{N}))\otimes_{\mathcal{O}}A_1'))^{1\times N'}.$$
\[points\_on\_1\_cmp\] With the above notation, the points $\ddot{\hat{\gamma}}$ and $\hat{\gamma}'$ introduced in \[gammahat\] determine the local points $\hat{\gamma}$ of $P$ on $\hat{C}_1$ and $\gamma'$ of $P$ in $A'_1$ such that $\hat{\gamma}$ is a point of $P$ on $$(\hat{C}_1)_{\gamma'}:=(\mathcal{O}\otimes_{\mathcal{O}(1\times N')} ({\operatorname{End}}_{\mathcal{O}}({\operatorname{Ind}}_{\ddot{P}}^{\ddot G}(\ddot{N}))\otimes_{\mathcal{O}}(A_1')_{\gamma'}))^{1\times N'}.$$
It follows from [@Puig 6.8] that $\ddot{\hat{\gamma}}$ is a point of $P$ on $${\operatorname{Ind}}_{\sigma}({\operatorname{End}}_{\mathcal{O}}(\ddot{N})\otimes_{\mathcal{O}}{\operatorname{Res}}_{\sigma'}(A'_{\hat{\gamma}'})).$$ Since $A'$ is $\Gamma$-graded, Lemma \[point\_relation\] gives a point $\gamma'$ of $P'$ on $A'_1$ such that $\hat{i}'\cdot i'=\hat{i}'$ for some idempotents chosen as above. Now clearly $\ddot{\hat{\gamma}}$ becomes a point of $P$ on $${\operatorname{Ind}}_{\rho}({\operatorname{End}}_{\mathcal{O}}(\ddot{N})\otimes_{\mathcal{O}}{\operatorname{Res}}_{\sigma'}(A'_{\gamma'}))$$ via the $\mathcal{O}P$-interior algebra embedding $${\operatorname{Ind}}_{\sigma}({\operatorname{End}}_{\mathcal{O}}(\ddot{N})\otimes_{\mathcal{O}}{\operatorname{Res}}_{\sigma'}(A'_{\gamma'}))\to {\operatorname{Ind}}_{\rho}({\operatorname{End}}_{\mathcal{O}}(\ddot{N})\otimes_{\mathcal{O}}{\operatorname{Res}}_{\sigma'}(A'_{\gamma'})).$$ This last algebra is again $\Gamma$-graded, having as identity component the $P$-algebra $(\hat{C}_1)_{\gamma'}.$ The point $\hat{\gamma}$ is determined by applying again Lemma \[point\_relation\].
Recall that a pointed group $P_\gamma$ on $A_1$ determines, as in [@FP 1.16], the simple quotient $A_1(P_\gamma)$ of $A_1^P$, and a field extension $\hat k$ of $k$ and a crossed product (or a twisted $\hat k^*$-group algebra) $\hat k_*\hat{\bar N}_G(P_\gamma)$, such that the simple $A_1(P_\gamma)$-module becomes a $\hat k_*\hat{\bar N}_G(P_\gamma)$-module denoted by $V_{A_1}(P_\gamma)$, and called the [*multiplicity module*]{} of $P_\gamma$. When we apply this idea to the $\mathcal{O}\ddot{G}$-module $M$ and the $\mathcal{O}\ddot{G}$-interior algebra ${\operatorname{End}}_{\mathcal{O}}(M)$, then the pointed group corresponding to the source $\ddot{N}$ will be denoted by $\ddot{P}_{\ddot{N}}$, and the multiplicity $\bar k_* \hat{{\bar N}}_{\ddot{G}}(\ddot{P})$-module will be denoted by $V_{M}(\ddot{P}_{\ddot{N}})$.
With the above notations, we may state the following group graded version of [@Puig Theorem 6.9]. Note again that we only consider here Morita equivalences (and not Morita stable equivalences), so our result takes the simpler form mentioned at the end of the statement of [@Puig Theorem 6.9].
\[main\] The $(\mathcal{O}G,\mathcal{O}G')$-bimodule $\ddot{M}$ induces a $\Gamma$-graded Morita equivalence between $A$ and $A'$ if and only if $P$ is a defect group of $b$ regarded as a primitive idempotent of $(A_1)^G,$ $P'$ is a defect group of $b'$ regarded as a primitive idempotent of $(A_1')^{G'},$ and for a suitable local point $\gamma$ of $P$ on $A_1$, there is a $\Gamma$-graded $\mathcal{O}P$-interior algebra isomorphism $$e:A_{\gamma}\longrightarrow ({\operatorname{Ind}}_{\sigma}({\operatorname{End}}_{\mathcal{O}}(\ddot{N})\otimes_{\mathcal{O}}{\operatorname{Res}}_{\sigma'}(A'_{\gamma'})))_{\hat{\gamma}}$$ [(]{}where $\hat\gamma$ and $\gamma'$ are defined in Proposition \[points\_on\_1\_cmp\][)]{} such that we have an isomorphism $$V_{A_1}(P_{\gamma})\simeq {\operatorname{Res}}_{\widehat{({\operatorname{Ind}}_P^G e)_1(P_{\gamma})}}(V_{M}((P\times G'))_{\hat{\gamma}}))$$ of $\hat k_*\hat{\bar{N}}_{G}(P_{\gamma})$-modules, where $$\widehat{({\operatorname{Ind}}_P^G e)_1(P_{\gamma})}: \hat k_*\hat{\bar N}_G(P_\gamma) \to \hat k_*\hat{\bar N}_G(P_{\hat\gamma})\simeq \hat k_*\hat{\bar N}_{G\times G'}((P\times G')_{\hat\gamma})$$ is the isomorphism of twisted group algebras induced by $e$.
By our assumptions, the $\Gamma$-graded Morita equivalence determines the $\Gamma$-graded $\mathcal{O}G$-interior algebra isomorphism $$\label{i}A\simeq {\operatorname{End}}_{\mathcal{O}}(\ddot{M})^{1\times G'},$$ given by the structural map. We have that $N\times 1$, $1\times N'$ and $N\times N'$ are normal subgroups of ${\ddot G},$ so the identity component of the above isomorphism restricts to the $\ddot G$-algebra isomorphism $$A_1\simeq {\operatorname{End}}_{\mathcal{O}}(M)^{1\times N'},$$ given by the structural map, as in [@Puig Theorem 6.5]. Therefore, we have $$A_1^{\ddot G}\simeq {\operatorname{End}}_{\mathcal{O}}(M)^{\ddot G}.$$ Note that the action of $\ddot G$ on $A_1$ coincides with the action of $G$ on $A_1,$ forcing $(A_1)^G\simeq (A_1)^{\ddot G}.$
Let $D$ be a defect group in $G$ of $b.$ If $D'$ is the subgroup of $G'$ such that $\omega(D)=\omega'(D')$ (so $\omega$ and $\omega'$ induce the isomorphism $DN/N\simeq D'N'/N'$), then $$(A_1)^G_D\simeq (A_1)^{\ddot G}_{D\times D'}\simeq {\operatorname{End}}_{\mathcal{O}}(M)^{\ddot G}_{D\times D'}.$$ We deduce that for a suitable pair $(x,x')\in {\ddot G}$, we have $\ddot P^{(x,x')}\leq D\times D',$ that is, $P^x\leq D.$
Now let $\bar{\ddot{M}}$ be an indecomposable $\mathcal{O}(G\times G')$-summand of $\ddot{M}$, having vertex $\ddot{Q}$ and source $\bar{\ddot{N}}$, and let $Q$ and $Q'$ the projections of $\ddot{Q}$ in $G$ and $G'$ respectively. Then $\bar{\ddot{M}}$ determines a block $B$ of $A$ such that by isomorphism (\[i\]) we obtain the isomorphism $$A\cdot B\simeq {\operatorname{End}}_{\mathcal{O}}(\bar{\ddot{M}})^{1\times G'}$$ of $\mathcal{O}G$-interior algebras. Moreover, the block $B$ determines a point $\tilde{\alpha}$ of $G$ on $${\operatorname{Ind}}_{\ddot{Q}}^{G\times G'}({\operatorname{End}}_{\mathcal{O}}(\bar{\ddot{N}}))^{1\times G'}$$ such that the $\mathcal{O}G$-interior algebra embedding $${\operatorname{End}}_{\mathcal{O}}(\bar{\ddot{M}})^{1\times G'}\to {\operatorname{Ind}}_{\ddot{Q}}^{G\times G'}({\operatorname{End}}_{\mathcal{O}}(\bar{\ddot{N}}))^{1\times G'}$$ determines the $\mathcal{O}G$-interior algebra isomorphism $$A\cdot B\simeq ({\operatorname{Ind}}_{\ddot{Q}}^{G\times G'}({\operatorname{End}}_{\mathcal{O}}(\bar{\ddot{N}}))^{1\times G'})_{\tilde{\alpha}}.$$ Using this last isomorphism and [@Puig Proposition 5.3], we deduce that $Q$ is actually a defect group of $B.$ On the other hand, we have $\bar{\ddot{M}}\mid \ddot{M}\mid {\operatorname{Ind}}_{\ddot{P}}^{G\times G'}(\ddot{N})$, and thus we may choose $Q$ to be a subgroup of $P.$ So far we have obtained the inclusions $Q^x\leq P^x\leq D$ for some $x\in G.$
By changing the choice of $\bar{\ddot{M}}$, we get, as above, all the blocks $B$ of $G$ that cover $b.$ It is known (see [@Alp IV.15, Theorem 1]) that there is at least one block that covers $b$ and has defect group $D$. If that is the case for $B$, then we deduce that $P^x=D$ for some $x\in G.$ Hence $P$ is a defect group of $b$ in $G.$ By symmetry, we get that $P'$ is a defect group of $b'$ in $G'.$
Let $\bar{B}$ be the block of $A$ corresponding to $\hat{\alpha}$ via (\[i\]). Then we have the isomorphisms $$\bar{B}A\simeq ({\operatorname{End}}_{\mathcal{O}}(\bar{\ddot{M}})^{1\times G'})_{\hat{\alpha}}\simeq \hat{A}_{\hat{\alpha}},$$ of $\mathcal{O}G$-interior algebras. Clearly, the defect pointed group $P_{\ddot{\hat{\gamma}}}$ of $G_{\hat{\alpha}}$ determines a defect pointed group $P_{\bar{\gamma}}$ of $G_{\bar{B}}$, and $\bar{\gamma}$ is still a local point of $P$ on $A$ via the embedding $\bar{B}A\to A.$ Again Lemma \[point\_relation\] gives a point $\gamma$ of $P$ on $A_1,$ which is also local since $\bar{\gamma}$ is. It is now easy to see that $P_{\gamma}$ is a defect pointed group of $G_{\{b\}}.$
The $\Gamma$-graded $\mathcal{O}G$-interior algebra embedding $$\label{4}A\simeq {\operatorname{End}}_{\mathcal{O}}(\ddot{M})^{1\times G'}\to \hat{A}\cdot (b')^{o}\simeq {\operatorname{Ind}}_{\rho}({\operatorname{End}}_{\mathcal{O}}(\ddot{N})\otimes_{\mathcal{O}}{\operatorname{Res}}_{\rho'}(A'))$$ restricts to a $G$-algebra embedding $A_1\to \hat{C}_1,$ so that the correspondence between $\bar{\gamma}$ and $\ddot{\hat{\gamma}}$ determines the correspondence between $\gamma$ and $\hat{\gamma}.$ Proposition \[points\_on\_1\_cmp\] gives the $\mathcal{O}P$-interior algebra isomorphism $$\label{5}(A_1)_{\gamma}\simeq ((\hat{C}_1)_{\gamma'})_{\hat{\gamma}}\simeq ({\operatorname{Ind}}_{\sigma}({\operatorname{End}}_{\mathcal{O}}(\ddot{N})\otimes_{\mathcal{O}}{\operatorname{Res}}_{\sigma'}((A'_1)_{\gamma'})))_{\hat{\gamma}}.$$ Recall that the embedding of $\mathcal{O}P$-interior algebras
$$({\operatorname{Ind}}_{\ddot{P}}^{P\times G'}({\operatorname{End}}_{\mathcal{O}}(\ddot{N}))^{1\times G'})\cdot (b')^o\to \hat{A}\cdot (b')^o$$ is grade-preserving, while $${\operatorname{Ind}}_{\ddot{P}}^{P\times G'}({\operatorname{End}}_{\mathcal{O}}(\ddot{N}))^{1\times G'}\cdot (b')^{o}\simeq {\operatorname{Ind}}_{\sigma}({\operatorname{End}}_{\mathcal{O}}(\ddot{N})\otimes_{\mathcal{O}}({\operatorname{Res}}_{\rho'}(A')))$$ is an isomorphism of $\Gamma$-graded $\mathcal{O}P$-interior algebras, and $${\operatorname{Ind}}_{\ddot{P}}^{G\times G'}({\operatorname{End}}_{\mathcal{O}}(\ddot{N}))^{1\times G'}\cdot (b')^o\simeq {\operatorname{Ind}}_{\rho}({\operatorname{End}}_{\mathcal{O}}(\ddot{N})\otimes_{\mathcal{O}}({\operatorname{Res}}_{\rho'}(A')))$$ is an isomorphism of $\Gamma$-graded $\mathcal{O}G$-interior algebras. We obtain the $\Gamma$-graded $\mathcal{O}P$-interior algebra homomorphism $$({\operatorname{Ind}}_{\sigma}({\operatorname{End}}_{\mathcal{O}}(\ddot{N})\otimes_{\mathcal{O}}{\operatorname{Res}}_{\sigma'}((A')_{\gamma'})))_{\hat{\gamma}}\to ({\operatorname{Ind}}_{\rho}({\operatorname{End}}_{\mathcal{O}}(\ddot{N})\otimes_{\mathcal{O}}{\operatorname{Res}}_{\sigma'}((A')_{\gamma'})))_{\hat{\gamma}},$$ which is an isomorphism by the second part of isomorphism (\[5\]). Now the homomorphism (\[4\]) restricts to the $\Gamma$-graded $\mathcal{O}P$-interior algebra homomorphism $$A_{\gamma}\to ({\operatorname{Ind}}_{\sigma}({\operatorname{End}}_{\mathcal{O}}(\ddot{N})\otimes_{\mathcal{O}}{\operatorname{Res}}_{\sigma'}((A')_{\gamma'})))_{\hat{\gamma}},$$ and (\[5\]) shows that this is in fact an isomorphism.
Conversely, let us first denote $$\hat{B}:=({\operatorname{Ind}}_{\sigma}({\operatorname{End}}_{\mathcal{O}}(\ddot{N})\otimes_{\mathcal{O}}{\operatorname{Res}}_{\sigma'}((A')_{\gamma'})))_{\hat{\gamma}},$$ and we have seen that $\hat{B}$ is a $\Gamma$-graded $\mathcal{O}P$-interior algebra. Further, since $G_{\{b\}}$ is projective relative to $P_{\gamma}$ on $A_1$, we still have $b\in {\operatorname{Tr}}_P^G(A^P\gamma A^P).$ It follows, according to [@Puig2 Proposition 3.6], that there exists a $\Gamma$-graded $\mathcal{O}G$-interior algebra embedding $$h:A\to {\operatorname{Ind}}_P^G(A_{\gamma}),$$ sending $a\in A$ to $${\operatorname{Tr}}_P^G(1\otimes i\otimes 1)\cdot (1\otimes iai\otimes 1)\cdot {\operatorname{Tr}}_P^G(1\otimes i\otimes 1),$$ where $i\in \gamma.$
By our assumptions we obtain the $\Gamma$-graded $\mathcal{O}G$-interior algebra isomorphism $$g:{\operatorname{Ind}}_P^G(A_{\gamma})\to \hat{B},$$ and then the obvious $\Gamma$-graded $\mathcal{O}G$-interior algebra embedding $$\hat{B}\to {\operatorname{Ind}}_P^G({\operatorname{Ind}}_{\sigma}({\operatorname{End}}_{\mathcal{O}}(\ddot{N})\otimes_{\mathcal{O}}{\operatorname{Res}}_{\rho'}(A')))\simeq \hat{A}.$$ Clearly, we have the $\Gamma$-graded $\mathcal{O}G$-interior algebra embedding $$\phi:A\to \hat{A},$$ given by the composition of the above maps. Then $\phi_1:A_1\to \hat{A}_1$ is an embedding of $G$-algebras. It is also clear that $\phi_1(\gamma)=\hat{\gamma}$, and then $b$ determines a point $\hat{\beta}$ of $G$ on $\hat{A}_1,$ which is already a point of $G$ on $\hat{B}_1,$ with defect group $P_{\hat{\gamma}}$, so that we have the isomorphism $$\label{6}\hat{\phi}(P_{\gamma}):\hat k_*\hat{\bar{N}}_{G}(P_{\gamma})\simeq \hat k_*\hat{\bar{N}}_{G}(P_{\hat{\gamma}})$$ of ${\bar{N}}_{G}(P_{\gamma})$-graded algebras, and the isomorphisms $$\begin{aligned}
V_{A_1}(P_{\gamma}) & \simeq {\operatorname{Res}}_{\hat{\phi}(P_{\gamma})}(V_{(\hat{A}_1)_{\hat{\beta}}}(P_{\hat{\gamma}})) \\
& \simeq {\operatorname{Res}}_{\hat{\phi}(P_{\gamma})}(V_{(\hat{B}_1)_{\hat{\beta}}}(P_{\hat{\gamma}})).\end{aligned}$$ Let $\hat{\alpha}$ be the point of $G$ on $\hat{A}_1$ such that $b\cdot \hat{\alpha}=\hat{\alpha}$, obtained via the $G$-algebra embedding $$\label{7}{\operatorname{End}}_{\mathcal{O}}(M)^{1\times N'}\to \hat{A}_1.$$ Using this and our assumption, we have the $\hat k_*{\hat{\bar{N}}}_G(P_{\gamma})$-module isomorphisms $$\begin{aligned}
V_{A_1}(P_{\gamma}) & \simeq {\operatorname{Res}}_{\hat{\phi}(P_{\gamma})}(V_{M}((P\times N')_{\hat{\gamma}})) \\
& \simeq {\operatorname{Res}}_{\hat{\phi}(P_{\gamma})}(V_{(\hat{A}_1)_{\hat{\alpha}}}(P_{\hat{\gamma}})).\end{aligned}$$ Hence, by (\[6\]) and (\[7\]) we can identify $\hat{\alpha}$ with $\hat{\beta}$ and the structural morphisms $f:A_1\to (\hat{A}_1)_{\hat{\alpha}}$ with the restriction $A_1\to (\hat{B}_1)_{\hat{\beta}},$ of $\phi_1$.
Since $g_1\circ h_1$ is an embedding, there is a homomorphism of $(\mathcal{O}N,\mathcal{O}N)$-bimodules $$r:(\hat{B}_1)_{\hat{\beta}}:=j(\hat{B}_1)j\to A_1,$$ which is also $\mathcal{O}G$-linear, such that $r\circ g_1\circ h_1=\mathrm{id}_{A_1},$ where $g_1(h_1(b))=j.$ Now, for any $x\in N$, we get $$xb=x(r (g_1(h_1(b)))=xr(j)=r(f(xb)),$$ hence $f$ is an isomorphism.
Finally, we obtain the $G$-algebra isomorphism $$A_1\simeq {\operatorname{End}}_{\mathcal{O}}(M)^{1\times N'}.$$ Now, by using [@Puig Theorem 6.5], the fact that $M$ extends to $\ddot G$, and [@M Theorem 5.1.2], the required statement follows.
\[sufficiency\_for\_morita\_eq\] Similarly to [@Puig 6.12], we show how by applying Theorem \[main\], we obtain all the possible choices of an indecomposable $\mathcal{O}\ddot{G}$-module inducing a Morita equivalence between $A_1$ and $A'_1$, and whose induction to $G\times G'$ induces a $\Gamma$-graded Morita equivalence between $A$ and $A'.$
Let $b$ and $b'$ be as in \[assumption\_on\_M\], and assume that $Q\leq G$ is a defect group of $b$ in $G$, while $Q'$ is a defect group of $b'$ in $G'$, such that $\ddot{Q}\leq \ddot G,$ where $\ddot{Q}$ runs through all the subgroups of $Q\times Q'$ such that the projections $\tau$ and $\tau'$ satisfy $\tau(\ddot{Q})=Q$ and $\tau'(\ddot{Q})=Q'.$
We consider the defect pointed group $Q_{\delta}$ of $b$ and the defect pointed group $Q'_{\delta'}$ of $b'$, and we look for all indecomposable $\mathcal{O}\ddot{Q}$-modules $\ddot{L}$ having vertex $\ddot{Q},$ such that the restrictions to ${\operatorname{Ker}}(\tau)$ and to ${\operatorname{Ker}}(\tau')$ are projective, and there is a $\Gamma$-graded $\mathcal{O}Q$-interior algebra embedding $$A_{\delta}\to {\operatorname{Ind}}_{\tau}({\operatorname{End}}_{\mathcal{O}}(\ddot{L})\otimes_{\mathcal{O}}({\operatorname{Res}}_{\tau'}(A'_{\delta'}))).$$
With these assumptions, denoting by $\xi'$ the restriction to $\ddot{Q}$ of the projection $G\times G'\to G'$, note that there is the composition $$\begin{aligned}
f:A_{\delta}&\to {\operatorname{Ind}}_{\tau}({\operatorname{End}}_{\mathcal{O}}(\ddot{L})\otimes_{\mathcal{O}}({\operatorname{Res}}_{\tau'}(A'_{\delta'})))\\
&\to {\operatorname{Ind}}_{\tau}({\operatorname{End}}_{\mathcal{O}}(\ddot{L})\otimes_{\mathcal{O}}({\operatorname{Res}}_{\xi'}(\mathcal{O}G')))\\
&\to {\operatorname{Ind}}_{\ddot{Q}}^{Q\times G'}({\operatorname{End}}_{\mathcal{O}}(\ddot{L}))^{1\times G'} \\
&\to {\operatorname{End}}_{\mathcal{O}}({\operatorname{Ind}}_{\ddot{Q}}^{G\times G'}(\ddot{L}))^{1\times G'}\end{aligned}$$ of $\Gamma$-graded $\mathcal{O}Q$-interior algebra embeddings, that restricts to the embedding $$f_1:(A_1)_{\delta}\to {\operatorname{End}}_{\mathcal{O}}({\operatorname{Ind}}_{\ddot{Q}}^{\ddot G}(\ddot{L}))^{1\times N'}$$ of $Q$-algebras between the identity components.
Now $\hat{\delta}:=f_1(\delta)$ is a local point of $Q$ on ${\operatorname{End}}_{\mathcal{O}}({\operatorname{Ind}}_{\ddot{Q}}^{\ddot G}(\ddot{L}))^{1\times N'},$ and at the same time, a point of $Q\times N'$ on ${\operatorname{End}}_{\mathcal{O}}({\operatorname{Ind}}_{\ddot{Q}}^{\ddot G}(\ddot{L}))$, or equivalently, a point of $Q\times G'$ on ${\operatorname{End}}_{\mathcal{O}}({\operatorname{Ind}}_{\ddot{Q}}^{G\times G'}(\ddot{L}))_1.$ Let $\bar{\delta}$ be a local point of $Q$ on $A_{\delta}.$ According to [@Puig 6.13], $\bar{\delta}$ determines a unique local pointed group $Q_{\ddot{\hat{\delta}}}$ on ${\operatorname{End}}_{\mathcal{O}}({\operatorname{Ind}}_{\ddot{Q}}^{G\times G'}(\ddot{L}))^{1\times G'}.$ The point $\ddot{\hat{\delta}}$ is actually a point of $Q\times G'$ on $({\operatorname{End}}_{\mathcal{O}}({\operatorname{Ind}}_{\ddot{Q}}^{G\times G'}(\ddot{L})))_{\hat{\delta}}$ satisfying $$\ddot{Q}_{\ddot{L}}\leq (Q\times G')_{\ddot{\hat{\delta}}}.$$
Let $\ddot{\hat{\beta}}$ be a point of $G\times G'$ on ${\operatorname{End}}_{\mathcal{O}}({\operatorname{Ind}}_{\ddot{Q}}^{G\times G'}(\ddot{L}))$ such that $$(Q\times G')_{\ddot{\hat{\delta}}}\leq (G\times G')_{\ddot{\hat{\beta}}}.$$ There is a unique indecomposable $\mathcal{O}\ddot G$-summand $Y$ of ${\operatorname{Ind}}_{\ddot{Q}}^{\ddot G}(\ddot{L})$ such that $\ddot{\hat{\beta}}$ corresponds to a unique isomorphisms class of indecomposable $\mathcal{O}(G\times G')$-modules determined by a direct summand of ${\operatorname{Ind}}_{\ddot G}^{G\times G'}(Y).$ Explicitly, we still have the inclusion $$(Q\times G')_{\ddot{\hat{\delta}}}\leq (G\times G')_{\ddot{\hat{\beta}}}$$ on the $\Gamma$-graded $\mathcal{O}(G\times G')$-interior algebra ${\operatorname{End}}_{\mathcal{O}}({\operatorname{Ind}}_{\ddot G}^{G\times G'}(Y)).$ Hence $\ddot{\hat{\delta}}$ is a point of $Q\times G'$ on $({\operatorname{End}}_{\mathcal{O}}({\operatorname{Ind}}_{\ddot G}^{G\times G'}(Y)))_{\hat{\delta}}.$
According to Lemma \[point\_relation\], $\ddot{\hat{\beta}}$ determines a point $\hat{\beta}$ of $G\times G'$ on the identity component ${\operatorname{End}}_{\mathcal{O}}({\operatorname{Ind}}_{\ddot G}^{G\times G'}(Y))_1$ of ${\operatorname{End}}_{\mathcal{O}}({\operatorname{Ind}}_{\ddot G}^{G\times G'}(Y)).$ We clearly have the inclusion $$(Q\times G')_{\hat{\delta}}\leq (G\times G')_{\hat{\beta}}$$ of pointed groups on ${\operatorname{End}}_{\mathcal{O}}({\operatorname{Ind}}_{\ddot G}^{G\times G'}(Y))_1.$ The $\mathcal{O}\ddot{Q}$-interior algebra embedding $${\operatorname{End}}_{\mathcal{O}}(\ddot{L})\to {\operatorname{End}}_{\mathcal{O}}({\operatorname{Ind}}_{\ddot{Q}}^{\ddot G}(\ddot{L}))$$ shows that $\ddot{L}$ lies in the restriction to $\ddot{Q}$ of a unique $\mathcal{O}{\ddot G}$-indecomposable direct summand of ${\operatorname{Ind}}_{\ddot{Q}}^{\ddot G}(\ddot{L}).$ Taking into account that $$\ddot{Q}_{\ddot{L}}\leq (G\times G')_{\ddot{\hat{\beta}}},$$ we deduce that $\ddot{L}\mid {\operatorname{Res}}_{\ddot{Q}}^{\ddot G}(Y),$ and then $\ddot{Q}_{\ddot{L}}$ is still a local pointed group on ${\operatorname{End}}_{\mathcal{O}}({\operatorname{Ind}}_{\ddot G}^{G\times G'}(Y)),$ hence on ${\operatorname{End}}_{\mathcal{O}}({\operatorname{Ind}}_{\ddot G}^{G\times G'}(Y))_1,$ where we have the inclusions $$\ddot{Q}_{\ddot{L}}\leq(Q\times G')_{\hat{\delta}}\leq (G\times G')_{\hat{\beta}}.$$
Now, the natural embedding of $\mathcal{O}\ddot G$-interior algebras $${\operatorname{End}}_{\mathcal{O}}(Y)\to {\operatorname{End}}_{\mathcal{O}}({\operatorname{Ind}}_{\ddot{Q}}^{\ddot G}(\ddot{L}))$$ and [@Puig 2.11.3], since $\ddot{Q}_{\ddot{L}}\leq {\ddot G}_Y,$ prove that $\ddot{Q}_{\ddot{L}}$ is a defect pointed group of ${\ddot G}_Y,$ hence $\ddot{Q}$ is a vertex of $Y.$ Let $\ddot{R}_{\hat{\epsilon}}$ be a local pointed group on ${\operatorname{End}}_{\mathcal{O}}({\operatorname{Ind}}_{\ddot{Q}}^{G\times G'}(\ddot{L}))_1$ such that $$\ddot{Q}_{\ddot{L}}\leq \ddot{R}_{\hat{\epsilon}}\leq (Q\times G')_{\hat{\delta}}.$$ Then $\ddot{R}_{\hat{\epsilon}}$ determines a local pointed group $\ddot{R}_{\ddot{\hat{\epsilon}}}$ on ${\operatorname{End}}_{\mathcal{O}}({\operatorname{Ind}}_{\ddot{Q}}^{G\times G'}(\ddot{L})),$ which, according to [@Puig 2.11.3], is included in $\ddot{Q}_{\ddot{L}}.$ This shows that $\ddot{Q}_{\ddot{L}}$ is a defect pointed group for both $(Q\times G')_{\hat{\delta}}$ and $(G\times G')_{\hat{\beta}}.$
We have obtained the embedding $$f_1:(A_1)^Q_{\delta}\to (({\operatorname{End}}_{\mathcal{O}}({\operatorname{Ind}}_{\ddot G}^{G\times G'}(Y)))_1)_{\hat{\delta}}^{Q\times G'}={\operatorname{End}}_{\mathcal{O}}(Y)_{\hat{\delta}}^{Q\times N'},$$ and then $$\begin{aligned}
N_G(Q_{\delta})/Q &\simeq N_{G\times G'}((Q\times G')_{\hat{\delta}})/Q\times G' \\
&\simeq N_{G\times G'}((Q\times N')_{\hat{\delta}})/Q\times G'.\end{aligned}$$ The embedding $f_1$ induces the isomorphisms $$\begin{aligned}
\hat{f_1}(Q_{\delta}):\hat k_*\hat{\bar{N}}_{G}(Q_{\delta})&\simeq \hat k_*\hat{\bar{N}}_{G\times G'}((Q\times G')_{\hat k_*{\delta}}) \\
&\simeq \hat k_*\hat{\bar{N}}_{G\times G'}((Q\times N')_{\hat{\delta}})/Q\times G',\end{aligned}$$ of ${\bar{N}}_{G}(Q_{\delta})$-graded algebras, and then we have $$V_{A_1}(Q_{\delta})\simeq {\operatorname{Res}}_{\hat{f_1}(Q_{\delta})}(V_Y((Q\times N')_{\hat{\delta}})).$$
Finally, notice that the indecomposable $\mathcal{O}\ddot G$-summand $Y$ of ${\operatorname{Ind}}_{\ddot{Q}}^{\ddot G}(\ddot{L})$ introduced here is projective when restricted to $\mathcal{O}(N\times 1)$ and to $\mathcal{O}(1\times N').$ As a conclusion of the above discussion, by Theorem \[main\] we obtain:
The $\Gamma$-graded $\mathcal{O}(G\times G')$-module $\ddot{Y}:={\operatorname{Ind}}_{\ddot G}^{G\times G'}(Y)$ induces a $\Gamma$-graded Morita equivalence between $b\mathcal{O}G$ and $b'\mathcal{O}G'.$
Graded basic Morita equivalences {#sect:basic}
================================
By using the constructions of the previous section, we get the following graded version of [@Puig Corollary 7.4 ], which allows us to define the notion of basic graded Morita equivalence. The notations are those of Section \[sect:prelim\] and of Theorem \[main\].
\[equiv\_statements\_for\_basic\] Assume that $\ddot{M}$ defines a graded Morita equivalence between $A$ and $A'.$ Then the following assertions are equivalent:
1. $\sigma$ is a group isomorphism;
2. $\sigma'$ is a group isomorphism;
3. $p$ does not divide the rank of $\ddot{N}$ over $\mathcal{O};$
4. $S$ is a Dade $\ddot{P}$-algebra.
The implication $4)\Rightarrow 3)$ follows exactly as in the proof of [@Puig Corollary 7.4]. Further, we assume that $3)$ holds. Then, since ${\operatorname{Ker}}(\sigma)\leq 1\times N',$ ${\operatorname{Ker}}(\sigma')\leq N\times 1$ and $\ddot{N}\mid {\operatorname{Res}}^{\ddot G}_{\ddot{P}}(M),$ the assumptions made in \[s:ddotM\] show that $\ddot{N}$ is a projective $\mathcal{O}{\operatorname{Ker}}(\sigma)$-module and a projective $\mathcal{O}{\operatorname{Ker}}(\sigma')$-module. This fact forces $\sigma$ and $\sigma'$ to be injective group homomorphisms. Now Proposition \[points\_on\_1\_cmp\] and the proof of Theorem \[main\] show that $\hat{A}_{\hat{\gamma}}$ is a direct $\mathcal{O}$-summand of $\hat{A}_{\alpha},$ since $\gamma$ corresponds to $\hat{\gamma},$ $b\cdot \alpha=\alpha,$ $P_{\gamma}$ is a defect pointed group of $G_{\{b\}}$, and since the isomorphism $A_1\simeq (\hat{A}_1)_{\alpha}$ forces $A\simeq \hat{A}_{\alpha}.$ The argument used in [@Puig Remark 6.11] guarantees that $\hat{A}_{\hat{\gamma}}$ has a $P$-stable $\mathcal{O}$-basis. The $\mathcal{O}P$-interior algebra embedding $$\hat{A}_{\hat{\gamma}}\simeq ({\operatorname{Ind}}_{\sigma}(S\otimes_{\mathcal{O}}{\operatorname{Res}}_{\sigma'}(A'_{\gamma'})))_{\hat{\gamma}},$$ given by Proposition \[points\_on\_1\_cmp\], determines the $\mathcal{O}P$-interior algebra embedding $$\hat{A}_{\hat{\gamma}}\to {\operatorname{Res}}_{\sigma^{-1}}(S)\otimes_{\mathcal{O}}{\operatorname{Res}}_{\sigma'\circ \sigma^{-1}}(A'_{\gamma'}).$$ Finally, we apply [@Puig Theorem 7.2] to this last embedding.
\[basic\_graded\_def\] The $\Gamma$-graded module $\ddot{M}$ determines a [*basic graded Morita equivalence*]{} between $A$ and $A'$ if $\ddot{M}$ determines a graded Morita equivalence between $A$ and $A'$, and any of the equivalent statements in Corollary \[equiv\_statements\_for\_basic\] hold.
\[c:truncation\] Assume that the $\Gamma$-graded $\mathcal{O}(G\times G')$-module $\ddot{M}$ determines a basic graded Morita equivalence between $A$ and $A'.$ Let $\Lambda$ be a subgroup of $\Gamma$, and let $H:=\omega^{-1}(\Lambda)$ and $H':={\omega'}^{-1}(\Lambda)$.
Then the $\Lambda$-graded $\mathcal{O}(H\times H')$-module $\ddot{M}_\Lambda=\bigoplus_{x\in \Lambda}\ddot{M}_x$ determines a basic graded Morita equivalence between $A_\Lambda$ and $A'_\Lambda.$
We have the isomorphism $$A_\Lambda\simeq {\operatorname{End}}_{\mathcal{O}}(\ddot{M}_\Lambda)^{1\times H'}$$ of $\Lambda$-graded algebras, which means that $M$ induces an $\Lambda$-graded Morita equivalence between $A_\Lambda$ and $A'_\Lambda.$ Denote $$\ddot H:=\{(g,g')\in H\times H'\mid \omega(g)=\omega(g') \}.$$ As $\mathcal{O}(H\times H')$-modules, we have that $$M\mid \sum_{(h,h')\in [\ddot{H}\setminus \ddot{G}/\ddot{P}]}{\operatorname{Ind}}_{\ddot{P}^{(h,h')}\cap\ddot{H}}^{\ddot H}(\ddot{N}^{(h,h')}),$$ where $\ddot{P}\leq \ddot{G}$ is the above vertex of $M$ in $\ddot{G}$. We may choose a vertex $\ddot{Q}$ of $M$ in $\ddot{H}$ such that $\ddot{Q}\leq \ddot{P}.$ Denote by $\sigma_{\ddot{Q}}$ and $\sigma'_{\ddot{Q}}$ the projections $\ddot{Q}\to Q$ and $\ddot{Q}\to Q'$ respectively, determined by the projections $\ddot{H}\to H$ and $\ddot{H}\to H'.$ We obtain the commutative diagram $$\begin{xy} \xymatrix{ \ddot{P} \ar[r]^{\sigma} &P
\\ \ddot{Q}\ar[u]\ar[r]^{\sigma_{\ddot{Q}}} &Q\ar[u] } \end{xy},$$ where the vertical maps are the inclusions. A similar commutative diagram exists for $P'$ and $Q'.$ Finally, if $\sigma$ is an isomorphism, $\sigma_{\ddot{Q}}$ is also an isomorphism.
The graded structure of the extended Brauer quotient
====================================================
We keep the notations introduced in \[sect:prelim\]. Let $Q$ be a $p$-subgroup of $G$ and consider the subgroup $$K:={\operatorname{Aut}}^\Gamma(Q)=\{\varphi\in {\operatorname{Aut}}(Q)\mid \varphi(u)\in uN \mbox{ for all } u\in Q\}$$ of grade-preserving automorphisms of $Q$. For any $\varphi\in K$, consider the $\varphi$-twisted diagonal map $$\Delta_\varphi:Q\to Q\times Q, \qquad u\mapsto(\varphi(u),u).$$ Denote by $N^K_G(Q)$ the inverse image of $K$ via the group homomorphism $$N_G(Q)\to {\operatorname{Aut}}(Q).$$ It is clear that $C_N(Q)$ is a normal subgroup of $N^K_G(Q).$ Let $N_{\mathcal{O}G}^\varphi(Q)$ be the set of $\Delta_\varphi(Q)$-fixed elements of $\mathcal{O}G$, and let $$\bar N^\varphi_{kG}(Q)=(\mathcal{O}G)(\Delta_\varphi(Q)).$$ The aim of this section is to point out that the extended Brauer quotient $$\bar{N}^K_{kG}(Q)=\bigoplus_{\varphi\in K}\bar{N}^{\varphi}_{kG}(Q)$$ introduced in [@Puig3] (see also [@CT] for the generalization to $N$-interior $G$-algebras), and the isomorphism of [@Puig3 Theorem 3.5] admit the following $N_G^K(Q)/C_N(Q)$-graded structure.
\[gr\_ext\_quotient\] There exists an isomorphism $$\psi^K:kN_G^K(Q)\simeq \bar{N}^K_{kG}(Q)$$ of $N_G^K(Q)/C_N(Q)$-graded $N_G^K(Q)$-interior $N_G(Q)$-algebras.
The set $G$ is a $Q\times Q$-invariant $\mathcal{O}$-basis of $\mathcal{O}G$, so it is also $\Delta_{\varphi}(Q)$-invariant for any $\varphi\in K.$ Hence for any $\varphi\in K,$ the subalgebra $\mathcal{O}G^{\Delta_{\varphi}(Q)}$ of $\Delta_{\varphi}(Q)$-fixed elements of $\mathcal{O}G$ has as basis the class sums $$\sum_{u\in Q}\varphi (u)gu^{-1},$$ where $g\in G.$ Now, if $\varphi (u)gu^{-1}=g$, we get $\varphi (u)=u^{g^{-1}}$ for any $u\in Q,$ forcing $g\in N_G^K(Q).$
For any $x\in N_G^K(Q)$ and any $u\in Q$, denote $\varphi_x(u)=u^{x^{-1}}$. Then it is easy to see that we get $$(\mathcal{O}Nz)^{\Delta_{\varphi_x}(Q)}\subseteq \mathcal{O}Nz,$$ as $\mathcal{O}$-modules, for any $z,x\in N_G^K(Q).$ We can organize the algebra $$\mathcal{O}N_G^K(Q)N=\bigoplus_{z\in [N_G^K(Q)/C_N(Q)] }\mathcal{O}Nz,$$ such that any class $Nz$ is a $\Delta_{\varphi_x}(Q)$-invariant $\mathcal{O}$-basis of the module $\mathcal{O}Nz$, which has a $\Delta_{\varphi_x}(Q)$-fixed element (for a unique representative $x\in [N_G^K(Q)/C_G(Q)]$) if and only if $zC_G(Q)=xC_G(Q).$
Since in our situation the map sending $x$ to $\varphi_x$ induces the monomorphism $$N_{G}^K(Q)/C_G(Q)\simeq K,$$ and since for any $z\in [N_G^K(Q)/C_N(Q)]$ there is $t\in [C_G(Q)/C_N(Q)]$ and $x\in [N_{G}^K(Q)/C_G(Q)]$ such that $z=xt,$ from the above and [@CT Proposition 2.5] we obtain that $$\begin{aligned}
\bar{N}^K_{kG}(Q)&=\bar{N}_{kN}^{K}(Q)\bigoplus\left(\bigoplus_{\substack{z\in [N_G^K(Q)/C_N(Q)]\\z\notin N} }\bar{N}_{\mathcal{O}Nz}^{\varphi_x}(Q)\right)\\
&=kN_N(Q)\bigoplus\left(\bigoplus_{\substack{z\in [N_G^K(Q)/C_N(Q)]\\z\notin N} }\bar{N}_{\mathcal{O}Nz}^{\varphi_x}(Q)\right).\end{aligned}$$ It is also easy to check that $$kC_N(Q)z\subseteq N^{\varphi_x}_{\mathcal{O}Nz}(Q)$$ for any $z\in [N_G^K(Q)/C_N(Q)].$ We define $\psi^K$ to be the sum $$\psi^K=\bigoplus_{z\in [N_G^K(Q)/C_N(Q)]}\psi^K_z$$ of the family $\psi^K_z$ of module homomorphisms, where each $\psi^K_z$ is the restriction to $kC_N(Q)z$ of ${\operatorname{Br}}_{\Delta_{\varphi_x}(Q)}^{kG}.$ Since $\psi^K_1$ is an isomorphism, so is $\psi^K.$
We can define an $N_G(Q)$-action on $K$ as follows. For any $\varphi\in K$ and any $y\in N_G(Q),$ $\varphi^y$ is the automorphism of $Q$ given by $$\varphi^y(u)=y^{-1}\varphi(yuy^{-1})y,$$ for any $u\in Q.$ So, for any $\varphi\in K$ and any $y\in N_G(Q),$ the element $a\in \mathcal{O}G^{\Delta_{\varphi}(Q)}$ verifies $a^y\in \mathcal{O}G^{\Delta_{\varphi^y}(Q)}.$ Hence $\bar{N}^K_{kG}(Q)$ is a $N_G(Q)$-algebra. Now, by its definition, $\psi^K$ is an isomorphism of $N_G(Q)$-algebras.
\[ext\_Br\_Quot\_for\_local\] Let $T$ be a subgroup of $N^K_G(Q)$ such that $Q\leq T$, and consider a pointed group $T_{\mu}$ on $\mathcal{O}N$ such that ${\operatorname{Br}}^{\mathcal{O}N}_Q(\mu)\neq 0.$ Then we also have that ${\operatorname{Br}}^{\mathcal{O}G}_Q(\mu)\neq 0.$ Choose a primitive idempotent $j\in \mu.$ The $\mathcal{O}Q$-interior algebra $j\mathcal{O}Gj$ is also $G/N$-graded, and consequently $\bar{N}^K_{i\mathcal{O}Gi}(Q)$ is $N_G^K(Q)/C_N(Q)$-graded as well. Moreover, the $G/N$-graded $\mathcal{O}T$-interior algebra embedding $$j\mathcal{O}Gj\to \mathcal{O}G$$ gives the $N_G^K(Q)/C_N(Q)$-graded $\mathcal{O}T$-interior algebra embedding $$\bar{N}^K_{jkGj}(Q)\to \bar{N}^K_{kG}(Q).$$ In $kC_G(Q)$, we clearly have the equality $${\operatorname{Br}}_{Q}^{\mathcal{O}N}(j)={\operatorname{Br}}_{Q}^{\mathcal{O}G}(j),$$ and then [@Puig3 Corollary 3.7] still holds in our case, giving the isomorphism $$(kN_G^K(Q))_{\widehat{{\operatorname{Br}}_Q(\mu)}}\simeq \bar{N}^K_{(kG)_{\mu}}(Q)$$ of $N_G^K(Q)/C_N(Q)$-graded $\mathcal{O}T$-interior algebras.
\[ext\_Br\_Quot\_for\_S\] Let $P$ be a $p$-subgroup of $G$ and denote in this section, for the moment, the $kP$-interior algebra $$S= {\operatorname{End}}_k(W),$$ where $W$ be an endopermutation $kP$-module such that $$({\operatorname{End}}_k(W))(P)\neq 0.$$ Let $Q$ be subgroup of $P,$ and let $x\in [N_G^K(Q)/C_G(Q)].$ Then the $\mathcal{O}P$-interior algebra $S\otimes_{\mathcal{O}}\mathcal{O}G$ is $G/N$-graded and we have that $$\begin{aligned}
(S\otimes_{\mathcal{O}}\mathcal{O}G)(\Delta_{\varphi_x}(Q))&=S(\Delta_{\varphi_x}(Q))\otimes_k (\mathcal{O}G)(\Delta_{\varphi_x}(Q))\\
&=\bigoplus_{t\in [C_G(Q)/C_N(Q), z=xt]}(S(\Delta_{\varphi_x}(Q))\otimes_k \bar{N}_{\mathcal{O}Nz}^{\varphi_x}(Q)).\end{aligned}$$ Consequently, the $N_P^K(Q)$-interior algebra $$\bar{N}^K_{S\otimes_{\mathcal{O}}\mathcal{O}G}(Q)=\bigoplus_{z\in [N_G^K(Q)/C_N(Q)]}S(\Delta_{\varphi_x}(Q))\otimes_k \bar{N}_{\mathcal{O}Nz}^{\varphi_x}(Q)$$ is also $N_G^K(Q)/C_N(Q)$-graded. We get yet another $N_G^K(Q)/C_N(Q)$-graded $N_P^K(Q)$-interior algebra by setting $$S(Q)\otimes_k\bar{N}^K_{\mathcal{O}G}(Q)=\bigoplus _{z\in [N_G^K(Q)/C_N(Q)]}S(Q)\otimes_k \bar{N}_{\mathcal{O}Nz}^{\varphi_x}(Q).$$
By using the notation introduced in Remarks \[ext\_Br\_Quot\_for\_local\] and \[ext\_Br\_Quot\_for\_S\], and by adapting the proof of [@Puig3 Proposition 3.9] we obtain:
\[gr\_on\_tens\_simple\_alg\] Let $Q_{\delta''}$ be a local pointed group on $A_1$, and let $$\tilde{K}:=F_{S}(Q_{\delta''})\cap K.$$ Then there is an isomorphism $$\bar{N}_{S\otimes j(\mathcal{O}G)j}^K(Q)\simeq S(Q)\otimes_k\bar{N}^K_{j(\mathcal{O}G)j}(Q)$$ of $N_G^{\tilde{K}}(Q)/C_N(Q)$-graded $T$-interior algebras.
Local basic graded Morita equivalences
======================================
In this section we assume that the bimodule $\ddot{M}$ induces a basic graded Morita equivalence between $A$ and $A'$, as defined in Section \[sect:basic\].
For any subgroup $Q$ in $P,$ denoting $\ddot{Q}=\sigma^{-1}(Q)$ and $\sigma'(\ddot{Q})=Q',$ we have that $$N_P(Q)\simeq N_{\ddot{P}}(\ddot{Q})\simeq N_{P'}(Q'),$$ and the action of $N_{\ddot{G}}(\ddot{Q})$ on $\ddot{Q}$ determines the group isomorphisms $$N_G(Q)/C_N(Q)\simeq N_{G'}(Q')/C_{N'}(Q')$$ and $$N_G(Q)/C_G(Q)\simeq N_{G'}(Q')/C_{G'}(Q').$$
\[s:pointbij\] We claim that there is a bijection between the set of local points of $Q$ on $(A_1)_{\gamma}$ and the set of local points of $Q'$ on $(A'_1)_{\gamma'}.$ Indeed, by our assumption, the isomorphism $$e:A_{\gamma}\longrightarrow ({\operatorname{Ind}}_{\sigma}({\operatorname{End}}_{\mathcal{O}}(\ddot{N})\otimes_{\mathcal{O}}{\operatorname{Res}}_{\sigma'}(A'_{\gamma'})))_{\hat{\gamma}}$$ of $\Gamma$-graded $\mathcal{O}P$-interior algebras introduced in Theorem \[main\] determines the embedding $$\label{8} A_{\gamma}\to {\operatorname{Res}}_{\sigma^{-1}}(S)\otimes_{\mathcal{O}}{\operatorname{Res}}_{\sigma'\circ \sigma^{-1}}(A')_{\gamma'}$$ of $\Gamma$-graded $\mathcal{O}P$-interior algebra, and in particular, the embedding $$\label{9} (A_1)_{\gamma}\to {\operatorname{Res}}_{\sigma^{-1}}(S)\otimes_{\mathcal{O}}{\operatorname{Res}}_{\sigma'\circ \sigma^{-1}}(A'_1)_{\gamma'}$$ of $\mathcal{O}(P\cap N)$-interior $P$-algebras.
Since $S(\ddot{Q})$ is a simple $k$-algebra, following [@Puig 7.6], we still have the embedding $$\label{10}(A_1)_{\gamma}(Q)\to {\operatorname{Res}}_{\sigma^{-1}}(S(\ddot{Q}))\otimes_{\mathcal{O}}{\operatorname{Res}}_{\sigma'\circ \sigma^{-1}}(A_1')_{\gamma'}(Q').$$ of $N_P(Q)$-algebras, and this proves that claim.
Furthermore, if $Q_{\delta}$ and $Q'_{\delta'}$ correspond under this bijection, then by (\[9\]) we also get the embedding $$\label{11}A_{\delta}\to {\operatorname{Res}}_{\sigma^{-1}}(S_{\ddot{\delta}})\otimes_{\mathcal{O}}{\operatorname{Res}}_{\sigma'\circ \sigma^{-1}}A_{\delta'}$$ of $\Gamma$-graded $Q$-interior algebras, where $\ddot{\delta}$ is the unique point of $\ddot{Q}$ on $S$ lifting the identity of $S(\ddot{Q}).$
Consider the local pointed group $Q_{\delta}$ on $A_1$ such that $Q_{\delta}\leq P_{\gamma}$ corresponding to the local pointed group $Q'_{\delta'}$ on $A'_1$ such that $Q'_{\delta'}\leq P'_{\gamma'}.$ Let $b_{\delta}$ be the unique block of $kC_N(Q)$ determined by $\delta$, and let $b'_{\delta'}$ be the unique block of $kC_{N'}(Q')$ determined by $\delta'.$ If $N_G(Q)_{b_{\delta}}$ denotes the stabilizer in $N_G(Q)$ of the block $b_{\delta},$ and similarly $N_{G'}(Q')_{b'_{\delta'}}$ denotes the stabilizer of $b'_{\delta'},$ then we have the inclusions $$N_G(Q_{\delta})\subseteq N_G(Q)_{b_{\delta}}$$ and $$N_{G'}(Q'_{\delta'})\subseteq N_{G'}(Q')_{b'_{\delta'}}.$$
We define $(G,\Gamma)$-fusions in terms of $\Gamma$-graded bimodules, in a way similar to [@L2 7.2–7.5]
We say that an element $\varphi\in K={\operatorname{Aut}}^\Gamma(Q)$ is a $(G,\Gamma)$-fusion of $Q_\delta$ in $A_\gamma$ if there is an homogeneous isomorphism of some degree of $\Gamma$-graded bimodules $$i\mathcal{O}Gj\simeq(i\mathcal{O}Gj)_\varphi,$$ where $i\in \gamma$ and $j\in\delta$.
\[the\_tricky\_lemma\] With the above notations, and identifying $Q$, $Q'$ and $\ddot Q$ via $\sigma$ and $\sigma'$, the following statements hold.
[1)]{} We have the group isomorphism $$N^{{K}}_G(Q_{\delta})/C_N(Q)\simeq F_{A_\gamma}^{\Gamma}(Q_{\delta}).$$
[2)]{} The $\Gamma$-graded $\mathcal{O}Q$-interior algebra embedding [(\[11\])]{} gives the inclusion $$F_{A_\gamma}^{\Gamma}(Q_{\delta})\subseteq F_S({\ddot Q}_{\ddot \delta})$$ and the equality $$F_{A_\gamma}^{\Gamma}(Q_{\delta})=F_{{A'}_{\gamma'}}^{\Gamma} (Q'_{\delta'}).$$
1\) As in [@Puig2 Theorem 3.1] (see also [@L2 7.2–7.5]), an homogeneous isomorphism between the $(i\mathcal{O}Gi,\mathcal{O}Q)$-bimodules $i\mathcal{O}Gj$ and $(i\mathcal{O}Gj)_\varphi$ is given by right multiplication with an element $x\in N_G^K(Q_\delta)$. The condition $x\in C_N(Q)$ means that $x$ defines a graded bimodule isomorphism of degree $1$ and a trivial automorphism of $Q$.
2\) Both statements follow by the argument of [@KP 1.17] and by taking into account that homogeneous isomorphisms of graded bimodules are induced by right multiplication with homogeneous invertible elements.
Assume that $Q\subseteq N_G^K(Q_\delta)$. Let $T$ be a defect group of $b_{\delta}$ regarded as a primitive idempotent of $kC_N(Q)^{N^{K}_G(Q_{\delta})}.$ Then the group $T$ can be chosen such that $Q_\delta \le T_{\mu}\leq P_{\gamma},$ where $\nu $ is a local point of $T$ on $kC_N(Q).$
Furthermore, in this situation we have that $T=N^{K}_P(Q_{\delta})$, and $T_{{\operatorname{Br}}_Q(\mu)}$ is a defect pointed group of $N^{K}_G(Q_{\delta})_{\{b_{\delta}\}}.$
Let $\beta\in \mathcal{O}N^{N^{K}_G(Q_{\delta})}$ be the unique point lifting $b_{\delta}$ via the $\mathcal{O}$-algebra epimorphism $${\operatorname{Br}}_Q:\mathcal{O}N^{N^{K}_G(Q_{\delta})}\to kC_N(Q)^{N^{K}_G(Q_{\delta})}.$$ It is not difficult to see that this epimorphism actually restricts to the epimorphism $${\operatorname{Br}}_Q:\mathcal{O}N_T^{N^{K}_G(Q_{\delta})}\to kC_N(Q)_T^{N^{K}_G(Q_{\delta})},$$ and then $T$ is also a defect group of $\beta.$ The equality $${\operatorname{Br}}_Q(\delta)b_{\delta}={\operatorname{Br}}_Q(\delta){\operatorname{Br}}_Q(\beta)={\operatorname{Br}}_Q(\delta)$$ shows that we have $Q_{\delta}\leq N^{K}_G(Q_{\delta})_{\beta},$ and then we also get $N^{K}_G(Q_{\delta})_{\beta}\leq G_{\{b\}}.$ Now $b$ lies in $$(\mathcal{O}N)^G_P\subseteq \sum_{x\in [N^{K}_G(Q_{\delta})\setminus G/P]}(\mathcal{O}N)^{N^{K}_G(Q_{\delta})}_{N^{K}_G(Q_{\delta})\cap P^x},$$ and since $bj=j=jb$ for any $j\in \beta$ we obtain $T\leq P^x,$ for some $x\in G.$
Hence replacing each $P_{\gamma}$ and $Q_{\delta}$ by a $G$-conjugate, such that we still have $Q_{\delta}\leq P_{\gamma},$ we may assume that $T\leq P.$ More exactly $T\leq N_P(Q_{\delta}).$ The local pointed group $P_{\gamma}$ forces the existence of at least one local point $\mu \subseteq ({\mathcal{O}}H)^{N_P(Q_{\delta})}$ with the property $$N_P(Q_{\delta})_{\mu}\leq P_{\gamma}.$$ The inclusion $Q\leq N_P(Q_{\delta})\leq N_G(Q_{\delta})$ shows that we may find, if necessary, some $G$-conjugate of $Q_{\delta }$ with $$Q_{\delta}\leq N_P(Q_{\delta})_{\mu}\leq P_{\gamma}$$ and $$Q_{\delta}\leq N_P(Q_{\delta})_{\mu}\leq N_G(Q_{\delta})_{\beta},$$ since $\delta$ determines $\beta.$
Now $T,$ the defect group of $\beta,$ verifies $T\leq N_P(Q_{\delta}).$ The local point $\mu$ determines a local point $\bar{\mu}$ of $T$ on ${\mathcal{O}}H$ with $T_{\bar{\mu}}\leq N_P(Q_{\delta})_{\mu}.$ The maximality of $T_{\bar{\mu}}$ forces the equality $T_{\bar{\mu}}=N_P(Q_{\delta})_{\mu}.$ The commutative diagram
$$\begin{xy} \xymatrix{ kC_N(Q)^T \ar[r]^{{\operatorname{Br}}^{kC_N(Q)}_T} &kC_N(T)
\\ \mathcal{O}N^T\ar[u]_{{\operatorname{Br}}^{\mathcal{O}N}_Q}\ar[ur]_{{\operatorname{Br}}_T^{\mathcal{O}N}} } \end{xy}$$ proves ${\operatorname{Br}}^{kC_N(Q)}_T({\operatorname{Br}}_Q(\mu))\neq 0$ and then $T_{{\operatorname{Br}}_Q(\mu)}$ is a defect pointed group of $N_{G}(Q_{\delta})_{b_{\delta}}.$
Let $T'_{\mu'}\leq P'_{\gamma'}$ be the pointed group corresponding to $T_{\mu}$ under the bijection given in \[s:pointbij\] (\[10\]). It is not difficult to prove that $T'_{{\operatorname{Br}}_{Q'}(\mu')}$ is a defect pointed group of $N^{K}_{G'}(Q'_{\delta'})_{\{b'_{\delta'}\}}$ and that $T'=N^{K}_{P'}(Q'_{\delta'}).$ By Lemma \[the\_tricky\_lemma\] we have $$N^{K}_{G}(Q_{\delta })/C_N(Q)= N^{K}_{G'}(Q'_{\delta'})/C_{N'}(Q')\subseteq F_{S}({\ddot Q}_{\ddot{\delta}}).$$ Moreover, embedding (\[8\]) provides the embedding $$(\mathcal{O}G)_{\mu}\to S\otimes_{\mathcal{O}}(\mathcal{O}G')_{\mu'}$$ of $\Gamma$-graded $\mathcal{O}T$-interior algebras, which according to Proposition \[gr\_ext\_quotient\] and Proposition \[gr\_on\_tens\_simple\_alg\], determines the embedding $$(kN_G^{K}(Q_{\delta}))_{\widehat{{\operatorname{Br}}_Q(\mu)}}\to S(\ddot{Q})\otimes_k(kN_{G'}^{K}(Q'_{\delta'}))_{\widehat{{\operatorname{Br}}_{Q'}(\mu')}}$$ of $\mathcal{O}T$-interior $N^{K}_G(Q_{\delta})/C_N(Q)$-graded algebras. Recall also that here $$S(\ddot{Q})={\operatorname{End}}_{k}(\ddot{N}_{\ddot{Q}}),$$ where ${\ddot N}_{\ddot{Q}}$ is a uniquely determined endo-permutation $kT$-module.
Consider the natural maps $\bar\omega:N^{K}_G(Q_{\delta})\to N^{K}_G(Q_{\delta})/C_N(Q)$ and $\bar\omega':N^{K}_{G'}(Q'_{\delta'})\to N^{K}_{G'}(Q'_{\delta'})/C_{N'}(Q')$, so in view of Lemma \[the\_tricky\_lemma\] 2), we have, as in \[s:GN\], the corresponding subgroup $$\{(g,g')\in N^{K}_G(Q_{\delta})\times N^{K}_{G'}(Q'_{\delta'})\mid \bar\omega (g)=\bar\omega(g') \}$$ of $N^{K}_G(Q_{\delta})\times N^{K}_{G'}(Q'_{\delta'})$. It is clear that this subgroup coincides with $N_{\ddot{G}}^K(\ddot{Q}_{\delta\times \delta'})$, where $\delta\times\delta'$ is the unique point of $\ddot{Q}$ on $A_1\otimes A'_1$ such that ${\operatorname{Br}}_Q(\delta)\times {\operatorname{Br}}_{Q'}(\delta')\subseteq {\operatorname{Br}}_{\ddot{Q}}(\delta\times \delta')$.
By applying Theorem \[main\] in this situation, we deduce the main result of this section.
\[t:localequiv\] Assume that $Q\subseteq N_G^K(Q_\delta)$. The indecomposable $\hat k_*\hat{\bar{N}}_{N^K_G(Q_{\delta})}(T_{\mu})$-module $V_{A_1}(T_{\mu})$ determines an indecomposable $kN_{\ddot{G}}^K(\ddot{Q}_{\delta\times \delta'})$-direct summand $Y$ of $${\operatorname{Ind}}_{\Delta(T)}^{N_{\ddot{G}}^K(\ddot{Q}_{\delta\times \delta'})}(\ddot{N}_{\ddot{Q}}),$$ such that the bimodule $${\operatorname{Ind}}_{N_{\ddot{G}}^K(\ddot{Q}_{\delta\times \delta'})}^{N^{K}_G(Q_{\delta})\times N^{K}_{G'}(Q'_{\delta'})}(Y)$$ induces a basic $N^K_G(Q_\delta)/C_N(Q)$-graded Morita equivalence between $kN^{K}_G(Q_{\delta})b_{\delta}$ and $kN^{K}_{G'}(Q'_{\delta'})b'_{\delta'}.$
1\) If in Theorem \[t:localequiv\] we have $G=N$, then we obtain the main result of Puig and Zhou [@Puig3 Theorem 1.4], and also of Hu [@Hu Theorem 1.1].
2\) On the other hand, if we assume that $b$ and $b'$ are the principal blocks of $\mathcal{O}N$ and $\mathcal{O}N'$ respectively, $\Gamma$ is a $p'$-group, $M$ is a splendid $\mathcal{O}(N\times N')$-module, and we take $K=1$ instead of $K={\operatorname{Aut}}^\Gamma(Q)$ (hence $N^K_G(Q)=C_G(Q)$), the we obtain [@Marcus Corollary 3.9.a)].
[99]{}
<span style="font-variant:small-caps;">Alperin, J.L.</span>, [*Local Representation Theory*]{}, Cambridge University Press, 1986. <span style="font-variant:small-caps;">Coconeţ, T.</span> and <span style="font-variant:small-caps;">Marcus, A.</span>, [*Group graded Hecke interior algebras*]{}, Mathematica [**56 (79)**]{} (2014), 117–126. <span style="font-variant:small-caps;">Coconeţ, T.</span> and <span style="font-variant:small-caps;">Todea, C.</span>, [*The extended Brauer quotient of $N$-interior $G$-algebras*]{}, J. Algebra [**396**]{} (2013), 10–17. <span style="font-variant:small-caps;">Fan, Y.</span> and <span style="font-variant:small-caps;">Puig, L.</span>, [*On Blocks with Nilpotent Coefficient Extensions*]{}, Algebras and Representation Theory [**1**]{} (1998) 27–73. <span style="font-variant:small-caps;">Harris, M.E.</span>, [*Splendid derived equivalences for blocks of finite groups.*]{} J. Lond. Math. Soc., [**(2) 60**]{} (1999), 71–82. <span style="font-variant:small-caps;">Hu, Xueqin</span>, [*A note on the basic Morita equivalences.*]{} Sci. China, Math. [**57**]{} (2014), 483-490. <span style="font-variant:small-caps;">Külshammer, B.</span> and <span style="font-variant:small-caps;">Puig, L.</span>, [*Extensions of nilpotent blocks*]{}. Invent. Math. [**102**]{} (1990), 17–71. <span style="font-variant:small-caps;">Linckelmann, M.</span>, [*On derived equivalences and local structure of blocks of finite groups*]{}, Turkish J. Math. [**22**]{} (1998) 93–107. <span style="font-variant:small-caps;">Linckelmann, M.</span>, [*On splendid derived and stable equivalences between blocks and finite groups*]{}. J. Algebra [**242**]{} (2001), 819–843 . <span style="font-variant:small-caps;">Marcus, A.</span>, *On equivalences between blocks of group algebras: reduction to the simple components*, J. Algebra [**184**]{} (1996), 372–396. <span style="font-variant:small-caps;">Marcus, A.</span>, [*Representation Theory of Finite Groups Graded Algebras*]{}, Nova Science Publishers, Commack, NY, 1999. <span style="font-variant:small-caps;">Puig, L.</span>, [*Local fusions in block source algebras*]{}, J. Algebra [**104**]{} (1986), 358–369. <span style="font-variant:small-caps;">Puig, L.</span>, [*On the Local Structure of Morita and Rickard Equivalences between Brauer Blocks*]{}, Progress in Mathematics, vol. 178. Birkhauser, Basel 1999. <span style="font-variant:small-caps;">Puig, L.</span>, [*Blocks of Finite Groups. The Hyperfocal Subalgebra of a Block*]{}, Springer, Berlin, 2002. <span style="font-variant:small-caps;">Puig, L.</span> and <span style="font-variant:small-caps;">Zhou, Y.</span>, [*A local property of basic Morita equivalences.*]{} Math. Z. [**256**]{} (2007), 551–562. <span style="font-variant:small-caps;">Puig, L.</span> and <span style="font-variant:small-caps;">Zhou, Y.</span>, [*A local property of basic Rickard equivalences.*]{} J. Algebra [**322**]{} (2009), 1946-1973. <span style="font-variant:small-caps;">Rickard, J.</span>, [*Splendid equivalences: Derived categories and permutation modules.*]{} Proc. Lond. Math. Soc., [**(3) 72**]{} (1996), 331–358 . <span style="font-variant:small-caps;">Thévenaz, J.</span>, [*$G$-Algebras and Modular Representation Theory*]{}, Clarendon Press, Oxford 1995.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We address the problem of severe class imbalance in unsupervised domain adaptation, when the class spaces in source and target domains diverge considerably. Till recently, domain adaptation methods assumed the aligned class spaces, such that reducing distribution divergence makes the transfer between domains easier. Such an alignment assumption is invalidated in real world scenarios where some source classes are often under-represented or simply absent in the target domain. We revise the current approaches to class imbalance and propose a new one that uses latent codes in the adversarial domain adaptation framework. We show how the latent codes can be used to disentangle the silent structure of the target domain and to identify under-represented classes. We show how to learn the latent code reconstruction jointly with the domain invariant representation and use them to accurately estimate the target labels.'
author:
- |
Boris Chidlovskii\
Naver Labs Europe, Meylan 38240, France
bibliography:
- 'DomAd-all.bib'
- 'mybib-da1.bib'
- 'MyLibrary.bib'
- 'DomAd.bib'
title: Using Latent Codes for Class Imbalance Problem in Unsupervised Domain Adaptation
---
Conclusion {#sec:conclusion}
==========
We propose a novel adversarial network for unsupervised domain adaptation, specific for severe class imbalance when the class distributions in the source and target domains diverge considerably. We address real world scenarios when source classes are under-represented of absent at all in the target domain. We test domain adaptation methods under the class imbalance using an extended evaluation protocol. We use latent codes in the adversarial learning aimed at disentangling the salient structure of the target domain. The auxiliary network is introduced in the adversarial domain adaptation framework to reconstruct latent codes; it plays the role of additional predictor when facing the severe class imbalance.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
This paper surveys some well-established approaches on the approximation of Bayes factors used in Bayesian model choice, mostly as covered in [@chen:shao:ibrahim:2000]. Our focus here is on methods that are based on importance sampling strategies—rather than variable dimension techniques like reversible jump MCMC—, including: crude Monte Carlo, maximum likelihood based importance sampling, bridge and harmonic mean sampling, as well as Chib’s method based on the exploitation of a functional equality. We demonstrate in this survey how these different methods can be efficiently implemented for testing the significance of a predictive variable in a probit model. Finally, we compare their performances on a real dataset.
[**Keywords:**]{} [Bayesian inference; model choice; Bayes factor; Monte Carlo; Importance Sampling; bridge sampling; Chib’s functional identity; supervised learning; probit model]{}
author:
- |
[[J.-M. Marin]{}$^{1,3,}$[^1]]{} and [[C.P. Robert]{}$^{2,3,}$[^2]]{}\
$^1$Institut de Mathématiques et Modélisation de Montpellier,\
$^2$Université Paris Dauphine, and\
$^3$Centre de Recherche en Economie et Statistique, INSEE, Paris
title: '[**Importance sampling methods for Bayesian discrimination between embedded models**]{}'
---
Introduction
============
The contribution of Jim Berger to the better understanding of Bayesian testing is fundamental and wide-ranging, from establishing the fundamental difficulties with $p$-values in [@berger:sellke:1987] to formalising the intrinsic Bayes factors in [@berger:pericchi:1996], to solving the difficulty with improper priors in [@berger:pericchi:varshavsky:1998], and beyond! While our contribution in this area is obviously much more limited, we aim at presenting here the most standard approaches to the approximation of Bayes factors.
The Bayes factor indeed is a fundamental procedure that stands at the core of the Bayesian theory of testing hypotheses, at least in the approach advocated by both [@jeffreys:1939] and by [@jaynes:2003]. (Note that [@robert:chopin:rousseau:2009], provides a reassessment of the crucial role of [@jeffreys:1939] in setting a formal framework for Bayesian testing as well as for regular inference.) Given an hypothesis $H_0:\,\theta\in\Theta_0$ on the parameter $\theta\in\Theta$ of a statistical model, with observation $y$ and density $f(y|\theta)$, under a compatible prior of the form $$\pi(\Theta_0) \pi_0(\theta) + \pi(\Theta_0^c) \pi_1(\theta)\,,$$ the [*Bayes factor*]{} is defined as the posterior odds to prior odds ratio, namely $$B_{01}(y) = \displaystyle{\frac{\pi(\Theta_0|y)}{ \pi(\Theta_0^c|y)}\bigg/\frac{\pi(\Theta_0)}
{\pi(\Theta_0^c)}} = {\displaystyle{\int_{\Theta_0} f(y|\theta) \pi_0(\theta) \text{d}\theta}}\bigg/ {
\displaystyle{\int_{\Theta_0^c} f(y|\theta) \pi_1(\theta) \text{d}\theta} }\,.$$Model choice can be considered from a similar perspective, since, under the Bayesian paradigm (see, e.g., [@robert:2001]), the comparison of models $$\mathfrak{M}_i : y \sim f_i(y|\theta_i), \quad\theta_i \sim \pi_i(\theta_i), \quad\theta_i\in\Theta_i , \quad i \in \mathfrak{I}\,,$$ where the family $\mathfrak{I}$ can be finite or infinite, leads to posterior probabilities of the models under comparison such that $$\mathbb{P}\left(\mathfrak{M}=\mathfrak{M}_i|y\right)\propto p_i \int_{\Theta_i}
f_i(y|\theta_i)\pi_i(\theta_i) \text{d}\theta_i\,,$$ where $p_i=\mathbb{P}(\mathfrak{M}=\mathfrak{M}_i)$ is the prior probability of model $\mathfrak{M}_i$.
In this short survey, we consider some of the most common Monte Carlo solutions used to approximate a generic Bayes factor or its fundamental component, the [*evidence*]{} $${m}_i = \int_{\Theta_i} \pi_i(\theta_i) f_i(y|\theta_i)\,\text{d}\theta_i\,,$$ aka the marginal likelihood. Longer entries can be found in [@carlin:chib:1995], [@chen:shao:ibrahim:2000], [@robert:casella:2004], or [@friel:pettitt:2008]. Note that we only briefly mention here trans-dimensional methods issued from the revolutionary paper of [@green:1995], since our goal is to demonstrate that within-model simulation methods allow for the computation of Bayes factors and thus avoids the additional complexity involved in trans-dimensional methods. While ameanable to an importance sampling technique of sorts, the alternative approach of nested sampling [@skilling:2007a] is discussed in [@chopin:robert:2007] and [@robert:wraith:2009].
The Pima Indian benchmark model
===============================
In order to compare the performances of all methods presented in this survey, we chose to evaluate the corresponding estimates of the Bayes factor in the setting of a single variable selection for a probit model and to repeat the estimation in a Monte Carlo experiment to empirically assess the variability of those estimates.
We recall that a probit model can be represented as a natural latent variable model in that, if we consider a sample $z_1,\ldots,z_n$ of $n$ independent latent variables associated with a standard regression model, i.e. such that $z_i|{{\boldsymbol \theta}}\sim\mathcal{N}\left({\mathbf{x}}_i^\text{T}{{\boldsymbol \theta}},1\right)$, where the ${\mathbf{x}}_i$’s are $p$-dimensional covariates and ${{\boldsymbol \theta}}$ is the vector of regression coefficients, then $y_1,\ldots,y_n$ such that $$ y_i = \mathbb{I}_{z_i>0}$$ is a probit sample. Indeed, given ${{\boldsymbol \theta}}$, the $y_i$’s are independent Bernoulli rv’s with $\mathbb{P}(y_i=1|{{\boldsymbol \theta}})=\Phi\left({\mathbf{x}}_i^\text{T}{{\boldsymbol \theta}}\right)$ where $\Phi$ is the standard normal cdf. The choice of a reference prior distribution for the probit model is open to debate, but the connection with the latent regression model induced [@marin:robert:2007] to suggest a $g$-prior model, ${{\boldsymbol \theta}}\sim\mathcal{N}\left(0_p,n({\mathbf{X}}^\text{T}{\mathbf{X}})^{-1}\right)$, with $n$ as the $g$ factor and ${\mathbf{X}}$ as the regressor matrix. The corresponding posterior distribution is then associated with the density $$\pi({{\boldsymbol \theta}}|{\mathbf{y}},{\mathbf{X}})\propto \prod_{i=1}^n
\left\{1-\Phi\left({\mathbf{x}}_i^\text{T}{{\boldsymbol \theta}}\right)\right\}^{1-y_i}
\Phi\left({\mathbf{x}}_i^\text{T}{{\boldsymbol \theta}}\right)^{y_i}
\times
\exp\left\{-{{\boldsymbol \theta}}^\text{T}({\mathbf{X}}^\text{T}{\mathbf{X}}){{\boldsymbol \theta}}/2n\right\}\,,
\label{eq:postprobit}$$ where ${\mathbf{y}}=(y_1,\ldots,y_n)$. In the completed model, i.e. when including the latent variables ${\mathbf{z}}=\left(z_1,\ldots,z_n\right)$ into the model, the $y_i$’s are deterministic functions of the $z_i$’s and the so-called completed likelihood is $$f({\mathbf{y}},{\mathbf{z}}|{{\boldsymbol \theta}})=(2\pi)^{-n/2}\exp\left(-\sum_{i=1}^n\left(z_i-{\mathbf{x}}_i^\text{T}{{\boldsymbol \theta}}\right)^2/2\right)
\prod_{i=1}^n\left(\mathbb{I}_{y_i=0}\mathbb{I}_{z_i\leq 0}+\mathbb{I}_{y_i=1}\mathbb{I}_{z_i>0}\right)\,.$$ The derived conditional distributions $$z_i|y_i,\theta\sim\left\{\begin{array}{ll}
\mathcal{N}_+\left({\mathbf{x}}_i^\text{T}{{\boldsymbol \theta}},1,0\right) & \mbox{if}\quad y_i=1\,, \\
\mathcal{N}_-\left({\mathbf{x}}_i^\text{T}{{\boldsymbol \theta}},1,0\right) & \mbox{if}\quad y_i=0\,,
\end{array}\right.
\label{gibbs1}$$ are of interest for constructing a Gibbs sampler on the completed model, where $\mathcal{N}_+\left({\mathbf{x}}_i^\text{T}{{\boldsymbol \theta}},\allowbreak 1,\allowbreak 0\right)$ denotes the Gaussian distribution with mean ${\mathbf{x}}_i^\text{T}{{\boldsymbol \theta}}$ and variance $1$ that is left-truncated at $0$, while $\mathcal{N}_-\left(x_i^\text{T}\theta,1,0\right)$ denotes the symmetrical normal distribution that is right-truncated at $0$. The corresponding full conditional on the parameters is given by $${{\boldsymbol \theta}}|{\mathbf{y}},{\mathbf{z}}\sim\mathcal{N}\left(\frac{n}{n+1}({\mathbf{X}}^\text{T}{\mathbf{X}})^{-1}{\mathbf{X}}^\text{T}{\mathbf{z}},\frac{n}{n+1}({\mathbf{X}}^\text{T}{\mathbf{X}})^{-1}\right)\,.
\label{gibbs2}$$ Indeed, since direct simulation from the posterior distribution of ${{\boldsymbol \theta}}$ is intractable, [@albert:chib:1993b] suggest implementing a Gibbs sampler based on the above set of full conditionals. More precisely, given the current value of ${{\boldsymbol \theta}}$, one cycle of the Gibbs algorithm produces a new value for ${\mathbf{z}}$ as simulated from the conditional distribution (\[gibbs1\]), which, when substituted into (\[gibbs2\]), produces a new value for ${{\boldsymbol \theta}}$. Although it does not impact the long-term properties of the sampler, the starting value of ${{\boldsymbol \theta}}$ may be taken as the maximum likelihood estimate to avoid burning steps in the Gibbs sampler.
Given this probit model, the dataset we consider covers a population of women who were at least 21 years old, of Pima Indian heritage and living near Phoenix, Arizona. These women were tested for diabetes according to World Health Organization (WHO) criteria. The data were collected by the US National Institute of Diabetes and Digestive and Kidney Diseases, and is available with the basic [R]{} package [@rmanual]. This dataset, used as a benchmark for supervised learning methods, contains information about $332$ women with the following variables:
- `glu`: plasma glucose concentration in an oral glucose tolerance test;
- `bp`: diastolic blood pressure (mm Hg);
- `ped`: diabetes pedigree function;
- `type`: Yes or No, for diabetic according to WHO criteria.
For this dataset, the goal is to explain the diabetes variable `type` by using the explanatory variables `glu`, `bp` and `ped`. The following table is an illustration of a classical (maximum likelihood) analysis of this dataset, obtained using the [R]{} [glm()]{} function with the probit link:
Deviance Residuals:
Min 1Q Median 3Q Max
-2.1347 -0.9217 -0.6963 0.9959 2.3235
Coefficients:
Estimate Std. Error z value Pr(>|z|)
glu 0.012616 0.002406 5.244 1.57e-07 ***
bp -0.029050 0.004094 -7.096 1.28e-12 ***
ped 0.350301 0.208806 1.678 0.0934 .
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
(Dispersion parameter for binomial family taken to be 1)
Null deviance: 460.25 on 332 degrees of freedom
Residual deviance: 386.73 on 329 degrees of freedom
AIC: 392.73
Number of Fisher Scoring iterations: 4
This analysis sheds some doubt on the relevance of the covariate `ped` in the model and we can reproduce the study from a Bayesian perspective, computing the Bayes factor $B_{01}$ opposing the probit model only based on the covariates `glu` and `bp` (model 0) to the probit model based on the covariates `glu`, `bp`, and `ped` (model 1). This is equivalent to testing the hypothesis $H_0:\theta_3=0$ since the models are nested, where $\theta_3$ is the parameter of the probit model associated with covariate `ped`. (Note that there is no intercept in either model.) If we denote by ${\mathbf{X}}_0$ the $332\times 2$ matrix containing the values of `glu` and `bp` for the $332$ individuals and by ${\mathbf{X}}_1$ the $332\times 3$ matrix containing the values of the covariates `glu`, `bp`, and `ped`, the Bayes factor $B_{01}$ is given by $$\begin{aligned}
\label{eq:bfprobit}
(2\pi)^{1/2}n^{1/2}&\frac{|({\mathbf{X}}_0^\text{T}{\mathbf{X}}_0)|^{-1/2}}{|({\mathbf{X}}_1^\text{T}{\mathbf{X}}_1)|^{-1/2}}\\
&\frac{{\displaystyle}\int_{\mathbb{R}^2}\prod_{i=1}^n \{1-\Phi\left(({\mathbf{X}}_0)_{i,\cdot}{{\boldsymbol \theta}}\right)\}^{1-y_i}
\Phi\left(({\mathbf{X}}_0)_{i,\cdot}{{\boldsymbol \theta}}\right)^{y_i}
\exp\left\{-{{\boldsymbol \theta}}^\text{T}({\mathbf{X}}_0^\text{T}{\mathbf{X}}_0){{\boldsymbol \theta}}/2n\right\}\text{d}{{\boldsymbol \theta}}}
{{\displaystyle}\int_{\mathbb{R}^3}\prod_{i=1}^n \{1-\Phi\left({\mathbf{X}}_1)_{i,\cdot}{{\boldsymbol \theta}}\right)\}^{1-y_i}
\Phi\left({\mathbf{X}}_1)_{i,\cdot}{{\boldsymbol \theta}}\right)^{y_i}
\exp\left\{-{{\boldsymbol \theta}}^\text{T}({\mathbf{X}}_1^\text{T}{\mathbf{X}}_1){{\boldsymbol \theta}}/2n\right\}\text{d}{{\boldsymbol \theta}}} \nonumber\\
&=\frac{{\displaystyle}\mathbb{E}_{\mathcal{N}_2(0_2,n({\mathbf{X}}_0^\text{T}{\mathbf{X}}_0)^{-1})}\left[\prod_{i=1}^n
\{1-\Phi\left(({\mathbf{X}}_0)_{i,\cdot}{{\boldsymbol \theta}}\right)\}^{1-y_i}\Phi\left(({\mathbf{X}}_0)_{i,\cdot}{{\boldsymbol \theta}}\right)^{y_i}\right]}
{{\displaystyle}\mathbb{E}_{\mathcal{N}_3(0_3,n({\mathbf{X}}_1^\text{T}{\mathbf{X}}_1)^{-1})}\left[\prod_{i=1}^n
\{1-\Phi\left(({\mathbf{X}}_1)_{i,\cdot}{{\boldsymbol \theta}}\right)\}^{1-y_i}\Phi\left(({\mathbf{X}}_1)_{i,\cdot}
{{\boldsymbol \theta}}\right)^{y_i}\right]}\nonumber\end{aligned}$$ using the shortcut notation that $A_{i,\cdot}$ is the $i$-th line of the matrix $A$.
The basic Monte Carlo solution
==============================
As already shown above, when testing for a null hypothesis (or a model) $H_0:\theta\in\Theta_0$ against the alternative hypothesis (or the alternative model) $H_1:\theta\in\Theta_1$, the Bayes factor is defined by $$B_{01}(y)={\displaystyle{\int_{\Theta_0} f(y|\theta_0) \pi_0(\theta_0) \text{d}\theta_0} }\bigg/
{\displaystyle{\int_{\Theta_1} f(y|\theta_1) \pi_1(\theta_1) \text{d}\theta_1} }\,.$$ We assume in this survey that the prior distributions under both the null and the alternative hypotheses are proper, as, typically, they should be. (In the case of common nuisance parameters, a common improper prior measure can be used on those, see [@berger:pericchi:varshavsky:1998; @marin:robert:2007]. This obviously complicates the computational aspect, as some methods like crude Monte Carlo cannot be used at all, while others are more prone to suffer from infinite variance.) In that setting, the most elementary approximation to $B_{01}(y)$ consists in using a ratio of two standard Monte Carlo approximations based on simulations from the corresponding priors. Indeed, for $i=0,1$: $$\int_{\Theta_i} f(y|\theta) \pi_i(\theta) \text{d}\theta=\mathbb{E}_{\pi_i}\left[f(y|\theta)\right]\,.$$ If $\theta_{0,1},\ldots,\theta_{0,n_0}$ and $\theta_{1,1},\ldots,\theta_{1,n_1}$ are two independent samples generated from the prior distributions $\pi_0$ and $\pi_1$, respectively, then $$\frac{n_0^{-1}\sum_{j=1}^{n_0}f(y|\theta_{0,j})}{n_1^{-1}\sum_{j=1}^{n_1}f(y|\theta_{1,j})}
\label{eq:bfmc}$$ is a strongly consistent estimator of $B_{01}(y)$.
In most cases, sampling from the prior distribution corresponding to either hypothesis is straightforward and fast. Therefore, the above estimator is extremely easy to derive as a brute-force evaluation of the Bayes factor. However, if any of the posterior distributions is quite different from the corresponding prior distribution—and it should be for vague priors—, the Monte Carlo evaluation of the corresponding evidence is highly inefficient since the sample will be overwhelmingly producing negligible values of $f(y|\theta_{i,j})$. In addition, if $f^2(y|\theta)$ is not integrable against $\pi_0$ or $\pi_1$, the resulting estimation has an infinite variance. Since importance sampling usually requires an equivalent computation effort, with a potentially highy efficiency reward, crude Monte Carlo approaches of this type are usually disregarded.
Figure \[fig:bfmc\] and Table \[tab:res\] summarize the results based on $100$ replications of Monte Carlo approximations of $B_{01}(y)$, using equation (\[eq:bfmc\]) with $n_0=n_1=20,000$ simulations. As predicted, the variability of the estimator is very high, when compared with the other estimates studied in this survey. (Obviously, the method is asymptotically unbiased and, the functions being square integrable in , with a finite variance. A massive simulation effort would obviously lead to a precise estimate of the Bayes factor.)
![\[fig:bfmc\] Pima Indian dataset: boxplot of 100 Monte Carlo estimates of $B_{01}(y)$ based on simulations from the prior distributions, for $2\times 10^4$ simulations.](bfmc.pdf){width="10cm" height="5cm"}
Usual importance sampling approximations
========================================
Defining two importance distributions with densities $\varpi_0$ and $\varpi_1$, with the same supports as $\pi_0$ and $\pi_1$, respectively, we have: $$B_{01}(y)={\mathbb{E}_{\varpi_0}\left[f(y|\theta)\pi_0(\theta)\big/\varpi_0(\theta)\right]}
\bigg/
{\mathbb{E}_{\varpi_1}\left[f(y|\theta)\pi_1(\theta)\big/\varpi_1(\theta)\right]}\,.$$ Therefore, given two independent samples generated from distributions $\varpi_0$ and $\varpi_1$, $\theta_{0,1},\ldots,\theta_{0,n_0}$ and $\theta_{1,1},\ldots,\theta_{1,n_1}$, respectively, the corresponding importance sampling estimate of $B_{01}(y)$ is $$\dfrac{n_0^{-1} \sum_{j=1}^{n_0} f_0(x|\theta_{0,j}) \pi_0(\theta_{0,j})/\varpi_0(\theta_{0,j})}
{n_1^{-1} \sum_{j=1}^{n_1} f_1(x|\theta_{1,j} \pi_1(\theta_{1,j})/\varpi_1(\theta_{1,j})}\,.
\label{eq:bfis}$$ Compared with the standard Monte Carlo approximation above, this approach offers the advantage of opening the choice of the representation in that it is possible to pick importance distributions $\varpi_0$ and $\varpi_1$ that lead to a significant reduction in the variance of the importance sampling estimate. This implies choosing importance functions that provide as good as possible approximations to the corresponing posterior distributions. Maximum likelihood asymptotic distributions or kernel approximations based on a sample generated from the posterior are natural candidates in this setting, even though the approximation grows harder as the dimension increases.
For the Pima Indian benchmark, we propose for instance to use as importance distributions, Gaussian distributions with means equal to the maximum likelihood (ML) estimates and covariance matrices equal to the estimated covariance matrices of the ML estimates, both of which are provided by the [R glm()]{} function. While, in general, those Gaussian distributions provide crude approximations to the posterior distributions, the specific case of the probit model will show this is an exceptionally good approximation to the posterior, since this leads to the best solution among all those compared here. The results, obtained over $100$ replications of the methodology with $n_0=n_1=20,000$ are summarized in Figure \[fig:bfmcis\] and Table \[tab:res\]. They are clearly excellent, while requiring the same computing time as the original simulation from the prior.
![\[fig:bfmcis\] Pima Indian dataset: boxplots of 100 Monte Carlo and importance sampling estimates of $B_{01}(y)$, based on simulations from the prior distributions, for $2\times 10^4$ simulations.](bfmcis.pdf){width="10cm" height="5cm"}
Bridge sampling methodology
===========================
The original version of the bridge sampling approximation to the Bayes factor [@gelman:meng:1998; @chen:shao:ibrahim:2000] relies on the assumption that the parameters of both models under comparison belong to the same space: $\Theta_0=\Theta_1$. In that case, for likelihood functions $f_0$ and $f_1$ under respectively models $\mathfrak{M}_0$ and $\mathfrak{M}_1$, the bridge representation of the Bayes factor is $$B_{01}(y)={\displaystyle{\int_{\Theta_0} f_0(y|\theta) \pi_0(\theta) \text{d}\theta} }\bigg/
{\displaystyle{\int_{\Theta_1} f_1(y|\theta) \pi_1(\theta) \text{d}\theta}}=
\mathbb{E}_{\pi_1}\left[\left.\frac{f_0(y|\theta) \pi_0(\theta)}{f_1(y|\theta) \pi_1(\theta)}\right|y\right]\,.
\label{eq:bridge1}$$ Given a sample from the posterior distribution of $\theta$ under model $\mathfrak{M}_1$, $\theta_{1,1},\ldots,\theta_{1,N}\sim \pi_1(\theta|y)$, a first bridge sampling approximation to $B_{01}(y)$ is $$N^{-1}\sum_{j=1}^{N}\frac{f_0(y|\theta_{1,j}) \pi_0(\theta_{1,j})}{f_1(y|\theta_{1,j}) \pi_1(\theta_{1,j})}\,.$$ From a practical perspective, for the above bridge sampling approximation to be of any use, the constraint on the common parameter space for both models goes further in that, not only must both models have the same complexity, but they must also be parameterised on a common ground, i.e. in terms of some specific moments of the sampling model, so that parameters under both models have a common meaning. Otherwise, the resulting bridge sampling estimator will have very poor convergence properties, possibly with infinite variance.
Equation (\[eq:bridge1\]) is nothing but a very special case of the general representation [@torrie:valleau:1977] $$B_{01}(y) = {\mathbb{E}_\varphi\left[f_0(y|\theta) \pi_0(\theta) / \varphi(\theta)\right]}\bigg/
{\mathbb{E}_\varphi\left[f_1(y|\theta) \pi_1(\theta) / \varphi(\theta)\right]} \,,$$ which holds for any density $\varphi$ with a sufficiently large support and which only requires a single sample $\theta_1,\ldots,\theta_{N}$ generated from $\varphi$ to produce an importance sampling estimate of the ratio of the marginal likelihoods. Apart from using the [*same*]{} importance function $\varphi$ for both integrals, this method is therefore a special case of importance sampling.
Another extension of this bridge sampling approach is based on the general representation $$\begin{aligned}
B_{01}(y) & = & {{\displaystyle}\int f_0(y|\theta) \pi_0(\theta) \alpha(\theta) {\pi}_1(\theta|y) \text{d}\theta }\bigg/
{{\displaystyle}\int f_1(y|\theta) \pi_1(\theta) \alpha(\theta) {\pi}_0(\theta|y) \text{d}\theta } \\
& \approx &
\frac{{\displaystyle}{n_1}^{-1} \sum_{j=1}^{n_1} f_0(y|\theta_{1,j}) \pi_0(\theta_{1,j}) \alpha(\theta_{1,j})}
{{\displaystyle}{n_0}^{-1} \sum_{j=1}^{n_0} f_1(y|\theta_{0,j}) \pi_1(\theta_{0,j}) \alpha(\theta_{0,j})}\end{aligned}$$ where $\theta_{0,1},\ldots,\theta_{0,n_0}$ and $\theta_{1,1},\ldots,\theta_{1,n_1}$ are two independent samples coming from the posterior distributions $\pi_0(\theta|y)$ and $\pi_1(\theta|y)$, respectively. That applies for any positive function $\alpha$ as long as the upper integral exists. Some choices of $\alpha$ lead to very poor performances of the method in connection with the harmonic mean approach (see Section \[sec:harmo\]), but there exists a quasi-optimal solution, as provided by [@gelman:meng:1998]: $${\alpha^\star(y) \propto \dfrac{1}{n_0{\pi}_0(\theta|y) + n_1 {\pi}_1(\theta|y)}} \,.$$ This optimum cannot be used [*per se*]{}, since it requires the normalising constants of both $\pi_0(\theta|y)$ and $\pi_1(\theta|y)$. As suggested by [@gelman:meng:1998], an approximate version uses iterative versions of $\alpha^\star$, based on iterated approximations to the Bayes factor. Note that this solution recycles simulations from both posteriors, which is quite appropriate since one model is selected via the Bayes factor, instead of using an importance weighted sample common to both approximations. We will see below an alternative representation of the bridge factor that bypasses this difficulty (if difficulty there is!).
Those derivations are, however, restricted to the case where both models have the same complexity and thus they do not apply to embedded models, when $\Theta_0\subset\Theta_1$ in such a way that $\theta_1=(\theta,\psi)$, i.e., when the submodel corresponds to a specific value $\psi_0$ of $\psi$: $f_0(y|\theta)=f(y|\theta,\psi_0)$.
The extension of the most advanced bridge sampling strategies to such cases requires the introduction of a [*pseudo-posterior density,*]{} $\omega(\psi|\theta,y)$, on the parameter that does not appear in the embedded model, in order to reconstitute the equivalence between both parameter spaces. Indeed, if we augment $\pi_0(\theta|y)$ with $\omega(\psi|\theta,y)$, we obtain a joint distribution with density $\pi_0(\theta|y)\times\omega(\psi|\theta,y)$ on $\Theta_1$. The Bayes factor can then be expressed as $$\label{eq:psudo}
B_{01} (y) =
\dfrac{{\displaystyle}\int_{\Theta_1} f(y|\theta,\psi_0) \pi_0(\theta)\alpha(\theta,\psi)\pi_1(\theta,\psi|y)
\text{d}\theta\omega(\psi|\theta,y)\,\text{d}\psi}
{{\displaystyle}\int_{\Theta_1} f(y|\theta,\psi) \pi_1(\theta,\psi) \alpha(\theta,\psi)\pi_0(\theta|y)
\times\omega(\psi|\theta,y) \text{d}\theta \,\text{d}\psi}\,,$$ for all functions $\alpha(\theta,\psi)$, because it is clearly independent from the choice of both $\alpha(\theta,\psi)$ and $\omega(\psi|\theta,y)$. Obviously, the performances of the approximation $$\dfrac{{\displaystyle}(n_1)^{-1} \sum_{j=1}^{n_1} f(y|\theta_{1,j},\psi_0)
\pi_0(\theta_{1,j}) \omega(\psi_{1,j}|\theta_{1,j},y)\alpha(\theta_{1,j},\psi_{1,j})}
{{\displaystyle}(n_0)^{-1} \sum_{j=1}^{n_0} f(y|\theta_{0,j},\psi_{0,j}) \pi_1(\theta_{0,j},
\psi_{0,j}) \alpha(\theta_{0,j},\psi_{0,j})}\,,$$ where $(\theta_{0,1},\psi_{0,1}),\ldots,(\theta_{0,n_0},\psi_{0,n_0})$ and $(\theta_{1,1},\psi_{1,1}),\ldots,(\theta_{1,n_1},\psi_{1,n_1})$ are two independent samples generated from distributions $\pi_0(\theta|y)\times\omega(\psi|\theta,y)$ and $\pi_1(\theta,\psi|y)$, respectively, do depend on this completion by the pseudo-posterior as well as on the function $\alpha(\theta,\psi)$. [@chen:shao:ibrahim:2000] establish that the asymptotically optimal choice for $\omega(\psi|\theta,y)$ is the obvious one, namely $$\omega(\psi|\theta,y) = \pi_1(\psi|\theta,y) \,,$$ which most often is unavailable in closed form (especially when considering that the normalising constant of $\omega(\psi|\theta,y)$ is required in ). However, in latent variable models, approximations of the conditional posteriors often are available, as detailed in Section \[sec:chibberies\].
While this extension of the basic bridge sampling approximation is paramount for handling embedded models, its implementation suffers from the dependence on this pseudo-posterior. In addition, this technical device brings the extended bridge methodology close to the cross-model alternatives of [@carlin:chib:1995] and [@green:1995], in that both those approaches rely on completing distributions, either locally [@green:1995] or globally [@carlin:chib:1995], to link both models under comparison in a bijective relation. The density $\omega(\psi|\theta_0,y)$ is then a pseudo-posterior distribution in Chib and Carlin’s (1995) sense, and it can be used as Green’s (1995) proposal in the reversible jump MCMC step to move (or not) from model $\mathfrak{M}_0$ to model $\mathfrak{M}_1$. While using cross-model solutions to compare only two models does seem superfluous, given that the randomness in picking the model at each step of the simulation is not as useful as in the setting of comparing a large number or an infinity of models, the average acceptance probability for moving from model $\mathfrak{M}_0$ to model $\mathfrak{M}_1$ is related to the Bayes factor since $$\mathbb{E}_{\pi_0\times\omega} \left[\frac{f(y|\theta,\psi)
\pi_1(\theta,\psi)}{f(y|\theta,\psi_0)\pi_0(\theta)\omega(\psi|\theta,y)} \right] = B_{01}(y)$$ even though the average $$\mathbb{E}_{\pi_0\times\omega} \left[ \min\left\{
1,\frac{f(y|\theta,\psi)\pi_1(\theta,\psi)}{f(y|\theta,\psi_0)\pi_0(\theta)\omega(\psi|\theta,y)}\right\}\right]$$ does not provide a closed form solution.
For the Pima Indian benchmark, we use as pseudo-posterior density $\omega(\theta_3|\theta_1,\theta_2,y)$, the conditional Gaussian density deduced from the asymptotic Gaussian distribution on $(\theta_1,\theta_2,\theta_3)$ already used in the importance sampling solution, with mean equal to the ML estimate of $(\theta_1,\theta_2,\theta_3)$ and with covariance matrix equal to the estimated covariance matrix of the ML estimate. The quasi-optimal solution $\alpha^\star$ in the bridge sampling estimate is replaced with the inverse of an average between the asymptotic Gaussian distribution in model $\mathfrak{M}_1$ and the product of the asymptotic Gaussian distribution in model $\mathfrak{M}_0$ times the above $\omega(\theta_3|\theta_1,\theta_2,y)$. This obviously is a suboptimal choice, but it offers the advantage of providing a non-iterative solution. The results, obtained over $100$ replications of the methodology with $n_0=n_1=20,000$ are summarized in Figure \[fig:bfbs\] and Table \[tab:res\]. The left-hand graph shows that this choice of bridge sampling estimator produces a solution whose variation is quite close to the (excellent) importance sampling solution, a considerable improvement upon the initial Monte Carlo estimator. However, the right-hand-side graph shows that the importance sampling solution remains far superior, especially when accounting for the computing time. (In this example, running 20,000 iterations of the Gibbs sampler for the models with both two and three variables takes approximately 32 seconds.)
![\[fig:bfbs\] Pima Indian dataset: [*(left)*]{} boxplots of 100 importance sampling, bridge sampling and Monte Carlo estimates of $B_{01}(y)$, based on simulations from the prior distributions, for $2\times 10^4$ simulations; [*(right)*]{} same comparison for the importance sampling versus bridge sampling estimates only.](bfisbsmc.pdf "fig:"){width="5cm" height="5cm"}![\[fig:bfbs\] Pima Indian dataset: [*(left)*]{} boxplots of 100 importance sampling, bridge sampling and Monte Carlo estimates of $B_{01}(y)$, based on simulations from the prior distributions, for $2\times 10^4$ simulations; [*(right)*]{} same comparison for the importance sampling versus bridge sampling estimates only.](bfisbs.pdf "fig:"){width="5cm" height="5cm"}
Harmonic mean approximations {#sec:harmo}
============================
While using the generic harmonic mean approximation to the marginal likelihood is often fraught with danger [@neal:1994], the representation [@gelfand:dey:1994] $(k=0,1)$ $$\label{eq:harmony}
\mathbb{E}_{\pi_k}\left[\left.\frac{\varphi_k(\theta) }{\pi_k(\theta)f_k(y|\theta)}\right| y \right]
= \int \frac{\varphi_k(\theta) }{\pi_k(\theta)f_k(y|\theta)} \,
\frac{\pi_k(\theta)f_k(y|\theta)}{{m}_k(y)}\,\text{d}\theta
= \frac{1}{{m}_k(y)}$$ holds, no matter what the density $\varphi_k(\theta)$ is—provided $\varphi_k(\theta)=0$ when $\pi_k(\theta)f_k(y|\theta)=0$—. This representation is remarkable in that it allows for a direct processing of Monte Carlo or MCMC output from the posterior distribution $\pi_k(\theta|y)$. As with importance sampling approximations, the variability of the corresponding estimator of $B_{01}(y)$ will be small if the distributions $\varphi_k(\theta)$ ($k=0,1$) are close to the corresponding posterior distributions. However, as opposed to usual importance sampling constraints, the density $\varphi_k(\theta)$ must have lighter—rather than fatter—tails than $\pi_k(\theta)f_k(y|\theta)$ for the approximation of the marginal $m_k(y)$ $$1\Bigg/ N^{-1}\,\sum_{j=1}^N \frac{\varphi_k(\theta_{k,j})}{\pi_k(\theta_{k,j}) f_k(y|\theta_{k,j})}$$ to enjoy finite variance. For instance, using $\varphi_k(\theta)=\pi_k(\theta)$ as in the original harmonic mean approximation [@newton:raftery:1994] will most usually result in an infinite variance estimator, as discussed by [@neal:1994]. On the opposite, using $\varphi_k$’s with constrained supports derived from a Monte Carlo sample, like the convex hull of the simulations corresponding to the $10\%$ or to the $25\%$ HPD regions—that again is easily derived from the simulations—is both completely appropriate and implementable [@robert:wraith:2009].
However, for the Pima Indian benchmark, we propose to use instead as our distributions $\varphi_k(\theta)$ the very same distributions as those used in the above importance sampling approximations, that is, Gaussian distributions with means equal to the ML estimates and covariance matrices equal to the estimated covariance matrices of the ML estimates. The results, obtained over $100$ replications of the methodology with $N=20,000$ simulations for each approximation of ${m}_k(y)$ ($k=0,1$) are summarized in Figure \[fig:bfhm\] and Table \[tab:res\]. They show a very clear proximity between both importance solutions in this special case and a corresponding domination of the bridge sampling estimator, even though the importance sampling estimate is much faster to compute. This remark must be toned down by considering that the computing time due to the Gibbs sampler should not necessarily be taken into account into the comparison, since samples are generated under both models.
![\[fig:bfhm\] Pima Indian dataset: [*(left)*]{} boxplots of 100 bridge sampling, harmonic mean and importance sampling estimates of $B_{01}(y)$, based on simulations from the prior distributions, for $2\times 10^4$ simulations; [*(right)*]{} same comparison for the harmonic mean versus importance sampling estimates only.](bfbshmis.pdf "fig:"){width="5cm" height="5cm"}![\[fig:bfhm\] Pima Indian dataset: [*(left)*]{} boxplots of 100 bridge sampling, harmonic mean and importance sampling estimates of $B_{01}(y)$, based on simulations from the prior distributions, for $2\times 10^4$ simulations; [*(right)*]{} same comparison for the harmonic mean versus importance sampling estimates only.](bfhmis.pdf "fig:"){width="5cm" height="5cm"}
Exploiting functional equalities {#sec:chibberies}
================================
Chib’s (1995) method for approximating a marginal (likelihood) is a direct application of Bayes’ theorem: given $y\sim f_k(y|\theta)$ and $\theta\sim\pi_k(\theta)$, we have that $${m}_k = \frac{f_k(y|\theta)\,\pi_k(\theta)}{\pi_k(\theta|y)}\,,$$ for all $\theta$’s (since both the lhs and the rhs of this equation are constant in $\theta$). Therefore, if an arbitrary value of $\theta$, say $\theta^*_k$, is selected and if a good approximation to $\pi_k(\theta|y)$ can be constructed, denoted $\hat{\pi}(\theta|y)$, Chib’s ([-@chib:1995]) approximation to the evidence is $$\label{eq:chib}
{m}_k = \frac{f_k(y|\theta^*_k)\,\pi_k(\theta^*_k)}{\hat{\pi_k}(\theta^*_k|y)}\,.$$ In a general setting, $\hat{\pi}(\theta|y)$ may be the Gaussian approximation based on the MLE, already used in the importance sampling, bridge sampling and harmonic mean solutions, but this is unlikely to be accurate in a general framework. A second solution is to use a nonparametric approximation based on a preliminary MCMC sample, even though the accuracy may also suffer in large dimensions. In the special setting of latent variables models (like mixtures of distributions but also like probit models), Chib’s (1995) approximation is particularly attractive as there exists a natural approximation to $\pi_k(\theta|y)$, based on the Rao–Blackwell [@gelfand:smith:1990] estimate $$\hat{\pi_k}(\theta^*_k|y) = \frac{1}{T}\,\sum_{t=1}^T \pi_k(\theta^*_k|y,z_k^{(t)})\,,$$ where the $z_k^{(t)}$’s are the latent variables simulated by the MCMC sampler. The estimate $\hat{\pi_k}(\theta^*_k|y)$ is a parametric unbiased approximation of $\pi_k(\theta^*_k|y)$ that converges with rate $\text{O}(\sqrt{T})$. This Rao–Blackwell approximation obviously requires the full conditional density $\pi_k(\theta^*_k|y,z)$ to be available in closed form (constant included) but, as already explained, this is the case for the probit model.
Figure \[fig:bfchi\] and Table \[tab:res\] summarize the results obtained for $100$ replications of Chib’s approximations of $B_{01}(y)$ with $T=20,000$ simulations for each approximation of ${m}_k(y)$ ($k=0,1$). While Chib’s method is usually very reliable and dominates importance sampling, the incredibly good approximation provided by the asymptotic Gaussian distribution implies that, in this particular case, Chib’s method is dominated by both the importance sampling and the harmonic mean estimates.
![\[fig:bfchi\] Pima Indian dataset: boxplots of 100 Chib’s, harmonic mean and importance estimates of $B_{01}(y)$, based on simulations from the prior distributions, for $2\times 10^4$ simulations.](bfchihmis.pdf){width="10cm" height="5cm"}
--------------------- -------- ------------ ---------- ---------- ---------------
Monte Importance Bridge Harmonic Chib’s
Carlo sampling sampling mean approximation
Median 3.277 3.108 3.087 3.107 3.104
Standard deviation 0.7987 0.0017 0.1357 0.0025 0.0195
Duration in seconds 7 7 71 70 64
--------------------- -------- ------------ ---------- ---------- ---------------
: \[tab:res\] Pima Indian dataset: Performances of the various approximation methods used in this survey.
Conclusion
==========
In this short evaluation of the most common estimations to the Bayes factor, we have found that a particular importance sampling and its symmetric harmonic mean counterpart are both very efficient in the case of the probit model. The bridge sampling estimate is much less efficient in this example, due to the approximation error resulting from the pseudo-posterior. In most settings, the bridge sampling is actually doing better than the equivalent importance sampler [@robert:wraith:2009], while Chib’s method is much more generic than the four alternatives. The recommendation resulting from the short experiment above is therefore to look for handy approximations to the posterior distribution, whenever available, but to fall back on Chib’s method as a backup solution providing a reference or better.
J.-M. Marin and C.P. Robert are supported by the 2009–2012 grant ANR-09-BLAN-0218 “Big’MC".
[25]{} natexlab\#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\]\[\][[\#2](#2)]{}
<span style="font-variant:small-caps;">Albert, J.</span> and <span style="font-variant:small-caps;">Chib, S.</span> (1993). ayesian analysis of binary and polychotomous response data. *J. American Statist. Assoc.*, **88** 669–679.
<span style="font-variant:small-caps;">Berger, J.</span> and <span style="font-variant:small-caps;">Pericchi, L.</span> (1996). The intrinsic [B]{}ayes factor for model selection and prediction. *J. American Statist. Assoc.*, **91** 109–122.
<span style="font-variant:small-caps;">Berger, J.</span>, <span style="font-variant:small-caps;">Pericchi, L.</span> and <span style="font-variant:small-caps;">Varshavsky, J.</span> (1998). ayes factors and marginal distributions in invariant situations. *Sankhya A*, **60** 307–321.
<span style="font-variant:small-caps;">Berger, J.</span> and <span style="font-variant:small-caps;">Sellke, T.</span> (1987). Testing a point-null hypothesis: the irreconcilability of significance levels and evidence (with discussion). *J. American Statist. Assoc.*, **82** 112–122.
<span style="font-variant:small-caps;">Carlin, B.</span> and <span style="font-variant:small-caps;">Chib, S.</span> (1995). ayesian model choice through [M]{}arkov chain [M]{}onte [C]{}arlo. *J. Royal Statist. Society Series B*, **57** 473–484.
<span style="font-variant:small-caps;">Chen, M.</span>, <span style="font-variant:small-caps;">Shao, Q.</span> and <span style="font-variant:small-caps;">Ibrahim, J.</span> (2000). *[M]{}onte [C]{}arlo Methods in [B]{}ayesian Computation*. Springer-Verlag, New York.
<span style="font-variant:small-caps;">Chib, S.</span> (1995). Marginal likelihood from the [G]{}ibbs output. *J. American Statist. Assoc.*, **90** 1313–1321.
<span style="font-variant:small-caps;">Chopin, N.</span> and <span style="font-variant:small-caps;">Robert, C.</span> (2007). Contemplating evidence: properties, extensions of, and alternatives to nested sampling. Tech. Rep. 2007-46, CEREMADE, Universit[é]{} Paris Dauphine. ArXiv:0801.3887.
<span style="font-variant:small-caps;">Friel, N.</span> and <span style="font-variant:small-caps;">Pettitt, A.</span> (2008). Marginal likelihood estimation via power posteriors. *J. Royal Statist. Society Series B*, **70(3)** 589–607.
<span style="font-variant:small-caps;">Gelfand, A.</span> and <span style="font-variant:small-caps;">Dey, D.</span> (1994). ayesian model choice: asymptotics and exact calculations. *J. Royal Statist. Society Series B*, **56** 501–514.
<span style="font-variant:small-caps;">Gelfand, A.</span> and <span style="font-variant:small-caps;">Smith, A.</span> (1990). Sampling based approaches to calculating marginal densities. *J. American Statist. Assoc.*, **85** 398–409.
<span style="font-variant:small-caps;">Gelman, A.</span> and <span style="font-variant:small-caps;">Meng, X.</span> (1998). Simulating normalizing constants: From importance sampling to bridge sampling to path sampling. *Statist. Science*, **13** 163–185.
<span style="font-variant:small-caps;">Green, P.</span> (1995). Reversible jump [MCMC]{} computation and [B]{}ayesian model determination. *Biometrika*, **82** 711–732.
<span style="font-variant:small-caps;">Jaynes, E.</span> (2003). *Probability Theory*. Cambridge University Press, Cambridge.
<span style="font-variant:small-caps;">Jeffreys, H.</span> (1939). *Theory of Probability*. 1st ed. The Clarendon Press, Oxford.
<span style="font-variant:small-caps;">Marin, J.-M.</span> and <span style="font-variant:small-caps;">Robert, C.</span> (2007). *Bayesian Core*. Springer-Verlag, New York.
<span style="font-variant:small-caps;">Neal, R.</span> (1994). Contribution to the discussion of “Approximate Bayesian inference with the weighted likelihood bootstrap" by Michael A. Newton and Adrian E. Raftery. *J. Royal Statist. Society Series B*, **56 (1)** 41–42.
<span style="font-variant:small-caps;">Newton, M.</span> and <span style="font-variant:small-caps;">Raftery, A.</span> (1994). Approximate [B]{}ayesian inference by the weighted likelihood bootstrap (with discussion). *J. Royal Statist. Society Series B*, **56** 1–48.
<span style="font-variant:small-caps;">[R Development Core Team]{}</span> (2008). *R: A Language and Environment for Statistical Computing*. R Foundation for Statistical Computing, Vienna, Austria. 3-900051-07-0, <http://www.R-project.org>.
<span style="font-variant:small-caps;">Robert, C.</span> (2001). *The [B]{}ayesian Choice*. 2nd ed. Springer-Verlag, New York.
<span style="font-variant:small-caps;">Robert, C.</span> and <span style="font-variant:small-caps;">Casella, G.</span> (2004). *[M]{}onte [C]{}arlo Statistical Methods*. 2nd ed. Springer-Verlag, New York.
<span style="font-variant:small-caps;">Robert, C.</span>, <span style="font-variant:small-caps;">Chopin, N.</span> and <span style="font-variant:small-caps;">Rousseau, J.</span> (2009). Theory of [P]{}robability revisited (with discussion). *Statist. Science*. (to appear).
<span style="font-variant:small-caps;">Robert, C.</span> and <span style="font-variant:small-caps;">Wraith, D.</span> (2009). Computational methods for [B]{}ayesian model choice. In *MaxEnt 2009 proceedings* (A. I. of Physics, ed.).
<span style="font-variant:small-caps;">Skilling, J.</span> (2006). Nested sampling for general [B]{}ayesian computation. *Bayesian Analysis*, **1(4)** 833–860.
<span style="font-variant:small-caps;">Torrie, G.</span> and <span style="font-variant:small-caps;">Valleau, J.</span> (1977). Nonphysical sampling distributions in [M]{}onte [C]{}arlo free-energy estimation: Umbrella sampling. *J. Comp. Phys.*, **23** 187–199.
[^1]: Institut de Mathématiques et Modélisation de Montpellier, Université Montpellier 2, Case Courrier 51, 34095 Montpellier cedex 5, France. Email: `[email protected]`
[^2]: CEREMADE - Université Paris Dauphine, 75775 Paris, and CREST, INSEE, Paris, France. Email: `[email protected]`
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We propose an efficient approach to generate the superposed macroscopically squeezed states with enhanced squeezing in a two-mode optomechanical system. This can be achieved by introducing a sinusoidal modulation to either the cavity frequencies or the coupling strengths between two cavity modes. The squeezement of the oscillator can be significantly enhanced to 12.16 dB with single photon, once the relative ratio of coupling strength is optimized under proper conditions. Further enhanced squeezing can be obtained by carefully adjusting the system parameters. In terms of the Wigner quasi-probability distribution, we show the squeezed error ellipses and interference fringes of the Yurke-Stoler-type squeezed states, denoting the squeezing and superposition properties. Our state generation scheme show reliable performance and robust resistance to finite environmental fluctuations, which implies applications for both fundamental interest and practical value.'
author:
- 'Sai-Nan Huai'
- Wei Nie
- 'Yun-bo Zhang'
- 'Yu-xi Liu'
title: 'Superposition of macroscopically squeezed states with enhanced squeezing in cavity optomechanical system at single-photon level'
---
INTRODUCTION {#Sec.III}
============
Quantum superposition [@Romero2011; @Asadian2014; @Arndt2014; @Pikovski2015], which is the intrinsic distinction between the classical and quantum worlds, has attracted considerable interest since its discovery. In the past few years, various superpositions of quantum states, e.g., superposition of number or coherent states [@Schrödinger1935; @Leibfried2005; @Ourjoumtsev2007; @Duan2019; @Johnson2017], have been explored in various systems, ranging from micro-scale, such as laser-trapped ions [@Monroe1996], ultracold atoms [@Berrada2013], superconducting systems [@Wal2000; @Liu2005; @Clarke2008; @Devoret2013], to meso-scale, e.g., microwave cavities interacting with Rydberg atoms [@Noel1996; @Brune1996; @Raimond1997]. However, it remains freshly studied to generate macroscopically superposition states in macroscopic systems with up to $10^{10}$ atoms, e.g., qubit-oscillator [@Liao2016PRA] and cavity optomechanical systems [@Liao2016PRL; @Hoff2016; @Aspelmeyer2014]. Decoherence to such huge number of atoms, induced by the quantum and thermal fluctuations, can often destroy and make it challenging to preserve the superposition. Meanwhile, squeezed state appeals a lot to researchers due to its high performance in the aspects of the continuous variable quantum information processing [@Caves1980; @Lorenz2001] and the quantum metrology [@Bocko1996; @Bollinger1996; @Kessler2014; @Barish1999]. Its intrinsic property of quadrature squeezement permits the reduction of quantum fluctuation in one quadrature to below the ground state (vacuum) level. Particularly, it plays an important role in ultrasensitive measurement, such as the detection of gravitational wave [@Aasi2013], photon scattering recoil event at the single photon level [@Hempel2013], and possible future astronomical observations such as supernova explosions. The aspiration for higher measurement precision inspires researchers to explore further enhancement of squeezing to beyond the quantum limit. Taking overall consideration, the state engineering of the quantum superposition of squeezed states [@Hach1993; @Fan1994; @Obada1997; @Barbosa2000; @El-Orany2011; @Balamurugan2017; @Sanders1998; @Schleich1988; @Xin1994] with both the macro-scale and enhanced squeezing degree, remains challenging but holds much promise for both the fundamental interest and practical value.
Many pioneering schemes have been proposed during the process of improving the degree of squeezing. As shown in Ref. [@LaHaye2004], position resolution with a factor of 4.3 above the quantum limit is achieved at millikelvin temperatures. Efficient squeezing can also be achieved via the backaction evading measurement (BAE) [@Clerk2008; @Hertzberg2010], or detuned mechanical parametric amplification (DMPA) with continuous weak measurement and feedback to the system [@Szorkovszky2011]. Besides, the generation of squeezing can be realized via the reservoir engineering [@Jahne2009; @Kronwald2013; @Kronwald2014; @Woolley2014; @Gu2013; @Liao2018; @Mari2009; @Farace2012; @Qu2014; @Lei2016; @Zhang2019], which does not require any feedback. For example, when the cavity is driven by the squeezed light, a squeezed heat bath can be provided, and transferred to the membrane via the radiation pressure efficiently [@Jahne2009]. An intermediate mode can also be engineered, via the modulated driving [@Mari2009; @Farace2012; @Qu2014] induced beam-splitter-like interactions, and then serve as a reservoir to cool the Bogoliubov mode of the target system operators to the near ground state [@Kronwald2013; @Kronwald2014; @Woolley2014; @Gu2013; @Liao2018; @Zhang2019]. However, these schemes are somewhat constrained by the requirements of ultra-low temperature or absolute strong coupling to the cavity mode, which are kind of difficult to be realized in current experiments. Additional side-effect is the possible parametric instability, which may induce the system entering the chaotic regime [@Monifi2016; @Lu2015; @Bakemeier2015]. Moreover, large squeezing can be achieved by increasing the effective coupling strength, in the way of achieving large number of average photons [@Shi2013; @Nunnenkamp2010], or adding atoms into the cavity [@Han2013], i.e., the atom-assisted squeezing. Nevertheless, the growing number of photons or atoms is accompanied with intensive dissipation effects, actually leading to the increase of threshold temperature of the mirror’s bath. Thus, in this paper, we focus on the single-photon induced squeezing to avoid from the intensive dissipation caused by large number of photons. Meanwhile, we resort to the modulated scheme [@Liao2014; @Liao2015; @Liao2016PRA; @Liao2016PRL; @Hoff2016; @Aspelmeyer2014; @Huang2017; @Huang2019; @Wu2018; @Quijandria2018], i.e., modulating the system parameters, such as frequencies or coupling strengths, to optimize relative ratios of coupling strengths, rather than simply increase their magnitudes.
Motivated by works in Refs. [@Liao2014; @Liao2015; @Liao2016PRA; @Liao2016PRL; @Hoff2016; @Aspelmeyer2014; @Huang2017; @Huang2019; @Wu2018; @Quijandria2018], here we revisit the standard quadratically coupled optomechanical system [@Murch2008; @Purdy2010; @Thompson2008] with the so-called “membrane-in-the-middle” (MIM) configuration [@Thompson2008; @Teng2012; @Bai2016], and propose an efficient approach to generate the superposed macroscopically squeezed states with enhanced squeezing. This goal can be realized by applying a sinusoidal modulation to the cavity frequency or the photon hopping rate and cavity optomechanical coupling strengths. Once the modulating frequencies and amplitudes satisfy specific conditions, the system can be approximately described by a quadratic-coupling optomechanical system with modified coupling strengths and mechanical resonator frequency. Then an effective strong coupling regime is achieved and the squeezement of the mechanical resonator is enhanced remarkably below the zero-point level to 12.16 dB. As far as we know, this degree of squeezing goes beyond to the best experimental realization 11.5 dB in Refs. [@Vinante2013; @Agarwal2016]. By carefully designing the initial state of the system and measuring the cavity at proper time, we can arrive at the superposed squeezed state of mechanical mode. We will show its enhanced squeezing and superposition properties in terms of Wigner function. We will also show the robust performance of this state generation scheme in open system.
The rest of our paper is organized as follows. In Sec. \[Sec.system\], we first give an explicit description for the quadratically coupled MIM optomechanical system. Then, the effective Hamiltonian in presence of a sinusoidal modulation to cavity frequencies are presented in Sec. \[Mfrequency\]. In Sec. \[enhance\], we show the mechanism of enhanced squeezing with single photon, and verify its validity through fidelity. In Sec. \[superpose\], we study the generation of superposed enhanced macroscopically squeezed states and its interference effect in terms of Winger quasi-probability distribution. In Sec. \[dissipation\], we consider the effect of environmental fluctuations to the performance of the state generation program. Finally, we summarize our research results in Sec. \[conclusion\]. Besides, the explicit processes with modulated photon hopping rate and optomechanical strengths are shown explicitly in Appendix \[Mcoupling\]. Meanwhile, the details in calculating the Wigner distributions are shown in Appendix \[wigner\].
THEORETICAL MODEL {#Sec.system}
=================
![(Color online) (a) Schematic diagram of a modulated hybrid membrane-in-the-middle (MIM) configuration of cavity optomechanical system. A thin membrane, with the phonon mode $b$ and located inside the cavity, separates the cavity into left and right modes, i.e., $a_L$, $a_R$. The cavity frequencies are modulated by the adjustion of cavity lengths $L_L(t)$ and $L_R(t)$. (b) Equivalent coupled-harmonic-resonator model. The parameters $\omega_L$, $\omega_R$, and $\omega_M$ represent the resonant frequencies of left, right cavities and mechanical resonator, respectively. The strength $J$ denotes the photon tunneling between the two cavities, while $g_0$ is the quadratic cavity optomechanical coupling.[]{data-label="Figure1"}](fig1.eps)
The standard quadratically coupled MIM optomechanical system we considered is schematically shown in Fig. \[Figure1\](a), where a thin dielectric membrane is placed inside a Fabry-Pérot cavity. We assume that the reflection coefficient of the membrane is large enough to separate the cavity into the right and left halves, and permits photon tunneling with a coefficient $J$ at the same time. The independent modes in these two halve cavities are labeled by $R$ and $L$. We also assume that the membrane is located in the node of the intra-cavity standing wave, such that the cavity field is coupled quadratically to the mechanical displacement [@Aspelmeyer2014]. The equivalent coupled-harmonic-resonator model is shown in Fig. \[Figure1\](b), i.e., the right, left cavities and the mechanical resonator mutually interact. The Hamiltonian of this hybrid MIM optomechanical system can be described as $$\begin{aligned}
H_0&=&\omega _{c} (a_{L}^{\dagger }a_{L}+a_{R}^{\dagger
}a_{R})+J\left( a_{L}^{\dagger }a_{R}+a_{R}^{\dagger }a_{L}\right) \nonumber \\
&&+\omega_{M}b^{\dagger }b+g_{0}\left( a_{L}^{\dagger }a_{L}-a_{R}^{\dagger
}a_{R}\right) \left( b^{\dagger }+b\right) ^{2}. \label{H0}\end{aligned}$$ Here we have assumed the same frequency $\omega_c$ for the left and right cavity modes $a_{L}$ and $a_{R}$, and $\omega_M$ is the resonance frequency for mechanical mode $b$. Specifically, the quadratic cavity-membrane coupling strength $g_0={\tau\zeta^2}[(1-T)/{T}]^{1/2}/2$ [@Teng2012; @Bai2016], with $T$ denoting the membrane transmissivity. The parameter $\tau=2L/c$ signifies the round trip time for each cavity, and $\zeta=\omega_c/L$, with $L$ the cavity length when the membrane and the outside mirror remain static. In fact, our model is general and system independent, it can be implemented with various experimental setups [@Murch2008; @Purdy2010; @Thompson2008; @Xue2007]. In addition to the MIM configuration [@Thompson2008], the cold-atom optomechanical system [@Murch2008; @Purdy2010] and superconducting quantum interference device (SQUID) [@Xue2007] are also potential platforms to demonstrate our scheme with carefully design.
In order to get the effective strong coupling and enhanced squeezing, we introduce a sinusoidal modulation to the cavity frequencies $\omega_c$ or the photon-hopping interaction constant $J$ and optomechanical coupling constant $g_0$. We note that this method has been widely applied to various tasks, such as generating distinct coherent states [@Liao2016PRA; @Liao2016PRL], manipulating counter-rotating interactions [@Huang2017] to achieve ultrastrong Jaynes-Cummings interaction [@Huang2019], realizing efficient and compact switch for quantum circuits [@Wu2018], and designing $\mathcal{PT}$-symmetric circuit QED [@Quijandria2018]. Here, we will mainly study the situation where a sinusoidal modulation is introduced to the cavity frequencies, while the case when the modulations are apllied to the photon-hopping rate $J$ and optomechanical coupling $g_0$ will be shown in Appendix \[Mcoupling\] explicitly.
HAMILTONIAN ENGINEERING {#Mfrequency}
=======================
Now we assume that the cavity frequencies of the left and right cavities are modulated as $$\begin{aligned}
\omega _{L} (t)&=&\omega _{c}+\Delta _{0}\cos 2Jt, \\
\omega _{R} (t)&=&\omega _{c}-\Delta _{0}\cos 2Jt,\end{aligned}$$respectively, and the modulated Hamiltonian is given by $$H_M(t)=\Delta _{0}\cos 2Jt\left(a_{L}^{\dagger }a_{L}-a_{R}^{\dagger }a_{R}\right).$$ Here, we assume that the modulating amplitude $\Delta_0$ is far less than the modulation frequency $J$, i.e., $\Delta_0\ll J$. Considering that cavity frequency is proportional to $L^{-1}$ [@Aspelmeyer2014], the modulated cavity frequencies can be realized by modulating the lengths of the left and right cavities as $L\left[1-\left(\Delta _{0}\cos 2Jt\right)/2\omega_c\right]^2$ and $L\left[1+\left(\Delta _{0}\cos 2Jt\right)/2\omega_c\right]^2$, respectively. In terms of experimental practicability, the cavity lengths can be modulated with the help of two lasers with beat frequency $2J$ [@Schmidt2015], or the piezoelectric transducer, which converts the electrical signal to mechanical vibration [@Svelto1982]. Besides, the rapid response of cavity frequencies makes it reasonable to neglect the back-action of the length modulation to the membrane. Thus, the full Hamiltonian of this hybrid MIM system under the frequency modulation can be engineered as $$\begin{aligned}
H(t)&=&H_0+H_M(t)\nonumber\\
&=&\omega_{L}(t) a_{L}^{\dagger }a_{L}+\omega _{R}(t) a_{R}^{\dagger
}a_{R}+J\left( a_{L}^{\dagger }a_{R}+a_{R}^{\dagger }a_{L}\right) \nonumber \\
&&+\omega_{M}b^{\dagger }b+g_{0}\left( a_{L}^{\dagger }a_{L}-a_{R}^{\dagger
}a_{R}\right) \left( b^{\dagger }+b\right) ^{2}. \label{H}\end{aligned}$$
The state-evolution of this system can be described by the Schrödinger equation $$i\frac{\partial }{\partial t}\left\vert \psi \left( t\right) \right\rangle
=H\left( t\right) \left\vert \psi \left( t\right) \right\rangle. \label{psi}$$ Usually, this equation can be solved by introducing an unitary transformation $Q(t)$, i.e., $\left\vert \psi \left( t\right) \right\rangle=Q(t)\left\vert \tilde{\psi} \left( t\right) \right\rangle.$ As a consequence, Eq. (\[psi\]) can be rewritten as $$i\frac{\partial }{\partial t}\left\vert \tilde{\psi} \left( t\right) \right\rangle
=\tilde{H}\left( t\right) \left\vert \tilde{\psi} \left( t\right) \right\rangle,$$ with $\tilde{H}(t)$ determined by [@Scully1997] $$\tilde{H}(t)=Q^{\dagger}(t)H(t)Q(t)-iQ^{\dagger}(t)\frac{\partial Q(t)}{\partial t}.\label{Htilde}$$ In our case, the unitary transformation $Q(t)$ is defined as $Q(t)=P_1(t)P_2(t)P_3(t)$ [@Silveri2017], with $$\begin{aligned}
P_{1}\left( t\right) &=&e^{-i\omega _{c}\left( a_{L}^{\dagger}a_{L}+a_{R}^{\dagger }a_{R}\right) t}, \label{P1}\\
P_{2}\left( t\right) &=&e^{-iJ\left( a_{L}^{\dagger }a_{R}+a_{R}^{\dagger}a_{L}\right)t}, \label{P2}\\
P_{3}\left( t\right) &=&e^{-i\left[ \frac{\Delta _{0}}{2}\left(a_{L}^{\dagger }a_{L}-a_{R}^{\dagger }a_{R}\right) +\omega _{M}b^{\dagger }b%
\right] t}.\label{P3}\end{aligned}$$ With the transformation $Q(t)$, the Hamiltonian shown in Eq. (\[Htilde\]) can be written explicitly as $$\begin{aligned}
\tilde{H}(t)&&=\left( a_{L}^{\dagger
}a_{L}-a_{R}^{\dagger }a_{R}\right)\left[\frac{\Delta _{0}}{4}F_1(b)+gF_2(b)+\rm{H.c.}\right]\nonumber\\
&&+\left(a_{L}^{\dagger }a_{R}e^{i\Delta_0t}-\rm{H.c.}\right)\left[-\frac{\Delta _{0}}{4}F_1(b)+gF_3(b)-\rm{H.c.}\right],\nonumber\\ \label{HTILDE}\end{aligned}$$ with the functions $$\begin{aligned}
F_1(b)&=&e^{4iJt},\nonumber\\
%F_2(b)&=&-2i{\rm{sin}}(4Jt)\nonumber\\
F_2(b)&=&b^{2}\left(e^{-2i\delta_{-} t}+e^{-2i\delta_{+} t}\right)+\left( 2b^{\dagger }b+1\right) e^{-2iJt},\nonumber\\
F_3(b)&=&b^{2}\left(e^{-2i\delta_{+} t}-e^{-2i\delta_{-} t}\right)+\left( 2b^{\dagger }b+1\right)e^{-2iJt}.\nonumber\end{aligned}$$
Besides, we have defined the modified coupling constant $g$ and detuning $\delta_{\pm}$ as, $$g=\frac{g_0}{2},\ \ \ \delta_{\pm}=\omega_M\pm J,$$ respectively. It is obvious that more interaction components show up due to cavity frequencies modulation, e.g., terms whose frequencies or amplitudes are related to $\Delta_0$ and $J$. However, we can always find a way to manipulate the desired terms into near-resonant ones, such that the resulted effective strong coupling terms actually dominate the evolution of the system, while other fast-osillating and relatively weak coupling ones can be neglected. In our work here, we would like to study the generation of the squeezing states for mechanical resonator, thus we consider the following conditions $$\begin{aligned}
J\gg\frac{5}{16}\Delta_0, \ \ \ \Delta_0&\gg& 2g, \left\vert\delta_{-}\right\vert \nonumber\\
\left\vert\Delta_0-2J\right\vert&\gg& g. \label{condition}\end{aligned}$$ In this situation, the terms other than those with frequency $\delta_{-}$ in Eq. (\[HTILDE\]) turn out to be fast-oscillating and weak coupling, while the coupling coefficients in the desired near-resonant terms are comparable to the oscillating frequencies, i.e., the effective strong coupling is realized. Thus, it is reasonable to conduct the rotating-wave approximation (RWA) to Eq. (\[HTILDE\]), leading to $$\tilde{H}_{\rm{eff}}(t)\simeq g\left( a_{L}^{\dagger
}a_{L}-a_{R}^{\dagger }a_{R}\right) \left( b^{2}e^{-2i\delta t}+b^{\dagger
2}e^{2i\delta t}\right). \label{Heff'}$$ For convenience, we have neglected the subscript and rewritten $\delta_{-}$ as $\delta$ to represent the slow oscillating frequency, i.e., $\delta=\delta_{-}=\omega_M-J$. We can further arrive at the effective Hamiltonian $$\begin{aligned}
H_{\rm{eff}}\left( t\right)&=&\omega _{c}\left( a_{L}^{\dagger }a_{L}+a_{R}^{\dagger }a_{R}\right)
+\delta b^{\dagger }b \nonumber\\
&&+g\left( a_{L}^{\dagger }a_{L}-a_{R}^{\dagger
}a_{R}\right) \left( b^{2}+b^{\dagger 2}\right). \label{Heff}\end{aligned}$$ by performing the unitary transformation $P_4(t)$, which is defined as $$P_{4}\left( t\right) =e^{i\left[ \omega _{c}\left( a_{L}^{\dagger
}a_{L}+a_{R}^{\dagger }a_{R}\right) +\delta b^{\dagger }b\right] t}. \label{P4}$$
As a result, under the conditions shown in Eq. (\[condition\]), we engineer an effective strong quadratically coupled Hamiltonian, i.e., Eq. (\[Heff\]), for the modulated cavity frequency case. Additionally, the conditions and engineered Hamiltonian correspond to the modulated coupling strengths case are shown in Eqs. (\[condition2\]) and (\[Heff2\]), respectively. The flexibility of modulation permits the detuning $\delta=\omega_M-J$ to be small enough than the modified coupling strength $g$, which further leads to effective strong and even ultra-strong quadratically coupling. This Hamiltonian engineering mechanism actually lays the foundation of generating the superposed macroscopically squeezed states with enhanced squeezing at single-photon level, which we would like to give a detailed description in the rest part of our work.
ENHANCED SQUEEZING WITH SINGLE-PHOTON {#enhance}
=====================================
Let us first focus on how to enhance the degree of squeezing for the states of the mechanical resonator utilizing the Hamiltonian in Eq. (\[Heff\]). We hope that the enhanced squeezing can go far beyond the quantum limit than most of reported schemes. With carefully designed initial state and dynamical evolution for a proper time, we can obtain the enhanced squeezing state for the mechanical mode.
For a closed system without dissipation, we can get analytical results, according to the effective Hamiltonian $H_{\rm{eff}}\left( t\right)$ in Eq. (\[Heff\]) and the corresponding time-dependent Schrödinger equation $i{\partial }\left\vert \psi_{\rm{eff}}\left( t\right) \right\rangle/{\partial t}
=H_{\rm{eff}}\left( t\right) \left\vert \psi_{\rm{eff}}\left( t\right) \right\rangle$. Here the instantaneous state $\left\vert\psi_{\rm{eff}}\left( t\right)\right\rangle$ is related to the analytical state $\left\vert\psi\left( t\right)\right\rangle$ in Eq. (\[psi\]) with several unitary transformations, which have been shown in Eqs. (\[P1\])-(\[P3\]) and (\[P4\]), i.e., $$\left\vert\psi\left( t\right)\right\rangle\simeq P_1\left( t\right)P_2\left( t\right)P_3\left( t\right)P_4\left( t\right) \left\vert\psi_{\rm{eff}}\left( t\right)\right\rangle. \label{psit}$$
Using the standard disentangling techniques utilized in quantum optics [@Mufti1993], we can evaluate the time evolution operator, corresponding to the effective Hamiltonian $H_{\rm{eff}}$, as $$e^{-iH_{\rm{eff}}t}=e^{-i\omega _{c}\left( a_{L}^{\dagger }a_{L}+a_{R}^{\dagger
}a_{R}\right) t}e^{\frac{i}{2}\delta t}e^{\frac{1}{2}\left( \xi ^{\ast
}b^{2}-\xi b^{\dagger 2}\right) }e^{i\eta \left( b^{\dagger }b+\frac{1}{2}%
\right)}. \label{Hefft}$$ The time-dependent parameters $\xi\equiv\xi \left( t\right)$ and phase factor $\eta\equiv\eta \left( t\right)$ are defined as $$\begin{aligned}
\xi \left( t\right)&=&e^{i\left( \eta +\frac{\pi }{2}\right) }\sinh
^{-1}\left( \frac{r\sin \chi t}{\chi }\right), \label{xi}\\
\eta \left( t\right)&=&\tan ^{-1}\left( \frac{p\tan \chi t}{2\chi }\right),\label{eta}\end{aligned}$$ respectively. It is obvious that $\xi \left( t\right)$ represent the degree of the squeezing as shown in quantum optics [@Mufti1993]. We also note that $\eta\left(t\right)$ is a real number, and the other parameters can be evaluated as $$\begin{aligned}
p &=&-2\delta, \\
r({N}_{LR})&=&2g{N}_{LR}, \label{rNLR}\\
\chi({N}_{LR}) &=&\sqrt{\delta ^{2}-4g^{2}{N}_{LR} ^{2}}.\label{chiNLR}\end{aligned}$$ Here, ${N}_{LR}=a_{L}^{\dagger }a_{L}-a_{R}^{\dagger }a_{R}$ denotes the photon number inversion between the left and right cavities. The Eqs. (\[xi\])-(\[chiNLR\]) indicate that the degree of squeezing is related to photon number. For the convenience of calculation and in oder to reduce the intensive dissipative effects caused by large number of photons for the open system which we will study in Sec. \[dissipation\], we will mainly consider the squeezing of the mechanical resonator with single photon, i.e., $n_L+n_R=1$. Here, $n_L~(n_R)=0,1,2\cdots$ denotes the photon excitation number corresponding to the photon number operator $a_{L}^{\dagger }a_{L}~(a_{R}^{\dagger }a_{R})$ in the left (right) cavity. According to Eqs. (\[psit\]) and (\[Hefft\]), the dynamical change from the initial state $\left\vert\psi\left( 0\right)\right\rangle$ to the final state $\left\vert\psi\left( t\right)\right\rangle$, i.e., $\left\vert\psi\left( 0\right)\right\rangle\to\left\vert\psi\left( t\right)\right\rangle$, can be described as $$\left\vert \psi \left( t\right) \right\rangle\simeq U(t)\left\vert \psi \left( 0\right)\right\rangle. \label{psitTheory}$$ Here, $$U(t)=P_1(t)P_2(t)P_3(t)P_4(t)e^{-iH_{\rm{eff}}t} \label{Ut}$$ denotes the approximate time evolution operator associated with the original total Hamiltonian in Eq. (\[H\]). Meanwhile, the relation $\left\vert\psi\left( 0\right)\right\rangle\equiv\left\vert\psi_{\rm{eff}}\left( 0\right)\right\rangle$ has been used. In the process of getting the approximate analytical state $\left\vert \psi \left( t\right) \right\rangle$ in Eq. (\[psitTheory\]), rotating-wave approximation has been made under the condition Eq. (\[condition\]). Thus we need to verify its validity by checking the fidelity between the approximate analytical state $\left\vert \psi \left( t\right) \right\rangle$, and the numerical state $\left\vert \Psi \left( t\right) \right\rangle$ associated with the original total Hamiltonian in Eq. (\[H\]). The numerical results can be obtained by solving the Schrödinger equation $i{\partial }\left\vert \Psi \left( t\right) \right\rangle/{\partial t}
=H\left( t\right) \left\vert \Psi \left( t\right) \right\rangle $ based on the general single-photon state of the system $$\left\vert \Psi \left( t\right) \right\rangle =\sum\limits_{n=0}^{\infty }
\left[ A_{m}\left( t\right) \left\vert 1\right\rangle _{L}\left\vert
0\right\rangle _{R}+B_{m}\left( t\right) \left\vert 0\right\rangle
_{L}\left\vert 1\right\rangle _{R}\right] \left\vert m\right\rangle _{M}. \label{Psit}$$ Here $A_m(t)$ and $B_m(t)$ denote the probability amplitudes of the states $\left\vert 1\right\rangle _{L}\left\vert0\right\rangle _{R}\left\vert m\right\rangle _{M}$ and $\left\vert 0\right\rangle _{L}\left\vert1\right\rangle _{R}\left\vert m\right\rangle _{M}$, respectively. The subscripts $L$, $R$, and $M$ represent the left, right cavities, and the mechanical resonator, respectively. The equations of motion can be expressed as $$\begin{aligned}
\dot{A}_{m}\left( t\right) &=&-i\left( \omega _{c}+\Delta _{0}\cos
2Jt+m\omega _{M}\right) A_{m}\left( t\right) -iJB_{m}\left( t\right) \nonumber\\
&&-ig_{0} [ \left( 2m+1\right) A_{m}\left( t\right) +\sqrt{m\left(
m-1\right) }A_{m-2}\left( t\right) \nonumber\\
&&+\sqrt{\left( m+1\right) \left(
m+2\right) }A_{m+2}\left( t\right) ],\label{Amt} \\
\dot{B}_{m}\left( t\right) &=&-i\left( \omega _{c}-\Delta _{0}\cos
2Jt+m\omega _{M}\right) B_{m}\left( t\right) -iJA_{m}\left( t\right) \nonumber\\
&&+ig_{0} [ \left( 2m+1\right) B_{m}\left( t\right) +\sqrt{m\left(
m-1\right) }B_{m-2}\left( t\right) \nonumber\\
&&+\sqrt{\left( m+1\right) \left(
m+2\right) }B_{m+2}\left( t\right) ]. \label{Bmt}\end{aligned}$$As long as the coefficients $A_m(0)$ and $B_m(0)$ $(m\geq 0)$ for the initial state $\left\vert \Psi \left( 0\right) \right\rangle$ are given, we can obtain the state $\left\vert \Psi \left( t\right) \right\rangle $ at any given time by solving Eqs. (\[Amt\]) and (\[Bmt\]) numerically.
![(Color online) (a) The degree of mechanical squeezing $R(t)$ under different rescaled detunings, e.g., $\delta/g_0=0.25,\,1,\,1.13,\,1.5$. Here, the rescaled modulating amplitude $\Delta_0/g_0=100$. (b) The fidelity $F(t)$ for the given detuning $\delta/g_0=1.13$, under different rescaled modulating amplitude $\Delta_0/g_0=20,\,60,\,100$. The rescaled photon tunneling rate $J/g_0=398.6$.[]{data-label="Figure2"}](fig2.eps)
We now consider the mechanical squeezing with single-photon by assuming the initial state $\left\vert \psi \left( 0\right) \right\rangle=\left\vert \Psi \left( 0\right) \right\rangle=\left\vert 1\right\rangle_L\left\vert 0\right\rangle_R\left\vert 0\right\rangle_M$, i.e., there is only one photon in the left cavity while vacuum in the right one, and the mechanical membrane is in the ground state $\left\vert 0\right\rangle_M$. According to Eq. (\[psitTheory\]), the wave function $\left\vert \psi(t)\right\rangle$ at time $t$ can be analtically written as $$\begin{aligned}
\left\vert \psi \left( t\right) \right\rangle&=&e^{-i\omega _{c}t}e^{-i\theta \left( t\right) }[ \cos \left(
Jt\right) \left\vert 1\right\rangle _{L}\left\vert 0\right\rangle _{R} \nonumber \\
&&-i\sin\left( Jt\right) \left\vert 0\right\rangle _{L}\left\vert 1\right\rangle _{R}%
] \left\vert Z\left( t\right) \right\rangle _{M},\end{aligned}$$ with the global phase factor $$\theta \left( t\right)=-\left[ \delta t+\eta \left( t\right)-\Delta _{0}t\right]/2. \label{thetat}$$ Here $ \left\vert Z\left( t\right) \right\rangle _{M}$ is the squeezed vacuum state of the membrane with $$Z\left( t\right)=\xi \left( t\right) e^{-2i\left( \omega _{M}-\delta
\right) t},$$ and the parameter $\xi \left( t\right)$ have been given in Eq. (\[xi\]). Meanwhile, the corresponding parameters in Eqs. (\[rNLR\]) and (\[chiNLR\]) can be specifically written as $$\begin{aligned}
%\theta \left( t\right) &=&-\frac{1}{2}\left( \delta t+\eta \left( t\right)-\Delta _{0}t\right), \\
r=2g,\text{ \ \ \ }\chi =\sqrt{\delta ^{2}-4g^{2}},\end{aligned}$$ based on the single photon assumption shown in Eq. (\[Psit\]). An expression of the squeezed vacuum state $\left\vert Z\left( t\right)\right\rangle_M$ in terms of the number state $\left\vert m\right\rangle$ is given by [@Zayed2005; @Gerry2004]$$\left\vert Z\left( t\right)\right\rangle_M=\sum\limits_{m=0}^{\infty }S_{m}\left( Z\left( t\right)\right) \left\vert2m\right\rangle,$$where $$S_{m}\left( Z\left( t\right)\right)=\frac{\left( -1\right) ^{m}}{\sqrt{\cosh R\left( t\right)}}\frac{\sqrt{\left( 2m\right) !}}{2^{m}m!}e^{im\Phi\left( t\right) }\tanh ^{m}R\left( t\right).$$
![(Color online) The fidelity $F(T_M)$ at the time $T_M$, where the maximum squeezing happens, versus the rescaled modulating amplitude $\Delta_0/g_0$, under different rescaled photon tunneling rates $J/g_0=118.6,\, 258.6, \,398.6$. Here the rescaled detuning $\delta/g_0$ is settled down to 1.13. The mechanical frequency is determined by $\omega_M=\delta+J$ for different tunnel rates $J$.[]{data-label="Figure3"}](fig3.eps)
Here the modulus ${R}\left( t\right)$ and argument $\Phi\left( t\right)$ of $Z\left( t\right)$ are defined as $$\begin{aligned}
{R}\left( t\right)&=&\left\vert \sinh^{-1}\left( \frac{r\sin \chi t}{\chi }\right)\right\vert,\\
\Phi \left( t\right)&=&\arg \left[ Z\left( t\right)\right],\end{aligned}$$ respectively. It is obvious that if $\delta> 2g$, then $\chi>0$, and ${R}\left( t\right)$ shows up as a periodical function, with the period $T=\pi/\chi$. The first maximum squeezing ${R}\left( T_M\right)=\sinh^{-1}\left( {r}/{\chi }\right)$ is achieved at $T_M=T/2$, as pointed out by the vertical red-dotted line in Fig. \[Figure2\](a). Besides, we also show how the degree of squeezing ${R}\left( t\right)$ varies versus time, under different detunings $\delta$ in Fig. \[Figure2\](a). It is shown that with the decreasing of the $\delta$, e.g., from $\delta/g_0=1.5$ (green dotted line) to 1.13 (red solid line), the period $T$ gets longer and ${R}\left( T_M\right)$ becomes larger. At the singular point $\delta=2g$, ${R}\left( t\right)=\sinh^{-1}\left(rt\right)$ grow exactly in the form of hyperbolic sine function (black dash-dotted line). As for the smaller detuning case $\delta< 2g$, for example, $\delta/g_0=0.25$ (blue dashed line), the mechanical squeezing ${R}\left( t\right)= \sinh^{-1}\left( {r\sinh \left({\sqrt{4g^{2}-\delta ^{2} } t}\right)}/{\sqrt{4g^{2}-\delta ^{2} }}\right)$ grows monofonically with time until the Hamitonian breaks down. We find that, although the larger detuning case ($\delta>2g$), which corresponds to periodic squeezing, can not get the given squeezing in shortest time, it shows higher stability to the perturbation around $T_M$. Thus, in this work, without loss of generality, we mainly focus on the larger detuning case with periodic squeezing, e.g., $\delta/g_0=1.13$ (red solid line). In this case, the maximum degree of mechanical squeezing can be enhanced to $-10{\rm{log}}_{10}[{\rm{exp}}(-2*R(T_M))]\simeq12.16$ dB [@Scully1997]. Next, we will show its reliability by checking the fidelity between the analytical and numerical states.
According to Eq. (\[Psit\]), the initial state $\left\vert \Psi \left( 0\right) \right\rangle$ can be expressed as $$A_0(0)=1,\text{\ \ \ } A_{m>0}(0)=B_{m\geq0}=0.$$ Then by solving the equations of motion in Eqs. (\[Amt\]) and (\[Bmt\]) numerically, we can get the state at any given time $\left\vert \Psi \left( t\right) \right\rangle$. In terms of the analytical expression $\left\vert \psi \left( t\right) \right\rangle$ and numerical result $\left\vert \Psi \left( t\right) \right\rangle$, the fidelity $F\left(t\right)$, which is defined by $\left\vert \left\langle \Psi \left( t\right)
|\psi \left( t\right) \right\rangle \right\vert ^{2}$, can be written as $$F\left( t\right) =\left\vert \sum\limits_{m=0}^{\infty }\left[ \cos \left( Jt\right)
A_{m}^{\ast }\left( t\right) -i\sin \left( Jt\right) B_{m}^{\ast }\left(
t\right) \right] \left\langle m\vert Z\left( t\right)
\right\rangle _{M}\right\vert ^{2}. \label{Ft}$$
In Fig. \[Figure2\](b), we plot the fidelity $F\left( t\right)$ as a function of time $t$ for different rescaled modulating amplitudes $\Delta_0/g_0$, under given detuning $\delta/g_0=1.13$. Note that the squeezing of this given detuning corresponds to the red-solid curve in Fig. \[Figure2\](a). We also would like to mention that the photon tunneling rate $J/g_0=398.6$ is chosen in order to get the strongest interference effect, which will be illustrated in Sec. \[superpose\]. Here, we focus on the evolution of $F\left( t\right)$ within $T_M=T/2=\pi/2\chi$, considering that it is long enough for us to get the macroscopically squeezed state with enhanced degree of squeezing. It can be seen that higher fidelity is arrived with the increase of the rescaled modulating amplitude $\Delta_0/g_0$, e.g., from 20 (blue line) ranges to 60 (red line), and then 100 (black line). However, it does not mean that the modulating amplitude $\Delta_0$ is allowed to increase unlimitedly, which actually has been implied in Eq. (\[condition\]). Figure \[Figure3\] helps us get more intuitional understanding about its restriction by $J$. Here we focus on the fidelity $F\left( T_M\right)$ at time $T_M$, where we get the maximum squeezing, as denoted by the red-dashed line in Fig. \[Figure2\](a). In Fig. \[Figure3\], we show the response of $F\left( T_M\right)$ to the variation of rescaled modulating amplitude $\Delta_0/g_0$, under different modulating frequencies $J/g_0$. Meanwhile, we fix the rescaled detuning $\delta/g_0=1.13$, and adjust the mechanical frequency following the relation $\omega_M=\delta+J$. It shows that high and stable fidelity $F\left( T_M\right)$ can only be obtained within a certain range of $\Delta_0/g_0$. Moreover, the range shrinks with the reduction of the rescaled tunneling rate $J/g_0$, e.g., from 398.6 (black solid line) to 258.6 (red dash-dotted line), and then 118.6 (blue dotted line). Once the rescaled modulating amplitude $\Delta_0/g_0$ goes out of the range, the fidelity $F\left( T_M\right)$ decreases in an oscillating way to zero, which means the validity of RWA needs to be reconsidered. Obviously, the characteristics shown in Fig. \[Figure2\](b) and Fig. \[Figure3\] are in good accordance with the condition in Eq. (\[condition\]).
![(Color online) (a) The fidelity $F^{\prime}(t)$ for the given detuning $\delta/g_0=1.13$, under different rescaled modulating amplitude $\Delta_0/g_0=20,\,60,\,100$. The rescaled mechanical frequency $\omega_M/g_0=400$. (b) The fidelity $F^{\prime}(T_M)$ at the time $T_M$, where the maximum squeezing happens, versus the rescaled modulating amplitude $\Delta_0/g_0$, under different rescaled photon tunneling rates $J/g_0=118.6,\, 258.6, \,398.6$. Here the rescaled detuning $\delta/g_0$ is settled down to 1.13. The mechanical frequency is determined by $\omega_M=\delta+J$ for different tunnel rates $J$.[]{data-label="Figure4"}](fig4.eps)
The high fidelity under the condition in Eq. (\[condition\]), makes it reliable to generate the mechanical squeezed state with enhanced squeezing degree in our approach. Provided that the system was initialized as $\left\vert \psi \left( 0\right) \right\rangle=\left\vert 1\right\rangle_L\left\vert 0\right\rangle_R\left\vert 0\right\rangle_M$ and evolved for time $T_M$, the mechanical resonator can be projected into an enhanced squeezed vacuum state $\left\vert Z(T_M)\right\rangle$. Meanwhile, the degree of squeezing can be enhanced to 12.16 dB and even further. Next, we would like to show the superposition of the squeezed state with enhanced squeezing degree and its interference property.
Generation of superposed enhanced squeezed states {#superpose}
=================================================
In this section, we focus on the superposition of the macroscopically squeezed state with enhanced squeezing degree, which has been discussed in Sec. \[enhance\]. Similarly, we now assume that the initial state of the whole system has the form of $$\left\vert \psi^{\prime} \left( 0\right) \right\rangle=\frac{1}{\sqrt2}( \left\vert 1\right\rangle_L\left\vert 0\right\rangle_R+\left\vert 0\right\rangle_L\left\vert 1\right\rangle_R) \left\vert 0\right\rangle_M,$$ which means that the photons in the left and right cavities are entangled with each other, and the mechanical resonator is initialized in the ground state. Utilizing the unitary evolution operator $U(t)$ in Eq. (\[Ut\]) and following the evolution equation in Eq. (\[psitTheory\]), the state $\left\vert \psi^{\prime} \left( t\right) \right\rangle$ of the system at time $t$ can be obtained analytically as $$\begin{aligned}
\left\vert \psi^{\prime} \left( t\right) \right\rangle&=&\frac{e^{-i\left[ \theta
\left( t\right) +\omega _{c}t\right] }}{\sqrt{2}}[ \left\vert
1\right\rangle _{L}\left\vert 0\right\rangle _{R}\left\vert \varphi
_{L}\left( t\right) \right\rangle _{M}\nonumber\\
&&+\left\vert 0\right\rangle
_{L}\left\vert 1\right\rangle _{R}\left\vert \varphi _{R}\left( t\right)
\right\rangle _{M}].\label{psipt}\end{aligned}$$ Here, the global phase factor $\theta(t)$ has been defined in Eq. (\[thetat\]), and $$\begin{aligned}
\left\vert \varphi _{L}\left( t\right) \right\rangle _{M} &=&\cos \left(
Jt\right) \left\vert Z\left( t\right) \right\rangle _{M}-ie^{i\Delta_{0}t}\sin \left(
Jt\right) \left\vert -Z\left( t\right) \right\rangle _{M},\nonumber \\
\left\vert \varphi _{R}\left( t\right) \right\rangle _{M} &=&e^{i\Delta_{0}t}\cos \left(
Jt\right) \left\vert -Z\left( t\right) \right\rangle _{M}-i\sin \left(
Jt\right) \left\vert Z\left( t\right) \right\rangle _{M}, \label{phiLR}\nonumber \\\end{aligned}$$ represent two Yurke-Stoler-like [@Yurke1986; @Yurke1997; @Yurke1990] quantum superpositions of the enhanced squeezing states with different relative phases, respectively. Following the same procedure in the last subsection, i.e., by measuring the photon in the left or right cavities at time $T_M$, we can get the mechanical resonator in the superposed squeezed states $\left\vert \varphi _{L}\left(T_M\right) \right\rangle _{M}$ or $\left\vert \varphi _{R}\left(T_M\right) \right\rangle _{M}$.
Meanwhile, the numerical solution $\left\vert \Psi^{\prime} \left( t\right) \right\rangle$ can also be obtained from the corresponding initial condition $\left\vert \Psi^{\prime} \left( 0\right) \right\rangle=\left\vert \psi^{\prime} \left( 0\right) \right\rangle$, i.e. $$A_0(0)=B_0(0)=\frac{1}{\sqrt{2}}, \text{\ \ \ }A_{m>0}(0)=B_{m>0}=0,$$ and the evolution equations shown in Eqs. (\[Amt\]) and (\[Bmt\]). In Figs. \[Figure4\](a) and \[Figure4\](b), the reliability of $\left\vert \psi^{\prime} \left( t\right) \right\rangle$ in Eq. (\[psipt\]) is verified by checking the fidelity dynamically and instantaneously, i.e., $F^{\prime}(t)$ and $F^{\prime}(T_M)$ under different conditions, respectively. Corresponding to the conditions in Eq. (\[condition\]), they behave similarly with $F(t)$ and $F(T_M)$ in Fig. \[Figure2\](b) and Fig. \[Figure3\], respectively, except for different absolute amplitudes and ranges. For example, under given photon tunneling rate $J$, the reliable range of modulating amplitudes $\Delta_0$ in Fig. \[Figure4\](b) is smaller than that in Fig. \[Figure3\], this may be induced by the interference between different channels shown in Eq. (\[psipt\]) and Eq. (\[phiLR\]).
![(Color online) Wigner functions (a) $W_L(\alpha)$ and (b) $W_R(\alpha)$ for the superposed squeezed states $\left\vert \varphi _{L}\left( T_M\right) \right\rangle _{M}$ and $\left\vert \varphi _{R}\left( T_M\right) \right\rangle _{M}$, respectively. The other parameters are $J/g_0=398.6$, $\Delta_0/g_0=100$ and $\delta/g_0=1.13$. []{data-label="Figure5"}](fig5a.eps "fig:")![(Color online) Wigner functions (a) $W_L(\alpha)$ and (b) $W_R(\alpha)$ for the superposed squeezed states $\left\vert \varphi _{L}\left( T_M\right) \right\rangle _{M}$ and $\left\vert \varphi _{R}\left( T_M\right) \right\rangle _{M}$, respectively. The other parameters are $J/g_0=398.6$, $\Delta_0/g_0=100$ and $\delta/g_0=1.13$. []{data-label="Figure5"}](fig5b.eps "fig:")
The interference effect in the superposed squeezed states can be seen from their Wigner quasi-probability distributions, which are defined as [@Buzek1995; @Barnett1997; @Walls1995] $$W_i(\alpha)=\frac{2}{\pi}{\rm{Tr}}[D^{\dagger}\left(\alpha\right)\rho_i D\left(\alpha\right)\left(-1\right)^{b^{\dagger}b}]. \label{Wi}$$ Here, $D\left(\alpha\right)={\rm{exp}}\left(\alpha b^{\dagger}-\alpha^{*}b\right)$ is the displacement operator, and $\rho_i$ is the density operator of state $\left\vert \varphi_i \left(T_M\right) \right\rangle _{M}$ with the subscript $i=L, R$. Substituting Eq. (\[phiLR\]) into Eq. (\[Wi\]), the Wigner functions for the superposed squeezed states $\left\vert \varphi _{L}\left( T_M\right) \right\rangle _{M}$ and $\left\vert \varphi _{R}\left( T_M\right) \right\rangle _{M}$, can be written as $$\begin{aligned}
W_{L}\left(\alpha\right) &=&\cos^{2}(JT_M) W_{+}+\sin^2(JT_M) W_{-}
\nonumber\\
&&+i\frac{\sin(2JT_M)}{2}(e^{-i\Delta_0T_M}W_{+-}-e^{i\Delta_0T_M}W_{-+}), \nonumber\\
W_{R}\left(\alpha\right) &=&\sin^{2}(JT_M) W_{+}+\cos^2(JT_M) W_{-}\nonumber\end{aligned}$$ $$+i\frac{\sin(2JT_M)}{2}(e^{i\Delta_0T_M}W_{-+}-e^{-i\Delta_0T_M}W_{+-}), \label{WLR}$$
![(Color online) Fidelity $F_{\rm{E}}(t)$ versus time with different (a) thermal phonon numbers $n_{\rm{th}}$, (b) mechanical decay rates $\gamma_m$, and (c) cavity decay rates $\gamma_c$ in the open system, respectively. The reference case is $n_{\rm{th}}=1$, $\gamma_m/g_0=10^{-4}$, and $\gamma_c/g_0=0.1$, as denoted by the blue solid curves in (a), (b) and (c). When one of them is varied, the others remain unchanged. Other parameters are $J/g_0=398.6$, $\Delta_0/g_0=100$ and $\delta/g_0=1.13$. []{data-label="Figure6"}](fig6.eps)
respectively. The parameter $W_{\pm}$ denotes the Wigner function of the squeezed state $\left \vert \pm Z(T_M)\right\rangle_M$, while $W_{\pm\mp}$ throws light on the quantum interference between two different squeezed states. For the compactness of the paper, the explicit expressions of $W_{+}$, $W_{-}$, $W_{+-}$ and $W_{-+}$ are shown in Appendix B. In Figs. \[Figure5\](a) and \[Figure5\](b), we show the Wigner functions $W_L(\alpha)$ and $W_R(\alpha)$ for the superposed squeezed states $\left\vert \varphi _{L}\left( T_M\right) \right\rangle _{M}$ and $\left\vert \varphi _{R}\left( T_M\right) \right\rangle _{M}$, respectively. In each figure, we find two error ellipses, which mainly come from the contribution of $W_{\pm}$, rotated from the horizontal axis by $\Phi/2$ and $(\Phi+\pi)/2$, respectively. The quantum fluctuation is strongly squeezed by $e^{-{R}}$ in one quadrature, while stretching by $e^{{R}}$ in the other one. Besides, the fringes appear around the two error ellipses, mainly come from the contribution of $W_{\pm\mp}$, denoting the quantum interference between the two different squeezed states. Moreover, it is obvious that the Wigner function $W_{R}\left(\alpha\right)$ in Fig. \[Figure5\](b), can be obtained by a rotation of $\pi/2$ from $W_{L}\left(\alpha\right)$ in Fig. \[Figure5\](a). This phenomenon actually is consistent with the relation, i.e., $\left\vert \varphi _{R}\left( T_M\right) \right\rangle _{M}=-e^{-ib^{\dagger}b\pi/2}\left\vert \varphi _{L}\left( T_M\right) \right\rangle _{M}$, which can be inferred from Eq. (\[phiLR\]). We would also like to mention that, without loss of generality, in Fig. \[Figure5\], we consider the maximum interference by choosing the tunneling rate $J$ satisfying $\left\vert\cos\left(JT_M\right)/ ie^{i\Delta_{0}t}\sin \left(JT_M\right)\right\vert\simeq1$. Till now, we have shown delicately the generations and quasi-probability distributions of Yurke-Stoler-like [@Yurke1986; @Yurke1997; @Yurke1990] enhanced squeezed vacuum states in the closed system. Actually, the method we used here can be extended to more general cases, e.g., generating the superposition of squeezed coherent or number states of mechanical resonator, but here we mainly focus on and take the vacuum case as an example. Next, we would like to study the effect of the environment fluctuations on the performance of this state generation scheme.
![(Color online) Wigner functions $W_{\rm{E}}(\alpha)$ for the superposed enhanced squeezed states $\left\vert \varphi _{L}\left( T_M\right) \right\rangle _{M}$ in the open system. The quasi-probability distribution of the reference case, which corresponds to $n_{\rm{th}}=1$, $\gamma_m/g_0=10^{-4}$, and $\gamma_c/g_0=0.1$, is shown in (a), while the cases when one of the dissipative parameters is varied is shown in (c-d). For example, the case corresponding to $n_{\rm{th}}=12$ is shown in (b), $\gamma_m/g_0=12\times10^{-4}$ in (c), and $\gamma_c/g_0=1.2$ in (d).[]{data-label="Figure7"}](fig7.eps)
EFFECTS OF DISSIPATIONS {#dissipation}
=======================
Let us now study the effects of environmental fluctuations on the state generation, including the fidelity and Wigner quasi-probability distribution in open system.
We study the environmental effect by using the master equation and including the dissipative terms, which can be expressed in the form of Lindblad superoperator, i.e., $$\begin{aligned}
\mathcal{L}_{\rm{diss}}&\simeq&\frac{\gamma_{m}}{2} \left[\left(n_{\rm{th}}+1\right)\mathcal{D}\left[b\right]+n_{\rm{th}}\mathcal{D}\left[b^{\dagger}\right] \right] \nonumber\\
&&+\gamma_{L}\mathcal{D}\left[a_L\right]+\gamma_{R}\mathcal{D}\left[a_R\right]. \label{Ldiss}\end{aligned}$$ Here, $\gamma_{L}$, $\gamma_{R}$, and $\gamma_{m}$ represent the decay rates of the left, right cavity modes and mechanical resonator, respectively. For simplicity and without loss of generality, we will assume that the photons in left and right cavities decay in the same rate in the following part of our paper, i.e., $\gamma_{L}=\gamma_{R}=\gamma_{c}$. The superoperator has the standard form of $$\mathcal{D}\left[o\right]\rho=o\rho o^{\dagger}-\left(o^{\dagger}o\rho+\rho o^{\dagger}o\right)/2,$$ where $\rho$ is the reduced density matrix of the system, and $o$ can be any system operator, i.e., $a_L, a_R, b$. The first line in Eq. (\[Ldiss\]) describes the coupling of the mechanical resonator to a thermal bath, and $n_{\rm{th}}=1/[\exp(\hbar\omega_{m}/k_{B}T)-1]$ denotes the thermal phonon number at temperature $T$, with $k_{B}$ the Boltzmann constant. The superoperators $\mathcal{D}\left[a_L\right]$ and $\mathcal{D}\left[a_R\right]$ represent the leakages of the cavity modes $a_L$ and $a_R$, respectively. Here, we assume that the cavity frequencies are much higher than the mechanical frequency, such that the thermal effect on the cavities can be neglected. Thus, the evolution of the system can be given by the quantum master equation, i.e., $$\begin{aligned}
\dot{\rho}=i\left[ \rho,H\left(t\right)\right]+\mathcal{L}_{\rm{diss}}\rho,\label{rho}\end{aligned}$$ with the full Hamiltonian $H\left(t\right)$ given in Eq. (\[H\]). Once the initial condition corresponding to the initial state is given, e.g., $\rho\left(0\right)=\left\vert \psi^{\prime} \left( 0\right) \right\rangle\left\langle \psi^{\prime} \left( 0\right) \right\vert$, Eq. (\[rho\]) can be solved numerically in the complete basis set $\left\vert n_{L}\right\rangle\otimes \left\vert n_{R}\right\rangle\otimes \left\vert n_{b}\right\rangle$. Here, $n_{L}$, $n_{R}$, and $n_{b}=0,1,2\cdots$ denote the excitation numbers in cavity modes $a_L, a_R$, and mechanical resonator mode $b$, respectively. In this work, the numerical solution $\rho\left(t\right)$ of the master equation in Eq. (\[rho\]) can be obtained utilizing the quantum toolbox [@Johansson2012; @Johansson2013] within a truncated Fock state space.
In Fig. \[Figure6\] and Fig. \[Figure7\], we show the effects of environmental fluctuations on the fidelity $F_{\rm{E}}(t)$ and Wigner distribution $W_{\rm{E}}(\alpha)$, respectively. The subscript E denotes that the effect of the environment is involved. For the purpose of convenience, here we only take the state $\left\vert \varphi _{L}\left( T_M\right) \right\rangle _{M}$ as an example, and the environment has similar effect on the state $\left\vert \varphi _{R}\left( T_M\right) \right\rangle _{M}$. In Figs. \[Figure6\](a-c) and Figs. \[Figure7\](a-d), we consider different thermal phonon numbers $n_{\rm{th}}$, mechanical decay rates $\gamma_m$ and cavity decay rates $\gamma_c$, respectively. We take the case $n_{\rm{th}}=1$, $\gamma_m/g_0=10^{-4}$, and $\gamma_c/g_0=0.1$ as a reference, whose fidelity and Wigner function are denoted by the blue solid lines in Figs. \[Figure6\](a-c) and the quasi-probability distribution in Fig. \[Figure7\](a), respectively. Meanwhile, the cases when one of the dissipative parameters is varied, are denoted by the red dash-dotted lines in Figs. \[Figure6\](a-c) and the distributions in Figs. \[Figure7\](b-d). For example, the fidelity and Wigner function corresponding to $n_{\rm{th}}=12$ are shown in Figs. \[Figure6\](a) and \[Figure7\](b), $\gamma_m/g_0=12\times10^{-4}$ in Figs. \[Figure6\](b) and \[Figure7\](c), and $\gamma_c/g_0=1.2$ in Figs. \[Figure6\](c) and \[Figure7\](d).
Comparing the unperturbed cases in Figs. \[Figure4\](a) and \[Figure5\](a) with the perturbed ones in Figs. \[Figure6\] and \[Figure7\] respectively, we can see great correspondence in the overall behavior, except for larger and faster oscillation, which are induced by faster photon and phonon leakages. Besides, in Figs. \[Figure6\](a) and \[Figure6\](b), when $n_{\rm{th}}$ and $10^{4}\gamma_M/g_0$ are increased from $1$ (blue-solid curve), to $12$ (red dash-dotted curve), the fidelity $F_{\rm{E}}(t)$ remains almost unchanged. The corresponding Wigner distributions in Figs. \[Figure7\](a), (b), (c) show almost the same patterns and absolute probability amplitudes. The system shows slightly obvious but reliability guaranteed degradation due to the increase of cavity dissipation, as shown in the decreasing amplitudes of the red dash-dotted curve in Fig. \[Figure6\](c) and quasi-probabilities in Fig. \[Figure7\](d). This phenomenon can be attributed to the fact that the photon leakage is the foremost resource of system dissipation.
Moreover, our approach remains robust even with further strong dissipation, e.g., $n_{\rm{th}}=50$, $\gamma_m/g_0=50\times10^{-4}$, or $\gamma_c/g_0=5$, which are checked but not included in our paper for the clearness of figures. When the system dissipations are further increased, the quasi-probability amplitudes show small change, e.g., the amplitudes get smaller in diagonal areas while larger in non-diagonal interference fringes, which means the probability distribution of the system becomes diffused. However, the overall patterns remain similar with less disturbed cases. Thus, we can see that this state generation scheme shows robust performance and has good resistance to environmental dissipations.
CONCLUSIONS {#conclusion}
===========
In summary, we study the superposition of the macroscopically squeezed states with enhanced squeezing in a weakly coupled two-mode optomechanical system. Notably enhanced squeezing (12.16dB) for the mechanical resonator can be achieved with single-photon, in the way of modulating either the cavity frequencies or the photon hopping rate and optomechanical coupling strength. In this work, we mainly focus on the first case, i.e., the system with modulated cavity frequencies.
Once the modulating amplitude and frequency satisfy certain conditions, it is reasonable to conduct rotating wave approximation (RWA) and neglect far-off-resonant terms. Meanwhile, the system with near-resonant terms can be approximated by an effective quadratically coupled two-mode optomechanical system with high fidelity. The Hamiltonian with flexibly modified ratio of coupling strength indicates that further considerable larger squeezing can be obtained by carefully adjusting the system parameters. Moreover, by initializing the two cavity modes into proper states and measuring the state of photon after time $T_M$, the mechanical resonator can be designed into a superposition state. In terms of Wigner quasi-probability distribution, we show the enhanced squeezing and superposition properties of the Yurke-Stoler-type [@Yurke1986; @Yurke1997; @Yurke1990] squeezed state, from the squeezed error ellipses and interference fringes, respectively. We also verify the validity of our approach in both the closed and open systems, i.e., without and with the dissipative fluctuations. The fidelity and Wigner distribution remain almost unchanged with small perturbations, while show little diffusion but with reliability guaranteed and pattern protected under further strong dissipation.
Our research show that this state generation scheme is reliable and exhibit comparable tolerance to environmental fluctuations. The generated states, which have enhanced squeezing and superposition properties, hold great promise for both fundamental interest, e.g., quantum superposition principle, and practical value, e.g., ultrasensitive measurement.
Y.X.L. acknowledges the financial support of the National Basic Research Program of China 973 Program under Grant No. 2014CB921401, the Tsinghua University Initiative Scientific Research Program, and the Tsinghua National Laboratory for Information Science and Technology (TNList) Cross-discipline Foundation. Y. Zhang would like to acknowledge the financial support of the NSF of China under Grants No. 11674201.
MODULATED COUPLING STRENGTHS {#Mcoupling}
============================
In this appendix, we consider the case when sinusoidal modulations are applied both to the photon-hopping interaction $J$ and quadratically coupling strength $g_0$, but with different modulating strength and frequency. Then the Hamiltonian has the form of $$\begin{aligned}
H^{\prime}\left( t\right) &=&\omega _{c}\left( a_{L}^{\dagger }a_{L}+a_{R}^{\dagger
}a_{R}\right)+J(t) \left(
a_{L}^{\dagger }a_{R}+a_{R}^{\dagger }a_{L}\right) \nonumber\\
&&+\omega _{M}b^{\dagger }b+g(t) \left( a_{L}^{\dagger
}a_{L}-a_{R}^{\dagger }a_{R}\right) \left( b^{\dagger }+b\right) ^{2}, \nonumber\\\end{aligned}$$ where the time-dependent coupling strengths $$J(t)=J\omega _{0}\cos \left( \omega _{0}t\right),\ \ \ g(t)=g_{0}\cos \left( 2\omega _{p}t\right).$$ Similarly, to see clearly the effect of coupling strength modulation, we perform a series of transformation defined by $$\begin{aligned}
V_{1}\left( t\right) &=&e^{-i\left[ \omega _{c}\left( a_{L}^{\dagger
}a_{L}+a_{R}^{\dagger }a_{R}\right) +\omega _{M}b^{\dagger }b\right] t}, \nonumber\\
V_{2}\left( t\right) &=&e^{-iJ\sin \left( \omega _{0}t\right) \left(
a_{L}^{\dagger }a_{R}+a_{R}^{\dagger }a_{L}\right) }.\end{aligned}$$ Then the transformed Hamiltonian can be written as $$\begin{aligned}
H^{\prime}_{2}\left( t\right)&=&\left[
\begin{array}{c}
\cos \left( 2J \sin \left( \omega _{0}t\right) \right) \left(
a_{L}^{\dagger }a_{L}-a_{R}^{\dagger }a_{R}\right) \\
+i\sin \left( 2J \sin \left( \omega _{0}t\right) \right) \left(
a_{R}^{\dagger }a_{L}-a_{R}a_{L}^{\dagger }\right)
\end{array}%
\right]\nonumber\\ &&\times \frac{g_{0}}{2}\left[G_1(b)+\rm{H.c.}\right]. \label{H2pt}
%\left[
%\begin{array}{c}
%b^{\dagger 2}e^{2i\left( \omega _{m}-\omega _{p}\right) t}+b^{2}e^{-i\left(
%\omega _{m}-\omega _{p}\right) t} \\
%b^{\dagger 2}e^{2i\left( \omega _{p}+\omega _{m}\right) t}+b^{2}e^{-i\left(
%\omega _{p}+\omega _{m}\right) t} \\
%\left( e^{2i\omega _{p}t}+e^{-2i\omega _{p}t}\right) \left( 2b^{\dagger
%}b+1\right)
%\end{array}%
%\right] \end{aligned}$$ Utilizing the Jacobi-Anger expansions $$\begin{aligned}
\cos \left( 2J \sin \left( \omega _{0}t\right) \right) &=&J_{0}\left(
2J \right) +2\sum\limits_{n=0}^{\infty }J_{2n}\left( 2J \right) \cos
\left( 2n\omega _{0}t\right), \\
\sin \left( 2J \sin \left( \omega _{0}t\right) \right)
&=&2\sum\limits_{n=0}^{\infty }J_{2n-1}\left( 2J \right) \sin \left(
\left( 2n-1\right) \omega _{0}t\right), \end{aligned}$$ with $J_m(x)$ being the Bessel function of the first kind and $m$ being an integer, the Hamiltonian $H^{\prime}_{2}(t)$ in Eq. (\[H2pt\]) can be expanded into $$\begin{aligned}
H^{\prime}_{2}(t)&=&\frac{g_{0}}{2}J_{0}\left( 2\xi \right) \left( a_{L}^{\dagger
}a_{L}-a_{R}^{\dagger }a_{R}\right) \left[G_1(b)+\rm{H.c.}\right] \nonumber\\
&&+\frac{g_{0}}{2}\left( a_{L}^{\dagger }a_{L}-a_{R}^{\dagger }a_{R}\right)
\sum\limits_{n=0}^{\infty }J_{2n}\left( 2\xi \right) \left[G_{2n}(b)+\rm{H.c.}\right] \nonumber \\
&&+\frac{g_{0}}{2}\left( a_{R}^{\dagger }a_{L}-a_{R}a_{L}^{\dagger }\right)
\sum\limits_{n=0}^{\infty }J_{2n-1}\left( 2\xi \right)\left[G_{3n}(b)-\rm{H.c.}\right]. \nonumber\\\end{aligned}$$ Here the functions have the form of $$\begin{aligned}
G_1(b)&=&b^{2}\left(e^{-2i\delta_{-}^{\prime} t}+e^{-2i\delta_{+}^{\prime} t}\right)+\left( 2b^{\dagger }b+1\right)e^{2i\omega _{p}t},\\
G_{2n}(b)&=&b^{2}\left(e^{-2i\delta _{n--} t}+e^{-2i\delta _{n-+} t}+e^{-2i\delta _{n+-} t}+e^{-2i\delta _{n++} t}\right)\nonumber\\
&&+\left( 2b^{\dagger }b+1\right)\left(e^{2i\delta p_{n-}t}+e^{2i\delta p_{n+}t}\right),\\
G_{3n}(b)&=&b^{2}\left(e^{-2i\delta _{n--} t}+e^{-2i\delta _{n+-} t}\right)e^{-i\omega_0t}\nonumber\\
&&-\left(e^{-2i\delta _{n++} t}+e^{-2i\delta _{n-+} t}\right)e^{i\omega_0t}\nonumber\\
&&+\left( 2b^{\dagger }b+1\right)\left(e^{2i\delta p_{n+}t}e^{-i\omega_0t}-e^{2i\delta p_{n-}t}e^{i\omega_0t}\right),\end{aligned}$$ with the detunings defined as $$\begin{aligned}
\delta^{\prime}_{\pm}&=&\omega_M\pm\omega_p,\\
\delta _{n\pm\pm}&=&\delta^{\prime}_{\pm}\pm n\omega_0,\\
\delta p_{n\pm}&=&\omega_p\pm n\omega_0.\end{aligned}$$ Obviously, more frequency terms appear because of the modulation of coupling strength, but we can always identify fast oscillating terms under specific conditions and perform the rotating-wave approximation (RWA). Consider the case $$\delta^{\prime}, g^{\prime} \ll \omega_0, \omega_M-\omega_p, \omega_p, \label{condition2}$$ with $$\delta^{\prime}=\delta _{n_0--}, \ \ \ g^{\prime}=\frac{g_0}{2}J_{2n_0}(2J), %\omega_m-\omega_p-n_0\omega_0$$ then we can arrive at $$\tilde{H}_{2}\left( t\right) \simeq g^{\prime}\left( a_{L}^{\dagger
}a_{L}-a_{R}^{\dagger }a_{R}\right) \left( b^{2}e^{-2i\delta^{\prime} t}+b^{\dagger
2}e^{2i\delta^{\prime} t}\right), \label{Heff2}$$ which is in the same form with Eq. (\[Heff\]).
INTERFERENCE EFFECT IN WIGNER DISTRIBUTION {#wigner}
==========================================
In this section, we show the details in the process of getting Eq. (\[WLR\]). Utilizing Eq. (\[phiLR\]), Eq. (\[Wi\]) can be written in the form of Eq. (\[WLR\]), where the explicit expressions of $W_{+}$, $W_{-}$, $W_{+-}$ and $W_{-+}$ can be written deliberately as $$\begin{aligned}
W_{+}&=&\frac{2}{\pi}\rm{Tr}[D^{\dagger}\left(\alpha\right)\left\vert Z\right\rangle \left\langle Z\right\vert D\left(\alpha\right)\left(-1\right)^{b^{\dagger}b}]\nonumber\\
&=&\frac{2}{\pi}\sum_{k}\left(-1\right)^k\left\vert S\left(k,-\alpha,Z\right) \right\vert^2;\\
W_{+-}&=&\frac{2}{\pi}\rm{Tr}[D^{\dagger}\left(\alpha\right)\left\vert Z\right\rangle \left\langle -Z\right\vert D\left(\alpha\right)\left(-1\right)^{b^{\dagger}b}]\nonumber\\
&=&\frac{2}{\pi}\sum_{k}\left(-1\right)^k S\left(k,-\alpha,Z\right)S^{*}\left(k,-\alpha,-Z\right);\\
W_{-+}&=&\frac{2}{\pi}\rm{Tr}[D^{\dagger}\left(\alpha\right)\left\vert -Z\right\rangle \left\langle Z\right\vert D\left(\alpha\right)\left(-1\right)^{b^{\dagger}b}]\nonumber\\
&=&\frac{2}{\pi}\sum_{k}\left(-1\right)^k S\left(k,-\alpha,-Z\right)S^{*}\left(k,-\alpha,Z\right);\\
W_{-}&=&\frac{2}{\pi}\rm{Tr}[D^{\dagger}\left(\alpha\right)\left\vert -Z\right\rangle \left\langle -Z\right\vert D\left(\alpha\right)\left(-1\right)^{b^{\dagger}b}]\nonumber\\
&=&\frac{2}{\pi}\sum_{k}\left(-1\right)^k\left\vert S\left(k,-\alpha,-Z\right) \right\vert^2;
\end{aligned}$$ Here, the parameter $S\left(k,\alpha,Z\right)$ denotes the amplitude of finding $k$ phonons in the squeezed coherent state $\left\vert \alpha, Z\right\rangle$ [@Gerry2004], i.e., $$\begin{aligned}
S\left(k,\alpha,Z\right)&=&\left\langle k \vert \alpha, Z\right\rangle \nonumber\\
&=&\frac{1}{\sqrt{\cosh R}}\rm{exp}\left[-\frac{1}{2}\left\vert\alpha \right\vert^2-\frac{1}{2}\alpha^{*2}e^{i\theta}\tanh R\right]\nonumber\\
&&\times\frac{\left[ \frac{1}{2}e^{i\theta}\tanh R\right]^{k/2}}{\sqrt{n!}}H_n\left[\gamma \left( e^{i\theta}\sinh(2R)\right)^{-1/2} \right] \nonumber\\\end{aligned}$$
[10]{} O. Romero-Isart, A. C. Pflanzer, F. Blaser, R. Kaltenbaek, N. Kiesel, M. Aspelmeyer, and J. I. Cirac, Phys. Rev. Lett. **107**, 020405 (2011). A. Asadian, C. Brukner, and P. Rabl, Phys. Rev. Lett. **112**, 190402 (2014). M. Arndt and K. Hornberger, Nature Physics **10**, 271 (2014). I. Pikovski, M. Zych, F. Costa, and C. Brukner, Nature Physics **11**, 668 (2015). E.Schrödinger, Naturwissenschaften **23**, 807 (1935). D. Leibfried, E. Knill, S. Seidelin, J. Britton, R. B. Blakestad, J. Chiaverini, D. B. Hume, W. M. Itano, J. D. Jost, C. Langer, R. Ozeri, R. Reichle and D. J. Wineland, Nature **438**, 639 (2005). A. Ourjoumtsev, H. Jeong, R. Tualle-Brouri, P. Grangier, Nature **448**, 784 (2007). L.-M. Duan, Nat. Photon. **13**, 73 (2019). K. G. Johnson, J. D. Wong-Campos, B. Neyenhuis, J. Mizrahi and C. Monroe, Nat. Commun **8**, 697 (2017).
C. Monroe, D. M. Meekhof, B. E. King, and D. J. Wineland, Science **272**, 1131 (1996). T. Berrada, S. van Frank, R. Bücker, T. Schumm, J.-F. Schaff, and J. Schmiedmayer, Nat. Commun. **4**, 2077 (2013).
C. H. van der Wal, A. C. J. ter Haar, F. K. Wilhelm, R. N. Schouten, C. J. P. M. Harmans, T. P. Orlando, S. Lloyd, and J. E. Mooij, Science **290**, 773 (2000); S. Han, Y. Yu, X. Chu, S. Chu, and Z. Wang, *ibid*. 293, 1457 (2001); J. R. Friedman, V. Patel, W. Chen, S. K. Tolpygo, and J. E. Lukens, Nature (London) **406**, 43 (2000). Y.-X. Liu, L. F. Wei, and F. Nori, Phys. Rev. A **71**, 063820 (2005). J. Clarke and F. K. Wilhelm, Nature **453**, 1031 (2008). M. Devoret and R. J. Schoelkopf, Science **339**, 1169 (2013). N. W. Noel and C. R. Stround, Jr., Phys. Rev. Lett. **77**, 1913 (1996). M. Brune, E. Hagley, J. Dreyer, X. Maître, A. Maali, C. Wunderlich, J. M. Raimond, and S. Haroche, Phys. Rev. Lett. **77**, 4887 (1996). J. M. Raimond, M. Brune, and S. Haroche, Phys. Rev. Lett. **79**, 1964 (1997). J.-Q. Liao, J.-F. Huang, and L. Tian, Phys. Rev. A **93**, 033853 (2016). J.-Q. Liao, and L. Tian, Phys. Rev. Lett. **116**, 163602 (2016). U. B. Hoff, J. Kollath-Bönig, J. S. Neergaard-Nielsen, U. L. Andersen, Phys. Rev. Lett. **117**, 143601 (2016). M. Aspelmeyer, T. J. Kippenberg, and F. Marquardt, Rev. Mod. Phys. **86**, 1391 (2014).
C. M. Caves, K. S. Thorne, R. W. P. Drever, V. D. Sandberg, and M. Zimmermann, Rev. Mod. Phys. **52**, 341 (1980). S. Lorenz, C. Silberhorn, N. Korolkova, R. S. Windeler, G. Leuchs, Appl. Phys. B **73**, 855 (2001). M. F. Bocko and R. Onofrio, Rev. Mod. Phys. **68**, 755–799 (1996). J. J. Bollinger, W. M. Itano, D. J. Wineland, and D. J. Heinzen, Phys. Rev. A **54**, R4649 (1996). E. M. Kessler, P. Kómár, M. Bishof, L. Jiang, A. S. Sørensen, J. Ye, and M. D. Lukin, Phys. Rev. Lett. **112**, 190403 (2014). B. C. Barish and R. Weiss, Phys. Today **52**, 44 (1999). J. Aasi, J. Abadie, B. P. Abbott, R. Abbott, Nature Photon. **7**, 613 (2013). C. Hempel, B. P. Lanyon, P. Jurcevic, R. Gerritsma, R. Blatt, C. F. Roos, Nature Photon. **7**, 630 (2013).
E. E. Hach III and C. C. Gerry, J. mod. Optics **40**, 2351 (1993). H. Fan and Z. Zhang, Quant. Opt. **6**, 411 (1994). A.-S. F. Obada and Z. M. Omar, Phys. Lett. A **227**, 349 (1997). Y. A. Barbosaa, G. C. Marquesa, B. Baseia, Physica A **280**, 346 (2000). F. A. A. El-Orany, arXiv: 1108.4765v1. M. Balamurugan, R. Chakrabarti, B. Virgin Jenisha, Physica A **473**, 428 (2017). B. C. Sanders, Phys. Rev. A **39**, 4284 (1998). W. Schleich, D. F. Walls, and J. A. Wheeler, Phys. Rev. A **38**, 1177 (1988). Z. Z. Xin, D. B. Wang, M. Hirayama, and K. Matumots, Phys. Rev. A **50**, 2865 (1994).
M. D. LaHaye, O. Buu, B. Camarota, K. C. Schwab, Science **304**, 74 (2004).
A. A. Clerk, F. Marquardt, and K. Jacobs, New J. Phys. **10**, 095010 (2008). J. B. Hertzberg, T. Rocheleau, T. Ndukum, M. Savva, A. A. Clerk and K. C. Schwab, Nature Phys. **6**, 213 (2010). A. Szorkovszky, A. C. Doherty, G. I. Harris, and W. P. Bowen, Phys. Rev. Lett. **107**, 213603 (2011). K. Jähne, C. Genes, K. Hammerer, M. Wallquist, E. S. Polzik, and P. Zoller, Phys. Rev. A **79**, 063819 (2009).
A. Mari and J. Eisert, Phys. Rev. Lett. **103**, 213603 (2009). A. Farace and V. Giovannetti, Phys. Rev. A **86**, 013820 (2012). K. Qu and G. S. Agarwal, New J. Phys. **16**, 113004 (2014). C. U. Lei, A. J. Weinstein, J. Suh, E. E. Wollman, A. Kronwald, F. Marquardt, A. A. Clerk, and K. C. Schwab, Phys. Rev. Lett **117**, 100801 (2016). A. Kronwald, F. Marquardt, and A. A. Clerk, Phys. Rev. A **88**, 063833 (2013). A. Kronwald, F. Marquardt, and A. A. Clerk, New J. Phys. **16**, 063058 (2014). M. J. Woolley and A. A. Clerk, Phys. Rev. A **89**, 063805 (2014). W.-J. Gu and G.-X. Li, Opt. Express **21**, 20423 (2013) C.-G. Liao, R.-X. Chen, H. Xie and X.-M. Lin, Phys. Rev. A **97**, 042314 (2018). R. Zhang, Y. Fang, Y.-Y. Wang, S. Chesi, and Y.-D. Wang, Phys. Rev. A **99**, 043805 (2019).
F. Monifi, J. Zhang, Ş. K. Özdemir, B. Peng,Y.-X. Liu, F. Bo, F. Nori, and L. Yang, Nat. Photonics **10**, 399 (2016). X. Y. Lü, H. Jing, J. Y. Ma, and Y. Wu, Phys. Rev. Lett. **114**, 253601 (2015). L. Bakemeier, A. Alvermann, and H. Fehske, Phys. Rev. Lett. **114**, 013601 (2015). H. Shi, and M. Bhattacharya, Phys. Rev. A **87**, 043829 (2013). A. Nunnenkamp, K. Børkje, J. G. E. Harris, and S. M. Girvin, Phys. Rev. A **82**, 021806(R) (2010). Y. Han, J. Cheng, and L. Zhou, Eur. Phys. J. D **67**, 20 (2013). J.-Q. Liao, K. Jacobs, F. Nori, and R. W. Simmonds, New J. Phys. **16** 072001 (2014). J.-Q. Liao, C. K. Law, L.-M. Kuang, and F. Nori, Phys. Rev. A **92**, 013822 (2015).
J.-F. Huang, J.-Q. Liao, L. Tian, and L. M. Kuang, Phys. Rev. A **96**, 043849 (2017). J.-F. Huang, J.-Q. Liao, and L.-M. Kuang, arXiv:1902.05779v1. Y.-L. Wu, L.-P. Yang, M. Gong, Y.-R. Zheng, H. Deng, Z.-G. Yan, Y.-J. Zhao, K.-Q. Huang, A. D. Castellano, W. J. Munro, K. Nemoto, D.-N. Zheng, C. P. Sun, Y.-X. Liu, X.-B. Zhu, and L. Lu, npj Quantum Information **4**, 50 (2018). F. Quijandría, U. Naether, S. K. Özdemir, F. Nori, and D. Zueco, Phys. Rev. A **97**, 053846 (2018).
K. W. Murch, K. L. Moore, S. Gupta and D. M. Stamper-Kurn, Nat. Phys. **4**, 561 (2008). T. P. Purdy, D. W. C. Brooks, T. Botter, N. Brahms, Z.-Y. Ma, and D. M. Stamper-Kurn, Phys. Rev. Lett. **105**, 133602 (2010). J. D. Thompson, B. M. Zwickl, A. M. Jayich, F. Marquardt, S. M. Girvin, and J. G. E. Harris, Nature **452**, 72 (2008). J. H. Teng, S. L. Wu, B. Cui, and X. X. Yi, J. Phys. B: At. Mol. Opt. Phys. **45**, 185506 (2012). C. Bai, B. P. Hou, D. G. Lai, and D. Wu, Phys. Rev. A **93**, 043804 (2016). A. Vinante and P. Falferi, Phys. Rev. Lett. **111**, 207203 (2013). G. S. Agarwal and S. Huang, Phys. Rev. A **93**, 043844 (2016).
F. Xue, Y.-X. Liu, C. P. Sun, and F. Nori, Phys. Rev. B **76**, 064305 (2007).
M. Schmidt, S. Kessler, V. Peano, O. Painter, and F. Marquardt, Optica , 635 (2015). O. Svelto, *Principles of Laser* (Plenum, New York, 1982). M. O. Scully and M. S. Zubairy, Quantum Optics (Cam- bridge University Press, Cambridge, 1997). M. P. Silveri, J. A. Tuorila, E. V. Thuneberg, and G. S. Paraoanu, Rep. Prog. Phys. **80**, 056002 (2017). A. Mufti, H. A. Schmitt, and M. Sargent III, Am. J. Phys. **61**, 729 (1993). B. Yurke and D. Stoler, Phys. Rev. Lett. **57**, 13 (1986). B. Yurke and D. Stoler, Phys. Rev. Lett. **79**, 4941 (1997). B. Yurke, W. Schleich, and D. F. Walls, Phys. Rev. A **42**, 1703 (1990).
E. M. E. Zayed, A. S. Daoud, M. A. AL-Laithy, E. N. Naseem, Chaos, Solitons and Fractals **24**, 967 (2005). C. Gerry and P. Knight, *Introductory Quantum Optics* (Cambridge Press, 2004). V. Buzek and P. L. Knight, in *Quantum Interference, Superposition States of Light, and Nonclassical Effects*, Progress in Optics Vol. XXXIV, edited by E. Wolf (Elsevier, Amsterdam, 1995). S. M. Barnett and P. M. Radmore, *Methods in Theoretical Quantum Optics* (Clarendon Press, Oxford, 1997). D. F. Walls and G. Milburn, *Quantum optics*, Springer Study Edition (Springer-Verlag, 1995). J. R. Johansson, P. D. Nation, and F. Nori, Comp. Phys. Comm. **183**, 1760 (2012) J. R. Johansson, P. D. Nation, and F. Nori, Comp. Phys. Comm. **184**, 1234 (2013)
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Dong-Won Jung [^1]'
- Jae Yong Lee
title: '1-loop Corrections to the $\rho$ Parameter in the Left-Right Twin Higgs Model'
---
Introduction {#intro}
============
The Standard Model (SM) has excellently described high energy physics up to energies of ${\mathcal O}(100)$ GeV. The only undetected constituent of the SM, up to now, is a Higgs boson which is required to explain the generation of fermion and gauge boson masses. Theoretically, the Higgs boson mass squared is quadratically sensitive to any new physics scale beyond the Standard Model (BSM) which may arise at higher energy scales and hence stabilization of the Higgs mass squared prefers the energy scale at which the BSM turns up to be lowered to ${\mathcal O}(1)$ TeV. On the other hand, electroweak precision measurements with naive naturalness assumption raise the energy scale of the BSM up to 100 TeV or even higher. Hence, there remains a tension between theory and experiment associated with the stabilization of the SM Higgs mass. But with the start-up of the LHC the tension may be relaxed by direct observation of the BSM at TeV energy scale. The idea of little Higgs originates in the speculation that the SM Higgs may be a pseudo-Nambu-Goldstone boson [@Georgi:1974yw; @Kaplan:1983fs; @Arkani-Hamed:2001nc; @Arkani-Hamed:2002pa; @Arkani-Hamed:2002qx; @Arkani-Hamed:2002qy; @Schmaltz:2002wx]. Stabilization of the Higgs mass in the little Higgs theories is achieved by the “collective symmetry breaking” which naturally renders the SM Higgs mass much smaller than the symmetry breaking scale. The distinct elements of little Higgs models are a vector-like heavy top quark and various scalar and vector bosons. The former is universal while the latter is model-dependent. Both of them contribute significantly to one-loop processes and hence establish strict constraints on the parameter space of little Higgs models. At worst, electroweak precision tests push up the symmetry breaking scale to $5$ TeV or higher, and regenerate significant fine-tuning in the Higgs potential. Twin Higgs idea shares the same origin with that of little Higgs in that the SM Higgs is a pseudo-Nambu-Goldstone boson [@Chacko:2005pe]. But rather than using collective symmetry breaking to stabilize the Higgs mass squared it makes use of additional discrete symmetry. In other words, the discrete symmetry ensures the absence of quadratic divergence in the Higgs mass squared. The twin Higgs mechanism is realized by identifying the discrete symmetry with left-right symmetry in the left-right model [@Chacko:2005un]. The left-right twin Higgs (LRTH) model contains $U(4)_1\times U(4)_2$ global symmetry as well as $SU(2)_L\times SU(2)_R\times U(1)_{B-L}$ gauge symmetry. The left-right symmetry acts on only the two $SU(2)$’s gauge symmetry. A pair of vector-like heavy top quarks play a key role at triggering electroweak symmetry breaking just as that of the little Higgs theories. Besides, the other Higgs particles acquire large masses not only at quantum level but also at tree level. These heavy Higgs bosons make the model deliver rich phenomenology at the LHC [@Goh:2006wj]. But theoretically, they lead to large radiative corrections to one-loop processes and, in return, the allowed parameter space can be reduced significantly. In this paper, we perform a one-loop analysis of the $\rho$-parameter in the LRTH model to reduce the parameter space. This is based on the original work with Jae Yong Lee, KIAS [@Jung:2007ea].
Left-Right Twin Higgs Model [@Goh:2006wj] {#sec:1}
=========================================
The LRTH model is based on the global $U(4)_1\times U(4)_2$ symmetry, with a locally gauged subgroup $SU(2)_L\times SU(2)_R\times U(1)_{B-L}$. A pair of Higgs fields, $H$ and $\hat H$, are introduced and each transforms as $(4,1)$ and $(1,4)$ respectively under the global symmetry. They are written as $$H=\left(\begin{array}{c} H_L\\ H_R\end{array}\right),\qquad
\hat H=\left(\begin{array}{c} \hat H_L\\ \hat H_R\end{array}\right),$$ where $H_{L,R}$ and $\hat H_{L,R}$ are two component objects which are charged under the $SU(2)_L\times SU(2)_R\times U(1)_{B-L}$ as $$H_L\mbox{ and }\hat H_L\,:\,(2,1,1),\qquad
H_R\mbox{ and }\hat H_R\,:\,(1,2,1).$$ The global $U(4)_1\,(U(4)_2)$ symmetry is spontaneously broken down to its subgroup $U(3)_1\,(U(3)_2)$ with VEVs $$\langle H\rangle =\left(\begin{array}{c} 0\\ 0 \\ 0 \\ f\end{array}\right),\qquad
\langle \hat H \rangle = \left(\begin{array}{c} 0\\ 0\\ 0\\ \hat f \end{array}\right).$$ Each spontaneous symmetry breaking results in seven Nambu-Goldstone bosons, which are parameterized as $$H=fe^{\pi/f}\left(\begin{array}{c} 0 \\ 0\\ 0\\ 1\end{array}\right),~
\pi=\left(\begin{array}{cccc} -\frac{N}{2\sqrt{3}} & 0 & 0 & h_1 \\
0 & -\frac{N}{2\sqrt{3}} & 0 & h_2 \\
0 & 0 & -\frac{N}{2\sqrt{3}} & C \\
h^\ast_1 & h^\ast_2 & C^\ast & \frac{\sqrt{3}N}{2}\end{array}\right),$$ where $\pi$ is the corresponding Goldstone field matrix. $N$ is a neutral real field $C$ and $C^\ast$ are a pair of charged complex scalar fields, and $h_{SM}=(h_1,h_2)$ is the SM $SU(2)_L$ Higgs doublet. $\hat H$ is parameterized in the identical way by its own Goldstone boson matrix, $\hat \pi$, which contains $\hat N$, $\hat C$, and $\hat h=(\hat h^+_1,\hat h^0_2)$. In turn, two $U(4)/U(3)$’s symmetry breaking left with fourteen Nambu-Goldstone bosons. The linear combination of $C$ and $\hat C$, and the linear combination of $N$ and $\hat N$ are eaten by the gauge bosons of $SU(2)_R\times U(1)_{B-L}$, which is broken down to the $U(1)_Y$. The orthogonal linear combinations, a charged complex scalar $\phi^\pm$ and a neutral real pseudoscalar $\phi^0$, remain as Nambu-Goldstone bosons. On top of that, the SM Higgs acquires a VEV, $\langle h_{SM}\rangle=(0,v/\sqrt{2})$, and thereby electroweak symmetry $SU(2)_L\times U(1)_Y$ is broken down to $U(1)_{EM}$. But $\hat h$’s do not get a VEV and remain as Nambu Goldstone bosons. These Nambu Goldstone bosons acquire masses through quantum effects and/or soft symmetry breaking terms, so called $\mu$-terms, $$V_{\mu}=-\mu^2_r(H^\dagger_R\hat H_R+c.c.)+\hat \mu^2 \hat H^\dagger_L \hat H_L,$$ which contribute to the Higgs masses at tree level. Because of the extend gaugue symmetry, there are extra gauge bosons besides the SM gauge bosons, $W_H$ and $Z_H$, masses of which are proportional to $f$ and $\hat f$. The existence of the extra gauge bosons would be the typical feature of the generic left-right symmetric models.
To cancel the quadratic sensitivity of the Higgs mass to the top quark loops, a pair of vector-like, charge $2/3$ fermion $({\mathcal Q}_L,{\mathcal Q}_R)$ are incorporated into the top Yukawa sector, $$\begin{aligned}
&&{\mathcal L}_{Yuk}=\\ \nonumber
&&y_L\bar Q_{L3}\tau_2 H^\ast_L {\mathcal Q}_R+
y_R\bar Q_{R3}\tau_2 H^\ast_R {\mathcal Q}_L-M\bar{\mathcal Q}_L {\mathcal Q_R}+h.c.,\end{aligned}$$ where $\tau_2=\left(\begin{array}{cc} 0 & -1 \\ 1 & 0
\end{array}\right)$, $Q_{L3}=-i(u_{L3},d_{L3})$ and $Q_{R3}=(u_{R3},d_{R3})$ are the third generation up- and down-type quarks, respectively. The left-right parity indicates $y_L=y_R(\equiv y)$. The mass parameter $M$ is essential to the top mixing. The value of $M$ is constrained by the $Z\to b\bar b$ branching ratio. It can be also constrained by the oblique parameters, which we will do in the letter. Furthermore, it yields large log divergence of the SM Higgs mass. To compensate for it the heavy gauge bosons also get large masses by increasing the value of $\hat f$. Therefore it is natural for us to take $M \lesssim yf$.
Results and Discustion
======================
The $Z$-pole, $W$-mass, and neutral current data can be used to search for and set limits on deviations from the SM. In the article we concentrate particularly on the the $\rho$-parameter, which is defined as $$\label{eq:defrho}
\rho\equiv\frac{M^2_W}{M^2_Zc^2_\theta}.$$ The effective leptonic mixing angle $s^2_\theta(=1-c^2_\theta)$ at the $Z$-resonance is defined as the ratio of the electron vector to axial vector coupling constants to the $Z$-boson, $$\frac{Re(g^e_V)}{Re(g^e_A)}\equiv4s^2_\theta-1,$$ where the coupling constants of a fermion $\psi$ to the gauge boson $X$ is given as, $${\mathcal L}=i \bar \psi_1\gamma_\mu(g_V+g_A\gamma_5)\psi_2 X^\mu.$$ Using the procedure in Ref. [@Chen:2003fm], we can calculate the 1-loop corrected W boson mass $$\label{eq:mw}
M^2_W=\frac{1}{2}\Big[a(1+\Delta \hat r)+\sqrt{a^2(1+\Delta \hat r)^2+4a\Pi^{WW}(0)}\Big],$$ with $ \quad a\equiv\frac{\pi\alpha(M_Z)}{\sqrt{2}G_Fs^2_\theta}$, and the definition of $\Delta \hat r$ is $$\begin{aligned}
\Delta {\hat r}=
&&-\frac{\Delta s^2_\theta}{s^2_\theta}-\frac{Re(\Pi^{ZZ}(M^2_Z))}{M^2_Z}
+\Pi^{\gamma\gamma'}(0) \\ \nonumber
&&+2(\frac{g^e_V-g^e_A}{Q_e})\frac{\Pi^{\gamma Z}(0)}{M^2_Z}
-\frac{c^2_\theta-s^2_\theta}{c_\theta s_\theta}
\frac{Re(\Pi^{\gamma Z}(M^2_Z))}{M^2_Z}. \end{aligned}$$ The 1-loop corrected $\rho$ parameter is then obtained using Eq. (\[eq:defrho\]) with the $M^2_W$ value predicted in Eq. (\[eq:mw\]). For doing the calculation concerning the precision measurements, the standard experimental values are necessary which play as input parameters. Here, we use the following experimentally measured values for the input parameters [@Yao:2006px; @:2005em]: $$\begin{aligned}
G_F&=& 1.16637(1)\times 10^{-5}\mbox{ GeV}^{-2},\\
M_Z&=& 91.1876(21)\mbox{ GeV},\\
\alpha(M_Z)^{-1}&=&127.918(18),\\
s^2_\theta&=&0.23153(16).\end{aligned}$$ We also take the top and bottom quark masses as [@Yao:2006px; @Rodrigo:1997gy] $$m_t = 172.3 ~{\rm GeV}, \qquad m_b = 3 ~{\rm GeV},$$ where $m_t$ is the central value of the electroweak fit and $m_b$ is the running mass at the $M_Z$ scale with $\overline{MS}$ scheme. Including all the SM corrections (top quark loop, bosonic loops), we take the allowed range of $\rho$ parameter as [@Yao:2006px] $$1.00989 ~\leq ~\rho^{exp}~\leq ~1.01026.$$ The input parameters of the LRTH model [@Chacko:2005un] are as follows: $$f, ~M, ~\mu_r, ~\hat \mu,$$ where $M$ is the heavy top quark mass scale, both $\mu_r$ and $\hat \mu$ are soft symmetry breaking terms. The masses of the top and heavy top quarks are determined by $f$ and $M$ while those of the scalar particles $\hat h_1,~\hat h_2,~\phi^\pm$ and $\phi^0$ largely depend on $\hat \mu,~\mu_r$ and $f$. Another scale $\hat
f$, which is associated with the masses of the heavy gauge bosons, can be determined from the electroweak symmetry breaking condition: there is a generic relation between $\hat f$ and $f$ since Coleman-Weinberg potential of the Higgs boson mostly depends on $M, f$ and $\hat f$. For scalar potential, there is a tree level mass term proportional to $\mu_r^2$. So we may not acquire negative mass squared term which is necessary for electroweak symmetry breaking and it gives an upper bound for the value of $\mu_r$. For a given $f$, $\hat f$ becomes larger as $M$ increases. It is because the increase of $M$ contributes positively to the Higgs mass through the top loop while the increase of $\hat f$ contributes negatively to the Higgs mass through the gauge boson loop, and thereby these contributions cancel out themselves in order to retain $v=246$ GeV. To draw a meaningful information on the model parameters from the $\rho$-parameter , we scan the parameter space generally, i.e., $$500 ~{\rm GeV} ~\leq ~f ~\leq ~2500 ~{\rm GeV},~~~ 0~ \leq ~M,
~\mu_r, ~\hat \mu ~\leq ~f.$$ Even though too large $f$ makes the model unviable, we take the rather large value of $f$, 2.5 TeV, as an upper limit for completeness of the scanning. As a result of $\rho$-parameter calculation, we can obtain the allowed regions of parameter space. As an example, Fig. \[fig:fm\] shows the allowed regions of parameter space for $f$ versus $M$. It is interesting to notice that the allowed parameter space is divided into two regions; less than 670 GeV and larger than 1100 GeV roughly, for $f$. This can be figured out as follows. The loop corrections tend to be larger as $f$ increases. It is because the masses of the particles involved in one-loop correction increase in general as $f$ increases. But at the same time, the mixing angles of top-heavy top quarks also vary. Since the mixing angles depend on not only $f$ but also $M$, these two effects compete during the increase of $f$. Because of this interplay of top mixing angles and masses, we have two distinct allowed parameter spaces. For small $f$, solution points prefer very small values of $M$. It means there is no large mixing between the top and heavy top quarks. In general, $\Pi^{WW}(0)$ is large for small $f$, and decreases as $f$ increases. So for fitting the observed W-boson mass in the small $f$ region, which is directly related to the $\rho$-parameter, we restrict the $\Delta \hat r$ within rather small range. Because the $\Delta \hat r$ is mostly determined by $\Pi^{ZZ}(M_Z^2)$, it should be also small. For doing that, we should take the small value of $M$, which makes the masses and mixing angles of heavy top quark small. We find that in the small $f$ region, $M$ should be smaller than about $22~{\rm GeV}$. Soft symmetry breaking parameter $\mu_r$ is restricted to the values less than around $60~{\rm GeV}$. This bound arises mainly from the electroweak symmetry breaking condition, and is generically independent of the $\rho$-parameter. Another free parameters $\hat \mu$ is not restricted from the one-loop corrected $\rho$-parameter. The reason is that $\hat \mu$ only contributes to the masses of $\hat h_1$ and $\hat h_2$, and their contributions are effectively cancelled among the relevant loop diagrams. This region of parameter space can provide constraints on the masses of many particles appear in this model. First, let us consider the masses of the heavy top and heavy gauge bosons. Their masses generically increase as $f$ increases. The mass of the heavy top quark is uniquely determined when $f, \hat f$ and $M$ are fixed. So does top Yukawa coupling. Basically $\hat f$ is determined by the electroweak symmetry breaking condition, but their $M$ and $\mu_r$ dependence provoke the ambiguity on its value. For small $f$ region, since $M$ is also very small, the $m_T$ is almost determined by $f$ alone. For large $f$ region, it becomes spread due to the top mixing angles. The plots of the heavy $Z$ and $W$ boson masses versus $f$ are quite similar to that of the heavy top mass versus $f$. In the case of heavy $W$ boson, the strongest constraint come from $K_L -
K_S$ mixing. The strongest bound ever known is $m_{W_H} > 1.6$ TeV, with the assumption of $g_L = g_R$ [@Beall:1981ze]. This can exclude some region from Fig. \[fig:fw\]. In this case, small $f$ region can be completely excluded. If the lower bound for $f$ is confirmed, we can give the lower bound for $f$ as $1.1 ~{\rm
TeV}$ from our calculation of the $\rho$-parameter and also for many particles appear in the model. Another constraints on the $m_{W_H}$ from CDF and D0 are about $650 \sim 786 {\rm GeV}$, as lower bound [@Affolder:2001gr; @Abachi:1995yi]. For Our results remain safe from these experimental bounds. Heavy $Z$ boson has also been studied in detail by many experimentalists. Current experimental bound is about $500 \sim 800 {\rm GeV}$ from precision measurements [@Yao:2006px] and $\sim 630 {\rm GeV}$ from CDF [@Yao:2006px]. In this case, also safe is the mass of heavy $Z$ boson. With the parameters allowed by the $\rho$-parameter, the masses of new scalar bosons $\hat h_{1,2}, \phi^0$ and $ \phi^\pm$ are also constrained. $\hat h_{1,2}$ has almost degenerate masses, and are dependent on both $\mu_r$ and $\hat \mu$, unlike the $\phi^{0,\pm}$ which depend only on $\mu_r$. Their masses are seriously constrained according to the value of $f$. Unfortunately, we cannot give a lower bound on the mass of $\phi^0$. In fact, its mass, though it is quite small, arise from radiative corrections. For $\phi^\pm$, the loop contribution is rather large so it acquire larger mass compared to the neutral one. The $\rho$-parameter cannot give a strong restriction on the Higgs mass. In the whole space, Higgs mass is restricted below about $167~{\rm GeV}$. We cannot give a lower bound for Higgs boson mass from $\rho$-parameter itself. Here, we adopt the LEP bound for Higgs mass, $114.4~{\rm GeV}$ [@Barate:2003sz], since its structure is same as the SM. The generic behavior of Higgs mass as a function of $f$ is shown in Fig .\[fig:fh\].
![$M ~{\bf vs.} f$, allowed range.[]{data-label="fig:fm"}](f-m-p.eps){width="45.00000%" height="36.00000%"}
![$m_{W_H} ~{\bf vs.} f$, allowed range.[]{data-label="fig:fw"}](f-wh-p.eps){width="45.00000%" height="36.00000%"}
![$m_h ~{\bf vs.} f$, allowed range.[]{data-label="fig:fh"}](f-higgs-p.eps){width="45.00000%" height="36.00000%"}
Summary
=======
We summarize the results of our analysis as follows. With the observed $\rho$-parameter, we see that the allowed parameter space is divided into two separate regions: $f$ smaller than about 670 GeV and larger than about 1.1 TeV. We give the bounds on the mass spectrum of many particles for either region. Especially the heavy gauge bosons remain safe from the experimental constraints. Unlike the other particles, we cannot set a lower bound for the neutral $\phi^0$ scalar. But loop correction plays an important role for the charged $\phi^\pm$ scalars, yielding mass difference between the charged and neutral scalars. Further analysis is required in order to reduce the allowed region. If the small $f$ region is excluded, for example by Ref. [@Goh:2006wj], we can provide exact lower bounds for the masses of $T, Z_H, W_H, \hat h_{1,2}$, and $ \phi^\pm$. But even in that case, we cannot do so for $\phi^0$ and SM Higgs boson.
[999]{} H. Georgi and A. Pais, Phys. Rev. D [**10**]{} (1974) 539. D. B. Kaplan and H. Georgi, Phys. Lett. B [**136**]{} (1984) 183.
N. Arkani-Hamed, A. G. Cohen and H. Georgi, Phys. Lett. B [**513**]{} (2001) 232 \[arXiv:hep-ph/0105239\]. N. Arkani-Hamed, A. G. Cohen, T. Gregoire and J. G. Wacker, JHEP [**0208**]{} (2002) 020 \[arXiv:hep-ph/0202089\]. N. Arkani-Hamed, A. G. Cohen, E. Katz, A. E. Nelson, T. Gregoire and J. G. Wacker, JHEP [**0208**]{} (2002) 021 \[arXiv:hep-ph/0206020\].
N. Arkani-Hamed, A. G. Cohen, E. Katz and A. E. Nelson, JHEP [**0207**]{} (2002) 034 \[arXiv:hep-ph/0206021\]. M. Schmaltz, Nucl. Phys. Proc. Suppl. [**117**]{} (2003) 40 \[arXiv:hep-ph/0210415\].
Z. Chacko, H. S. Goh and R. Harnik, Phys. Rev. Lett. [**96**]{} (2006) 231802 \[arXiv:hep-ph/0506256\]. Z. Chacko, H. S. Goh and R. Harnik, JHEP [**0601**]{} (2006) 108 \[arXiv:hep-ph/0512088\]. H. S. Goh and S. Su, arXiv:hep-ph/0611015. D. W. Jung and J. Y. Lee, arXiv:hep-ph/0701071. M. C. Chen and S. Dawson, Phys. Rev. D [**70**]{} (2004) 015003 \[arXiv:hep-ph/0311032\]. W. M. Yao [*et al.*]{} \[Particle Data Group\], J. Phys. G [**33**]{} (2006) 1. \[ALEPH Collaboration\], Phys. Rept. [**427**]{} (2006) 257 \[arXiv:hep-ex/0509008\]. G. Rodrigo, A. Santamaria and M. S. Bilenky, Phys. Rev. Lett. [**79**]{} (1997) 193 \[arXiv:hep-ph/9703358\]. G. Beall, M. Bander and A. Soni, Phys. Rev. Lett. [**48**]{} (1982) 848. A. A. Affolder [*et al.*]{} \[CDF Collaboration\], Phys. Rev. Lett. [**87**]{} (2001) 231803 \[arXiv:hep-ex/0107008\]. S. Abachi [*et al.*]{} \[D0 Collaboration\], Phys. Rev. Lett. [**76**]{} (1996) 3271 \[arXiv:hep-ex/9512007\]. R. Barate [*et al.*]{} \[LEP Working Group for Higgs boson searches\], Phys. Lett. B [**565**]{} (2003) 61 \[arXiv:hep-ex/0306033\].
[^1]: *Email:* [email protected], now moved to Nat’l Central Univ., Taiwan, after Sep. 1.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The Depth-Image-Based-Rendering (DIBR) is one of the main fundamental technique to generate new views in 3D video applications, such as Multi-View Videos (MVV), Free-Viewpoint Videos (FVV) and Virtual Reality (VR). However, the quality assessment of DIBR-synthesized views is quite different from the traditional 2D images/videos. In recent years, several efforts have been made towards this topic, but there lacks a detailed survey in literature. In this paper, we provide a comprehensive survey on various current approaches for DIBR-synthesized views. The current accessible datasets of DIBR-synthesized views are firstly reviewed. Followed by a summary and analysis of the representative state-of-the-art objective metrics. Then, the performances of different objective metrics are evaluated and discussed on all available datasets. Finally, we discuss the potential challenges and suggest possible directions for future research.'
author:
- 'Shishun Tian, Lu Zhang, Wenbin Zou, Xia Li, Ting Su, Luce Morin, and Olivier Déforges[^1][^2] [^3]'
bibliography:
- 'bare\_jrnl.bib'
title: 'Quality Assessment of DIBR-synthesized views: An Overview'
---
[Shell : Bare Demo of IEEEtran.cls for IEEE Journals]{}
DIBR, Multi-view videos (MVV), view synthesis, distortions, quality assessment.
Introduction
============
more immersive experiences with depth perception to the observers, the 3D applications, such as the Multi-View Video (MVV) and Free-Viewpoint Video (FVV), have drawn great public attention in recent years. These 3D applications allow the users to view the same scene at various angles which may result in a huge information redundancy and cost tremendous bandwidth or storage space. To reduce these limitations, researchers attempt to transmit and store only a subset of these views and synthesize the others at receiver by using the Multiview-Video-Plus-Depth (MVD) data format and Depth-Image-Based-Rendering (DIBR) techniques [@fehn2004depth]. Only limited viewpoints (both texture images and depth maps) are included in the MVD data format, the other view images are synthesized through DIBR.
This MVD plus DIBR scenario greatly reduces the burden on the storage and transmission of 3D video content. However, the DIBR view synthesis technique also raises new challenges in the quality assessment of virtual synthesized views. During the DIBR process, the pixels in the texture image at the original viewpoint are back-projected to the real 3D space, and then re-projected to the target virtual viewpoint using the depth map, which is named 3D image warping in the literature. As shown in Fig. \[Fig: DIBR\], DIBR view synthesis can be divided into two parts: 3D image warping and hole filling. During the 3D image warping procedure, the pixels in the original view are warped to the corresponding position in the target view. Owing to the changing of viewpoint, some objects which are invisible in the original view may become visible in the target one, which is called dis-occlusion and causes black holes in the synthesized view. Then, the second step is to fill the black holes. The holes can be filled by typical image in-painting algorithms. Most of the image in-painting algorithms use the pixels around the “black holes" to search the similar regions in the same image, and then use this similar region to fill the “black holes". Due to the imprecise depth map and imperfect image in-painting method, various distortions, which are quite different from the traditional ones in 2D images/videos, may be caused. Most of the 2D objective quality metrics [@Wangssim; @wang2003multiscale; @wang2011information; @ma2011reduce; @liu2019blind; @liu2018reduce; @li2016blindimage; @gu2015usingfree] which focus on the traditional distortions will fail to evaluate the quality of DIBR-synthesized views. Subjective test is the most accurate and reliable way to assess the quality of media content since the human observers are the ultimate users in most applications. The subjective tests offer the datasets along with subjective quality scores. The objective metrics are designed to mathematically model and predict the subjective quality scores. In other words, an ideal objective model is expected to be consistent with the subjective results. Since the subjective test is time consuming and practically not suitable for real-time applications, effective objective metrics are highly desired.
![Procedure of DIBR[]{data-label="Fig: DIBR"}](DIBR_cropped){width="0.9\linewidth"}
Although several efforts have been made targeting at the objective quality assessment of DIBR-synthesized views in recent years, to the best of our knowledge, there is not a detailed survey on these works in current literature. In this paper, we provide a comprehensive survey on the quality assessment approaches for DIBR-synthesized views ranging from the subjective to objective methods. The main contributions can be summarized as follows: (1) the state-of-the-art metrics are introduced and classified based on their used approaches. (2) the metrics in terms of the contributions, advantages and disadvantages are analysed in depth. (3) the performances of these metrics are evaluated on different datasets, and the essential reasons of their performance on different type of distortions are analysed. (4) furthermore, the limitations of current works are discussed and the possible directions for future research are given.
The rest of this paper is organized as follows. The subjective methods are first surveyed in Section . Section introduces the state-of-the-art objective quality metrics in detail. The experimental results are presented and discussed in Section . Finally, the conclusions are given in Section .
Subjective image/video quality assessment of DIBR-synthesized views
===================================================================
Subjective test is the most direct method for image/video quality assessment. During the test, a group of human observers are asked to rate the quality of each tested image or video. The subjective test results obtained from the subjective ratings are recognized as the quality of the tested images/videos. In different subjective test methodologies, the acquisition of subjective scores are also different. The Absolute Category Rating (ACR) method used in IVC image / video datasets [@bosc2011towards; @bosc2013visual] randomly present the test sequences to the observers and ask them to rate on five-scales quality judgement (excellent, good, fair, poor, bad). The subjective quality scores are calculated by simply averaging the ratings. The Single Stimulus Continuous Quality Evaluation (SSCQE) in SIAT [@liu2015subjective] dataset allows the observer to rate on a continuous scale instead of a discrete five-scales evaluation. The IVY [@jung2016critical] image dataset uses the Double Stimulus Continuous Quality Scale (DSCQS). The test image along with its associated reference image are presented in succession. It is usually used when the test and reference images are similar. Pairwise Comparison (PC) method directly performs a one-to-one comparison of every image pair in the dataset. It is the most accurate and reliable way to get the subjective quality scores, but it costs too much time since all the image pairs need to be tested. Compared to PC method, the Subjective Assessment Methodology for VIdeo Quality (SAMVIQ) method used in IETR dataset can achieve much higher accuracy than ACR method for the same number of observers and cost less time than PC since it allows the observer to freely view several image multiple times and a continuous rating scale. Besides, the IVY [@jung2016critical], IETR [@tian2019benchmark] and SIAT datasets normalize the obtained scores to *z-score* to make the results more intuitive. The IVC and MCL-3D [@song2014mcl] datasets directly use the average scores. Except for the subjective test methodology, as shown in Table \[Tab:dibrdatabase\], they use different sequences, DIBR algorithms etc. In the following part, we will introduce them respectively in detail.
-- ----------------- -------------------- ---- ---------------- ------ -------------------------- -- -- --
Name Year
[BookArrival]{} 1024 $\times$ 768 Fehn’s 2004
Lovebird1 1024 $\times$ 768 Telea’s 2003
Newspaper 1024 $\times$ 768 VSRS 2009
Müller 2008
PC Ndjiki-Nya 2010
Köppel 2010
Black hole —
H.264
(QP: 26,34,44)
BookArrival 1024 $\times$ 768 Criminisi 2004
Lovebird1 1024 $\times$ 768 VSRS 2009
Newspaper 1024 $\times$ 768 LDI 2011
Balloons 1024 $\times$ 768 HHF 2012
Kendo 1024 $\times$ 768 Ahn’s 2013
Dancer 1920 $\times$ 1088 Luo’s 2016
Shark 1920 $\times$ 1088 Zhu’s 2016
Poznan\_Street 1920 $\times$ 1088
PoznanHall2 1920 $\times$ 1088
GT\_fly 1920 $\times$ 1088
Aloe 1280 $\times$ 1100 Criminisi 2004
Dolls 1300 $\times$ 1100 Ahn’s 2013
Reindeer 1300 $\times$ 1100 VSRS 2009
Laundry 1300 $\times$ 1100 Yoon 2014
Lovebird1 1024 $\times$ 768
Newspaper 1024 $\times$ 768
BookArrival 1024 $\times$ 768
Kendo 1024 $\times$ 768 Fehn’s 2004 [Additive White Noise]{}
Lovebird1 1024 $\times$ 768 Telea’s 2003 Blur
Balloons 1024 $\times$ 768 HHF 2012 Down sampling
Dancer 1920 $\times$ 1088 Black hole — JPEG
Shark 1920 $\times$ 1088 JPEG2k
Poznan\_Street 1920 $\times$ 1088 Translation Loss
PoznanHall2 1920 $\times$ 1088
GT\_fly 1920 $\times$ 1088
Microworld 1920 $\times$ 1088
BookArrival 1024 $\times$ 768
Balloons 1024 $\times$ 768
Kendo 1024 $\times$ 768
Lovebird1 1024 $\times$ 768
Newspaper 1024 $\times$ 768
Dancer 1920 $\times$ 1088
PoznanHall2 1920 $\times$ 1088 (QP: 24-50)
Poznan\_Street 1920 $\times$ 1088
GT\_fly 1920 $\times$ 1088
Shark 1920 $\times$ 1088
-- ----------------- -------------------- ---- ---------------- ------ -------------------------- -- -- --
PVS: Processed Video Sequences.
ACR: Absolute Categorical Rating.
PC: Pairwise Comparison.
SAMVIQ: Subjective Assessment Methodology for VIdeo Quality.
DSCQS: Double Stimulus Continuous Quality Scale.
SSCQ: Single Stimulus Continuous Quality Scale.
IVC DIBR datasets
-----------------
The IVC DIBR-image dataset [@bosc2011towards] was proposed by Bosc *et al.* in 2011. It contains 84 DIBR-synthesized view images synthesized by 7 DIBR algorithms [@fehn2004depth; @telea2004image; @mori2009view; @mueller2009view; @ndjiki2010depth; @ndjiki2011depth; @koppel2010temporally]. 3 Multi-view plus Depth (MVD) sequences, $BookArrival$, $Lovebird1$ and $Newspaper$, are extracted as the source contents. MVD sequences contain several view-points acquired with real cameras in rectified configuration capturing the same scene, together with a depth image for each acquired view. For each sequence, 4 virtual views are synthesized from the adjacent viewpoint by using the above algorithms. Note that in this dataset, virtual views were only generated by single-view-based synthesis, which means that the virtual view is synthesized with only one image and its associated depth map. The IVC DIBR-video dataset [@bosc2013visual] uses almost the same contents and methodologies except that it adds the H.264 compression (with 3 quantization levels) distortion for each test sequence. In other words, there are 93 distorted videos in this dataset, among which 84 ones only contain the DIBR view synthesis distortions. As one of the first DIBR related image datasets, the IVC datasets play an important role in the first research phase of this topic. However, because of the fast development of DIBR view synthesis algorithms, some of the distortions in these datasets do not exist in the state-of-the-art view synthesis algorithms.
IETR DIBR image dataset
-----------------------
Similar to the IVC datasets, the IETR dataset [@tian2019benchmark] is dedicated to investigate the DIBR view synthesis distortions as well. Compared to the IVC datasets, it uses more and newer DIBR view synthesis algorithms [@criminisi2004region; @jantet2011object; @ahn2013novel; @luo2016hole; @solh2012hierarchical; @mori2009view; @zhu2016depth], including both interview synthesis and single view based synthesis, to exclude some “old fashioned” distortions, such as “black holes”. The interview DIBR algorithms use two neighbouring views to synthesize the virtual viewpoint. In addition, the IETR dataset also uses more MVD sequences, of which 7 sequences ($Balloons$, $BookArrival$, $Kendo$, $Lovebird1$, $Newspaper$, $Poznan\ {Street}$ and $Poznan{Hall}$) are natural images and 3 sequences ($Undo\ {Dancer}$, $Shark$ and $Gt\ {Fly}$) are computer animation images. It contains 140 synthesized view images and their associated 10 reference images which are also the images captured by real cameras at the virtual viewpoints.
IVY stereoscopic image dataset
------------------------------
Jung *et al.* proposed the IVY stereoscopic 3D image dataset for the quality assessment of DIBR-synthesized stereoscopic images [@jung2016critical]. Different from the above two datasets, besides the DIBR view synthesis distortion, the IVY dataset also explores binocular perception by showing the synthesized image pairs on a stereoscopic display. A total of 7 sequences and three MVD sequences are selected. 84 stereo images are synthesized by four DIBR algorithms [@criminisi2004region], [@ahn2013novel], [@tanimoto2008reference], [@yoon2014inter] in this dataset. All the virtual view images in the IVY dataset are generated by single-view-based synthesis.
MCL-3D image dataset
--------------------
Song *et al.* proposed the MCL-3D stereoscopic image dataset [@song2014mcl] to evaluate the quality of DIBR-synthesized stereoscopic images. Although 4 DIBR algorithms are included, the number of images synthesized by these algorithms is quite limited (36 pairs). The major part of this dataset focuses on the traditional distortions in the synthesized views. 6 types of traditional distortions are considered in this dataset: additive white noise, Gaussian blur, down sampling blur, JPEG, JPEG2000 and transmission loss. Nine MVD sequences are collected, among which $Kendo$, $Lovebird1$, $Balloons$, $PoznanStreet$ and $PoznanHall2$ are natural images; $Shark$, $Microworld$, $GT\_Fly$ and $Undodancer$ are Computer Graphics images. For each sequence, these traditional distortions are first applied on the base views. Then, the left and right view images are synthesized from these distorted base view images by using the view synthesis reference software (VSRS) [@mori2009view]. Different from the above IVC, IETR and IVY datasets, the reference images in the MCL-3D dataset are the images synthesized from undistorted base view images instead of the ones captured by real cameras.
SIAT synthesized video dataset
------------------------------
The SIAT synthesized video dataset [@liu2015subjective] focuses on the distortions caused by compressed texture and depth images in the synthesized views. It uses the same 10 MVD sequences as the IETR image dataset. For each sequence, 4 different texture and depth image quantization levels and their combinations are applied on the base views. Then, the videos at the virtual viewpoints are synthesized using the VSRS-1D-Fast software [@sullivan2013standardized]. This dataset uses the real images (captured by real cameras at the virtual viewpoint) as references. Only interview synthesis is used in this dataset.
In the above datasets, the distortions in the DIBR-synthesized views come from not only the DIBR view synthesis algorithms, but also from the distorted texture and depth images. The IVC [@ivcDIBRimagedatabase; @bosc2011towards; @bosc2013visual], IVY [@jung2016critical] and IETR [@tian2019benchmark] datasets focus on the distortions caused by different DIBR view synthesis algorithms; while the MCL-3D [@song2014mcl] and SIAT [@liu2015subjective] datasets explore the influence of traditional 2D distortions of original texture and depth map on the DIBR-synthesized views. These datasets were usually used to evaluate and validate several quality metrics. In the next section, we will introduce the objective approaches for the quality assessment of DIBR-synthesized views.
Objective image/video quality assessment of DIBR-synthesized views
==================================================================
Several methods have been proposed to evaluate the quality of DIBR-synthesized views in the past decade. Based on the amount of reference information, these methods can be divided into 4 categories: Full-reference (FR), Reduced-reference (RR), Side View based Full-reference (SV-FR) and No-reference (NR), as shown in Fig \[Fig: FRNR\]. The FR methods use the original undistorted image/video at the virtual viewpoint as reference to assess the quality of synthesized views, while the RR methods only use some features extracted from the original reference. Especially, the SV-FR methods use the undistorted image/video at the original viewpoint, from which the virtual view is synthesized, as the reference. The NR methods need no access to the original image/video.
{width="0.7\linewidth"}
\(a) [FR]{} metrics
{width="0.75\linewidth"}
\(b) [RR]{} metrics
{width="0.8\linewidth"}
\(c) Side view based FR metrics
{width="0.9\linewidth"}
\(d) [NR]{} metrics
Table \[tab:metrics\] classify the metrics based on their used approaches. Most of them (VSQA, MP-PSNR, MW-PSNR, EM-IQA and CT-IQA) evaluate the quality of synthesized views by considering the contour or gradient degradation between the synthesized and the reference images which is one of the most annoying characteristics of geometric distortions. Meanwhile some metrics (DSQM, 3DSwIM) calculate the quality score by comparing the extracted perceptual features between the synthesized and the reference images. Especially, the APT metric uses a local image description model to reconstruct the synthesize image, and evaluates the quality of the synthesized view based on the reconstruction error. These metrics are introduced as follows.
HF DF C/G JND MSD LID DE DR SE SC IC ML
-- ------------------------------------------------- ---- ----- ----- ----- ----- ---- ---- ---- ---- ---- ---- ---
[Bosc *et al.* 2012 [@bosc2012edge]]{} - - - - - - - - - - -
[[VSQA]{} [@conze2012objective]]{} - - - - - - - - - - -
[3DSwIM [@battisti2015objective]]{} - - - - - - - - - -
[MW-PSNR [@sandic2016free; @sandic2015dibrw]]{} - - - - - - - - - -
[MP-PSNR [@sandic2016multi]]{} - - - - - - - - - -
[CT-IQA [@ling2017quality]]{} - - - - - - - - - - -
[ST-SIAQ [@ling2017icme]]{} - - - - - - - - -
[EM-IQA [@ling2017image]]{} - - - - - - - - -
[PSPTNR [@zhao2010perceptual]]{} - - - - - - - - - - -
[VQA-SIAT [@liu2015subjective]]{} - - - - - - - - - -
[SR-3DVQA [@Zhang2019Sparse]]{} - - - - - - - - -
[3VQM [@solh20113vqm]]{} - - - - - - - - - - -
SDRD [@zhou2016quality] - - - - - - - - - -
[SCDM [@tian2018dibrfr]]{} - - - - - - - - - -
[SC-IQA [@tian2018sc]]{} - - - - - - - - - - -
[CBA [@jung2016critical]]{} - - - - - - - - - - -
[Zhou [@zhou2019quality]]{} - - - - - - - -
[Ling [@ling2019perceptual]]{} - - - - - - - - -
[Wang [@wang2019qualityaccess]]{} - - - - - - - -
[MP-PSNRr [@sandic2016dibr]]{} - - - - - - - - - -
[MW-PSNRr [@sandic2016dibr]]{} - - - - - - - - - -
[RRLP [@jakhetiya2017prediction]]{} - - - - - - - - -
[LOGS [@li2018quality]]{} - - - - - - - - -
[DSQM [@farid2017perceptual]]{} - - - - - - - - - -
[SIQE [@farid2015objective]]{} - - - - - - - - - -
[SIQM [@farid2018evaluating]]{} - - - - - - - - - -
[APT [@gu2017model]]{} - - - - - - - - - - -
[OUT [@jakhetiya2018highly]]{} - - - - - - - - - - -
[MNSS [@gu2019multiscale]]{} - - - - - - - - -
[NR\_MWT [@sandic2019fast]]{} - - - - - - - -
[NIQSV [@shishun2017]]{} - - - - - - - - - -
[NIQSV+ [@tian2018niqsv+]]{} - - - - - - - - - -
[HEVSQP[@shao2017no]]{} - - - - - - - - - -
[CLGM [@yue2018combining]]{} - - - - - - - - -
[GDIC [@wang2019blind]]{} - - - - - - - - -
[Wang [@wang2019blindtip]]{} - - - - - - - -
[SET [@zhou2019no]]{} - - - - - - - -
[FDI [@zhou2018no]]{} - - - - - - - - - -
[CSC-NRM [@ling2018learn]]{} - - - - - - - - - - -
[SIQA-CFP [@wang2019deepicip]]{} - - - - - - - - - -
[GANs-NRM [@ling2019gannrm]]{} - - - - - - - - - -
FR and RR metrics
-----------------
In this subsection, we review 19 well-known FR metrics and 3 RR metrics.
### **[Edge/Contour based FR metrics]{}**
the distortions in DIBR-synthesized views are mostly geometrical and structural distortions, which may degrade the object shape in the synthesized image. It can be measured by the change of object edges. In addition, the sharp edges in the depth map may also induce large dis-occlusions in the synthesized views which may result in dramatic distortions. Thus, a few edge-based methods have been proposed to evaluate the quality of DIBR-synthesized views.
The FR metric proposed by Bosc *et al.* in [@bosc2012edge] indicates the structural degradations by calculating the contour displacement between the synthesized and the reference images. Firstly, a Canny edge detector is used to extract the image contours; then, the contour displacements between the synthesized and reference images are estimated. Based on the contour displacement map, three parameters are computed: the mean ratio of inconsistent displacement vectors per contour pixel, the ratio of inconsistent vectors, the ratio of new contours. The final quality score is obtained as a weighted sum of these three parameters.
In [@ling2017icme], Ling *et al.* proposed a contour-based FR metric ST-SIAQ for the quality assessment of DIBR-synthesized views. Instead of directly using the contour information in [@bosc2012edge], ST-SIAQ uses mid-level contour descriptor called “Sketch Token” [@Lim2013sketch]. The “Sketch Token” stands as a codebook of image contour representation, of which each dimension can be recognized as the possibility which indicates how likely the current patch belongs to one certain category of contour from the codebook. To reduce the shifting effect in the feature comparison stage, the patches in the reference image are firstly matched to the synthesized image. The “Sketch Token” is clustered into 151 categories, which means the “Sketch Token” descriptor has 151 dimensions. A Random Forests decision model associated with a set of low-level features (including oriented gradient channels [@dollar2009integral], color channels, and self-similarity channels [@shechtman2007matching]) are used to obtain the “Sketch Token” descriptor. The geometric distortion strength in the synthesized view is calculated as the Kullback Leibler divergence of “Sketch Token” descriptors between the synthesized and reference images. In [@ling2019prediction], this metric is improved to evaluate the quality of DIBR-synthesized videos by considering the temporal dissimilarity.
Ling *et al.* also proposed another contour-based FR metric EM-IQA in [@ling2017image]. Different from ST-SIAQ metric, EM-IQA uses interest points matching and elastic metric [@mio2007shape], instead of block matching and “Sketch Token” descriptor, to compensate the shifting and evaluate the contour degradation respectively. After interest points matching, a Simple Linear Iterative Clustering (SLIC) is used to extract the contours in the image. SLIC is originally proposed for image segmentation, in the EM-IQA metric, the boundaries of the segmented objects are considered as contours. Then, the elastic metric proposed in [@mio2007shape; @srivastava2010shape] is used to finally measure the degradation between the contours of synthesized and reference images, which provides the quality score of DIBR-synthesized view.
In [@ling2017quality], Ling *et al.* proposed a variable-length context tree based image quality assessment metric CT-IQA, dedicated to quantify the overall structure dissimilarity and dissimilarities in various contour characteristics. Firstly, the contours of the reference and synthesized images are converted to differential chain code (DCC) [@freeman1978application] which represents the direction of object contours. Then, an optimal context tree [@zheng2016context] is learned from the DCC in the reference image. The overall structural dissimilarity is calculated by subtracting the encoding cost of DCC in the synthesized image and reference images. In addition, the overall dissimilarity in contour characteristics is also obtained by measuring the difference of total contour number, total contour start information and total number of symbols between the reference and synthesized image. The final quality score is calculated by combining the overall structure dissimilarity and contour characteristics dissimilarity.
Liu *et al.* proposed a gradient-based FR video quality assessment metric VQA-SIAT [@liu2015subjective] by considering the “Activity” and “Flickering” which is the most annoying temporal distortion in the DIBR-synthesized views. The main contribution of this metric is the two following proposed structures: Quality Assessment Group of Pictures (QA-GoP) and Spatio-Temporal (S-T) tube. The QA-GoP acts as a process unit on a whole video sequence, it contains a group of 2N+1 frames (N frames before and N frames after the central frame). Besides, a block matching method is used to search the corresponding blocks of the central frame blocks in the forward and backward frames. The 2N + 1 blocks along the motion trajectory construct a S-T tube. The distortion of “Activity” is calculated from the difference of the spatial gradient in the (S-T) tube and (QA-GoP) between the synthesized and reference videos. The “Flickering” distortion is measured from the difference of temporal gradient, which is defined below: $$\vec {\bigtriangledown} I_{x,y,i}^{temporal} = I(x,y,i) - I(x', y',i-1),$$ where ($x', y'$) is the coordinate in frame $i-1$ corresponding to ($x,y$) along the motion trajectory in previous frame $i$. The final quality score of DIBR-synthesized view video is obtained by integrating both “Activity” and “Flickering” distortions.
Furthermore, in [@Zhang2019Sparse], Zhang *et al.* proposed a FR metric SR-3DVQA combining the “Activity” measurement module in VQA-SIAT with a sparse representation-based flicker estimation method. In the SR-3DVQA metric, a DIBR-synthesized video is treated as a 3D volume data by stacking the frames sequentially. Then, the volume data is decomposed as a number of spatially neighboring temporal layers i.e. X-T or Y-T planes where X, Y are the spatial coordinate and T is the temporal coordinate. In order to effectively evaluate the flicker distortion in the synthesized video, the gradient in the temporal layers and sharp edges in the associate depth map are extracted as key features for the dictionary learning and sparse representation. The rank-based method in [@li2018quality] is used to pool the flicker score from the temporal layers. The final quality score is calculated by combining the flicker score and “Activity” score in the previous VQA-SIAT [@liu2015subjective].
Jakhetiya *et al.* proposed a free-energy-principle-based IQA metric RRLP for Screen Content and DIBR-synthesized view images based on prediction model and distortion categorization [@jakhetiya2017prediction]. The image quality is measured by calculating the disorder and sharpness similarity between the distorted and reference images. The disorder is obtained from a prediction model. As shown in Eq. \[eq:obf\], an observation-model-based bilateral filter (OBF) [@Jakhetiya2014fast] is firstly used to divide the predicted and disorder parts. $$\hat{X_{d_i}} = \frac{X_{d_i}\lambda + \sum_{k \in N_i}\omega_{k_i}I_{k_i}}{\lambda + \sum_{k \in N_i}\omega_{k_i}}
\label{eq:obf}$$ where $\hat{X_{d_i}}$ represents the predicted part, $I_{k_i}$ and $\omega_{k_i}$ are respectively the pixels and their associated weights in the surrounding $3 \times 3$ window $N_i$ of the $i$th pixel, $\lambda$ is a parameter. The disorder part is computed as the difference between the predicted part and the original image: $$R_{d_i} = |\hat{X_{d_i}} - {X_{d_i}}|$$ Then, the sharpness (edge structures) is calculated by four filters in[@Wu2013reduce]. Finally, the disorder and sharpness similarity between the distorted and reference images are estimated by using the similarity function in SSIM [@Wangssim].
### **[Wavelet transform based FR metrics]{}**
in the previous part, we introduced the metrics that use the edge/contour in luminance domain to evaluate the geometric distortions in DIBR-synthesized views. According to previous research, the wavelet transform representation can not only capture the image edges, but also some other texture unnaturalness. In this part, the wavelet transform based FR metrics will be reviewed.
Battisti *et al.* proposed an FR metric (3DSwIM) for DIBR-synthesized views based on the comparison of statistical features of wavelet sub-bands [@battisti2015objective; @3dswimsoft]. The same as EM-IQA [@ling2017image] and VQA-SIAT [@liu2015subjective], 3DSwIM uses a block matching to ensure the “shifting-resilience”. The distortions in each block of the synthesized view is measured by the Kolmogorov–Smirnov [@lilliefors1967kolmogorov] distance between the histograms of the two matched blocks. In addition, since the Human Vision System (HVS) pays more attention on the human body, a skin detector is used to weight the skin regions in the matched blocks.
Sandić-Stanković *et al.* proposed another multi-scaled decomposition based FR metric MW-PSNR [@sandic2015dibrw; @sandic2016free]. The MW-PSNR uses morphological wavelet filters for decomposition. Then a multi-scale wavelet mean square error (MW-MSE) is calculated as the average MSE of all sub-bands and finally the MW-PSNR is calculated from it.
The wavelet transform based FR metrics can be recognized as a kind of edge/contour based metrics. For example, the higher sub-bands of the wavelet transformed image represent the edge information of the original image. Compared to the pixel level edge/contour used in the previous subsection, the metrics in this subsection use the features in wavelet transformed domain to represent both the image edges and other characteristics.
### **[Morphological operation based FR metrics]{}**
morphological operations are widely used in image processing, especially a couple of erosion and dilation operations can be used to detect the image edges [@maragos1990morphological]. In [@sandic2016multi], Sandić-Stanković *et al.* proposed the MP-PSNR based on multi-scaled pyramid decomposition using morphological filters. The basic erosion and dilation operations used in MP-PSNR are calculated as maximum and minimum in the neighbourhood defined by the structure element, as shown in the following equation: $$D: dilation_{SE}(f)(x) = max_{y \in SE}{f(x-y)}$$ $$E: erosion_{SE}(f)(x) = min_{y \in SE}{f(x+y)}$$ where $f$ is a gray-scale image and $SE$ is binary structure element. Then, they use the Mean Square Error (MSE) between the reference and synthesized images in all pyramids’ sub-bands to quantify the distortion. As shown in Fig \[fig:mppsnr\], during the decomposition, the dilation is used as expand operation and the erosion is used as reduce operation, the detail image of each scale is calculated as the difference between the original and processed (erosion and dilation) images. Finally, the overall quality is calculated by averaging the MSE of detail images in all the sub-bands and expressing it as PSNR.
![Decomposition scheme of MP-PSNR. $S_j$ represents the image at scale $j$ ($j \in [1,5]$), $D_j$ represent the detail image at scale $j$ [@sandic2016multi].[]{data-label="fig:mppsnr"}](MPPSNR_cropped.pdf){width="1\linewidth"}
In [@sandic2016dibr], Sandić-Stanković *et al.* also proposed the reduced version of MP-PSNR, and MW-PSNR. Only detail images from higher decomposition scales are taken into account to measure the difference between the synthesized image and the reference image. The reduced version achieved significant improvement over the original FR metrics with lower computational complexity.
### **[Depth estimation based FR metrics]{}**
Solh *et al.* proposed a full reference metric 3VQM [@solh20113vqm] to evaluate synthesized view distortions by deriving an “ideal” depth map from the virtual synthesized view and the reference view at a different viewpoint. The “ideal” depth is the depth map that would generate the distortion-free image given the same reference image and DIBR parameters. Three distortion measurements, spatial outliers, temporal outliers and temporal inconsistency are calculated from the difference between the “ideal” depth map and the distorted depth map: $$SO = STD(\triangle Z)$$ $$TO = STD(\triangle Z_{t+1} +\triangle Z_{t})$$ $$TI = STD(Z_{t+1} + Z_{t})$$ where $SO$, $TO$ and $TI$ denote the spatial outliers, temporal outliers and temporal inconsistencies respectively, $STD$ represents the standard deviation. $\triangle Z$ is the difference between the “ideal” and the distorted depth maps and $t$ is the frame number. These three measurements are then integrated into a final quality score.
Since the calculation of the “ideal” depth map is based on the assumption that the horizontal shift of the synthesized view and the original view is small, this metric would not work well when the baseline distance increases.
### **[Dis-occlusion region based FR metrics]{}**
since the DIBR view synthesis distortions mainly occur in the dis-occlusion regions, some of the FR metrics improve the performance of 2D FR metrics by using dis-occlusion maps[@zhou2016quality; @tian2018dibrfr] instead of using weighting maps.
The SDRD metric proposed by Zhou in [@zhou2016quality] detects the dis-occlusion regions by simply comparing the absolute difference between the synthesized and reference images. Before that, a self-adaptive scale transform model is used to eliminate the effect of view distance, and a SIFT flow-based warping is adopted to compensate the global shift in the synthesized view image. The final quality score is obtained by weighting the dis-occlusion regions with their size since the distortions with bigger size are more annoying to human vision system.
Tian *et al.* proposed a full-reference quality assessment model (SCDM) for 3D synthesized views by considering global shift compensation and dis-occlusion regions [@tian2018dibrfr]. This model can be used on any pixel-based FR metrics. SCDM firstly compensates the shift by using a SURF + RANSAC approach instead of the SIFT flow used in SDRD. Then, the dis-occlusion regions are directly extracted from the depth map. It is more precise and uses more resources compared to SDRD. The final quality score is obtained as a weighted PSNR or weighted SSIM. It is reported to improve the performance of PSNR and SSIM by 36.85% and 13.33% in terms of Pearson Linear Correlation Coefficients (PLCC).
Since the distortions in the DIBR-synthesized views are not restricted in the dis-occlusion regions only, they may occur around these regions as well. In [@wang2019qualityaccess], Wang *et al.* proposed a critical region based metric by dilating the dis-occlusion region with a morphological operator. Similar to SDRD, the dis-occlusion region map is extracted by a SIFT-flow based approach. Then a Discrete Cosine Transform (DCT) decomposition method is used to partition and classify the critical regions into edge blocks, texture blocks and smooth blocks. Based on the perceptual properties of these three types of blocks, their distortions are measured differently. The edge and texture blocks contain more complex edges or texture information, the blur distortions in these regions would be much more annoying than that in the smooth regions. On the other hand, the smooth regions are sensitive to color degradations. Thus, the texture similarity and color contrast similarity between the synthesized and reference images are calculated to measure the local distortions in the edge, texture and smooth blocks respectively. Finally, a global sharpness detection is combined with the local distortion measurement to obtain the overall quality score.
### **[2D related FR metrics]{}**
the main reason of the ineffectiveness of 2D quality assessment metrics on DIBR-synthesized views can be analysed as follows. Firstly, there exists large object shift in the synthesized views and this kind shift effect can be easily penalized by 2D metrics even though the HVS is not sensitive to the global shift in the image. The second reason is the distribution of distortions. The distortions in traditional 2D images often scatter over the whole image while the DIBR view synthesis distortions are mostly local, especially in the dis-occluded regions. The 2D related metrics are based on the traditional 2D FR metrics, such as PSNR, SSIM etc. They try to improve the performance of 2D metrics by considering HVS and the characteristics of DIBR view synthesis distortions.
The VSQA metric proposed by Conze *et al.* in [@conze2012objective] tries to improve the performance of SSIM [@Wangssim] by taking advantage of known characteristics of the human visual system (HVS). It aims to handle areas where disparity estimation may fail, such as thin objects, object borders, transparency etc. by applying three weighting maps on the SSIM distortion map. The main purpose of these three weighting maps is to characterize the image complexity in terms of textures, diversity of gradient orientations and presence of high contrast since the HVS is more sensitive to the distortions in such areas. For example, the distortions in an untextured area are much more annoying than the ones located in a high texture complexity area. It is reported that this method approaches a gain of 17.8% over SSIM in correlation with subjective measurements.
Zhao *et al.* proposed the PSPTNR metric to measure the perceptual temporal noise of the synthesized sequence [@zhao2010perceptual]. The temporal noise is defined as the the difference between inter-frame change in the processed sequence and that in the reference sequence: $$TN_{i,n} = ((P_{i,n} - P_{i, n-1}) - (R_{i,n} - R_{i,n-1})^2 ,$$ where $TN$ indicates the temporal noise, $P$ and $R$ represent the distorted and reference sequence respectively. In order to better predict the perceptual quality of synthesized videos, temporal noise is filtered by a Just Noticeable Distortion (JND) model [@Chou1996] and a motion mask [@westerink1995perceived], since the human can observe noise only beyond certain level and motion may decrease the texture sharpness in the video.
The shift compensation methods included in SDRD and SCDM only consider the global shift, but according to the recent research[@moorthy2009visual], Human Visual System (HVS) is more sensitive to local artefacts compared to the global object shift. In [@tian2018sc], Tian *et al.* proposed a shift-compensation based image quality assessment metric (SC-IQA) for DIBR-synthesized views. The same as SCDM, a SURF + RANSAC approach is used to roughly compensate the global shift. In addition, a multi-resolution block matching method is proposed to precisely compensate the global shift and penalize the local shift at the same time. A saliency map [@jiang2013salient] is also considered to weight the distortion map of the synthesized view. Furthermore, only the blocks with the worst quality are used to calculate the final quality score since HVS tends to perceive poor regions in an image with more severity than the good ones [@moorthy2009visual; @liu2015subjective]. SC-IQA achieves the performance of SCDM without access to the depth map.
The metrics introduced above consider only the view synthesis and compression artefacts which occur on applications that show the synthesized views on a 2D display, the binocular effect in the synthesized stereoscopic images is not taken into consideration. In [@jung2016critical], Jung *et al.* proposed a SSIM-based FR metric to measure the critical binocular asymmetry (CBA) in the synthesized stereo images. Firstly, the disparity inconsistency between the two different views is generated to detect the critical areas in terms of Left-Right image mismatches. Then, only the SSIM value on the critical areas of each view are computed to measure the asymmetry in the corresponding view image. The final binocular asymmetry score is obtained by averaging the asymmetry score in the left and right views.
Side view based FR metrics
--------------------------
The major limitation of the FR metrics is that they always need the reference view which may be unavailable in some circumstances (eg. FVV). In other words, there is no ground truth for a full comparison with the distorted synthesized view. In this part, four side view based FR metrics will be reviewed. This kind of metrics use the real image/video at the original viewpoint, from which the virtual view is synthesized, as reference to evaluate the quality of DIBR-synthesized virtual views. These metrics are named as side view based FR metrics in this paper.
Li *et al.* proposed a side view based FR metric for DIBR-synthesized views by measuring local geometric distortions in dis-occluded regions and global sharpness (LOGS) [@li2018quality]. This metric consists of three parts. Firstly, the dis-occlusion regions are detected by using SIFT-flow based warping. These dis-occluded regions are extracted from the absolute difference map between the synthesized view $I_{syn}$ and the warped reference view $I_{ref}^{w}$ followed by an additional threshold. Then, the distortion size and strength in the local dis-occlusion regions are combined to obtain the overall local geometric distortion. The distortion size is simply measured by the number of pixels in the dis-occluded regions and the distortion strength is defined as the mean value of the dis-occluded regions in the whole difference map $M$. The next part is to measure the global sharpness by using a reblurring-based method. The synthesized image is firstly blurred by a Gaussian smoothing filter. Both the synthesized image and its reblurred version are divided into blocks. The sharpness of each block is calculated by its textural complexity, which is represented by its variance $\sigma^2$. Then, the overall sharpness score is computed by averaging the textural distance of all blocks. Finally, the local geometric distortion and the global sharpness are pooled to generate the final quality score.
Farid *et al.* proposed a side view based FR metric (DSQM) for the DIBR-synthesized view in [@farid2017perceptual]. A block matching is firstly used to estimate the shift between the reference and synthesized image. Then the difference of Phase congruency (PC) in these two matched blocks is used to measure the quality of the block in the synthesized image, which is defined as follows: $$PC(x) = \max_{\bar{\phi(x)} \in [0,2\pi]}{\frac{\sum_n{A_n cos(\phi_n(x)} - \bar{\phi(x)})}{\sum_n{A_n}}}$$ where $A_n$ and $\phi_n(x)$ represent the amplitude and the local phase of the $n$-th Fourier component at position $x$ respectively. The implementation of phase congruency is based on an logarithmic Gabor wavelet method proposed in [@kovesi1999image]. The quality score of each block is calculated as the absolute difference between the mean values of the phase congruency maps of the matched blocks in the synthesized and reference image: $$Q_i = |\mu(PC_{si} - PC_{ri})|$$ where $\mu()$ represents the mean value of the corresponding phase congruency map, the $PC_{si}$ and $PC_{ri}$ indicate the PC map of the matched blocks in the synthesized and reference image. The final image quality is obtained by averaging the quality score of all the blocks.
Farid *et al.* proposed a cyclopean eye theory [@julesz1972cyclopean] and divisive normalization (DN) transform [@teo1994perceptual] based Synthesized Image Quality Evaluator (SIQE) in [@farid2015objective]. The DIBR-synthesized view image associated with the left and right side views are firstly transformed by DN. Then, the statistical characteristics of the cyclopean image are estimated from the DN representations of the left and right side views while the statistical characteristics of the synthesized image are obtained directly from itself. The similarity (Bhattacharyya coefficient [@patra2015new]) between the distribution of the cyclopean and the synthesized image’s DN representations is computed to measure the quality score of the synthesized image.
The SIQE metric only considers the texture information, in [@farid2018evaluating], Farid *et al.* proposed an extended version of SIQM by considering both the texture and depth information. The depth distortion estimation is based on the fact that the edge regions in a depth image are more sensitive to noise than the flat homogeneous regions since the distorted edge in the depth map may cause very annoying structural distortions in the synthesized image. Firstly, the pixels in the depth map with a high gradient value are extracted as noise sensitive pixels (NSP). Then, for each NSP, a local histogram from the distorted depth map is constructed and analysed to estimate the distortion in the depth image. The overall depth distortions are calculated by averaging the distortions in the left and right depth image. The final quality of the synthesized view is pooled from the texture and depth distortions.
NR metrics
----------
In this part, we will review the NR metrics which do not need ground truth images/videos to evaluate the quality of DIBR-synthesized views.
### **[Local image description based NR metrics]{}**
due to the distorted depth map and imperfect rendering method, there exists a large number of structural and geometric distortions in the DIBR-synthesized views. As introduced in the RRLP metric [@jakhetiya2017prediction], the structural distortions may result in local disorder in the image. Similarly, several local image description based NR metrics have been proposed to evaluate the structural distortions by measuring the local inconsistency via different models.
Gu *et al.* proposed an auto-regression (AR) based model (APT) to capture the geometric distortions in the DIBR-synthesized views. For each pixel, a local AR model (3$\times$3) is first used to construct a relationship between this pixel and its neighbouring pixels. $$x_i = \Omega(x_i)s + d_i$$ where $\Omega(x_i)$ denotes a vector which is composed of the neighbouring pixels of $x_i$ in the (3$\times$3) patch, $s$ is a vector of AR parameters and $d_i$ represents the error difference between the current pixel value and its corresponding AR prediction. The AR parameters are solved on the assumption that the 7$\times$7 local patch, which consists of the current pixel and its 48 adjacent pixels, shares the same AR model. The error difference map between the synthesized and the reconstructed images is obtained as the distortion map. Then, a Gaussian filter and a saliency map [@gu2015visual] associated with a maximum pooling are used to obtain the final image quality score. Due to its computational complexity, this method owns a high computing cost.
Different from the APT metric, the OUT (outliers) metric [@jakhetiya2018highly] proposed by Jakhetiya *et al.* uses a median filter to calculate the difference map. Then, two thresholds are used to extract the structural and geometric distortion regions. The quality score is finally obtained from the standard deviation of the structural and geometric distortion regions.
These local image description based metrics can only detect thin distortions or local noise, they do not work well on the large size distortions.
### **[Morphological operation based NR metrics]{}**
the morphological operations show their effectiveness in the FR metric MP-PSNR [@sandic2016multi]. In [@shishun2017; @tian2018niqsv], Tian *et al.* proposed two metrics NIQSV and NIQSV+ to detect the local thin structural distortions through morphological operations. These two metrics assume that the “perfect” image consists of flat areas and sharp edges, so such images are insensitive to the morphological operations while the local thin structural distortions can be easily detected by these morphological operations. The NIQSV metric firstly uses an opening operation to detect the thin distortions and followed by a closing operation with larger Structural Element (SE) to file the black holes. The NIQSV+ extend the NIQSV by proposing two additional measurements: black hole detection and stretching detection. The black hole distortion is estimated by counting the black hole pixels proportion in the image while the stretching distortion is evaluated by calculating the gradient decrease of the stretching region and its adjacent non-stretching region.
Due to the limitation of the assumption and the SE size, these two metrics do not work well on the distortions in complex texture and the distortions with large size.
### **[Sharpness detection based NR metric]{}**
sharpness detection has been widely used in 2D image quality assessment [@tang2016blind; @zhang2017reduced; @gu2015nosharpness] and also in the side view based FR metric LOGS [@li2018quality]. In this part, we will introduce its usage in NR metrics. Sharpness is one of the most important measurements in NR image quality assessment [@Gu2015noref; @Min2018blind; @Ferzli2009]. The DIBR view synthesis may introduce multiple distortions such as blur, geometric distortions around the object edges, which may significantly result in the degradation of sharpness. Nonlinear morphological wavelet decomposition can extract high-pass image content while preserving the unblurred geometric structures [@sandic2016multi; @sandic2015dibrw]. In the transform domain, geometry distorted areas introduced by DIBR-synthesis are characterized by coefficients of higher value compared to the coefficients of smooth, edge and textural areas. In [@sandic2019fast], Sandić-Stanković *et al.* proposed a wavelet-based NR metric (NR\_MWT) for the DIBR-synthesized view videos. The sharpness is measured by quantifying the high frequency components in the image, which are represented by the high-high wavelet sub-band. The final quality is obtained from the sub-band coefficients whose value are higher than the threshold. Similar to MW-PSNR and MP-PSNR [@sandic2016multi; @sandic2015dibrw], the NR\_MWT also achieved a very low computational complexity.
Differently, in CLGM [@yue2018combining], the sharpness is measured as the distance of standard deviations between the synthesized image and its down-sampled version. Besides, two additional distortions, dis-occluded regions and stretching, are also taken into consideration in CLGM. The dis-occluded regions are detected through an analysis of local image similarity. Similar to NIQSV+ [@tian2018niqsv+], the stretching distortion is estimated by computing the similarity between the stretching region and its adjacent non-stretching region.
In [@wang2019blind], Wang *et al.* also proposed a NR metric (GDIC) to measure the geometric distortions and image complexity. Firstly, different from the wavelet transform based metrics introduced above, this GDIC metric uses the edge map of wavelet sub-bands to obtain the shape of geometric distortions. Then, the geometric distortion is measured by edge similarity between the wavelet low-level and high-level sub-bands [@cohen1992biorthogonal]. Besides, the image complexity is also an important factor in human visual perception. In order to evaluate the image complexity of the DIBR-synthesized images, hybrid filter [@gu2017no; @gu2016no; @gu2017evaluating], which combines the Autoregressive (AR) and bilateral (BL), is used. The final image quality score is computed by normalizing the geometric distortion with image complexity. Furthermore, in [@wang2019blindtip], this metric is extended to achieve higher performance by adding a log-energy based sharpness detection module.
### **[Flicker region based video NR metrics]{}**
in DIBR-synthesized videos, temporal flicker is one of the most annoying distortions. Extracting the flicker regions may help to evaluate the quality of DIBR-synthesized videos.
In [@kim2016measurement], Kim *et al.* also proposed a NR metric (CTI) to measure the temporal inconsistency and flicker regions in the DIBR-synthesized video. First, the flicker regions are detected from the difference between motion-compensated consecutive frames. Then, the structural similarity between consecutive frames are calculated on the flicker regions to measure the structural distortions in each frame. At the same time, the number of pixels in the flicker regions is used to weight the distortion of each frame. The final quality score is obtained as the weighted sum of the quality scores of all the frames in the DIBR-synthesized video.
In [@zhou2018no], Zhou *et al.* proposed a NR metric FDI to measure the temporal flickering distortion in the DIBR-synthesized videos. Firstly, the gradient variations between each frame are used to extract the potential flickering regions. Followed by a refinement to precisely obtain the flickering regions through calculating the correlation between the candidate flickering regions and their neighbours. Then, the flickering distortion is estimated in SVD domain from the difference between the singular vectors of the flickering block and their associated block in the previous frame. The final video quality is computed as the average quality of all the frames.
### **[Natural Scene Statistics based NR metrics]{}**
Natural Scene Statistics (NSS) based approaches, which assume that the natural images contain certain statistics and these statistics may be changed by different distortions, have achieved great success in the quality assessment of traditional 2D images [@saad2012blind; @saad2011dct; @moorthy2011blind; @LIU2014494]. Due to the big difference between the DIBR view synthesis distortions and the traditional 2D ones, these NSS based metrics do not work well on the quality assessment of DIBR-synthesized views. Recently, several efforts have been made to fix this gap.
As introduced in the previous Edge/Contour based FR metrics part, the edge image is significantly degraded by structural and geometric distortions in DIBR-synthesized images, and the edge based FR metrics have shown their superiority. With this view, Zhou *et al.* proposed a NR metric (SET) for DIBR-synthesized images via edge statistics and texture naturalness based on Difference-of-Gaussian (DoG) in [@zhou2019no]. The orientation selective statistics (similar to the metric in [@moorthy2011blind]) are extracted from different scale DoG images while the texture naturalness features are obtained based on the Gray level Gradient Co-occurrence Matrix (GGCM) [@Li2015blind] which represents the joint distribution relation of pixel gray level and edge gradient. A Random Forest (RF) regression model is finally trained based on these two groups of features to predict the quality of DIBR-synthesized images.
Gu *et al.* proposed a self-similarity and main structure consistency based Multiscale Natural Scene Statistics (MNSS) in [@gu2019multiscale]. The multiscale analysis on the DIBR-synthesized image and its associated reference image indicates that the distance (SSIM value [@Wangssim]) between the synthesized and the reference image decreases significantly when the scale reduces. It is assumed that the synthesized image at a higher scale holds a better quality, which means the higher scale images can be approximately used as reference. Thus, the similarity between the lower scale image (first scale is used in this metric) and the higher scale images (self similarity) are used to measure the quality of DIBR-synthesized image. Besides, in the main structure NSS model, the authors use 300 natural images from the Berkeley segmentation dataset [@Martin2001adatabase] to obtain the general statistical regularity of main structure in natural images. The similarity between the main structure map of the synthesized image and the obtained prior NSS vector is calculated to evaluate the structure degradation of the DIBR-synthesized image. Finally, the statistical regularity of main structure and the structure degradation are combined to get the overall quality score.
Shao *et al.* propose a NR metric (HEVSQP) for DIBR-synthesized videos based on color-depth interactions in [@shao2017no]. Firstly, the video sequence is divided into Group of Frames (GoF). Through an analysis of color-depth interactions, more than 90 features from both texture and depth videos, including gradient magnitude, asymmetric generalized Gaussian distribution (AGGD) [@saad2011dct], local binary pattern (LBP), are extracted. Then, a principal component analysis (PCA) is applied to reduce the feature dimension. Then, two dictionaries, color dictionary and depth dictionary, are learned to establish the relationship between the features and video quality. The final quality score is pooled from the color and depth quality.
In [@ling2018learn], Ling *et al.* proposed a NR learning based metric for DIBR-synthesized views, which focuses on the non-uniform distortions. Firstly, a set of convolutional kernels are learned by using the improved fast convolutional sparse coding (CSC) algorithms. Then, the convolutional sparse coding (CSC) based features of the DIBR-synthesized images are extracted, from which the final quality score is obtained via support vector regression (SVR).
Although the NSS models have made great progress for the NR IQA, the hand-craft features may not be sufficient to represent complex image textures and artefacts, there still exists a large gap between objective quality measurement and human perception [@yang2019survey].
### **[Deep feature based NR metrics]{}**
the deep learning techniques, especially the Convolutional Neural Networks (CNN), have shown their great advantages in various computer vision tasks. They make it possible to directly learn the representative features from image [@kim2017deep]. Unfortunately, owing to the limitation of size of DIBR-synthesized view datasets, there is not enough data to train the deep model straightforwardly. However, it is shown in the recent published literature that the deep neural network models trained on large-scale datasets, eg. ImageNet [@ILSVRC15], can be used to extract effective representative features of human perception. In [@wang2019deepicip], Wang *et al.* proposed a NR metric SIQA-CFP which uses the ResNet-50 [@he2016resnet] model pre-trained on ImageNet to extract multi-level features of DIBR-synthesized images. Then, a contextual multi-level feature pooling strategy is designed to encode the high-level and low-level features, and finally to get the quality scores.
As introduced in Section , various distortions may be introduced during the dis-occlusion region filling stage. Meanwhile, in current literature, several Generative Adversarial Networks (GAN) [@NIPS2014_5423] based models have been proposed for image in-painting. As the generator is trained to in-paint the missing part, the discriminator is supposed to have the capability to capture the perceptual information which reflects the in-painted image quality. Based on this assumption, Ling *et al.* proposed a GAN based NR metric (GANs-NRM) for DIBR-synthesized images. In GANs-NRM, a generative adversial network for image in-painting is firstly trained on two large-scale datasets (PASCAL [@Everingham2015] and Places [@zhou2018places]). Then, the features extracted from the pre-trained discriminator are used to learn a Bag-of-Distortion-Word (BDW) codebook. A Support Vector Regression (SVR) is trained on the encoded information of each image to predict the final quality of DIBR-synthesized images. Instead of simply using the general models trained for other tasks, eg. object detection, this metric is more targeted, and it also proposes a new way to obtain the semantic features for image quality assessment.
[Summary]{}
-----------
In this section, 19 FR, 3 RR, 4 SV-FR and 15 NR DIBR quality metrics have been reviewed and categorized based on their used approaches and on the amount of reference information used. As shown in Table \[tab:metrics\], most of the metrics consist of multiple parts, it is thus difficult to classify them into a single specific category thoroughly, we just classify them into the most related one instead. Besides, there are also some other ways to do the classification. For example, if we focus on the image structural representation used in these metrics, they can be classified into low-level [@liu2015subjective]), mid-level [@ling2017icme; @ling2017image] and high-level [@ling2018learn; @wang2019deepicip; @ling2019gannrm] metrics. As introduced in [@manassi2013crowding], the low-level representations indicate the pixel level edges or contours; the mid-level representations mean the shapes and texture information; the high-level representations refer to the complex features eg. objects, unnatural structures. Besides, there are also some hierarchical metrics which combine the above features, such as the LMS metric proposed in [@zhou2019quality] which uses both low-level and mid-level features [@ling2017icme] and the metric in [@ling2019perceptual] which integrates the features on each level.
-- ------ -------- -------- -------- -------- -------- -------- ------------ ------------ ------------ ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------- --
PLCC RMSE SROCC PLCC RMSE SROCC PLCC RMSE SROCC PLCC RMSE SROCC
PSNR 0.4557 0.5927 0.4417 0.6012 0.1985 0.5356 [0.7852]{} [1.6112]{} [0.7915]{} **[0.6311]{} & **[19.1227]{} & **[0.6668]{}\
& SSIM & 0.4348 & 0.5996 & 0.4004 & 0.4016 & 0.2275 & 0.2395 & 0.7331 & 1.7693 & 0.7470 & 0.3786 & 22.8172 & 0.3742\
&BIQI & 0.5150 & 0.5708 & 0.3248 & 0.4427 & 0.2223 & 0.4321& 0.3347 & 2.4516 & 0.3696 & 0.5686 & 20.2791 & 5754\
& BLIINDS2 & 0.5709 & 0.5467 & 0.4702 & 0.2020 & 0.2428 & 0.1458 & 0.6338 & 2.0124 & 0.5893 & 0.3508 & 23.0855 & 0.2569\
& Bosc & 0.5841 & 0.5408 & 0.4903 & — & — & — & 0.4536 & 2.2980 & 0.4330 & — & — & —\
&3DSwIM & 0.6864 & 0.4842 & 0.6125 & — & — & — & 0.6519 & 1.9729 & 0.5683 & — & — & —\
&VSQA & 0.6122 & 0.5265 & 0.6032 & 0.5576 & 0.2062 & [0.4719]{} & 0.5078 & 2.9175 & 0.5120 & —& — & —\
&ST-SIAQ & 0.6914 & 0.4812 & 0.6746 & 0.3345 & 0.2336 & 0.4232 & 0.7133 & 1.8233 & 0.7034 & — & — & —\
&EM-IQA & 0.7430 & 0.4455 & 0.6282 & 0.5627 & 0.2020 & 0.5670 &— & — & — & — & — & —\
&MP-PSNR & 0.6729 & 0.4925 & 0.6272 & [0.5753]{} & [0.2032]{} & [0.5507]{} & 0.7831 & 1.6179 & 0.7899 & [0.5947]{} & [19.8182]{} & [0.5707]{}\
&MW-PSNR & 0.6200 & 0.5224 & 0.5739 & 0.5301 & 0.2106 & [0.4845]{} & 0.7654 & 1.6743 & 0.7721 & 0.5373 & 20.7910 & 0.5051\
&SCDM & 0.8242 & 0.3771 & 0.7889 & **[0.6685]{} & **[0.1844]{} & **[0.5903]{} & 0.7166 & 1.8141 & 0.7197 & — & — & —\
&[[SC-IQA]{}]{} & **[0.8496]{} & **[0.3511]{} & [0.7640]{} & **[0.6856]{} & **[0.1805]{} & **[0.6423]{} & **[0.8194]{} & **[1.4913]{} & **[0.8247]{} & 0.4326 & 22.2256 & 0.3135\
&[Wang [@wang2019qualityaccess]]{} & **[0.8512]{} & **[0.3146]{} & **[0.8346]{} & [0.6118]{} & [0.1961]{} & [0.6136]{} & **[0.7910]{} & **[1.5917]{} & **[0.7929]{} & — & — & —\
&CBA & — & — & — & — & — & — & — & — & — & **[0.826]{} & **[8.181]{} & **[0.829]{}\
&MP-PSNRr & 0.6954 & 0.4784 & 0.6606 & [0.6061]{} & [0.1976]{} & [0.5873]{} & 0.7740 & 1.6474 & 0.7802 & 0.5384 & 20.7733 & 0.5454\
& MW-PSNRr & 0.6625 & 0.4987 & 0.6232 & 0.5403 & 0.2090 & [0.4946]{} & 0.7579 & 1.7012 & 0.7665 & 0.5304 & 20.8993 & 0.5138\
&[SIQE]{}& 0.7650 & 0.5382 & 0.4492 & 0.3144 & 0.2353 & 0.3418 & 0.6734 & 1.9233 & 0.6976 & — & — & —\
&[LOGS]{}& **[0.8256]{} & **[0.3601]{} & [0.7812]{} & **[0.6687]{} & **[0.1845]{} & **[0.6683]{} & 0.7614 & 1.6873 & 0.7579 & **[0.6442]{} & **[18.8553]{} & **[0.6385]{}\
& [DSQM]{}& 0.7430 & 0.4455 & 0.7067 & 0.2977 & 0.2367 & 0.2369 & 0.6995 & 1.8593 & 0.6980 & — & — & —\
&[APT]{}& 0.7307 & 0.4546 & 0.7157 & 0.4225 & 0.2252 & 0.4187 & 0.6433 & 1.9870 & 0.6200 & 0.5156 & 21.1239 & 0.4754\
&[OUT]{}& 0.7243 & 0.4591 & 0.7010 & 0.2007 & 0.2429 & 0.1924 & 0.4208 & 2.3601 & 0.3171 & 0.2525 & 23.8530 & 0.2409\
&[MNSS]{}& 0.7700 & 0.4120 & 0.7850 & 0.3387 & 0.2333 & 0.2281 & 0.3766 & 2.4101 & 0.3531 & 0.3834 & 22.7681 & 0.2282\
&[NR\_MWT]{}& 0.7343 & 0.4520 & 0.5169 & 0.4769 & 0.2179 & 0.4567 & 0.1373 & 2.5771 & 0.0110 & 0.4848 & 21.5614 & 0.4558\
&[NIQSV]{}& 0.6346 & 0.5146 & 0.6167 & 0.1759 & 0.2446& 0.1473 & 0.6460 & 1.9820 & 0.5792 & 0.4113 & 22.4706 & 0.2717\
&[NIQSV+]{}& 0.7114 & 0.4679 & 0.6668 & 0.2095 & 0.2429 & 0.2190 & 0.6138 & 2.0375 & 0.6213 & 0.2823 & 23.6491 & 0.3823\
&[SET]{}& **[0.8586]{} & **[0.3015]{} & **[0.8109]{} & — & — & — & **[0.9117]{} & **[1.0631]{} & **[0.9108]{} & — & — & —\
&[GANs-NRM]{}& [0.826]{} & [0.386]{} & **[0.807]{} & 0.646 & 0.198 & 0.571 & — & — & — & — & — & —\
****************************************************************************
-- ------ -------- -------- -------- -------- -------- -------- ------------ ------------ ------------ ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------- --
: Due to the unavailability of source code or reference resources eg. depth map and side view reference image, we just use the reported results in their corresponding publications instead, their associated results on other datasets are marked by the symbol “—” in the table.
Experimental results and discussions
====================================
In this section, the performance of different objective quality assessment metrics are presented and analysed. Besides, some potential challenges and possible directions for future work will be discussed.
Performance evaluation methodologies
------------------------------------
The subjective test results can be recognized as the ground truth visual quality since the human observer is the ultimate receiver of image/video content. The accuracy of an objective quality metric can be evaluated based on its consistencies with the subjective quality scores. In this part, we will introduce the Video Quality Expert Group (VQEG) [@vqeg] recommended correlation based methods and the recently proposed Krasula’ model [@krasula2016accuracy] in detail.
### **[Correlation coefficients based methods]{}**
the reliability of objective metrics can be evaluated through their correlation with subjective test scores. Three widely used criteria, Pearson Linear Correlation Coefficients (PLCC) and Root-Mean-Square-Error (RMSE) and Spearman Rank-Order Correlation Coefficients (SROCC), are recommended by VQEG to evaluate the prediction accuracy, prediction monotonicity and prediction consistency of the objective metrics respectively, which are defined as follows: $$PLCC(X,Y) = \frac{\sum_{i=1}^n(X_i - \bar{X})(Y_i - \bar{Y})}{\sqrt{\sum_{i=1}^n(X_i - \bar{Y})^2}\sqrt{\sum_{i=1}^n(Y_i - \bar{Y})^2}}$$ $$RMSE(X,Y) = \sqrt{\frac{1}{m}\sum_{i=1}^m(X_i - Y_i)^2}$$ $$SROCC(X,Y) = 1- \frac{6 \sum{d_i^2}}{n(n^2 - 1)}$$ where $d_i$ indicates the difference of ranking of $X$ and $Y$. Higher PLCC and SROCC value indicate higher accuracy and better monotonicity respectively. On the contrary, a higher RMSE value refers to a lower prediction accuracy.
![Example relationship between DMOS and objective quality scores. This figure is from [@tian2019image][]{data-label="Fig: fitting"}](fitting){width="1\linewidth"}
\(a) Before regression
![Example relationship between DMOS and objective quality scores. This figure is from [@tian2019image][]{data-label="Fig: fitting"}](after_fitting){width="1\linewidth"}
\(b) After regression
Before computing these three criteria, the objective scores are recommended by VQEG to be mapped to the predicted subjective score $DMOS_p$ to remove the nonlinearties due to the subjective rating processing and to facilitate comparison of the metrics in a common analysis space [@vqeg]. The nonlinear function for regression mapping is shown as follows: $$DMOS_p = \beta_1 (0.5-\frac{1}{(1+e^{(\beta_2 (s-\beta_3))})})+\beta_4 s+\beta_5
\label{eq:regression}$$ where $s$ is the score obtained by the objective metric and $\beta_1, \beta_2, \beta_3, \beta_4, \beta_5$ are the parameters of these regression functions. They are obtained through regression to minimize the difference between $DMOS_p$ and $DMOS$. As shown in Fig. \[Fig: fitting\], the nonlinearity has been removed after the regression.
![Krasula’s model for performance evaluation of objective quality metrics [@krasula2016accuracy].[]{data-label="fig:Krasula"}](Lucas.pdf){width="1\linewidth"}
### **[Analysis of Krasula’s model]{}**
the above methods compare the performance of each metric by calculating their correlations with the subjective results. However they only consider the mean value of subjective scores, the uncertainty of the subjective scores are ignored. In addition, the quality scores need to be regressed by a regression function cf. Eq. \[eq:regression\], that is not the way they are exactly used in real scenarios. Thus, we further conduct a statistical test proposed by Krasula *et al.* in [@krasula2016accuracy] which does not suffer from the drawbacks of the above methods. The performances of objective metrics are evaluated by their classification abilities.
As shown in Fig. \[fig:Krasula\], firstly, the tested image pairs in the dataset are divided into two groups: different and similar according to their subjective scores. The cumulative distribution function (cdf) of the normal distribution is used to calculate the probability of image pairs. Then, we consider the pairs with higher than the selected significance level 0.95 to be significantly different. The others will be recognized as similar.
There are two performance analysis. The first performance analysis is conducted by by evaluating how well the objective metric succeeds to distinguish significantly different image pairs from unsignificantly different video pairs, in a consistent way with subjective evaluation of significant difference. In the case of the two videos in the pair are significantly different according to the subjective results. The second analysis determines whether the objective metric can correctly identify the image of higher quality in the pair.
Compared to simply calculating the correlation coefficients, this model considers not only the mean value of subjective scores, but also their uncertainties. Besides, since no regression is used, this model less depends on the quality ranges of different datasets. Another advantage of Krasula’s model is that it can easily combine the data from multiple datasets and evaluate a comprehensive performance on multiple datasets instead of simply averaging the results on different datasets.
Performance on DIBR image datasets
----------------------------------
### **[Results of PLCC, RMSE and SROCC]{}**
the obtained PLCC, RMSE and SROCC values of the objective image quality assessment metrics on the DIBR-synthesized image datasets are given in Table \[tab:plcc\_ivc\], in which four 2D metrics [@moorthy2009modular; @saad2012blind; @Wangssim] and 24 DIBR metrics are tested. The best three performances among the blind IQA methods are shown in bold. We can easily observe that the DIBR-synthesized view dedicated metrics significantly outperform the traditional 2D metrics on the IVC and IETR image datasets which focus on the DIBR view synthesis distortions. In other words, the metrics initially designed for traditional 2D image distortions can not well evaluate the DIBR view synthesis distortions.
The shift compensation based FR and SV-FR metrics obtain great improvement compared to the original 2D FR metrics, eg. the SC-IQA compared to PSNR. One main reason is that the global object shift existing in the DIBR-synthesized images may not be perceived by human observers but can be easily detected by the original 2D pixel-based FR metrics. So, this shift distortions are often overestimated by the 2D pixel-based FR metrics.
If we focus on the wavelet transform-based metrics (NR\_MWT and MW-PSNR), the NR metric (NR\_MWT etc.) perform better than the FR metric (MW-PSNR) on the IVC dataset. It is surprising that the FR metric performs even worse than the NR metric since these metrics use similar features and FR metric has access to the ground truth. While on the IETR dataset, the NR metric perform worse than the FR metrics. The main reason is probably also be the global shift distortion in the IVC image dataset.
To further explore the object shift effect, we have made an additional experiment on the IVC dataset while excluding the A1 view synthesis algorithm [@telea2004image] which causes great object shift in the synthesized views. The A1 algorithm fills the black holes in the dis-occlusion regions by simply stretching the adjacent texture which may cause great global object shift in the synthesized views. The results are shown in Table \[tab:IVC\_exclude\_A1\]. We can observe that the performance of FR and RR metrics increase significantly when large global shift artefacts are excluded.
PLCC RMSE SROCC
----------------------- ---------- -------- -------- --------
PSNR 0.7519 0.4525 0.6766
SSIM 0.5956 0.5513 0.4424
FR DIBR image metrics MW-PSNR 0.8545 0.3565 0.7750
RR DIBR image metrics MW-PSNRr 0.8855 0.3188 0.8298
: Performance on the IVC DIBR image dataset excluding A1 algorithm[]{data-label="tab:IVC_exclude_A1"}
The edge/contour based metrics also perform much better than the 2D pixel-based FR metrics since the edge/contour features can better represent the geometric degradations in the DIBR-synthesized images compared to simple pixel information.
The NR metrics do not need any reference information to evaluate the image quality, so the global shift does not have effect on the NR metrics. Besides, since the real reference images at virtual viewpoints are not always available in real applications, the NR metrics are more practical and useful. From table \[tab:plcc\_ivc\], we can easily find that the performance of the DIBR-synthesized view dedicated metrics decrease greatly in IETR dataset compared to their performance in IVC dataset. Among these metrics, the NR ones decrease the most, especially the learning based NR metrics. This is because of the fact that these NR metrics focus on the distortions in the IVC dataset, but in the IETR dataset, many “old fashioned” distortions are excluded.
{width="1\linewidth"}
\(a) Different / Similar analysis on IVC image dataset
{width="1\linewidth"}
\(b) Better / Worse analysis on IVC image dataset
{width="1\linewidth"}
\(c) Different / Similar analysis on IETR image dataset
{width="1\linewidth"}
\(d) Better / Worse analysis on IETR image datasets
{width="1\linewidth"}
\(e) Different / Similar analysis combining two datasets
{width="1\linewidth"}
\(f) Better / Worse analysis combining two datasets
As introduced in Section II, the MCL-3D dataset does not focus on the DIBR view synthesis distortions, but on the traditional distortion effects on the synthesized views. Thus, the performance of the tested objective metrics are quite different. Some of the metrics (Bosc, VSQA and NR\_MWT) that only consider the DIBR view synthesis distortions perform not as good as the traditional 2D metrics. Some 2D related FR metrics perform even worse than their original version. For instance, VSQA and 3DSwIM metrics can not achieve the performance of SSIM; the SCDM, MP-PSNR and MW-PSNR metrics perform worse than PSNR. Among these metrics, the feature-based FR metrics perform better than the simple edge/contour based metrics. It can be inferred that the frequency domain features can represent not only the edge/contour information, but also some other texture characteristics. The SET metric contains not only the DoG features for the DIBR view synthesis distortions, but also the GGCM based features for the texture naturalness. That may explain its good performance on both IVC and MCL-3D datasets.
The IVY dataset considers not only the view synthesis distortion, but also de binocular asymmetry in synthesized stereoscopic images. The baseline distance between the virtual viewpoint and the original viewpoint is much bigger than that in the other datasets. Thus, the metrics which do not consider the binocular asymmetry perform not well on this dataset.
### **[Results of Krasula’s model]{}**
only the IVC and IETR datasets are tested in this part since the MCL-3D and IVY datasets do not provide the standard deviation which represents the subject uncertainty. The obtained Area Under the Curves (AUC) and significant test results on IVC and IETR are shown in Table \[Fig: IVC\_IETR\_lucas\] (a) (b) (c) (d). The Fig. \[Fig: IVC\_IETR\_lucas\] (e) and (f) demonstrate the results on the combination of IVC and IETR datasets. A higher AUC value indicates a higher performance. In the significant test results, the white block indicates that the metric in the row performs significantly better that the metric in the column and vice versa for the black block. The gray block means these two metrics are statistically equivalent.
In the first different / similar analysis on the IVC dataset cf. \[Fig: IVC\_IETR\_lucas\] (a), none of these metrics perform well since most AUC values are below 0.7 and there even exist some metrics whose AUC values are under 0.5. Generally, the DIBR FR metrics perform better than the other metrics.
In the second different / similar analysis on the IVC dataset cf. \[Fig: IVC\_IETR\_lucas\] (b), the DIBR-synthesized view dedicated metrics perform significantly better than the 2D metrics (first and last 2 metrics) since the DIBR metrics can achieve higher AUC values. Among these metrics, the SCDM and SC-IQA metrics perform the best, they achieve AUC values higher than 0.9.
The results on the IETR dataset cf. \[Fig: IVC\_IETR\_lucas\] (c) (d) and the combination of the two datasets cf. \[Fig: IVC\_IETR\_lucas\] (e) (f) show that most of the FR metrics outperform the NR metrics except the SSIM metric. The 2D NR metrics achieve similar results compared to their performance on IVC dataset, while the performance of the DIBR NR metrics decrease greatly compared to their performance on IVC dataset. The results of Krasula’s model are consistent with the correlation coefficients results in the previous part.
-- --------- -------- -------- -------- ------------------------------------------------------------------------------------ -------- --------
PLCC RMSE SROCC PLCC RMSE SROCC
PSNR 0.5104 0.5690 0.4647 0.6525 0.0972 0.6366
SSIM 0.4081 0.6041 0.3751 0.4528 0.1144 0.4550
MOVIE 0.4971 0.4903 0.3877 0.646 0.097 0.693
ST-RRED 0.2025 0.6480 0.5777 **[0.7164]{} & **[0.0895]{} & **[0.6971]{}\
& SpEED & 0.3771 & 0.6128 & 0.5952 & **[0.7236]{} & **[0.0885]{} & **[0.6987]{}\
&VIIDEO & 0.5971 & 0.5308 & 0.5877 & 0.2586 & 0.1239 & 0.2535\
&Bosc & 0.5856 & 0.4602 & 0.2654 & 0.453 & 0.114 & 0.431\
&MP-PSNR & 0.5026 & 0.5720 & 0.5478 & 0.5681 & 0.1056 & 0.5044\
&MW-PSNR & 0.4911 & 0.4638 & 0.4558 & 0.5745 & 0.1050 & 0.5024\
&3DSwIM & 0.4822 & 0.4974 & 0.3320 & 0.5677 & 0.1057 & 0.2762\
&MP-PSNRr & 0.4617 & 0.5869 & 0.5307 & 0.5640 & 0.1059 & 0.5040\
&MW-PSNRr & 0.4802 & 0.5804 & 0.5038 & 0.5757 & 0.1049 & 0.5853\
& SIQE & 0.4084 & 0.5138 & 0.0991 & 0.3627 & 0.1195 & 0.2586\
&DSQM & 0.5241 & 0.4857 & 0.3157 & 0.4001 & 0.1071 & 0.3994\
& OUT & 0.6762 & 0.4874 & 0.6151 & 0.0945 & 0.1277 & 0.0926\
& NR\_MWT & **[0.7530]{} & **[0.4354]{} & **[0.7145]{} & 0.5051 & 0.1107 & 0.3092\
& NIQSV & 0.6505 & 0.5025 & 0.5963 & 0.5144 & 0.1100 & 0.4562\
&MNSS & 0.5180 & 0.5660 & 0.5371 & 0.1591 & 0.1266 & 0.2463\
&CQM & 0.4102 & 0.5101 & 0.3265 & 0.4021 & 0.1070 & 0.4064\
&PSPTNR & 0.4321 & 0.5002 & 0.4152 & 0.4461 & 0.1069 & 0.4305\
&VQA-SIAT & 0.5943 & 0.5321 & 0.5879 & **[0.8527]{} & **[0.0670]{} & **[0.8583]{}\
&CTI & **[0.6821]{} & **[0.4372]{} & **[0.6896]{} & 0.5736 & 0.1053 & 0.5425\
&FDI & **[0.7576]{} & **[0.4319]{} & **[0.7162]{} & 0.5952 & 0.1033 & 0.5425\
************************************
-- --------- -------- -------- -------- ------------------------------------------------------------------------------------ -------- --------
{width="1\linewidth"}
\(a) Different / Similar analysis on IVC Video dataset
{width="1\linewidth"}
\(b) Better / Worse analysis on IVC Video dataset
Performance on DIBR video datasets
----------------------------------
The DIBR-synthesized videos contain some temporal distortions, such as flickering, in addition to the spatial distortions in images. In this experiment, 12 state-of-the-art DIBR image metrics in addition to 5 DIBR video metrics are tested. To compare the performance of DIBR metrics and traditional 2D metrics, 5 widely used 2D video metrics and 2 2D image metrics are tested. The quality scores of image metrics are obtained by averaging the quality of all the frames. The three metrics which performance the best among the BIQA methods are marked in bold.
The obtained PLCC, RMSE and SROCC values on IVC video and SIAT video datasets are given in table \[tab:VIDEO\_plcc\]. Only the results of Krasula’s model on IVC video dataset are shown in Fig. \[Fig: IVC\_Video\_lucas\] since the SIAT video dataset does not provide the uncertainty of subject ratings.
The IVC video dataset focuses on the DIBR view synthesis distortions while the SIAT dataset focuses on the compression effect on the synthesized views. We can easily observe that the best three metrics on IVC and SIAT datasets are the DIBR metrics and 2D metrics respectively except VQA-SIAT metric. The VQA-SIAT metric mainly focuses on the compression effect which may lead obvious flicker in the DIBR-synthesized views. The spatial view synthesis distortions considered in this metric are very limited. That may explain why it significantly outperforms the other metrics on SIAT dataset while it can not obtain a very good performance on the IVC dataset. When we focus on the IVC video dataset, none of FR metrics achieve high correlation with the subjective results. Moreover, there is no significant difference between the performance of DIBR FR and 2D FR metrics. However, the DIBR NR metrics perform the best compared to other metrics. The main reason is the same as that on IVC image dataset: the global shift effect.
Discussions
-----------
The experimental results show that although great progress has been made towards the quality assessment of synthesized views, there is still significant room for improvement.
### **[Synthesized video quality assessment]{}**
the DIBR-synthesized videos contain not only the compression distortions but also the distortions induced by DIBR. The VQA-SIAT metric works well on capturing the temporal flicker caused by video compression, but it fails to assess the DIBR view synthesis distortions in the synthesized video frames. In addition, the imperfect view synthesis algorithms may also result in great miss-match between the adjacent frames in the synthesized video, which causes very annoying temporal distortions that the 8 by 8 block matching (in VQA-SIAT) may fail to detect. Therefore, we could try to further analyse the specific spatial-temporal distortions in the synthesized videos and design a complete metric for the DIBR-synthesized videos.
### **[Quality assessment of synthesized views in real applications]{}**
as introduced previously, DIBR can be used in various applications, but the quality assessment for these applications are rarely researched. For example, the free viewpoint videos (FVV) and multi-view videos (MVV) provide the images from multiple viewpoints at the same time instant. The temporal distortions in FVV or MVV are mainly introduced by the changing of viewpoints instead of timeline [@ling2019prediction; @ling2019perceptual]. This type of distortions are different from that in normal DIBR-synthesized views videos. Besides, in order to provide immersive perception for the observer, the AR or VR applications need to generate multiple synthesized images and change the viewpoint with the motion of the observer. The synthesized video contains both the inter-frame and inter-viewpoint temporal distortions, as well as the binocular asymmetric distortions which may happen in stereoscopic applications [@jung2016critical]. It could be interesting to try to design the metrics for these applications since they are currently rarely explored.
### **[Deep learning approaches]{}**
the main limitation of the usage of deep learning on the quality assessment of DIBR-synthesized views is the limited size of available dataset. Unlike the homogeneous distortions in the traditional 2D images, the distortions in the DIBR-synthesized views mostly occur in the dis-occlusion regions. In other words, the major part of the DIBR-synthesized view holds a perfect quality. The synthesized image can not be split into several patches and directly use the quality of the whole image as the quality of all the patches. Creating a very large-scale dataset may significantly help train a good deep model, but unlike the datasets for other tasks eg. object recognition, creating an image quality dataset necessarily requires subjective tests which are quite expensive and time-consuming. Thus, exploring how to train a comprehensive model on limited data could be more practical, eg. one-shot learning and few-shot learning [@fei2006one; @snell2017prototypical]. The fact that quality score of the whole synthesized image can not directly be distributed to all the image patches does not mean that the image can not be processed patch by patch. The main challenge is to find a proper pooling method to get the overall quality score. Although the pre-trained deep features have been successfully used in metrics [@wang2019deepicip; @ling2019gannrm], more effort could be made to create a more general and effective end-to-end deep model.
Conclusion
==========
In this paper, we present an up-to-date overview for the quality assessment methods of DIBR-synthesized views. We firstly described the existing DIBR-synthesized view datasets. Secondly, we analysed and discussed the recently proposed state-of-the-art objective quality metrics for DIBR-synthesized views, and classified them into different categories based on their used approaches. Then, we conducted a reliable experiment to compare the performance of each metric, and analysed their advantages and disadvantages at the same time. Furthermore, we discussed the potential challenges and directions for future research. We hope this overview can help to better understand the state-of-the-art of this research topic and provide insights to design better metrics and experiments for effective DIBR-synthesized images/videos quality evaluation.
Acknowledgment {#acknowledgment .unnumbered}
==============
The authors would like to thank Dr. Suiyi Ling and Dr. Yu Zhou for sharing their code. We would also like to thank Prof. Patrick Le Callet and Dr. Lucas Krasula for their kind advices on the experiment.
[^1]: Shishun Tian, Wenbin Zou and Xia Li are with the Guangdong Key Laboratory of Intelligent Information Processing, College of Electronics and Information Engineering, Shenzhen University, Shenzhen, 518060, China. ([email protected]; [email protected]; [email protected])
[^2]: Lu Zhang, Luce Morin and Olivier Déforges are with National Institute of Applied Sciences of Rennes (INSA de Rennes) and Institute of Electronics and Telecommunications of Rennes (IETR), Rennes, France. ([email protected]; [email protected]; [email protected])
[^3]: Ting Su is with Research Center for Medical Articial Intelligence, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, Guangdong 518055, China.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Intelligent detecting processing capabilities can be instrumental to improving safety and efficiency for firefighters and victims during firefighting activities. The objective of this research, is to create an automated system that is capable of real-time object detection and recognition that improves the situational awareness of firefighters on the scene. We have explored state of the art Machine Learning (ML) techniques to achieve this objective. The goal for this work is to maximize the situational awareness for fire-fighters by effectively exploiting the information gathered from the infrared camera and use a trained deep Convolutional Neural Network (CNN) system to classify and identify objects of interest in real time. In the midst of such critical circumstances created by a fire, this system is able to accurately inform the decision making process of firefighters in real time by extracting crucial information for processing. It is then able to make inferences about these circumstances to aid firefighters in safely navigating such hazardous and catastrophic environments.'
author:
- 'Manish Bhattarai$^1$, and Manel Martínez-Ramón$^2$ [^1] [^2] [^3] [^4]'
bibliography:
- 'lite.bib'
title: Detection and Identification of Objects and Humans in Thermal Images
---
Deep Convolutional Neural Networks, Infrared Images, Firefighting Environment, Fire fighters, Situational Awareness.
Introduction {#sec:introduction}
============
Current fire-fighting modality does not involve any automated detection mechanism and the target is identified solely by firefighter. The resulting detection process might be slowed down and obscured due to high anxiety levels of the active firefighter and due to presence of debris, smoke, fire, etc respectively. Hence under such hazardous conditions, a firefighter and others lives may be compromised due to decisions pertaining to the rescue operation that are based on incomplete or inaccurate environmental conditions. Fire fighter mortality statistics are largely due to inefficient decision making protocol which are derived from data informed by human senses that are affected by the on-the-ground conditions. To minimize these statistics, we propose an Artificial Neural Network based system which can autonomously identify objects and humans on the scene of the event, and in real time to better guide decision making protocol with respect to navigating a fire safely while rescuing victims. This system is capable of accurately predicting human posture to inform the firefighter about the hostage condition to better assess those in critical condition and could help in prioritization of rescue. In this paper, we demonstrate a CNN based autonomous system developed to generate situational awareness to the fighters and the commanders about the firefighting environment on which they are deployed. In this context, the situational awareness is generated by classifying fire, smoke, and other thermal conditions based on the data collected from the infrared video recorded by one on the firefighter. The CNN system is capable of making inferences that inform important decisions by detecting and classifying the crucial targets. This enables the fighter to assess the local critical conditions and choose a safer path in a fire environment by autonomously identifying the objects of interest such as human targets, excessive smoke, high temperatures, etc. and making right decisions based on real time conditions.
The application of CNN technology is predominately found in the fields of Surveillance and defense [@4; @5; @6; @7] but little has been done surrounding firefighting applications.
In the conventional firefighting system, different sensors such as temperature, UV, and fire detectors are used to determine the presence of fire, smoke, and other hazards [@8; @9]. Their long-established usage is evidenced by the prevalence of such sensors in all buildings, and is a requirement in building codes, to generate timely response for first responders. However, such detectors typically have a long response time in large spaces [@10]. Furthermore, they do not provide any spatial information regarding the presence of hazards in the given scenario. More contemporary firefighting modalities for detection of fire, smoke and other targets under fire environment rely on color [@10; @11; @12; @13], motion [@14; @15] and texture [@13; @16] features of the captured image. These vision-based approaches use histogram thresholding, optical flow- based motion vector computations, and texture analysis. Although these more modern techniques provide enhanced performance using RGB imagery and are an improvement over the more conventional techniques described above, they do not perform as robustly on IR or darker imagery. They also require higher computational times and lack sufficient complexity needed by algorithms to perform well. Currently, RGB images are hard to classify in fire and smoke environments. This gap can be filled by IR image technology. Also the presence of fire and smoke,by its very nature, creates a non-stationary environment and renders most existing stationary vision-based detection systems ineffectual in informing decision-making processes in real time. To address this issue, a robust real-time detection system based on CNN is proposed which is able to detect and localize the target of interest instantaneously. A similar work can be found in [@17], which describes the usage of infrared images to extract motion and statistics features in real-time using a Bayesian classifier for multi class identification. Significant research has also been done in human detection in other dark/ low visibility environments utilizing a single visible camera and fusion of the generated RGB image with an IR data set. Single IR camera based detection mechanisms have been presented in [@18; @19; @20; @21]. Most of this research is based on HOG based feature extraction and a classifier using SVM or other ML-based techniques [@20; @22]. Paper [@18] presents the use of GMM system in human detection. Template matching techniques and thresholding techniques have also been reported in [@20; @21; @23].
Other works related to our research have been published that perform classification tasks in thermal imaging. In [@Cho2018], a CNN is used in material recognition with non fire fighter grade thermal cameras. A related research is reported in [@Vandecasteele2017], where transfer learning is used to detect objects in a fireground. Nevertheless, the number of data samples used in this paper is low, so deep learning cannot be applied. Saliency detection and convolutional neural networks are applied to detect wildfires in [@Zhao2018]. In [@Kim2016] a Bayesian procedure to detect fire and smoke and discriminate them from thermal reflections in infrared is used in [@Yun2018]. In the recent work [@Ajith19] authors introduce a methodology based on Random Markov fields to segment fire, smoke and background in a sequence of images.
Further works that use thermal imagery and deep learning include [@Zhang2018], where authors use a CNN to detect known objects in infrared surveillance cameras. In [@Wang2016] a CNN is applied to the detection of vehicles in thermal imagery. Researchers in [@DeOliveira2018] introduce the use of CNNs to detect pedestrians in order to apply the detection to unmanned aerial vehicles (UAV). Another application in UAV is presented in [@Rodin2018], where authors train a CNN structure to detect objects of interest, as bodies or body parts or objects pertaining to victims of avalanches in he snow. A similar approach using CNNs in long wave infrared imagery to detect objects. Work [@Valldor2014] uses deep learning to detect persons with semisupervised approach that takes advantage of a large quantity of non labelled images containing humans.
In spite of the large quantity of works related to the processing of infrared images in fire ground or related applications, up to our knowledge, there are no works that attempt to construct a system that integrates online detection of interest targets in these scenarios, including humans, objects, poses or the presence of fire, that explore the possibilities of different network configurations trained with a large quantity of images recorded in real fire training situations by fire fighters. Thus, there is a need for effective automatic target detection generated in real-time in a firefighting environment as well as the associated need for a high accuracy classifier. Our research seeks to address this need. To do so, we have adapted and enhanced an existing state of the art CNN based automatic classifier system to improve it’s efficacy in identifying and classifying humans and objects of interest in real time in a firefighting environment. Also, we have improved upon the creation process of a data set so that it may be used effectively train the neural network (NN) system. We have also trained the system to detect objects and human simultaneously. Prior technique capabilities were limited to one or the other.
System Description
==================
The convolutional neural network presented is based on the structure of the VGG16 neural network presented in [@simonyan2014very]. Several different depht of the neural network have been tested, ranging from 1 to 5 convolutional sections as shown in Fig \[fig:1and2\], all of them followed by a fully connected section of two layers. The CNNs have an input of dimension $224\times 224$, corresponding to which the IR images are scaled to the this size. They are convolved by $3\times3$ filters to produce 64 channels of dimension $224\times 224$. The process is repeated and the resulting outputs are then passed through a set of ReLU activations, which are then pooled to produce 64 channels of dimension $112 \times 112$ pixels. Next, the output is passed through a $1 \times 1$ convolution to produce two channels of $112\times 112$ pixels. This is then processed through a fully connected layer with 4096 outputs, ReLU activation and dropping, a second identical layer, and then a layer with 5 outputs and soft max activation that gives the classification among 5 different classes.
For the model with two outputs, the first convolution layer is identical to the previous model, and after the first pool, a two more convolutions are added that produce 128 channels of size $112\times 112$, reduced to 128 channels of dimension $56\times 56$. The subsequent layers have identical structure as in the 1 layer model, where the input to the first fully connected layer has 8 channels of dimension $56\times56$. The models with 3, 4 and 5 layers are constructed using the same methodology.
The networks have been trained and tested in three different modalities The first one was object classification, which includes persons, ladders, windows, doors and combination of windows and fire fighters. The second modality was pose classification, which included standing, sitting and lying. The third modality is binary and included the presence or not of fire. For the modalities with 5 and 3 classes, the training was performed using a categorical cross entropy loss, and for two classes, binary cross entropy loss, combined with stochastic gradient descent. The learning rate for all trainings was $1^{-4}$ and a decay of $0.009$. Early stopping with cross validation was applied to avoid overfitting.
Table is before augmentation. After augmentation they were in around 17000 each for five classes. And here the augmentation is done for only training purpose. ALso augmentation for the poses to make each class around 8000.
Training and test datasets
==========================
The data set used in this project was recorded at the Santa Fe Firefighting Facility, located in Santa Fe, New Mexico. Extensive video data was acquired using an IR MSA 5200HD2TIC Camera. This camera is a multipurpose firefighting tool designed for search and rescue and structural firefighting. It uses an uncooled microbolometer vanadium oxide(Vox) detector which comprises of 320x240 FPA with the pitch of 38um and spatial resolution of 7.5 to 13.5um. This resolution is sufficient to capture necessary features for target detection. It records the image a with 320x240 focal plane array sensor and it has the ability to record imagery in two different modes, i.e. low and high sensitivity modes. This device also features high score imagery generating 76000 pixels of image detail in low and high sense modes. Dense spectral resolution is (7.5 to 13.5 $\mu m$) and the output video is in NTSC format with a frame rate of 30 frames per second/ The scene temperature has a range of 560 degrees Celsius or 1040 degree Fahrenheit.
The recordings were taken in both open and closed environments for a total time of over 6 hours of recorded video. The recording sessions produced more than 400 infrared video files, each one between 2 and 3 minutes long. All videos contain sequences with the the subjects to be detected. In some scenarios single objects of interest are present, and in others there are multiple objects of interest that need to be detected in the same frame.
The objects of interest in this setup are humans, doors, windows, ladders, smoke and fire, and the objective of the above described structure is to detect all objects in the scene at the same time.
![Demonstration of Image augmentation for training set.[]{data-label="fig:augment"}](Images/examples1_grid.jpg)
Since data sets of sufficient details are needed to accurately train the neural network, the videos were used to extract a large quantity of images for training and test purposes. The training and test sequences were extracted from different videos in order to avoid over fitting. The video extracted from the camera is produced in grayscale and 8 bits. In order to generate the training data set the images have been preprocessed with data augmentation techniques such as skewing, translation, zooming, cropping and rotation as shown in fig \[fig:augment\] (False color for better visualization).
Tables \[tab:data\_objects\], \[tab:data\_fire\] and \[tab:data\_poses\] show he total number of images acquired for training and test before the augmentation procedure. After the data augmentation of the training dataset, the total number of resulting images was of three sets (objects, fire, poses) of about 17000 images per set. The number of data produced was higher for those objects with less samples in order to compensate for the assimetry of the different classes.
The training has been performed in an Alienware Aurora R6 Desktop computer configured with 32GB RAM memory, and a Dual GTX 1080 with 16GB GPU memory.
Object of Interest number
------------------------ --------
door 322
firefighter and window 4663
firefighter 15484
ladder 1589
window 1620
: Data quantification for poses classification[]{data-label="tab:data_poses"}
Object of Interest number
-------------------- --------
fire 7603
No fire 7950
: Data quantification for poses classification[]{data-label="tab:data_poses"}
Object of Interest number
-------------------- --------
Crawling 8678
sitting 1803
standing 9928
: Data quantification for poses classification[]{data-label="tab:data_poses"}
The test data consists of 1/10 of the data set. The rest of the data has been used for training purposes. A validation has been performed with a 9 fold procedure with the rest 9/10th of data.
![t-SNE of the CNN when trained to detect objects. The number of convolutional layers was 1.[]{data-label="fig:TSNE"}](Images/TSNE2.png)
Results
=======
In the experiments conducted the network is trained to detect objects, including ladders, doors and windows, the presence of fighters and fire. The networks is also trained to detect the pose of the detected fire fighters. In this experiment we use three different poses, corresponding to standing, sitting and crawling, which are the poses available in the videos. Other poses that are important can be detected with if there are available images for training. For example, the presence of a laying person is very important in a fire scenario, as it is significant of a possible victim in the fire ground. The results below include the visualization of the features extracted by the convolutional section of the network, the F1 scores and accuracy of the network, confusion matrices and ROC curves to estimate the false alarm versus the detection probability of the network. For this experiment, we change the detection probability by sweeping the detection threshold of the network.
Visualization of the extracted features
---------------------------------------
All the configurations have been tested against the available data. Figure \[fig:features\] show the visualization of the features obtained at the end of the last convolutional layer, which are fed into the fully connected layer. The upper left picture shows the input image with a fire fighter going inside a room from a window. The picture from the network with one layer (Depth = 1) shows that the convolution simply detects the edges of the objects of interest. With depth equal to 2 and 3, the network produces a poor feature extraction, showing that these configurations are inadequate, as it is stated in the results below. Nevertheless, for depths equal to 4 and 5, a higher level of abstraction features are extracted that produce satisfactory results.
Figure \[fig:TSNE\] shows the visualization of the output of the neural networks using the t-distributed stochastic neighbor embedding technique presented in [@maaten2008visualizing]. This technique allows to obtain a low dimensional representation of the data in order to have an intuition of its distribution. The upper left pane shows the distribution of the 5-dimensional output of the network when the depth is of one convolutional layer. In this configuration, the separability of the data is intuitively fair. Right upper and left lower panes show the distribution for depth 2 and 3. These representations show a high level of overlapping of the different classes, which is consistent with the poor features extracted, as shown in figure \[fig:features\]. The representation for a depth of 4 layers is highly improved, as the overlapping is dramatically decreased with respect to the 2 and 3 layer networks.
The green dots are images corresponding to fire fighters and the orange ones are images containing fighters in windows. These take almost all the surface but they show a low overlapping among them. The blue dots correspond to training images with doors. All doors even in different perspectives have extracted features which are very similar, thus all features appear clustered in a small area. Nevertheless, they are highly overlapped with the images containing fire fighters, which occupy the largest area. The overlapping decreases for 4 and 5 layers. The violet spots are the windows, which appear to be overlapping with the images of fighters and windows. The accuracy and classification rates are in high agreement with this visualization.
Accuracy and precision
----------------------
The accuracy of the neural network for different depths is consistent with the visualization of the features and the t-SNE. Figures \[fig:F1Scores\] and \[fig:precision\] show the F1 scores and the precision of the network with depths form 1 to 5. In this experiment the networks have been trained to detect objects. In particular, the targets were doors, fire fighters, ladders, windows and combination of fighters and windows. The network with one layer has a reasonable between accuracy close to 75% if we taken into account the fact that it has a low computational burden in training and test, although it shows a high variance in the results. Nevertheless, the use of two and three layers produces an unacceptable performance, which dramatically improves when using 4 and 5 layers. It is apparent that the use of 5 layers is not necessary since its performance is almost identical to the one with 4 layers.
![F1 scores for the classification of objects with CNNs of different depths.[]{data-label="fig:F1Scores"}](Images/F1-score_objects.png)
![Precision curves for object detection with CNNs of different depths.[]{data-label="fig:precision"}](Images/Precision_objects.png)
Figure \[fig:fireAccuracy\] shows the accuracy of the network for fire detection. In this case it is only necessary a network of one layer. As in the previous experiments, the network with three layers produces a poor performance, while the rest show an accuracy that is almost identical. The average test accuracy for all objects is depicted in \[fig:testAccuracy\]. From this graph, we can conclude that a good trade off between computational burden and accuracy is obtained with a network of only one convolutional layer, and a network with 4 layers has an excellent accuracy of more than 97%.
![Test accuracy in fire detection with CNNs of different dephts.[]{data-label="fig:fireAccuracy"}](Images/accuracy_fire.png)
![Test accuracy in object detection with CNNs of different dephts.[]{data-label="fig:testAccuracy"}](Images/Test_Accuracy_ObjectsBox_plot.png)
An interesting use of the network is related to the pose classification of the humans detected in the scene. This is important in order to determine if the person is in any kind of trouble. In the example, we trained the network to detect persons standing, sitting or crawling. from labelled images. The network shows a similar performance with accuracies ranging the 95% in average.
Confusion matrices
------------------
The confusion matrices for all detection modalities have been computed. Tables \[tab:confusionObjects1\] to \[tab:confusionObjects5\] shows the confusion matrices for the task of objects and human detection. As stated before, the network with just one layer of convolutional blocks has a reasonable performance. The detection probabilities are around 75%. Nevertheles, when the network is challenged to detect fire fighters and windows, it has a high confusion rate in detecting windows and ladders, as there is a confusion of about 19% with fire fighters in the classification of these elements. When the number of convolutional layers is increased to two, the confusion is even worse, reaching a 35%. In these cases, the detection of fire fighters is between 97 and 99%. If the number of layers is increased to 4, the confusion matrices show a confusion between 0 and 1.7% for all elements. As stated above, there are no significant differences between 4 and 5 layers. Note that the classification performance with these number of layers ranges between 97 and 99.8%.
Door 73.2 1.1 23.5 0.0 2.2
--------- ------ ------ ------ ------ ------
F/W 0.0 77.7 19.0 0.0 3.3
Fighter 0.3 1.8 97.0 0.1 0.8
Ladder 00.0 1.2 19.1 79.6 0.1
Window 0.3 4.3 19.9 0.0 75.5
: Confusion matrix for object detection, depth=1[]{data-label="tab:confusionObjects1"}
Door 61.3 0.0 38.2 0.0 0.5
--------- ------ ------ --------- -------- --------
F/W 0.0 62.4 36.0 0.0 1.7
Fighter 0.1 0.2 99.4 0.1 0.2
Ladder 0.0 0.0 34.3 65.7 0
Window 0.1 1.0 35.9 0 63.0
Door F/W Fighter Ladder Window
: Confusion matrix for object detection, depth=2[]{data-label="tab:confusionObjects2"}
Door 97.1 0.2 0.9 0.0 1.7
--------- ------ ------ --------- -------- --------
F/W 0.0 98.3 1.0 0.0 0.7
Fighter 0.1 0.1 99.8 0.0 0.0
Ladder 0.0 0.0 0.2 99.8 0.0
Window 0.3 0.5 0.3 0.0 98.9
Door F/W Fighter Ladder Window
: Confusion matrix for object detection, depth=4[]{data-label="tab:confusionObjects4"}
Door 98.0 0.0 0.7 0.0 1.3
--------- ------ ------ --------- -------- --------
F/W 0.0 99.2 0.5 0.0 0.3
Fighter 0.0 0.1 99.9 0.0 0.0
Ladder 0.0 0.0 0.1 99.9 0.0
Window 0.3 0.6 0.2 0.0 98.9
Door F/W Fighter Ladder Window
: Confusion matrix for object detection, depth=5[]{data-label="tab:confusionObjects5"}
A similar performance can be seen in the detection of poses. Table \[tab:confusionPoses1\] shows a poor performance of the network with 1 and similarly for two layers (Table \[tab:confusionPoses2\]), mainly in the detection of fighters sitting and crawling. This is possibly due to the fact that the relative positions of the fire fighters in these two poses are similar but simply rotated. The results increase significantly with 4 layers (Table \[tab:confusionPoses4\]), where the confusion decreases to a range between 0 and 0.4 in all cases except for the detection of the sitting pose, which stands at a confusion rate of 5.7 (in the network with 4 layers) and 5.1% with the crawling pose with 5 layers (Table \[tab:confusionPoses5\]). Again, it is evident that there are no significant differences between the results of these two network configurations.
Crawling 85.6 1.2 13.2
---------- ------ ------ ------
Sitting 41.4 45.9 12.7
Standing 10.8 0.4 88.8
: Confusion matrix for pose detection, depth=1[]{data-label="tab:confusionPoses1"}
Crawling 86.3 0.8 12.9
---------- ---------- --------- ----------
Sitting 19.3 67.8 12.9
Standing 7.5 0.2 92.3
Crawling Sitting Standing
: Confusion matrix for pose detection, depth=2[]{data-label="tab:confusionPoses2"}
Crawling 99.5 0.2 0.3
---------- ---------- --------- ----------
Sitting 5.7 94.1 0.2
Standing 0.5 0.0 99.5
Crawling Sitting Standing
: Confusion matrix for pose detection, depth=4[]{data-label="tab:confusionPoses4"}
Crawling 99.4 0.2 0.4
---------- ---------- --------- ----------
Sitting 5.1 94.7 0.2
Standing 0.5 0.0 99.5
Crawling Sitting Standing
: Confusion matrix for pose detection, depth=5[]{data-label="tab:confusionPoses5"}
Detection versus false alarm probabilities
------------------------------------------
Finally, Figures \[fig:ROC\_object\], \[fig:ROC\_pose\] and \[fig:ROC\_fire\] show the probability of detection versus false alarm for the classification of objects and humans, poses and fire. The graphs have been obtained by sweeping the detection threshold from 0 to 1. In all cases, it is observed that the probability of false alarm is negligible when the networks have a depth of 4 or 5 convolutional sections. The results with depth 2 and 3 are not interesting, and for the case of a depth of one convolutional layer, the results evidence flaws in the probability of detection versus probability of false alarm. While the object detection a detection rate of about 90% is achieved with a false alarm of about 10% (\[fig:ROC\_fire\]), the optimal false alarm rates in object detection are of about 20%, in particular, in the detection of humans. Similar results are obtained in pose detection. Therefore, the lower computational burden network should be used only in cases where such false alarm rates can be tolerated. This,m of course, can be decreased if the detection is performed not in one image but in a sequence of images, and then a voting procedure is applied to all detections. This would be done at expenses of a latency in the object or pose detection procedures.
![ROC curve for object detection.[]{data-label="fig:ROC_object"}](Images/ROC_Overall_objects.png)
![ROC curve for pose detection.[]{data-label="fig:ROC_pose"}](Images/ROC_CUrve_Poses.png)
![ROC curve for fire detection.[]{data-label="fig:ROC_fire"}](Images/ROC_CUrve_Fire.png)
Conclusion
==========
There is a need for mechanisms that are able to automatically interpret the fire ground from the information provided by the hand held thermal cameras used by fire fighters. These systems should be able to automatically and reliably detect all objects and situations of interest in the fire ground. Convolutional neural networks have demonstrated outstanding performance in object detection in RGB and thermal imaging in a large variety of works. In this paper we introduce the use of CNNs to simultaneously detect objects, persons, person poses and fire from thermal image cameras. The detection can be performed online and locally if a camera is connected to a simple single board commercial computer endowed with GPU capabilities (for example, an NVIDIA JETSON).
The data that is used in this investigation has been recorded ad hoc by fire fighters in training sessions in facilities in Santa Fe, New Mexico (USA). The sequences have been preprocessed to be segmented and labelled manually in order to obtain a large training and test database that has been further augmented using augmentation techniques described above.
The CNN chosen for this investigation is the well known VGG16. This network shows a heavy computational load in training, but an affordable test burden that can be held in a small GPU computer. The test results show a performance that exceeds a classification probability above 95% in all classifications when the configuration has 4 layers of convolutional sections.
Future work include the real time segmentation and tracking of the detected objects as well as the use of other deep learning structures to produce human readable descriptions of the fire ground scenario.
[Manish Bhattarai]{} received the Masters degree in Electrical Engineering with specialization in signal processing from the University of New Mexico in 2017.He is currently working towards his Ph.D. degree in Electrical Engineering from the same school as well as employed as a full time research assistant at the Los Alamos National Laboratory. His research interests are Machine Learning, deep learning ,Computer Vision, AI and HPC.
[Manel Martínez Ramón]{} is a professor with the ECE department of The University of New Mexico. He holds the King Felipe VI Endowed Chair of the University of New Mexico, a chair sponsored by the Household of the King of Spain. He is a Telecommunications Engineer (Universitat Politecnica de Catalunya, Spain, 1996) and PhD in Communications Technologies (Universidad Carlos III de Madrid, Spain, 1999). His research interests are in Machine Learning applications to smart antennas, neuroimage, first responders and other cyber-human systems, smart grid and others. His last work is the monographic book “Signal Processing with Kernel Methods”, Wiley, 2018.
[^1]: $^1$Los Alamos National Labs, New Mexico, USA
[^2]: $^2$Department of Electrical and Computer Engineering, The University of New Mexico, New Mexico 87106, USA.
[^3]: This work has been supported by NSF S&CC EAGER grant 1637092.
[^4]: Corresponding author: Manish Bhattarai (e-mail:[email protected]).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study uncertainty and certainty relations for two successive measurements of two-dimensional observables. Uncertainties in successive measurement are considered within the following two scenarios. In the first scenario, the second measurement is performed on the quantum state generated after the first measurement with completely erased information. In the second scenario, the second measurement is performed on the post-first-measurement state conditioned on the actual measurement outcome. Induced quantum uncertainties are characterized by means of the Tsallis entropies. For two successive projective measurement of a qubit, we obtain minimal and maximal values of related entropic measures of induced uncertainties. Some conclusions found in the second scenario are extended to arbitrary finite dimensionality. In particular, a connection with mutual unbiasedness is emphasized.'
author:
- 'Alexey E. Rastegin'
title: 'Uncertainty and certainty relations for successive projective measurements of a qubit in terms of Tsallis’ entropies'
---
Introduction {#sec1}
============
The Heisenberg uncertainty principle [@wh27] is one of the fundamentals of quantum theory. Despite of its wide popularity, there is no general consensus over the scope and validity [@lahti; @hall99]. It is typically said that measuring some observable will inevitably disturb the system, whence the context for further observations is raised. There are many ways to quantify uncertainties as well as few scenarios of measuring observables. The first form of explicit mathematical formulation has been given for the position and momentum by Kennard [@kennard]. For this pair, a product of the standard deviations cannot be less than $\hbar/2$. For any pair of observables, this direction was realized by Robertson [@robert]. However, the traditional approach deals with quantum uncertainties raised in two different experiments with the same pre-measurement state. So, this approach does not reveal many details about a disturbance of the system due to performed measurements. Rather, Heisenberg’s initial reasons are better formulated in terms of noise and disturbance [@ozawa04]. As was discussed in Refs. [@mdsrin03; @paban13], studies of quantum uncertainties in the results of successive measurements have received less attention than they deserve.
Much attention to uncertainty relations is stimulated by a progress in using quantum systems as informational recourses [@ww10]. Hence, formulations in information-theoretic terms including entropies are of interest. Traditional uncertainty relations for quantum measurements deal with probability distributions calculated for one and the same input state. This treatment was studied in entropic terms [@deutsch; @maass; @brud11] and recently by means of majorization technique [@prz13; @fgg13; @rpz14]. In quantum information processing, other situations are rather typical. Our subsequent manipulations deal with an output state of the latter stage. In the case of two successive measurements, the following two scenarios are typically addressed [@bfs2014; @zzhang14]. In the first scenario, the second measurement is performed on the quantum state generated after the first measurement with completely erased information. In the second scenario, the second measurement is performed on the post-first-measurement state conditioned on the actual measurement outcome. Thus, the scenarios are related to a realistic situation, when subsequent actions deal with an output state of the latter stage.
In the present work, we study uncertainties in successive measurements with the use of Tsallis entropies. We also focus on maximal possible values of related entropic measures. Indeed, certainty relations for successive measurements seem to be not considered previously. For successive measurements of a pair of qubit observables, we obtain uncertainty and certainty entropic bounds. The paper is organized as follows. In Sec. \[sec2\], the background material is reviewed. Here, definitions of the used entropic functions are recalled. In Sec. \[sec3\], we generally discuss quantum uncertainties induced by successive measurements of a pair of observables. Two scenarios of such measurements and related $\alpha$-entropic measures are introduced. In Section \[sec4\], we derive tight uncertainty and certainty relations for a qubit within the first scenario. The conditions for equality are obtained and discussed. Tight lower and upper bounds on the conditional $\alpha$-entropies within the second scenario are presented in Sec. \[sec5\]. We also observe a connection of the equality conditions for upper bounds with mutual unbiasedness. This observation remains valid in arbitrary finite dimensions. In Sec. \[sec6\], we conclude the paper with a summary of results.
Preliminaries {#sec2}
=============
To quantify uncertainties in generated probability distributions, we will use Tsallis’ entropies. Let discrete random variable $X$ take values on a finite set $\Omega_{X}$ of cardinality $\#\Omega_{X}$. The Tsallis entropy of degree $\alpha>0\neq1$ is defined by [@tsallis] $$H_{\alpha}(X):=\frac{1}{1-\alpha}\left(\sum_{{\,}x\in\Omega_{X}} p(x)^{\alpha}
- 1\right)
{\,}. \label{tsaent}$$ With other factor, this function was examined by Havrda and Charvát [@havrda] and later by Daróczy [@ZD70]. In statistical physics, the entropy (\[tsaent\]) is extensively used due to Tsallis [@tsallis]. For other multidisciplinary applications, see the book [@gmt] and references therein. The Rényi entropies [@renyi61] form another especially important family of one-parametric extensions of the Shannon entropy. For $\alpha>0\neq1$, Rényi’s $\alpha$-entropy can be expressed via Tsallis’ entropy as $$R_{\alpha}(X)=\frac{1}{1-\alpha}{\>}
\ln\bigl(1+(1-\alpha)H_{\alpha}(X)\bigr)
\ . \label{exrt}$$ Despite of a direct relation, the entropies (\[tsaent\]) and (\[exrt\]) differ in essential properties. As a rule, formulation in terms of one of the two entropies cannot immediately be recast for other. Such a situation takes place for entropic uncertainty relations too. In Ref. [@rast14a], uncertainty and certainty relations for the Pauli observables were derived in terms of the Rényi entropies. Using Rényi’s entropies, uncertainty relations for successive projective measurements were studied in Ref. [@zzhang14]. Properties and applications of both the types of entropies in quantum theory are considered in the book [@bengtsson].
Obviously, the function $\bigl(\xi^{\alpha}-\xi\bigr)/(1-\alpha)$ is concave for all $\alpha>0$. Hence, the entropy (\[tsaent\]) is a concave function of probability distribution. The maximal value $\ln_{\alpha}(\#\Omega_{X})$ is reached by the uniform distribution. It is often convenient to rewrite Eq. (\[tsaent\]) as $$\begin{aligned}
H_{\alpha}(X)&=-\sum_{x\in\Omega_{X}}p(x)^{\alpha}{\,}\ln_{\alpha}{p}(x)
\nonumber\\
&=\sum_{x\in\Omega_{X}}p(x){\>\,}{\ln_{\alpha}}{\left(\frac{1}{p(x)}\right)}
\, . \label{tsaln}\end{aligned}$$ Here, we used the $\alpha$-logarithm defined for $\alpha>0\not=1$ and $\xi>0$ as $$\ln_{\alpha}(\xi)=\frac{\xi^{1-\alpha}-1}{1-\alpha}
\ . \label{lnadf}$$ In the limit $\alpha\to1$, we have $\ln_{\alpha}(\xi)\to\ln\xi$, so that the Shannon entropy $H_{1}(X)=-\sum_{x}p(x)\ln{p}(x)$ is raised. For brevity, we will typically write entropic sums without set symbols such as $\Omega_{X}$.
The above definitions are also applied in the quantum regime. Let the state of a quantum system be described by the density matrix ${\boldsymbol{\rho}}$. It is a positive semi-definite matrix with the unit trace. The quantum $\alpha$-entropy of ${\boldsymbol{\rho}}$ is defined as $${{\textup{H}}}_{\alpha}({\boldsymbol{\rho}}):=\frac{1}{1-\alpha}{\>}\Bigl({{\textup{Tr}}}({\boldsymbol{\rho}}^{\alpha})-1\Bigr)
\, . \label{qtsdf}$$ In the case $\alpha=1$, we deal with the von Neumann entropy ${{\textup{H}}}_{1}({\boldsymbol{\rho}})=-{{\textup{Tr}}}({\boldsymbol{\rho}}\ln{\boldsymbol{\rho}})$. For general properties on the von Neumann entropy, see Refs. [@wehrl; @ohya]. In the following, the entropies (\[tsaent\]) and (\[qtsdf\]) will be used in studying uncertainties in successive projective measurements.
In one of the two scenarios of successive measurements, the second measurement is performed on the actual post-first-measurement state. Analyzing this scenario, conditional form of entropies will be utilized. The conditional entropies are widely used in information theory [@CT91] and in applied disciplines. The standard conditional entropy is defined as follows. Let $X$ and $Z$ be random variables. For each $z\in\Omega_{Z}$, we take the function $$H_{1}(X|z)=-\sum_{x\in\Omega_{X}}p(x|z){\,}\ln{p}(x|z)
\ . \label{csheny}$$ By $p(x|z)$, we mean the conditional probability that $X=x$ given that $Z=z$. By Bayes’ rule, it obeys $p(x|z)=p(x,z)/p(z)$. The entropy of $X$ conditional on knowing $Z$ is defined as [@CT91] $$\begin{aligned}
H_{1}(X|Z)&:=\sum_{z\in\Omega_{Z}} p(z){\,}H_{1}(X|z)
\nonumber\\
&=-\sum_{x\in\Omega_{X}}\sum_{z\in\Omega_{Z}} p(x,z){\,}\ln{p}(x|z)
\ . \label{cshen}\end{aligned}$$ For brevity, we will further write entropic sums without mentioning that $x\in\Omega_{X}$ and $z\in\Omega_{Z}$. In the context of quantum theory, the conditional entropy (\[cshen\]) was used in information-theoretic formulations of Bell’s theorem [@BC88] and noise-disturbance uncertainty relations [@bhow13].
In the literature, two kinds of the conditional THC entropy were considered [@sf06; @rastkyb]. These forms are respectively inspired by the two expressions shown in Eq. (\[tsaln\]). The first conditional form is defined as [@sf06] $$H_{\alpha}(X|Z):=\sum_{z} p(z)^{\alpha}{\,}H_{\alpha}(X|z)
\ , \label{hct1}$$ where $$H_{\alpha}(X|z):=\frac{1}{1-\alpha}{\,}\left(\sum_{x} p(x|z)^{\alpha}-1\right)
{\,}. \label{hctm0}$$ The conditional entropy (\[hct1\]) is, up to a factor, the quantity originally introduced by Daróczy [@ZD70]. For any $\alpha>0$, the conditional entropy (\[hct1\]) satisfies the chain rule [@sf06; @ZD70]. The chain rule is essential in some applications of conditional entropies in quantum information science. For example, this rule for the standard conditional entropy is used in deriving the Braunstein–Caves inequality [@BC88]. This inequality expresses an entropic version of the Bell theorem [@bell64]. The $\alpha$-entropies were applied in formulating non-locality, contextuality and non-macrorealism inequalities [@rastqic14; @rastq14].
Using the particular function (\[hctm0\]), the second form of conditional $\alpha$-entropy is written as [@sf06; @rastkyb] $${\widetilde{H}}_{\alpha}(X|Z):=\sum_{z} p(z){\,}H_{\alpha}(X|z)
\ . \label{hct2}$$ It should be noted that this form of conditional entropy does not share the chain rule of usual kind [@sf06]. Nevertheless, the conditional entropy (\[hct2\]) is interesting at least as an auxiliary quantity [@rastkyb]. Moreover, it can be used in studying the Bell theorem [@rastrian], even though the chain rule is not applicable here.
On successive projective measurements in finite dimensions {#sec3}
==========================================================
In this section, we consider several facts concerning uncertainty and certainty relations for successive measurements in finite dimensions. Let ${\boldsymbol{Z}}$ be an observable of some finite-level quantum system. By ${{\textup{spec}}}({\boldsymbol{Z}})$, we denote the spectrum of ${\boldsymbol{Z}}$. The spectral decomposition is written as $${\boldsymbol{Z}}=\sum_{z\in{{\textup{spec}}}({\boldsymbol{Z}})}{z{\,}{\boldsymbol{P}}_{z}}
\ , \label{xspc}$$ where ${\boldsymbol{P}}_{z}$ denotes the orthogonal projection on the corresponding eigenspace of ${\boldsymbol{Z}}$. The operators ${\boldsymbol{P}}_{z}$ are mutually orthogonal and satisfy the completeness relation $$\sum_{z\in{{\textup{spec}}}({\boldsymbol{Z}})}{{\boldsymbol{P}}_{z}}={\openone}\ . \label{cmrl}$$ By ${\openone}$, we denote the identity operator in the Hilbert space of studied system. Let the pre-measurement state be described by the density matrix ${\boldsymbol{\rho}}$. The probability of each outcome $z$ is expressed as ${{\textup{Tr}}}({\boldsymbol{P}}_{z}{\,}{\boldsymbol{\rho}})$. Due to Eq. (\[cmrl\]), these probabilities are summarized to $1$. Calculating the $\alpha$-entropy of the generated probability distribution, we will deal with the quantity $$H_{\alpha}({\boldsymbol{Z}};{\boldsymbol{\rho}})=\frac{1}{1-\alpha}
\left(\sum_{{\,}z\in{{\textup{spec}}}({\boldsymbol{Z}})} \bigl({{\textup{Tr}}}({\boldsymbol{P}}_{z}{\,}{\boldsymbol{\rho}})\bigr)^{\alpha}
- 1\right)
{\,}. \label{haby}$$ Let ${\boldsymbol{X}}$ be another observable with the spectral decomposition $${\boldsymbol{X}}=\sum_{x\in{{\textup{spec}}}({\boldsymbol{X}})}{x{\,}{\boldsymbol{Q}}_{x}}
\ . \label{yspc}$$ Here, the operator ${\boldsymbol{Q}}_{x}$ is the orthogonal projection on the corresponding eigenspace of ${\boldsymbol{X}}$. We consider an amount of uncertainties induced by two successive measurements of the observables, ${\boldsymbol{Z}}$ at first and ${\boldsymbol{X}}$ later. As was above mentioned, there are two possible scenarios of interest.
In the first scenario, the second measurement is performed on the quantum state generated after the first measurement with completely erased information. That is, the second measurement is performed with the pre-measurement state $${\mathcal{E}}_{{\boldsymbol{Z}}}({\boldsymbol{\rho}})=\sum_{{\,}z\in{{\textup{spec}}}({\boldsymbol{Z}})}{\boldsymbol{P}}_{z}{\,}{\boldsymbol{\rho}}{\,}{\boldsymbol{P}}_{z}
\ . \label{smpm}$$ The linear map ${\mathcal{E}}_{{\boldsymbol{Z}}}$ describes the action of the projective measurement of ${\boldsymbol{Z}}$. The right-hand side of Eq. (\[smpm\]) is an operator-sum representation of the map [@nielsen]. The completeness relation (\[cmrl\]) provides that the map (\[smpm\]) preserves the trace for all inputs. The uncertainty in the second measurement is quantified by means of the entropy $H_{\alpha}\bigl({\boldsymbol{X}};{\mathcal{E}}_{{\boldsymbol{Z}}}({\boldsymbol{\rho}})\bigr)$. The latter is expressed similarly to Eq. (\[haby\]), but with the probabilities ${{\textup{Tr}}}\bigl({\boldsymbol{Q}}_{x}{\,}{\mathcal{E}}_{{\boldsymbol{Z}}}({\boldsymbol{\rho}})\bigr)$. We will characterize a total amount of uncertainty in the first scenario by the sum of the classical entropies of two generated probability distributions. Note that the post-first-measurement state (\[smpm\]) obeys the following property. If we have measured ${\boldsymbol{Z}}$ in the state ${\mathcal{E}}_{{\boldsymbol{Z}}}({\boldsymbol{\rho}})$, we again deal with probabilities ${{\textup{Tr}}}({\boldsymbol{P}}_{z}{\,}{\boldsymbol{\rho}})$. Thus, we write the formula $$\begin{aligned}
&H_{\alpha}({\boldsymbol{Z}};{\boldsymbol{\rho}})+H_{\alpha}\bigl({\boldsymbol{X}};{\mathcal{E}}_{{\boldsymbol{Z}}}({\boldsymbol{\rho}})\bigr)
\nonumber\\
&={{\textup{H}}}_{\alpha}\bigl({\mathcal{E}}_{{\boldsymbol{Z}}}({\boldsymbol{\rho}})\bigr)+{{\textup{H}}}_{\alpha}\bigl({\mathcal{E}}_{{\boldsymbol{X}}}\circ{\mathcal{E}}_{{\boldsymbol{Z}}}({\boldsymbol{\rho}})\bigr)
\ . \label{hhcq}\end{aligned}$$ By ${\mathcal{E}}_{{\boldsymbol{X}}}\circ{\mathcal{E}}_{{\boldsymbol{Z}}}$, we mean the composition of two quantum operations [@nielsen]. Further, the projectors ${\boldsymbol{Q}}_{x}$ give rise to the map $${\mathcal{E}}_{{\boldsymbol{X}}}({\boldsymbol{\omega}})=\sum_{{\,}x\in{{\textup{spec}}}({\boldsymbol{X}})}{\boldsymbol{Q}}_{x}{\,}{\boldsymbol{\omega}}{\,}{\boldsymbol{Q}}_{x}
\ , \label{pmsm}$$ where ${\boldsymbol{\omega}}$ is a density matrix. The left-hand side of Eq. (\[hhcq\]) is the sum of classical $\alpha$-entropies of the two probability distributions, whereas the right-hand side is the sum of quantum $\alpha$-entropies. To formulate uncertainty and certainty relations for successive measurements, we aim to have a two-sided estimate on the quantity (\[hhcq\]).
Let us recall one of physically important properties of the von Neumann entropy related to the measurement process. In effect, projective measurements cannot decrease the von Neumann entropy (see, e.g., theorem 11.9 in Ref. [@nielsen]), that is $${{\textup{H}}}_{1}\bigl({\mathcal{E}}_{{\boldsymbol{Z}}}({\boldsymbol{\rho}})\bigr)\geq{{\textup{H}}}_{1}({\boldsymbol{\rho}})
\ . \label{vnpm}$$ In the paper [@rastjst], we extended the above property to the family of quantum unified entropies. In particular, for all $\alpha>0$ we have [@rastjst] $${{\textup{H}}}_{\alpha}\bigl({\mathcal{E}}_{{\boldsymbol{Z}}}({\boldsymbol{\rho}})\bigr)\geq{{\textup{H}}}_{\alpha}({\boldsymbol{\rho}})
\ . \label{tepm}$$ It follows from Eqs. (\[hhcq\]) and (\[tepm\]) that $$H_{\alpha}({\boldsymbol{Z}};{\boldsymbol{\rho}})+H_{\alpha}\bigl({\boldsymbol{X}};{\mathcal{E}}_{{\boldsymbol{Z}}}({\boldsymbol{\rho}})\bigr)\geq
{{\textup{H}}}_{\alpha}({\boldsymbol{\rho}})+{{\textup{H}}}_{\alpha}\bigl({\mathcal{E}}_{{\boldsymbol{Z}}}({\boldsymbol{\rho}})\bigr)
\ . \label{hhcq1}$$ This inequality can be treated as an uncertainty relation expressed in terms of the quantum $\alpha$-entropies of ${\boldsymbol{\rho}}$ and ${\mathcal{E}}_{{\boldsymbol{Z}}}({\boldsymbol{\rho}})$. Certainty relations are formulated as upper bounds on the sum of considered entropies. At this stage, only simple bounds may be given. We merely recall that the quantum $\alpha$-entropy is not more than the $\alpha$-entropy of the completely mixed state. Thus, we obtain $$H_{\alpha}({\boldsymbol{Z}};{\boldsymbol{\rho}})+H_{\alpha}\bigl({\boldsymbol{X}};{\mathcal{E}}_{{\boldsymbol{Z}}}({\boldsymbol{\rho}})\bigr)\leq
2{\,}{{\textup{H}}}_{\alpha}({\boldsymbol{\varrho}}_{*})
\ , \label{hhcq2}$$ where the completely mixed state ${\boldsymbol{\varrho}}_{*}={\openone}/{{\textup{Tr}}}({\openone})$. We aim to obtain uncertainty and certainty relations connected with the purity ${{\textup{Tr}}}({\boldsymbol{\rho}}^{2})$ of the input state. In the following sections, more detailed relations for two successive measurements will be formulated in the qubit case.
In another scenario of successive measurements, the second measurement is performed on the post-first-measurement state conditioned on the actual measurement outcome. Let the first measurement has given the outcome $z$. The probability of this event is written as $p(z)={{\textup{Tr}}}({\boldsymbol{P}}_{z}{\,}{\boldsymbol{\rho}})$. According to the projection postulate, the state right after the measurement is described by the projector ${\boldsymbol{P}}_{z}$. In the second measurement, therefore, the outcome $x$ is obtained with the probability $p(x|z)={{\textup{Tr}}}({\boldsymbol{Q}}_{x}{\boldsymbol{P}}_{z})$.[^1] The latter is the conditional probability of outcome $x$ given that the previous measurement of ${\boldsymbol{Z}}$ has resulted in $z$. In our case, the function (\[hctm0\]) is obviously expressed as $$H_{\alpha}({\boldsymbol{X}}|z)=\frac{1}{1-\alpha}
\left(\sum_{{\,}x\in{{\textup{spec}}}({\boldsymbol{X}})}\bigl({{\textup{Tr}}}({\boldsymbol{Q}}_{x}{\boldsymbol{P}}_{z})\bigr)^{\alpha}
- 1\right)
\, . \label{haazb}$$ It should be emphasized that this quantity does not depend on ${\boldsymbol{\rho}}$. For the scenario considered, an amount of uncertainties is characterized by means of the conditional entropies $$\begin{aligned}
H_{\alpha}({\boldsymbol{X}}|{\boldsymbol{Z}};{\boldsymbol{\rho}})&=\sum_{z\in{{\textup{spec}}}({\boldsymbol{Z}})}
\bigl({{\textup{Tr}}}({\boldsymbol{P}}_{z}{\,}{\boldsymbol{\rho}})\bigr)^{\alpha}{\,}H_{\alpha}({\boldsymbol{X}}|z)
\ , \label{ahct1}\\
{\widetilde{H}}_{\alpha}({\boldsymbol{X}}|{\boldsymbol{Z}};{\boldsymbol{\rho}})&=\sum_{z\in{{\textup{spec}}}({\boldsymbol{Z}})}
{{\textup{Tr}}}({\boldsymbol{P}}_{z}{\,}{\boldsymbol{\rho}}){\,}H_{\alpha}({\boldsymbol{X}}|z)
\ . \label{ahct2}\end{aligned}$$ Taking $\alpha=1$, both the $\alpha$-entropies are reduced to the standard conditional entropy $H_{1}({\boldsymbol{X}}|{\boldsymbol{Z}};{\boldsymbol{\rho}})$. In Ref. [@bfs2014], the latter entropy was examined as a measure of uncertainties in successive measurements. In the following, we will give minimal and maximal values of the conditional entropies (\[ahct1\]) and (\[ahct2\]) in the qubit case.
First-scenario relations for successive qubit measurements {#sec4}
==========================================================
In this section, we will obtain uncertainty and certainty relations for a qubit within the first scenario. The formulation of uncertainty relations for successive measurements in terms of the Shannon entropies was given in Ref. [@bfs2014]. The authors of Ref. [@zzhang14] studied corresponding uncertainty relations in terms of the Rényi entropies. Although the Rényi and Tsallis entropies are closely connected, relations for one of them are not immediately applicable to other. In particular, the formula (\[exrt\]) does not allow to move between the Rényi-entropy and Tsallis-entropy formulations of the uncertainty principle. This fact also holds for the case of successive measurements. We will formulate uncertainty relations for two successive projective measurements in terms of Tsallis’ entropies.
Each density matrix in two dimensions can be represented in terms of its Bloch vector as [@nielsen] $${\boldsymbol{\rho}}=\frac{1}{2}{\,}\bigl({\openone}+\vec{r}\cdot\vec{{\boldsymbol{\sigma}}}\bigr)
\, . \label{bvrn}$$ By $\vec{{\boldsymbol{\sigma}}}$, we denote the vector of the Pauli matrices ${\boldsymbol{\sigma}}_{1}$, ${\boldsymbol{\sigma}}_{2}$, ${\boldsymbol{\sigma}}_{3}$. The three-dimensional vector $\vec{r}=(r_{1},r_{2},r_{3})$ obeys $|\vec{r}|\leq1$, with equality if and only if the state is pure. By calculations, the obtain the purity $${{\textup{Tr}}}({\boldsymbol{\rho}}^{2})=\frac{1+|\vec{r}|^{2}}{2}
\ . \label{bvpy}$$ The language of Bloch vectors is very convenient in description of qubit states and their transformations [@nielsen; @bengtsson]. Following Ref. [@zzhang14], we will aslo use this aproach in representing projectors of the observables ${\boldsymbol{Z}}$ and ${\boldsymbol{X}}$.
Without loss of generality, we assume the observables to be non-degenerate. Indeed, in two dimensions any degenerate observable is inevitably proportional to the identity operator. We will exclude this trivial case. Further simplification is reached by rescaling eigenvalues of the observables. Working in information-theoretic terms, we mainly deal with probability distributions. In such a consideration, we can turn observables to be dimensionless. Moreover, we can further shift eigenvalues of the observables without altering the probabilities. Of course, such actions are not appropriate for more traditional approach, using the mean value and the variance. Thus, the spectral decompositions are written as $$\begin{aligned}
{\boldsymbol{Z}}&=z_{+}{\boldsymbol{P}}_{+}+z_{-}{\boldsymbol{P}}_{-}
\ , \label{2dax}\\
{\boldsymbol{X}}&=x_{+}{\boldsymbol{Q}}_{+}+x_{-}{\boldsymbol{Q}}_{-}
\ . \label{2day}\end{aligned}$$ Taking the observables ${\boldsymbol{Z}}$ and ${\boldsymbol{X}}$ with eigenvalues $\pm1$, we will arrive at dimensionless spin observables. Any projector describes a pure state and, herewith, is represented by means of some Bloch vector. We introduce two unit vectors $\vec{p}$ and $\vec{q}$ such that $$\begin{aligned}
{\boldsymbol{P}}_{\pm}&=\frac{1}{2}{\,}\bigl({\openone}\pm\vec{p}\cdot\vec{{\boldsymbol{\sigma}}}\bigr)
\, , \label{pprn}\\
{\boldsymbol{Q}}_{\pm}&=\frac{1}{2}{\,}\bigl({\openone}\pm\vec{q}\cdot\vec{{\boldsymbol{\sigma}}}\bigr)
\, . \label{qqrn}\end{aligned}$$ When eigenvalues of each of the observables are $\pm1$, we simply have ${\boldsymbol{Z}}=\vec{p}\cdot\vec{{\boldsymbol{\sigma}}}$ and ${\boldsymbol{X}}=\vec{q}\cdot\vec{{\boldsymbol{\sigma}}}$. In the first measurement, we measure ${\boldsymbol{Z}}$ in the pre-measurement state (\[bvrn\]). The generated probability distribution is written as $$p(z=\pm1)=\frac{1\pm\vec{p}\cdot\vec{r}}{2}
\ . \label{fmpd}$$ In the considered scenario, information contained in the qubit after the first measurement is completely erased. By calculations, we now have $$\begin{aligned}
{\mathcal{E}}_{{\boldsymbol{Z}}}({\boldsymbol{\rho}})&={\boldsymbol{P}}_{+}{\,}{\boldsymbol{\rho}}{\,}{\boldsymbol{P}}_{+}+{\boldsymbol{P}}_{-}{\,}{\boldsymbol{\rho}}{\,}{\boldsymbol{P}}_{-}
\nonumber\\
&=\frac{1}{2}{\,}\bigl({\openone}+(\vec{p}\cdot\vec{r}){\,}\vec{p}\cdot\vec{{\boldsymbol{\sigma}}}\bigr)
\ , \label{clar}\end{aligned}$$ with the Bloch vector $(\vec{p}\cdot\vec{r}){\,}\vec{p}$. Here, the map turns the Bloch vector into its projection on $\vec{p}$. The density matrix (\[clar\]) describes the pre-measurement state of the measurement of ${\boldsymbol{X}}$. Similarly to Eq. (\[fmpd\]), we write generated probabilities in the form $$p(x=\pm1)=\frac{1\pm(\vec{q}\cdot\vec{p})(\vec{p}\cdot\vec{r})}{2}
\ . \label{smpd}$$ To avoid bulky expressions, we put the function of positive variable with the parameter $\alpha>0$, $$\eta_{\alpha}(\xi):=\frac{\xi^{\alpha}-\xi}{1-\alpha}
=-\xi^{\alpha}\ln_{\alpha}(\xi)
\ , \label{etdf}$$ including $\eta_{1}(\xi)=-\xi\ln\xi$. Due to the space isotropy, we can choose the frame of references in such a way that $\vec{p}=\vec{e}_{3}$. Then the entropic quantity (\[hhcq\]) is written as $$\begin{aligned}
&\sum_{m=\pm1}\eta_{\alpha}\!\left(\frac{1+m(\vec{p}\cdot\vec{r})}{2}\right)
+\sum_{n=\pm1}\eta_{\alpha}\!\left(\frac{1+n(\vec{q}\cdot\vec{p})(\vec{p}\cdot\vec{r})}{2}\right)
\nonumber\\
&=\sum_{m=\pm1}\eta_{\alpha}\!\left(\frac{1+m{\,}r_{3}}{2}\right)
+\sum_{n=\pm1}\eta_{\alpha}\!\left(\frac{1+n{\,}\mu{\,}r_{3}}{2}\right)
\, . \label{fnrz}\end{aligned}$$ For brevity, we denote $\mu=(\vec{q}\cdot\vec{p})\in[-1;+1]$. To obtain purity-based uncertainty and certainty relations, we will search the minimum and the maximum of (\[fnrz\]) under the restriction that the Bloch vector length $|\vec{r}|$ is fixed. Due to Eq. (\[bvpy\]), the purity of a quantum state then remains unchanged. The following statement takes place.
\[pan1\] Let the length $|\vec{r}|$ of the Bloch vector of ${\boldsymbol{\rho}}$ be fixed. For all $\alpha>0$, the sum of $\alpha$-entropies for successive measurements of non-degenerate observables ${\boldsymbol{Z}}$ and ${\boldsymbol{X}}$ is bounded from below as $$\begin{aligned}
&H_{\alpha}({\boldsymbol{Z}};{\boldsymbol{\rho}})+H_{\alpha}\bigl({\boldsymbol{X}};{\mathcal{E}}_{{\boldsymbol{Z}}}({\boldsymbol{\rho}})\bigr)
\nonumber\\
&\geq
\sum_{m=\pm1}\eta_{\alpha}\!\left(\frac{1+m{\,}|\vec{r}|}{2}\right)
+\sum_{n=\pm1}\eta_{\alpha}\!\left(\frac{1+n{\,}\mu{\,}|\vec{r}|}{2}\right)
\, . \label{minr}\end{aligned}$$ The equality in Eq. (\[minr\]) holds if and only if ${\boldsymbol{\rho}}$ commutes with ${\boldsymbol{Z}}$, i.e., ${\boldsymbol{\rho}}{\,}{\boldsymbol{Z}}={\boldsymbol{Z}}{\,}{\boldsymbol{\rho}}$. For all $\alpha>0$, the sum of $\alpha$-entropies is bounded from above as $$H_{\alpha}({\boldsymbol{Z}};{\boldsymbol{\rho}})+H_{\alpha}\bigl({\boldsymbol{X}};{\mathcal{E}}_{{\boldsymbol{Z}}}({\boldsymbol{\rho}})\bigr)\leq
2\ln_{\alpha}(2)
\ . \label{maxr}$$ The equality in Eq. (\[maxr\]) holds if and only if ${{\textup{Tr}}}({\boldsymbol{Z}}{\boldsymbol{\rho}})={{\textup{Tr}}}({\boldsymbol{Z}})/2$.
[**Proof.**]{} For the given $|\vec{r}|$, the component $r_{3}=(\vec{p}\cdot\vec{r})$ obeys $-|\vec{r}|\leq{r}_{3}\leq+|\vec{r}|$. Since the right-hand side of Eq. (\[fnrz\]) is an even function of $r_{3}$, we can restrict our consideration to the interval $r_{3}\in\bigl[0;|\vec{r}|\bigr]$. For all $\alpha>0$, the function (\[etdf\]) is concave. Thus, the right-hand side of Eq. (\[fnrz\]) is concave with respect to $r_{3}$. A concave function reaches the minimal value at one or more least points of the interval. To minimize the function (\[fnrz\]), we should therefore compare its values for $r_{3}=0$ and for $r_{3}=|\vec{r}|$. The latter actually leads to the minimum, whence the claim (\[minr\]) is proved. Indeed, substituting $r_{3}=0$ leads to the uniform distribution and, therefore, to the maximal value $\ln_{\alpha}(2)$ of each of two entropies. The last comment justifies the claim (\[maxr\]).
We shall now prove conditions for the equality. Substituting $r_{3}=\pm|\vec{r}|$, the right-hand side of Eq. (\[fnrz\]) is equal to the right-hand side of Eq. (\[minr\]). Some inspection shows that deviating $r_{3}$ from the points $\pm|\vec{r}|$ will certainty increase the term (\[fnrz\]). Thus, these points are the only case when the inequality (\[minr\]) is saturated. In this case, we have $\vec{r}\parallel\vec{p}$, whence the density matrix is diagonal with respect to the common eigenbasis of the projectors ${\boldsymbol{P}}_{\pm}$. Hence, the operators ${\boldsymbol{\rho}}$ and ${\boldsymbol{Z}}$ commute. To saturate the inequality (\[maxr\]), the two entropies must reach the maximal value. The only case is $r_{3}=0$, when the probability distributions $(1\pm{r}_{3})/2$ and $(1\pm\mu{\,}r_{3})/2$ are both uniform. Since $${{\textup{Tr}}}\bigl(
(\vec{p}\cdot\vec{{\boldsymbol{\sigma}}})(\vec{r}\cdot\vec{{\boldsymbol{\sigma}}})
\bigr)=2(\vec{p}\cdot\vec{r})
\ , \label{prbb}$$ for $\vec{r}\perp\vec{p}$ we obtain ${{\textup{Tr}}}({\boldsymbol{P}}_{\pm}{\boldsymbol{\rho}})=1/2$. Combining the latter with (\[2dax\]) gives ${{\textup{Tr}}}({\boldsymbol{Z}}{\boldsymbol{\rho}})={{\textup{Tr}}}({\boldsymbol{Z}})/2$. $\square$
The result (\[minr\]) is a Tsallis-entropy family of uncertainty relations for successive projective measurements of a qubit. The pre-measurement density matrix ${\boldsymbol{\rho}}$ describes a mixed state of the purity (\[bvpy\]). For the given $|\vec{r}|$, this purity is constant. In two dimensions, a geometrical description in terms of the Bloch vector is more convenient. The equality in (\[minr\]) takes place, if and only if $\vec{r}\parallel\vec{p}$. Among states of the fixed purity, minimal uncertainties are revealed by states, whose Bloch vector is collinear to the Bloch vector associated with the projectors on the eigenspaces of ${\boldsymbol{Z}}$. In terms of operators, this condition implies the commutativity of ${\boldsymbol{Z}}$ and ${\boldsymbol{\rho}}$. Further, the density matrix is then a fixed point of ${\mathcal{E}}_{{\boldsymbol{Z}}}$, i.e., ${\mathcal{E}}_{{\boldsymbol{Z}}}({\boldsymbol{\rho}})={\boldsymbol{\rho}}$ for $\vec{r}\parallel\vec{p}$.
We further note that the lower bound (\[hhcq1\]) is typically not saturated for $\vec{r}\parallel\vec{p}$. Indeed, the first entropic sum in the quantity (\[minr\]) actually becomes equal to the quantum $\alpha$-entropy of ${\boldsymbol{\rho}}$. This is not the case for the second one. In general, the second entropic sum in the formula (\[minr\]) is strictly larger than the quantum $\alpha$-entropy of ${\mathcal{E}}_{{\boldsymbol{Z}}}({\boldsymbol{\rho}})$. The equality takes place only for $\mu=\pm1$ that is equivalent to $\vec{q}\parallel\vec{p}$. In such a situation, the operators ${\boldsymbol{\rho}}$, ${\boldsymbol{Z}}$, and ${\boldsymbol{X}}$ are all diagonal in the same common eigenbasis. Then the picture becomes purely classical in character.
As we see, uncertainties in considered successive measurements are minimized for $\vec{r}\parallel\vec{p}$. It is natural to expect that the condition $\vec{r}\perp\vec{p}$ will lead to an opposite case of maximal uncertainties. In effect, this case actually lead to the maximal values to both the entropies of the left-hand side of (\[maxr\]). Then the quantum operation ${\mathcal{E}}_{{\boldsymbol{Z}}}$ maps the input ${\boldsymbol{\rho}}$ into the completely mixed state. Here, we have ${\mathcal{E}}_{{\boldsymbol{X}}}\circ{\mathcal{E}}_{{\boldsymbol{Z}}}({\boldsymbol{\rho}})={\mathcal{E}}_{{\boldsymbol{Z}}}({\boldsymbol{\rho}})={\boldsymbol{\varrho}}_{*}$. Thus, the upper bound (\[hhcq2\]) is actually saturated. When the two eigenvalues of ${\boldsymbol{Z}}$ are symmetric with respect to $0$, as for spin observables, the condition $\vec{r}\perp\vec{p}$ implies that the mean value of ${\boldsymbol{Z}}$ in the state ${\boldsymbol{\rho}}$ is zero, i.e., ${{\textup{Tr}}}({\boldsymbol{Z}}{\boldsymbol{\rho}})=0$. Indeed, in such a case we deal with the traceless observable.
Bounds for the second scenario of successive qubit measurements {#sec5}
===============================================================
In this section, we will obtain uncertainty and certainty relations for a qubit within the second scenario. Here, the second measurement is performed on the post-first-measurement state conditioned on the actual measurement outcome. Using the conditional Rényi entropy, corresponding uncertainty relations were derived in Ref. [@zzhang14]. To quantify uncertainties, we will use the conditional Tsallis entropies (\[hct1\]) and (\[hct2\]). They cannot be related immediately to the conditional Rényi entropy. So, a formulation in terms the conditional entropies (\[hct1\]) and (\[hct2\]) is of own interest. It turns out that the entropy (\[hct1\]) rather gives a more sensitive measure than (\[hct2\]). Entropic certainty bounds for successive measurements seem to be not studied in the literature.
We first measure the observable ${\boldsymbol{Z}}$ in the state (\[bvrn\]), for which the probabilities of outcomes is calculated according to Eq. (\[fmpd\]). If the first measurement has given the outcome $m$, then the post-first-measurement state is described by the projector ${\boldsymbol{P}}_{m}$ with the Bloch vector $m\vec{p}$. Then we perform the measurement of ${\boldsymbol{X}}$ with generating the probability distribution $$p(x=n|z=m)=\frac{1+n(\vec{q}\cdot{m}\vec{p})}{2}
\ , \label{gepd}$$ where $m,n=\pm1$. These quantities are conditional probabilities used in Eq. (\[haazb\]). For both $m=\pm1$, the entropic function (\[haazb\]) is expressed as $$H_{\alpha}({\boldsymbol{X}}|z=m)={{\textup{H}}}_{\alpha}\bigl({\mathcal{E}}_{{\boldsymbol{X}}}({\boldsymbol{P}}_{m})\bigr)
=\sum_{n=\pm1}\eta_{\alpha}\!\left(\frac{1+n{\,}\mu}{2}\right)
\, . \label{gehf}$$ Hence, the two conditional Tsallis entropies are represented as $$\begin{aligned}
H_{\alpha}({\boldsymbol{X}}|{\boldsymbol{Z}};{\boldsymbol{\rho}})&=
\sum_{m=\pm1}\left(\frac{1+m{\,}r_{3}}{2}\right)^{\!\alpha}
\sum_{n=\pm1}\eta_{\alpha}\!\left(\frac{1+n{\,}\mu}{2}\right)
\, , \label{fcen}\\
{\widetilde{H}}_{\alpha}({\boldsymbol{X}}|{\boldsymbol{Z}};{\boldsymbol{\rho}})&=
\sum_{n=\pm1}\eta_{\alpha}\!\left(\frac{1+n{\,}\mu}{2}\right)
\, . \label{scen}\end{aligned}$$ Thus, the second form of conditional $\alpha$-entropies does not depend on the input state ${\boldsymbol{\rho}}$. This entropic quantity is completely given by taking the observables ${\boldsymbol{Z}}$ and ${\boldsymbol{X}}$. Except for $\alpha=1$, the first conditional entropy (\[fcen\]) depends on the state ${\boldsymbol{\rho}}$ and both the observables. The right-hand side of Eq. (\[fcen\]) gives a general expression of the entropy. Let us examine an interval, in which this quantity ranges as a function of inputs of the fixed purity.
\[prop1\][Proposition]{}
\[pan2\] Let the length $|\vec{r}|$ of the Bloch vector of ${\boldsymbol{\rho}}$ be fixed. For $\alpha\in(0;1)$, the conditional $\alpha$-entropy for successive measurements of non-degenerate observables ${\boldsymbol{Z}}$ and ${\boldsymbol{X}}$ obeys $$\begin{aligned}
&\bigl(1+(1-\alpha){{\textup{H}}}_{\alpha}({\boldsymbol{\rho}})\bigr)\sum_{n=\pm1}\eta_{\alpha}\!\left(\frac{1+n{\,}\mu}{2}\right)
\leq{H}_{\alpha}({\boldsymbol{X}}|{\boldsymbol{Z}};{\boldsymbol{\rho}})
\ , \label{cea011}\\
&H_{\alpha}({\boldsymbol{X}}|{\boldsymbol{Z}};{\boldsymbol{\rho}})\leq2^{1-\alpha}\sum_{n=\pm1}\eta_{\alpha}\!\left(\frac{1+n{\,}\mu}{2}\right)
\, . \label{cea012}\end{aligned}$$ The lower bound (\[cea011\]) is reached if and only if ${\boldsymbol{\rho}}$ commutes with ${\boldsymbol{Z}}$, i.e., ${\boldsymbol{\rho}}{\,}{\boldsymbol{Z}}={\boldsymbol{Z}}{\,}{\boldsymbol{\rho}}$. The upper bound (\[cea012\]) is reached if and only if ${{\textup{Tr}}}({\boldsymbol{Z}}{\boldsymbol{\rho}})={{\textup{Tr}}}({\boldsymbol{Z}})/2$.
For $\alpha\in(1;\infty)$, the conditional $\alpha$-entropy for successive measurements of non-degenerate observables ${\boldsymbol{Z}}$ and ${\boldsymbol{X}}$ obeys $$\begin{aligned}
&2^{1-\alpha}\sum_{n=\pm1}\eta_{\alpha}\!\left(\frac{1+n{\,}\mu}{2}\right)\leq{H}_{\alpha}({\boldsymbol{X}}|{\boldsymbol{Z}};{\boldsymbol{\rho}})
\ , \label{cea021}\\
&H_{\alpha}({\boldsymbol{X}}|{\boldsymbol{Z}};{\boldsymbol{\rho}})
\leq\bigl(1+(1-\alpha){{\textup{H}}}_{\alpha}({\boldsymbol{\rho}})\bigr)\sum_{n=\pm1}\eta_{\alpha}\!\left(\frac{1+n{\,}\mu}{2}\right)
\, . \label{cea022}\end{aligned}$$ The lower bound (\[cea021\]) is reached if and only ${{\textup{Tr}}}({\boldsymbol{Z}}{\boldsymbol{\rho}})={{\textup{Tr}}}({\boldsymbol{Z}})/2$. The upper bound (\[cea022\]) is reached if and only if ${\boldsymbol{\rho}}{\,}{\boldsymbol{Z}}={\boldsymbol{Z}}{\,}{\boldsymbol{\rho}}$.
[**Proof.**]{} We need only find those least values that exactly bound the first sum in the right-hand side of Eq. (\[fcen\]). For brevity, we define the function $$g_{\alpha}(r_{3}):=\left(\frac{1+r_{3}}{2}\right)^{\!\alpha}+\left(\frac{1-r_{3}}{2}\right)^{\!\alpha}
\, . \label{dfga}$$ As this function is even, the aim is to find its minimum and maximum on the interval $r_{3}\in\bigl[0;|\vec{r}|\bigr]$.
For $\alpha\in(0;1)$, the function $r_{3}\mapsto{g}_{\alpha}(r_{3})$ is concave. So, we have $g_{\alpha}(r_{3})\leq2^{1-\alpha}=g_{\alpha}(0)$ by using Jensen’s inequality with the weights $1/2$. Hence, the claim (\[cea012\]) follows. Due to concavity, the minimum is reached for one of two least points of the interval $r_{3}\in\bigl[0;|\vec{r}|\bigr]$. So, the desired minimum is $$\begin{aligned}
g_{\alpha}\bigl(|\vec{r}|\bigr)&=
\left(\frac{1+|\vec{r}|}{2}\right)^{\!\alpha}+\left(\frac{1-|\vec{r}|}{2}\right)^{\!\alpha}
\nonumber\\
&=1+(1-\alpha){{\textup{H}}}_{\alpha}({\boldsymbol{\rho}})
\ , \label{gadf}\end{aligned}$$ as the eigenvalues of ${\boldsymbol{\rho}}$ are $\bigl(1\pm|\vec{r}|\bigr)/2$. Combining this with Eq. (\[fcen\]) finally gives Eq. (\[cea011\]). Focusing on the conditions for equality, $r_{3}=\pm|\vec{r}|$ is equivalent to $\vec{r}\parallel\vec{p}$, and $r_{3}=0$ is equivalent to $\vec{r}\perp\vec{p}$. The former implies ${\boldsymbol{\rho}}{\,}{\boldsymbol{Z}}={\boldsymbol{Z}}{\,}{\boldsymbol{\rho}}$, the latter implies ${{\textup{Tr}}}({\boldsymbol{Z}}{\boldsymbol{\rho}})={{\textup{Tr}}}({\boldsymbol{Z}})/2$.
For $\alpha\in(1;\infty)$, the function $r_{3}\mapsto{g}_{\alpha}(r_{3})$ is convex. Now, we have $g_{\alpha}(r_{3})\geq2^{1-\alpha}=g_{\alpha}(0)$ by using Jensen’s inequality with the weights $1/2$. By convexity, the maximum is reached for one of two least points of the interval $r_{3}\in\bigl[0;|\vec{r}|\bigr]$. In this case, the right-hand side of Eq. (\[gadf\]) represents the desired maximum. The conditions for equality are treated similarly. $\square$
The formulas (\[cea011\]) and (\[cea021\]) describe the minimal values of the conditional $\alpha$-entropy (\[fcen\]) for the states of fixed purity. The bounds (\[cea012\]) and (\[cea022\]) present the maximal values. Unlike Eq. (\[fcen\]), the inequalities (\[cea011\])–(\[cea022\]) do not depend on a direction of the Bloch vector. For the given observables ${\boldsymbol{Z}}$ and ${\boldsymbol{X}}$, the mentioned results involve the constant factor equal to Eq. (\[gehf\]). For the case $\alpha\in(0;1)$, the conditional entropy is minimized or maximized under the same conditions as the sum of two $\alpha$-entropies in the first scenario (see Proposition \[pan1\] above). For $\alpha\in(1;\infty)$, the conditions for the equality are swapped. For $\alpha=1$, both the forms of conditional entropies give $$H_{1}({\boldsymbol{X}}|{\boldsymbol{Z}};{\boldsymbol{\rho}})=\sum_{n=\pm1}\eta_{1}\!\left(\frac{1+n{\,}\mu}{2}\right)
\leq\ln2
\, . \label{fcen1}$$ Using the standard conditional entropy, this formula expresses an upper bound for the second scenario of two successive measurements. Overall, we can say the following. The conditional $\alpha$-entropy (\[fcen\]) seems to be more appropriate, since it depends also on the pre-measurement state. This state is not involved by the conditional $\alpha$-entropy (\[scen\]) including the standard case (\[fcen1\]). However, such a property is two-dimensional in character. In more dimensions, entropic functions of the form (\[haazb\]) will generally depend on the label $z$. Hence, the $\alpha$-entropy (\[ahct2\]) will be dependent on the pre-measurement state.
We shall now consider the entropic quantity (\[gehf\]). It reaches its maximal value $\ln_{\alpha}(2)$ for $\vec{p}\perp\vec{q}$, when $\mu=0$. Here, we have arrived at an interesting observation. For the given input ${\boldsymbol{\rho}}$, both the measures (\[fcen\]) and (\[scen\]) are maximized, when the eigenbases of ${\boldsymbol{Z}}$ and ${\boldsymbol{X}}$ are mutually unbiased. The condition $\vec{p}\perp\vec{q}$ actually implies that $\bigl|\langle{z}|x\rangle\bigr|=1/\sqrt{2}$ for all labels $z$ and $x$. For instance, this property takes place for eigenbases of any two of the three Pauli matrices. Such eigenstates are indistinguishable in the following sense. The detection of a particular basis state reveals no information about the state, which was prepared in another basis. Indeed, two possible outcomes are then equiprobable. As is known, this property is used in the BB84 protocol of quantum key distribution [@bb84].
Thus, formulation of certainty relations for successive measurement again emphasizes a role of mutual unbiasedness. The concept of mutually unbiased bases is naturally posed in arbitrary finite dimensions. If two bases are mutually unbiased, then the overlaps between any basis state in one basis and all basis states in the other are the same. Mutually unbiased bases have found use in many questions of quantum information. They also connected with important mathematical problems (see the review [@bz10] and references therein). Some of the above conclusions can be extended to an arbitrary dimensionality. The generalization is formulated as follows.
\[prop1\][Proposition]{}
\[pan3\] Let ${\boldsymbol{Z}}$ and ${\boldsymbol{X}}$ be two non-degenerate $d$-dimensional observables. For each density $d\times{d}$-matrix ${\boldsymbol{\rho}}$, the conditional entropies of successive measurements of ${\boldsymbol{Z}}$ and ${\boldsymbol{X}}$ obey $$\begin{aligned}
H_{\alpha}({\boldsymbol{X}}|{\boldsymbol{Z}};{\boldsymbol{\rho}})
&\leq{{\textup{Tr}}}\bigl({\mathcal{E}}_{{\boldsymbol{Z}}}({\boldsymbol{\rho}})^{\alpha}\bigr){\,}\ln_{\alpha}(d)
\ , \label{pr31}\\
{\widetilde{H}}_{\alpha}({\boldsymbol{X}}|{\boldsymbol{Z}};{\boldsymbol{\rho}})&\leq\ln_{\alpha}(d)
\ . \label{pr32}\end{aligned}$$ If the eigenbases of ${\boldsymbol{Z}}$ and ${\boldsymbol{X}}$ are mutually unbiased, then the equality is reached in both Eqs. (\[pr31\]) and (\[pr32\]). For strictly positive ${\boldsymbol{\rho}}$, mutual unbiasedness of the eigenbases is the necessary condition for equality.
[**Proof.**]{} By $\bigl\{|z\rangle\}$ and $\bigl\{|x\rangle\}$, we respectively denote eigenbases of the non-degenerate observables ${\boldsymbol{Z}}$ and ${\boldsymbol{X}}$. Recall that the Tsallis $\alpha$-entropy is not larger than $\ln_{\alpha}(d)$, where $d$ is the number of outcomes. For each eigenvalue $z$, we then have $$H_{\alpha}({\boldsymbol{X}}|z)\leq\ln_{\alpha}(d)
\ . \label{hxzml}$$ In the case considered, we clearly have $\sum_{z}p(z)^{\alpha}={{\textup{Tr}}}\bigl({\mathcal{E}}_{{\boldsymbol{Z}}}({\boldsymbol{\rho}})^{\alpha}\bigr)$. Combining the latter with Eq. (\[hxzml\]) provides the claims (\[pr31\]) and (\[pr32\]). We shall now proceed to the conditions for equality. The maximal entropic value $\ln_{\alpha}(d)$ is reached only for the uniform distribution, when the probabilities are all $1/d$. To saturate Eq. (\[hxzml\]) with the given $z$, we should therefore have $$\forall{\,}x\in{{\textup{spec}}}({\boldsymbol{X}}):
{\>}\bigl|\langle{x}|z\rangle\bigr|=\frac{1}{\sqrt{d}}
\ . \label{xtwb}$$ When ${\boldsymbol{\rho}}$ is strictly positive, $p(z)\neq0$ for all $z\in{{\textup{spec}}}({\boldsymbol{Z}})$. To reach the equality in both Eqs. (\[pr31\]) and (\[pr32\]), the condition (\[xtwb\]) should be provided for all $z\in{{\textup{spec}}}({\boldsymbol{Z}})$. Then the two eigenbases are mutually unbiased. $\square$
Although the BB84 scheme of quantum cryptography is primarily important, other protocols have been studied in the literature [@grtz02]. Some of them are based on mutually unbiased bases (see, e.g., Refs. [@bruss98; @mdg13]). We have seen above that the conditional entropies for a pair of successive projective measurements are maximized just in the case of mutual unbiasedness. Entropic uncertainty relations for several mutually unbiased bases were studied in Refs. [@ballester; @molm09; @MWB10; @rast13b]. It would be interesting to consider main questions of this section for many mutually unbiased bases. Such investigations may give an additional perspective of possible use of mutually unbiased bases in quantum information processing. It could be a theme of separate investigation.
Conclusion {#sec6}
==========
We have studied Tsallis-entropy uncertainty and certainty relations for two subsequent measurements of a qubit. Despite of very wide prevalence of the Heisenberg principle, there is no general consensus over its scope and validity. The following claim is commonly accepted. It is not possible to assign jointly exact values for two or more incompatible observables. There are several ways to fit this claim as a quantitative statement. Heisenberg’s original argument is adequately formulated in terms of noise and disturbance [@ozawa04]. Using the two scenarios of successive measurements, we are able to fit the question of measuring uncertainties in a different way. Such an approach may be more significant in the sense of its potential applications in studying protocols of quantum information. Indeed, subsequent manipulations with qubits rather deal with an output state of the latter stage. Uncertainty relations for two successive measurements were already examined in terms of the standard entropies [@bfs2014] and the Rényi entropies [@zzhang14]. At the same time, certainty relations for successive measurements seem to be not addressed in the literature.
The following two scenarios of successive measurements were considered. In the first scenario, a subsequent measurement is performed on the quantum state generated after the previous stage with completely erased information. In the second scenario, a subsequent measurement is performed on the post-measurement state conditioned on the actual measurement outcome. for both the scenarios, we derived uncertainty and certainty bounds on $\alpha$-entropic functions related to successive measurements of a pair of qubit observables. The conditions for equality in these bounds were obtained as well. Some of found results in the second scenario were extended to arbitrary finite dimensionality. They are connected with a frequently used property of mutually unbiased bases. In effect, the detection of a particular basis state reveals no information about the state, which was prepared in another basis. It would be interesting to study uncertainty and certainty relations for successive measurements in several mutually unbiased bases.
[77]{}
W. Heisenberg, Z. Phys. [**43**]{} (1927) 172.
M.J.W. Hall, Phys. Rev. A [**59**]{} (1999) 2602.
P. Busch, T. Heinonen, and P.J. Lahti, Phys. Rep. [**452**]{} (2007) 155.
E.H. Kennard, Z. Phys. [**44**]{} (1927) 326.
H.P. Robertson, Phys. Rev. [**34**]{} (1929) 163.
M. Ozawa, Ann. Phys. [**311**]{} (2004) 350.
M.D. Srinivas, Pramana–J. Phys. [**60**]{} (2003) 1137.
J. Distler and S. Paban, Phys. Rev. A [**87**]{} (2013) 062112.
S. Wehner and A. Winter, New J. Phys. [**12**]{} (2010) 025009.
D. Deutsch, Phys. Rev. Lett. [**50**]{} (1983) 631.
H. Maassen and J.B.M. Uffink, Phys. Rev. Lett. [**60**]{} (1988) 1103.
I. Bia[ł]{}ynicki-Birula and [Ł]{}. Rudnicki, [*Entropic Uncertainty Relations in Quantum Physics*]{}, In: [*Statistical Complexity*]{}, ed. K.D. Sen, Springer, Berlin (2011) 1.
Z. Pucha[ł]{}a, [Ł]{}. Rudnicki, and K. Życzkowski, J. Phys. A: Math. Theor. [**46**]{} (2013) 272002.
S. Friedland, V. Gheorghiu, and G. Gour, Phys. Rev. Lett. [**111**]{} (2013) 230401.
[Ł]{}. Rudnicki, Z. Pucha[ł]{}a, and K. Życzkowski, Phys. Rev. A [**89**]{} (2014) 052115.
K. Baek, T. Farrow, and W. Son, Phys. Rev. A [**89**]{} (2014) 032108.
J. Zhang, Y. Zhang, and C.-S. Yu, Quantum Inf. Process. (2015), in press. http://dx.doi.org/10.1007/s11128-015-0950-z
C. Tsallis, J. Stat. Phys. [**52**]{} (1988) 479.
J. Havrda and F. Charvát, Kybernetika [**3**]{} (1967) 30.
Z. Daróczy, Inform. Control [**16**]{} (1970) 36.
M. Gell-Mann and C. Tsallis, eds., [*Nonextensive Entropy – Interdisciplinary Applications*]{}, Oxford University Press, Oxford (2004).
A. Rényi, [*On Measures of Entropy and Information*]{}, In: [*Proc. 4th Berkeley Symposium on Mathematical Statistics and Probability*]{}, University of California Press, Berkeley, CA (1961) 547.
A.E. Rastegin, Commun. Theor. Phys. [**61**]{} (2014) 293.
I. Bengtsson and K. Życzkowski, [*Geometry of Quantum States: An Introduction to Quantum Entanglement*]{}, Cambridge University Press, Cambridge (2006).
A. Wehrl, Rev. Mod. Phys. [**50**]{} (1978) 221.
M. Ohya and D. Petz, [*Quantum Entropy and its Use*]{}, Springer-Verlag, Heidelberg (1993).
T.M. Cover and J.A. Thomas, [*Elements of Information Theory*]{}, John Wiley & Sons, New York (1991).
S.L. Braunstein and C.M. Caves, Phys. Rev. Lett. [**61**]{} (1988) 662.
F. Buscemi, M.J.W. Hall, M. Ozawa, and M.M. Wilde, Phys. Rev. Lett. [**112**]{} (2014) 050401.
S. Furuichi, J. Math. Phys. [**47**]{} (2006) 023302.
A.E. Rastegin, Kybernetika [**48**]{} (2012) 242.
J.S. Bell, Physics [**1**]{} (1964) 195.
A.E. Rastegin, Quantum Inf. Comput. [**14**]{} (2014) 0996.
A.E. Rastegin, Commun. Theor. Phys. [**62**]{} (2014) 320.
A.E. Rastegin, Ann. Phys. [**355**]{} (2015) 241.
M.A. Nielsen and I.L. Chuang, [*Quantum Computation and Quantum Information*]{}, Cambridge University Press, Cambridge (2000).
A.E. Rastegin, J. Stat. Phys. [**143**]{} (2011) 1120.
C.H. Bennett and G. Brassard, [*Quantum cryptography: public key distribution and coin tossing*]{}, In: [*Proc. IEEE International Conference on Computers, Systems, and Signal Processing, Bangalore, India*]{}, IEEE, New York (1984) 175.
T. Durt, B.-G. Englert, I. Bengtsson, and K. Życzkowski, Int. J. Quantum Inf. [**8**]{} (2010) 535.
N. Gisin, G. Ribordy, W. Tittel, and H. Zbinden, Rev. Mod. Phys. [**74**]{} (2002) 145.
D. Bru[ß]{}, Phys. Rev. Lett. [**81**]{} (1998) 3018.
M. Mafu, A. Dudley, S. Goyal, [*et al.*]{}, Phys. Rev. A [**88**]{} (2013) 032305.
M.A. Ballester and S. Wehner, Phys. Rev. A [**75**]{} (2007) 022319.
S. Wu, S. Yu, and K. M[ø]{}lmer, Phys. Rev. A [**79**]{} (2009) 022104.
P. Mandayam, S. Wehner, and N. Balachandran, J. Math. Phys. [**51**]{} (2010) 082201.
A.E. Rastegin, Eur. Phys. J. D [**67**]{} (2013) 269.
[^1]: In the non-degenerate case.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We tackle the task of question generation over knowledge bases. Conventional methods for this task neglect two crucial research issues: 1) the given predicate needs to be expressed; 2) the answer to the generated question needs to be definitive. In this paper, we strive toward the above two issues via incorporating diversified contexts and answer-aware loss. Specifically, we propose a neural encoder-decoder model with multi-level copy mechanisms to generate such questions. Furthermore, the answer aware loss is introduced to make generated questions corresponding to more definitive answers. Experiments demonstrate that our model achieves state-of-the-art performance. Meanwhile, such generated question can express the given predicate and correspond to a definitive answer.'
author:
- |
Cao Liu$^{1,2}$, Kang Liu$^{1,2}$, Shizhu He$^{1,2}$, Zaiqing Nie$^3$, Jun Zhao$^{1,2}$\
$^1$ National Laboratory of Pattern Recognition, Institute of Automation,\
Chinese Academy of Sciences, Beijing, 100190, China\
$^2$ University of Chinese Academy of Sciences, Beijing, 100049, China\
$^3$ Alibaba AI Labs, Beijing, 100029, China\
{cao.liu, kliu, shizhu.he, jzhao}@nlpr.ia.ac.cn\
[email protected]
bibliography:
- 'emnlp-2019.bib'
title: |
Generating Questions for Knowledge Bases\
via Incorporating Diversified Contexts and Answer-Aware Loss
---
Introduction {#Setion: Introduction}
============
Question Generation over Knowledge Bases (KBQG) aims at generating natural language questions for the corresponding facts on KBs, and it can benefit some real applications. Firstly, KBQG can automatically annotate question answering (QA) datasets. Secondly, the generated questions and answers will be able to augment the training data for QA systems. More importantly, KBQG can improve the ability of machines to actively ask questions on human-machine conversations [@D17-1090; @D18-1427]. Therefore, this task has attracted more attention in recent years [@P16-1056; @N18-1020].
Specifically, KBQG is the task of generating natural language questions according to the input facts from a knowledge base with triplet form, like $<$subject, predicate, object$>$. For example, as illustrated in Figure \[fig:example\], KBQG aims at generating a question “Which city is Statue of Liberty located in?" (Q3) for the input factual triplet “$<$Statue of Liberty, location/containedby[^1], New York City$>$". Here, the generated question is associated to the subject “*Statue of Liberty*" and the predicate `fb:location/containedby`) of the input fact, and the answer corresponds to the object “*New York City*".
![Examples of KBQG. We aims at generating questions like Q3 which expresses (matches) the given predicate and refers to a definitive answer.[]{data-label="fig:example"}](example.pdf){width="218pt"}
As depicted by , KBQG is required to transduce the triplet fact into a question about the subject and predicate, where the object is the correct answer. Therefore, it is a key issue for KBQG to correctly understand the knowledge symbols (subject, predicate and object in the triplet fact) and then generate corresponding text descriptions. More recently, some researches have striven toward this task, where the behind intuition is to construct implicit associations between facts and texts. Specifically, designed an encoder-decoder architecture to generate questions from structured triplet facts. In order to improve the generalization for KBQG, utilized extra contexts as input via distant supervisions [@Mintz:2009:DSR:1690219.1690287], then a decoder is equipped with attention and part-of-speech (POS) copy mechanism to generate questions. Finally, this model obtained significant improvements. Nevertheless, we observe that there are still two important research issues (RIs) which are not processed well or even neglected.
***RI-1:** The generated question is required to **express the given predicate** in the fact*. For example in Figure \[fig:example\], Q1 does not express (match) the predicate (`fb:location/containedby`) while it is expressed in Q2 and Q3. Previous work [@N18-1020] usually obtained predicate textual contexts through distant supervision. However, the distant supervision is noisy or even wrong (e.g. “X is the husband of Y" is the relational pattern for the predicate `fb:marriage/spouse`, so it is wrong when “X" is a woman). Furthermore, many predicates in the KB have no predicate contexts. We make statistic in the resources released by , and find that only 44% predicates have predicate textual context[^2]. Therefore, it is prone to generate error questions from such without-context predicates. ***RI-2:** The generated question is required to contain a **definitive answer***. A definitive answer means that one question only associates with a determinate answer rather than alternative answers. As an example in Figure \[fig:example\], Q2 may contain ambiguous answers since it does not express the refined answer type. As a result, different answers including “*United State*", “*New York City*", etc. may be correct. In contrast, Q3 refers to a definitive answer (the object “*New York City*" in the given fact) by restraining the answer type to a city. We believe that Q3, which expresses the given predicate and refers to a definitive answer, is a better question than Q1 and Q2. In previous work, only regarded a most frequently mentioned entity type as the textual context for the subject or object in the triplet. In fact, most answer entities have multiple types, where the most frequently mentioned type tends to be universal (e.g. a broad type “administrative region" rather than a refined type “US state" for the entity “*New York*"). Therefore, generated questions from may be difficult to contain definitive answers.
To address the aforementioned two issues, we exploit more **diversified contexts** for the given facts as textual contexts in an encoder-decoder model. Specifically, besides using predicate contexts from the distant supervision utilized by , we further leverage the domain, range and even topic for the given predicate as contexts, which are off-the-shelf in KBs (e.g. the range and the topic for the predicate `fb:location/containedby` are “location" and “containedby", respectively^\[omit\_domain\]^). Therefore, 100% predicates (rather than 44%^\[predicate\ context\ percentage\]^ of those in @N18-1020) have contexts. Furthermore, in addition to the most frequently mentioned entity type as contexts used by , we leverage the type that best describes the entity as contexts (e.g. a refined entity type[^3] “US state" combines a broad type “administrative region" for the entity “New York"), which is helpful to refine the entity information. Finally, in order to make full use of these contexts, we propose context-augmented fact encoder and multi-level copy mechanism (KB copy and context copy) to integrate diversified contexts, where the multi-level copy mechanism can copy from KB and textual contexts simultaneously. For the purpose of further making generated questions correspond to definitive answers, we propose the **answer-aware loss** by optimizing the cross-entropy between the generated question and answer type words, which is beneficial to generate precise questions. We conduct experiments on an open public dataset. Experimental results demonstrate that the proposed model using diversified textual contexts outperforms strong baselines (+4.5 BLEU4 score). Besides, it can further increase the BLEU score (+5.16 BLEU4 score) and produce questions associated with more definitive answers by incorporating answer-aware loss. Human evaluations complement that our model can express the given predicate more precisely. In brief, our main contributions are as follows:
\(1) We leverage diversified contexts and multi-level copy mechanism to alleviate the issue of incorrect predicate expression in traditional methods.
\(2) We propose an answer-aware loss to tackle the issue that conventional methods can not generate questions with definitive answers.
\(3) Experiments demonstrate that our model achieves state-of-the-art performance. Meanwhile, such generated question can express the given predicate and refer to a definitive answer.
{width="446pt"}
Task Description
================
We leverage textual contexts concerned with the triplet fact to generate questions over KBs. The task of KBQG can be formalized as follows:$$\begin{split}
P(Y|F)=\prod \nolimits_{t=1}^{|Y|}P(y_{t}|y_{<t},F, C)
\end{split}
\label{eq:3411}$$ where $F=(s,p,o)$ represents the subject ($s$), predicate ($p$) and object ($o$) of the input triplet, $C=\{\textbf{x}^{s}, \textbf{x}^{p}, \textbf{x}^{o}\}$ denotes a set of additional textual contexts, $Y=(y_1,y_2,...,y_{|Y|})$ is the generated question, $y_{<t}$ represents all previously generated question words before time step $t$.
Methodology
===========
Our model extends the encoder-decoder architecture [@cho-EtAl:2014:EMNLP2014] with three encoding modules and two copy mechanisms in the decoder. The model overview is shown in Figure \[fig:overall-structure\] along with its caption. It should be emphasized that we additionally design an answer-aware loss to make the generated question associated with a definitive answer (Sec. \[subsubsec:ans\_loss\]).
Context Encoder {#subsec:ctx_enc}
---------------
Inspired by the great success of transformer [@NIPS2017_7181] in sequence modeling [@shen2018disan], we adopt a transformer encoder to encode each textual context separately. Take the subject context $\textbf{x}^{s}$ as an example, $\textbf{x}^{s}=(x_{1}^{s}, x_{2}^{s},...,x_{|s|}^{s})$ is concatenated from diversified types for the subject, and $x_{i}^{s}$ is the $i$-th token in the subject context, $|s|$ stands for the length of the subject context. Firstly, $\textbf{x}^{s}$ is mapped into a query matrix $\textbf{Q}$, where $\textbf{Q}$ is constructed by summing the corresponding token embeddings and segment embeddings. Similar to BERT [@cotterell-eisner-2018-deep], segment embeddings are the same for tokens of $\textbf{x}^{s}$ but different for that of $\textbf{x}^{p}$ (predicate context) or $\textbf{x}^{o}$ (object context). Based on the query matrix, transformer encoder works as follows: $$\begin{gathered}
\footnotesize
\textbf{Q}_{j}=\textbf{QW}_{j}^{Q}, \textbf{K}_{j}=\textbf{KW}_{j}^{K}, \textbf{V}_{j}=\textbf{VW}_{j}^{V} \label{eq:subspace} \\
head_{j} = {\rm softmax} (\textbf{Q}_{j}\textbf{K}_{j}^{T} / \sqrt{d/h})\textbf{V}_{j} \label{eq:head} \\
\textbf{H}^{s}={\rm Concat}(head_1,head_2,...,head_{h})\textbf{W}_{0} \label{eq:concat} \\
\textbf{N}^{s} = {\rm LayerNorm}(\textbf{Q} + \textbf{H}^{s}) \label{eq:layernorm} \\
\textbf{C}^{s} = {\rm max}(0, \textbf{N}^{s} \textbf{W}_{1} + b_{1}) \textbf{W}_{2} + b_{2} \label{eq:FFN}\end{gathered}$$ where $\textbf{K}$ and $\textbf{V}$ are the key matrix and value matrix, respectively. It is called self-attention because $\textbf{K}$ and $\textbf{V}$ are equal to the query matrix $\textbf{Q} \in \mathbb{R} ^{|s|, d}$ in the encoding stage, where $d$ represents the number of hidden units. And $h$ denotes the number of the heads in multi-head attention mechanism of the transformer encoder. It first projects the input matrixes ($\textbf{Q}$, $\textbf{K}$, $\textbf{V}$) into subspaces $h$ times mapped by different linear projections $\textbf{W}_{j}^{Q}$, $\textbf{W}_{j}^{K}$, $\textbf{W}_{j}^{V} \in \mathbb{R} ^{|s|, d/h}$ ($j=1,2,...,h$) in Equation \[eq:subspace\]. And then $h$ projections perform the scaled dot-product attention to obtain the representation of each head in parallel (Equation \[eq:head\]). Representations for all parallel heads are concatenated together in Equation \[eq:concat\]. After residual connection, layer normalization (Equation \[eq:layernorm\]) and feed forward operation (Equation \[eq:FFN\]), we can obtain the subject context matrix $\textbf{C}^{s}=\{\textbf{c}_{1}^{s},\textbf{c}_{2}^{s},...,\textbf{c}_{|s|}^{s}\} \in \mathbb{R} ^{|s|, d}$.
Similarly, $\textbf{C}^{p}$ and $\textbf{C}^{o}$ are obtained from the same transformer encoder for the predicate and object, respectively.
Fact Encoder {#subsec:fact_enc}
------------
In contrast to general Sequence-to-Sequence (Seq2Seq) model [@sutskever2014sequence], the input fact is not a word sequence but instead a structured triplet $F=(s,p,o)$. We employ a fact encoder to transform each atom in the fact into a fixed embedding, and the embedding is obtained from a KB embedding matrix. For example, the subject embedding $\textbf{e}^{s} \in \mathbb{R} ^{d}$ is looked up from the KB embedding matrix $\textbf{E}_{f} \in \mathbb{R}^{k,d}$, where $k$ represents the size of KB vocabulary, and the size of KB embedding is equal to the number of hidden units ($d$) in Equation \[eq:head\]. Similarly, the predicate embedding $\textbf{e}^{p}$ and the object embedding $\textbf{e}^{o}$ are mapped from the KB embedding matrix $\textbf{E}_{f}$, where $\textbf{E}_{f}$ is pre-trained using *TransE* [@NIPS2013_5071] to capture much more fact information in previous work [@N18-1020]. In our model, $\textbf{E}_{f}$ can be pre-trained or randomly initiated (Details in Sec. \[subsec:Pre-trained KB\]).
Context-Augmented Fact Encoder {#subsec:ctx_fact_enc}
------------------------------
In order to combine both the context encoder information and the fact encoder information, we propose a context-augmented fact encoder which applies the gated fusion unit [@W18-2601] to integrate the context matrix and the fact embedding. For example, the subject context matrix $\textbf{C}^{s}=\{\textbf{c}_{1}^{s},\textbf{c}_{2}^{s},...,\textbf{c}_{|s|}^{s}\}$ and the subject embedding vector $\textbf{e}^{s}$ are integrated by the following gated fusion: $$\begin{gathered}
\textbf{f} = {\rm tanh}(\textbf{W}_f[\textbf{c}^{s},\textbf{e}^{s}]) \label{eq:f} \\
\textbf{g} = {\rm sigmoid}(\textbf{W}_g[\textbf{c}^{s},\textbf{e}^{s}]) \label{eq:g} \\
\textbf{h}^{s} = \textbf{g} \ \odot \ \textbf{f} + (1-\textbf{g}) \ \odot \ \textbf{e}^{s} \label{eq:fusion}\end{gathered}$$where $\textbf{c}^{s}$ is an attentive vector from $\textbf{e}^{s}$ to $\textbf{C}^{s}$, which is similar to . The attentive vector $\textbf{c}^{s}$ is combined with original subject embedding $\textbf{e}^{s}$ as a new enhanced representation $\textbf{f}$ (Equation \[eq:f\]). And then a learnable gate vector, $\textbf{g}$ (Equation \[eq:g\]), controls the information from $\textbf{c}^{s}$ and $\textbf{e}^{s}$ to the final augmented subject vector $\textbf{h}^{s} \in \mathbb{R} ^{d}$ (Equation \[eq:fusion\]), where $\odot$ denotes the element-wise multiplication. Similarly, the augmented predicate vector $\textbf{h}^{p}$ and the augmented object vector $\textbf{h}^{o}$ are calculated in the same way. Finally, the context-augmented fact representation $\textbf{H}_{f} \in \mathbb{R} ^{3,d}$ is the concatenation of augmented vectors as follows:$$\begin{split}
\textbf{H}_{f} = [\textbf{h}^{s}; \textbf{h}^{p}; \textbf{h}^{o}]
\end{split}
\label{eq:3411}$$
Decoder {#subsec:dec}
-------
The decoder aims at generating a question word sequence. As shown in Figure \[fig:overall-structure\], we also exploit the transformer as the basic block in our decoder. Then we use a multi-level copy mechanism (KB copy and context copy), which allows copying from KBs and textual contexts. Specifically, we first map the input of the decoder into an embedding representation by looking up word embedding matrix, then we use position embedding [@NIPS2017_7181] to enhance sequential information. Compared with the transformer encoder in Sec. \[subsec:ctx\_enc\], transformer decoder has an extra sub-layer: a fact multi-head attention layer, which is similar to Equation \[eq:subspace\]-\[eq:FFN\], where the query matrix is initiated with previous decoder sub-layer while both the key matrix and the value matrix are the augmented fact representation $\textbf{H}_{f}$. After feedforward and multiple transformer layers, we obtain the decoder state $\textbf{s}_t$ at time step $t$, and then $\textbf{s}_t$ could be leveraged to generate the target question sequence word by word.
As depicted in Figure \[fig:overall-structure\], we propose multi-level copy mechanism to generate question words. At each time step $t$, given decoder state $\textbf{s}_t$ together with input fact $F$, textual contexts $C$ and vocabulary $V$, the probabilistic function for generating any target question word $y_t$ is calculated as: $$\begin{gathered}
\begin{split}
P(y_{t}|\textbf{s}_t, \! y_{t-1}, \! F, \! C) \! = \! p_{genv} P_{genv}(y_{t}|\textbf{s}_t, \! V) +\!\\
p_{cpkb} P_{cpkb}(y_{t}|\textbf{s}_t, \!F) \! + \! p_{cpctx} P_{cpctx}(y_{t}|\textbf{s}_t,\! C) \label{eq:yt}
\end{split} \\
p_{genv}, p_{cpkb}, p_{cpctx} \! = \! {\rm softmax}([\textbf{s}_t, \textbf{y}_{t-1}]) \label{eq:mode}\end{gathered}$$where $genv$, $cpkb$ and $cpctx$ denote the vocab generation mode, the KB copy mode and the context copy mode, respectively. In order to control the balance among different modes, we employ a 3-dimensional switch probability in Equation \[eq:mode\], where $\textbf{y}_{t-1}$ is the embedding of previous generated word, $P_{\cdot}(\cdot | \cdot)$ indicates the probabilistic score function for generated target word of each mode. In the three probability score functions, $P_{vocab}$ is typically performed by a $softmax$ classifier over a fixed vocabulary $V$ based on the word embedding similarity, and the details of $P_{cpkb}$ and $P_{cpctx}$ are in the following.
### KB Copy
Previous study found that most questions contain the subject name or its aligns in SimpleQuestion [@D18-1051]. However, the predicate name and object name hardly appear in the question. Therefore, we only copy the subject name in the KB copy, where $P_{cpkb}(y_{t}|\textbf{s}_t,f)$, the probability of copying the subject name, is calculated by a neural network function with a multi-layer perceptron (MLP) projected from $\textbf{s}_t$.
### Context Copy
demonstrated the effectiveness of POS copy for the context. However, such a copy mechanism heavily relies on POS tagging. Inspired by the CopyNet [@P16-1154], we directly copy words in the textual contexts $C$, and it does not rely on any POS tagging. Specifically, the input sequence $\chi$ for the context copy is the concatenation of all words in the textual contexts $C$. Unfortunately, $\chi$ is prone to contain repeated words because it consists of rich contexts for subject, predicate and object. The repeated words in the input sequence tend to cause repetition problems in output sequences [@P16-1008]. We adopt the maxout pointer [@D18-1424] to address the repetition problem. Instead of summing all the probabilistic scores for repeated input words, we limit the probabilistic score of repeated words to their maximum score as Equation \[eq:cpctx\]:$$\begin{split}
P_{cpctx}(y_t)=\left\{\begin{matrix}
\underset{m,\, where\, \chi_{m}=y_{t}}{{\rm max}}\, sc_{t,m} \quad \quad y_{t} \in \chi \\
~~~~~~~~~~0 ~~~~~~~~~~~~~~~~~~~~~ otherwise
\end{matrix}\right.
\end{split}
\label{eq:cpctx}$$ $$\begin{split}
P_{cpctx}(y_t|.)\!\!=\!\!\left\{\!\begin{matrix}
\underset{m,\, where\, \chi_{m}=y_{t}}{{\rm max}} sc_{t,m} \quad y_{t} \! \in \! \chi \\
~~~~~~~~~~0 ~~~~~~~~~~~~~~~ otherwise
\end{matrix}\right.
\end{split}
\label{eq:cpctx}$$ where $\chi_{m}$ represents the $m$-th token in the input context sequence $\chi$, $sc_{t,m}$ is the probabilistic score of generating the token $\chi_m$ at time step $t$, and $sc_{t,m}$ is calculated by a softmax function over $\chi$.
Learning
--------
### Question-Aware Loss {#subsubsec:ques_loss}
It is totally differential for our model to obtain question words, and it can be optimized in an end-to-end manner by back-propagation. Given the input fact $F$, additional textual context $C$ and target question word sequence $Y$, the object function is to optimize the following negative log-likelihood:$$\begin{split}
\mathcal{L}_{ques\_loss}\!=\!\frac{-1}{|Y|}\! \sum _{t=1}^{|Y|} \! {\rm log}[P(y_t|\textbf{s}_t,\!y_{t-1},\!F,\!C)]
\end{split}
\label{eq:ques_loss}$$
The question-aware loss $\mathcal{L}_{ques\_loss}$ does not require any additional labels to optimize because the three modes share a same ${\rm softmax}$ classifier to keep a balance (Equation \[eq:mode\]), and they can learn to coordinate each other by minimizing $\mathcal{L}_{ques\_loss}$.
### Answer-Aware loss {#subsubsec:ans_loss}
It is able to generate questions similar to the labeled questions by optimizing the question-aware loss $\mathcal{L}_{ques\_loss}$. However, there is an ambiguous problem in the annotated questions where the questions have alternative answers rather than determinate answers [@D18-1051]. In order to make generated questions correspond to definitive answers, we propose a novel answer-aware loss. By answer-aware loss, we aim at generating an answer type word in the question, which contributes to generating a question word matching the answer type. Formally, the answer-aware loss is in the following: $$\begin{split}
\mathcal{L}_{ans\_loss}=\underset{a_n,a_n \in A}{{\rm min}} \ \ \underset{y_t,y_t \in Y}{{\rm min}} \ H_{a_n,y_t}
\end{split}
\label{eq:ans_loss}$$ where $A=\{a_n\}_{n=1}^{|A|}$ is a set of answer type words. We treat object type words as the answer type words because the object is the answer. $H_{a_n,y_t}$ denotes the cross entropy between the answer type word $a_n$ and the generated question word $y_t$. Finally, the minimum cross entropy is regarded as the answer-aware loss $\mathcal{L}_{ans\_loss}$. Optimizing $\mathcal{L}_{ans\_loss}$ means that the model aims at generating an answer type word in the generated question sequence. For example, the model tends to generate Q3 rather than Q2 in Figure \[fig:example\], because Q3 contains an answer type word—“city". Similarly, $\mathcal{L}_{ans\_loss}$ could be optimized in an end-to-end manner, and it can integrate $\mathcal{L}_{ques\_loss}$ by a weight coefficient $\lambda$ to the total loss as follows: $$\begin{split}
\mathcal{L}_{total\_loss}=\mathcal{L}_{ques\_loss} + \lambda \mathcal{L}_{ans\_loss}
\end{split}
\label{eq:toal_loss}$$
Experiment {#Setion: Experiment}
==========
Experimental Settings {#subsec:exp_setting}
---------------------
### Experimental Data Details
We conduct experiments on the SimpleQuestion dataset [@DBLP:journals/corr/BordesUCW15], and there are 75910/10845/21687 question answering pairs (QA-pairs) for training/validation/test. In order to obtain **diversified contexts**, we additionally employ domain, range and topic of the predicate to improve the coverage of predicate contexts. In this way, 100% predicates (rather than 44%^\[predicate\ context\ percentage\]^ of those in @N18-1020) have contexts. For the subject and object context, we combine the most frequently mentioned entity type [@N18-1020] with the type that best describe the entity^\[note:notable\_type\]^. The KB copy needs subject names as the copy source, and we map entities with their names similar to those in . The data details are in Appendix A and submitted Supplementary Data.
### Evaluation Metrics {#subsubsec:metrics}
Following [@P16-1056; @N18-1020], we adopt some word-overlap based metrics (WBMs) for natural language generation including BLEU-4 [@Papineni:2002:BMA:1073083.1073135], ROUGE$_{\rm L}$ [@Lin:2004] and METEOR [@W14-3348]. However, such metrics still suffer from some limitations [@novikova-EtAl:2017:EMNLP2017]. Crucially, it might be difficult for them to measure whether generated questions that express the given predicate and refer to definitive answers. To better evaluate generated questions, we run two further evaluations as follows. (1) ***Predicate identification***: Following , we employ annotators to judge whether the generated question expresses the given predicate in the fact or not. The score for predicate identification is the percentage of generated questions that express the given predicate. (2) ***Answer coverage***: We define a novel metric called answer coverage to identify whether the generated question refers to a definitive answer. Specifically, answer coverage is obtained by automatically calculating the percentage of questions that contain answer type words, and answer type words are object contexts (entity types for the object are regarded as answer type words). Furthermore, it is hard to automatically evaluate the naturalness of generated questions. Following , we adopt human evaluation to measure the naturalness by a score of 0-5.
### Comparison with State-of-the-arts
We compare our model with following methods.
\(1) *Template*: A baseline in , it randomly chooses a candidate fact $F_c$ in the training data to generate the question, where $F_c$ shares the same predicate with the input fact.
\(2) **: We compare our methods with the single placeholder model, which performs best in . (3) **: We compare our methods with the model utilizing copy actions, the best performing model in . Although this model is designed to a zero-shot setting (for unseen predicates and entity type), it has good abilities to generate better questions (on known or unknown predicates and entity types) represented in the additional context input and SPO copy mechanism.
### Implementation Details
To make our model comparable to the comparison methods, we keep most parameter values the same as . We utilize RMSProp algorithm with a decreasing learning rate (0.001), batch size (200) to optimize the model. The size of KB embeddings is 200, and KB embeddings are pre-trained by TransE [@NIPS2013_5071]. The word embeddings are initialized by the pre-trained Glove word vectors[^4] with 200 dimensions. In the transformer, we set the hidden units $d$ to 200, and we employ 4 paralleled attention head and a stack of 5 identical layers. We set the weight ($\lambda$) of the answer-aware loss to 0.2.
Overall Comparisons {#subsec:Overall Comparisons}
-------------------
[p[2.6cm]{}<p[1.2cm]{}<p[1.3cm]{}<p[1.3cm]{}<]{} Model $\quad \ \ $ & BLEU4 & ROUGE$\rm _{L}$ & METEOR Template $\quad \ \ $ & 31.36 & \* & 33.12 & 33.32 & \* & 35.38 & 36.56 & 58.09 & 34.41 Our Model $\quad \ \ \,$ & 41.09 & 68.68 & 47.75 $\ \ \,$ Our Model$_{\rm ans\_loss}$ & **41.72** & **69.31** & **48.13**
In Table \[tab: overall performance\], we compare our model with the typical baselines on word-overlap based metrics. It is evident that our model is remarkably better than baselines on all metrics, where the BLEU4 score increases 4.53 compared with the strongest baseline [@N18-1020]. Especially, incorporating answer-aware loss (the last line in Table \[tab: overall performance\]) further improves the performance (+5.16 BLEU4).
Performances on Predicate Identification {#subsec:Predicate Identification}
----------------------------------------
[p[2.8cm]{}<p[2.5cm]{}<]{} Model & Pred. Identification & 53.5 & 71.5 Our Model$_{\rm ans\_loss}$ & **75.5**
To evaluate the ability of our model on predicate identification, we sample 100 generated questions from each model, and then two annotators are employed to judge whether the generated question expresses the given predicate. The Kappa for inter-annotator statistics is 0.611, and p-value for all scores is less than 0.005. As shown in Table \[tab: pred\_identification\], we can see that our model has a significant improvement in the predicate identification.
Performances on Answer Coverage — The Effectiveness of Answer-Aware Loss {#subsec:Answer Coverage}
------------------------------------------------------------------------
[p[2.6cm]{}<p[1.0cm]{}<p[1.2cm]{}<p[1.3cm]{}<]{} Model & $\lambda$ & BLEU4 & ${\rm Ans_{cov}}$ & 0 & 36.56 & 59.49 Our Model& 0 & 41.09 & 61.65 Our Model$_{\rm ans\_loss}$ & 0.05 & 41.55 & 62.27 Our Model$_{\rm ans\_loss}$ & 0.2 & **41.72** & 64.23 Our Model$_{\rm ans\_loss}$ & 0.5 & 41.57 & **65.50** Our Model$_{\rm ans\_loss}$ & 1.0 & 41.34 & 65.25
Table \[tab: answer-aware loss\] reports performances on BLUE4 and answer coverage (Ans$_{\rm cov}$). We can obtain that: (1) When answer-aware loss is not leveraged ($\lambda=0$), advantages of performance are obvious in our model. Note that the answer coverage is 55.23 on the human-labeled questions. Although our model does not explicitly capture answer information, it still obtains a high answer coverage, which may be because our diversified contexts contain rich answer type words. (2) To demonstrate the effectiveness of answer-aware loss, we set the weight of answer-aware loss ($\lambda$) to 0.05/0.2/0.5/1.0 (the last four lines in Table \[tab: answer-aware loss\]). It can be seen that our model, incorporating answer-aware loss, has a significant improvement on answer coverage while there is no performance degradation on BLEU4 compared with $\lambda=0$, which indicates that answer-aware loss contributes to generating better questions. Especially, the generated questions are more precise because they refer to more definitive answers with high Ans$_{\rm cov}$.
\(3) It tends to correspond to alternative answers (object in the triplet fact) for some predicates such as `fb:location/containedby`, while other predicates (e.g. `fb:person/gender`) may refer to a definitive answer. To investigate our model, by incorporating answer-aware loss, does not generate an answer type word in a mandatory way, we found 20.5% predicate corresponds to the generated questions without answer type words when our model obtains the highest Ans$_{\rm cov}$ ($\lambda$=0.5), and it is very close to 21.7% for the one in human-annotated questions. This demonstrates that the answer-aware loss does not force all predicates to generate questions with answer type words.
The Effectiveness of Generated Questions for Enhancing QA over KBs {#subsec:Question Answering}
------------------------------------------------------------------
[p[4.1cm]{}<p[1.5cm]{}<]{} Data Type & Accuracy human-labeled data & 68.97
+ gen\_data [@P16-1056] & 68.53 + gen\_data [@N18-1020] & 69.13 + gen\_data (Our Model$_{\rm ans\_loss}$) & **69.57**
Previous experiments demonstrate that our model is able to deliver more precise questions. To further prove the effectiveness of our model, we will see how useful the generated questions are for training a question answering system over knowledge bases. Specifically, we combine human-labeled data with the same amount (15180) of model-generated data to a typical QA system [@N18-2047]. The accuracy of QA is shown in Table \[tab: Question Answering\]. We can observe that adding generative questions may weaken the performance of QA (drop from 68.97 to 68.53 in Table \[tab: Question Answering\]). Our generated questions achieve the best performance on QA system. It indicates that our model generates more precise question and has improved QA performances greatly.
Ablation Study {#subsec:Ablation Study}
--------------
[p[3.0cm]{}<p[1.15cm]{}<p[1.15cm]{}<p[1.2cm]{}<]{} Model $\quad \ \ $ & BLEU4 & ROUGE$\rm _{L}$ & METEOR Our Model$_{\rm ans\_loss}$ & **41.72** & **69.31** & **48.13** w/o context copy & 41.27 & 68.36 & 47.54 w/o KB copy & 41.04 & 68.66 & 47.72 w/o answer-aware loss& 41.09 & 68.68 & 47.75 w/o diversified contexts & 40.53 & 68.52 & 47.66
In order to validate the effectiveness of model components, we remove some important components in our model, including context copy, KB copy, answer-aware loss and diversified contexts. The results are shown in Table \[tab: Ablation\]. We can see that removing any component brings performance decline on all metrics. It demonstrates that all these components are useful. Specifically, the last line in Table \[tab: Ablation\], replacing diversified contexts with contexts used in , has more obvious performance degradation.
Performances on Naturalness {#subsec:Human Evaluation}
---------------------------
[p[2.8cm]{}<p[2.5cm]{}<]{} Model & Naturalness & 2.96 & 2.23 Our Model$_{\rm ans\_loss}$ & **3.56**
Human evaluation is important for generated questions. Following , we sample 100 questions from each system, and then two annotators measure the naturalness by a score of 0-5. The Kappa coefficient for inter-annotator is 0.629, and p-value for all scores is less than 0.005. As shown in Table \[tab: naturalness\], perform poorly on naturalness, while our model obtains the highest score on naturalness, which demonstrates our model can deliver more natural questions than baselines.
Discussion
----------
### Without Pre-trained KB Embeddings {#subsec:Pre-trained KB}
[p[2.58cm]{}<p[1.0cm]{}<p[1.13cm]{}<p[1.33cm]{}<p[1.36cm]{}<]{} Model& TransE & BLEU4 & ROUGE$\rm _{L}$ & METEOR & True & 36.56 & 58.09 & 34.41 & False & 33.67 & 55.57 & 33.20 Our Model$_{\rm ans\_loss}$ & True & 41.72 & 69.31 & 48.13 Our Model$_{\rm ans\_loss}$ & False & 41.55 & 68.59 & 47.52
Pre-trained KB embeddings may provide rich structured relational information among entities. However, it heavily relies on large-scale triplets, which is time and resource-intensive. To investigate the effectiveness of pre-trained KB embedding for KBQG, we report the performance of KBQG whether using pre-trained KB embeddings by simply applying TransE. Table \[tab: pre-train\] shows that the performance of KBQG is degraded without TransE embeddings. In comparison, obtain obvious degradation on all metrics while there is only a slight decline in our model. We believe that it may owe to the context-augmented fact encoder since our model drops to 40.87 on the BLEU4 score without context-augmented fact encoder and transE embeddings.
### The Effectiveness of Generated Questions for Enhancing Question Answering over Knowledge Bases {#sec:Question Answering}
[p[4.1cm]{}<p[1.5cm]{}<]{} Data Type & Accuracy human-labeled data & 68.97
+ gen\_data (Serban et al., 2016) & 68.53 + gen\_data (Elsahar et al., 2018) & 69.13 + gen\_data (Our Model$_{\rm ans\_loss}$) & **69.57**
Previous experiments demonstrate that our model can deliver more precise questions. To further prove the effectiveness of our model, we will see how useful the generated questions are for training a question answering system over knowledge bases. Specifically, we combine human-labeled data with the same amount of model-generated data to a typical QA system (Mohammed et al., 2018). The accuracy of QA is shown in Table \[tab: Question Answering\]. We can observe that adding generative questions may weaken the performance of QA (drop from 68.97 to 68.53 in Table \[tab: Question Answering\]). Our generated questions achieve the best performance on the QA system. It indicates that our model generates more precise question and has improved QA performances greatly.
### Speed {#subsec:Speed}
![Performance on valid data through epochs, where “base" is the method in Elsahar et al. (2018).[]{data-label="fig:epoch"}](epoch.pdf){width="6.6cm"}
In order to further explore the convergence speed, we plot the performances on valid data through epochs in Figure \[fig:epoch\]. Our model has much more information to learn, and it may have a bad impact on the convergence speed. Nevertheless, our model can copy KB elements and textual context simultaneously, which may accelerate the convergence speed. As demonstrated in Figure \[fig:epoch\], our model achieves the best performances on almost epochs. After about 6 epochs, performances on our model become stable and convergent.
### Case Study {#subsec:case_study}
![Examples of questions by different models.[]{data-label="fig:case"}](case.pdf){width="7cm"}
Figure \[fig:case\] lists referenced question and generated questions by different models. It can be seen that our generated questions can better express the target predicate such as ID 1 (marked as ). In ID 2, although all questions express the target predicate correctly, only our question refers to a definitive answer since it contains an answer type word “city" (marked as **bold**). It should be emphasized that the questions, generated by our method with answer-aware loss, do not always contain answer type words (ID 1 and 3).
[|p[2.0cm]{}<|p[2.0cm]{}<|p[2.0cm]{}<|]{} Eval Problem & Label Problem & Error Predicate\
57 & 16 & 14\
Related Work
============
Our work is inspired by a large number of successful applications using neural encoder-decoder frameworks on NLP tasks such as machine translation [@conf/ssst/ChoMBB14] and dialog generation [@journals/corr/VinyalsL15]. Our work is also inspired by the recent work for KBQG based on encoder-decoder frameworks. first proposed a neural network for mapping KB facts into natural language questions. To improve the generalization, introduced extra contexts for the input fact, which achieved significant performances. However, these contexts may make it difficult to generate questions that express the given predicate and associate with a definitive answer. Therefore, we focus on the two research issues: expressing the given predicate and referring to a definitive answer for generated questions. Moreover, our work also borrows the idea from copy mechanisms. Point network [@NIPS2015_5866] predicted the output sequence directly from the input, and it can not generate new words while CopyNet [@P16-1154] combined copying and generating. proposed to copy elements in the table (KB). exploited POS copy action to better capture textual contexts. To incorporate advantages from above copy mechanisms, we introduce KB copy and context copy which can copy KB element and textual context, and they do not rely on POS tagging.
Conclusion and Future Work
==========================
In this paper, we focus on two crucial research issues for the task of question generation over knowledge bases: generating questions that express the given predicate and refer to definitive answers rather than alternative answers. For this purpose, we present a neural encoder-decoder model which integrates diversified off-the-shelf contexts and multi-level copy mechanisms. Moreover, we design an answer-aware loss to generate questions that refer to definitive answers. Experiments show that our model achieves state-of-the-art performance on automatic and manual evaluations. For future work, we investigate error cases by analyzing the error distributions of 100 examples. We find that most generated questions (51%) are judged by the human to correctly express the input facts, but they unfortunately obtain low scores on the widely used metrics. It implies that it is still intractable to evaluate generated questions. Although we additionally evaluate on predicate identification and answer coverage, these metrics may be coarse and deserve further study.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work is supported by the National Natural Science Foundation of China (No.61533018), the Natural Key R&D Program of China (No.2018YFC0830101), the National Natural Science Foundation of China (No.61702512) and the independent research project of National Laboratory of Pattern Recognition. This work was also supported by CCF-Tencent Open Research Fund.
Input Data — Contexts for Predicate and Entity {#sec:Input Data}
==============================================
In order to improve the reproducibility of this paper, we release the context files for predicate and entity in the supplementary materials. The *“Predicate Context.xlsx"* file is the context for the predicate, and *“Entity Context.xlsx"* file is the context for the subject and the object. The file descriptions are as follows:
1\. *Predicate Context.xlsx*
*predicate* — predicate in Freebase
*dep\_w* — relation (predicate) pattern obtained from
*domain/range/topic* — domain, range and topic in Freebase for the predicate, note that only one is reserved for the same tokens, and this is our additionally added predicate context. In this way, 100% predicates (rather than 44%^\[predicate\ context\ percentage\]^ of those in @N18-1020) have contexts.
2\. *Entity Context.xlsx*
*entity* — entity ID in Freebase
*entity name* — entity name for the entity
*type* — the most frequently mentioned entity type as contexts used by
*notable type* — the most notable type for the entity, and this is our additionally added entity context for the subject and object.
The Effectiveness of Generated Questions for Enhancing Question Answering over Knowledge Bases {#sec:Question Answering}
==============================================================================================
[p[4.1cm]{}<p[1.5cm]{}<]{} Data Type & Accuracy human-labeled data & 68.97
+ gen\_data [@P16-1056] & 68.53 + gen\_data [@N18-1020] & 69.13 + gen\_data (Our Model$_{\rm ans\_loss}$) & **69.57**
Previous experiments demonstrate that our model is able to deliver more precise questions. To further prove the effectiveness of our model, we will see how useful the generated questions are for training a question answering system over knowledge bases. Specifically, we combine human-labeled data with the same amount (15180) of model-generated data to a typical QA system [@N18-2047]. The accuracy of QA is shown in Table \[tab: Question Answering\]. We can observe that adding generative questions may weaken the performance of QA (drop from 68.97 to 68.53 in Table \[tab: Question Answering\]). Our generated questions achieve the best performance on QA system. It indicates that our model generates more precise question and has improved QA performances greatly.
Speed {#subsec:Speed}
=====
![Performance on valid data through epochs, where “base" is the method in .[]{data-label="fig:epoch"}](epoch.pdf){width="6.6cm"}
In order to further explore the convergence speed, we plot the performances on valid data through epochs in Figure \[fig:epoch\]. Our model has much more information to learn, and it may have a bad impact on the convergence speed. Nevertheless, our model can copy KB elements and textual context simultaneously, which may accelerate the convergence speed. As demonstrated in Figure \[fig:epoch\], our model achieves the best performances on almost epochs. After about 6 epochs, performances on our model become stable and convergent.
[^1]: We omit the domain of the predicate for sake of brevity.\[omit\_domain\]
[^2]: We map the “prop\_text\_evidence.csv" file to the “property.vocab" file in \[predicate context percentage\]
[^3]: We obtain such representative entity types through the predicate `fb:topic/notable_types` in freebase.\[note:notable\_type\]
[^4]: http://nlp.stanford.edu/data/glove.6B.zip
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The aim of this paper is to investigate strong convergence of modified truncated Euler-Maruyama method for neutral stochastic differential delay equations introduced in Lan (2018). Strong convergence rates of the given numerical scheme to the exact solutions at fixed time $T$ are obtained under local Lipschitz and Khasminskii-type conditions. Moreover, convergence rates over a time interval $[0,T]$ are also obtained under additional polynomial growth condition on $g$ without the weak monotonicity condition (which is usually the standard assumption to obtain the convergence rate). Two examples are presented to interpret our conclusions.'
author:
- |
Guangqiang Lan[^1]and Qiushi Wang\
School of Science, Beijing University of Chemical Technology, Beijing 100029, China
title: '**Strong convergence rates of modified truncated EM methods for neutral stochastic differential delay equations [^2]**'
---
**MSC 2010:** 60H10, 65C30, 65L20.
**Key words:** neutral stochastic differential delay equations; local Lipschitz condition; modified truncated Euler-Maruyama method; strong convergence rate.
Introduction
============
Let $(\Omega,\mathscr{F},\{\mathscr{F}_t\}_{t\geq 0},P)$ be a complete filtered probability space satisfying usual conditions (i.e. $\{\mathscr{F}_t\}_{t\geq 0}$ is right continuous and $\mathscr{F}_0$ contains all $P$-null sets). Consider the following neutral stochastic differential delay equations (short for NSDDEs)
$$\label{nsdde}d(x(t)-D(x(t-\tau))
=f(x(t),x(t-\tau))dt+g(x(t),x(t-\tau))dB_t$$
on $t\ge0$ with initial value $\{x(\theta):-\tau\le\theta\le0\}=\xi\in C_{\mathscr{F}_0}^b([-\tau,0],\mathbb{R}^d)$ (the family of all $\mathscr{F}_0$ measurable bounded $C([-\tau,0],\mathbb{R}^d)$-valued random variables), for any $\xi\in C_{\mathscr{F}_0}^b([-\tau,0],\mathbb{R}^d), p\ge1$ define $||\xi||_p=(\mathbb{E}(\sup_{-\tau\le\theta\le0}|\xi(\theta)|^p))^{\frac{1}{p}}$, $\{B_t,t\geq0\}$ is an $n$-dimensional standard $\mathscr{F}_t$-Brownian motion, $D:x\in\mathbb{R}^d\mapsto D(x)\in\mathbb{R}^d$, $f:(x,y)\in\mathbb{R}^d\times\mathbb{R}^d\mapsto f(x,y)\in\mathbb{R}^d$ and $g:(x,y)\in\mathbb{R}^d\times\mathbb{R}^d\mapsto g(x,y)\in\mathbb{R}^d\otimes\mathbb{R}^n$ are measurable functions.
Recently, such neutral stochastic differential delay equations (short for NSDDEs) have been found more and more applications in many fields such as control theory, electrodynamics, biomathematics and so on. However, most NSDDEs can not be solved explicitly except some special ones. Thus numerical methods for NSDDEs (\[nsdde\]) have been playing more and more important roles.
The convergence of the numerical methods for NSDDEs (\[nsdde\]) have been discussed intensively by many researchers, for example, Gan et. al [@GSZ] investigated mean square convergence of stochastic $\theta$ method under global Lipschitz condition, [@ZG] studied Mean square convergence of one-step methods under the same assumptions, [@M] considered convergence in probability of the backward Euler approximate solution for a class of stochastic differential equations with constant delay, Zhang et.al considered strong convergence of the partially truncated Euler-Maruyama method for a class of stochastic differential delay equations in [@ZSL], there are also many other literatures concerning with this topic, see e.g. [@EXT; @Kloeden; @mao2; @MY; @ZSL1; @ZW].
Recently, in [@mao], Mao developed a new explicit numerical simulation method, called truncated EM method. Strong convergence theory were established there under local Lipschitz condition plus the Khasminskii-type condition. And then he obtained sufficient conditions for the strong convergence rate of it in [@mao1]. Motivated by these two works, Lan and Xia introduced in [@LX] modified truncated Euler-Maruyama (MTEM) method and obtained the strong convergence rate under given conditions. Then in [@L], the author generalized the MTEM method from SDEs cases to the NSDDEs cases and obtain asymptotic exponential stability of it under given conditions. However, the strong convergence of the MTEM method is still not known, which is the main topic of this paper.
Although strong convergence of the given numerical methods are considered in many papers such as [@BT; @EXT; @GSZ; @mao; @YM; @ZG] and so on, the convergence rates are not known. Some other papers considered strong convergence rates of the given numerical methods under weak monotonicity condition. For example, in Guo et al [@GMY], Assumption 5.1 is necessary to obtain the convergence rate of the truncated EM method at fixed time $T$, in Tan and Yuan [@TY], A8 is needed to obtain the convergence rate of the truncated EM method over a time interval $[0,T]$, for the weak monotonicity, one can also see [@mao1; @LX; @TY1] and so on. In this paper, we will consider the strong convergence rates of MTEM methods to exact solutions both at fixed time $T$ and over a time interval $[0,T]$ without such weak monotonicity conditions.
The organization of the paper is as the following. In Section 2, the MTEM method for NSDDE is introduced, and main results are presented. In Section 3, some useful lemmas are presented to prove the convergence theorems. In Section 4, convergence rates at fixed time $T$ are obtained. The convergence rates over the time interval $[0,T]$ will be proved with additional polynomial growth condition on $g$ in Section 5. Then in Section 6, two examples are presented to interpret the Theorems. We will conclude our paper in Section 7.
The settings and main results
=============================
Assume that both the coefficients $f$ and $g$ in (\[nsdde\]) are locally Lipschitz continuous, that is, for each $R$ there is $L_R>0$ (depending on $R$) such that $$\label{local}|f(x,y)-f(\bar{x},\bar{y})|
\vee|g(x,y)-g(\bar{x},\bar{y})|\le L_{R}(|x-\bar{x}|+|y-\bar{y}|)$$ for all $|x|\vee|y|\vee|\bar{x}|\vee|\bar{y}|\le R>0$. Here the norm of a matrix $A$ is denoted by $|A|=\sqrt{\textrm{trace}(A^\textrm{T}A)}$.
Assume also that there is a positive constant $u\in(0,1)$ such that
$$\label{Lip}|D(x)-D(y)|\le u|x-y|,\ \forall x,y\in\mathbb{R}^d$$
It is well known that there is a unique strong solution (might explode at finite time) to equation (\[nsdde\]) under conditions (\[local\]) and (\[Lip\]), see e.g. [@MY].
As interpreted in [@LX], we can always choose $\Delta^*>0$ small enough and a strictly positive decreasing function $h:(0,\Delta^*]\to(0,\infty)$ such that $$\label{tiaoj}\lim_{\Delta\to0}h(\Delta)=\infty\ \textrm{and}\ \lim_{\Delta\to0}L_{h(\Delta)}^4\Delta=0.$$
For any $\Delta\in(0,\Delta^*),$ we define the the modified truncated function of $f$ as the following: $$\label{dy}f_\Delta(x,y)=\left\{\begin{array}{ll} f(x,y),\qquad\qquad\qquad\ |x|\vee|y|\le h(\Delta),\\
\frac{|x|\vee|y|}{h(\Delta)} f\left(\frac{h(\Delta)}{|x|\vee|y|}(x,y)\right),|x|\vee|y|>h(\Delta).
\end{array}
\right.$$ $g_\Delta$ is defined in the same way as $f_\Delta$. Here $f(a(x,y))\equiv f(ax,ay)$ for any $a\in \mathbb{R}, x,y\in\mathbb{R}^d.$
It is obvious that the functions $f_\Delta$ and $g_\Delta$ defined above are different from the truncated functions defined in [@GMY].
We have defined the discrete MTEM method in [@L]. However, we recall it here for readers’ convenience.
Let $\Delta$ be a stepsize such that $\tau=m\Delta$ for some positive integer $m$. Then by using $f_\Delta$ and $g_\Delta$, we can define the MTEM method $X_k$ of (\[nsdde\]) as the following: $$\label{num}\aligned X_{k+1}-D(X_{k+1-m})&=X_k-D(X_{k-m})+f_\Delta(X_k,X_{k-m})\Delta\\&\quad
+g_\Delta(X_k,X_{k-m})\Delta B_k,\quad k=0,1,2,\cdots,\\&
X_k=\xi(k\Delta),\quad k=-m,-m+1,\cdots,0.\endaligned$$ Here $\Delta B_k=B((k+1)\Delta)-B(k\Delta)$ is the increment of the $n$-dimensional standard Brownian motion.
The two versions of the continuous-time MTEM solutions are defined as the following:
$$\label{num1}\bar{x}_\Delta(t)=\sum_{k=-m}^\infty X_k1_{[k\Delta,(k+1)\Delta)}(t),\quad t\ge-\tau,$$
and $x_\Delta(t)=\xi(t),$ $t\in[-\tau,0],$ $$\label{num2}\aligned x_\Delta(t)&=D(\bar{x}_\Delta(t-\tau))+\xi(0)-D(\xi(-\tau))+\int_0^tf_\Delta(\bar{x}_\Delta(s),\bar{x}_\Delta(s-\tau))ds\\&
\quad+\int_0^tg_\Delta(\bar{x}_\Delta(s),\bar{x}_\Delta(s-\tau))dB(s),\quad t\ge0.\endaligned$$
Obviously, $x_\Delta(k\Delta)=\bar{x}_\Delta(k\Delta)=X_k$ for all $k\ge0.$
To study the strong convergence of continuous version of MTEM (\[num2\]), let us first consider the following condition:
Assume that there exist positive constants $K$ and $p>2$ such that $$\label{dandiao}
2\langle
x-aD(\frac{1}{a}y),f(x,y)\rangle+(p-1)|g(x,y)|^2\le K(1+|x|^2+|y|^2)$$ holds for all $x,y\in\mathbb{R}^d, a\in(0,1].$
Notice that when $a=1,$ (\[dandiao\]) reduces to the well known Khasminskii condition $$\label{dandiao1}
2\langle
x-D(y),f(x,y)\rangle+(p-1)|g(x,y)|^2\le K(1+|x|^2+|y|^2).$$
Suppose for fixed $\Delta$ ($\tau=m\Delta$) the initial value $\xi$ satisfies $$\label{chuzhi}\mathbb{E}\sup_{-m\le k\le -1}\sup_{k\Delta\le s\le (k+1)\Delta}|\xi(s)-\xi(k\Delta)|^q\le \hat{K}\Delta^\frac{q}{2}$$ for $2<q<p$.
Now we are ready to state our first result on the strong convergence rate for MTEM method at fixed time $T.$
\[conv\] Assume that (\[local\]), (\[Lip\]), (\[dandiao\]) and (\[chuzhi\]) hold for some $2<q<p$. If there exist $0<\Delta_0\ (\le\Delta^*)$ and $h(\Delta)$ such that (\[tiaoj\]) and $$\label{tj}
h(\Delta)\ge(L^{2q}_{h(\Delta)}\Delta^{\frac{q}{2}})^{-\frac{1}{p-q}}$$ holds for any $\Delta\le\Delta_0$, then the continuous-time MTEM methods satisfy $$\label{shou}\mathbb{E}|x(T)-x_\Delta(T)|^q\le C(q,T)L^{2q}_{h(\Delta)}\Delta^{q/2}\ \textrm{and}\ \mathbb{E}|x(T)-\bar{x}_\Delta(T)|^q\le C(q,T)L^{2q}_{h(\Delta)}\Delta^{q/2}.$$
For the convergence rates over the time interval $[0,T],$ we have to introduce an additional assumption.
Suppose there exist $r\ge2$ and $\bar{K}>0$ such that $$\label{gzeng}
|g(x,y)|^2\le\bar{K}(1+|x|^r+|y|^r), \forall x,y\in \mathbb{R}^d.$$
\[conv1\] Assume that all conditions in Theorem \[conv\] hold. If (\[gzeng\]) holds for some $r$ satisfies $2\le r<p-2$ and $2< q\le p-r$ then there exists $C(q,T)$ (independent of $\Delta$) such that $$\label{shou1}\mathbb{E}\sup_{0\le t\le T}|x(t)-x_\Delta(t)|^q\le C(q,T)L^{2q}_{h(\Delta)}\Delta^{q/2}$$ and if further $2<q<4$, then $$\label{shou2}
\mathbb{E}\sup_{0\le t\le T}|x(t)-\bar{x}_\Delta(t)|^q\le C(q,T)L^{q}_{h(\Delta)}\Delta^{q/2-1}.$$
In [@GMY], the authors considered strong convergence of the truncated EM method for SDDEs. However, they only obtained the strong convergence rate at fixed time $T$, while the strong convergence rate over a time interval $[0,T]$ is not considered. Moreover, the weak monotonicity condition is needed (see Assumption 5.1 in [@GMY]). In [@TY], the authors obtained the strong convergence rate over a time interval $[0,T]$, but they also need the weak monotonicity condition A8 (similar to Assumption 5.1 in [@GMY]). We only need the local Lipschitz condition (\[local\]), Khasminskii-type condition (\[dandiao\]) and (\[gzeng\]) to make sure the numerical scheme $x_{\Delta}(t)$ strongly converges to $x(t)$ on $[0,T]$ in the sense (\[shou1\]) with the rate $C(q,T)L^{2q}_{h(\Delta)}\Delta^{q/2}$.
Some useful lemmas
==================
\[l0\] Suppose the local Lipschitz condition (\[local\]) holds. Then for any fixed $\Delta>0$, $$\label{global}|f_\Delta(x,y)-f_\Delta(\bar{x},\bar{y})|
\vee|g_\Delta(x,y)-g_\Delta(\bar{x},\bar{y})|\le 5L_{h(\Delta)}(|x-\bar{x}|+|y-\bar{y}|)$$ for any $x,y,\bar{x},\bar{y}\in\mathbb{R}^d$.
For the proof, see [@L] Lemma 3.1.
\[l1\] For $\Delta$ small enough, condition (\[dandiao\]) implies $$\label{bijin}\aligned 2\langle x-D(y),f_\Delta(x,y)\rangle+(p-1)|g_\Delta(x,y)|^2&\le 2K(1+|x|^2+|y|^2)\endaligned$$ for any $x,y\in\mathbb{R}^d.$
**Proof** On one hand, (\[bijin\]) holds naturally by (\[dandiao\]) and the definitions of $f_\Delta$ and $g_\Delta$ if $|x|\vee|y|\le h(\Delta)$.
On the other hand, if $|x|\vee|y|>h(\Delta)$, then $$\begin{array}{*{20}{l}}
2\langle x-D(y),f_\Delta(x,y)\rangle+|g_\Delta(x,y)|^2 &=2\left\langle x-D(y),\frac{|x|\vee|y|}{h(\Delta)}f\left(\frac{h(\Delta)}{|x|\vee|y|}x,\frac{h(\Delta)}{|x|\vee|y|}y\right)\right\rangle\\
&\quad+\frac{|x|^2\vee|y|^2}{h^2(\Delta)}\left|g\left(\frac{h(\Delta)}{|x|\vee|y|}x,\frac{h(\Delta)}{|x|\vee|y|}y\right)\right|^2\\
&=2\langle x-D(y),\frac{1}{a}f(ax,ay)\rangle+\frac{1}{a^2}|g(ax,ay)|^2\\&
=\frac{1}{a^2}(2\langle ax-aD(\frac{ay}{a}),f(ax,ay)\rangle+|g(ax,ay)|^2)
\end{array}$$ where $a=\frac{h(\Delta)}{|x|\vee|y|}$. Since $h(\Delta)\ge1$ for sufficiently small $\Delta,$ then by using (\[dandiao\]), it follows that $$\begin{array}{*{20}{l}}
2\langle x-D(y),f_\Delta(x,y)\rangle+|g_\Delta(x,y)|^2&\le\frac{1}{a^2}\cdot K(1+|ax|^2+|ay|^2)\\
&=\frac{K_1(|x|^2+|y|^2)}{h^2(\Delta)}+K(|x|^2+|y|^2)\\&
\le2K(1+|x|^2+|y|^2)
\end{array}$$ as required. $\square$
Now let us state the following two important lemmas. First, we have
\[ju\] Under conditions (\[local\]), (\[Lip\]) and (\[dandiao\]), the NSDDE (\[nsdde\]) has a unique global solution $x(t)$ and, moreover, there exists constant $C(p,T)$ (independent of $\Delta$) such that $$\sup_{0\le t\le T}\mathbb{E}|x(t)|^p\le C(p,T)<\infty,\ \forall T>0.$$ If we define the stopping time $$\tau_R=\inf\{t\ge0,|x(t)|\ge R\},\ \inf\emptyset=\infty,$$ then for any $T>0,$ $$P(\tau_R\le T)\le\frac{C}{R^p}.$$
**Proof** By Itô’s formula, we have $$\aligned\mathbb{E}|x(t)-D(x(t-\tau))|^p&\le \mathbb{E}|\xi(0)-D(\xi(-\tau))|^p+p\mathbb{E}\int_0^t|x(s)-D(x(s-\tau))|^{p-2}\\&
\qquad\qquad\qquad\qquad\qquad\qquad\qquad\times F(x(s),x(s-\tau))ds,
\endaligned$$ where $F(x,y)=\langle x-D(y),f(x,y)\rangle+\frac{p-1}{2}|g(x,y)|^2.$
Then by (\[dandiao\]), it follows that $$\aligned&\quad\mathbb{E}|x(t)-D(x(t-\tau))|^p\\&
\le \mathbb{E}|\xi(0)-D(\xi(-\tau))|^p+p K\mathbb{E}\int_0^t|x(s)-D(x(s-\tau))|^{p-2}\\&
\quad\qquad\qquad\qquad\qquad\qquad\qquad\times(1+|x(s)|^2+|x(s-\tau)|^2)ds\\&
\le \mathbb{E}|\xi(0)-D(\xi(-\tau))|^p+p K\mathbb{E}\int_0^t(1+|x(s)-D(x(s-\tau))|^{p})ds\\&
\quad+pK\mathbb{E}\int_0^t\left[\frac{p-2}{p}|x(s)-D(x(s-\tau))|^{p}+|x(s)|^p+|x(s-\tau)|^p\right]ds.
\endaligned$$ We have used Young’s inequality in the last inequality.
Thus $$\aligned\mathbb{E}|x(t)-D(x(t-\tau))|^p&
\le \mathbb{E}|\xi(0)-D(\xi(-\tau))|^p+pKT+pK||\xi||^p_p\\&\quad
+ (2p-2)K\mathbb{E}\int_0^t|x(s)-D(x(s-\tau))|^{p}ds\\&
\quad+2pK\mathbb{E}\int_0^t|x(s)|^pds.\\&
\le \mathbb{E}|\xi(0)-D(\xi(-\tau))|^p+pKT+pK(1+2pu^p)||\xi||^p_p\\&\quad
+[(2p-2)K(1+pu^2)+2pK]\mathbb{E}\int_0^t|x(s)|^pds.
\endaligned$$
Then we have $$\label{sqj}
\aligned \sup_{0\le s\le t}\mathbb{E}|x(s)-D(x(s-\tau))|^p&\le \mathbb{E}|\xi(0)-D(\xi(-\tau))|^p+pKT+pK(1+2pu^p)||\xi||^p_p\\&\quad
+[(2p-2)K(1+pu^2)+2pK]\int_0^t\sup_{0\le r\le s}\mathbb{E}|x(r)|^pds.\endaligned$$
On the other hand, for any $c>0$, $$\label{sqj1}\aligned\sup_{0\le s\le t}\mathbb{E}|x(s)|^p&\le (1+c)^{p-1}\sup_{0\le s\le t}\mathbb{E}|x(s)-D(x(s-\tau))|^p\\&
\quad+(\frac{1+c}{c})^{p-1}u^p(||\xi||^p_p+\sup_{0\le s\le t}\mathbb{E}|x(s)|^p).\endaligned$$
Since $0<u<1$, then we can take $c$ large enough such that $(\frac{1+c}{c})^{p-1}u^p<1$. So $$\label{sqj2}\aligned\sup_{0\le s\le t}\mathbb{E}|x(s)|^p&\le \frac{(1+c)^{p-1}\sup_{0\le s\le t}\mathbb{E}|x(s)-D(x(s-\tau))|^p+(\frac{1+c}{c})^{p-1}u^p||\xi||^p_p}{1-(\frac{1+c}{c})^{p-1}u^p}.\endaligned$$
Gronwall lemma and (\[sqj\]) and (\[sqj2\]) imply that $$\sup_{0\le t\le T}\mathbb{E}|x(t)|^p\le C(p,T),$$ as required.
Now let us prove $$P(\tau_R\le T)\le\frac{C}{R^p}.$$
Since $$\mathbb{E}|x(t\wedge\tau_R)|^p\ge R^pP(\tau_R\le T),$$ so we only need to prove $$\label{sqj5}\sup_{0\le t\le T}\mathbb{E}|x(s\wedge\tau_R)|^p\le C.$$
As in the above proof, we have $$\label{sqj30}
\aligned \sup_{0\le s\le t}\mathbb{E}|y(s\wedge\tau_R)|^p&\le \mathbb{E}|\xi(0)-D(\xi(-\tau))|^p+pKT+pK(1+2pu^p)||\xi||^p_p\\&\quad
+[(2p-2)K(1+pu^2)+2pK]\int_0^t\sup_{0\le r\le s}\mathbb{E}|x(r\wedge\tau_R)|^pds,\endaligned$$ where $y(s\wedge\tau_R)=x(s\wedge\tau_R)-D(x(s\wedge\tau_R-\tau))$ and $$\label{sqj40}\aligned\sup_{0\le s\le t}\mathbb{E}|x(s\wedge\tau_R)|^p&\le \frac{(1+c)^{p-1}\sup_{0\le s\le t}\mathbb{E}|y(s\wedge\tau_R)|^p+(\frac{1+c}{c})^{p-1}u^p||\xi||^p_p}{1-(\frac{1+c}{c})^{p-1}u^p}.\endaligned$$
Gronwall lemma, (\[sqj30\]) and (\[sqj40\]) yield the required (\[sqj5\]). $\square$
As a similar result of Lemma \[ju\], we have the following moment property for the MTEM method (\[num2\]).
\[temju\] Assume that (\[local\]), (\[Lip\]) and (\[dandiao\]) hold for $p>2$. Then there exist $0<\Delta_0\le\Delta^*$ and a constant $C(p,T)>0$ (independent of $\Delta$) such that for any $\Delta\in(0,\Delta_0]$, the MTEM method (\[num2\]) satisfies $$\label{sj}\sup_{0<\Delta\le\Delta_0}\sup_{0\le t\le T}\mathbb{E}|x_\Delta(t)|^p\le C(p,T)<\infty,\ \forall T>0.$$
Define the stopping time $$\rho_{\Delta,R}=\inf\{t\ge0,|x_\Delta(t)|\ge R\}.$$ Then for any $R>|x_0|$ and $\Delta\in(0,\Delta^*)$ ($\Delta^*$ small enough), we have $$\label{sj1}P(\rho_{\Delta,R}\le T)\le\frac{C}{R^p}.$$
**Proof** Let us first prove (\[sj\]).
Denote $y_\Delta(t)=x_\Delta(t)-D(\bar{x}_\Delta(t-\tau))$, $\bar{y}_\Delta(t)=\bar{x}_\Delta(t)-D(\bar{x}_\Delta(t-\tau))$. By Itô formula and Lemma \[l1\], for any $0\le t\le T,$ $$\aligned&\quad\mathbb{E}(|y_\Delta(t)|^p)\\&\le|y_\Delta(0)|^p+\frac{p}{2}\mathbb{E}\int_0^{t}|y_\Delta(s)|^{p-2}
(2\langle \bar{y}_\Delta(s),f_\Delta(\bar{x}_\Delta(s),\bar{x}_\Delta(s-\tau))\rangle\\&
\qquad\qquad\qquad\qquad\qquad+(p-1)|g_\Delta(\bar{x}_\Delta(s),\bar{x}_\Delta(s-\tau))|^2)ds\\&
\quad+p\mathbb{E}\int_0^{t}|y_\Delta(s)|^{p-2}
\langle x_\Delta(s)-\bar{x}_\Delta(s),f_\Delta(\bar{x}_\Delta(s),\bar{x}_\Delta(s-\tau))\rangle ds\\&
\le|y_\Delta(0)|^p+\frac{p}{2}\mathbb{E}\int_0^{t}|y_\Delta(s)|^{p-2}\cdot 2K(1+|\bar{x}_\Delta(s)|^2+|\bar{x}_\Delta(s-\tau)|^2)ds\\&
\quad+p\mathbb{E}\int_0^{t}|y_\Delta(s)|^{p-2}
|x_\Delta(s)-\bar{x}_\Delta(s)|\cdot (5L_{h(\Delta)}(|\bar{x}_\Delta(s)|+|\bar{x}_\Delta(s-\tau)|)+|f(0,0)|)ds\\&
\le|x_0|^p+C_p\mathbb{E}\int_0^{t}(1+|y_\Delta(s)|^{p}+|\bar{x}_\Delta(s)|^p)ds\\&
\quad+C_p\mathbb{E}\int_0^{t}(|y_\Delta(s)|^{p}+L_{h(\Delta)}^p|x_\Delta(s)-\bar{x}_\Delta(s)|^p+
|\bar{x}_\Delta(s)|^p+|\bar{x}_\Delta(s-\tau)|^p)ds
\\&\quad+C_p|f(0,0)|\mathbb{E}\int_0^{t}(1+|y_\Delta(s)|^{p}+|x_\Delta(s)-\bar{x}_\Delta(s)|^p)ds\\&
\le C_p+C_p\mathbb{E}\int_0^{t}|y_\Delta(s)|^{p}ds+C_p\mathbb{E}\int_0^{t}\sup_{0\le r\le s}\mathbb{E}|\bar{x}_\Delta(r)|^pds\\&
\quad+(C_p L_{h(\Delta)}^p+1)\int_0^{t}\sup_{0\le r\le s}\mathbb{E}|x_\Delta(r)-\bar{x}_\Delta(r)|^pds.\endaligned$$
Notice that for any $0\le s\le t,$ there exists $k\le[\frac{t}{\Delta}]$ such that $k\Delta\le s<(k+1)\Delta.$ Thus $$x_\Delta(s)-\bar{x}_\Delta(s)=x_\Delta(s)-X_k=f_\Delta(X_k,X_{k-m})(s-k\Delta)+g_\Delta(X_k,X_{k-m})(B(s)-B(k\Delta)).$$
So we have $$\aligned\mathbb{E}|x_\Delta(s)-\bar{x}_\Delta(s)|^p&\le C_p\left[\Delta^p\mathbb{E}(|f_\Delta(X_k,X_{k-m})|^p)\right.\\&
\qquad\left.+\mathbb{E}(|g_\Delta(X_k,X_{k-m})|^p)\mathbb{E}(|B(s)-B(k\Delta)|^p|\mathscr{F}_{k\Delta})\right].\endaligned$$
Since $f_\Delta$ and $g_\Delta$ satisfy the global Lipschitz condition (\[global\]), and notice that $B(t)-B(k\Delta)$ is independent of $\mathscr{F}_{k\Delta},$ then $$\aligned\mathbb{E}|x_\Delta(s)-\bar{x}_\Delta(s)|^p&\le C_p\big(\Delta^p\mathbb{E}(5L_{h(\Delta)}(|X_k|+|X_{k-m}|)+|f(0,0)|)^p\\&\quad
+\mathbb{E}(5L_{h(\Delta)}(|X_k|+|X_{k-m}|)+|g(0,0)|)^p\Delta^\frac{p}{2}\big)\\&
\le C_pL_{h(\Delta)}^p\Delta^p(\mathbb{E}(|X_k|^p)+\mathbb{E}(|X_{k-m}|^p))+C_p\Delta^p|f(0,0)|^p\\&\quad
+C_pL_{h(\Delta)}^p\Delta^\frac{p}{2}(\mathbb{E}(|X_k|^p)+\mathbb{E}(|X_{k-m}|^p))+C_p\Delta^\frac{p}{2}|g(0,0)|^p.\endaligned$$
Therefore, for any $t\le T,$ $$\label{k2}\aligned\sup_{0\le s\le t}\mathbb{E}|x_\Delta(s)-\bar{x}_\Delta(s)|^p&\le C_p(L^p_{h(\Delta)}\Delta^\frac{p}{2}\sup_{0\le k\le[\frac{t}{\Delta}]}\mathbb{E}(|X_k|^p)+\Delta^\frac{p}{2}),\endaligned$$ where $C_p$ is a positive constant (independent of $\Delta$) which might change values from line to line.
Since $L^4_{h(\Delta)}\Delta\to0$ as $\Delta\to0$, then $(C_p L_{h(\Delta)}^p+1)L^p_{h(\Delta)}\Delta^\frac{p}{2}$ is bounded for $\Delta\in(0,\Delta_0]$, therefore we have $$\aligned\quad\mathbb{E}(|y_\Delta(t)|^p)&\le
C_p+C_p\mathbb{E}\int_0^{t}|y_\Delta(s)|^{p}ds+C_p\mathbb{E}\int_0^{t}\sup_{0\le r\le s}\mathbb{E}|\bar{x}_\Delta(r)|^pds\\&
\quad+(C_p L_{h(\Delta)}^p+1)[C_pL^p_{h(\Delta)}\Delta^\frac{p}{2}\int_0^{t}\sup_{0\le s\le t}\mathbb{E}(|x_\Delta(s)|^p)ds+C_pT\Delta^\frac{p}{2}]\\&
\le C_p+C_p\int_0^{t}\sup_{0\le s\le t}\mathbb{E}(|x_\Delta(s)|^p)ds.\endaligned$$
On the other hand, similar to (\[sqj2\]), we can take $c$ large enough such that $(\frac{1+c}{c})^{p-1}u^p<1$. So $$\label{sqj33}\aligned\sup_{0\le s\le t}\mathbb{E}|x_\Delta(s)|^p&\le \frac{(1+c)^{p-1}\sup_{0\le s\le t}\mathbb{E}|y_\Delta(s)|^p+(\frac{1+c}{c})^{p-1}u^p||\xi||^p_p}{1-(\frac{1+c}{c})^{p-1}u^p}.\endaligned$$
Thus $$\aligned\quad\sup_{0\le s\le t}\mathbb{E}|x_\Delta(s)|^p&\le
C_p+C_p\mathbb{E}\int_0^{t}\sup_{0\le r\le s}\mathbb{E}|x_\Delta(r)|^pds.\endaligned$$
Gronwall inequality yields (\[sj\]).
Now let us prove (\[sj1\]).
Let $y_\Delta(t)$, $\bar{y}_\Delta(t)$ are defined as above and $\rho_{\Delta,R}=\rho.$ By Itô formula and Lemma \[l1\], for any $0\le t\le T,$
$$\aligned\mathbb{E}(|y_\Delta(t\wedge\rho)|^p)&
\le|x_0|^p+C_p\mathbb{E}\int_0^{t\wedge\rho}(1+|y_\Delta(s)|^{p}+|\bar{x}_\Delta(s)|^p)ds\\&
\quad+C_p\mathbb{E}\int_0^{t\wedge\rho}(|y_\Delta(s)|^{p}+L_{h(\Delta)}^p|x_\Delta(s)-\bar{x}_\Delta(s)|^p+
|\bar{x}_\Delta(s)|^p+|\bar{x}_\Delta(s-\tau)|^p)ds
\\&\quad+C_p|f(0,0)|\mathbb{E}\int_0^{t\wedge\rho}(1+|y_\Delta(s)|^{p}+|x_\Delta(s)-\bar{x}_\Delta(s)|^p)ds\\&
\le C_p+C_p\mathbb{E}\int_0^{t\wedge\rho}|y_\Delta(s)|^{p}ds+C_pT\sup_{0\le s\le T}\mathbb{E}|\bar{x}_\Delta(s)|^p\\&
\quad+(C_p T L_{h(\Delta)}^p+1)\sup_{0\le s\le T}\mathbb{E}|x_\Delta(s)-\bar{x}_\Delta(s)|^p.\endaligned$$
Then by (\[k2\]) and (\[sj\]), we have
$$\mathbb{E}(|y_\Delta(t\wedge\rho)|^p)\le C_p+C_p\int_0^{t}\mathbb{E}|y_\Delta(s\wedge\rho)|^pds+(C_p T L_{h(\Delta)}^p+1)L_{h(\Delta)}^p\Delta^\frac{p}{2}.$$
Gronwall’s lemma yields that $$\sup_{0\le s\le T}\mathbb{E}(|y_\Delta(s\wedge\rho)|^p)\le C(p,T)<\infty.$$
Then similar to (\[sqj2\]), we can take $c$ large enough such that $(\frac{1+c}{c})^{p-1}u^p<1$. So $$\label{sqj4}\aligned\sup_{0\le s\le t}\mathbb{E}|x_\Delta(s\wedge\rho)|^p&\le \frac{(1+c)^{p-1}\sup_{0\le s\le t}\mathbb{E}|y_\Delta(s\wedge\rho)|^p+(\frac{1+c}{c})^{p-1}u^p||\xi||^p_p}{1-(\frac{1+c}{c})^{p-1}u^p}.\endaligned$$
This implies the required assertion easily. $\square$
Convergence rate at fixed time $T$
==================================
Let us first present a lemma which will play a key role in the proof of Theorem \[conv\].
\[jubu\] Suppose (\[local\]) and (\[Lip\]) hold for $p>2,$ and for any $2<q<p,$ $$\mathbb{E}\sup_{-m\le k\le -1}\sup_{k\Delta\le s\le (k+1)\Delta}|\xi(s)-\xi(k\Delta)|^q\le \hat{K}\Delta^\frac{q}{2}.$$ Set $$\theta_{\Delta,R}=\tau_R\wedge\rho_{\Delta,R}\quad and\quad e_\Delta(t)=x(t)-x_\Delta(t)\ for\ t\ge0.$$ Then for any $\Delta\in(0,\Delta^*)$ and any $R\le h(\Delta^*)$, there exists $C(q,T)>0$ (independent of $\Delta$) such that $$\sup_{0\le t\le T}\mathbb{E}(|e_\Delta(t\wedge\theta_{\Delta,R})|^q)\le C(q,T)L^{2q}_{h(\Delta)}\Delta^{\frac{q}{2}}.$$
**Proof** Define the truncated functions $$F_R(x,y)=f_{h^{-1}(R)}(x,y)\ \textrm{and}\ G_R(x,y)=g_{h^{-1}(R)}(x,y), \forall x,y\in\mathbb{R}^d,$$ where $f_{h^{-1}(R)}$ is defined in (\[dy\]) with $\Delta$ replaced by $h^{-1}(R).$ By Lemma (3.1), $F_R$ and $G_R$ are globally Lipschitz continuous for any fixed $R$ ($\ge L^{-1}$(1)), where $L^{-1}$ is the inverse function of $L_R$ when it is seen as a function of $R.$
Without loss of generality, suppose $\Delta^*$ is sufficiently small such that $$h(\Delta^*)=L^{-1}(L_R\exp\{\frac{2^{q-1}(T^\frac{1}{q-1}+4)}{q}\})\ge R.$$
Then for those $x,y\in\mathbb{R}^d$ with $|x|\vee|y|\le R$ and all $\Delta\in(0,\Delta^*]$, we have $$F_R(x,y)=f_{h^{-1}(R)}(x,y)=f(x,y)=f_\Delta(x,y).$$
Similarly, we have $$G_R(x,y)=g_\Delta(x,y).$$
Now consider NSDDE $$\label{fz}
d[z(t)-D(z(t-\tau))
=F_R(z(t),z(t-\tau))dt+G_R(z(t),z(t-\tau))dB_t,t\ge0$$ with $z(\theta)=\xi(\theta)$ on $\theta\in[-\tau,0].$ Since $F_R$ and $G_R$ are globally Lipschitz continuous (with Lipschitz constant $5L_R$) for any fixed $R$, then (\[fz\]) has a unique global solution $z(t)$ on $t\ge\tau.$ Thus $$\label{weiyi}P(x(t\wedge\tau_R)=z(t\wedge\tau_R), \forall t\in[0,T])=1.$$
On the other hand, similar to (\[num1\]) and (\[num2\]), we can define $\bar{z}_\Delta(t), z_\Delta(t)$ in the same way for NSDDE (\[fz\]). We also have $$\label{weiyi1}P(x_\Delta(t\wedge\tau_R)=z_\Delta(t\wedge\tau_R), \forall t\in[0,T])=1.$$
We claim that $$\label{zhengti}\mathbb{E}\sup_{0\le t\le T}|z(t)-z_\Delta(t)|^q\le C(q,T)L^{2q}_{h(\Delta)}\Delta^\frac{q}{2}.$$
Let $y(t)=z(t)-D(z(t-\tau)), y_\Delta(t)=z_\Delta(t)-D(\bar{z}_\Delta(t-\tau))$. Then for any $c,c'>0$ $$\aligned|z(t)-z_\Delta(t)|^q&\le (1+c)^{q-1}|y(t)-y_\Delta(t)|^q+(\frac{1+c}{c})^{q-1}u^q|z(t-\tau)-\bar{z}_\Delta(t-\tau)|^q\\&
\le (1+c)^{q-1}|y(t)-y_\Delta(t)|^q+(\frac{(1+c)(1+c')}{c})^{q-1}u^q|z(t-\tau)-z_\Delta(t-\tau)|^q\\&
\quad+(\frac{(1+c)(1+c')}{cc'})^{q-1}u^q|z_\Delta(t-\tau)-\bar{z}_\Delta(t-\tau)|^q.\endaligned$$
Choose $c$ sufficiently large and $c'$ sufficiently small such that $c_0:=(\frac{(1+c)(1+c')}{c})^{q-1}u^q<1$, and denote $c_1=(\frac{(1+c)(1+c')}{cc'})^{q-1}u^q$. Then we have $$\aligned\mathbb{E}\sup_{0\le s\le t}|z(s)-z_\Delta(s)|^q&\le (1+c)^{q-1}\mathbb{E}\sup_{0\le s\le t}|y(s)-y_\Delta(s)|^q+c_0\mathbb{E}\sup_{0\le s\le t}|z(s-\tau)-z_\Delta(s-\tau)|^q\\&
\quad+c_1\mathbb{E}\sup_{0\le s\le t}|z(s-\tau)-\bar{z}_\Delta(s-\tau)|^q\\&
\le (1+c)^{q-1}\mathbb{E}\sup_{0\le s\le t}|y(s)-y_\Delta(s)|^q+c_0\mathbb{E}\sup_{0\le s\le t}|z(s)-z_\Delta(s)|^q\\&
\quad+c_1\mathbb{E}\sup_{0\le s\le t}|z_\Delta(s)-\bar{z}_\Delta(s)|^q\\&
\quad+c_1\mathbb{E}\sup_{-m\le k\le -1}\sup_{k\Delta\le s\le (k+1)\Delta}|\xi(s)-\xi(k\Delta)|^q.\endaligned$$
So $$\label{kz0}\aligned\mathbb{E}\sup_{0\le s\le t}|z(s)-z_\Delta(s)|^q&\le \frac{(1+c)^{q-1}}{1-c_0}\mathbb{E}\sup_{0\le s\le t}|y(s)-y_\Delta(s)|^q\\&
\quad+\frac{c_1}{1-c_0}\mathbb{E}\sup_{0\le s\le t}|z_\Delta(s)-\bar{z}_\Delta(s)|^q+\frac{c_1}{1-c_0}\hat{K}\Delta^{\frac{q}{2}}.\endaligned$$
As in (\[k2\]), we have $$\label{k4}\sup_{0\le t\le T}\mathbb{E}|z_\Delta(t)-\bar{z}_\Delta(t)|^p\le C(p,T)L^p_{h(\Delta)}\Delta^\frac{p}{2},$$
Now by Itô’s formula, Hölder’s inequality and BDG inequality, it follows that for $0\le t\le T$ $$\aligned&\quad\mathbb{E}\sup_{0\le s\le t}|y(s)-y_\Delta(s)|^q\\&\le 2^{q-1}T^{\frac{1}{q-1}}\mathbb{E}\int_0^t|F_R(z(s),z(s-\tau))-F_R(\bar{z}_\Delta(s),\bar{z}_\Delta(s-\tau))|^qds\\&
\quad+2^{q-1}\mathbb{E}\int_0^t|G_R(z(s),z(s-\tau))-G_R(\bar{z}_\Delta(s),\bar{z}_\Delta(s-\tau))|^qds\\&
\le 2^{q-1}(T^{\frac{1}{q-1}}+4)L_R^q\int_0^t(\mathbb{E}|z(s)-\bar{z}_\Delta(s)|^q+\mathbb{E}|z(s-\tau)-\bar{z}_\Delta(s-\tau)|^q)ds\\&
\le 2^{q}(T^{\frac{1}{q-1}}+4)L_R^q\int_0^t\mathbb{E}|z(s)-z_\Delta(s)|^qds+C(q,T)L_R^q\sup_{0\le t\le T}\mathbb{E}|z_\Delta(t)-\bar{z}_\Delta(t)|^q\\&
\quad+C(q,T)L_R^q\sum_{k=-m}^{-1}\int_{k\Delta}^{(k+1)\Delta}\mathbb{E}|\xi(s)-\xi(k\Delta)|^qds\\&
\le 2^{q}(T^{\frac{1}{q-1}}+4)L_R^q\int_0^t\mathbb{E}|z(s)-z_\Delta(s)|^qds+C(q,T)L_R^{2q}\Delta^{\frac{q}{2}}\\&
\quad+C(q,T)L_R^q\hat{K}\Delta^{\frac{q}{2}}.
\endaligned$$
Then (\[kz0\]) and Gronwall lemma yields $$\label{zhengti1}\mathbb{E}\sup_{0\le t\le T}|z(t)-z_\Delta(t)|^q\le H(R,T,\xi)\Delta^\frac{q}{2},$$ where $$\aligned H(R,T,\xi)&:=(C(q,T)L_R^{2q}+C(q,T)L_R^q\hat{K})e^{2^{q}(T^{\frac{1}{q-1}}+4)TL_R^q}\\&
\le C(q,T) L_R^{2q}\exp\{2^{q}(T^{\frac{1}{q-1}}+4)TL_R^q\}\\&
=C(q,T)L^{2q}_{h(\Delta^*)}\le C(q,T)L^{2q}_{h(\Delta)}.\endaligned$$
Hence (\[weiyi\]), (\[weiyi1\]) and (\[zhengti\]) implies $$\label{kz01}\aligned\mathbb{E}\sup_{0\le s\le t}|x(s\wedge\theta_{\Delta,R})-x_\Delta(s\wedge\theta_{\Delta,R})|^q&\le C(q,T)L^{2q}_{h(\Delta)}\Delta^{\frac{q}{2}},\endaligned$$
This completes the proof. $\square$
Now we are ready to prove Theorem \[conv\].
**Proof of Theorem \[conv\]** Let $\tau_R, \rho_{\Delta,R}, \theta_{\Delta,R}$ and $e_\Delta(t)$ be the same as before. Then by Young’s inequality, we have that for any $\delta>0,$ $$\label{bds}\aligned\mathbb{E}(|e_\Delta(T)|^q)&\le \mathbb{E}(|e_\Delta(T)|^q1_{\{\theta_{\Delta,R}>T\}})+ \mathbb{E}(|e_\Delta(T)|^q1_{\{\theta_{\Delta,R}\le T\}})\\&\le\mathbb{E}(|e_\Delta(T)|^q1_{\{\theta_{\Delta,R}>T\}})+ \frac{q\delta}{p}\mathbb{E}(|e_\Delta(T)|^p)+\frac{p-q}{p\delta^{q/(p-q)}}P(\theta_{\Delta,R}\le T)\\&
\le\mathbb{E}(|e_\Delta(T\wedge\theta)|^q)+ \frac{q\delta C}{p}\left(\mathbb{E}(|x_\Delta(T)|^p)+\mathbb{E}(|x(T)|^p)\right)+\frac{p-q}{p\delta^{q/(p-q)}}P(\theta_{\Delta,R}\le T)\endaligned$$
where $C$ is a positive constant (independent of $\Delta$) which might change the value from line to line. We have used the fact that
$$\mathbb{E}(|e_\Delta(T)|^p)\le C(\mathbb{E}(|x_\Delta(T)|^p)+\mathbb{E}(|x(T)|^p))$$ in the last inequality.
By Lemma \[ju\] and \[temju\], we have $$\mathbb{E}(|x_\Delta(T)|^p)+\mathbb{E}(|x(T)|^p)\le C,$$ and $$P(\theta_{\Delta,R}\le T)\le P(\tau_R\le T)+P(\rho_{\Delta,R}\le T)\le\frac{C}{R^p}.$$
Thus, $$\mathbb{E}(|e_\Delta(T)|^q)\le\mathbb{E}(|e_\Delta(T\wedge\theta)|^q)+\frac{qC\delta}{p}+\frac{C(p-q)}{pR^p\delta^{q/(p-q)}}$$ holds for any $\Delta\in(0,\Delta^*), R>|x_0|$ and $\delta>0$. Then we can choose $\delta=L^{2q}_{h(\Delta)}\Delta^\frac{q}{2}$ and $R=(L^{2q}_{h(\Delta)}\Delta^\frac{q}{2})^{-\frac{1}{p-q}}$ to get $$\mathbb{E}(|e_\Delta(T)|^q)\le\mathbb{E}(|e_\Delta(T\wedge\theta)|^q)+CL^{2q}_{h(\Delta)}\Delta^\frac{q}{2}.$$
But by condition (\[tj\]), we have $$h(\Delta)\ge (L^{2q}_{h(\Delta)}\Delta^\frac{q}{2})^{-\frac{1}{p-q}}=R.$$
Then by Lemma \[jubu\], $$\mathbb{E}(|e_\Delta(T)|^q)\le C(q,T)L^{2q}_{h(\Delta)}\Delta^\frac{q}{2},$$ where $C$ is a positive constant depends on $q$ and $T$. This is the first inequality of (\[shou\]).
For the second inequality, since $q<p,$ by Hölder inequality, it follows easily from the above inequality and (\[k2\]) in Lemma \[temju\] that $$\aligned\mathbb{E}(|x(T)-\bar{x}_\Delta(T)|^q)&\le C_q\left(\mathbb{E}(|e_\Delta(T)|^q)+\mathbb{E}(|x_\Delta(T)-\bar{x}_\Delta(T)|^q)\right)\\&\le C_q\left(\mathbb{E}(|e_\Delta(T)|^q)+[\sup_{0\le t\le T}\mathbb{E}(|x_\Delta(t)-\bar{x}_\Delta(t)|^p)]^\frac{q}{p}\right)\\&
\le C(q,T)L^{2q}_{h(\Delta)}\Delta^\frac{q}{2}+(C(q,T)L^{p}_{h(\Delta)}\Delta^\frac{p}{2})^\frac{q}{p}\le C(q,T)L^{2q}_{h(\Delta)}\Delta^\frac{q}{2}.\endaligned$$ We complete the proof. $\square$
Convergence rates over the time interval $[0,T]$
================================================
First of all, let us prove a similar Lemma to Lemma \[ju\].
\[zuida\] Let (\[local\]), (\[dandiao\]) and (\[gzeng\]) hold for $p>2$ and $2<r<p.$ Set $\bar{p}=p-r+2.$ Then $$\mathbb{E}(\sup_{0\le t\le T}|x(t)|^{\bar{p}})\le C,\ \forall T>0.$$
**Proof** Since for sufficiently large $c>0$, $$\mathbb{E}(\sup_{0\le t\le T}|x(t)|^{\bar{p}})\le \frac{(1+c)^{\bar{p}-1}\mathbb{E}\sup_{0\le t\le T}|y(t)|^{\bar{p}}+(\frac{1+c}{c})^{\bar{p}-1}u^{\bar{p}}||\xi||^{\bar{p}}}{1-(\frac{1+c}{c})^{\bar{p}-1}u^{\bar{p}}},$$ then we only need to prove $$\mathbb{E}\sup_{0\le t\le T}|y(t)|^{\bar{p}}\le C(\bar{p},T).$$
Indeed, Itô’s formula and (\[dandiao\]) imply that $$\aligned\mathbb{E}(\sup_{0\le t\le T}|y(t)|^{\bar{p}})&\le |y(0)|^{\bar{p}}+\bar{p}K\mathbb{E}\int_0^T|y(s)|^{\bar{p}-2}(1+|x(s)|^2+|x(s-\tau)|^2)ds\\&
\quad+\mathbb{E}(\sup_{0\le t\le T}|M(t)|),\endaligned$$ where $y(t)=x(t)-D(x(t-\tau)),$ and $$M(t)=\bar{p}\int_0^t|y(s)|^{\bar{p}-2}y^T(s)g(x(s),x(s-\tau))dB(s)$$ is a local martingale with $M(0)=0.$
It is obvious that $$\mathbb{E}\int_0^T|y(s)|^{\bar{p}-2}(1+|x(s)|^2+|x(s-\tau)|^2)ds\le C(\bar{p},T)<\infty.$$
On the other hand, by Burkholder-Davis-Gundy (BDG) inequality (see e.g. [@Ikeda]), it follows that $$\aligned\mathbb{E}(\sup_{0\le t\le T}|M(t)|)&\le C'\mathbb{E}\left(\int_0^T|y(s)|^{2\bar{p}-2}|g(x(s),x(s-\tau))|^2ds\right)^\frac{1}{2}\\&
\le C'\mathbb{E}\left(\sup_{0\le t\le T}|y(t)|^{\bar{p}}\int_0^T|y(s)|^{\bar{p}-2}\cdot \bar{K}(1+|x(s)|^r+|x(s-\tau)|^r)ds\right)^\frac{1}{2}\\&
\le \frac{1}{2}\mathbb{E}\sup_{0\le t\le T}|y(t)|^{\bar{p}}+\frac{C'^2}{2}\mathbb{E}\int_0^T|y(s)|^{\bar{p}-2}\cdot \bar{K}(1+|x(s)|^r+|x(s-\tau)|^r)ds.\endaligned$$
As in the proof of Lemma \[ju\], we have $$\frac{C'^2}{2}\mathbb{E}\int_0^T|y(s)|^{\bar{p}-2}\cdot \bar{K}(1+|x(s)|^r+|x(s-\tau)|^r)ds\le C''.$$
Thus $$\aligned\mathbb{E}(\sup_{0\le t\le T}|y(t)|^{\bar{p}})&\le C(\bar{p},T),\endaligned$$
This completes the proof. $\square$
For the discontinuous and continuous-time MTEM methods (\[num\]) and (\[num2\]), we have
\[zuida1\] Let (\[local\]), (\[tiaoj\]), (\[dandiao\]) and (\[gzeng\]) hold for $p>2$ and $p>r\ge2.$ Set $\bar{p}=p+2-r.$ Then $$\label{zuida2}\sup_{0<\Delta\le\Delta^*}\mathbb{E}(\sup_{0\le t\le T}|x_\Delta(t)|^{\bar{p}})\le C,\ \forall T>0,$$ and therefore, $$\label{zuida3}\sup_{0<\Delta\le\Delta^*}\mathbb{E}(\sup_{0\le k\le [\frac{T}{\Delta}]}|X_k^\Delta|^{\bar{p}})\le C,\ \forall T>0,$$
**Proof** Similar to the proof of Lemma \[zuida\], we only need to prove $$\sup_{0<\Delta\le\Delta^*}\mathbb{E}(\sup_{0\le t\le T}|y_\Delta(t)|^{\bar{p}})\le C(\bar{p},T),$$ where $y_\Delta(t)=x_\Delta(t)-D(\bar{x}_\Delta(t-\tau)).$
For any $\Delta\in(0,\Delta^*],$ by Itô formula and Lemma \[l1\], we have $$\aligned\sup_{0\le t\le T}|y_\Delta(t)|^{\bar{p}}&\le|y_0|^{\bar{p}}+2K\bar{p}\int_0^T|y_\Delta(s)|^{\bar{p}-2}
(1+|\bar{x}_\Delta(s)|^2+\bar{x}_\Delta(s-\tau)|^2)ds\\&
\quad+\bar{p}\int_0^T|y_\Delta(s)|^{\bar{p}-2}|x_\Delta(s)-\bar{x}_\Delta(s)||f_\Delta(\bar{x}_\Delta(s),\bar{x}_\Delta(s-\tau))|ds\\&
\quad+\bar{p}\sup_{0\le t\le T}\left|\int_0^t|y_\Delta(s)|^{\bar{p}-2}\langle y_\Delta(s),g_\Delta(\bar{x}_\Delta(s),\bar{x}_\Delta(s-\tau))dB(s)\rangle\right|.\endaligned$$
Young’s inequality and Lemma \[temju\] imply that $$\int_0^T|y_\Delta(s)|^{\bar{p}-2}
(1+|\bar{x}_\Delta(s)|^2+\bar{x}_\Delta(s-\tau)|^2)ds<C(\bar{p},T)$$
Moreover, since for $\Delta>0$ small enough, $$|f_\Delta(x,y)|\le 5L_{h(\Delta)}(|x|+|y|)+|f(0,0)|,$$ then as in the proof of Lemma \[temju\], we have $$\mathbb{E}\int_0^T|y_\Delta(s)|^{\bar{p}-2}|x_\Delta(s)-\bar{x}_\Delta(s)||f_\Delta(\bar{x}_\Delta(s),\bar{x}_\Delta(s-\tau))|ds\le C.$$
So by BDG inequality again and (\[gzeng\]), we have $$\aligned&\quad\mathbb{E}\sup_{0\le t\le T}|y_\Delta(t)|^{\bar{p}}\\&
\le C+C(\bar{p},\bar{K})\mathbb{E}\left(\int_0^T|y_\Delta(s)|^{2\bar{p}-2}(1+|\bar{x}_\Delta(s)|^r+|\bar{x}_\Delta(s-\tau)|^r)ds\right)^\frac{1}{2}\\&
\le C+\mathbb{E}\left|\sup_{0\le t\le T}|y_\Delta(t)|^{\bar{p}}\cdot C(\bar{p},\bar{K})\left(\int_0^T|y_\Delta(s)|^{\bar{p}-2}
(1+|\bar{x}_\Delta(s)|^r+|\bar{x}_\Delta(s-\tau)|^r)ds\right)\right|^\frac{1}{2}\\&
\le C+\frac{1}{2}\mathbb{E}\sup_{0\le t\le T}|y_\Delta(t)|^{\bar{p}}+\frac{C^2(\bar{p},\bar{K})}{2}\int_0^T|y_\Delta(s)|^{\bar{p}-2}
(1+|\bar{x}_\Delta(s)|^r+|\bar{x}_\Delta(s-\tau)|^r)ds,
\endaligned$$ where $C$ is a constant (independent of $\Delta$).
Then we have $$\aligned\mathbb{E}\sup_{0\le t\le T}|y_\Delta(t)|^{\bar{p}}
&\le C+C^2(\bar{p},\bar{K})\mathbb{E}\int_0^T|y_\Delta(s)|^{\bar{p}-2}
(1+|\bar{x}_\Delta(s)|^r+|\bar{x}_\Delta(s-\tau)|^r)ds\le C.\endaligned$$
Then the required assertion (\[zuida2\]) follows. $\square$
\[close1\] Assume that (\[local\]), (\[tiaoj\]), (\[dandiao\]) and (\[gzeng\]) hold for $p>2$ and $2\le r< p.$ If $q\le p+2-r,$ then for any $\Delta\in(0,\Delta^*)$, there exists $C>0$ (independent of $\Delta$) such that $$\mathbb{E}\left(\sup_{0\le t\le T}|x_\Delta(t)-\bar{x}_\Delta(t)|^q\right)\le CL_{h(\Delta)}^q\Delta^{\frac{q}{2}-1}.$$
The proof is almost the same as that of Lemma 5.5 in [@LX], so we omit it here.
Now let us prove Theorem \[conv1\].
**Proof of Theorem \[conv1\]** Let $\theta_{\Delta,R}$ and $e_\Delta(t)$ be the same as before. As in the proof of Theorem \[conv\], by Young’s inequality, we have that for any $\delta>0,$ $$\aligned\mathbb{E}(\sup_{0\le t\le T}|e_\Delta(t)|^q)&\le \mathbb{E}(1_{\{\theta_{\Delta,R}>T\}}\sup_{0\le t\le T}|e_\Delta(t)|^q)+ \frac{q\delta}{p}\mathbb{E}(\sup_{0\le t\le T}|e_\Delta(t)|^p)\\&\quad+\frac{p-q}{p\delta^{q/(p-q)}}P(\theta_{\Delta,R}\le T).\endaligned$$
By Lemma \[zuida\], \[zuida1\], $$\mathbb{E}(\sup_{0\le t\le T}|e_\Delta(t)|^p)\le C(\mathbb{E}(\sup_{0\le t\le T}|x(t)|^p)+\mathbb{E}(\sup_{0\le t\le T}|x_\Delta(t)|^p))\le C.$$
Then similar to the proof of Theorem \[conv\], we have $$\aligned\mathbb{E}(\sup_{0\le t\le T}|e_\Delta(t)|^q)&\le \mathbb{E}(\sup_{0\le t\le T}|e_\Delta(t\wedge\theta_{\Delta,R})|^q)+ \frac{Cq\delta}{p}+\frac{C(p-q)}{pR^p\delta^{q/(p-q)}}\endaligned$$ holds for any $\Delta\in(0,\Delta^*), \delta>0$ and $R>|x_0|.$
Since we have proved Lemma \[jubu\], repeat the proof of Theorem \[conv\], we have $$\mathbb{E}(\sup_{0\le t\le T}|e_\Delta(t)|^q)\le CL^{2q}_{h(\Delta)}\Delta^\frac{q}{2}$$ for $\delta=L^{2q}_{h(\Delta)}\Delta^\frac{q}{2},\quad R=(L^{2q}_{h(\Delta)}\Delta^\frac{q}{2})^{-\frac{1}{p-q}}\le h(\Delta),$ as required. $\square$
Examples
========
Now let us present two examples to illustrate our theory.
**Example 1** Let $d=1, \tau=1.$ Consider the following scalar NSDDE: $$\label{sde1}\aligned d[x(t)-\frac{1}{2}\sin x(t-1)]&=(2x(t)-x(t)e^{3x(t)}+\frac{1}{2}\sin
x(t-1))dt\\&\quad+\sqrt{\frac{1}{5}x^2(t)e^{3x(t)}+x^2(t-1)+1}dB_t.\endaligned$$
Here $f(x,y)=2x-xe^{3x}+\frac{1}{2}\sin y$, $g(x,y)=\sqrt{\frac{1}{5}x^2e^{3x}+y^2+1}$ and $D(y)=\frac{1}{2}\sin y.$ Then neither $f$ nor $g$ is polynomial growing (although both are local Lipschitz continuous).
Moreover, for any $a\in(0,1]$, we have $$\aligned2\langle
x-aD(\frac{1}{a}y),f(x,y)\rangle+5|g(x,y)|^2&=4x^2-2x^2e^{3x}+x\sin y-2ax\sin \frac{y}{a}\\&\quad+axe^{3x}\sin \frac{y}{a}-\frac{a}{2}\sin y\sin\frac{y}{a}+x^2e^{3x}+5y^2+5\\&
\le4x^2-x^2e^{3x}+|x|+2|x|\\&\quad+|x|e^{3x}+\frac{1}{2}+5y^2+5\\&
\le4x^2+3\cdot\frac{1+x^2}{2}+\frac{11}{2}+5y^2+e^{3x}(|x|-x^2).\endaligned$$
Notice that if $x\le0$, we have $$e^{3x}(|x|-x^2)=e^{3x}[\frac{1}{4}-(|x|-\frac{1}{2})^2]\le \frac{1}{4}e^{3x}\le \frac{1}{4}.$$
If $x>0$, then $$e^{3x}(|x|-x^2)\le\sup_{0\le x\le1}e^{3x}(x-x^2)=e^{3x}(x-x^2)|_{x=\frac{\sqrt{5}-1}{2}}=(\sqrt{5}-2)e^\frac{3\sqrt{5}-3}{2}<e^2.$$
Thus $$\aligned2\langle
x-aD(\frac{1}{a}y),f(x,y)\rangle+3|g(x,y)|^2&
\le4x^2+3\cdot\frac{1+x^2}{2}+\frac{11}{2}+5y^2+e^{2}\\&
\le (7+e^2)(1+|x|^2+|y|^2).\endaligned$$
We have shown that condition (\[dandiao\]) holds for $p=6$ and $K=7+e^2$ for any $x,y.$
Moreover, since $f$ and $g$ are differential on $\mathbb{R}^2$, by mean value theorem, for any $|x|\vee|x'|\vee|y|\vee|y'|\le R,$ let $h=x-\bar{x}$, $k=y-\bar{y}$, then there exists $\theta\in(0,1)$ such that $$\aligned|f(x,y)-f(x',y')|&=f'_x(\bar{x}+\theta h,\bar{y}+\theta k)h+f'_y(\bar{x}+\theta h,\bar{y}+\theta k)k\\&
\le(2+(1+3R)e^R)(|x-x'|+|y-y'|).\endaligned$$ Similarly, $$|g(x,y)-g(x',y')|\le Re^R(1+\frac{3}{2}R)(|x-x'|+|y-y'|)$$ for all $R>0$ and $|x|\vee|x'|\vee|y|\vee|y'|\le R.$ Thus (\[local\]) holds for $L_{R}=3(1+R+R^2)e^R.$
Then for any $0<\varepsilon<1,$ we can define $l(x):=\frac{1}{x^{1-\varepsilon}L^4_x}$ for $x>0$. It is clear that $l$ is a strict decreasing function in the interval $(0,\infty)$. Let $h$ be the inverse function of $l.$ Then $h$ is also a strict decreasing function in the interval $(0,\Delta^*)$ and $h(\Delta)\to\infty$ as $\Delta\to0$.
Now $$L^4_{h(\Delta)}\Delta=L^4_Rl(R)=\frac{1}{R^{1-\varepsilon}},$$ where $R:=h(\Delta)$. Therefore, $$L^4_{h(\Delta)}\Delta=\frac{1}{h(\Delta)^{1-\varepsilon}}\to0\qquad \textrm{as}\ \Delta\to0.$$
And $$(L^{2q}_{h(\Delta)}\Delta^{\frac{q}{2}})^{-\frac{1}{p-q}}=(L^{4}_{h(\Delta)}\Delta)^{-\frac{q}{2(p-q)}}
=h(\Delta)^{\frac{q(1-\varepsilon)}{2(p-q)}}\le h(\Delta)$$ for $\Delta$ small enough if $\frac{q}{2(p-q)}\le 1$ (i.e. $q\le\frac{2p}{3}=4$). Then by Theorem \[conv\], for any $T>0$, $2<q\le4$ and sufficient small $\Delta$, we have $$\label{b1}\mathbb{E}|x(T)-x_\Delta(T)|^q\le C(q,T)L^{2q}_{h(\Delta)}\Delta^\frac{q}{2}=C(q,T)h(\Delta)^{\frac{q(\varepsilon-1)}{2}}$$ and $$\label{b2}\mathbb{E}|x(T)-\bar{x}_\Delta(T)|^q\le C(q,T)L^{2q}_{h(\Delta)}\Delta^\frac{q}{2}=C(q,T)h(\Delta)^{\frac{q(\varepsilon-1)}{2}}.$$
Hölder inequality implies that (\[b1\]) and (\[b1\]) holds for any $0<q\le4$.
Since $f$ does not satisfy polynomial growth condition in this case, then the strong convergence result Theorem 3.7 in [@ZSL] does not be hold here. However, for the continuous-time MTEM methods (\[num1\]) and (\[num2\]), the strong convergence results still holds for the given NSDDE.
**Example 2** Consider the scalar NSDDE $$\label{sde2}\aligned d[x(t)-\frac{1}{2}\sin x(t-1)]&=(2x(t)-x^5(t)+\frac{1}{2}\sin
x(t-1))dt+\frac{2x^3(t)x(t-1)}{1+x^2(t-1)}dB_t.\endaligned$$
Here $f(x,y)=2x-x^5+\frac{1}{2}\sin y$, $g(x,y)=\frac{x^3y}{2(1+y^2)}$ and $D(y)=\frac{1}{2}\sin y.$ It is obvious that $f$ and $g$ are both locally Lipschitz continuous functions with respect to $x$ and $y$.
Moreover, for any $a\in(0,1]$, we have $$\aligned2\langle
x-aD(\frac{1}{a}y),f(x,y)\rangle+5|g(x,y)|^2&=4x^2-2x^6+x\sin y-2ax\sin \frac{y}{a}+x^5a\sin \frac{y}{a}\\&\quad-\frac{a}{2}\sin y\sin\frac{y}{a}+\frac{5x^6y^2}{4(1+y^2)^2}\\&
\le4x^2-2x^6+|xy|+|2ax\cdot\frac{y}{a}|\\&\quad+\frac{5}{6}x^6+\frac{a^6}{6}\sin^6\frac{y}{a}+\frac{a}{2}|y\cdot\frac{y}{a}|+\frac{5}{16}x^6\\&
\le4x^2+3\frac{x^2+y^2}{2}+\frac{y^2}{6}+\frac{y^2}{2}\\&
\le\frac{11}{2}(1+x^2+y^2).\endaligned$$
So (\[dandiao\]) holds for $p=6$.
Moreover $|g(x,y)|^2\le|x|^3\le 1+|x|^3+|y|^3.$ That is, (\[gzeng\]) also holds for $r=3$ and $\bar{K}=1$.
On the other hand, we have $$|f(x,y)-f(x',y')|\le (2+5R^4)(|x-x'|+|y-y'|)$$ and $$|g(x,y)-g(x',y')|\le(3R^2+R^3)(|x-x'|+|y-y'|)$$ for all $R>0$ and $|x|\vee|x'|\vee|y|\vee|y'|\le R.$ Thus $f$ and $g$ are local Lipschitz continuous with local Lipschitz constant $L_R=5R^4+4.$
For $\varepsilon>0$ small enough, choose $$h(\Delta)=\sqrt[4]{\frac{\Delta^{-\varepsilon}-4}{5}},\quad \Delta<4^{-\frac{1}{\varepsilon}}.$$ Then we have $h(\Delta)\to \infty$ and $L^4_{h(\Delta)}\Delta=\Delta^{1-\varepsilon}\to0$ as $\Delta\to0.$ That is, (\[tiaoj\]) holds for such defined $h$.
Choose $q=4$. If we take $\frac{4}{5}<\varepsilon<1,$ then for sufficiently small $\Delta,$ $$(L^{2q}_{h(\Delta)}\Delta^{\frac{q}{2}})^{-\frac{1}{p-q}}=(L^{4}_{h(\Delta)}\Delta)^{-1}
=\Delta^{\varepsilon-1}\le \sqrt[4]{\frac{\Delta^{-\varepsilon}-4}{5}}=h(\Delta),$$ i.e. (\[tj\]) holds for small $\Delta$. So by Theorem \[conv1\], we have $$\mathbb{E}\sup_{0\le t\le T}|x(t)-x_\Delta(t)|^4\le C\Delta^{2(1-\varepsilon)},$$ and $$\mathbb{E}\sup_{0\le t\le T}|x(t)-\bar{x}_\Delta(t)|^4\le CL^4_{h(\Delta)}\Delta=C\Delta^{1-\varepsilon}.$$
Conclusions
===========
We have investigated the strong convergence rates of so called two versions of continuous-time MTEM methods (i.e., $x_\Delta(t)$ and $\bar{x}_\Delta(t)$) for nonlinear NSDDE $d[x(t)-D(x(t-\tau))]
=f(x(t),x(t-\tau))dt+g(x(t),x(t-\tau))dB_t$ in this paper. Roughly speaking, $x_\Delta(t)$ and $\bar{x}_\Delta(t)$ strongly converge (in the sense of $q$-th moment) to the exact solution $x(t)$ at fixed time $T$ (with rate $L^{2q}_{h(\Delta)}\Delta^{\frac{q}{2}}$) if local Lipschitz condition and the Khasiminskii-type condition hold. Moreover, if $g$ satisfies polynomial growth condition (\[gzeng\]), then $x_\Delta(t)$ and $\bar{x}_\Delta(t)$ strongly converge to the exact solution $x(t)$ over a time interval $[0,T]$ (with rates $L^{2q}_{h(\Delta)}\Delta^{\frac{q}{2}}$ and $L^{q}_{h(\Delta)}\Delta^{\frac{q}{2}-1}$, respectively).
[99]{}
A.F. Bastani, M. Tahmasebi, Strong convergence of split-step backward Euler method for stochastic differential equations with non-smooth drift, J. Comput. Appl. Math. 236 (2012) 1903-1918.
M. Eissa, Y. Xiao, B. Tian, Convergence and Stability of Two Classes of Theta Methods with Variable Step Size for a Stochastic Pantograph Equations, Journal of Advanced Mathematics and Applications, 5(2016)95-106.
Q. Guo, X. Mao, R. Yue, The truncated Euler¨CMaruyama method for stochastic differential delay equations, Numer. Algor. 78(2018)599-624.
S. Gan, H. Schurz, H. Zhang, Mean square convergence of stochastic $\theta$-methods for nonlinear neutral stochastic differential delay equations, International Journal of Numerical Analysis and Modeling, 8(2011) 201-213.
I. Ikeda, S. Watanabe, Stochastic differential equations and diffusion processes, North-Holland, Amsterdam, 1981.
P.E. Kloeden, E. Platen, Numerical Solution of Stochastic Differential Equations, Applications of Mathematics, Springer, Berlin, 1992.
G. Lan, Asymptotic exponential stability of modified truncated EM method for neutral stochastic differential delay equations, J. Comput. Appl. Math., 2018(340), 334-341.
G. Lan, F. Xia, Strong convergence rates of modified truncated EM method for stochastic differential equations, J. Comput. Appl. Math., 2018(334), 1-17.
X. Mao, The truncated Euler-Maruyama method for stochastic differential equations, J. Comput. Appl. Math. 290 (2015) 370-384.
X. Mao, Convergence rates of the truncated Euler-Maruyama method for stochastic differential equations, J. Comput. Appl. Math. 296 (2016) 362-375.
X. Mao, Stochastic differential equations and applications, 2nd edition, Horwood, Chichester, 2007.
X. Mao, C. Yuan, Stochastic differential equations with Markovian switching, Imperial College Press, 2006.
M. Milo$\check{\textrm{s}}$evi$\acute{\textrm{c}}$, Convergence and almost sure exponential stability of implicit numerical methods for a class of highly nonlinear neutral stochastic differential equations with constant delay, Journal of Computational and Applied Mathematics 280 (2015) 248-264.
L. Tan, C. Yuan, Convergence rates of truncated EM scheme for NSDDEs, arXiv:1801.05952.
L. Tan, C. Yuan, Convergence rates of theta-method for neutral SDDEs under non-globally Lipschitz continuous coefficients, arXiv:1701.00223.
B. Yin, Z. Ma, Convergence of the semi-implicit Euler method for neutral stochastic delay differential equations with phase semi-Markovian switching, Applied Mathematical Modelling 35 (2011) 2094-2109.
H. Zhang, S. Gan, Mean square convergence of one-step methods for neutral stochastic differential delay equations, Applied Mathematics and Computation, 204 (2008) 884-890.
W. Zhang, M. Song, M. Liu, Strong convergence of the partially truncated Euler¨CMaruyama method for a class of stochastic differential delay equations, Journal of Computational and Applied Mathematics 335 (2018) 114-128.
W. Zhang, M. Song, M. Liu, Strong convergence of the truncated Euler-Maruyama method for stochastic functional differential equation, International Journal of Computer Mathematics, DOI: 10.1080/00207160.2017.1395871
S. Zhou, F. Wu, Convergence of numerical solutions to neutral stochastic delay differential equations with Markovian switching, Journal of Computational and Applied Mathematics 229 (2009)85-96.
[^1]: Corresponding author: Email: [email protected].
[^2]: Supported by Natural Science Foundation of China (NSFC 11601025).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We consider effective action for the Einstein gravity and show that dressed mean fields are actual variables of the effective action. Kernels of this effective action expressed in terms of dressed effective fields are constituent parts of scattering amplitudes for gravitons. Possible applications to the graviton scattering and black hole formation are discussed at the semiclassical level. In particular, we consider graviton scattering in four dimensions based on the Lipatov effective action for quantum gravity, shock waves of particles moving on the brane in Randall-Sundrum scenario with fifth extra dimension, and Giddings’ estimation of Froissart bound. Keywords: Einstein gravity, 1PI effective action, Slavnov–Taylor identity, eikonal amplitude'
---
.3in .3in 12 1g2
\
\
[**Semiclassical scattering amplitudes of dressed gravitons**]{}\
Kyungsik Kang$^{(a)}$ and Igor Kondrashuk$^{(b)}$\
\
Introduction
============
There is an intriguing belief that the high energy scattering in gauge theories in four space-time dimensions can be described by the bulk physics of supergravity with a higher dimension in AdS space. In other words, one can extract information about quantum process amplitudes in four spacetime dimensions starting with the classical physics of wider theories including gravity in higher dimensions. This is called gauge/string-gravity duality and is based on Maldacena conjecture [@Maldacena:1997re].
The total cross section for the particle absorption by a black hole is estimated by the area of the black hole horizon in Ref. [@Das] for four dimensions in a semiclassical way. Another estimation has been given in Ref.[@Solodukhin:2002ui]. In Ref. [@Matschull:1998rv] it has been shown that the horizon is necessarily formed in three dimensions at sufficiently high energies when two particles collide. The amplitude of this process has been calculated at the semiclassical level in Ref. [@Jevicki:2002fq], in which classical Hamiltonian can be found.
In four dimensions there are no exact results about the cross section of the process [*two particles $\rar$ black hole*]{} even at the classical level. One has to rely on approximate estimates such as in Ref. [@Das; @Solodukhin:2002ui]. In unrestricted four-dimensional spacetime the horizon area is proportional to the square of total energy in center-mass frame. However, in brane models as it has been shown by Giddings [@Giddings] radius of horizon does modify its form to logarithm of c.m. energy. This is the basic source of the Froissart bound (FB) for the total cross sections which has been known since 1961 as a consequence of unitarity, positivity of the imaginary part of partial wave amplitudes of analytic S-matrix [@Froissart]. Giddings [@Giddings] concluded that bulk theory feels in some way unitarity of the four-dimensional theory on the boundary.
At the same time there is approach to consider the effective action of quantum gravity as a functional of the effective field of metric convoluted with unspecified dressing function [@Kondrashuk:2000br; @Cvetic:2002dx; @Cvetic:2002in; @Kondrashuk:2003tw]. We will call this construction “dressed effective field”. In this paper we make steps to embed the results of semiclassical and eikonal estimations of the amplitudes [*two particles $\rar$ black hole*]{} process and graviton scattering processes into the approach of Ref. [@Cvetic:2002dx; @Kondrashuk:2003tw; @Cvetic:2004kx].
The effective action for quantum gravity is Legendre transform of the logarithm of path integral. Tilded fields in this paper will mean dressed effective fields $\tilde{h}$ and $\tilde{\phi}$ of graviton and matter fields, that is the effective fields convoluted with unspecified dressing function [@Kondrashuk:2000br; @Cvetic:2002dx; @Cvetic:2002in; @Kondrashuk:2003tw], $$\begin{aligned}
& \dis{ \tilde{h}_{\mu\nu}(x) = \int~d^D x G^{-1}_{h}(x-x')h_{\mu\nu}(x')} \\
& \dis{\tilde{\phi}(x) = \int~d^D x G^{-1}_{\phi}(x-x') \phi(x') },\end{aligned}$$ $h_{\mu\nu}$ is the graviton tensor, $g_{\mu\nu} = \eta_{\mu\nu} +
h_{\mu\nu},$ where $\eta_{\mu\nu}$ is Minkowski tensor.
In addition to analyzing size of horizon along the lines of Ref. [@Giddings] we consider graviton scattering in four dimensions based on the Lipatov effective action for quantum gravity [@Lipatov:1991nf], scattering of one particle by a shock wave of another one both of which move on on the brane in Randall-Sundrum scenario with fifth extra dimension. We start our analysis with introducing the concept of the dressed effective fields.
Effective action of dressed gravitons in four dimensions {#s2}
========================================================
There is a way to write a functional structure of the effective action for dressed gravitons in supergravity [@Kondrashuk:2003tw]. For example, one can consider supergravity in the component formulation. Strictly speaking, if one works with supergravity the vielbeins must be introduced. Since we do not want to overload notation in this paper we work with the metric as an independent field which is variable of integration in the path integral. This would be sufficient for the purpose of the present paper. To clarify the idea we elaborate usual four-dimensional Einstein gravity as an example in this section.
Let us take the path integral for the theory of four dimensional gravity in the following form [@Hawking'76]: $$\begin{aligned}
Z = \int~dg~d{\phi}~\exp i\left[\dis{S[g,\phi]} \right],\end{aligned}$$ where $S[g,\phi]$ is the classical action of D4 gravity $g$ coupled to matter fields $\phi.$ The action of the gravitational field is usually taken to be $$\begin{aligned}
S = \frac{1}{16\pi}\int~R\sqrt{-g}~d^4x. \end{aligned}$$
In such approach we can fix the symmetry of the diffeomorphism group by imposing some linear gauge fixing condition on the graviton field $g_{\mu\nu}.$ One can take the most general form of a linear gauge fixing condition $$\begin{aligned}
F[g_{\mu\nu}] = 0. \end{aligned}$$ To fix the gauge we have to introduce into the path integral the gauge fixing term and Faddeev–Popov ghost field at the same time in order to factorize the volume of diffeomorphism out [@SF]. This procedure brings out additional symmetry for the classical action extended by gauge fixing term and by the ghost term which is called BRST symmetry [@BRST]. Total action including gauge fixing, FP ghost action, at the classical level can be written as $$\begin{aligned}
& \dis{S = S_{\rm cl} + S_{\rm gf} + S_{\rm gh}} \no\\
& = \dis{\frac{1}{16\pi}\int~d^4x~R\sqrt{-g}
- \int~d^4x ~\frac{1}{2\a}\le F[g_{\mu\nu}] \ri^2} \no\\
& - \dis{\int~d^4x~i~b~\frac{\del F}{\del g_{\mu\nu}}~{\cal L}_c g_{\mu\nu}
+ S_{I}\le g_{\mu\nu},\phi\ri,}
\label{total-action-gravity}\end{aligned}$$ where $c^{\mu}(x)$ here is ghost field, $b(x)$ is antighost field, and ${\cal L}_c$ is Lie derivative associated with ghost field $c ^{\mu}(x)$ that acts on the metric field as $$\begin{aligned}
{\cal L}_c g^{\mu\nu} = c^\lambda\pd_\lambda g^{\mu\nu}- (\pd_\lambda c^\mu)
g^{\nu\lambda} - (\pd_\lambda c^\nu)g^{\mu\lambda},\end{aligned}$$ and $S_{I} (g, \phi)$ is the action term containing the interaction between gravity and matter. The BRST symmetry for the action (\[total-action-gravity\]) can be written as $$\begin{aligned}
& \dis{g_{\mu\nu} \rar g_{\mu\nu} + i{\cal L}_c g_{\mu\nu} \ve }, \no \\
& \dis{\phi \rar \phi + i{\cal L}_c \phi \ve }, \no \\
& \dis{c \rar c - \frac{1}{2}{\cal L}_c{c}\ve}, \no \\
& \dis{b \rar b + \frac{1}{\a}F\ve,}
\label{BRST}\end{aligned}$$ $\ve$ is Grassmannian constant parameter of the BRST symmetry, $\ve^2= 0.$ This invariance at the level of quantum theory can be transformed in a usual way [@SF] to Slavnov–Taylor (ST) identity that is the equation for the Legendre transform of the logarithm of the path integral. This Legendre transform is performed with respect to external sources of the theory which are coupled in the path integral to the quantum fields from the measure of the path integral. To do this one defines the path integral extended by the dependence on the following external sources $$\begin{aligned}
& \dis{Z[T,~\eta,~\rho,~K,~k,~L]
= \int~dg_{\mu\nu}~dc_{\lambda}~db~d\phi~
\exp i}\left[\dis{S} \right. \no \\
& \left. + \dis{\int~d^4x~T^{\mu\nu}g_{\mu\nu} + \int~d^4x~j~\phi
+ i\int~d^4x~\eta_{\mu}c^{\mu}
+ i\int~d^4x~\rho b} \right. \label{path-gravity}\\
& \left. + ~\dis{i\int~d^4x~K^{\mu\nu}{\cal L}_c g_{\mu\nu} +
i\int~d^4x~k{\cal L}_c \phi -
\int~d^4x~L_\mu \frac{1}{2} {\cal L}_c{c^\mu}}\right],
\no\end{aligned}$$ where new external sources $K^{\mu\nu},$ $k,$ and $L_\mu$ coupled to the BRST variations of the metric, matter field and the ghost field under group of diffeomorphisms are introduced, respectively. The effective action $\G$ is related to $W = i~ln~Z$ by the Legendre transformation $$\begin{aligned}
& \dis{g_{\mu\nu} \equiv - \frac{\del W}{\del T^{\mu\nu}},~~~
\phi \equiv - \frac{\del W}{\del j},
~~~ ic^\mu \equiv - \frac{\del W}{\del \eta_\mu}, ~~
ib \equiv - \frac{\del W}{\del \rho}}, ~~\label{defphi}\\
& \dis{\G = - W - \int~d^4x~T^{\mu\nu}g_{\mu\nu}
- \int~d^4x~j~\phi
- i\int~d^4x~\eta_\mu c^\mu - i\int~d^4x~\rho b}\no \end{aligned}$$ If all equations Eq. (\[defphi\]) can be inverted, $$\begin{aligned}
& \Omega = \Omega[\varphi,K^{\mu\nu},k,L_\mu], \\
& \dis{\Omega \equiv \le T^{\mu\nu},j,\eta_\mu,\rho\ri, ~~~
\varphi \equiv \le g_{\mu\nu}, \phi, c^\mu, b\ri}. \no\end{aligned}$$ the effective action can be defined in terms of new variables, $\G = \G[\varphi,K^{\mu\nu},k,L_\mu].$ Hence the following equalities hold: $$\begin{aligned}
& \dis{\frac{\del \G}{\del g_{\mu\nu} } = - T^{\mu\nu} , ~~~
\frac{\del \G}{\del \phi} = - j, ~~~
\frac{\del \G}{\del K^{\mu\nu}} = - \frac{\del W}{\del K^{\mu\nu}}},
\label{GW}\\
& \dis{\frac{\del \G}{\del k}
= - \frac{\del W}{\del k}, ~~~
\frac{\del \G}{\del c^\mu} = i\eta_\mu, ~~ \frac{\del \G}{\del b} =
i\rho, ~~ \frac{\del \G}{\del L_\mu} = - \frac{\del W}{\del L_\mu}}. \no\end{aligned}$$
If the transformation Eq. (\[BRST\]) is made in the path integral Eq. (\[path-gravity\]) one obtains (as the result of the invariance of the integral Eq. (\[path-gravity\]) under a change of variables) the ST identity: $$\begin{aligned}
& \dis{\left[\int~d^4x~T^{\mu\nu}\frac{\del}{\del K^{\mu\nu}}
+ \int~d^4x j\frac{\del }{\del k}
- \int~d^4x~i\eta_\mu\le\frac{\del}{\del L_\mu}\ri \right.} \no\\
& + \dis{\left.
\int~d^4x~i\rho\frac{1}{\a}F\left[\frac{\del}{\del T^{\mu\nu}}\right]
\right]W} = 0,\end{aligned}$$ or, taking into account the relations Eq. (\[GW\]), we have $$\begin{aligned}
& \dis{\int~d^4x~\frac{\del \G}{\del g_{\mu\nu}}\frac{\del \G}{\del K^{\mu\nu}} +
\int~d^4x~\frac{\del \G}{\del \phi}
\frac{\del \G}{\del k}
+ \int~d^4x~\frac{\del \G}{\del c^\mu}\frac{\del \G}{\del L_\mu}} \no\\
& \dis{- \int~d^4x~\frac{\del \G}{\del b}\le\frac{1}{\a}F[g_{\mu\nu}]\ri}
= 0.
\label{STrM}\end{aligned}$$ In addition to ST identity also there is the ghost equation that can be derived by shifting the antighost field $b$ by an arbitrary field $\ve(x)$ in the path integral. The consequence of invariance of the path integral with respect to such a change of variable is (in terms of the variables (\[defphi\])) [@SF] $$\begin{aligned}
\frac{\del \G}{\del b} + \frac{\del F}{\del g_{\mu\nu}}
\frac{\del \G}{\del K^{\mu\nu} } = 0.
\label{ghost}\end{aligned}$$ The ghost equation (\[ghost\]) restricts the dependence of $\G$ on the antighost field $b$ and the external source $K_M$ to an arbitrary dependence on the combination $$\begin{aligned}
\frac{\del F}{\del g_{\mu\nu}} b(x) - K^{\mu\nu}(x). \label{comb}\end{aligned}$$ The main idea of Refs. [@Cvetic:2002dx; @Cvetic:2002in] is that the momentum-dependent part of the $Lcc$ correlator related to the superficial divergence (divergent in the limit of removing regularization) is invariant itself with respect to ST identity in each order of the perturbation theory. According to Ref. [@Cvetic:2002dx; @Cvetic:2004kx], this invariance results in the following integral equation for the part of the correlator $Lcc$ corresponding to the superficial divergence $\sim \ln\frac{p^2}{\Lambda^2}$ $$\begin{aligned}
& \dis{\int~dx~\G_{\Lambda}(y',x,z')\G_{\Lambda} (x,y,z) = \int~dx~\G_{\Lambda}
(y',y,x)\G_{\Lambda} (x,z,z')} \no\\
& = \dis{\int~dx~ \G_{\Lambda} (y',x,z)\G _{\Lambda} (x,z',y),} \label{integral}\end{aligned}$$ where $\G_{\Lambda} (x,y,z)$ is the scale-dependent part of the correlator $Lcc$ corresponding to the superficial divergence in each order of the perturbation theory [@Cvetic:2004kx], $\Lambda$ is a scale of ultraviolet regularization. The most general parametrization of this correlator is $\G (x,y,z)$ [^1], $$\begin{aligned}
\dis{\G \sim \int~dx~dy~dz~\G (x,y,z)
L_\mu(x)c^\lambda (y) \pd_\lambda c^{\mu}}. \label{A1}\end{aligned}$$ As has been shown in Ref. [@Cvetic:2002dx], the only solution to the integral equation (\[integral\]) is $$\begin{aligned}
& \dis{\G_{\Lambda}(x,y,z) = \int~dx'~G_{c}(x'-x)~G^{-1}_{c}(x'-y)
~G^{-1}_{c}(x'-z),} \label{result}\end{aligned}$$ where $G_c(x)$ is some unspecified function, $\int d^4x' G_c(x-x')G^{-1}_c(x'-y) =
\del(x-y).$ The complete correlator $Lcc$ can be then written as $$\begin{aligned}
& \dis{\int~dx~dy~dz~\G (x,y,z) L_\mu(x)
c^\lambda (y) \pd_\lambda c^{\mu}(z) } = \\
& = \dis{\int~dx'dy'dz'dxdydz~\tilde{\G} (x',y',z')G_{c}(x'-x)~G^{-1}_{c}(y'-y)}
\times \no\\
& \times~\dis{G^{-1}_{c}(z'-z) L_\mu(x)
c^\lambda (y) \pd_\lambda c^{\mu}(z)} \no\\
& = \dis{\int~dx'dy'dz'~\tilde{\G} (x',y',z')
\tilde{L}_\mu(x')\tilde{c}^\lambda (y') \pd_\lambda \tilde{c}^{\mu}(z')}.\end{aligned}$$ Here $\tilde{\G} (x',y',z')$ is the kernel of $Lcc$ correlator written in terms of dressed fields $\tilde{L}_\mu$ and $\tilde{c}^\mu$ defined by the convolutions $$\begin{aligned}
& \dis{ \tilde{L}_{\mu}(x) = \int~d^4 x' G_{c}(x-x')L_\mu(x'),} \\
& \dis{ \tilde{c}_{\mu}(x) = \int~d^4 x' G^{-1}_{c}(x-x')c_\mu(x').}\end{aligned}$$
As has been shown in Ref. [@Cvetic:2004kx], the consequence of such a structure for the $Lcc$ correlator is that the effective action can be expressed in terms of the dressed effective fields for the rest of proper correlators. The effective action expressed in terms of the dressed effective fields possesses some kernels which in fact are constituent parts of the scattering amplitudes. In ${\cal N} = 4$ super-Yang–Mills theory such kernels do not depend on the ultraviolet regularization scale [@Cvetic:2004kx]. It is known for a long time that gravity cannot be renormalized by adding finite number of counterterms to remove all the divergences. One can think that gravity is regularized in some way that conserves symmetry of the classical action, for example by dimensional regularization. These kernels will depend on the UV scale and on mutual distances of $n$ spacetime coordinates of some $n$-point proper Green function. For scattering amplitudes of gravitons in multi-Regge kinematic region of the momentum space these kernels can be calculated in the leading logarithmic approximations as it has been shown in Ref. [@Lipatov:1991nf]. We will use this action in the next section.
The structure of background metric in quantum theory of gravity is open question [@Hawking'96]. To analyze the structure of the background metric would be possible if we knew physical part of the effective action exactly. The minimum of the physical part determines the background metric. The background is to be found by solving the equation of motion for the effective action in terms of the dressed effective fields of matter sector and gravity sector. We will assume that the solution to such equations of motion gives Minkowski or AdS metric on the gravity side. However, metric perturbations about the background geometry in course of particle collisions can become strong and horizon forms.
Graviton scattering in four dimensions
======================================
Based on the construction of the previous section, let us begin with the pure four-dimensional gravity without any higher dimensional additions. Typical examples considered in literature concern collision of two shock waves in Aichelburg-Sexl gauge [@Aichelburg:1970dh]. There are two physically different cases, both of which have been analyzed in references at the semiclassical level. First case is the scattering of one particle by a shock wave of another. Second case is gravitational collapse of two particles into black hole. The effective parameter to differentiate these two cases is fraction of the effective Schwartzschild radius to the impact parameter of the problem. In the first case the black hole does not appear.
Lipatov [@Lipatov:1991nf] has proposed to calculate scattering amplitudes of gravitons from some effective action which is local. This action restores unitarity which is lost in the leading approximation. According to Ref. [@Amati:1993tb] the eikonal amplitude can be calculated by solving equations of motion that come from the extremizing the Lipatov action and estimating the value of this action on the effective trajectories of motion. The action of Ref. [@Lipatov:1991nf] gives the same results for the tree level scattering amplitudes of gravitons. In Ref. [@Amati:1993tb] it has been checked that the Lipatov action reproduces eikonal amplitudes correctly.
The Lipatov effective action [@Lipatov:1991nf] is $$\begin{aligned}
\int d^4x~ L_{eff} = \int d^4x~\le L_0 + L_e + L_r \ri, \end{aligned}$$ where $L_0$ is the kinetic term of relevant degrees of freedom, $L_e$ is the graviton emission-absorption term, $L_r$ is rescattering term, which is usually neglected [@Amati:1993tb]. The exact form of these terms is $$\begin{aligned}
L_0 = -2\pd h^{--} \pd^{*} h^{++} + \pd_{+}{\pd^{*}}^2~~\phi
\pd_{-}\pd^2\phi^{*}. \end{aligned}$$ Here $h^{--},$ $h^{++}$ are longitudinal degrees of freedom of the metric $h^{\mu\nu}.$ They have been defined by using the only non-zero components of the energy-momentum tensor $$\begin{aligned}
& T_{--} = kE\del(x^{-})\del^2({\bf{x}}), ~~~
T_{++} = kE\del(x^{+})\del^2({\bf{x-b}}), \\
& x^{\pm} = x^{0} \pm x^{3}, ~~~ {\bf{x}} = (x^1,x^2), ~~~ z = x^1 + ix^2, ~~~
\pd^{*} = {\pd}/{\pd{z^*}}, ~~~ k^2 = 8\pi G,\end{aligned}$$ in the following way $$\begin{aligned}
T_{\mu\nu}(x)h^{\mu\nu}(x) = T_{--}(x)h^{--}(x) + T_{++}(x)h^{++}(x),\end{aligned}$$ where $x$ is four-dimensional coordinate. The $L_e$ is interaction Lagrangian. It is important for estimating two-loop correction. In order to estimate leading order correction $L_e$ is not necessary. Equations of motion that come from the Lipatov action can be factorized [@Amati:1993tb] by the substitution $$\begin{aligned}
& h^{++}(x) = kE\del(x^{-})a(z,z^*), ~~~
h^{--}(x) = kE\del(x^{+})\bar{a}(z,z^*) \\
& \phi = \frac{1}{2}k(kE)^2\frac{1}{2}\t(x^{+}x^{-})\varphi(z,z^*) \\
& \bar{a}(b-z,b-z^*) = a(z,z^*).\end{aligned}$$
The physical picture depends on how large is impact parameter $b = |\bf{b}|$ in comparison with the effective Schwartzschild radius $R_S.$ If $b$ is much larger than $R_S$ the black hole does not appear and one has pure gravitational scattering process [@Amati:1993tb]. In the limit case $b >> R_S,$ the effective equations of motion are free field solutions which are Aichelburg-Sexl shock waves [@'tHooft:rb], $$\begin{aligned}
a(z,z^*) = - \frac{1}{2\pi}\log \frac{zz^*}{L^2},~~~\varphi = 0, \label{eqs}\end{aligned}$$ where $L$ is a large-distance scale, which plays the role of infrared cutoff. Moreover, the Lipatov action has reproduced in the next approximation with finite impact parameter $b$ the result of direct calculations of Ref. [@Amati:1990xe] of the eikonal amplitude up to two-loop level [@Amati:1993tb]. The value of the effective action on the effective equations of motion (\[eqs\]) is [@Amati:1993tb] $$\begin{aligned}
& \dis{A(a,\bar{a},\varphi) = 2\pi Gs \le \bar{a}(0) + a(b) +
\int d^2x\left[\bar{a}\pd^*\pd a + a\pd^*\pd\bar{a}\right]\ri} \\
& \dis{= 2\pi Gs \frac{1}{2} \le \bar{a}(0) + a(b)\ri = -2Gs\log\frac{b}{L}},\end{aligned}$$ which is phase of eikonal $\del(s,b).$ Then the leading eikonal function is equal to [@Amati:1993tb] $$\begin{aligned}
S = \exp\le-2\frac{i}{\hbar}Gs\log\frac{b}{L}\ri. \label{eikonal}\end{aligned}$$ Then by going from $s,b$ representation to $s,t$ representation one can write [*two particles $\rar$ two particles*]{} process scattering amplitude $$\begin{aligned}
& \dis{a(s,b) = e^{i\del(s,b)}}, \no\\
& \dis{ a(s,t) = \frac{1}{(2\pi)^2} \int~d^2x~e^{iq_ix^i}
e^{i\del(s,b)}} \no\\
& = \dis{\frac{1}{2\pi}\int_0^{\infty}dbbJ_{0}(bq) e^{i\del(s,b)}}. \label{st}\end{aligned}$$ Taking into account (\[eikonal\]), we can reproduce the ’t Hooft result [@'tHooft:rb] for the amplitude, $$\begin{aligned}
\dis{a(s,t) \sim G \frac{s}{t}}.\end{aligned}$$ However, this is only scattering amplitude of two particles in Minkowski background geometry. When the scattering parameter is small, the black hole formation can occur. This case has been investigated in Refs. [@D'Eath:hb; @D'Eath:hd; @D'Eath:qu] for the case of axisymmetric ($b = 0$) two black holes collision. The form of the trapped surface which appears in such a collision has been found there. The criteria of the black hole formation is appearance of this trapped surface [@HAWELL]. The form of the trapped surface in four-dimensional collisions of two particles for $b \ne 0$ has been found in Ref. [@Eardley:2002re]. It has been found there that the correction changes form of the geometrical cross-section a little by some factor of the order of 1.
Graviton scattering in RS2 case
===============================
The same idea can be applied to five-dimensional case. Scattering amplitudes can be considered as value of the effective action calculated on the solution to the equations of motion. Here we have Randall-Sundrum model with infinite fifth dimension (RS2 scenario). Emparan [@Emparan:2001ce] has extended the analysis of the shock wave of an ultrahigh-energy particle to scenarios with extra dimensions. The metric of shock wave in five dimensions has the form $$\begin{aligned}
ds^2 = dy^2 + e^{-2y/R}(-dudv + dx^idx^i + h_{uu}(u,x_i,y)du^2), \end{aligned}$$ where $u,v$ are light-cone coordinates and $x_i$ are coordinates in the directions transverse to propagation, $y$ is the coordinate in fifth dimension. The equations of motion for the shock wave addition to AdS space $h_{uu}$ has been solved analytically [@Emparan:2001ce] in the linearized approximation and then it has been argued that the solution to the linearized equation is in fact exact in this model [@Emparan:2001ce]. The analysis analogous to the four-dimensional case can be applied again and one obtains that in case $b > R$ the eikonal phase is $$\begin{aligned}
\dis{\del(s,b) = -G_4s\le\log\frac{b}{R} - \frac{R^2}{2b^2}\ri}.\end{aligned}$$ The second term is the Kaluza-Klein correction to the pure four-dimensional eikonal term, $G_4$ is four-dimensional coupling of gravity.
In the limit $r<<R$ the solution for $h_{uu}$ is [@Emparan:2001ce] $$\begin{aligned}
\dis{h_{uu} = -4G_4 p\del (u)\left[-\frac{R}{r} +
\frac{3}{2}\log(r/R) + \frac{3r}{8R}
+ \dots \right],}\end{aligned}$$ and this means that eikonal scattering phase is $$\begin{aligned}
\dis{\del(s,b) = \frac{G_4sR}{2b}.}\end{aligned}$$ In the $s,t$ representation (\[st\]) the amplitude is [@Emparan:2001ce] $$\begin{aligned}
a(s,t) \sim G_4R \frac{s^2}{\sqrt{-t}}.\end{aligned}$$ This is the case of scattering. Again, as in the four dimensional case one has to search for the trapped surface to trace the black hole formation. At present, the result for trapped surface are absent in literature. Some estimations of the shape of horizon for the point static source on the brane have been done by Giddings [@Giddings].
Giddings’ estimation for FB in gravity with extra dimensions
============================================================
As is suggested, there is a minimum of the effective action $\Gamma$ in the form of AdS metric $$\begin{aligned}
ds^2 = dy^2 + e^{-2y/R}ds_M^2,\end{aligned}$$ where $ds_M^2$ is Minkowski interval and that equations of motion of two particles in this background are non-linearized Einstein equations. One can consider static source of perturbations from the background metric on the brane as product of collision. In linearized approximation we just follow by the idea of Giddings when estimating size of horizon by the requirement $h_{00} \sim -1.$ The field describing the perturbation about background is related to the variable of the effective action whose derivatives with respect to that variable are vertices of gravitons in this background.
The metric of the theory in extra dimension is taken in the form of perturbed AdS metric. It can be considered as some solution in low energy action of string theory. The perturbed AdS metric takes the form [@Giddings] $$\begin{aligned}
ds^2 = \le 1+h_{yy}\ri dy^2 +
e^{-2y/R}\le\eta_{\mu\nu} + h_{\mu\nu}\ri{dx^\mu}{dx^\nu}\end{aligned}$$ where $x$ are coordinates of our four dimensional world and $y$ is a coordinate in fifth extra dimension. The energy-momentum tensor is a source localized on the IR brane, $$\begin{aligned}
T_{\mu\nu}(x,y) = S_{\mu\nu}(x)\del(y), ~~~ T_{yy} = T_{y\mu} = 0.\end{aligned}$$ The point mass source considered in [@Giddings] is $$\begin{aligned}
S_{\mu\nu}(x) = 2m\del^{d-1}(x)\del_{\mu}^0\del_{\nu}^0.\end{aligned}$$ Linearized equations of motion can be solved by using Neumann Green’s function $$\begin{aligned}
\Box\Delta_{d+1}(x,y;x',y') = \frac{\del^d(x-x')\del(y-y')}{\sqrt{-G}}. \end{aligned}$$ Under Neumann boundary conditions, the solution takes the form $$\begin{aligned}
& \dis{h_{\mu\nu}(x,y) =
-\frac{1}{2M_P^{d-1}}\int~d^dx'\sqrt{-g}\Delta_{d+1}(x,y;x',0)
\left[S_{\mu\nu}(x') - \eta_{\mu\nu}\frac{S_\lambda^\lambda(x')}{d-1} \right.} \no\\
& + \dis{\left.\frac{\pd_\mu\pd_\nu}{\pd^2}\frac{S_\lambda^\lambda(x')}{d-1}\right].} \end{aligned}$$ This behaviour of the metric fluctuations $h_{\mu\nu}$ significantly simplifies at the large distances in four-dimensional space. The most interesting component $h_{00}$ has the following behaviour $$\begin{aligned}
h_{00}(x,y) = -\frac{km}{RM_P^{d-1}}\frac{e^{-M_1r+yd/R}}
{r^{d-3}}. \end{aligned}$$ Here $k$ is a numerical constant and $M_1 = C/R,$ where $C$ is some number. This component of the metric is important since it is responsible for the horizon formation, when $h_{00} = -1.$ Taking logarithm of the both parts, we come to the equation $$\begin{aligned}
M_1r -yd/R = \ln\left[\frac{km}{RM_P^{d-1}r^{d-3}}\right].\end{aligned}$$ By solving this equation, one comes to the estimation of the area of the horizon in this model, $$\begin{aligned}
r_h(m) \sim \frac{1}{M_1} \ln\left[\frac{kmM_1^{d-3}}{RM_P^{d-1}}\right].\end{aligned}$$ Thus, the radius of horizon has logarithmic behavior in this brane model, in complete correspondence to the Froissart bound [@Giddings].
Conclusion
==========
In this paper we have analysed general structure of the effective action for quantum gravity. We have shown that the effective action has the dressed effective fields as actual variables of the effective action. Kernels of the effective action written in terms of the dressed effective fields are constituent parts of the scattering amplitudes. In the eikonal approximation the scattering amplitudes can be found by solving the equations of motion derived from the Lipatov effective action.
Work of I.K. is supported by Ministry of Education (Chile) under grant Mecesup FSM9901, by DGIP UTFSM, and by Fondecyt (Chile) grants \#8000017 and \#1040368. I.K. is grateful to Department of Physics of Brown University where this work has been started for the kind hospitality and financial support during his stay.
[99]{} J. M. Maldacena, Adv. Theor. Math. Phys. [**2**]{} (1998) 231 \[Int. J. Theor. Phys. [**38**]{} (1999) 1113\] \[arXiv:hep-th/9711200\].
S. Das, G. Gibbons, and S. Mathur, Phys.Rev.Lett. [**78**]{} (1997) 417.
S. N. Solodukhin, Phys. Lett. B [**533**]{} (2002) 153.
A. Jevicki and J. Thaler, Phys. Rev. D [**66**]{} (2002) 024041 \[arXiv:hep-th/0203172\].
H. J. Matschull, Class. Quant. Grav. [**16**]{} (1999) 1069 \[arXiv:gr-qc/9809087\].
B. Giddings, Phys.Rev.D [**67**]{} (2003) 126001.
M. Froissart, Phys. Rev. [**23**]{} (1961) 1053.
I. Kondrashuk, JHEP [**0011**]{} (2000) 034 \[arXiv:hep-th/0007136\].
G. Cvetic, I. Kondrashuk and I. Schmidt, Phys. Rev. D [**67**]{} (2003) 065006 \[arXiv:hep-ph/0203014\].
G. Cvetic, I. Kondrashuk and I. Schmidt, Phys. Rev. D [**67**]{} (2003) 065007 \[arXiv:hep-ph/0210185\].
I. Kondrashuk, Mod. Phys. Lett. A [**19**]{} (2004) 1291 \[arXiv:gr-qc/0309075\].
A.A. Slavnov and L.D. Faddeev, [*Introduction to the quantum theory of gauge fields,*]{} Moscow, Nauka, 1988.
C. Becchi, A. Rouet, R. Stora, Comm.Math.Phys. [**42**]{} (1975) 127; I.V. Tyutin, preprint FIAN [*“Gauge invariance in field theory and statistical physics in operator formalism”,*]{} Lebedev-75-39 (in Russian), 1975.
G.W. Gibbons and S.W. Hawking, Phys.Rev.D [**15**]{} (1977) 2752
S.W. Hawking, Phys.Rev.D [**53**]{} (1996) 3099
G. Cvetic, I. Kondrashuk and I. Schmidt, arXiv:hep-th/0407251.
P. C. Aichelburg and R. U. Sexl, Gen. Rel. Grav. [**2**]{} (1971) 303.
L. N. Lipatov, Nucl. Phys. B [**365**]{} (1991) 614.
D. Amati, M. Ciafaloni and G. Veneziano, Nucl. Phys. B [**403**]{} (1993) 707.
G. ’t Hooft, Phys. Lett. B [**198**]{} (1987) 61.
D. Amati, M. Ciafaloni and G. Veneziano, Nucl. Phys. B [**347**]{} (1990) 550.
P. D. D’Eath and P. N. Payne, Phys. Rev. D [**46**]{} (1992) 658.
P. D. D’Eath and P. N. Payne, Phys. Rev. D [**46**]{} (1992) 675.
P. D. D’Eath and P. N. Payne, Phys. Rev. D [**46**]{} (1992) 694.
S.W. Hawking and G.F.R. Ellis, [*The large scale structure of space time,*]{} Cambridge University Press, 1973. D. M. Eardley and S. B. Giddings, Phys. Rev. D [**66**]{} (2002) 044011
R. Emparan, Phys. Rev. D [**64**]{} (2001) 024025
[^1]: The structure of indices in (\[A1\]) is appropriate for the part of the correlator $Lcc$ related to its superficial divergence at least.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We obtained \[O III\] narrow-band imaging and multi-slit MXU spectroscopy of the blue compact dwarf (BCD) galaxy NGC 1705 with FORS2@VLT to derive chemical abundances of PNe and H II regions and, more in general, to characterize the properties of the ionized gas. The auroral \[O III\]$\lambda4363$ line was detected in all but one of the eleven analyzed regions, allowing for a direct estimate of their electron temperature. The only object for which the \[O III\]$\lambda4363$ line was not detected is a possible low-ionization PN, the only one detected in our data. For all the other regions, we derived the abundances of Nitrogen, Oxygen, Neon, Sulfur and Argon out to $\sim$1 kpc from the galaxy center. We detect for the first time in NGC 1705 a negative radial gradient in the oxygen metallicity of $-0.24 \pm 0.08$ dex kpc$^{-1}$. The element abundances are all consistent with values reported in the literature for other samples of dwarf irregular and blue compact dwarf galaxies. However, the average (central) oxygen abundance, $12 + \log(O/H)=7.96 \pm 0.04$, is $\sim$0.26 dex lower than previous literature estimates for NGC 1705 based on the \[O III\]$\lambda4363$ line. From classical emission-line diagnostic diagrams, we exclude a major contribution from shock excitation. On the other hand, the radial behavior of the emission line ratios is consistent with the progressive dilution of radiation with increasing distance from the center of NGC 1705. This suggests that the strongest starburst located within the central $\sim$150 pc is responsible for the ionization of the gas out to at least $\sim$1 kpc. The gradual dilution of the radiation with increasing distance from the center reflects the gradual and continuous transition from the highly ionized H II regions in the proximity of the major starburst into the diffuse ionized gas.'
author:
- 'F. Annibali , M. Tosi , A. Pasquali , A. Aloisi , M. Mignoli, D. Romano'
title: Chemical abundances and properties of the ionized gas in NGC 1705
---
Introduction
============
Dwarf galaxies play a key role in our understanding of galaxy formation and evolution. Within the framework of hierarchical formation, they are considered the first systems to collapse, supplying the building blocks for the formation of more massive galaxies through merging and accretion [e.g. @kau93]. Moreover, late-type dwarf galaxies (dwarf irregular (dIrr) and blue compact dwarf (BCD) galaxies), with their low metallicity, high gas content, and blue colors, resemble the properties of primordial galaxies [e.g. @it99], and are the best laboratories where to study star formation processes similar to those occurring in the early Universe.
Because of their poorly evolved nature, dIrrs and BCDs where initially suggested to be young galaxies at their first bursts of star formation. However, studies of the star formation histories (SFHs) based on color-magnitude diagrams (CMDs) of the resolved stars in a large number of dwarf galaxies within the Local Group and beyond [see @tht09 for a review] showed that these systems started forming stars as long ago as the look-back time set by the available photometry. In practice, all of them were already active at least $\sim 1-2$ Gyr ago, and, when the photometry was deep enough to reach the horizontal branch or the most ancient main-sequence turnoffs, or when RR Lyrae stars were found, at epochs as old as $\sim$ a Hubble time [e.g. @dolphin00; @clementini03; @mcconnachie06]. This poses a major challenge for chemical evolution models of dIrrs and BCDs, which have to reconcile the low observed metallicity with the old ages, roughly continuous SFHs, and fairly high star formation rates of the most metal-poor systems.
Since three decades, metal enriched winds seem to be the most viable mechanism to reconcile the high rates of star formation observed in starburst dwarf galaxies with their low metal abundances [e.g. @mt85; @pilyugin93; @marconi94; @carigi95; @romano06]. Outflows driven by supernova explosions and stellar winds are in fact expected to play an especially important role in these systems, whose relatively shallow potential wells make them susceptible to wind-driven loss of gas and newly created metals [e.g. @dekel86; @martin99]. In practice, even the accretion of large amounts of metal-free gas is usually insufficient to explain the low chemical abundances observed in dwarf galaxies with very intense episodes of star formation, where the metal production is quite high. Alternative models have sometimes been proposed, but never succeded in reproducing all the observed properties of starburst dwarfs. For instance, recently, a scenario that considers infall of primordial external gas, but does not include any outflow or galactic wind, has been proposed by [@gavilan13] to explain the observed properties of gas-rich dwarf galaxies. However, these models have proven to be successful in reproducing the gas fractions, chemical abundances and photometric properties only for “quiet” field dwarf galaxies, with recent star formation activity but not strong intensity episodes (i.e. dIrr galaxies). Moreover, NGC1705, as well as a few other starburst dwarfs, does show observational evidence of galactic winds [@meurer92; @papaderos94; @marlowe95].
Element abundance estimates in late-type dwarf galaxies are fundamental to constrain chemical evolution models. While H II regions define the present-day galaxy properties, chemical abundances for individual stars or planetary nebulae (PNe) offer a view of the galaxy chemical properties at earlier epochs, but are challenging outside the Local Group, where the most actively star-forming BCDs are found. In addition to the evolution of elements in time, the spatial behaviour of the chemical abundances plays a key role in constraining dwarf galaxy models. The majority of literature studies indicate that, within the observational uncertainties, dIrrs and BCDs show nearly spatially constant chemical abundances [e.g., @kobul97; @croxall09; @lagos13; @haurberg13]. This could indicate that the ejecta from stellar winds and supernovae are dispersed and mixed across the interstellar medium (ISM) on timescales of $<10^7$ yr, but an alternative possibility is that the freshly synthesized elements remain unmixed with the surrounding interstellar medium and reside in a hot $10^6$ K phase or a cold, dusty, molecular phase [@kobul97].
NGC 1705 is a late-type dwarf galaxy classified as BCD, famous for its high recent star formation (SF) activity and for the unambiguous evidence of galactic winds [@meurer92; @heckman01]. Several H II regions have been found in NGC 1705 [@melnick85; @meurer92], and most recently studied with EFOSC2 by [@lee04], hereafter LS04. Among their 16 analyzed H II regions, LS04 were able to measure the \[O III\]$\lambda$4363 line in 5 central (distances $\lesssim$300 pc from the galaxy center) regions, from which they inferred a spatially homogeneous oxygen abundance of $12+\log(O/H)=8.21 \pm0.05$ (i.e. Z$\sim$0.005), $\log(N/O)=-1.63\pm0.07$ and $\log(Ar/O)=-2.31\pm0.11$; they noticed large differences with the corresponding abundances estimated by [@heckman01] for the neutral gas observed with the FUSE satellite, where N, O and Ar are all significantly lower than in the H II regions [^1]. The chemistry of the neutral gas in NGC 1705 was confirmed with further FUSE data by [@aloisi05]. We observed NGC 1705 in UBVIJH with the HST/WFPC2 and resolved its individual stars down to V$=$28 mag, $\sim$2 mag below the RGB tip [@tosi01]. The HST-derived CMDs led to a safe determination of the galaxy distance (5.1 Mpc, i.e. $DM=(m-M)_0=28.54$, [@tosi01]) from the luminosity of its RGB tip, and to a detailed description of its SF history [@anni03; @anni09 hereafter, A03, A09]. One-zone chemical evolution models for NGC 1705 based on the SFH of A03, and assuming metal-enriched galactic winds, were able to reproduce the observed H II region metallicity in NGC 1705, but could not explain the low N/O ratio as derived by LS04 [@romano06].
We acquired FORS2@VLT data (programme ID 084.D-0248 and 086.D-0761, PI Tosi) with the aim of deriving chemical abundances in PNe and H II regions and, more in general, to characterize the properties of the warm gas in NGC 1705. Pre-imaging and spectroscopy are described in Sections 2 and 3, respectively. The emission line fluxes, corrected for underlying Balmer absorption and reddening, are given in Section 4. In Section 5 we provide the derived chemical abundances. In Section 6 we discuss the comparison with classical emission-line diagnostic diagrams. In Section 7, we discuss the comparison with literature data for NGC 1705 and for other samples of late-type dwarf galaxies, discuss relative element abundances, and provide evidence for a metallicity gradient. Our conclusions are in Section 8.
Pre-imaging
===========
FORS2 [@fors] pre-imaging exposures were taken in order to identify the PNe and H II regions selected to build the mask for our multi-object spectroscopy. We imaged NGC 1705 on March 25, November 15 and December 1, 2010 for a total of $\sim$5 h. Seeing ranged between 0.63 arcsec in March, 0.97 in November and 0.81 in December. Observations were performed with the standard resolution collimator in the narrow-band filters FILT500$-$5$+$85 (sampling the \[OIII\]$\lambda$5007 emission line) and OIII/6000$+$52 (centered at $\lambda=5105$ Å, and sampling the continuum emission adjacent to the \[OIII\] line). The total exposure time was $\sim$2.5 h for each filter, split into three subexposures. The images have a pixel scale of 0.25”/pixel and a field of view of 6.8’ $\times$ 6.8’.
After bias subtraction and flat field calibration, we combined the individual images into a single frame for each filter using the DRIZZLE task [@drizzle] in the IRAF environment[^2]. We normalized the images to account for the different filter widths, and applied a Gaussian smoothing to bring them to the same resolution; then we constructed a \[O III\] continuum-subtracted image, which is shown in Figure \[o3\].
The morphology of the \[O III\] image resembles that in H$\alpha$ already discussed by [@meurer92]. The ionized gas presents a bipolar morphology extending out to $\sim$100 ” ($\sim$2.5 kpc) from the galaxy center, and it is organized into thin filaments, loops, and arcs. The presence of “cellular” structures or cavities is striking, making the gas appear as “fragmented” at a global look. The cells have sizes from $\sim$1” to $\sim$4” ($\sim$25 pc to $\sim$100 pc), with a median value of $\sim$2” ($\sim$50 pc).
The ionized gas extends much further away than the optical image of the galaxy. This is shown in Figure \[rgb\], where we overplot isocontours for a portion of the \[O III\] continuum-subtracted image over a color image of NGC 1705 obtained combing WFPC2 images in the F380W ($\sim$U), F555W ($\sim$V), and F814W ($\sim$I) filters. The field of view is $\sim 70''\times 60''$, i.e. just $\sim$3% of the total field of view in Figure \[o3\]. Indicated are also some of the regions that we targeted for spectroscopy (see next section). The figure shows a strong concentration of ionized gas organized in a sort of shell around the high surface brightness optical component of the galaxy. The central region of NGC 1705, with the highest density of young star formation, corresponds to a zone of significantly lower \[O III\] surface brightness compared to the surrounding shell. A few blue stars are observed in correspondence of some of the \[O III\] knots: this is the case for s7-1, for the double-peaked knot just nearby it, and for s8-1. However, we will show in Section \[dd\] that, despite the presence of local association of young stars, the main ionization source seems to come from the central starburst in NGC 1705.
The \[O III\] continuum-subtracted image was used to identify PN candidates and H II regions. At the distance of NGC 1705 [$5.1\pm0.6$ Mpc or $DM=28.54\pm0.26$, @tosi01], PNe, with typical diameters smaller than 1 pc, should appear as point-like objects emitting in the \[OIII\] line. Thus we selected point-like sources in the \[OIII\] continuum-subtracted image to identify PN candidates, although we could not exclude that these sources were compact H II regions instead. PN central stars typically have absolute visual magnitudes of $M_V \sim-2.0$ [@mendez92], corresponding to an apparent magnitude of $m_V\sim26.5$ in NGC 1705, too faint to be detected in our continuum image. Therefore, when selecting PN candidates, we also checked that no continuum emission was associated to the point-like sources in the \[OIII\]-continuum subtracted image. Following these criteria, we identified 17 PN candidates in the frame, whose positions are indicated in Figure \[o3\].
Spectroscopy
============
Spectroscopy was obtained with FORS2 in the MXU configuration from January 2011 until January 2012 in service mode. The slit mask (see Figure \[o3\]), was constructed to observe the largest possible number of PN candidates, giving priority to the brightest ones. We were able to target 9 out of the 17 identified candidates (slit number s1, s2, s3, s4, s5, s11, s12, s13, s14). Two slits (s7 and s8) were positioned on H II regions observed by [@lee04], while three slits (s6, s9, and s10) were positioned on extended \[O III\] emission regions and gas filaments. We also positioned six slits on regions where no diffuse gas emission was observed, to serve as sky templates. These were used to subtract the background in those slits totally occupied by the emission. To avoid confusion, the sky slits are not shown in the mask of Figure \[o3\].
To cover the spectral range from the \[O III\]$\lambda$3727 to the \[S II\]$\lambda\lambda$6716,31 doublet the grism GRIS$-$300V was employed, alone in the blue, and coupled with the GG435 filter in the red. NGC 1705 was observed in the blue with 7$\times$2700 s exposures, for a total of 5.2 h, and in the red with 20$\times$2700 s exposures, for a total of $\sim$15 h. The seeing varied from $\sim$0.7” to $\sim$1.3” (median value of 1.1”) for the blue observations, and from $\sim$0.6” to 1.5” (median value of 0.9”) for the red observations, higher than our requirement of a seeing $\leq$0.8” . The pixel scale and the dispersion are 0.126 arcsec/pixel and $\sim$2 Å/pixel, respectively. The effective resolution is $\lambda/FWHM \sim780$ at $\lambda \sim5500$ Å. Wavelength calibration was based on four arc lamps (a He, a HgCd, and two Ar arc lamps) to provide a sufficient number of lines over the whole spectral range. Data were reduced with the standard procedure using the [*onedspec*]{} and [*twodspec*]{} packages inside IRAF. The science and arc frames were bias subtracted. For each science exposure, the individual 2d spectra were extracted from the frame. The exact same extraction apertures used for the science frames were used for the arc frames. Then, the 2d spectra were wavelength calibrated using the LONGSLIT [*identify*]{}, [*fitcoords*]{} and [*transform*]{} tasks. After that we subtracted the background. This step deserves a particular explanation: in the case of point-like sources (PN candidates or compact H II regions), we defined the background using two windows at the opposite sides of the central source, and performed the subtraction with the [*[background]{}*]{} task in IRAF. Typically, the background windows were chosen at a distance larger than $\sim$ 2 times the FWHM of the spatial profile from the source peak. This procedure removes both the sky emission and the background emission from NGC 1705. However, in presence of extended emission occupying the entire slit, it was not possible to define the background windows; then, we subtracted the sky templates from the object spectra. Notice that this procedure allows to subtract the sky emission but does not remove the galaxy background. We applied the sky template subtraction to slits number 6, 8, 9 and 10. For all the other slits we used the iraf “background” task. After background subtraction, the individual exposures were combined into a median 2d spectrum, for the blue and for the red observations separately. Then 1d spectra were extracted using the [*apall*]{} task. In the case of PN candidates, we centered the extraction aperture on the point-like source, while in the case of extended emission, multiple apertures were selected within one slit (e.g., apertures s6-1, s6-2, s7-1, s7-2, s8-1, s8-2, s8-3, s8-4, see Figure \[o3bis\]).
The 1d spectra were flux calibrated using the spectrophotometric standard stars Feige 56, GD 108, LTT 3218, LTT 4816, and Feige 110, EG 21, LTT 1020, LTT 2415, LTT 3218, respectively for the blue and for the red observations. The standards were observed with a 5 arcsec wide slit positioned at the center of the MOS instrument. After bias subtraction, wavelength calibration, and background removal, a 1d spectrum was extracted for each standard. To obtain the sensitivity function we ran the iraf tasks [*standard*]{} and [*sensfunc*]{}. The [*standard*]{} task provides the standard star observed counts, along with the associated calibration fluxes, over some pre-defined bandpasses. Then the [*sensfunc*]{} uses the [*standard*]{} output to determine the system sensitivity as a function of wavelength. In the process of deriving the sensitivity curve, we assumed the atmospheric optical extinction curve derived by [@patat11] for Cerro Paranal.
The [*calibrate*]{} task was run to obtain flux calibrated spectra. The spectra obtained with the blue (GRIS$-$300V) and red (GRIS$-$300V$+$GG435) configurations are shown in Figures \[spectra1\], \[spectra2\], \[spectra3\] and \[spectra4\]. We do not show the regions below $\sim$3600 Å and below $\sim$4500 Å for the blue and red spectra, respectively, because the instrumental sensitivity curves significantly decrease below these wavelengths and are very uncertain. Also, below $\sim$3600 Å, the atmospheric transmission is low. Figs. \[spectra1\] to \[spectra4\] show that the blue spectra extend up to $\sim$7000 Å and, in some cases, even redder. The red spectra always cover wavelengths redder than 7000 Å. The agreement in flux between the red and blue spectra is excellent, with the only exception of the region below $\sim$5000 Å, due to the higher uncertainty of the sensitivity curve in the red configuration below these wavelengths.
With the calibrated spectra in hands, we were able to asses the nature of our PN candidates. Unfortunately, almost the totality of them turned out to be background objects. In particular, the targets in slit s1, s2, s11, s12 and s14 show just one prominent emission line around $\sim$5000 Å, and no (or very weak) stellar continuum, and are likely to be Ly$\alpha$ emitting galaxies at $z\sim3$. The target in slit s14 also shows another prominent line at $\lambda\sim6157$ Å, identified as the C IV $\lambda$1539 line at $z\sim3$. The targets in s3 and s13 are two background galaxies at $z\sim0.3$. The object in s4 is likely a QSO.
At this point we are left only with the PN candidate in s5. Some specific lines and line ratios can be used to discriminate between compact H II regions and PNe. For instance, an He II $\lambda$4686 emission exceeding a few percent of H$\beta$ unquestionably indicates a PN; also, \[O III\]$\lambda$5007/H$\beta$ ratios larger than $\sim$ 4 are found only in PNe [e.g. @pena07; @magrini09]. From the blue spectrum of s5, the \[O III\]$\lambda$5007/H$\beta$ ratio is $\sim$3.4, lower than what typically observed in PNe, but still consistent with a low-ionization PN classification [e.g. @depew11; @gorny14]. On the other hand, due to the low signal-to-noise, the detection of the He II $\lambda$4686 line in s5 is challenging. For this reason, we performed some statistical analysis on the bidimensional spectrum in order to asses the presence of the line, and in case, the significance of the detection. More specifically, we derived the average counts in 6 $\times$ 7 px$^2$ boxes centered on the line and on different regions of the background. The significance of the detection was computed as $(<C>_{line} - <C>_{bg})/STD_{bg}$ , where $<C>_{line}$ and $<C>_{bg}$ are the average counts for the line and for the background, respectively, and $STD_{bg}$ is the standard deviation of the background computed from the different measurements. We obtain that the He II $\lambda$4686 line is detected at $\sim$3 sigma. From the 1D blue and red spectra (see next section), we measure He II $\lambda$4686/H$\beta$ ratios of 0.28$\sim$0.11 and 0.35$\sim$0.14, respectively. Concluding, there is a significant possibility that s5 is a low-excitation PN.
Emission line measurement
=========================
Emission line fluxes were derived with the [*deblend*]{} option available in the [*splot*]{} iraf task. We used this function to fit single lines, group of lines, or blended lines. Lines were fitted with Gaussian profiles, treating the centroids and the widths as free parameters. In presence of blended lines we forced the lines to have all the same widths. The continuum was defined choosing two continuum windows to the left and to the right of the line or line complex, and fitted with a linear regression. The final emission fluxes were obtained repeating the measurement several times with slightly different continuum choices, and averaging the results. The errors were derived from the standard deviation of the different measurements. The emission line measurements for the blue and red spectra are given in Tables 1 and 2. The errors, which in the case of bright lines can be very small (even lower than $\sim$1%), do not account for flux calibration errors or for other systematic uncertainties. A comparison between the fluxes measured for the lines in common between the blue and red spectra provides more realistic errors: the differences are typically within $\sim$4% for the brightest lines, and within $\sim$10% for the faintest lines.
Balmer absorption
-----------------
Particular attention was paid to the correction of the emission lines for Balmer absorption due to the underlying NGC 1705’ s stellar population. Balmer absorption lines are visible in our spectra as absorption wings in correspondence of the H$\delta$, H$\gamma$, and H$\beta$ emission lines, while no absorption wings are visible in the H$\alpha$ region due to the blend with the \[N II\]$\lambda$6548,84 emission lines. We attempted a fit of the absorption wings in order to recover the full strength of the Balmer absorption lines, but these were not deep enough to allow for a “stable” solution: the resulting absorption strengths were highly dependent on the particular choice of the input parameters (for instance, the windows for the continuum evaluation). Thus we adopted a different approach than a direct spectral fit to all the regions. More specifically, we selected a specific region in NGC 1705 with the minimum possible contribution of the emission with respect to the underlying stellar absorption (and with the deepest absorption wings), determined the corrections directly from its spectrum, and then extended the derived corrections to other regions of NGC 1705 using population synthesis models. In the following subsections we describe in details the three different steps of our semi-empirical approach : a) Derivation of the absorption lines from an empirical spectrum in a region of NGC 1705; b) Comparison with predictions from population synthesis models; c) Extension of the correction to all regions in NGC 1705 through population synthesis models.
### a) Absorption lines from empirical spectral data
Unfortunately, it was not possible from our data to extract an emission-free spectrum with a signal to noise high enough to allow for a reliable characterization of the Balmer absorption lines in NGC 1705: in fact, all the slits positioned near the galaxy center, where the stellar surface brightness is higher, are severely contaminated by the strong emission from the ionized gas, while the galaxy stellar surface brightness is too low in more external regions. Given the impossibility of working with an emission-free spectrum, we looked for a region with the highest possible strength of the absorption lines in comparison with the emission lines, to allow for a good fit of the absorption wings. This region is located between the two emission peaks at s7-1 and s7-2 (see Figure \[o3bis\], panel c), and we will call it region s7-a hereafter. The strong absorption wings around H$\delta$, H$\gamma$ and H$\beta$ in region s7-a are well visible in Figure \[balmer\].
To compute the corrections, we adopted the same procedure adopted by [@lee03] for a sample of dwarf irregular galaxies: this consists in fitting the Balmer regions first with a model constructed with both absorption and emission lines, and then with a model that assumes only emission lines (see Figure \[balmer\]). The correction is then computed as the difference between the emission strengths obtained with the “emission only” and with the “emission $+$ absorption” fits. In our specific case, we used the IRAF [*splot*]{} task to fit the Balmer absorption features with Voigt profiles (resulting from the convolution of a Gaussian and a Lorentzian profile), and the emission lines with Gaussians, while the linear continuum was estimated from two windows located on both sides of the emission lines, away from the Balmer absorption wings. For H$\delta$ and H$\beta$, we assumed a line in absorption plus a line in emission, with the peaks and the FWHMs treated as free-parameters. Instead, the more complex H$\gamma$ region was fitted assuming two absorption profiles (a Gaussian for the G band around $\sim$ 4300 Å and a Voigt profile for the H$\gamma$ in absorption) plus two Gaussians for the H$\gamma$ and \[O III\]$\lambda$4363 emission lines. Finally, the H$\alpha$ region (not shown in Figure \[balmer\]) was fitted adopting three Gaussians in emission for the H$\alpha$ $+$ \[N II\]$\lambda\lambda6548,84$ blend, plus a Voigt profile for the H$\alpha$ in absorption. Due to the absence of visible absorption wings, the parameters of the H$\alpha$ absorption line were fixed: the absorption equivalent width was taken from population synthesis models (see point [*b)*]{} in Section 4.1.2), and the (Lorentzian and Gaussian) FWHMs were fixed to the values obtained from the fits to the H$\delta$, H$\gamma$, and H$\beta$ lines.
The “emission $+$ absorption” fits are shown in the bottom panels of Figure \[balmer\], and the results are given in columns 2 and 3 of Table 3. The values were obtained repeating the fitting procedure several times, and computing average values and standard deviations. The absorption strengths for the H$\delta$, H$\gamma$ and H$\beta$ lines are $9.9 \pm 0.4$ Å, $10.2 \pm 0.6$ Å, and $8.9 \pm 0.5$ Å, respectively. The “emission only” fits are instead shown in the top panels of Figure \[balmer\], and the results are given in column 4 of Table 3. For each line, the final correction (column 5) was computed as the difference between the emission equivalent widths obtained with the “emission only” and “emission $+$ absorption” fits.
We notice that the corrections are lower than the full strength of the Balmer absorption lines; this is because the absorption lines are wider than the emission ones, with the consequence that only part of the absorption affects the emission line (see Figure \[balmer\]). We obtain corrections of 4.0 $\pm$ 0.2 Å, 4.5 $\pm$ 0.3 Å, 3.1 $\pm$ 0.4 Å, and 1.5 $\pm$ 0.6 Å to the equivalent widths in H$\delta$, H$\gamma$, H$\beta$, and H$\alpha$, respectively. Our values are higher than the average correction of 1.59$\pm$0.56 Å to H$\beta$ derived by [@lee03] for their sample of dwarf irregular galaxies, and also higher than the constant 2 Å correction usually applied in the literature to all Balmer lines [e.g., @mccall85; @kennicutt96; @lee03; @lee04]. Furthermore, our correction to the H$\beta$ line is roughly twice as high as that in H$\alpha$. Thus, applying the same correction to H$\beta$ and H$\alpha$ may lead to significant uncertainties in reddening estimates, in particular if the emission lines are weak. We also notice from Table 3 that there are non negligible corrections to the \[O III\]$\lambda$4363, \[N II\]$\lambda6548$, and \[N II\]$\lambda6584$ lines (0.9 $\pm$ 0.1 Å, 0.5 $\pm$ 0.1 Å, and 0.5 $\pm$ 0.1 Å, respectively), whereas such lines are usually not corrected in the literature.
### b) Comparison with population synthesis models
Since our results are based on fits to the absorption wings rather than on direct measurements of the Balmer absorption lines, we performed a consistency-check with population synthesis models. To this purpose, we used the results by [@anni03], hereafter A03, and [@anni09], hereafter A09, who derived the star formation history (SFH) in different regions of NGC 1705 (regions 7 to 0 from the center outwards) from the CMDs of the resolved stars [@tosi01]. From a comparison of our FORS2-MXU slit mask in Figure \[o3bis\] with WFPC2 images of NGC 1705 (see Figure 2 in A09), we find that region s7-a defined in Section 4.1.1 is located within region 5 of A09. We thus compare our fits to the empirical spectrum in s7-a (see Section 4.1.1) with population synthesis models based on the SFH from A03 and A09 for region 5.
In order to construct the synthetic spectrum, simple stellar populations (SSPs) from the Padova group for a metallicity of $Z=0.004$, based on the tracks of [@marigo08] [see also @chavez09], were combined according to the SFH of region 5 in NGC 1705. The adopted SSPs are based on the MILES empirical spectral library [@sanchez06], which consists of 2.3 Å FWHM optical spectra of $\sim$1000 stars spanning a large range of atmospheric parameters. Balmer absorption lines were then measured on the synthetic spectrum using, as for the real data, the [*splot*]{} task in IRAF. The results are given in Column 2 of Table 4. The synthetic absorption equivalent widths [^3], $H\delta=8.8 \pm 0.2$ Å, $H\gamma=8.7 \pm 0.2$ Å, and $H\beta=8.7 \pm 0.3$ Å are consistent, within the errors, with the values of $9.9 \pm 0.4$ Å, $10.2 \pm 0.6$ Å, and $8.9 \pm 0.5$ Å derived from the empirical spectrum in s7-a. This indicates that our fit to the absorption wings provided sensible values for the Balmer lines, and that our corrections are reliable. From the synthetic spectrum of region 5, we also measured an absorption EW in H$\alpha$ of 5.9 $\pm$ 0.2 Å, which we used in Section 4.1.1 to compute the emission corrections.
### c) Extension of the correction to other regions
Our corrections, which are given in terms of equivalent widths, do not depend on the continuum absolute flux, but depend on the (luminosity-weighted) age and metallicity of the underlying stellar population. Thus, given the spatial variation of the SFH within NGC 1705, we expect different amounts of Balmer absorption over the galaxy field of view. To extend the corrections computed for region 5 to other regions, we constructed synthetic spectra for regions 6 to 0 in NGC 1705 using the SFHs derived by A03 and A09, and combining the SSPs as described in Section 4.1.2. We want to stress out that the corrections were not directly derived from the synthetic spectra: what we did was instead to scale the corrections derived from the empirical spectrum in s7-a (Table 3) by a factor predicted from population synthesis models. In doing this, we used [*[ratios]{}*]{} of absorption lines as predicted by the models and [*not absolute values*]{}; this minimizes some of the uncertainties related to the adopted models and to the assumed input parameters (e.g., the IMF).
More in details, for each $X$ region (6, 4-3, or 2-1-0), the new corrections were obtained by multiplying the values in col. 5 of Table 3 by the quantity $EW_{reg X}/EW_{reg5}$, i.e. by the ratio between the absorption strengths from the synthetic spectra. The results are given in Table 4. The absorption lines from the synthetic spectra, the $EW_{reg X}/EW_{reg5}$ ratios, and the final corrections are given in columns 2, 3 and 4, respectively. The comparison of the FORS2-MXU slit mask with the WFPC2 images of NGC 1705 provides the following spatial association: s8-1 with region 6; s6-2, s7-1, s7-2, s8-2, s8-3, and s8-4 with region 5; s6-1 and s9 with region 4-3; s5 and s10 with region 2-1-0. Notice that no correction needs to be applied to spectra s5, s7-1 and s7-2, where both the sky and the galaxy background were subtracted, as described in Section 3.
Reddening correction
--------------------
For each region, the reddening was estimated from the observed H$\alpha$/H$\beta$ ratio using the Cardelli et al. (1989) extinction law according to the formula:
$$E(B-V)= \frac {log R_o/ R_t }{0.4\times R_V \times [A_{H\beta}/A_V - A_{H\alpha}/A_V ]}$$
where $R_o$ and $R_t$ are the observed and theoretical H$\alpha$/H$\beta$ ratios, respectively; the magnitude attenuation ratio $A_{\lambda}/A_V$ was taken from the Cardelli’ s law adopting $R_V = 3.05$. To compute $R_o$, the H$\alpha$ and H$\beta$ emission fluxes were corrected for underlying Balmer absorption as described in Section 4.1. We assumed a theoretical Balmer decrement of $R_t=2.8$ from Table 4.4 of [@oster89], assuming $n_e=100 \ cm^{-3}$ and $T_e=15,000$ K. In fact, the measured \[S II\] $\lambda$6716 / $\lambda$6731 ratios imply densities lower than this ($n_e \lesssim 30 \ cm^{-3}$, Tables 7 and 8) but from models [@panuzzo03] we verified that $R_t$ does not significantly depend on density below $n_e\sim 100 \ cm^{-3}$.
The $E(B-V)$ values derived from the blue and the red spectra are given in Tables 7 and 8. When the H$\alpha$/H$\beta$ ratio resulted into a negative reddening, we adopted $E(B-V)=0$. For many regions (the most external ones), the derived reddenings are consistent with zero or very low levels of extinction. However, along some, more central, lines of sight, we do detect significant extinction. For example, the extinction derived for s7-1 and s7-2 is as high as $E(B-V)=0.10$ and 0.16, respectively, which is significantly higher than the Galactic reddening of $E(B-V)=0.045$ from [@meurer92], or the newer $E(B-V)\sim0.007$ from [@schlegel] and [@schlafly]. We notice that our red spectra tend to provide systematically lower reddening values than the blue ones.
The intrinsic emission line fluxes for the blue and red spectra were computed according to the formula :
$$F_c = F_o \times 10^{0.4\times A_{\lambda}},$$
where $F_o$ is the measured line flux, corrected only for the effect of underlying Balmer absorption, and $F_c$ is the extinction-corrected flux; $A_{\lambda}$, the magnitude attenuation at the line wavelength, is derived from the Cardelli et al. (1989) extinction law. We provide in Tables 5 and 6 the emission line fluxes corrected for both Balmer underlying absorption and dust extinction. The H$\gamma$/H$\delta$ ratios from the corrected fluxes result into 1.80 $\pm$ 0.06, 1.77 $\pm$ 0.27, 1.79 $\pm$ 0.08, 1.57 $\pm$ 0.08, 2.03 $\pm$ 0.21, 1.92 $\pm$ 0.15, 1.86 $\pm$ 0.08, 1.87 $\pm$ 0.08, 1.81 $\pm$ 0.10, 1.73 $\pm$ 0.40 for regions s6-1, s6-2, s7-1, s7-2, s8-1, s8-2, s8-3, s8-4, s9, and s10, respectively. With the exception of region s7-2, they are all consistent, within the errors, with the theoretical ratio of H$\gamma/H\delta \sim 1.8$ predicted for case B recombination in [@oster89].
Chemical abundances
===================
Temperatures, densities, and chemical abundances were derived using two different tools, the [*abund*]{} task in IRAF [@abund] and the PyNeb tool [@pyneb], both based upon the FIVEL program developed by [@fivel]. The electron density $n_e$ was computed using the density-sensitive \[S II\] $\lambda$6716/ $\lambda$6731 diagnostic line ratio, which provides the density of the low-ionization ions. This density was assumed to be constant all over the nebula. For all the regions but s5, where the \[O III\]$\lambda$4363 was not detected, the \[O III\]($\lambda$4959$+$5007)/$\lambda$4363 diagnostic line ratio, measured in the blue spectra, was used instead to derive the temperature of the $O^{+2}$ ions (the so-called “direct” $T_e$ method). Both the IRAF [*abund*]{} task and the [*getCrossTemDen*]{} option in PyNeb adopt an iterative process to derive temperatures and densities: the quantity derived with one line ratio is inputted into the other, and the process is iterated; at the end of the iteration process, the two temperature-sensitive and density-sensitive diagnostics give self-consistent results. The high \[S II\]$\lambda\lambda$6716,31 line ratios measured in NGC 1705 imply very low densities. In those, numerous, cases where the \[S II\] line ratio exceeded the theoretical value for $n_e\rightarrow0$, we provided an upper limit for the density at the $-1 \sigma$ value of the measured \[S II\] line ratios. Typically, the derived densities are $\lesssim 30 \ cm^{-3}$. In those cases where a solution could not be found at the $-1 \sigma$ level, we assumed a density of $n_e = 30 \ cm^{-3}$ to compute the temperature. We checked that assuming $n_e = 10 \ cm^{-3}$ or $n_e = 100 \ cm^{-3}$ had no effect on the temperatures. Direct temperatures could be derived for all regions in NGC 1705 but s5, where the \[O III\]$\lambda$4363 line was not detected. The derived values are in the range $\sim 12 000 - 15000$ K. We provide in Tables 7 and 8 the temperatures and densities obtained with PyNeb. These results are consistent, within the errors, with the values obtained with the IRAF [*[abund]{}*]{} task, but we notice a systematic shift of $\sim +300$ K for the temperatures derived with [*abund*]{} with respect to those obtained with PyNeb.
Starting from the temperature of the $O^{+2}$ ions directly obtained from the \[O III\]($\lambda$4959$+$5007)/$\lambda$4363 ratio, we derived the temperatures of the other ions assuming that $T_e(Ne^{+2})=T_e(O^{+2})$, and $T_e(N^{+})=T_e(S^{+})=T_e(O^{0})=T_e(O^{+})$. In the literature, the relation between $T_e(O^{+2})$ and $T_e(O^+)$ has been widely discussed [e.g. @campbell86; @perez03; @izotov06; @pilyugin07]. In this work, we used the formula in equation (14) of [@izotov06] for the intermediate metallicity regime, i.e. $t_e (O^+)= -0.744 + t_e (O^{+2}) \times (2.338 - 0.610\times t_e (O^{+2}))$, where $t_e = T_e/10^4$ K. The temperature of the $Ar^{+2}$ and $S^{+2}$ ions was assumed to be intermediate between that of $O^+$ and $O^{+2}$, and from [@izotov06], eq. (15), we adopted $t_e (Ar^{+2})=t_e (S^{+2})= -1.276 + t_e (O^{+2}) \times (2.645 - 0.546\times t_e (O^{+2}))$. We used the \[N II\] $\lambda$6548, 6584 lines for the determination of the $N^+$ abundance, and the \[O II\] $\lambda$3727 and \[O III\] $\lambda$4959, 5007 lines for the $O^+$ and $O^{+2}$ abundances, respectively; when detected (regions s7-1, s8-1, s8-2, s8-3, and s8-4), the \[O I\] $\lambda$ 6300 line provided the $O^{0}$ abundance. We used the \[Ne III\] $\lambda$ 3869 line for the determination of the $Ne^{+2}$ abundance; the \[S II\] $\lambda$ 6716, 6731 doublet and, when detected, the \[S III\] $\lambda$ 6312 line, for the determination of the $S^+$ and $S^{+2}$ abundances, respectively; from the \[Ar III\] $\lambda$ 7135 we derived the $Ar^{+2}$ abundance. The fluxes used for the abundance computations are those listed in Tables 5 and 6, i.e. corrected for the effect of underlying Balmer absorption and dust extinction.
For all regions but s5, the ion abundances obtained with the [*getIonAbundance*]{} option in PyNeb are given in Tables 7 and 8. Due to the wavelength coverage, the abundance of the $O^+$ and $Ne^{+2}$ ions was derived only from the blue spectra, while for all the other ions we derived two different values respectively from the blue and from the red spectra (although in both cases the temperature was that derived from the blue spectra). The ion abundances derived with PyNeb and with [*abund*]{} are generally consistent within the errors, nevertheless we notice systematic trends for some elements: the $O^{+2}$, $Ne^{+2}$, $N^{+}$, and $S^{+}$ abundances derived with PyNeb are systematically higher than those obtained with [*abund*]{} by $\sim$0.03, 0.04, 0.03, and 0.04 dex, respectively; on the other hand, the abundances of the $O^{+}$ and $S^{+2}$ ions are systematically lower by $\sim$0.02 and 0.01 dex, respectively.
The total element abundances can be derived from the abundances of ions seen in the optical spectra using ionization correction factors (ICFs). In the case of oxygen, the most abundant ions $O^+$ and $O^{+2}$ are seen in our blue spectra, allowing for an immediate derivation of the total abundance as $O/H= O^+/H + O^{+2}/H$. For regions s71, s81, s82, s83, and s84, where we measured the \[O I\] $\lambda$ 6300 line, we added the contribution from the $O^0$ ion to the computation of the total oxygen abundance. This contribution however is very small, amounting to $\lesssim$0.02 dex. Concerning the $O^{+3}$ ion, [@izotov06] estimated a contribution larger than $\sim$1% to the total oxygen abundance only in the highest-excitation H II regions with $O^+ /(O^+ + O^{+2}) \lesssim 0.1$. From our Tables 7 and 8, we notice that this condition does never occur in our regions, implying that the contribution from the $O^{+3}$ ions can be neglected. The oxygen abundances derived for regions s6-1 to s10 in NGC 1705 are in the range $7.75 \lesssim 12 + \log(O/H) \lesssim 8.05$ (see Tables 7 and 8), with an average value and standard deviation of $7.91 \pm 0.08$ dex. The tabulated values were computed from the blue spectra, due to the lack of the \[O II\] $\lambda$3727 line in the red spectra.
To compute the abundances of the other elements, we adopted the following ICFs from [@izotov06]:
$$ICF(N^+)= -0.809 v + 0.712 + 0.852/v,$$
$$ICF(Ne^{+2}) = -0.405 w + 1.382 + 0.021/w,$$
$$ICF(S^+ + S^{+2})=0.155 v + 0.849 + 0.062/ v,$$
$$ICF(Ar^{+2})=0.285 v + 0.833 + 0.051/v,$$
where
$$v = O^+/(O^+ + O^{+2}), w = O^{+2}/(O^+ + O^{+2}).$$
The ICFs were computed from the $O^+$ and $O^{+2}$ abundances derived from the blue spectra, but were used to correct the ions both for the blue and the red spectra. The total sulfur abundance could be computed only for those regions where we measured the \[S III\] $\lambda$6314 line, providing the $S^{+2}$ abundance. Our results for the total abundances of N, O, Ne, S, and Ar in regions s6-1 to s10 are given in Tables 7 and 8. The abundances of N, S, and Ar obtained from the blue and red spectra are consistent within the errors. Hereafter, we will use the abundances based on the blue spectra.
Line ratio diagnostic diagrams \[dd\]
=====================================
The [@bpt] and [@vo87] diagnostic diagrams are a powerful tool to classify ionized regions according to the main excitation mechanism. In the most commonly used $\log([O III]\lambda5007/H\beta)$ versus $\log([N II]\lambda\lambda6548,84/H\alpha)$, $\log([S II]\lambda\lambda6716,31/H\alpha)$, or $\log([O I]\lambda6300/H\alpha)$ diagrams, starbursts fall onto the left-hand region, while Seyfert galaxies are located in the upper right, and LINERs lie in the lower right zone. [@kewley01], hereafter Ke01, produced starburst grids on the optical diagnostic diagrams combing stellar population synthesis models with the MAPPINGS III photoionization and shock code [@dopita00 and references therein] for a wide range of metallicities ($Z = 0.05 -3$ times solar) and ionization parameters ($q=5\times10^6 - 3\times10^8 \ cm \ s^{-1}$). Using these grids, Ke01 defined a theoretical upper limit for starburst models on the optical diagnostic diagrams, called the “maximum starburst line” (MSL): objects lying above and to the right of the MSL can not be explained with pure starburst models, but require an additional contribution from a harder ionizing source such as an active galactic nucleus or shocks. On the other hand, objects located below and to the left of the MSL may have a non negligible contribution (up to $\sim30\%$, Ke01) from excitation mechanisms different than stellar photoionization.
In order to investigate if excitation mechanisms other than pure stellar photoionization are present in NGC 1705, we compared the line ratios observed in regions s5 to s10 against models. In Figure \[bpt\] we plotted the Ke01 models based on the PEGASE v2.0 population synthesis code [@pegase], utilizing the Padova group tracks [@bressan93], for a continuous star formation, density $n=10\ cm^{-3}$, and metallicities of 0.2 and 0.5 times solar. The Ke01 models suggest a higher metallicity for NGC 1705 than that derived by the direct method, but large differences are observed in the starburst grids depending on the adopted population synthesis code (STARBURST 99 or PEGASE v2.0), stellar evolution tracks, continuous or instantaneous burst, and on other input parameters. In the diagrams of panels a), b), and c), the models run from top left to bottom right with decreasing ionization parameter q, defined as the ratio between the ionizing photon flux through a unit area and the local number density of hydrogen atoms. We found that the Ke01 models based on the STARBURST99 [@sb99] code utilizing the Geneva group tracks [@geneva] provide a worst agreement with the NGC 1705 data. The thick solid line in Figure \[bpt\] defines the maximum starburst sequence by Ke01.
In panels a), b), and c) of Figure \[bpt\], we also plotted the fast shock$+$precursor models of [@allen08] for the SMC and LMC chemical compositions obtained with the MAPPINGS III code. In these models, the ionizing radiation generated by the cooling of hot gas behind the shock front generates a strong radiation field of extreme ultraviolet and soft X-ray photons, which leads to significant photoionizing effects. At shock velocities above a certain limit ($v\sim170$ km $s^{-1}$), the ionization front velocity exceeds that of the shock, and expands to form a precursor H II region ahead of the shock. For a given metallicity and pre-shock gas density, the emission line ratios are mainly determined by the shock velocity and by the magnetic field, which acts to limit the compression through the shock. For the shock models in Figure \[bpt\], the $[O III]/H\beta$ ratio increases with shock velocity (with plotted velocities in the range $125-800$ km $s^{-1}$ ), while the $[N II]/H\alpha$, $[S II]/H\alpha$, and $[O I]/H\alpha$ ratios increase with magnetic field (with plotted magnetic fields in the range $0.5 - 10 \ \mu G$). The models assume a density of $n=1 \ cm^{-3}$ (the only value available for non-solar abundance models). The diagrams in panels a), b), and c) clearly show that shocks are better separated from stellar photoionization in the $[S II]/H\alpha$ and $[O I]/H\alpha$ ratios, while photo-ionized and shock excited components are strongly degenerate in the $[O III]/H\beta$ vs $[N II]/H\alpha$ diagram [e.g. @mat72; @dopita97].
From the comparison of the NGC 1705 data with models in Figure \[bpt\] we can draw the following conclusions. First, the emission line ratios measured in regions s5 to s10 are below the MSL in all the a), b), and c) diagnostic diagrams, implying that the emission lines can in principle be explained with pure stellar photoionization. Furthermore, a major contribution from shock excitation can be excluded on the basis of diagrams b) and c), where the data show few or no overlap with the shock models. Indeed, the data completely avoid the shock grids in the $[O III]/H\beta$ versus $[O I]/H\alpha$ diagrams, where stellar photoionization and shock excitation are best separated (but notice that the most external regions s5, s6-1, s6-2, s9, and s10 are missing in this diagram). Of course, we can not exclude a minor contribution from shocks projected on relatively stronger stellar photoionized regions. Furthermore, it is possible that our data are biased toward stellar photoionization, and that shock-dominated zones are actually present in NGC 1705; in fact, non-radiative ionization is typically confined to regions of low H$\alpha$ surface brightness, as observed by many authors who studied the occurrence of shocks in star forming galaxies [@ferguson96; @martin97; @calzetti04], while our slits were positioned in regions of relatively high \[O III\] (and H$\alpha$) surface brightness in order to obtain spectra with a good signal-to-noise. Indeed, the presence of a kiloparsec-scale supershell of ionized gas centered on the nucleus of NGC 1705 and expanding at a velocity of $(53\pm10)$ km $s^{-1}$ [@meurer92; @sahu97; @heckman01] suggests that shock excitation is a possible mechanism.
The second consideration from an inspection of the diagrams in Figure \[bpt\] is that the data tend to move from top left to bottom right in diagrams a), b) and c) (decreasing $[O III]/H\beta$ values and increasing $[N II]/H\alpha$, $[S II]/H\alpha$, and $[O I]/H\alpha$ ratios) with increasing distance from the galaxy center. We can exclude that this effect is due to dust extinction, given the very weak effect of dust absorption on the BPT diagrams and the low $E(B-V)$ values derived for NGC 1705 (see Tables 7 and 8); we can also exclude a density effect, as that shown in [@califa] for a sample of $\sim$ 5000 H II regions in the CALIFA survey, given the homogeneously low $n_e$ derived for regions s6 to s10. The behaviour that we see in Figure \[bpt\] has already been observed for other galaxies, and it is commonly explained with the gradual and continuous transition from the core of the H II regions into the diffuse ionized gas (DIG) [e.g., @ferguson96; @martin97]. In fact, the trend of the data from top left to bottom right follows that of the starburst models of Ke01 with decreasing ionization parameter q. [@califa] have shown that there is a strong correlation between position in the BPT diagram, metallicity and ionization parameter: more metal-rich H II regions have the lowest ionization parameters, and are located in the bottom right part of the star-forming sequence, while low-metallicity H II regions display the highest ionization parameters and are located top left. However, in the case of NGC 1705, the observed trend from top left to bottom right in the BPT diagram with increasing galactocentric distance goes in the opposite direction, being associated to a decrease in metallicity as derived by the direct $T_e$ method. In this context, the observed gradual decrease in the ionization parameter could be explained with the dilution of radiation from a centralized source with increasing distance from the main star forming regions.
[@dopita00] showed that the $[O III]\lambda5007/[O II]\lambda3727$ versus $[N II]\lambda\lambda6548,84/[O II]\lambda3727$ diagram provides an optimal separation between the ionization parameter and the chemical abundance. This diagram is shown in panel d) of Figure \[bpt\], where the NGC 1705 data have been plotted against the Ke01 starburst models for metallicities of 0.2 and 0.5 solar. In this metallicity range, the $[O III]\lambda5007/[O II]\lambda3727$ ratio is almost independent of metallicity and decreases with decreasing ionization parameter q. For the plotted models, the data span ionization parameters from $q=2\times10^7 \ cm \ s^{-1}$ to $q=8\times10^7 \ cm \ s^{-1}$, with the most internal regions (s8-1, s8-2) having the largest q values, and the most external ones (s6-1, s9) corresponding to the lowest q values. To better show this radial trend, we have plotted in Figure \[ion\] the $[O III]\lambda5007/[O II]\lambda3727$ ratio against the projected distance of regions s5 to s10 from the galaxy center. This plot suggests a well defined decreasing trend of the ionization parameter with increasing distance from the galaxy center, with regions s6-2 and s10 as outliers. This trend indicates that the main ionizing source is located within the central $\sim$6.6 arcsec ($\sim$160 pc), which is the distance of region s8-1 from the galaxy center, and that the diluted radiation from this centralized source is able to ionize the gas out to at least $\sim$32.8 arcsec ($\sim$800 pc, the distance of region s6-1). Indeed, Figure \[rgb\] shows that the high surface brightness optical component of NGC 1705, where the young starburst is concentrated, is well within our innermost region s8-1. Young blue stars are found in correspondence of two of our targeted H II regions, i.e. regions s7-1 and s8-1. Through a comparison with the map of stars younger than $\sim$5 Myr derived from the CMDs in [@anni09], Figure 14, we ascertained that these stars are indeed young enough to be able to ionize the gas. Nevertheless, the trend of the ionization parameter with distance from the galaxy center suggests that the main ionizing source is the most central starburst, or perhaps the super star cluster (SSC), in NGC 1705, and that local association of young stars likely contribute to a minor extent. Regions s6-2 and s10, which deviate from the global decreasing trend (notice however the particularly large errors in their measured line ratios), could be affected by different excitation mechanisms than pure stellar photoionization in a larger amount.
Results and Discussion
======================
Comparison with literature abundances
-------------------------------------
We compared our results with the work of [@lee04] who derived chemical abundances for H II regions in NGC 1705. Lee et al. obtained long-slit spectra with the EFOSC2 instrument on the 3.6 m telescope at ESO La Silla Observatory for 16 H II regions, and detected the \[O III\]$\lambda$4363 line in five of them. There are a few H II regions in common with our sample, namely region s8-1 (their A1) and region s7-1 (their region C1); our regions s8-2 and s8-3 partially overlap with their regions A2 and A3, respectively. Among the regions in common, the \[O III\]$\lambda$4363 is detected only in region A3, while for regions A1 and C1 Lee et al. provide an upper limit for the flux. For region A3, our oxygen abundance is $12 + \log(O/H)=7.97\pm0.02$, while they obtain $8.21^{+0.11}_{-0.16}$, which is consistent with our results within 3$\sigma$ errors. For regions s8-1 and s7-1 our oxygen abundances are $12 + \log(O/H)=7.95\pm0.04$ and $7.93\pm0.04$, respectively, while for A1 and C1 they provide lower limits of $12 + \log(O/H)>7.57$ and $>8.16$, respectively. Thus, while our result for region s8-1 is consistent with their abundance, our value for region s7-1 turns out significantly lower than their lower limit. For the five regions with detected \[O III\]$\lambda$4363 line, the average oxygen abundance by [@lee04] is $12 + \log(O/H)=8.22\pm0.05$, while our average oxygen abundance is $12 + \log(O/H)=7.91\pm0.08$ (see Table 9). However, we notice that all the five H II regions in Lee et al. are relatively close to the galaxy center, while some of our slits are located at much higher distances; due to the evidence for a negative oxygen metallicity gradient in NGC 1705 (see Section 7.3), we re-computed the average abundance using only the most central regions s7-1, s7-2, s8-1, s8-2, s8-3, and s8-4, and obtained $\log(O/H)=7.96\pm0.04$, which is still $\sim0.26$ dex lower than the Lee et al.’s abundance. We identified a few reasons for the discrepancy between our abundances on those by Lee et al. : 1) First of all, our temperatures are systematically higher than those derived by them (their average $T_e(O^{+2})$ is $\sim11700 \ K$, while our average $T_e(O^{+2})$ is $\sim13000 \ K$ for the regions in slits 7 and 8); our higher temperatures imply $\sim0.15$ dex lower abundances. Although both temperature estimates are based on the direct detection of the \[OIII\]$\lambda$4363 line, we were able to detect this feature with a better signal-to-noise than Lee et al., and thus we tend to consider our temperature estimates more reliable; 2) In the second instance, we notice that the \[O II\]$\lambda3727$/H$\beta$ reddening-corrected fluxes published by Lee et al. are systematically higher than ours (almost twice as high for regions C1, A1, and A2 in common with us), while the \[O III\] fluxes are more in agreement with our values. Their higher \[O II\] fluxes imply $\sim$ 0.1 dex higher (total) oxygen abundances. Possible reasons for the discrepancy between the \[O II\] fluxes are the drop in the instrumental sensitivity below $\sim$4000 Å and the higher atmospheric extinction toward bluer wavelengths: systematic uncertainties in the determination of the sensitivity curve or in the atmospheric extinction function in this wavelength range could lead to significant differences in the estimated emission fluxes; 3) Finally, we notice that our $t(O^{+})$ temperatures are based on the [@izotov06] relation, while Lee et al. adopted [@campbell86]; the Izotov et al. formula provides $\sim$ 500 k higher $O^{+}$ temperatures, implying $\sim 0.01- 0.03$ dex lower total oxygen abundances.
We investigated how our results for NGC 1705 compare with the luminosity-metallicity relationship for other samples of dwarf galaxies [@ks96; @vanzee06; @zhao10; @haurberg13]. Our results are illustrated in Figure \[metallicity\], where we plotted the absolute B magnitude ($M_B=-15.45$ from $B=13.09$, corrected for Galactic extinction [@paz03], and ($m-m_0)=28.54$ [@tosi01]) versus our average oxygen abundance ($12 + \log(O/H)=7.91\pm0.08$), together with abundance estimates for samples of dIrrs and BCDs in the literature. The solid line in Figure \[metallicity\] is the 3 $\sigma$-rejection linear least squares fit to literature data, excluding NGC 1705 ($12 + \log(O/H)= (-0.092 \pm 0.010) \times M_B + (6.50 \pm 0.17)$, $rms=0.17$). This is flatter than the relations obtained by [@vanzee06], and by [@haurberg13] (equation 2). Our average abundance estimate for NGC 1705 is in agreement with all the three metallicity-luminosity relations.
[@lee04] derived also abundances for nitrogen, neon, and argon. From their Table 8, we computed average abundances and standard deviations of $12 + \log(N/H)=6.51\pm0.06$ ($\log(N/O)=-1.69\pm0.10$), $12 + \log(Ne/H)=7.81\pm0.07$ ($\log(Ne/O)=-0.41\pm0.05$), and $12 + \log(Ar/H)=5.93\pm0.05$ ($\log(Ar/O)=-2.30\pm0.07$) (see Table 9). Averaging our regions s6-1 to s10, we derive instead mean abundances and standard deviations of $12 + \log(N/H)=6.63\pm0.07$ ($\log(N/O)=-1.27\pm0.11$), $12 + \log(Ne/H)=7.19\pm0.12$ ($\log(Ne/O)=-0.71\pm0.06$), and $12 + \log(Ar/H)=5.62\pm0.06$ ($\log(Ar/O)=-2.31\pm0.03$). We also derive a sulfur average abundance of $12 + \log(S/H)=6.34\pm0.04$ ($\log(S/O)=-1.61\pm0.04$). If we use only regions s7-1, s7-2, s8-1, s8-2, s8-3, and s8-4 to compute the average abundances, we obtain $12 + \log(N/H)=6.62\pm0.05$ ($\log(N/O)=-1.34\pm0.06$), $12 + \log(Ne/H)=7.26\pm0.09$ ($\log(Ne/O)=-0.70\pm0.07$), and $12 + \log(Ar/H)=5.65\pm0.05$ ($\log(Ar/O)=-2.32\pm0.03$). Thus, while our average central $N/H$ abundance is marginally consistent within the errors with that of Lee et al., (the effect of their lower temperatures is partly compensated by their slightly lower \[N II\]$\lambda$6584 fluxes), our $\log(N/O)$ ratio is $\sim$0.35 dex higher, as a consequence of our lower oxygen abundance. There is significant discrepancy also for Ne and Ar , since the Lee et al. abundances are $\sim0.55$ and $\sim$0.28 dex higher than ours. For Ne, the particularly large difference is due to the combined effect of their lower temperatures and their $\sim$ 50% higher \[Ne III\]$\lambda$3869 fluxes. The authors themselves noticed the anomalously high Ne abundances derived in their study. Figures \[elements\] and \[nsuo\], where we compared our abundances with literature values for different samples of late-type dwarf galaxies [@ks96; @it99; @vanzee06; @guseva11; @haurberg13; @esteban14; @nicholls14], show that our abundances for NGC 1705 are in very good agreement with literature trends.
Relative abundances of N and $\alpha$-process elements
------------------------------------------------------
Since the $\alpha$-elements (such as O, Ne, S, and Ar) are all synthesized by massive stars on similar timescales, the relative abundance of these elements is expected to remain constant, independent of the metallicity of the H II regions. This has been observationally demonstrated by many authors [e.g., @vanzee97; @it99; @vanzee06], and it is also shown in Figure \[elements\], where we plotted the Ne/O, S/O, and Ar/O ratios as a function of total oxygen abundance for different samples of late-type dwarf galaxies. Figure \[elements\] shows that this trend is also valid within NGC 1705 for the different H II regions: within the $\sim$0.3 dex oxygen abundance range spanned by our data, the $\alpha$-element ratios remain roughly constant with total oxygen abundance.
The case of nitrogen is more complex since it may have a “primary” and a “secondary” origin. Nitrogen which is produced only out of the original hydrogen in a star is referred to as “primary” nitrogen. The major contributors of primary nitrogen are believed to be intermediate mass stars in the range of 4 to 8 $M_{\odot}$ [@rv81]. It has also been suggested that primary nitrogen is produced in low metallicity massive stars [e.g., @ww95; @meynet02]. “Secondary” nitrogen refers instead to nitrogen which has been produced from C and O originally incorporated into the star when it formed. In presence of a substantial primary nitrogen component, the N/O ratio is expected to be independent of the oxygen abundance. On the other hand, one expects secondary nitrogen to become more and more important as oxygen abundance increases. Indeed, this is what is observed: for very low-metallicity galaxies the N/O ratio is roughly constant at $\log(N/O)\sim-1.6$, suggesting a substantial primary nitrogen component. Then, as oxygen increases, the N/O ratio starts to increase with O, suggesting a major secondary nitrogen component [@vanzee98]. We see in Figure \[nsuo\] that our H II regions are in the regime where nitrogen has both a primary and a secondary component. The N/O ratios exhibit a significant scatter, with the more external regions in NGC 1705 (see also Figure \[radial\]) populating the upper envelope of the distribution of late-type dwarf galaxies.
Abundance radial trends
-----------------------
We illustrate in Figure \[radial\] the behaviour of the total oxygen and nitrogen abundances, and of the N/O, Ne/O, S/O, and Ar/O ratios, with distance from the galaxy center. While no gradients in the relative abundances of the $\alpha$-elements are detected, the plot suggests a decreasing trend of the total oxygen abundance with increasing distance from the center of NGC 1705. A linear least squares fit to the data provides the following relation: $12 + log(O/H) = (-0.24\pm0.06) \times R \ kpc^{-1} + (8.01\pm0.03)$, which implies a radial gradient of $d \log(O/H)/d r = -0.24 \pm 0.08$ dex $kpc^{-1}$. This is steeper than what usually found in spiral galaxies and in the Milky Way [see @hw99 for a review]. The presence of metallicity gradients in late-type dwarf galaxies has been widely investigated and discussed in the literature. The majority of studies indicate that, within the observational uncertainties, dIrr and BCD galaxies show nearly spatially constant chemical abundances [e.g., @kobul97; @croxall09; @lagos13; @haurberg13]. Two possible explanations for the absence of significant metallicity gradients in these systems are that a) the ejecta from stellar winds and supernovae are dispersed and mixed across the ISM on timescales of $<10^7$ yr , or that b) freshly synthesized elements remain unmixed with the surrounding interstellar medium and reside in a hot $10^6$ K phase or a cold, dusty, molecular phase [@kobul97]. Detections of negative metallicity gradients from stars and H II regions have been reported in the literature for the dIrr NGC 6822 [@venn04; @lee06], with an oxygen abundance gradient of $-0.16 \pm 0.05$ dex $kpc^{-1}$. More recent studies based on spectroscopy of individual RGB stars have detected slightly negative gradients in \[Fe/H\] for the SMC, the LMC and the dIrr WLM [@leaman14 and references therein]. A very recent study by [@pilyugin15] finds a correlation between H II region abundance gradients and surface brightness profiles in irregular galaxies: galaxies with a flat inner profile show shallow (if any) radial abundance gradients, while irregular galaxies with a steep inner profile show rather steep radial abundance gradients, in agreement with our result.
Finally, in Figure \[radial\] there is some evidence for larger N/O ratios in the external regions of NGC 1705 ($R>$ 0.45 kpc), in correspondence of which no young stars (age $<$ 15 Myr) are found [@anni09 and Figure \[bursts\] in this paper]. The N/O behaviour is a consequence of the decreasing radial oxygen trend, since the total nitrogen abundance does not show any definite trend, and it is consistent with a large scatter around a flat distribution. The innermost regions share common values (at a 2$\sigma$ level) of $\langle$log(N/O)$\rangle_{\mathrm{int}}$ = $-$1.34$\pm$0.06 and $\langle$12$+$log(O/H)$\rangle_{\mathrm{int}}$ = 7.96$\pm$0.04, while the external regions appear characterized by lower oxygen abundances and higher N/O ratios, $\langle$log(N/O)$\rangle_{\mathrm{ext}}$ = $-$1.17$\pm$0.07 and $\langle$12$+$log(O/H)$\rangle_{\mathrm{ext}}$ = 7.82$\pm$0.04. A higher oxygen abundance and lower N/O toward the galaxy center may be the natural consequence of a burst in the inner regions sufficiently recent to have polluted the medium in oxygen but not yet in nitrogen [see, e.g. @pilyugin92]. This scenario is consistent with the location of the most central ($R<$ 0.45 kpc) s7-1, s7-2, and s8-1 to s8-4 H II regions, which are at the periphery of a region that has been forming stars during the last $\sim$15 Myr (see Figure \[bursts\]), and that has consequently been enriched in oxygen. We will test this hypothesis quantitatively when the LEGUS HST Survey [@legus] will provide more precise information on the recent star formation activity of different regions in NGC 1705. This information will be used as an input to multi-zone chemical evolution models, thus improving upon the one-zone model approach adopted in [@romano06].
We also want to point out that the history of nitrogen enrichment in our Galaxy is still poorly constrained by N measurements in stars, especially in the halo [see e.g., @romano10 their figure 3, and references and discussion therein]. Therefore, the determination of N abundances in the gaseous component of low-metallicity external systems is of the highest importance to improve our understanding of N stellar nucleosynthesis.
Conclusions
===========
We obtained \[O III\] narrow-band imaging and multi-slit optical spectroscopy of NGC 1705 with FORS2@VLT to infer chemical abundances for planetary nebulae (PNe) and H II regions and, more in general, to characterize the properties of the ionized gas.
From the \[O III\] pre-imaging, the ionized gas presents a bipolar morphology extending out to $\sim$2.5 kpc from the galaxy center, and it is organized into thin filaments, loops, and arcs, in agreement with previous studies in the literature based on H$\alpha$ narrow-band imaging [e.g. @meurer92]. It also exhibits “cellular” structures or cavities, with a typical cell size of $\sim$50 pc.
Spectra in the wavelength range $\sim3600-7200$ Å were obtained in the MXU configuration for some selected regions. We positioned the slits on planetary nebulae (PN) candidates, extended \[O III\] emission regions, and gas filaments. Nine out of the seventeen PN candidates identified from the \[O III\] pre-imaging were targeted for spectroscopy. However, almost the totality of them ultimately turned out to be background objects, with the exception of one, which is a possible low-ionization PN. We were then left with five MXU slits positioned on the ionized gas in NGC 1705 out to $\sim$1 kpc from the galaxy center. From those, we extracted spectra for ten apertures (regions), plus the low-ionization PN. The auroral \[O III\]$\lambda4363$ line was detected in all regions but the PN, allowing for a direct estimate of the electron temperatures. We derived the abundances of nitrogen, oxygen, neon, sulfur and argon out to $\sim$1 kpc from the center of NGC 1705. The lack of detection of the \[O III\]$\lambda4363$ line in the PN, coupled with the low signal-to-noise of its spectra, prevented us from deriving any reliable characterization of the chemical properties for this object.
We detected for the first time in NGC 1705 a negative radial gradient for the oxygen abundance of $-0.24 \pm 0.08$ dex $kpc^{-1}$. On the other hand, nitrogen exhibits a large dispersion around a roughly flat spatial distribution. As a consequence, the external regions ($R>$ 0.45 kpc) of NGC 1705 present larger N/O ratios than the central ones. These trends may be the natural consequence of oxygen enrichment in the central regions of NGC 1705 from the starburst activity occurred over the last $\sim$15 Myr, as derived from the CMDs of the resolved stars.
The average abundances and standard deviations derived for O, N, Ne, S, and Ar are: $12 + \log(O/H)=7.91 \pm 0.08$, $12 + \log(N/H)=6.63\pm0.07$ ($\log(N/O)=-1.27\pm0.11$), $12 + \log(Ne/H)=7.19\pm0.12$ ($\log(Ne/O)=-0.71\pm0.06$), $12 + \log(S/H)=6.34\pm0.04$ ($\log(S/O)=-1.61\pm0.04$), and $12 + \log(Ar/H)=5.62\pm0.06$ ($\log(Ar/O)=-2.31\pm0.03$), in agreement with values reported in the literature for other samples of dwarf irregular and blue compact dwarf galaxies. On the other hand, our average oxygen abundance is $\sim$ 0.31 dex lower than previous literature estimates for NGC 1705 based on the \[O III\]$\lambda4363$ line, while the $\log(N/O)$ ratio is $\sim$0.42 dex higher. Part of the discrepancy with other estimates of NGC 1705 oxygen abundance is due to the presence of the negative radial gradient. However, if we recompute the average O abundance using only the most central regions (comparable in distance from the center to regions used in the literature), we obtain $12 + \log(O/H)=7.96 \pm 0.04$, which is still $\sim$0.26 dex lower than the average abundance reported in the literature.
Using classical emission-line diagnostic diagrams, we investigated the possible contribution from components different than pure stellar photoionization in NGC 1705. A major contribution from shock excitation can be excluded in our analyzed regions. Furthermore, the radial behavior of the emission line ratios is consistent with the progressive dilution of radiation with increasing distance from the center of NGC 1705. This suggests that the strongest starburst located within the central $\sim$150 pc is the major ionizing source out to distances of at least $\sim$1 kpc, and possibly even more. The gradual dilution of the radiation (decrease of the ionization parameter) with increasing distance from the galaxy center reflects the gradual and continuous transition from the highly ionized H II regions in the proximity of the major starburst into the diffuse ionized gas.
In order to explain the observed element abundances in NGC 1705, we plan to run in the future new multi-zone chemical evolution models based both on the SFH already derived from WFPC2 data, and on the more precise information on the recent star formation activity that will come from the LEGUS HST Survey. In particular, we will quantitatively test the hypothesis that the presence of a negative radial gradient in the oxygen abundance, coupled with a flat nitrogen distribution, can be explained by the active star formation derived from the CMDs of the resolved stars in the most central regions of NGC 1705 over the last $\sim$15 Myr.
F. A. and this work have been supported by PRIN MIUR through grant 2010LY5N2T\_006. D. R. and M. T. have also been supported by the same grant. We acknowledge support from the ESO telescope operator I. Condor. Based on observations made with ESO Telescopes at the La Silla Paranal Observatory under programme ID 084.D-0248 and 086.D-0761.
Aloisi, A., Heckman, T. M., Hoopes, C. G., et al. 2005, Starbursts: From 30 Doradus to Lyman Break Galaxies, Dordrecht, p.2
Allen, M. G., Groves, B. A., Dopita, M. A., Sutherland, R. S., & Kewley, L. J. 2008, , 178, 20
Annibali, F., Greggio, L., Tosi, M., Aloisi, A., & Leitherer, C. 2003, , 126, 2752
Annibali, F., Tosi, M., Monelli, M., et al. 2009, , 138, 169
Appenzeller, I., Fricke, K., F[ü]{}rtig, W., et al. 1998, The Messenger, 94, 1
Baldwin, J. A., Phillips, M. M., & Terlevich, R. 1981, , 93, 5
Berg, D. A., Skillman, E. D., Marble, A. R., et al. 2012, , 754, 98
Bressan, A., Fagotto, F., Bertelli, G., & Chiosi, C. 1993, , 100, 647
Calzetti, D., Harris, J., Gallagher, J. S., III, et al. 2004, , 127, 1405
Calzetti, D., Lee, J. C., Sabbi, E., et al. 2015, , 149, 51
Campbell, A., Terlevich, R., & Melnick, J. 1986, , 223, 811
Carigi, L., Colin, P., Peimbert, M., & Sarmiento, A. 1995, , 445, 98
Chavez, M., Bertone, E., Morales-Hernandez, J., & Bressan, A. 2009, , 700, 694
Clementini, G., Held, E. V., Baldacci, L., & Rizzi, L. 2003, , 588, L85
Croxall, K. V., van Zee, L., Lee, H., et al. 2009, , 705, 723
Dekel, A., & Silk, J. 1986, , 303, 39
Depew, K., Parker, Q. A., Miszalski, B., et al. 2011, , 414, 2812
De Robertis, M. M., Dufour, R. J., & Hunt, R. W. 1987, , 81, 195
Dolphin, A. E. 2000, , 531, 804
Dopita, M. A. 1997, , 485, L41
Dopita, M. A., Kewley, L. J., Heisler, C. A., & Sutherland, R. S. 2000, , 542, 224
Esteban, C., Garc[í]{}a-Rojas, J., Carigi, L., et al. 2014, , 443, 624
Ferguson, A. M. N., Wyse, R. F. G., & Gallagher, J. S. 1996, , 112, 2567
Fioc, M., & Rocca-Volmerange, B. 1997, , 326, 950
Fruchter, A. S., & Hook, R. N. 2002, , 114, 144
Gavil[á]{}n, M., Ascasibar, Y., Moll[á]{}, M., & D[í]{}az, [Á]{}. I. 2013, , 434, 2491
Gil de Paz, A., Madore, B. F., & Pevunova, O. 2003, , 147, 29
Gonz[á]{}lez Delgado, R. M., Leitherer, C., & Heckman, T. M. 1999, , 125, 489
G[ó]{}rny, S. K. 2014, , 570, A26
Guseva, N. G., Izotov, Y. I., Stasi[ń]{}ska, G., et al. 2011, , 529, AA149
Haurberg, N. C., Rosenberg, J., & Salzer, J. J. 2013, , 765, 66
Heckman, T. M., Sembach, K. R., Meurer, G. R., et al. 2001, , 554, 1021
Henry, R. B. C., & Worthey, G. 1999, , 111, 919
Izotov, Y. I., Thuan, T. X., & Lipovetsky, V. A. 1994, , 435, 647
Izotov, Y. I., & Thuan, T. X. 1999, , 511, 639
Izotov, Y. I., Stasi[ń]{}ska, G., Meynet, G., Guseva, N. G., & Thuan, T. X. 2006, , 448, 955
Kauffmann, G., White, S. D. M., & Guiderdoni, B. 1993, , 264, 201
Kennicutt, R. C., Jr., & Garnett, D. R. 1996, , 456, 504
Kewley, L. J., Dopita, M. A., Sutherland, R. S., Heisler, C. A., & Trevena, J. 2001, , 556, 121
Kobulnicky, H. A., & Skillman, E. D. 1996, , 471, 211
Kobulnicky, H. A., & Skillman, E. D. 1997, , 489, 636
Kobulnicky, H. A., Kennicutt, R. C., Jr., & Pizagno, J. L. 1999, , 514, 544
Lagos, P., & Papaderos, P. 2013, Advances in Astronomy, 2013, 20
Leaman, R., Venn, K., Brooks, A., et al. 2014, , 85, 504
Lee, H., Skillman, E. D., & Venn, K. A. 2006, , 642, 813
Lee, H., McCall, M. L., & Richer, M. G. 2003, , 125, 2975
Lee, H., McCall, M. L., Kingsburgh, R. L., Ross, R., & Stevenson, C. C. 2003, , 125, 146
Lee, H., & Skillman, E. D. 2004, , 614, 698
Leitherer, C., Ekstr[ö]{}m, S., Meynet, G., et al. 2014, , 212, 14
Luridiana, V., Morisset, C., & Shaw, R. A. 2012, IAU Symposium, 283, 422
Magrini, L., & Gon[ç]{}alves, D. R. 2009, , 398, 280
Marconi, G., Matteucci, F., & Tosi, M. 1994, , 270, 35
Marlowe, A. T., Heckman, T. M., Wyse, R. F. G., & Schommer, R. 1995, , 438, 563
Martin, C. L. 1999, , 513, 156
Matteucci, F., & Tosi, M. 1985, , 217, 391
McConnachie, A. W., Arimoto, N., Irwin, M., & Tolstoy, E. 2006, , 373, 715
McGaugh, S. S. 1991, , 380, 140
Marigo, P., Girardi, L., Bressan, A., et al. 2008, , 482, 883
Martin, C. L. 1997, , 491, 561
Martins, L. P., Gonz[á]{}lez Delgado, R. M., Leitherer, C., Cervi[ñ]{}o, M., & Hauschildt, P. 2005, , 358, 49
Martins, F., Schaerer, D., & Hillier, D. J. 2005, , 436, 1049
Mathewson, D. S., & Clarke, J. N. 1972, , 178, L105
McCall, M. L., Rybski, P. M., & Shields, G. A. 1985, , 57, 1
Melnick, J., Moles, M., & Terlevich, R. 1985, , 149, L24
Mendez, R. H., Kudritzki, R. P., & Herrero, A. 1992, , 260, 329
Meurer, G. R., Freeman, K. C., Dopita, M. A., & Cacciari, C. 1992, , 103, 60
Meynet, G., & Maeder, A. 2002, , 381, L25
Nicholls, D. C., Dopita, M. A., Sutherland, R. S., et al. 2014, , 786, 155
Osterbrock, D. E. 1989, Research supported by the University of California, John Simon Guggenheim Memorial Foundation, University of Minnesota, et al. Mill Valley, CA, University Science Books, 1989, 422 p.,
Panuzzo, P., Bressan, A., Granato, G. L., Silva, L., & Danese, L. 2003, , 409, 99
Papaderos, P., Fricke, K. J., Thuan, T. X., & Loose, H.-H. 1994, , 291, L13
Patat, F., Moehler, S., O’Brien, K., et al. 2011, , 527, A91
Pe[ñ]{}a, M., Richer, M. G., & Stasi[ń]{}ska, G. 2007, , 466, 75
P[é]{}rez-Montero, E., & D[í]{}az, A. I. 2003, , 346, 105
Pilyugin, L. S. 1992, , 260, 58
Pilyugin, L. S. 1993, , 277, 42
Pilyugin, L. S., & Thuan, T. X. 2005, , 631, 231
Pilyugin, L. S., & Thuan, T. X. 2007, , 669, 299
Pilyugin, L. S., Grebel, E. K., & Zinchenko, I. A. 2015, arXiv:1505.00337
Renzini, A., & Voli, M. 1981, , 94, 175
Romano, D., Tosi, M., & Matteucci, F. 2006, , 365, 759
Romano, D., Karakas, A. I., Tosi, M., & Matteucci, F. 2010, , 522, A32
S[á]{}nchez-Bl[á]{}zquez, P., Peletier, R. F., Jim[é]{}nez-Vicente, J., et al. 2006, , 371, 703
S[á]{}nchez, S. F., P[é]{}rez, E., Rosales-Ortega, F. F., et al. 2015, , 574, A47
Sahu, M. S., & Blades, J. C. 1997, , 484, L125
Shaw, R. A., & Dufour, R. J. 1994, Astronomical Data Analysis Software and Systems III, 61, 327
Schaller, G., Schaerer, D., Meynet, G., & Maeder, A. 1992, , 96, 269
Schlafly, E. F., & Finkbeiner, D. P. 2011, , 737, 103
Schlegel, D. J., Finkbeiner, D. P., & Davis, M. 1998, , 500, 525
Tolstoy, E., Hill, V., & Tosi, M. 2009, , 47, 371
Tosi, M., Sabbi, E., Bellazzini, M., et al. 2001, , 122, 1271
Vacca, W. D., Garmany, C. D., & Shull, J. M. 1996, , 460, 914
van Zee, L., Haynes, M. P., & Salzer, J. J. 1997, , 114, 2497
van Zee, L., Salzer, J. J., & Haynes, M. P. 1998, , 497, L1
van Zee, L., & Haynes, M. P. 2006, , 636, 214
Veilleux, S., & Osterbrock, D. E. 1987, , 63, 295
Venn, K. A., Tolstoy, E., Kaufer, A., & Kudritzki, R. P. 2004, Origin and Evolution of the Elements, 58
Woosley, S. E., & Weaver, T. A. 1995, , 101, 181
Zhao, Y., Gao, Y., & Gu, Q. 2010, , 710, 663
[lcccccc]{} $\lambda$3727 & 471.0 $\pm$ 62.0 & 334.4 $\pm$ 5.9 & 240.3 $\pm$ 4.5 & 210.7 $\pm$ 2.0 & 173.1 $\pm$ 0.9 & 142.3 $\pm$ 0.7\
[\[Ne III\]]{} $\lambda$3869 & $-$ & 35.8 $\pm$ 1.0 & 42.1 $\pm$ 1.3 & 22.7 $\pm$ 1.1 & 37.9 $\pm$ 0.6 & 45.9 $\pm$ 0.5\
H8$+$He I $\lambda$3889 & $-$ & 20.2 $\pm$ 0.8 & 14.4 $\pm$ 0.7 & 17.2 $\pm$ 0.8 & 23.5 $\pm$ 0.3 & 12.2 $\pm$ 0.2\
H$\epsilon$ $+$ He I $+$\[Ne III\] $\lambda$3970 & $-$ & 20.7 $\pm$ 1.0 & 15.6 $\pm$ 1.6 & 20.6 $\pm$ 0.1 & 30.8 $\pm$ 3.0 & 16.5 $\pm$ 0.8\
H$\delta$ $\lambda$4101 & $-$ & 19.4 $\pm$ 0.4 & 15.8 $\pm$ 0.2 & 23.0 $\pm$ 0.5 & 27.0 $\pm$ 0.1 & 10.9 $\pm$ 0.3\
H$\gamma$ $\lambda$4340 & $-$ & 42.0 $\pm$ 0.7 & 40.5 $\pm$ 0.9 & 42.4 $\pm$ 0.4 & 44.2 $\pm$ 1.8 & 36.0 $\pm$ 0.2\
[\[O III\]]{} $\lambda$4363 & $-$ & 3.3 $\pm$ 0.3 & 3.4 $\pm$ 0.5 & 4.2 $\pm$ 0.3 & 4.6 $\pm$ 0.5 & 3.7 $\pm$ 0.2\
He I $\lambda$4471 & $-$ & 5.1 $\pm$ 0.8 & $-$ & 3.2 $\pm$ 0.2 & $-$ & 3.3 $\pm$ 0.1\
He II $\lambda$4686 & 28.0 $\pm$ 11.1 & $-$ & $-$ & 0.8 $\pm$0.1 & $-$ & 1.5 $\pm$ 0.2\
& 35.0 $\pm$ 14.0 &$-$ & $-$ & $-$ & $-$& $-$\
H$\beta$ $\lambda$4861 & 100.0 $\pm$ 10.4 & 100.0 $\pm$ 1.4 & 100.0 $\pm$ 1.4 & 100.0 $\pm$ 0.5 & 100.0 $\pm$ 0.4 & 100.0 $\pm$ 0.2\
& 100.0 $\pm$ 5.6 & 100.0 $\pm$ 0.8 & 100.0 $\pm$ 0.7 & 100.0 $\pm$ 0.3 & 100.0 $\pm$ 0.2 & 100.0 $\pm$ 0.2\
[\[O III\]]{} $\lambda$4959 & 115.0 $\pm$ 14.9 & 90.6 $\pm$ 1.5 & 142.9 $\pm$ 2.4 & 112.5 $\pm$ 1.8 & 144.0 $\pm$ 0.7 & 163.8 $\pm$ 3.3\
& 141.8 $\pm$ 10.8 & 89.8 $\pm$ 1.0 & 139.5 $\pm$ 1.0 & 111.1 $\pm$ 1.0 & 132.8 $\pm$ 0.4 & 157.9 $\pm$ 0.5\
[\[O III\]]{} $\lambda$5007 & 343.9 $\pm$ 37.6 & 270.0 $\pm$ 4.0 & 423.2 $\pm$ 6.1 & 319.4 $\pm$ 2.2 & 420.2 $\pm$ 1.8 & 457.8 $\pm$ 3.6\
& 492.2 $\pm$ 29.5 & 264.3 $\pm$ 2.4 & 412.9 $\pm$ 2.9 & 312.7 $\pm$ 1.5 & 411.9 $\pm$ 0.9 & 443.2 $\pm$ 1.2\
He I $\lambda$5876 & $-$ & 22.7 $\pm$ 0.9 & 30.1 $\pm$ 0.7 & 11.7 $\pm$ 0.2 & 12.5 $\pm$ 0.2 & 13.3 $\pm$ 0.5\
& 44.6 $\pm$ 2.6 & 29.1 $\pm$ 0.3 & 35.0 $\pm$ 0.4 & 10.9 $\pm$ 0.1 & 10.9 $\pm$ 0.1 & 12.8 $\pm$ 0.1\
[\[OI\]]{} $\lambda$6302 & $-$ & $-$ & $-$ & 2.4 $\pm$ 0.1 & $-$ & 1.9 $\pm$ 0.3\
& $-$ & $-$ & $-$ & 2.7 $\pm$ 0.1 & $-$ & $-$\
[\[S III\]]{} $\lambda$6314 & $-$ & $-$ & $-$ & 1.7 $\pm$ 0.1 & $-$ & 2.2 $\pm$ 0.1\
& $-$ & $-$ & $-$ & 1.8 $\pm$ 0.1 & $-$ & $-$\
[\[OI\]]{} $\lambda$6365 & $-$ & $-$ & $-$ & 1.0 $\pm$ 0.1 & $-$ & $-$\
& $-$ & $-$ & $-$ & 0.8 $\pm$ 0.1 & $-$ & $-$\
[\[NII\]]{} $\lambda$6548 & $-$ & 7.5 $\pm$ 0.6 & 7.2 $\pm$ 0.7 & 3.2 $\pm$ 0.5 & 3.1 $\pm$ 0.4 & 2.2 $\pm$ 0.5\
& $-$ & 3.7 $\pm$ 0.2 & $-$ & $-$ & $-$ & $-$\
H$\alpha$ $\lambda$6563 & 283.6 $\pm$ 30.1 & 298.1 $\pm$ 4.2 & 318.1 $\pm$ 4.4 & 310.3 $\pm$ 1.7 & 330.0 $\pm$ 1.3 & 320.7 $\pm$ 3.2\
& 355.3 $\pm$ 22.8 & 278.2 $\pm$ 2.4 & 300.4 $\pm$ 2.1 & 306.3 $\pm$ 0.5 & 328.6 $\pm$ 0.9 & 308.0 $\pm$ 0.8\
[\[N II\]]{} $\lambda$6584 & $-$ & 18.5 $\pm$ 0.7 & 15.6 $\pm$ 0.9 & 12.8 $\pm$ 0.7 & 12.3 $\pm$ 0.6 & 9.1 $\pm$ 0.7\
& 32.0 $\pm$ 8.2 & 20.0 $\pm$ 0.4 & 16.2 $\pm$ 0.3 & 11.6 $\pm$ 0.3 & 12.4 $\pm$ 0.6 & 9.1 $\pm$ 0.1\
He I $\lambda$6678 & $-$ & 3.6 $\pm$ 0.1 & 3.9 $\pm$ 0.1 & 3.4 $\pm$ 0.2 & 3.9 $\pm$ 0.2 & 3.4 $\pm$ 0.1\
& $-$ & 3.8 $\pm$ 0.1 & 4.1 $\pm$ 0.1 & 3.5 $\pm$ 0.1 & 3.5 $\pm$ 0.1 & 3.8 $\pm$ 0.1\
[\[S II\]]{} $\lambda$6716 & 42.0 $\pm$ 5.1 & 43.9 $\pm$ 0.9 & 33.3 $\pm$ 0.6 & 19.0 $\pm$ 0.2 & 23.0 $\pm$ 0.1 & 17.5 $\pm$ 0.1\
& $-$ & 41.7 $\pm$ 0.4 & 32.3 $\pm$ 0.3 & 19.5 $\pm$ 0.1 & 25.1 $\pm$ 0.1 & 17.9 $\pm$ 0.2\
[\[S II\]]{} $\lambda$6731 & 25.4 $\pm$ 4.3 & 28.9 $\pm$ 1.1 & 22.3 $\pm$ 0.6 & 13.1 $\pm$ 0.1 & 15.4 $\pm$ 0.1 & 12.5 $\pm$ 0.1\
& $-$ & 28.1 $\pm$ 0.3 & 21.4 $\pm$ 0.2 & 13.6 $\pm$ 0.1 & 17.2 $\pm$ 0.1 & 12.7 $\pm$ 0.1\
He I $\lambda$7065 & $-$ & $-$ & $-$ & 2.4 $\pm$ 0.1 & $-$ & 2.4 $\pm$ 0.1\
& $-$ & $-$ & $-$ & 2.4 $\pm$ 0.1 & 2.0 $\pm$ 0.1 & 2.4 $\pm$ 0.2\
[\[Ar III\]]{} $\lambda$7136 & $-$ & $-$ & $-$ & 7.7 $\pm$ 0.1 & 9.4 $\pm$ 0.1 & 8.6 $\pm$ 0.2\
& $-$ & 7.3 $\pm$ 0.3 & 7.8 $\pm$ 0.3 & 8.3 $\pm$ 0.1 & 9.1 $\pm$ 0.2 & 8.6 $\pm$ 0.1\
[lccccc]{} $\lambda$3727 & 163.1 $\pm$ 1.0 & 222.7 $\pm$ 6.1 & 261.4 $\pm$ 4.1 & 305.1 $\pm$ 6.9 & 300.2 $\pm$ 21.5\
[\[Ne III\]]{} $\lambda$3869 & 45.9 $\pm$ 0.3 & 37.5 $\pm$ 0.6 & 36.1 $\pm$ 0.8 & 39.8 $\pm$ 0.2 & 36.7 $\pm$ 3.3\
H8$+$He I $\lambda$3889 & 13.8 $\pm$ 0.4 & 18.3 $\pm$ 0.2 & 18.5 $\pm$ 0.5 & 22.3 $\pm$ 0.1 & 14.7 $\pm$ 2.7\
H$\epsilon$ $+$ He I $+$\[Ne III\] $\lambda$3970 & 16.2 $\pm$ 0.6 & 23.2 $\pm$ 0.8 & 23.5 $\pm$ 0.6 & 26.7 $\pm$ 0.5 & 23.1 $\pm$ 3.2\
H$\delta$ $\lambda$4101 & 13.9 $\pm$ 0.1 & 19.4 $\pm$ 0.4 & 20.2 $\pm$ 0.1 & 23.6 $\pm$ 0.4 & 22.1 $\pm$ 0.7\
H$\gamma$ $\lambda$4340 & 39.0 $\pm$ 0.2 & 42.5 $\pm$ 0.1 & 43.5 $\pm$ 0.1 & 46.7 $\pm$ 0.3 & 44.1 $\pm$ 2.0\
[\[O III\]]{} $\lambda$4363 & 4.3 $\pm$ 0.1 & 3.9 $\pm$ 0.1 & 3.6 $\pm$ 0.1 & 4.2 $\pm$ 0.2 & 6.5 $\pm$ 1.1\
He I $\lambda$4471 & 4.3 $\pm$ 0.2 & 3.1 $\pm$ 0.1 & 3.5 $\pm$ 0.1 & 4.6 $\pm$ 0.3 & $-$\
He II $\lambda$4686 & 1.2 $\pm$ 0.1 & 0.6 $\pm$ 0.1 & $-$ & $-$ & $-$\
H$\beta$ $\lambda$4861 & 100.0 $\pm$ 0.2 & 100.0 $\pm$ 0.1 & 100.0 $\pm$ 0.1 & 100.0 $\pm$ 0.4 & 100.0 $\pm$3.1\
& 100.0 $\pm$ 0.3 & 100.0 $\pm$ 0.2 & 100.0 $\pm$ 0.3 & 100.0 $\pm$ 0.6 & 100.0 $\pm$ 1.9\
[\[O III\]]{} $\lambda$4959 & 156.7 $\pm$ 1.3 & 130.7 $\pm$ 1.2 & 113.1 $\pm$ 1.1 & 101.3 $\pm$ 1.3 & 125.9 $\pm$ 4.8\
& 149.8 $\pm$ 0.6 & 125.9 $\pm$ 0.3 & 109.6 $\pm$ 0.5 & 96.5 $\pm$ 0.8 & 133.3 $\pm$ 3.2\
[\[O III\]]{} $\lambda$5007 & 462.6 $\pm$ 1.4 & 390.4 $\pm$ 2.8 & 338.1 $\pm$ 2.3 & 304.9 $\pm$ 4.3 & 383.1 $\pm$ 12.4\
& 449.5 $\pm$ 1.5 & 377.4 $\pm$ 0.8 & 329.0 $\pm$ 1.1 & 286.6 $\pm$ 1.8 & 400.9 $\pm$ 8.2\
He I $\lambda$5876 & 12.8 $\pm$ 0.2 & 11.9 $\pm$ 0.3 & 12.6 $\pm$ 0.4 & 19.3 $\pm$ 1.0 & $-$\
& 13.0 $\pm$ 0.2 & 11.9 $\pm$ 0.1 & 12.1 $\pm$ 0.1 & $-$ & $-$\
[\[OI\]]{} $\lambda$6302 & 2.1 $\pm$ 0.4 & 3.8 $\pm$ 0.2 & 5.2 $\pm$ 0.1 & $-$ & $-$\
& $-$ & $-$ & $-$ & $-$ & $-$\
[\[S III\]]{} $\lambda$6314 & 1.7 $\pm$ 0.2 & 1.4 $\pm$ 0.2 & 1.9 $\pm$ 0.1 & $-$ & $-$\
& $-$ & $-$ & $-$ & $-$ & $-$\
[\[OI\]]{} $\lambda$6365 & 1.2 $\pm$ 0.1 & 1.5 $\pm$ 0.1 & 2.5 $\pm$ 0.2 & 2.3 $\pm$ 0.2 & $-$\
& $-$ & $-$ & $-$ & $-$ & $-$\
[\[NII\]]{} $\lambda$6548 & 2.4 $\pm$ 0.3 & 2.4 $\pm$ 0.3 & 4.8 $\pm$ 0.4 & 7.0 $\pm$ 0.9 & $-$\
& $-$ & $-$& 4.4 $\pm$ 0.4 & 6.9 $\pm$ 0.1 & $-$\
H$\alpha$ $\lambda$6563 & 310.4 $\pm$ 0.8 & 304.7 $\pm$ 0.5 & 309.5 $\pm$ 0.6 & 300.3 $\pm$ 1.3 & 295.2 $\pm$ 9.2\
& 303.4 $\pm$ 1.0 & 294.2 $\pm$ 0.7 & 294.5 $\pm$ 1.1 & 290.2 $\pm$ 2.1 & 296.4 $\pm$ 6.5\
[\[N II\]]{} $\lambda$6584 & 10.0 $\pm$ 0.8 & 13.0 $\pm$ 0.7 & 19.5 $\pm$ 0.7 & 30.1 $\pm$ 1.0 & 16.8 $\pm$ 0.9\
& 10.8 $\pm$ 0.2 & 13.9 $\pm$ 0.2 & 20.2 $\pm$ 0.5 & 32.8 $\pm$ 0.9 & 15.0 $\pm$ 2.4\
He I $\lambda$6678 & 3.4 $\pm$ 0.1 & 3.3 $\pm$ 0.1 & 3.6 $\pm$ 0.1 & 3.6 $\pm$ 0.4 & $-$\
& 3.5 $\pm$ 0.1 & 3.2 $\pm$ 0.1 & 3.6 $\pm$ 0.1 & 3.8 $\pm$ 0.1 & $-$\
[\[S II\]]{} $\lambda$6716 & 19.6 $\pm$ 0.1 & 26.6 $\pm$ 0.1 & 35.3 $\pm$ 0.2 & 48.9 $\pm$ 0.5 & 33.5 $\pm$ 2.4\
& 20.3 $\pm$ 0.1 & 27.0 $\pm$ 0.1 & 34.9 $\pm$ 0.1 & 49.3 $\pm$ 0.3 & 25.4 $\pm$ 1.1\
[\[S II\]]{} $\lambda$6731 & 13.9 $\pm$ 0.1 & 17.8 $\pm$ 0.1 & 23.3 $\pm$ 0.2 & 33.9 $\pm$ 0.5 & 19.7 $\pm$ 1.7\
& 14.1 $\pm$ 0.1 & 18.0 $\pm$ 0.1 & 23.5 $\pm$ 0.1 & 34.3 $\pm$ 0.2 & 18.3 $\pm$ 1.1\
He I $\lambda$7065 & 2.2 $\pm$ 0.1 & 1.9 $\pm$ 0.1 & 2.1 $\pm$ 0.1 & 2.4 $\pm$ 0.1 & $-$\
& 2.6 $\pm$ 0.1 & 2.2 $\pm$ 0.1 & 2.3 $\pm$ 0.1 & $-$ & $-$\
[\[Ar III\]]{} $\lambda$7136 & 8.4 $\pm$ 0.1 & 7.5 $\pm$ 0.1 & 7.3 $\pm$ 0.2 & 7.4 $\pm$ 0.2 & 9.0 $\pm$ 2.2\
& 8.7 $\pm$ 0.3 & 7.7 $\pm$ 0.1 & 8.0 $\pm$ 0.2 & 6.7 $\pm$ 0.2 & 8.9 $\pm$ 1.6\
[lcccc]{} H$\delta$ & $-4.5 \pm 0.2$ & & $-0.5 \pm 0.1$ &\
H$\gamma$ & $-9.7 \pm 0.3$ & & $-5.2 \pm 0.1$ &\
[\[OIII\]]{}$\lambda$4363 & $-1.6 \pm 0.1$& & &\
H$\beta$ & $-22.5 \pm 0.3$ & $8.9 \pm 0.5$ & &\
[\[NII\]]{} $\lambda$6548 & $-1.3 \pm 0.1$ & & &\
H$\alpha$ & $-100.5 \pm 0.4$ & $5.9 \pm 0.2$ & &\
[\[NII\]]{} $\lambda$6584 & $-5.1 \pm 0.1$ & & &\
[lccc]{}\
H$\delta$ & 8.1 $\pm$ 0.1 & 0.92 $\pm$ 0.02 & 3.7 $\pm$ 0.2\
H$\gamma$ & 8.4 $\pm$ 0.3 & 0.96 $\pm$ 0.04 & 4.3 $\pm$ 0.3\
[\[O III\]]{} $\lambda$4363 & & & 0.9 $\pm$ 0.1\
H$\beta$ & 7.5 $\pm$ 0.4 & 0.86 $\pm$ 0.05 & 2.7 $\pm$ 0.4\
[\[NII\]]{} $\lambda$6548 & & & 0.5 $\pm$ 0.1\
H$\alpha$ & 6.0 $\pm$ 0.1 & 1.02 $\pm$ 0.04 & 1.5 $\pm$ 0.6\
[\[NII\]]{} $\lambda$6584 & & & 0.5 $\pm$ 0.1\
\
H$\delta$ & 8.8 $\pm$ 0.2 & 1.00 & 4.0 $\pm$ 0.2\
H$\gamma$ & 8.7 $\pm$ 0.2 & 1.00 & 4.5 $\pm$ 0.3\
[\[O III\]]{} $\lambda$4363 & & & 0.9 $\pm$ 0.1\
H$\beta$ & 8.7 $\pm$ 0.3 & 1.00 & 3.1 $\pm$ 0.4\
[\[NII\]]{} $\lambda$6548 & & & 0.5 $\pm$ 0.1\
H$\alpha$ & 5.9 $\pm$ 0.2 & 1.00 & 1.5 $\pm$ 0.6\
[\[NII\]]{} $\lambda$6584 & & & 0.5 $\pm$ 0.1\
\
H$\delta$ & 8.7 $\pm$ 0.1 & 0.99 $\pm$ 0.02 & 3.9 $\pm$ 0.2\
H$\gamma$ & 9.1 $\pm$ 0.4 & 1.05 $\pm$ 0.05 & 4.7 $\pm$ 0.4\
[\[O III\]]{} $\lambda$4363 & & & 0.9 $\pm$ 0.1\
H$\beta$ & 8.2 $\pm$ 0.4 & 0.94 $\pm$ 0.06 & 2.9 $\pm$ 0.4\
[\[NII\]]{} $\lambda$6548 & & & 0.5 $\pm$ 0.1\
H$\alpha$ & 5.9 $\pm$ 0.1 & 1.0 $\pm$ 0.04 & 1.5 $\pm$ 0.6\
[\[NII\]]{} $\lambda$6584 & & & 0.5 $\pm$ 0.1\
\
H$\delta$ & 5.9 $\pm$ 0.2 & 0.67 $\pm$ 0.03 & 2.7 $\pm$ 0.2\
H$\gamma$ & 5.3 $\pm$ 0.6 & 0.61 $\pm$ 0.07 & 2.7 $\pm$ 0.4\
[\[O III\]]{} $\lambda$4363 & & & 0.5 $\pm$ 0.1\
H$\beta$ & 7.0 $\pm$ 0.4 & 0.80 $\pm$ 0.05 & 2.5 $\pm$ 0.4\
[\[NII\]]{} $\lambda$6548 & & & 0.4 $\pm$ 0.1\
H$\alpha$ & 4.9 $\pm$ 0.1 & 0.83 $\pm$ 0.03 & 1.2 $\pm$ 0.5\
[\[NII\]]{} $\lambda$6584 & & & 0.4 $\pm$ 0.1\
[lcccccc]{} $\lambda$3727 & 477.7 $\pm$ 285.5 & 310.2 $\pm$ 6.1 & 216.7 $\pm$ 25.0 & 235.8 $\pm$ 7.6 & 207.2 $\pm$ 4.7 & 145.8 $\pm$ 10.9\
[\[Ne III\]]{} $\lambda$3869 & $-$ & 33.2 $\pm$ 1.0 & 37.9 $\pm$ 4.4 & 25.1 $\pm$ 1.5 & 44.6 $\pm$ 1.2 & 46.6 $\pm$ 3.5\
H8$+$He I $\lambda$3889 & $-$ & 18.8 $\pm$ 0.7 & 13.0 $\pm$ 2.1 & 19.0 $\pm$ 1.1 & 27.6 $\pm$ 0.7 & 12.4 $\pm$ 0.9\
H$\epsilon$ $+$ He I $+$\[Ne III\] $\lambda$3970 & $-$ & 19.3 $\pm$ 0.9 & 14.1 $\pm$ 2.1 & 22.6 $\pm$ 0.7 & 35.8 $\pm$ 3.6 & 16.6 $\pm$ 1.4\
H$\delta$ $\lambda$4101 & $-$ & 27.6 $\pm$ 0.7 & 29.6 $\pm$ 3.3 & 25.0 $\pm$ 0.9 & 30.7 $\pm$ 0.6 & 23.5 $\pm$ 1.8\
H$\gamma$ $\lambda$4340 & $-$ & 49.8 $\pm$ 1.3 & 52.4 $\pm$ 5.7 & 44.8 $\pm$ 1.3 & 48.3 $\pm$ 2.2 & 47.7 $\pm$ 3.4\
[\[O III\]]{} $\lambda$4363 & $-$ & 5.2 $\pm$ 0.4 & 6.3 $\pm$ 0.9 & 4.4 $\pm$ 0.3 & 5.0 $\pm$ 0.6 & 6.2 $\pm$ 0.6\
He I $\lambda$4471 & $-$ & 4.8 $\pm$ 0.7 & $-$ & 3.4 $\pm$ 0.2 & $-$ & 3.2 $\pm$ 0.2\
He II $\lambda$4686 & 28.1 $\pm$ 18.1 & $-$ & $-$ & 0.9 $\pm$ 0.1 & $-$ & 1.4 $\pm$ 0.2\
& 36.3 $\pm$ 14.1 & $-$ & $-$ & $-$ & $-$ & $-$\
H$\beta$ $\lambda$4861 & 100.0 $\pm$ 51.7 & 100.0 $\pm$ 2.3 & 100.0 $\pm$ 10.0 & 100.0 $\pm$ 2.7 & 100.0 $\pm$ 1.9 & 100.0 $\pm$ 6.4\
& 100.0 $\pm$ 31.0 & 100.0 $\pm$ 1.7 & 100.0 $\pm$ 2.3 & 100.0 $\pm$ 0.8 & 100.0 $\pm$ 1.3 & 100.0 $\pm$ 4.6\
[\[O III\]]{} $\lambda$4959 & 114.9 $\pm$ 58.1 & 84.1 $\pm$ 1.5 & 126.0 $\pm$ 12.2 & 111.5 $\pm$ 3.4 & 141.9 $\pm$ 2.7 & 152.8 $\pm$ 10.1\
& 138.9 $\pm$ 42.4 & 83.6 $\pm$ 1.2 & 123.3 $\pm$ 2.0 & 110.2 $\pm$ 0.9 & 130.9 $\pm$ 1.6 & 148.0 $\pm$ 6.6\
[\[O III\]]{} $\lambda$5007 & 343.4 $\pm$ 170.9 & 250.4 $\pm$ 4.3 & 372.7 $\pm$ 35.8 & 315.2 $\pm$ 8.4 & 411.3 $\pm$ 7.7 & 425.8 $\pm$ 26.8\
& 477.2 $\pm$ 142.9 & 246.1 $\pm$ 3.1 & 364.8 $\pm$ 5.9 & 309.1 $\pm$ 2.5 & 403.4 $\pm$ 5.0 & 414.6 $\pm$ 18.4\
He I $\lambda$5876 & $-$ & 21.0 $\pm$ 0.9 & 26.2 $\pm$ 2.4 & 10.9 $\pm$ 0.3 & 11.1 $\pm$ 0.2 & 11.8 $\pm$ 0.8\
& 37.9 $\pm$ 10.5 & 27.1 $\pm$ 0.4 & 31.0 $\pm$ 0.6 & 10.3 $\pm$ 0.1 & 9.8 $\pm$ 0.1 & 11.7 $\pm$ 0.5\
[\[OI\]]{} $\lambda$6302 & $-$ & $-$ & $-$ & 2.2 $\pm$ 0.1 & $-$ & 1.7 $\pm$ 0.3\
& $-$ & $-$ & $-$ & 2.4 $\pm$ 0.1 & $-$ & $-$\
[\[S III\]]{} $\lambda$6314 & $-$ & $-$ & $-$ & 1.5 $\pm$ 0.1 & $-$ & 1.9$\pm$ 0.1\
& $-$ & $-$ & $-$ & 1.7 $\pm$ 0.1 & $-$ & $-$\
[\[OI\]]{} $\lambda$6365 & $-$ & $-$ & $-$ & 0.9 $\pm$ 0.1 & $-$ & $-$\
& $-$ & $-$ & $-$ & 0.7 $\pm$ 0.1 & $-$ & $-$\
[\[NII\]]{} $\lambda$6548 & $-$ & 8.0 $\pm$ 0.6 & 7.7 $\pm$ 0.9 & 2.9 $\pm$ 0.4 & 2.6 $\pm$ 0.4 & 2.5 $\pm$ 0.5\
& $-$ & 4.4 $\pm$ 0.3 & $-$ & $-$ & $-$ & $-$\
H$\alpha$ $\lambda$6563 & 280.0 $\pm$ 123.0 & 279.6 $\pm$ 4.8 & 280.0 $\pm$ 23.8 & 280.0 $\pm$ 6.5 & 280.0 $\pm$ 4.6 & 280.0 $\pm$ 15.7\
& 280.0 $\pm$ 74.4 & 262.0 $\pm$ 3.4 & 270.0 $\pm$ 4.8 & 280.0 $\pm$ 2.0 & 280.0 $\pm$ 3.1 & 280.0 $\pm$ 11.0\
[\[N II\]]{} $\lambda$6584 & $-$ & 18.2 $\pm$ 0.7 & 15.0 $\pm$ 1.5 & 11.6 $\pm$ 0.6 & 10.4 $\pm$ 0.6 & 8.5 $\pm$ 0.8\
& 25.1 $\pm$ 9.1 & 19.6 $\pm$ 0.4 & 15.9 $\pm$ 0.5 & 10.6 $\pm$ 0.3 & 10.5 $\pm$ 0.5 & 8.8 $\pm$ 0.4\
He I $\lambda$6678 & $-$ & 3.3 $\pm$ 0.1 & 3.4 $\pm$ 0.3 & 3.1 $\pm$ 0.2 & 3.3 $\pm$ 0.1 & 3.0 $\pm$ 0.2\
& $-$ & 3.5 $\pm$ 0.1 & 3.6 $\pm$ 0.1 & 3.2 $\pm$ 0.1 & 2.9 $\pm$ 0.1 & 3.4 $\pm$ 0.2\
[\[S II\]]{} $\lambda$6716 & 41.5 $\pm$ 18.2 & 40.7 $\pm$ 0.9 & 28.8 $\pm$ 2.4 & 17.1 $\pm$ 0.4 & 19.2 $\pm$ 0.3 & 15.1 $\pm$ 0.8\
& $-$ & 38.8 $\pm$ 0.5 & 28.5 $\pm$ 0.5 & 17.7 $\pm$ 0.1 & 21.1 $\pm$ 0.2 & 16.1 $\pm$ 0.6\
[\[S II\]]{} $\lambda$6731 & 25.1 $\pm$ 11.4 & 26.8 $\pm$ 1.0 & 19.2 $\pm$ 1.6 & 11.7 $\pm$ 0.3 & 12.9 $\pm$ 0.2 & 10.8 $\pm$ 0.6\
& $-$ & 26.1 $\pm$ 0.4 & 18.9 $\pm$ 0.3 & 12.3 $\pm$ 0.1 & 14.5 $\pm$ 0.2 & 11.5 $\pm$ 0.5\
[\[S II\]]{} $\lambda$6716/$\lambda$6731 & 1.65 $\pm$ 0.60 & 1.52 $\pm$ 0.06 & 1.49 $\pm$ 0.10 & 1.46 $\pm$ 0.03 & 1.50 $\pm$ 0.02 & 1.40 $\pm$ 0.06\
& $-$ & 1.48 $\pm$ 0.02 & 1.51 $\pm$ 0.01 & 1.44 $\pm$ 0.01 & 1.46 $\pm$ 0.02 & 1.41 $\pm$ 0.05\
He I $\lambda$7065 & $-$ & $-$ & $-$ & 2.1 $\pm$ 0.1 & $-$ & 2.0 $\pm$ 0.2\
& $-$ & $-$ & $-$ & 2.2 $\pm$ 0.1 & 1.7 $\pm$ 0.1 & 2.2 $\pm$ 0.2\
[\[Ar III\]]{} $\lambda$7136 & $-$ & $-$ & $-$ & 6.8 $\pm$ 0.1 & 7.7 $\pm$ 0.2 & 7.3 $\pm$ 0.4\
& $-$ & 6.8 $\pm$ 0.3 & 6.9 $\pm$ 0.3 & 7.4 $\pm$ 0.1 & 7.5 $\pm$ 0.1 & 7.7 $\pm$ 0.3\
[lccccc]{} $\lambda$3727 & 157.6 $\pm$ 8.9 & 225.6 $\pm$ 9.4 & 272.4 $\pm$ 8.9 & 305.6 $\pm$ 13.9 & 295.5 $\pm$ 54.3\
[\[Ne III\]]{} $\lambda$3869 & 44.2 $\pm$ 2.4 & 37.8 $\pm$ 1.3 & 37.3 $\pm$ 1.3 & 39.7 $\pm$ 1.6 & 36.1 $\pm$ 6.8\
H8$+$He I $\lambda$3889 & 13.3 $\pm$ 0.8 & 18.4 $\pm$ 0.6 & 19.1 $\pm$ 0.7 & 22.2 $\pm$ 0.9 & 14.4 $\pm$ 3.6\
H$\epsilon$ $+$ He I $+$\[Ne III\] $\lambda$3970 & 15.5 $\pm$ 1.0 & 23.3 $\pm$ 1.1 & 24.1 $\pm$ 0.9 & 26.5 $\pm$ 1.1 & 22.6 $\pm$ 4.9\
H$\delta$ $\lambda$4101 & 25.8 $\pm$ 1.5 & 26.3 $\pm$ 0.9 & 26.4 $\pm$ 0.8 & 28.7 $\pm$ 1.2 & 27.5 $\pm$ 4.5\
H$\gamma$ $\lambda$4340 & 49.5 $\pm$ 2.7 & 48.9 $\pm$ 1.5 & 49.5 $\pm$ 1.4 & 52.0 $\pm$ 2.0 & 47.5 $\pm$ 7.7\
[\[O III\]]{} $\lambda$4363 & 6.5 $\pm$ 0.4 & 5.2 $\pm$ 0.2 & 4.8 $\pm$ 0.2 & 5.3 $\pm$ 0.3 & 7.1 $\pm$ 1.5\
He I $\lambda$4471 & 4.0 $\pm$ 0.3 & 3.1 $\pm$ 0.1 & 3.5 $\pm$ 0.1 & 4.4 $\pm$ 0.3 & $-$\
He II $\lambda$4686 & 1.2 $\pm$ 0.1 & 0.6 $\pm$ 0.1 & $-$ & $-$ & $-$\
H$\beta$ $\lambda$4861 & 100.0 $\pm$ 4.9 & 100.0 $\pm$ 2.7 & 100.0 $\pm$ 2.5 & 100.0 $\pm$ 3.4 & 100.0 $\pm$ 15.0\
& 100.0 $\pm$ 4.9 & 100.0 $\pm$ 3.0 & 100.0 $\pm$ 3.0 & 100.0 $\pm$ 4.2 & 100.0 $\pm$ 11.0\
[\[O III\]]{} $\lambda$4959 & 145.0 $\pm$ 7.0 & 125.0 $\pm$ 3.5 & 108.6 $\pm$ 2.8 & 97.2 $\pm$ 3.4 & 120.9 $\pm$ 17.8\
& 139.0 $\pm$ 6.6 & 120.5 $\pm$ 3.5 & 105.5 $\pm$ 3.1 & 93.1 $\pm$ 3.9 & 126.6 $\pm$ 13.5\
[\[O III\]]{} $\lambda$5007 & 427.5 $\pm$ 20.1 & 372.7 $\pm$ 10.1 & 323.6 $\pm$ 8.0 & 291.9 $\pm$ 10.5 & 367.6 $\pm$ 53.1\
& 416.8 $\pm$ 19.7 & 361.1 $\pm$ 10.3 & 316.4 $\pm$ 9.2 & 276.4 $\pm$ 11.4 & 380.6 $\pm$ 40.2\
He I $\lambda$5876 & 11.6 $\pm$ 0.5 & 11.0 $\pm$ 0.4 & 11.6 $\pm$ 0.4 & 18.1 $\pm$ 1.1 & $-$\
& 12.0 $\pm$ 0.5 & 11.3 $\pm$ 0.3 & 11.5 $\pm$ 0.3 & $-$ & $-$\
[\[OI\]]{} $\lambda$6302 & 1.9 $\pm$ 0.3 & 3.5 $\pm$ 0.2 & 4.7 $\pm$ 0.1 & $-$ & $-$\
& $-$ & $-$ & $-$ & $-$ & $-$\
[\[S III\]]{} $\lambda$6314 & 1.5 $\pm$ 0.2 & 1.3 $\pm$ 0.2 & 1.7 $\pm$ 0.1 & $-$ & $-$\
& $-$ & $-$ & $-$ & $-$ & $-$\
[\[OI\]]{} $\lambda$6365 & 1.1 $\pm$ 0.1 & 1.4 $\pm$ 0.1 & 2.2 $\pm$ 0.2 & 2.2 $\pm$ 0.2 & $-$\
& $-$ & $-$ & $-$ & $-$ & $-$\
[\[NII\]]{} $\lambda$6548 & 2.9 $\pm$ 0.3 & 2.6 $\pm$ 0.3 & 4.7 $\pm$ 0.4 & 7.0 $\pm$ 0.9 & $-$\
& $-$ & $-$ & $-$ & 4.6 $\pm$ 0.4 & 7.1 $\pm$ 0.3\
H$\alpha$ $\lambda$6563 & 280.0 $\pm$ 11.7 & 280.0 $\pm$ 6.5 & 280.0 $\pm$ 6.0 & 280.0 $\pm$ 8.3 & 280.0 $\pm$ 35.7\
& 280.0 $\pm$ 11.8 & 280.0 $\pm$ 7.1 & 280.0 $\pm$ 7.2 & 280.0 $\pm$ 35.7 & 280.0 $\pm$ 26.2\
[\[N II\]]{} $\lambda$6584 & 9.7 $\pm$ 0.9 & 12.3 $\pm$ 0.7 & 18.0 $\pm$ 0.7 & 28.4 $\pm$ 1.2 & 16.4 $\pm$ 2.2\
& 10.6 $\pm$ 0.5 & 13.6 $\pm$ 0.4 & 19.5 $\pm$ 0.7 & 31.9 $\pm$ 1.4 & 14.7 $\pm$ 2.6\
He I $\lambda$6678 & 3.0 $\pm$ 0.1 & 3.0 $\pm$ 0.3 & 3.2 $\pm$ 0.1 & 3.3 $\pm$ 0.4 & 6.7 $\pm$ 1.5\
& 3.2 $\pm$ 0.1 & 3.0 $\pm$ 0.1 & 3.4 $\pm$ 0.2 & 3.7 $\pm$ 0.1 & $-$\
[\[S II\]]{} $\lambda$6716 & 17.5 $\pm$ 0.7 & 24.2 $\pm$ 0.5 & 31.6 $\pm$ 0.7 & 45.3 $\pm$ 1.4 & 31.5 $\pm$ 4.5\
& 18.5 $\pm$ 0.8 & 25.5 $\pm$ 0.6 & 33.0 $\pm$ 0.8 & 47.3 $\pm$ 1.7 & 23.7 $\pm$ 2.4\
[\[S II\]]{} $\lambda$6731 & 12.4 $\pm$ 0.5 & 16.3 $\pm$ 0.4 & 20.9 $\pm$ 0.5 & 31.4 $\pm$ 1.0 & 18.6 $\pm$ 2.8\
& 12.9 $\pm$ 0.5 & 17.0 $\pm$ 0.4 & 22.2 $\pm$ 0.6 & 33.0 $\pm$ 1.2 & 17.1 $\pm$ 1.8\
[\[S II\]]{} $\lambda$6716/$\lambda$6731 & 1.42 $\pm$ 0.05 & 1.49 $\pm$ 0.03 & 1.51 $\pm$ 0.03 & 1.44 $\pm$ 0.04 & 1.70 $\pm$ 0.24\
& 1.44 $\pm$ 0.05 & 1.50 $\pm$ 0.03 & 1.48 $\pm$ 0.03 & 1.44 $\pm$ 0.04 & 1.38 $\pm$ 0.14\
He I $\lambda$7065 & 2.0 $\pm$ 0.1 & 1.8 $\pm$ 0.1 & 1.9 $\pm$ 0.1 & 2.2 $\pm$ 0.1 & $-$\
& 2.3 $\pm$ 0.2 & 2.0 $\pm$ 0.1 & 2.2 $\pm$ 0.1 & $-$ & $-$\
[\[Ar III\]]{} $\lambda$7136 & 7.4 $\pm$ 0.3 & 6.8 $\pm$ 0.1 & 6.5 $\pm$ 0.2 & 6.8 $\pm$ 0.3 & 8.5 $\pm$ 2.3\
& 8.0 $\pm$ 0.4 & 7.3 $\pm$ 0.2 & 7.5 $\pm$ 0.3 & 6.5 $\pm$ 0.3 & 8.1 $\pm$ 1.7\
[lccccccc]{} $E(B-V)$ & b & 0.01 $\pm$ 0.11 & 0.00 $\pm$ 0.02 & 0.02 $\pm$ 0.02 & 0.10 $\pm$ 0.01 & 0.16 $\pm$ 0.01 & 0.08 $\pm$ 0.01\
& r & 0.24 $\pm$ 0.06 & 0.00 $\pm$ 0.01 & 0.00 $\pm$ 0.02 & 0.09 $\pm$ 0.01 & 0.16 $\pm$ 0.01 & 0.04 $\pm$ 0.01\
$n_e$ \[$cm^{-3}$\] & b & $-$ & $-$ & $<32$ & $<10$ & $-$ & $27^{+49}_{-27}$\
& r & $-$ & $-$ & $-$ & $<6$ & $<1$ & $20^{+40}_{-20}$\
$T_e (O^{+2})$ \[K\] & b & $-$ & $15300^{+560}_{-490}$ & $14030^{+900}_{-700}$ & $12950^{+370}_{-320}$ & $12230^{+580}_{-470}$ & $13080^{+480}_{-420}$\
$(N^+/H^+)\times 10^6$ & b & $-$ & $1.85_{- 0.07}^{+0.07}$ & $1.78_{-0.13}^{+ 0.13}$ & $1.26_{-0.09}^{+ 0.09}$ & $1.28_{-0.11}^{+ 0.11}$ & $0.95_{-0.11}^{+ 0.11}$\
& r & $-$ & $1.69 _{-0.06}^{+ 0.06}$ & $1.67_{-0.12}^{+0.12}$ & $1.23_{-0.05}^{+ 0.06}$ & $1.39_{-0.12}^{+0.12}$ & $1.02_{-0.06}^{+ 0.06}$\
$12+\log(N/H)$ & b & $-$ & $6.54 \pm 0.02$ & $6.73 \pm 0.03$ & $6.53 \pm 0.03$ & $6.63 \pm 0.04$ & $6.63 \pm 0.05$\
& r & $-$ & $6.49 \pm 0.02$ & $6.66 \pm 0.03$ & $6.54 \pm 0.02$ & $6.62 \pm 0.04$ & $6.61 \pm 0.03$\
$(O^0/H^+)\times 10^6$ & b & $-$ & $-$ & $-$ & $1.89_{-0.12}^{+0.13}$ & $-$ & $1.43_{-0.11}^{+ 0.12}$\
& r & $-$ & $-$ & $-$ & $-$ & $-$ & $-$\
$(O^+/H^+)\times 10^5$ & b & $-$ & $2.95_{-0.17}^{+0.17}$ & $2.44_{-0.28}^{+0.30}$ & $3.21_{-0.22}^{+0.24}$ & $3.37_{-0.40}^{+0.50}$ & $1.94_{-0.17}^{+0.18}$\
$(O^{+2}/H^+)\times 10^5$ & b & $-$ & $2.62_{-0.21}^{+0.21}$ & $4.90_{-0.74}^{+0.72}$ & $5.21_{-0.38}^{+0.37}$ & $7.95_{-0.99}^{+1.05}$ & $6.87_{-0.65}^{+0.62}$\
& r & $-$ & $2.58_{-0.21}^{+0.21}$ & $4.79_{-0.73}^{+0.71}$ & $5.12_{-0.38}^{+0.36}$ & $7.68_{-0.96}^{+1.01}$ & $6.68_{-0.63}^{+0.61}$\
$12+\log(O/H)$ & b & $-$ & $7.75 \pm 0.03$ & $7.87 \pm 0.06$ & $7.93 \pm 0.03$ & $8.05 \pm 0.05$ & $7.95 \pm 0.04$\
$(Ne^{+2}/H^+)\times 10^5$ & b & $-$ & $0.83_{-0.08}^{+0.08}$ & $1.24_{-0.21}^{+0.22}$ & $1.05_{-0.09}^{+0.09}$ & $2.27_{-0.33}^{+0.36}$ & $1.90_{-0.21}^{+0.20}$\
$12+\log(Ne/H)$ & b & $-$ & $7.01 \pm 0.04$ & $7.16 \pm 0.08$ & $7.09 \pm 0.04$ & $7.41 \pm 0.07$ & $7.32 \pm 0.05$\
$(S^+/H^+)\times 10^7$ & b & $-$ & $7.93_{-0.27}^{+0.27}$ & $6.23_{-0.52}^{+0.52}$ & $4.16_{-0.17}^{+0.18}$ & $5.18_{-0.44}^{+0.44}$ & $3.71_{-0.19}^{+0.20}$\
& r & $-$ & $7.62_{-0.26}^{+0.26}$ & $6.16_{-0.44}^{+0.44}$ & $4.32_{-0.18}^{+0.19}$ & $5.74_{-0.49}^{+0.49}$ & $3.95_{-0.21}^{+0.21}$\
$(S^{+2}/H^+)\times 10^6$ & b & $-$ & $-$ & $-$ & $1.49_{-0.17}^{+0.20}$ & $-$ & $1.80_{-0.26}^{+0.31}$\
& r & $-$ & $-$ & $-$ & $1.69_{-0.20}^{+0.23}$ & $-$ & $-$\
$12+\log(S/H)$ & b & $-$ & $-$ & $-$ & $6.31 \pm 0.04$ & $-$ & $6.40 \pm 0.06$\
& r & $-$ & $-$ & $-$ & $6.36 \pm 0.05$ & $-$ & $-$\
$(Ar^{+2}/H^+)\times 10^7$ & b & $-$ & $-$& $-$& $3.90_{-0.26}^{+0.29}$ & $5.22_{-0.66}^{+0.66}$ & $4.07_{-0.32}^{+0.38}$\
& r & $-$ & $2.73_{-0.15}^{+0.17}$ & $3.29_{-0.39}^{+0.39}$ & $4.24_{-0.28}^{+0.31}$ & $5.08_{-0.69}^{+0.67}$ & $4.29_{-0.34}^{+0.40}$\
$12+\log(Ar/H)$ & b & $-$ & $-$ & $-$ & $5.62 \pm 0.03$ & $5.75 \pm 0.03$ & $5.66 \pm 0.04$\
& r & $-$ & $5.47 \pm 0.02$ & $5.55 \pm 0.05$ & $5.66 \pm 0.03$ & $5.74 \pm 0.06$ & $5.68 \pm 0.04$\
[lccccccc]{} $E(B-V)$ & b & 0.04 $\pm$ 0.01& 0.05 $\pm$ 0.01 & 0.07 $\pm$ 0.01 & 0.04 $\pm$ 0.01 & 0.02 $\pm$ 0.03\
& r & 0.01 $\pm$ 0.01 & 0.01 $\pm$ 0.01 & 0.02 $\pm$ 0.01 & 0.01 $\pm$ 0.01 & 0.01 $\pm$ 0.02\
$n_e$ \[$cm^{-3}$\] & b & $14^{+39}_{-14}$ & $-$ & $-$ & $<22$ & $-$\
& r & $<35$ & $-$ & $-$ & $<26$ & $40^{+140}_{-40}$\
$T_e (O^{+2})$ \[K\] & b & $13420^{+380}_{-340}$ & $12990^{+240}_{-210}$ & $13290^{+200}_{-190}$ & $14470^{+350}_{-320}$ & $14940^{+1720}_{-1160}$\
$(N^+/H^+)\times 10^6$ & b & $1.04_{- 0.10}^{+ 0.10}$ & $1.31_{- 0.09}^{+ 0.09}$ & $1.92_{-0.09}^{+0.09}$ & $2.65_{-0.16}^{+0.16}$ & $1.59_{-0.21}^{+0.21}$\
& r & $1.17_{-0.05}^{+ 0.05}$ & $1.60_{-0.05}^{+0.05}$ & $2.21_{-0.08}^{+0.08}$ & $2.73_{-0.13}^{+ 0.13}$ & $2.11_{-0.28}^{+0.28}$\
$12+\log(N/H)$ & b & $6.64 \pm 0.04$ & $6.60 \pm 0.03$ & $6.69 \pm 0.02$ & $6.75 \pm 0.03$ & $6.58 \pm 0.06$\
& r & $6.65 \pm 0.02$ & $6.65 \pm 0.01$ & $6.71 \pm 0.01$ & $6.73 \pm 0.02$ & $6.68 \pm 0.06$\
$(O^0/H^+)\times 10^6$ & b & $1.50_{-0.24}^{+0.24}$ & $2.98_{-0.12}^{+0.13}$ & $3.80_{-0.12}^{+0.13}$ & $-$ & $-$\
& r & $-$ & $-$ & $-$ & $-$ & $-$\
$(O^+/H^+)\times 10^5$ &b& $1.96_{-0.13}^{+0.13}$ & $3.06_{-0.14}^{+0.15}$ & $3.48_{-0.12}^{+0.13}$ & $3.22_{-0.14}^{+0.16}$ & $2.93_{-0.54}^{+0.54}$\
$(O^{+2}/H^+)\times 10^5$ & b & $6.34_{-0.46}^{+0.45}$ & $6.02_{-0.29}^{+0.29}$ & $4.91_{-0.19}^{+0.20}$ & $3.52_{-0.22}^{+ 0.22}$ & $4.05_{-0.90}^{+ 0.98}$\
& r & $ 6.16_{-0.45}^{+0.44}$ & $5.82_{-0.28}^{+ 0.28}$ & $4.80_{-0.19}^{+0.20}$ & $3.34_{-0.21}^{+0.21}$ & $4.21_{-0.94}^{+1.02}$\
$12+\log(O/H)$ & b & $7.93 \pm 0.03$ & $7.97 \pm 0.02$ & $7.94 \pm 0.02$ & $7.83 \pm 0.02$ & $7.84 \pm 0.08$\
$(Ne^{+2}/H^+)\times 10^5$ & b & $1.66_{-0.14}^{+0.14}$ & $1.57_{-0.09}^{+0.09}$ & $1.44_{-0.07}^{+0.07}$ & $1.18_{-0.08}^{+0.09}$ & $0.97_{-0.18}^{+0.28}$\
$12+\log(Ne/H)$ & b & $7.27 \pm 0.04$ & $7.26 \pm 0.02$ & $7.24 \pm 0.02$ & $7.16 \pm 0.03$ & $7.06 \pm 0.12$\
$(S^+/H^+)\times 10^7$ & b & $4.10_{-0.16}^{+0.15}$ & $5.85_{-0.15}^{+0.16}$ & $7.33_{-0.17}^{+ 0.17}$ & $9.55_{-0.30}^{+0.30}$ & $6.03_{-0.88}^{+0.88}$\
& r & $4.31_{-0.18}^{+0.18}$ & $6.14_{-0.16}^{+0.17}$ & $7.71_{-0.20}^{+0.20}$ & $10.00_{-0.27}^{+0.29}$ & $4.91_{-0.51}^{+0.51}$\
$(S^{+2}/H^+)\times 10^6$ & b & $1.27_{-0.17}^{+0.17}$ & $1.27_{-0.20}^{+0.20}$ & $1.50_{-0.09}^{+0.10}$ & $-$ & $-$\
& r & $-$ & $-$ & $-$ & $-$ & $-$\
$12+\log(S/H)$ & b & $6.29 \pm 0.05$ & $6.31 \pm 0.05$ & $6.38 \pm 0.02$ & $-$ & $-$\
& r & $-$ & $-$ & $-$ & $-$ & $-$\
$(Ar^{+2}/H^+)\times 10^7$ & b & $3.89_{-0.24}^{+0.23}$ & $3.86_{-0.17}^{+0.20}$ & $3.49_{-0.11}^{+0.12}$ & $3.04_{-0.14}^{+0.15}$ & $3.57_{-0.97}^{+0.97}$\
& r & $4.21_{-0.26}^{+0.25}$ & $4.15_{-0.18}^{+0.21}$ & $4.03_{-0.14}^{+0.16}$ & $2.91_{-0.13}^{+0.14}$ & $3.40_{-0.71}^{+0.71}$\
$12+\log(Ar/H)$ & b & $5.64 \pm 0.02$ & $5.62 \pm 0.02$ & $5.57 \pm 0.01$ & $5.51 \pm 0.02$ & $5.58 \pm 0.12$\
& r & $5.67 \pm 0.02$ & $5.65 \pm 0.02$ & $5.64 \pm 0.02$ & $5.50 \pm 0.02$ & $5.56 \pm 0.10$\
[lccc]{} $12 + \log(N/H)$ & $6.63 \pm 0.07$ & $6.62 \pm 0.05$ & $6.51 \pm 0.06$\
$12 + \log(O/H)$ & $7.91 \pm 0.08$ & $7.96 \pm 0.04$& $8.22 \pm 0.05$\
$12 + \log(Ne/H)$ & $7.19 \pm 0.12$ & $7.26 \pm 0.09$& $7.81 \pm 0.07$\
$12 + \log(Ar/H)$ & $5.62 \pm 0.06$ & $5.65 \pm 0.05$& $5.93 \pm 0.05$\
$\log(N/O)$ & $-1.27 \pm 0.11$ & $-1.34 \pm 0.06$ & $-1.69 \pm 0.10$\
$\log(Ne/O)$ & $-0.71 \pm 0.06$ & $-0.70 \pm 0.07$ & $-0.41 \pm 0.05$\
$\log(Ar/O)$ & $-2.31 \pm 0.03$ & $-2.32 \pm 0.03$ & $-2.30 \pm 0.07$\
[^1]: Notice however that the lines of N, O, and Ar sampled by FUSE have issues (e.g., saturation) that likely give an underestimate of the real column density of these elements in the neutral gas.
[^2]: IRAF is distributed by the National Optical Astronomy Observatory, which is operated by the Association of Universities for Research in Astronomy, Inc., under cooperative agreement with the National Science Foundation.
[^3]: Absorption strengths obtained with the Starburst 99 [@sb99] SSPs turned out to be consistent with those obtained using the Padova models.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study effects of weak disorder with Gaussian correlation function on a dipolar Bose gas with three-body interactions using the Hartree-Fock-Bogoliubov theory. Corrections due to quantum, thermal and disorder fluctuations to the condensate depletion, the one-body density correlation function as well as to the equation of state and the ground state energy are properly calculated. We show that the intriguing interplay of the disorder, dipole-dipole interactions and three-body interactions plays a fundamental role in the physics of the system. Interestingly, we find that the three-body interactions release atoms localized in the respective minima of the random potential.'
author:
- 'Redaouia Keltoum$^{1}$ and Abdelâali Boudjemâa$^{1,2}$'
title: 'Dipolar Bose gas with three-body interactions in weak disorder'
---
Introduction
============
Since the first observation of a Bose-Einstein condensate (BEC) with dipole-dipole interactions (DDI) in 2005, quantum dilute atomic gases have attracted major attention both theoretically and experimentally [@Baranov; @Pfau; @Carr; @Pupillo]. Dipolar BECs provide rich new physics, not encountered in systems with the contact interaction, thanks to the long-range and anisotropic character of the DDI.
The properties of dipolar BECs in weak random potential have recently sparked a great interest [@Krum; @Nik; @Ghab; @Boudj; @Boudj1; @Boudj2; @Boudj3]. It is widely agreed that the stability and the shape of such systems depend critically on the interplay between the DDI and the disorder potential. For instance, the superfluidity acquires a characteristic direction dependence due to the anisotropy of the DDI [@Krum; @Nik; @Ghab; @Boudj; @Boudj1]. The studies of two-dimensional (2D) disordered dipolar gases may offer the possibility for the observation of the superglass state [@Boudj2; @Boudj3]. Most recently, impacts of the Lee-Huang-Yang (LHY) quantum corrections on a dirty dipolar Bose gas have been analyzed by one of us [@Boudj5] using a perturbative theory. It is found that the LHY quantum fluctuations lead to reduce the disorder effects inside the condensate preventing the formation of the Bose glass state.
On the other hand, three-body interactions (TBI) play a key role in a wide variety of interesting physical phenomena, and provide a new physics not existed in systems with two-body interactions. Inelastic three-body processes, including observations of Efimov quantum states and atom loss from recombination have been reported in Refs [@Eff; @Eff1; @Bed; @Kra; @Brut]. Weakly interacting Bose and Fermi gases with competing attractive two-body and large repulsive TBI may form droplets [@Bulg]. Effects of TBI in ultracold bosonic atoms loaded in an optical lattice or a superlattice were also studied in [@Daly1; @Mazz; @Singh; @Mahm]. It was shown also that the TBI in Bose condensate may singnificantly modify the collective excitations [@Abdul; @Hamid; @Chen], the transition temperature, the condensate depletion and the stability of a BEC [@Peng; @Mash]. In the context of ultracold atoms with DDI, it has been revealed that the combined effects of TBI and DDI may lead to the formation of a stable supersolid state [@Petrov1] and a quantum droplet state [@Pfau1; @Kui; @Blakie; @Chom; @BoudjDp]. Very recently, we have shown that the TBI may shift the density profiles, the condensed fraction and the collective modes of a dipolar condensate at finite temperature [@Boudj4].
Disordered dipolar Bose gases with TBI present a different physical picture and may open prospects to achieve a stable superfluid. The goal of this work is then to study, for the first time to the best of our knowledge, effects of a weak disorder potential with Gaussian correlation function on the properties of BEC with two-body interactions and TBI. To this end, we use the Hatree-Fock-Bogoliubov (HFB) theory which includes an additional TBI term in the momentum space. Our results show that the TBI are relevant in reducing the influence of the disorder potential in BEC. Impacts of the disorder potential and the TBI on the fluctuations, coherence and the thermodynamics of the condensate are also highlighted. We compare our findings with previous theoretical results.
The rest of the paper is organized as follows. Section \[Mod\] introduces the HFB model for a disordered dipolar Bose gases with two-body interactions and TBI. In Sec.\[GCD\], we use a correlated Gaussian disorder potential to illustrate our model and derive useful expressions for the disorder fraction and the noncondensed density. In Sec.\[OBCF\] we look at how the interplay of the TBI and the disorder potential enhance the coherence of the system by numerically analyzing the behavior of the one-body density matrix. In Sec.\[Therm\] we calculate corrections due to the disorder effects and TBI to the chemical potential and the ground state energy. Our conclusions are drawn in Sec.\[Conc\].
Model {#Mod}
=====
We consider the effects of an external random potential $U(\mathbf r)$ on a dilute 3D dipolar Bose gas with contact two- and three-body interactions. Assuming that dipoles are oriented along $z$-axis. The Hamiltonian of the system reads: $$\begin{aligned}
\label{Ham}
\nonumber \hat{H}&=\int d \mathbf r \, \hat{\psi}^\dagger (\mathbf r) \left[-\frac{\hbar^2}{2m} \Delta +U (\mathbf r) \right]\hat{\psi} (\mathbf r)\\
&+\frac{1}{2}\int d\mathbf r\int d\mathbf r' \hat{\psi}^\dagger (\mathbf r) \hat{\psi}^\dagger(\mathbf r') V(\mathbf r- \mathbf r')\hat{\psi} (\mathbf r') \hat{\psi} (\mathbf r)\\ \nonumber
&+\frac{g_{3}}{6}\int d\mathbf r\, \hat{\psi}^\dagger (\mathbf r) \hat{\psi}^\dagger (\mathbf r) \hat{\psi}^\dagger (\mathbf r) \hat{\psi} (\mathbf r) \hat{\psi} (\mathbf r) \hat{\psi} (\mathbf r),\end{aligned}$$ where $\hat{\psi}^\dagger$ and $\hat{\psi}$ denote, respectiveiy the usual creation and annihilation field operators, $m$ is the particle mass. The two-body interactions is described by the potential $ V(\mathbf r- \mathbf r') =g_2\delta (\mathbf r- \mathbf r')+V_{dd} (\mathbf r -\mathbf r')$, where $ g_2=4\pi\hbar^2a/m $ with $a$ being the $s$-wave scattering length is assumed to be positive. The DDI term $ V_{dd}(\mathbf r)= {\cal M}_0{\cal M}^2 (1-3 \cos^2 \theta)/4\pi r^3 $, where ${\cal M} $ is the magnetic moment and $ \theta $ is the angle between the relative position of particles $\mathbf r$ and $z$-axis. The three-body coupling constant $g_3$ is in general a complex number with $ Im (g_3)$ describing the three-body recombination loss and $ Re(g_3)$ quantifying the three-body scattering parameter. Here, we will assume that the imaginary part of $g_3$ is negligible [@Bed; @Hamid; @Braa; @Zhan; @Blakie] which means that the loss rate is sufficiently small and hence, the system is stable. This well coincides with the experimental conditions reported in Ref.[@Pfau1]. Note that the strength of the three-body coupling $g_3$ is related to the atomic species and can be adjusted by Feshbach resonance [@Kui; @Evrt]. It is therefore, hard to predict the exact value of $g_3$ (see e.g. [@Bulg; @Braa; @Braa1]).
In what follows, we suppose that the disorder potential is described by vanishing ensemble averages $\langle U(\mathbf r)\rangle=0$ and a finite disorder correlation function $\langle U (\mathbf r) U(\mathbf r')\rangle=R (\mathbf r,\mathbf r')$.
In the frame of the HFB formalism, the Bose-field operator can be written as $$\label{FO}
\hat \psi ({\bf r},t)=\Phi ({\bf r},t)+\hat {\bar\psi} ({\bf r},t),$$ where $\Phi$ is the condensate wavefunction, and $\hat {\bar\psi}$ stands for the field of the noncondensed thermal atoms. Working in Fourier space, the condensate wavefunction is taken as $\Phi ({\bf r},t)= \sqrt{n_c}$ with $n_c$ being the condensate density, and the field operator of noncondensed atoms can be expanded in terms of plane waves $\hat {\bar\psi}= (1/\sqrt{V}) \sum_k a_k e^{ i \bf k. r}$. The DDI potential in momentum space is given by: $V _{dd}(\mathbf k)=(\mu_0\mu^2/12 \pi) (3\cos^2\theta_{\mathbf k}-1)$, where the vector $\mathbf k$ represents the momentum transfer imparted by the collision.
Now we deal with a weakly interacting system and assume that the disorder is sufficiently weak. Then it is possible to use the Bogoliubov-Huang-Meng approach [@HM] which suggests the transformation: $$\begin{aligned}
\hat{a}_{k} =u_k \hat{b}_k-v_k\hat{b}_{-k}^\dagger -\beta_k,\quad \hat{a}_{k}^\dagger =u_k \hat{b}_k^\dagger-v_k\hat{b}_{-k}-\beta_k^*,\end{aligned}$$ where $ \hat{b}_k^\dagger$ and $\hat{b}_k $ are operators of elementary excitations, the functions $ u_k,v_k $ are defined as $ u_k,v_k=(\sqrt{\varepsilon_k/E_k}\pm \sqrt{E_k/\varepsilon_k})/2$ with $ E_k=\hbar^2k^2/2m $ being the free particle energy, and $$\beta_k=\sqrt{\frac{n_c}{V}}\frac{E_k}{\varepsilon_k^2} U_k.$$ The Bogoliubov excitations energy is given by $$\varepsilon_k=\sqrt{E_k^2+2n_c E_k \bar V(\mathbf k)},$$ where $\bar V ({\bf k})= g_2 (1+g_3n_c/g_2) [1+ \gamma (3 \cos^2 \theta_{\mathbf k}-1)]$ with $ \gamma =\epsilon_{dd}/(1+g_3n_c/g_2)$, $\epsilon_{dd}={\cal M}_0 {\cal M}^2/(3 g_2) $ is the relative strength between the DDI and short-range interactions. For $k \rightarrow 0$, the excitations are sound waves $ \varepsilon_k=\hbar c_s(\theta_{\bf k})k $, where $ c_s(\theta_{\bf k})=c_0\sqrt{(1+g_3n_c/g_2) [1+ \gamma (3 \cos^2 \theta_{\bf k}-1)]}$ with $ c_0=\sqrt{g_2n_c/m} $ being the sound velocity without DDI and TBI.\
The diagonal form of the Hamiltonian (\[Ham\]) can be written as $\hat{H}=E+\displaystyle\sum_k\varepsilon_k\hat{b}_k^\dagger \hat{b}_k$. The total energy $E=E_0 (\theta)+\delta E+E_R $, where the zeroth order term $$\label{EGY0}
E_0(\theta)=\bar V (\theta) n_c N_c/2,$$ which should be computed in the limit $k \rightarrow 0$ since it accounts for the condensate (lowest state). The ground-state energy shift due to quantum fluctuations is $$\label{EGY1}
\delta E=\frac{1}{2}\displaystyle\sum_k [\varepsilon_k-E_k-n_c \bar V ({\bf k})],$$ and $$\label{EGY2}
E_R=-\displaystyle\sum_k n_c \langle\mid U_k\mid^2\rangle\frac{E_k}{\varepsilon_k^2}=-\displaystyle\sum_k n_c R_k\frac{E_k}{\varepsilon_k^2},$$ gives the correction to the ground-state energy due to the external random potential.
The noncondensed density is defined as $$\begin{aligned}
\label{NCD}
\tilde{n}=\displaystyle\sum_k\langle\hat{a}_k^\dagger \hat{a}_k\rangle=\frac{1}{V}\displaystyle\sum_k \left[(u_k^2+v_k^2)N_k+v_k^2+\langle |\beta_{\bf k}|^2 \rangle \right],\end{aligned}$$ where $ N_k=\langle\hat{b}_k^\dagger\hat{b}_k\rangle=1/[\exp({\varepsilon_k}/T)-1] $ is the Bose-Einstein distribution function, and the rest of the expectation values equal to zero ($\langle\hat{b}_k^\dagger \hat{b}_k^\dagger \rangle=\langle\hat{b}_k \hat{b}_k\rangle=0$).
Inserting the expressions of $u_k$ and $v_k$ in Eq.(\[NCD\]), and working in the thermodynamic limit where the sum over $k$ can be replaced by the integral $\sum_{\bf k}=V\int d \mathbf k/(2\pi)^3$, we get
\[NCDD\] $$\begin{aligned}
\tilde n&= \tilde n_0+\tilde n_{th} +n_R, \\
&=\frac{1}{2}\int \frac{d \mathbf k} {(2\pi)^3} \left[\frac{E_k+\bar V(\mathbf k) n_c} {\varepsilon_k} -1\right] \label{NCD1} \\
&+\frac{1}{2}\int \frac{d \mathbf k} {(2\pi)^3} \frac{E_k+\bar V(\mathbf k) n_c} {\varepsilon_k}\left[\coth\left(\frac{\varepsilon_k}{2T}\right)-1\right] \label{NCD2} \\
&+n_c\int \frac{d \mathbf k} {(2\pi)^3} R_k \frac{ E_k^2}{\varepsilon_k^4}. \label{NCD3}\end{aligned}$$
The leading term (\[NCD1\]) denotes the zero temperature contribution to the noncondensed density. The subleading term (\[NCD2\]) stands for thermal fluctuation corrections to the noncondensed density. Whereas the third term (\[NCD3\]) represents the condensate fluctuations due to the disorder potential known as [*glassy fraction*]{} and originates from the accumulation of density near the potential minima and density depletion around the maxima.
Gaussian-correlated disorder {#GCD}
=============================
As a concrete example, we consider in this section the case of a correlated Gaussian disorder model, which allow for unique control of the interplay between the disorder potential and interactions in both dipolar and nondipolar BEC [@Krum; @Boudj1]. It can be written as $$\label {Gdis}
R(k)=R_0\exp[-\sigma^2k^2/2],$$ where $ R_0 $ is the disorder strength which has dimension (energy)$^2 \times$(length)$^3 $ and $ \sigma$ characterizes the correlation length of the disorder.
The glassy fraction can be calculated easily via Eq.(\[NCD3\]) $$n_R=n_{\text{HM}} (1+g_3n_c/g_2)^{-1/2} h(\gamma,\sigma/\xi),$$ where $n_{\text{HM}}=[m^2 R_0/8\pi^{3/2}\hbar^4]\sqrt{n_c/a} $ is the usual Huang-Meng result [@HM]. The anisotropic disorder function is given as $$\begin{aligned}
\label{HF}
h (\gamma,\sigma/\xi)=\int^\pi_0 \mathrm{d}\theta\frac{\sin\theta S(\alpha)}{2\sqrt{1+\gamma(3\cos^2\theta-1)}},\end{aligned}$$ where the function $ S(\alpha)=e^{2\alpha} (4\alpha+1) \left[1-\text{erf}(\sqrt{2\alpha})\right]-2\sqrt{2\alpha/\pi}$, and $\alpha=\sigma^2[\epsilon_{dd}/\gamma (1+\gamma(3\cos^2\theta-1))]/\xi^2 $ with $ \xi=\hbar/\sqrt{mn_cg_2} $ being the healing length. In the absence of the DDI ($\epsilon_{dd}=0$), and in the limit $ \sigma/\xi \rightarrow 0 $ and $g_3=0 $, one has $ h(\gamma,\alpha)\rightarrow 1$, thus, one recovers the well-known Hang and Meng result ($n_R=n_{\text{HM}}$) [@HM].
The effects of both correlation length and effective interaction parameter $\gamma$ on the behavior of the disorder function are presented in Fig.\[DisF\]. We observe that the function $h(\gamma,\alpha)$ is decreasing with $g_3 n_c/g_2$ indicating that the TBI lead to reduce the disorder fluctuations (glassy fraction) inside the condensate even in the limit $ \sigma<\xi $. As is expected, the disorder fraction becomes significant for large DDI in contrast to the case of a disordered dipolar BEC with Lee-Huang-Yang (LHY) quantum corrections [@Boudj5]. The main difference between the TBI and the LHY corrections is that these latter are valid only in the regime of weak disorder since they are computed within the local density approximation which assumes that the external random potential should vary smoothly in space on a length scale comparable to the healing length or the characteristic correlation length of the disorder [@Boudj5]. Whereas, the TBI still remain applicable for both weak and strong disorder potentials. For $ \sigma >\xi $, the disorder effects is not important (see Fig.\[DisF\].b).
![ Disorder function $ h(\gamma, \sigma/\xi) $, as a function of $g_3 n_c/g_2$ for several values of $\epsilon_{dd}$ for $\sigma/\xi=0.4$ (a) and $\sigma/\xi=1.2$ (b).[]{data-label="DisF"}](TBI05.eps "fig:") ![ Disorder function $ h(\gamma, \sigma/\xi) $, as a function of $g_3 n_c/g_2$ for several values of $\epsilon_{dd}$ for $\sigma/\xi=0.4$ (a) and $\sigma/\xi=1.2$ (b).[]{data-label="DisF"}](TBI06.eps "fig:")
For delta-correlated disorder where $ \sigma/\xi \rightarrow 0 $, the function $ h(\gamma,0) ={\cal Q}_{-1}(\gamma)$ and $ n_R=n_{HM} {\cal Q}_{-1}(\gamma) $, where the contribution of the DDI is expressed by the functions ${\cal Q}_j (x)=\int_0^1 dy (1-x+3xy^2)^{j/2}$ [@Krum; @Ghab; @Boudj; @Boudj1; @Boudj5]. Note that the functions ${\cal Q}_j (x)$ tend to unity for $ \gamma=0 $ ($ {\cal Q}_j(0)=1$), and become imaginary for $ \gamma>0 $.
Now, we focus ourselves to calculate quantum and thermal depletion in a disordered BEC. Integrals (\[NCD1\]) and (\[NCD2\]) yield, respectively $$\begin{aligned}
\label{NDQ1}
\frac{\tilde{n}_0}{n_c}=\frac{8}{3}\sqrt{\frac{n_ca^3}{\pi}} (1+g_3n_c/g_2)^{3/2} {\cal Q}_{3}(\gamma),\end{aligned}$$ and $$\begin{aligned}
\label{NDT1}
\frac{\tilde{n}_{th}}{n_c}=\frac{2}{3} \left(\frac{\pi T}{n_c g_2} \right)^2\sqrt{\frac{n_ca^3}{\pi}} (1+g_3n_c/g_2)^{-1/2} {\cal Q}_{-1}(\gamma).\end{aligned}$$ For $\epsilon_{dd}=0$ and $g_3=0$, we recover the standard expressions for $\tilde{n}_0$ and $\tilde{n}_{th}$ . When $g_3=0$, Eqs.(\[NDQ1\]) and (\[NDT1\]) reduce to that obtained in our previous work for a dipolar BEC without TBI [@Boudj1].
One-body density matrix {#OBCF}
========================
![(color online) One-body density matrix due to the disorder corrections, $\mathrm{g}_R^{(1)}(r) $, for $\sigma/\xi=0.2$ (a)-(b) and $\sigma/\xi=1.2$ (c)-(d).[]{data-label="corr"}](correlation1.eps "fig:") ![(color online) One-body density matrix due to the disorder corrections, $\mathrm{g}_R^{(1)}(r) $, for $\sigma/\xi=0.2$ (a)-(b) and $\sigma/\xi=1.2$ (c)-(d).[]{data-label="corr"}](correlation2.eps "fig:") ![(color online) One-body density matrix due to the disorder corrections, $\mathrm{g}_R^{(1)}(r) $, for $\sigma/\xi=0.2$ (a)-(b) and $\sigma/\xi=1.2$ (c)-(d).[]{data-label="corr"}](correlation3.eps "fig:") ![(color online) One-body density matrix due to the disorder corrections, $\mathrm{g}_R^{(1)}(r) $, for $\sigma/\xi=0.2$ (a)-(b) and $\sigma/\xi=1.2$ (c)-(d).[]{data-label="corr"}](correlation4.eps "fig:")
The one-body density matrix (first-order correlation function) is defined as $g^{(1)}(\mathbf {r},\mathbf{r'},t,t')=\langle \hat\psi^\dagger (\mathbf r,t) \hat \psi(\mathbf r',t')\rangle$. In uniform case it depends only on the difference $ \vert{\bf r-r'} \vert=r$. Using the decomposition (\[FO\]), expressing the noncondensed field operator as $\hat {\bar\psi}=(1/V) \sum_k [u_k \hat b_k e^{i \bf k. r}- v_k\hat b_k^\dagger e^{-i \bf k. r}]$, and then taking into account that $|\Phi ({\bf r},t)|=\sqrt{n_c}$. We thus, get $$\begin{aligned}
\label{corrF}
g^{(1)} ({\bf r})&=n_c+g_R^{(1)}({\bf r}) +\int_0^{\infty} \frac{d \mathbf k}{(2\pi)^d}\left[v_{k}^2 + (u_{k}^2+v_{k}^2)N_{k}\right] e^{i {\bf k \cdot r}},\end{aligned}$$ The second term $g_R^{(1)}({\bf r})= \int (d \mathbf k/(2\pi)^3) \langle |\beta_{\bf k}|^2 \rangle\, \mathrm{e}^{i\bf {k.r}}$ represents the disorder effects on the first order correlation function. The behavior of $g_R^{(1)}({\bf r})$ is displayed in Fig.\[corr\]. We observe that for small disorder correlation length ($\sigma/\xi =0.2$), $g_R^{(1)}(r) $ is decreasing with increasing the TBI and the DDI (see Fig.\[corr\]. a-b ). The same behavior holds for large $\sigma$. Importantly, $g_R^{(1)}(r) $ vanishes at large distance $r$ in both cases signaling the non-existence of mini condensates formed by the localized particles in the respective minima of the external random potential. This does not mean that the long-range order of the whole system is destroyed.
The last term in Eq.(\[corrF\]) accounts for the quantum and thermal contributions to the one-body correlation function. One can easily show that this term decays at $r \rightarrow \infty$ and thus, $g^{(1)}({\bf r})$ tends to $n_c$, revealing the existence of the long-range order (true condensate). Note that the DDI, the TBI and the temperature can also shift the one-body correlation function.
Thermodynamic quantities {#Therm}
========================
In this section, we calculate disorder corrections to some thermodynamic quantities such as the chemical potential and the ground state energy.\
Within the realm of the HFB theory, the chemical potential can be written as $$\label{EoS}
\mu= \mu_0+\delta\mu+2\mu_R,$$ where $$\label{EoS1}
\mu_0=\bar V(0) n_c,$$ is the first-order chemical potential [@Boudj6].\
Corrections to the chemical potential due to the disorder effects are given as $$\label{EoS2}
\mu_R= g_2 n_{\text{HM}} (1+g_3n_c/g_2)^{1/2} H (\gamma,\sigma/\xi),$$ where $$\begin{aligned}
H (\gamma,\sigma/\xi)=\frac{1}{2} \int^\pi_0 \mathrm{d}\theta\sin\theta\sqrt{1+\gamma(3\cos^2\theta-1)}S(\alpha),\end{aligned}$$ Corrections to the chemical potential due to the quantum and thermal fluctuations are defined as : $\delta \mu=\sum\limits_{\bf k} \bar V (\mathbf k) \left[v_k(v_k-u_k)+(v_k-u_k)^2N_k \right]$ [@BoudjDp; @Boudj6]. Nevertheless, this chemical potential cannot be evaluated straightforwardly since the zero-temperature term is ultraviolet divergent. Such a problem can be worked out either by using the dimensional regularization which is valid for very dilute gases [@Anders; @Yuk; @Boudjbook] or by renormalizing the contact interaction through the $T-$matrix approach [@FV]. After some algebra, the resulting corrections to the chemical potential read $$\begin{aligned}
\label{EoS3}
\frac{\delta\mu}{g_2 n_c}&=\frac{32}{3} \sqrt{\frac{n_ca^3}{\pi}} (1+g_3n_c/g_2)^{5/2} {\cal Q}_5 (\gamma) \nonumber\\
&+\frac{2}{3} \left(\frac{\pi T}{n_cg_2} \right)^2\sqrt{\frac{n_c a^3}{\pi}} (1+g_3n_c/g_2)^{1/2} {\cal Q}_{1}(\gamma) .
\end{aligned}$$
![(Color online) Disorder energy function $ h_1(\gamma,\sigma/\xi) $ versus $ g_3n_c/g_2 $ for $\sigma/\xi =0.5$. Black line: $ \epsilon_{dd}=0.16 $, blue line: $ \epsilon_{dd}=0.4 $, and red line: $ \epsilon_{dd}=0.8 $.[]{data-label="Efunc"}](energy.eps)
Importantly, for $g_3=0$, the total chemical potential (\[EoS\]) reduces to that obtained in our recent work [@Boudj6]. For a cleaned ($R_0=0$) condensate with two-body contact interactions ($g_3= \epsilon_{dd}=0$), the obtained corrections to the chemical potential coincide with the seminal Lee-Huang-Yang quantum corrected equation of state [@LHY].\
The energy shift (\[EGY2\]) due to the disorder effects is finite and it can be evaluated as $$\frac{E_R}{N}= \frac{2mR_0}{\hbar^2} (1+g_3n_c/g_2)^{1/2} \sqrt{\frac{n_c a }{\pi}} {\cal H} (\gamma,\sigma/\xi),$$ where the function $$\begin{aligned}
{\cal H} (\gamma,\sigma/\xi)=\frac{1}{2} \int^\pi_0 \mathrm{d}\theta\sin\theta\sqrt{1+\gamma(3\cos^2\theta-1)}S_1(\alpha),\end{aligned}$$ and the function $ S_1(\alpha)=e^{2\alpha} \text{erfc} (\sqrt{2\alpha})-\sqrt{1/2\alpha} $. The disorder energy function $ h_1(\gamma,\sigma/\xi) $ is decreasing with $ g_3 $ as is seen in Fig.\[Efunc\] indicating that the TBI lead to lower the energy due to the disorder fluctuations which is in agreement with the above results. We observe also that for $ g_3n_c/g_2 \leq 0.7 $, the DDI effects on the energy are more pronounced.
Corrections to the energy due to the quantum and thermal fluctuations can be calculated easily through Eq.(\[EGY1\]) or by integrating the chemical potential (\[EoS3\]) with respect to the density.
Conclusion {#Conc}
==========
In this paper, we investigated the properties of dipolar Bose gas with TBI subjected to a correlated Gaussian disorder. We showed that the DDI may lead to arrest transport of atoms under disorder augmenting the glassy fraction inside the condensate, while the presence of the TBI may lead to a diffusive motion of particles. We pointed out that the one-body density matrix is a decreasing function with the TBI. We calculated in addition the chemical potential of a disordered dipolar BEC and ultraviolet divergences are removed by means of dimensional regularization. The combined effects of the DDI, TBI, and temperature found to crucially affect the chemical potential and the ground state energy of the system.
Author Contributions {#author-contributions .unnumbered}
====================
All authors discussed the results and made critical contributions to the work. AB contributed to the writing of the manuscript.
[28]{}
See for review: M. A. Baranov, Physics Reports [**464**]{}, 71 (2008). See for review: T. Lahaye et al., Rep. Prog. Phys. [**72**]{}, 126401 (2009). See for review: L.D. Carr, D. DeMille, R.V. Krems, and J. Ye, New Journal of Physics [**11**]{}, 055049 (2009). See for review: M.A. Baranov, M. Delmonte, G. Pupillo, and P. Zoller, Chemical Reviews, [**112**]{}, 5012 (2012).
C. Krumnow and A. Pelster, Phys. Rev. A [**84**]{}, 021608(R) (2011). B. Nikolic, A. Balaz, and A. Pelster, Phys. Rev. A [**88**]{}, 013624 (2013). M. Ghabour and A. Pelster, Phys. Rev. A [**90**]{}, 063636 (2014). A. Boudjemâa, Phys. Rev. A [**91**]{}, 053619 (2015). A. Boudjemâa, Low Temp. Phys. [**180**]{}, 377 (2015). A. Boudjemâa, Phys. Lett. A [**379**]{}, 2484 (2015). A. Boudjemâa, J. Phys. B: At. Mol. Opt. Phys. [**49**]{}, 105301 (2016). A. Boudjemâa, arXiv:1707.07477 (2017).
V. Efimov Phys. Lett. B [**33**]{}, 563 (1970). V. Efimov, Sov. J. Nucl. Phys. [**12**]{} 589 (1971). P F Bedaque, E. Braaten and H-W-Hammer, Phys. Rev. Lett. [**85**]{} 908 (2000 ). T. Kraemer et [*a*l]{}. Nature [**440**]{}, 315 (2006). E. A. Burt, R. W. Ghrist, C. J. Myatt, M J. Holland, E. A. Cornell and C.E. Wieman, Phys. Rev. Lett. [**79**]{} 337 (1997). A. Bulgac, Phys. Rev. Lett. [**89**]{}, 050402 (2002). A. J. Daley, J.M. Taylor, S. Diehl, M. Baranov, and P. Zoller, Phys. Rev. Lett. [**102**]{}, 040402 (2009). L. Mazza, M. Rizzi, M. Lewenstein, and J. I. Cirac, Phys. Rev. A [**82**]{}, 043629 (2010). M. Singh, A. Dhar, T. Mishra, R. V. Pai, B. P. Das, Phys. Rev. A [**85**]{}, 051604 (2012). K.W. Mahmud and E. Tiesinga, Phys. Rev. A [**88**]{}, 023602 (2013). F.K. Abdullaev, A. Gammal, L. Tomio and T. Frederico, Phys. Rev. A [**63**]{} 043604 (2001). H. Al-Jibbouri, I. Vidanovic, A. Balaz, and A. Pelster, J. Phys. B: At. Mol. Opt. Phys. [**46**]{}, 065303 (2013). H-C Li, K-J. Chen and J-K Xue, Chin. Phys. Lett. [**27**]{} 030304 (2010) Peng P and Li G-Q Chin. Phys. B [**18**]{}, 3221 (2009). M. S. Mashayekhi, J.-S. Bernier, D. Borzov, J.-L. Song, and F. Zhou, Phys. Rev. Lett. [**110**]{}, 145301 (2013). Zhen-Kai Lu, Yun Li, D. S. Petrov, and G. V. Shlyapnikov, Phys. Rev. Lett. [**115**]{}, 075303 (2015). H. Kadau, M. Schmitt, M. Wenzel, C. Wink, T. Maier, I. Ferrier-Barbut and T. Pfau, Nature [**530**]{} ,194 (2016). Kui-Tian Xi and Hiroki Saito, Phys. Rev. A [**93**]{}, 011604(R) (2016). P. B. Blakie, Phys. Rev. A [**93**]{}, 033644 (2016). L. Chomaz, S. Baier, D. Petter, M. J. Mark, F. Wächtler, L. Santos and F. Ferlaino, Phys. Rev. X [**6**]{}, 041039 (2016). A. Boudjemâa, Annals of Physics, [**381**]{}, 68 (2017). A. Boudjemâa, J. Phys. B: At. Mol. Opt. Phys. [**51**]{}, 2 (2017). E. Braaten and A. Nieto, Eur. Phys. J. B [**11**]{}, 143 (1999). A. X. Zhangand and J. K. Xue, Phys.Rev. A [**75**]{}, 013624 (2007). P. J. Everitt, M. A. Sooriyabandara, G. D. McDonald, K. S.Hardman, C. Quinlivan, M. Perumbil, P. Wigley, J. E. Debs, J. D. Close, C. C. N. Kuhn, N. P. Robins, arXiv:1509.06844v1 (2015). E. Braaten, H.-W. Hammer, and T. Mehen, Phys. Rev. Lett. [**88**]{}, 040401 (2002).
K. Huang and H. F. Meng, Phys. Rev. Lett. [**69**]{}, 644 (1992). A. Boudjemâa, J. Phys. B: At. Mol. Opt. Phys. [**48**]{} 035302 (2015). J. O. Andersen, Rev. Mod. Phys [**76**]{}, 599 (2004). V. I. Yukalov, Phys. Part. Nucl. [**42**]{}, 460 (2011). A. Boudjemâa, Degenerate Bose Gas at Finite Temperatures, LAP LAMBERT Academic Publishing (2017).
Alexander L. Fetter and John Dirk Walecka, Theoretical Mechanics of Particles and Continua, Dover Publications (2003). T. D. Lee, K. Huang and C. N. Yang, Phys. Rev [**106**]{}, 1135 (1957).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study the conformally invariant variational problem for time-like curves in the $n$-dimensional Einstein universe defined by the conformal strain functional. We prove that the stationary curves are trapped into an Einsetin universe of dimension $2$, $3$ or $4$. We study the linearly-full stationary curves in a four-dimensional Einstein universe and we show that they can be integrated by quadratures in terms of elliptic functions, elliptic integrals and Jacobi’s theta functions.'
address:
- '(O. Eshkobilov) Dipartimento di Matematica, Università di Torino, Via Carlo Alberto, 10 - I-10123 Torino, Italy'
- '(E. Musso) Dipartimento di Scienze Matematiche, Politecnico di Torino, Corso Duca degli Abruzzi 24, I-10129 Torino, Italy'
author:
- Olimjon Eshkobilov
- Emilio Musso
date: Version of
title: 'A conformally invariant variational problem for time-like curves'
---
=1
[^1]
Introduction {#s:intro}
============
Lorentzian conformal geometry began in $1918$ with the seminal work of H.Weyl [@W]. In the 1980s, it has been considerably developed in connection with the twistor approach to gravity by Penrose and Rindler [@PR] and it is widely used in the cyclic cosmological models in general relativity [@Pn1; @Pn3; @Tod]. It is also related to regularization of the Kepler problem [@GS] and Lie sphere geometry [@Blaschke; @Ce; @JMN]. The purpose of the present paper is to investigate a variational problem for time-like curves in the conformal compactification $\mathcal{E}^{1,n-1}$ of the Minkowski n-space. In the literature, $\mathcal{E}^{1,n-1}$ is often referred to as the [*compact $n$-dimensional Einstein universe*]{} [@BCDG; @Fr; @GS]. The choice of working within the compact model is motivated by the fact that, in this case, the conformal group is a group of matrices. This greatly simplifies the computational aspects. However, our considerations apply as well in the non-compact case. It ’should be noted that any Lorentzian space-form or any Friedmann–Lemaitre–Robertson–Walker cosmological model can be realized via a conformal embedding, as an open domain of $\mathcal{E}^{1,n-1}$ [@HE].
Lorentzian conformal geometry shares, at least on a formal level, many common features, with Riemannian conformal geometry, a classical topic in differential geometry. The Lorentzian conformal geometry of the Einstein universe can be developed in analogy with the conformal geometry of a sphere (Möbius geometry). For instance, the existence of a canonical conformal invariant arc-element along a generic curve in $S^n$ is a well known fact that goes back to the works [@F; @Ha; @L; @T; @V] of Fialkov, Haantjes, Liebmann, Takasu and Vessiot, all published in the first decades of the past century. In [@DMN] the construction of the conformal arc-element has been extended to time-like curves in a $3$-dimensional Einstein universe. In this paper we show how to define a conformal invariant arc-element for a time-like curve in an Einstein universe of arbitrary dimension. The integral of the conformal arc-element defines the simplest conformal invariant variational problem on the space of generic time-like curves of $\mathcal{E}^{1,n-1}$, called the conformal strain functional. This is the Lorentzian counterpart of the conformal arc-length functional in Möbius geometry [@LO2010; @LS; @MMR; @M1; @M2; @MN] and generalizes the homonymous functional for time-like curves in $\mathcal{E}^{1,2}$, previously considered in [@DMN]. Proceeding in analogy with [@MMR] and using the method of moving frames we deduce the variational equations satisfied by the stationary curves, referred to as conformal world-lines. We prove that conformal world-lines in $\mathcal{E}^{1,n-1}$ are trapped into a totally geodesic Einstein universe of dimension $2$, $3$ or $4$. This is the Lorentzian analogue of a similar result in Möbius geometry [@MMR]. The $2$-dimensional case is rather trivial since the trajectories are orbits of $1$-parameter groups of conformal transformations. The $3$-dimensional case is more challenging and it has been partially investigated in [@DMN]. Here, we focus on the world-lines of $\mathcal{E}^{1,3}$ which are not trapped in a lower dimensional totally umbilical Einstein universe (linearly-full world lines). We prove that their trajectories can be explicitly integrated by quadratures in terms of elliptic functions and elliptic integrals.
The paper is organized as follows. In Section \[s1\], we collect from the literature few basic facts about conformal Lorentzian geometry [@BCDG; @GS] and we reformulate in the Lorentzian context the classical approach to the conformal geometry of curves in the Möbius space [@F; @Ha; @Mon; @SS; @S; @T; @Thomsen]. We define the conformal strain and the conformal line-element, which is the Lorentzian analogue of the conformal arc element of a curve in $S^n$ [@LO2010; @MMR; @Mon; @M1; @M2; @MN]. In Section \[s2\], we use the moving frame method to compute the Euler-Lagrange equations of the conformal strain functional (Theorem \[thmA\]). Consequently we show that the trajectory of a conformal world-line lies in a totally umbilical Einstein universe of dimension $2$, $3$ or $4$ (Theorem \[ThmB\]). In Section \[s3\] we study linearly full conformal world-lines of a four-dimensional Einstein universe. Given such a world-line we build a canonical lifting to the conformal group, the canonical conformal frame and we define the three conformal curvatures (Proposition \[propuniq\]). Then we use the variational equation to show that the curvatures are either constants or else can be expressed in terms of Jabobi’s elliptic functions (Proposition \[ConfCurv\]). In the first case the trajectory is an orbit of a $1$-parameter group of conformal transformations. Leaving aside the constant curvature case, we prove that the conformal equivalence classes of world lines can be parameterized by three real parameters (Proposition \[PHRS\]). Then we define the momentum operator, that is an element of the Lie algebra of the conformal group, intrinsically defined by the world-line. The existence of the momentum is a consequence of the conformal invariance of the functional and of the Nöther conservation theorem (Remark \[ThEx\]). We prove that the momentum operator is either a regular or an exceptional element of the conformal Lie algebra (Proposition \[CT\]). Then we define the integrating factors of an eigenvalue of the momentum and we build the principal vectors of the eigenvalues (Proposition \[PV\]). The explicit computation of the integrating factors is rather technical and is considered in the Appendix. In Section $5$ we use the integrating factors and the principal vectors to integrate by quadratures the trajectory of a linearly full conformal world-line with non-constant curvatures (Theorem \[ThmC\] and Theorem \[ThmD\]). At the end of the section we briefly comment the theoretical aspects underlying the integration by quadratures, we explain why a linearly full world-line with non-constant curvatures, can’t be closed and we show that the trajectory of a world-line is invariant by the action of an infinite cyclic group of conformal transformations.
Conformal geometry of a time-like curve {#s1}
=======================================
The Einstein universe and its restricted conformal group
--------------------------------------------------------
We denote by $\mathcal{E}^{1,n-1}$ the n-dimensional sub-manifold of $\R^{n+2}, n>1$, defined by the equations $x_0^2+x_1^2=1$ and $x_2^2+\dots + x_{n+1}^2=1$. As a manifold, $\mathcal{E}^{1,n-1}$ is the Cartesian product $S^1 \times S^{n-1}$. The restriction of the quadratic form $$g=-dx_0^2-dx_1^2+dx_2^2+\dots + dx_{n+1}^2$$ induces a Lorentzian pseudo-metric $g_{\textsl{e}}$ on $\mathcal{E}^{1,n-1}$. The normal bundle of $\mathcal{E}^{1,n-1}$ is spanned by the restrictions of the vector fields $\mathbf{n}_1=x_0\partial_{x_0}+x_1\partial_{x_1}$ and $\mathbf{n}_2=x_2\partial_{x_2}+\dots +x_{n+1}\partial_{x_{n+1}}$. Thus, contracting $dx_0\wedge\dots dx_{n+1}$ with $\mathbf{n}_1$ and $\mathbf{n}_2$, we get a volume form on $\mathcal{E}^{1,n-1}$ which in turn defines an orientation. The vector field $-x_1\partial_{x_0}+x_0\partial_{x_1}$ is tangent to $\mathcal{E}^{1,n-1}$ and induces a unit time-like vector field on $\mathcal{E}^{1,n-1}$. We time-orient $\mathcal{E}^{1,n-1}$ by requiring that such a vector field is future-oriented.
The Lorentzian manifold $(\mathcal{E}^{1,n-1},g_{\textsl{e}})$, with the above specified orientation and time-orientation, is called the [*$n$-dimensional Einstein universe*]{}.
In order to describe the conformal geometry of $\mathcal{E}^{1,n-1}$ it is convenient to consider in $\R^{n+2}$ the coordinates $$y_0=\frac{1}{\sqrt{2}}(x_0+x_{n+1}),\quad y_1=x_1, \dots , y_n=x_n,\quad y_{n+1}=\frac{1}{\sqrt{2}}(-x_0+x_{n+1}).$$ The corresponding basis of $\R^{n+2}$ is denoted by $(E_0,\dots, E_{n+1})$ and is said the [*standard (light-cone) basis*]{} of $\R^{n+2}$. With respect to $(y_0,\dots , y_{n+1})$ the scalar product associated to $g$ can be written as $$\label{sp}\langle Y,Y'\rangle=-(y_0y'_{n+1}+y_{n+1}y'_0)-y_1y'_1+\sum_{j=2}^{n}y_jy'_j.$$ In addition, $dV=dy_0\wedge \dots \wedge y_{n+1}$ is a positive volume form. From now on we will use the light-cone coordinates and the elements $V\in \R^{n+2}$ are thought of as column vectors constructed from the components of $V$ with respect to $(E_0,\dots ,E_{n+1})$. For each $V\in \R^{n+2}$, $V\neq 0$, we denote by $[V]$ the oriented line spanned by $V$. The map $ V\in \mathcal{E}^{1,n-1}\to [V]$ allow us to identify $\mathcal{E}^{1,n-1}$ and the manifold of the isotropic oriented lines through the origin of $\R^{n+2}$ (null rays). Using such an identification, the connected component of the identity $\mathrm{A}^{\uparrow}_+(2,n)$ of the pseudo-orthogonal group of (\[sp\]) acts transitively and effectively on the left of $\mathcal{E}^{1,n-1}$ by $\mathbf{X}[V]=[\mathbf{X}\cdot V]$. The action preserves the oriented, time-oriented conformal Lorentzian structure of $\mathcal{E}^{1,n-1}$. It is a classical result that, if $n>2$, every restricted conformal transformation of $\mathcal{E}^{1,n-1}$ is induced by a unique element of $\mathrm{A}^{\uparrow}_+(2,n)$ [@DNF; @Fr]. For this reason, we call $\mathrm{A}^{\uparrow}_+(2,n)$ the [*restricted conformal group of the $n$-dimensional Einstein universe*]{}.
[To distinguish the connected component of the identity we proceed as in [@GS] : we consider the cone $\mathcal{C}\subset \bigwedge^2(\R^{n+2})$ of all isotropic bi-vectors, ie the non-zero decomposable elements $V\wedge W$ of $\bigwedge^2(\R^{n+2})$ such that $\langle V,V\rangle = \langle W,W\rangle = \langle V,W\rangle = 0$. The function $$\mathfrak{V}: V\wedge W \in \mathcal{C} \to dV(V,W,E_1,\dots E_n,E_{n+1}-E_0)\in \R$$ never vanishes and the half cones $$\mathcal{C}_+=\{V\wedge W\in \mathcal{C} : \mathfrak{V}(V\wedge W)>0\},\quad \mathcal{C}_-=\{V\wedge W\in \mathcal{} : \mathfrak{V}(V\wedge W)<0\}$$ are the two connected components of $\mathcal{C}$. Then, $\mathrm{A}^{\uparrow}_+(2,n)$ is the group of all pseudo-orthogonal matrices $\mathbf{B}$ of the scalar product (\[sp\]) such that $\mathrm{det}(\mathbf{B})=1$ and $\mathbf{B}\cdot \mathcal{C}_+=\mathcal{C}_+$.]{}
We put $$\label{mm}\mathtt{m}=(\mathtt{m}_{ji}),\quad \mathtt{m}_{ji}=\langle \mathrm{E}_j,\mathrm{E}_i\rangle,\quad i,j=0,\dots, n+1.$$ Then the column vectors $B_0,\dots B_{n+1}$ of a matrix $\mathbf{B}\in \mathrm{A}^{\uparrow}_+(2,n)$ constitute a [*light-cone basis*]{} of $\R^{n+2}$, ie a positive-oriented basis such that $$\langle B_i,B_j\rangle = \mathtt{m}_{ji}, \quad i,j=0,\dots n+1,\quad B_0\wedge(B_1+B_2)\in \mathcal{C}_+.$$ This allow us to identify $\mathrm{A}^{\uparrow}_+(2,n)$ and the manifold of all light-cone basis of $\R^{n+2}$. We use the notation $\R^{2,n}$ to denote $\R^{n+2}$ equipped with the scalar product $(\ref{sp})$, the volume form $dV$ and the positive half cone $\mathcal{C}_+$. Differentiating the $\R^{2,n}$-valued maps $$\mathcal{B}_j:B \in \mathrm{A}^{\uparrow}_+(2,n)\to B_j\in \R^{2,n},\quad j=0,\dots,n+1,$$ yields $$d\mathcal{B}_j=\sum_{i=0}^{n+1}\mu_j^i \mathcal{B}_i,$$ where $\mu^i_j$ are left-invariant 1-forms. The conditions $\langle B_j,B_i\rangle= \mathtt{m}_{ji}$ imply that $\mu =(\mu^i_j)$ takes values in the Lie algebra $$\mathfrak{a}(2,n)=\{X\in \mathfrak{gl}(n+2,\R) / ^tX\cdot \mathtt m + \mathtt m\cdot X=0\}.$$ of $\mathrm{A}^{\uparrow}_+(2,n)$. Consequently we can write $$\mu = \left(
\begin{array}{cccc}
\mu^0_0 & -\mu^1_{n+1} & \mu_{n+1} & 0 \\
\mu^1_0 & 0 & ^t\mu_1 & \mu^1_{n+1} \\
\mu_0 & \mu_1 & \widetilde{\mu} & \mu_{n+1} \\
0 & -\mu^1_0 & ^t\mu_0 & -\mu^0_0 \\
\end{array}
\right),$$ where $$\mu_0=^t(\mu^2_0,\dots, \mu^n_0),\quad \mu_1=^t(\mu^2_1,\dots, \mu^n_1),\quad
\mu_{n+1}=^t(\mu^2_{n+1},\dots, \mu^n_{n+1})$$ and $^t\widetilde{\mu}+\widetilde{\mu}=0$. The left-invariant 1-forms $\mu^0_0$, $\mu^j_0$, $\mu^j_{n+1}$, $j=1,\dots, n$ and $\mu^i_j$, $1\le i<j=1,\dots n$ are linearly independent and span the dual of the Lie algebra $\mathfrak{a}(2,n)$. They satisfy the [*Maurer-Cartan equations*]{} $$\label{MC}d\mu^i_j=-\sum_{k=0}^{n+1}\mu^i_k\wedge \mu^k_j,\quad i,j=0,\dots, n+1.$$
Let $\mathbb{M}^{1,n-1}$ be [*Minkowski n-space*]{}, i.e., the affine space $\R^n$ with the Lorentzian scalar product $$(\mathbf{p},\mathbf{q})=-p_1q_1+p_2q_2+\dots + p_nq_n.$$ The map $$\label{j}
\mathbf{j}(\mathbf{p})= [^t(1,p_1,\dots,p_n,\frac{1}{2}(\mathbf{p},\mathbf{p}))]\in \mathcal{E}^{1,n-1}$$ is a conformal embedding whose image is said the [*Minkowski-chamber*]{} of $\mathcal{E}^{1,n-1}$ Let $\mathrm{P}^{\uparrow}_+(1,n-1)=\mathbb{M}^{1,n-1}\rtimes \mathrm{SO}^{\uparrow}_+(1,n-1)$ be the restricted Poincaré group of $\mathbb{M}^{1,n-1}$. For each $(\mathbf{p},\mathrm{L})\in \mathrm{P}^{\uparrow}_+(1,n-1)$ we put $^{*}\mathbf{p}=(-p_1,p_2,\dots, p_n)$. The matrix $$\mathbf{B}(\mathbf{p},\mathrm{L})=\left(
\begin{array}{ccc}
1 & 0 & 0 \\
\mathbf{p} & \mathrm{L} & 0 \\
\frac{1}{2}^{*}\mathbf{p}\cdot \mathbf{p} & ^{*}\mathbf{p}\cdot \mathrm{L} & 1 \\
\end{array}
\right)$$ belongs to $\mathrm{A}^{\uparrow}_+(2,n)$ and $$\label{J}\mathbf{J}:(\mathbf{p},\mathrm{L})\in \mathrm{P}^{\uparrow}_+(1,n-1)\to \mathbf{B}(\mathbf{p},\mathrm{L}))\in \mathrm{A}^{\uparrow}_+(2,n)$$ is a faithful representation. In a similar way, one can build conformal embeddings of the de-Sitter or anti de-Sitter $n$-spaces into $\mathcal{E}^{1,n-1}$. Also the Robertson-Walker $n$-spaces can be conformally mapped as open sub-manifolds of the Einstein universe [@HE].
[A $(2+h)$-dimensional vector subspace $\mathbb{W}\subset \R^{2,n}$, $h=1,...,n-1$ is said [*Lorentzian*]{} if the restriction of $\langle -, -\rangle $ to $\mathbb{W}$ is non-degenerate, of signature $(2,h)$. The set $\mathcal{E}(\mathbb{W})$ of all null rays belonging to $\mathbb{W}$ is a $h$-dimensional totally umbilical Lorentzian sub-manifold of $\mathcal{E}^{1,n-1}$. We call $\mathcal{E}(\mathbb{W})$ a [*$h$-dimensional Lorentzian cycle*]{} of $\mathcal{E}^{1,n-1}$. If we equip $\mathcal{E}(\mathbb{W})$ with the induced conformal structure we get a Lorentzian conformal space equivalent to a $h$-dimensional Einstein universe.]{}
Time-like curves
----------------
Let $\gamma : {I}\subset \R\to \mathcal{E}^{1,n-1}$ be a parameterized time-like curve, $n\ge 3$. A [*null-lift*]{} is a map $\Gamma :I\to \R^{2,n}$ such that $\gamma(t)=[\Gamma(t)]$, for every $t\in I$. Since $\gamma$ is a time-like immersion, its null-lifts satisfy $$\Gamma\wedge \Gamma'|_t\neq 0,\quad \langle \Gamma|_t,\Gamma|_t=\langle \Gamma|_t,\Gamma'|_t\rangle =0,\quad
\langle \Gamma'|_t,\Gamma'|_t\rangle < 0,$$ for every $t\in I$. For each $k=1,...,n+1$, the [*$k$-th osculating space*]{} of $\gamma$ at $\gamma(t)$ is the linear subspace $\mathcal{T}^k(\gamma)|_t\subset \R^{2,n}$ spanned by the vectors $\Gamma|_t,\Gamma'|_t,\dots ,\Gamma^k|_t$. Obviously the definition does not depend on the choice of the null lift. Putting together all the osculating spaces we get the k-th osculating sheaf $$\mathcal{T}^{k}(\gamma)=\{(t,V)\in I\times \R^{2,n} : V\in \mathcal{T}^k(\gamma)|_t\}.$$
[$\mathcal{T}^2(\gamma)|_t$ is a $3$-dimensional Lorentzian subspace of $\R^{2,n}$, for every $t\in I$.]{}
[Choose $\Gamma$ so that $\langle \Gamma',\Gamma'\rangle = -1$. Differentiating $\langle \Gamma,\Gamma'\rangle = 0$ and $\langle \Gamma',\Gamma'\rangle = -1$ we get $\langle \Gamma,\Gamma''\rangle = 1$ and $\langle \Gamma',\Gamma''\rangle = 0$. This implies that $\Gamma|_t,\Gamma'|_t$ and $\Gamma''|_t$ are linearly independent, for each $t\in I$. We put $$A_1= \frac{1}{2}(1+\langle \Gamma'',\Gamma''\rangle)\Gamma - \Gamma'',\quad A_2=\Gamma' \quad A_3=\frac{1}{2}(1-\langle \Gamma'',\Gamma''\rangle)\Gamma + \Gamma''.$$ Then, $(A_1|_t,A_2|_t,A_3|_t)$ is a basis of $\mathcal{T}^2(\gamma)|_t$ such that $$\begin{split}
\langle A_i|_t,A_j|_t\rangle &= 0,\quad i\neq j,\\
\langle A_1|_t,A_1|_t\rangle &= \langle A_2|_t,A_2|_t\rangle=-\langle A_3|_t,A_3|_t\rangle = -1.
\end{split}$$ This yields the result.]{}
The Lemma implies that $\mathcal{T}^k(\gamma)|_t$ is a Lorentzian subspace of $\R^{2,n}$, for every $k\ge 2$ and every $t\in I$. The Lorentzian cycle $\mathcal{E}(\mathcal{T}^k(\gamma)|_t)$ is denoted by $\mathcal{E}^k(\gamma)|_t$. We call $\mathcal{E}^k(\gamma)|_t$ the [*$k$-th osculating cycle*]{} of $\gamma$ at $\gamma(t)$. The orthogonal complement $\mathcal{N}^k(\gamma)|_t$ of $\mathcal{T}^k(\gamma)|_t$ is a space-like vector subspace such that $\R^{2,n}=\mathcal{T}^k(\gamma)|_t \oplus \mathcal{N}^k(\gamma)|_t$. We call $\mathcal{N}^k(\gamma)|_t$ the [*$k$-th normal space*]{} at $\gamma(t)$ and we define the [*k-th normal sheaf*]{} by $$\mathcal{N}^{k}(\gamma)=\{(t,V)\in I\times \R^{2,n} : V\in \mathcal{N}^k(\gamma)|_t\}.$$ The orthogonal projection of $\R^{n+2}$ onto $\mathcal{N}^k(\gamma)|_t$ is denoted by $pr_{(k)}|_t$.
[The [*conformal strain*]{} of a time-like curve is the quartic differential $$\label{strain-density}\mathcal{Q}_{\gamma}=
\frac{\langle pr_{(2)}(\Gamma'''),
pr_{(2)}(\Gamma''')\rangle}{|\langle \Gamma',\Gamma'\rangle|}dt^4$$]{}
The conformal strain is independent on the choice of the null lift and, in addition, if $\gamma$ and $\widetilde{\gamma}$ are two [*equivalent time-like curves*]{}[^2] then $\mathcal{Q}_{\widetilde{\gamma}}=h^*(\mathcal{Q}_{\gamma})$.
[If $\mathcal{Q}(\gamma)|_t=0$, the point $\gamma(t)$ is said to be a [*conformal vertex*]{}. A time-like curve without conformal vertices is said [*generic*]{}. If $\mathcal{Q}_{\gamma}=0$, the curve is said [*totally degenerate*]{}.]{}
[If $\gamma$ is totally degenerate then $\mathcal{E}^1(\gamma)|_t$ is constant and the trajectory of $\gamma$ is contained in a 1-dimensional conformal cycle.]{}
It can be shown [@DMN] that $\mathcal{E}^1(\gamma)|_t$ has second order analytic contact with $\gamma$ at $\gamma(t)$. Moreover, $\gamma(t)$ is a conformal vertex if and only if the order of contact is strictly bigger than 2. This highlights the fact that the conformal strain measures the infinitesimal distorsion of the curve from its osculating cycle.
[If $\gamma$ is non-degenerate then there exist a unique null lift, referred to as the [*canonical lift*]{} such that $\mathcal{Q}_{\gamma}= \langle \Gamma'',\Gamma\rangle^2dt^4$. The smooth positive function $\upsilon_{\gamma}=|\langle \Gamma',\Gamma'\rangle|^{1/2}$ is the [*conformal strain density*]{} and the exterior differential $1$-form $\sigma_{\gamma}=\upsilon_{\gamma}dt$ is called the [*conformal arc element*]{} of $\gamma$. By construction, $\sigma_{\gamma}$ is invariant under the action of the restricted conformal group and orientation-preserving changes of parameter. In particular, each generic time-like curve can be parameterized is such a way that $\upsilon_{\gamma}=1$. In this case, we say that the curve is [*parameterized by conformal parameter*]{}, which is usually denoted by $u$. Given a smooth map $f:I\to \R^h$, we define the [*derivative of $f$ with respect to the conformal arc element*]{} by $\dot{f}=\upsilon_{\gamma}^{-1}f'$.]{}
[The $1$-form $\sigma_{\gamma}$ is the Lorentzian analogue of the conformal arc element of a curve in the 3-dimensional round sphere [@LO2010; @MMR; @M1; @MN] and generalizes the analogue notion for generic time-like curves in the $3$-dimensional Einstein universe [@DMN].]{}
Let $\gamma$ be non-degenerate and $\Gamma$ be its canonical lift. Then, $\mathrm{dim}(\mathcal{T}^3(\gamma)|_t)=4$, for every $t$. Hence $\mathcal{T}^3(\gamma)$ is a vector bundle. Furthermore, the cross sections $$\label{cvf1}
M_0=\Gamma,\quad M_1=\frac{1}{|\langle \dot{\Gamma},\dot{\Gamma}\rangle|^{1/2}}\dot{\Gamma},\quad
M_2=\frac{pr_{(2)}{\dddot{(\Gamma)}}}{\langle pr_{(2)}(\dddot{\Gamma}),pr_{(2)}(\dddot{\Gamma})\rangle)^{1/2}},$$ and $$\label{cvf2}
M_{n+1}=-\frac{1}{\langle \Gamma,\ddot{\Gamma}\rangle}\ddot{\Gamma}+\frac{\langle \dot{\Gamma},\ddot{\Gamma}\rangle}{\langle \Gamma,\ddot{\Gamma}\rangle}\dot{\Gamma}+\frac{1}{2}(
\frac{\langle \ddot{\Gamma},\ddot{\Gamma}\rangle}{\langle \Gamma,\ddot{\Gamma}\rangle^2}
- \frac{\langle \ddot{\Gamma},\dot{\Gamma}\rangle^2}{\langle \Gamma,\ddot{\Gamma}\rangle^2})\Gamma$$ give rise to a [*canonical trivialization*]{} $(M_0,M_1,M_2,M_{n+1})$ of $\mathcal{T}^3(\gamma)$. The canonical trivialization satisfies $$\label{CT}
\begin{split}
& \langle M_0,M_0\rangle = \langle M_{n+1},M_{n+1}\rangle = \langle M_0,M_1\rangle = \langle M_0,M_2\rangle=0,\\
& \langle M_1,M_{n+1}\rangle=\langle M_2,M_{n+1}\rangle=\langle M_1,M_2\rangle=0,\\
&\langle M_0,M_{n+1}\rangle = \langle M_1,M_1\rangle = -\langle M_2,M_2\rangle=-1.
\end{split}$$ In particular, $M_0|_t\wedge (M_1|_t+M_2|_t)$ is an isotropic bi-vector, for every $t\in I$. If we revert the orientation along the curve, this bi-vector change sign. Thus, each non-degenerate curve possesses an [*intrinsic orientation*]{} such that $M_0|_t\wedge (M_1|_t+M_2|_t)\in \mathcal{C}_{+}$, for every $t\in I$.
From now on we implicitly assume that the time-like curves are non-degenerate and equipped with their intrinsic orientations. In addition, if $\gamma$ is such a curve we use the notation $\Gamma$ to denote its canonical null lifting.
[The scalar product (\[sp\]) induces a metric structure on the vector bundle $\mathcal{N}^3(\gamma)$. The [*covariant derivative*]{} of a cross section $V:I\to \mathcal{N}^3(\gamma)$ with respect to the conformal line-element is defined by $$\label{D}
D(V)=pr_{(3)}(\dot{V}):I\to \mathcal{N}^3(\gamma).$$]{}
The normal bundles $\mathcal{N}^2(\gamma)$ and $\mathcal{N}^3(\gamma)$ possess two [*canonical cross sections*]{}, denoted by $W_{\gamma}$ and $S_{\gamma}$ respectively. The cross section $S_{\gamma}$ is defined by $$\label{SS1} S_{\gamma}=pr_{(3)}(\dot{M}_2).$$ We set $$\label{cinv1}h_1=\langle \dot{M}_1,M_{n+1}\rangle, \quad h_2=\sqrt{\langle S_{\gamma},S_{\gamma}\rangle}.$$ and we define $W_{\gamma}$ by $$\label{VD}W_{\gamma}=(\dot{h_1}-3h_2\dot{h_2})M_2-(h_2^2-2h_1)S_{\gamma}+D^2(S_{\gamma}).$$ Note that $$\begin{split}
\mathcal{T}^4(\gamma)|_t&=\mathrm{span}(M_0|_t,M_1|_t,M_2|_t,S_{\gamma}|_t,B_{n+1}|_t),\\
\mathcal{T}^5(\gamma)|_t&=\mathrm{span}(M_0|_t,M_1|_t,M_2|_t,S_{\gamma}|_t,D(S_{\gamma})|_t,B_{n+1}|_t),
\end{split}$$ for every $t\in I$.
First and second-order frames
-----------------------------
A [*first-order frame*]{} along $\gamma$ is a smooth map $$A=(A_0,\dots A_{n+1}):I\to \mathrm{A}^{\uparrow}_+(2,n)$$ such that $A_0$ is a null lift and $A'_0\in \mathrm{Span}(A_0,A_1)$. First-order frame fields do exist along any time-like curve. If $A$ is a first-order frame, then any other is given by $$\widetilde{A}=A\cdot X(r,x,y,R),$$ where $$r,x : I\to \R, x>0,\quad y:I\to \R^{n-1},\quad R:I\to \mathrm{SO}(n-1)$$ are smooth functions and $$\label{gauge}X(r,x,y,R)=\left(
\begin{array}{cccc}
r &-x & ^ty\cdot R & \frac{^ty\cdot y-x^2}{2} \\
0 & 1 & 0 & x/r \\
0 & 0 & R & y/r \\
0 & 0 & 0 & r^{-1} \\
\end{array}
\right).$$
[A first-order frame is said of the [*second-order*]{} if $$A_0=M_0,\quad A_1=M_1,\quad A_2=M_2,\quad A_{n+1}=M_{n+1}.$$ Second-order frames do exist along any generic, time-like curve with its intrinsic orientation. Note that $(A_3,\dots A_n)$ is a trivialization of the third-order normal bundle of $\gamma$.]{}
If $A$ is a second order frame field, then $$\label{LS}A' = A\cdot \left(
\begin{array}{ccccc}
0 & -h_1& 1 & 0 & 0 \\
1 & 0 & 0 & 0 & h_1 \\
0 & 0 & 0 & -^ts & 1 \\
0 & 0 & s & \phi & 0 \\
0 & -1 & 0 & 0 & 0 \\
\end{array}
\right)\upsilon_{\gamma},$$ where $\phi =(\phi^j_i)_{i,j=3,\dots,n} :I\to \mathfrak{o}(n-2)$ is a smooth map, $h_1$ is as in (\[cinv1\]) and $s=^t(s^3,\dots,s^n)$ is defined by $$S_{\gamma}=\sum_{j=3}^{n}s^jA_j.$$ If $V=\sum_{j=3}^{n}V^jA_j$ is a cross-section of $\mathcal{N}^3(\gamma)$, then $$DV=\sum_{j=3}^{n}(\dot{V}^j+\sum_{i=3}^{n}\phi^j_i V^i)A_j.$$ We then have $$\label{D12S}DS_{\gamma}= \sum_{j=3}^n s^j_{(1)}A_j,\quad D^2(S_{\gamma})= \sum_{j=3}^n s^j_{(2)}A_j,$$ and $$\label{VDL}
W_{\gamma}=(\dot{h}_1-3^ts\cdot \dot{s})A_2+\sum_{j=3}^{n}(\dot{s}^j_{(2)}+(2h_1-^ts\cdot s)s^j)A_j,$$ where $s_{(1)},s_{(2)}:I\to \R^{n-3}$ are given by $$s^j_{(1)} = \dot{s}^j+\sum_{i=3}^{n}\phi^j_i s^i,\quad
s^j_{(2)} = \dot{s}^j_{(1)}+\sum_{i=3}^{n}\phi^j_i s^i_{(1)}.$$
The strain functional and conformal world-lines
-----------------------------------------------
We have seen that we may construct, for a generic time-like curve, a canonical line-element $\sigma_{\gamma}$ on $I$. If $K\subset I$ is a closed interval in $I$, we can define the [*total strain functional*]{} $$\mathcal{S}_K(\gamma)=\int_K \sigma_{\gamma}$$ on the space of smooth, generic, time-like immersions of $I$ into $\mathcal{E}^{1,n-1}$.
[We say that $\gamma$ is a [*conformal world-line*]{} if for any closed interval $K\subset I$ and any smooth variation $$\mathbf{g}:(t,\tau)\in I\times (-\epsilon,\epsilon)\to \gamma_{\tau}(t)\in \mathcal{E}^{1,n-1},$$ with $\gamma_0=\gamma$ and[^3] $\mathrm{supp}(\mathbf{g})
\subset K$, we have $$\frac{d}{d\tau}\left( \mathcal{S}_K(\gamma_\tau)\right)|_t=0.$$]{}
The variational equations {#s2}
=========================
The purpose of this section is to prove the following
\[thmA\][A generic time-like curve equipped with its intrinsic orientation is a conformal world-line if and only if $W_{\gamma}$ vanishes identically. ]{}
First we prove that a generic time-like curve satisfying $W_{\gamma}=0$ is a conformal world line. Without loss of generality we assume that $\gamma$ is parameterized by conformal parameter. Keep in mind that $W_{\gamma}=0$ if and only if $$\label{EL1}
\dot{h}_1=3h_2\dot{h}_2,\quad D^2(S_{\gamma})=(h_2^2-2h_1)S_{\gamma}.$$ Let fix a second-order frame $A:I\to \mathrm{A}^{\uparrow}_+(2,n)$ and $$\mathbf{g}:(u,\tau)\in I\times (-\epsilon,\epsilon)\to \gamma_{\tau}(u)\in \mathcal{E}^{1,n-1}$$ be a compactly supported variation of $\gamma$ such that $\mathrm{supp}(\mathbf{g})\subseteq K$. Eventually shrinking the interval $(-\epsilon, \epsilon)$, we may assume that $\gamma_{\tau}$ is generic and equipped with its intrinsic orientation, for every $\tau$. Then, there is a differentiable map $$\mathcal{A}:(u,\tau)\in I\times (-\epsilon,\epsilon)\to \mathcal{A}_{\tau}(u)\in \mathrm{A}^{\uparrow}_+(2,n)$$ such that $\mathcal{A}_0=A$ and $\mathcal{A}_{\tau}$ is a second-order frame along $\gamma_{\tau}$, for every $\tau\in (-\epsilon,\epsilon)$. We then have $$\mathcal{A}^{-1}d\mathcal{A} = Qdu+\Lambda dt,$$ where $$Q=\left(
\begin{array}{ccccc}
0 & -m & v & 0 & 0 \\
v & 0 & 0 & 0 & m \\
0 & 0 & 0 & -^tp & 0 \\
0 & 0 & p & \psi & 0 \\
0 & v & 0 & 0 & 0 \\
\end{array}
\right),$$ and $$\Lambda=\left(
\begin{array}{ccccc}
\lambda_0^0 & -\lambda_{n+1}^1 & \lambda^2_{n+1} & ^t\lambda_{n+1} & 0 \\
\lambda^1_0 & 0 & \lambda^2_1 & ^t\lambda_1 & \lambda^1_{n+1} \\
\lambda^2_0 & \lambda^2_1 & 0 & -^t\lambda_2 & \lambda^2_{n+1} \\
\lambda_0 & \lambda_1 & \lambda_2 & \lambda & \lambda_{n+1} \\
0 & -\lambda^1_0 & \lambda^2_0 & \lambda_0 & -\lambda^0_0 \\
\end{array}
\right).$$ The entries $$\begin{split}
& v,m,\lambda^0_0,\lambda^1_0,\lambda^2_0,\lambda_{n+1}^1,\lambda^2_{n+1}
:I\times (-\epsilon,\epsilon)\to \R,\\
& p, \lambda_0,\lambda_2,\lambda_2,\lambda_{n+1}:I\times (-\epsilon,\epsilon)\to \R^{n-2},\\
& \psi,\lambda :I\times (-\epsilon,\epsilon)\to \mathfrak{o}(n-2).
\end{split}$$ of $Q$ and $\Lambda$ are smooth maps. Their restrictions to $I\cong I\times \{0\}$ are denoted with the same symbols with an over-bar. Note that $\mathrm{supp}(\overline{\Lambda})\subseteq K$. The cross section $$V_{\mathbf{g}}=\sum_{a=2}^{n}\overline{\lambda}^a_0 A_a:I\to \mathcal{N}^2(\gamma)$$ does not depend on the choice of $A$ and $\mathcal{A}$. We call $V_{\mathbf{g}}$ the [*infinitesimal variation*]{} of $\mathbf{g}$. From $\dot{A}=A\cdot \overline{Q}$ and taking into account (\[LS\]), we obtain $\overline{\upsilon}=1$, $\overline{m}_1=h_1$, $\overline{p}=s$ and $\overline{\psi}=\phi$. From the Maurer-Cartan equations we have $$\partial_u\Lambda - \partial_tQ = \Lambda\cdot Q - Q\cdot \Lambda.$$ This implies $$\begin{split}
&(\partial_t v)|_I = \overline{\lambda}_0^0 + \dot{\overline{\lambda}}^{1}_{0},\\
&\overline{\lambda}_0^0 =\frac{1}{2}(\dot{\overline{\lambda}}^0_{2}-\dot{\overline{\lambda}}^1_{0}-
\overline{\lambda}_{n+1}\cdot s-h_1\overline{\lambda}^2_1),\\
&\dot{\overline{\lambda}}_{0}=\overline{\lambda}_1+\overline{\lambda}_0^2s-\phi \cdot \overline{\lambda}_0,\\
&\dot{\overline{\lambda}}_1=-h_1\overline{\lambda}_0-\overline{\lambda}_{n+1}-\overline{\lambda}^2_1s-\phi\cdot\overline{\lambda}_1,\\
& \dot{\overline{\lambda}}_{n+1}=h_1\overline{\lambda}_1+\overline{\lambda}_2-\overline{\lambda}^2_0s-\phi\cdot \overline{\lambda}_{n+1},\\
&\dot{\overline{\lambda}}_0^2=\overline{\lambda}^2_1-s\cdot \overline{\lambda}_0.
\end{split}$$ From these equalities, integrating by parts and taking into account that $\mathrm{supp}(\overline{\lambda}^i_j)\subset K$, $i,j=0,\dots, n+1$, we have $$\begin{split}
&\frac{d}{dt}\left(\mathcal{S}_K(\gamma_t)\right)|_{t=0}=\int_K (\partial_t v|_I)du =
\int_K (\overline{\lambda}_0^0+\dot{\overline{\lambda}}^1_0)du=\int_K \overline{\lambda}_0^0du=\\
&= -\frac{1}{2}\int_K (s\cdot(-\dot{\overline{\lambda}}_1-
h_1\overline{\lambda}_0-\overline{\lambda}_1^2 s-\phi\cdot \overline{\lambda}_1 )+h_1(\dot{\overline{\lambda}}_0^2+s\cdot \overline{\lambda}_0))du=\\
&=\frac{1}{2}\int_K(-\dot{s}\cdot\overline{\lambda}_1+h_1s\cdot \overline{\lambda}_0+(s\cdot s)\overline{\lambda}^2_1+s\cdot \phi\cdot \overline{\lambda}_1+\dot{h}_1\overline{\lambda}^2_0+h_1s\cdot\overline{\lambda}_0)du=\\
&=-\frac{1}{2}\int_K(\dot{s}\cdot(\dot{\overline{\lambda}}_0+\overline{\lambda}_0^2s+\phi\cdot \overline{\lambda}_0))du+\frac{1}{2}\int_K(s\cdot \phi\cdot (\dot{\overline{\lambda}}_0+\overline{\lambda}^2_0 s+\phi\cdot \overline{\lambda}_0))du\\
&+\frac{1}{2}\int_K(k_1s\cdot \overline{\lambda}_0+(^ts\cdot s)(\dot{\overline{\lambda}}^2_0-s\cdot \overline{\lambda}_0)+\dot{h}_1\overline{\lambda}^2_0+h_1s\cdot \overline{\lambda}_0)du=\\
&=\frac{1}{2}\int_K(\overline{\lambda}^2_0(\dot{h}_1-3 ^ts\cdot \dot{s})+
\overline{\lambda}_0\cdot (\dot{s}_{(1)}+\phi\cdot s_{(1)}-((^ts\cdot s)-2h_1)s))du=\\
&=\frac{1}{2}\int_K(\overline{\lambda}^2_0(\dot{h}_1-3 ^ts\cdot \dot{s})+
\overline{\lambda}_0\cdot (s_{(2)}+((2h_1-^ts\cdot s))s))du= \frac{1}{2}\int_K \langle V_{\mathbf{g}},W_{\gamma}\rangle du.
\end{split}$$ This proves the result.
Next we show that for each $u_0\in I$ there exist an open interval $J\subset I$ containing $u_0$ such that for every smooth function $\rho:I\to \R$ with compact support $K\subset J$ and every $j=2,\dots ,n$, there exist a compactly supported variation $\mathbf{g}$ such that $V_{\mathbf{g}}=\rho A_j$. This clearly implies that a conformal world-line satisfies $W_{\gamma}=0$.
Using the conformal invariance of the functional we may suppose, without loss of generality, that $\gamma(u)$ belongs to the Minkowski chamber, for every $u$ lying in an open interval $J\subset I$ containing $u_0$. Then, $\gamma|_J=\mathbf{j}\circ \alpha$ where $\alpha:J\to \mathbb{M}^{1,n-1}$ is a time-like curve of the Minkowski space. Let $\mathbf{t}:J\to \mathbb{M}^{1,n-1}$ be the future-directed time-like unit tangent vector along $\alpha$ and $$N(\alpha)=\{(u,\mathbf{v})\in J\times \mathbb{M}^{1,n-1} : ^{*}\mathbf{x}\cdot \mathbf{t}|_u=0\}$$ be the normal bundle of $\alpha$, equipped with the metric covariant derivative[^4] $\nabla(\mathbf{v})=pr(\dot{\mathbf{v}})$. Let $(\mathbf{b}_2,\dots,\mathbf{b}_n)$ be a flat orthogonal trivialization of $N(\alpha)$ such that $(\mathbf{t},\mathbf{b}_2,\dots,\mathbf{b}_n)$ is positive-oriented. Then $(\alpha,\mathbf{t},\mathbf{b}_2,\dots, \mathbf{b}_n):J\to P^{\uparrow}_+(1,n-1)$ is a lift of $\alpha$ to the restricted Poincaré group and[^5] $$F=\mathbf{J}\circ (\alpha,\mathbf{t},\mathbf{b}_2,\dots, \mathbf{b}_n):J \to \mathrm{A}^{\uparrow}_+(2,n)$$ is a first-order frame field along $\gamma|_J$. Let $A:J\to \mathrm{A}^{\uparrow}_+(2,n)$ be a second order frame along $\gamma$. We then have $A=F\cdot X(r,x,y,R)$, where $$x,r:J\to \R, r>0,\quad y : J\to \R^{n-1},\quad R=(R^i_j)_{i,j=2,\dots n} :J\to \mathrm{SO(n-1)}$$ are smooth maps and $X(r,x,y,R)$ is as in (\[gauge\]). We put $$m=^t(m^2,\dots,m^n)=\frac{\rho}{r}^t(R^2_j,\dots, R^n_j).$$ It is now an easy matter to check that the variation $$\begin{split}
\gamma_{\tau}(u) &= \mathbf{j}\circ(\alpha|_u+ \tau\sum_{j=2}^n m^j(u)\mathbf{b}_j|_u), \quad
(u,\tau)\in J\times (-\epsilon,\epsilon),\\
\gamma_{\tau}(u) &=\gamma(u),\quad (u,\tau)\in (I-J)\times (-\epsilon,\epsilon),
\end{split}$$ satisfies the required properties. This concludes the proof of the Theorem.
The next theorem shows that conformal world-lines lie in an Einstein universe of dimension 2,3 or 4.
\[ThmB\][The trajectory of a conformal world-line is contained in a $m$-dimensional Lorentzian cycle, with $2\le m \le 4$.]{}
Let $\gamma$ be a conformal world-line parameterized by conformal parameter. We consider a second-order frame field $A=(A_0,\dots A_{n+1})$ along $\gamma$. Then $$\mathcal{T}^5(\gamma)|_u=\mathrm{Span}(A_0|_u,A_1|_u,A_2|_u,S_{\gamma}|_u,DS_{\gamma}|_u,A_{n+1}|_u),$$ for every $u\in I$. From (\[LS\]) we have $$\label{STR1}\begin{split}
\dot{A}_0&=A_1,\quad \dot{A}_1=-h_1A_0-A_{n+1},\\
\dot{A}_2&=A_0+S_{\gamma},\quad \dot{A}_{n+1}=-h_1A_1+A_2,
\end{split}$$ and $$\label{STR2}
\dot{A}_j=-s_jA_2+\sum_{k=3}^{n}\phi_j^kA_k,\quad j=3,\dots, n.$$ The first equation in (\[EL1\]) implies $$\label{EL1Bis}(h_2)^2=\frac{2}{3}h_1+c_1,\quad c_1\in \R.$$ The remaining equations in (\[EL1\]) can be written as $$\label{EL2}
\dot{s}=-\phi\cdot s+s_{(1)},\quad \dot{s}_{(1)}=-(\frac{4}{3}h_1-c_1)s-\phi\cdot s_{(1)}$$ or, equivalently, in the form $$\label{EL3}\begin{cases}
\frac{d}{du}(S_{\gamma})=DS_{\gamma}-(\frac{2}{3}h_1+c_1)A_2,\\ \frac{d}{du}(DS|_{\gamma})=-(\frac{4}{3}h_1-c_1)S_{\gamma}-\frac{1}{3}\dot{h}_1A_2.
\end{cases}$$ From (\[EL2\]) we see that $s\wedge s_{(1)}$ is a solution of the linear system $$\frac{d}{du}(s\wedge s_{(1)}) = -(\phi\cdot s)\wedge s_{(1)}- s\wedge (\phi\cdot s_{(1)}).$$ Hence, two possibilities may occur :
- Case I : $s|_u\wedge s_{(1)}|_u\neq 0$, for every $u\in I$
- Case II : $s\wedge s_{(1)}$ vanishes identically.
[*Case I*]{}. The vectors fields $S_{\gamma}$ and $DS_{\gamma}$ are everywhere linearly independent. Then, the osculating spaces $\mathcal{T}^5(\gamma)|_u$ are six-dimensional. Using (\[STR1\]) and (\[EL3\]) we obtain $$\frac{d}{du}(A_0\wedge A_1\wedge A_2\wedge S_{\gamma}\wedge DS_{\gamma}\wedge A)=0.$$ Hence, $\mathcal{T}^5(\gamma)|_u$ coincides with a fixed $6$-dimensional Lorentzian subspace $\mathbb{W}_{\gamma}\subseteq \R^{2,n}$, for each $u\in I$. This implies that the trajectory of $\gamma$ is contained in the $4$-dimensional Lorentzian cycle $\mathcal{E}(\mathbb{W}_{\gamma})$.
[*Case II*]{}. Since $(s,s_{(1)})$ is a solution of the linear system (\[EL2\]), there are two possibilities :
- Case II.1 : $s=s_{(1)}=0$;
- Case II.2 : $s^t\cdot s+^ts_{(1)}\cdot s_{(1)}>0$.
[*Case II.1*]{}. If $s=\dot{s}=0$, then (\[STR1\]) implies $$\frac{d}{du}(A_0\wedge A_1\wedge A_2\wedge A_{n+1})=0.$$ Therefore, $\mathcal{T}^3(\gamma)|_u$ coincides with a fixed $4$-dimensional Lorentzian subspace $\mathbb{W}_{\gamma}\subseteq \R^{2,n}$, for every $u\in I$. Hence, the trajectory of $\gamma$ is contained in the $2$-dimensional Lorentzain cycle $\mathcal{E}(\mathbb{W}_{\gamma})$.
[*Case II.2* ]{} If $s^t\cdot s+^ts_{(1)}\cdot s_{(1)}>0$ and $s\wedge s_{(1)}=0$, then $\mathrm{Span}(S_{\gamma}|_u,DS_{\gamma}|_u)$ is a $1$-dimensional space-like subspace, for every $u\in I$. Thus $$\mathcal{P}=\{(u,V)\in I\times \R^{2,n}: V\in \mathrm{Span}(S_{\gamma}|_u,DS_{\gamma}|_u)\}$$ is a rank $1$ vector bundle. Let $P$ be a unit-length cross section of $\mathcal{P}$. The identity $s\wedge s_{(1)}=0$ implies that $S_{\gamma}$ and $DS_{\gamma}$ are both proportional to $P$. We put $\widetilde{I}=I-I_*$, where $I_*$ is the discrete set $\{u\in I : S_{\gamma}|_u=0\}$. On $\widetilde{I}$ the fifth and fourth-order osculating bundles are spanned by $A_0,A_1,A_2,P,A_{n+1}$. This implies the existence of a smooth function $f:\widetilde{I}\to \R$ such that $$\dot{P}|_{\widetilde{I}}=fP|_{\widetilde{I}},\quad \mathrm{mod}(A_0,A_1,A_2,A_{n+1}),$$ From (\[STR1\]) and taking into account the previous identity we get $$\label{CC}\frac{d}{du}(A_0\wedge A_1\wedge A_2\wedge P\wedge A_{n+1})|_{\widetilde{I}}= f( A_0\wedge A_1\wedge A_2\wedge P\wedge A_{n+1})|_{\widetilde{I}}.$$ Let $\mathrm{Gr}_{5}(\R^{n+2})$ be the Grassmannian of the $5$-dimensional vector subspace of $\R^{n+2}$. From (\[CC\]) it follows that the map $$u\in I\to \mathrm{Span}(A_0|_u\wedge A_1|_u\wedge A_2|_u\wedge P|_u\wedge A_{n+1}|_u)\in \mathrm{Gr}_{5}(\R^{n+2})$$ is constant on $\widetilde{I}$. By continuity is constant on $I$. Then, the fourth-order osculating spaces coincide with a fixed $5$-dimensional Lorentzian subspace $\mathbb{W}_{\gamma}\subseteq \R^{2,n}$. Hence, the trajectory of $\gamma$ is contained in the $3$-dimensional Lorentzain cycle $\mathcal{E}(\mathbb{W}_{\gamma})$.
This theorem shows that if $\gamma$ is a conformal world-line then three possibilities may occur :
- $\gamma$ is trapped in a $4$-dimensional Einstein universe and there are no 3-dimensional Lorentzian cycles containing the trajectory of $\gamma$. If this occurs, we say that $\gamma$ is a [*linearly full conformal world-line*]{}.
- the trajectory is trapped in a $3$-dimensional Lorentzian cycle but does not lie in any $2$-dimensional Lorentzian cycle.
- the trajectory is trapped in a $2$-dimensional Lorentzian cycle.
Linearly full conformal world-lines {#s3}
===================================
From now on we limit ourselves to consider linearly full conformal world-lines of the four-dimensional Einstein universe. We also suppose that the parametrization is by the conformal parameter and that the orientation is the intrinsic one. These assumptions will be implicitly assumed throughout the subsequent discussions.
Canonical frames and conformal curvatures {#ss:conf-fr}
-----------------------------------------
\[propuniq\][Let $\gamma : I\to \mathcal{E}^{1,3}$ be a conformal world-line. Then, there exist a unique second-order conformal frame $B:I\to A^{\uparrow}_+(2,3)$ such that $$\label{mcc}\dot{B}=B\cdot \mathcal{K},$$ where $$\label{mcc2}\mathcal{K}=\left(
\begin{array}{cccccc}
0 & -k_1 & 1 & 0 & 0 & 0 \\
1 & 0 & 0 & 0 & 0 & k_1 \\
0 & 0 & 0 & -k_2 & 0 & 1 \\
0 & 0 & k_2 & 0 & -k_3 & 0 \\
0 & 0 & 0 & k_3 & 0 & 0 \\
0 & -1 & 0 & 0 & 0 & 0 \\
\end{array}
\right).$$ and $k_1,k_2,k_3:I\to \R$ are smooth functions with $k_2>0$ and $k_3\neq 0$. In addition, if $k_1$, $k_2$ and $k_3$ are non-constant then there exist $c_1,c_2,c_3\in \R$ with $c_3\neq 0$ such that $$\label{eqmrk4}
k_1=\frac{3}{2}k_2^2+c_1,\quad k_3=c_3k_2^{-2},\quad \dot{k}_2^2+k_2^4+c_3^2k_{2}^{-2}+2c_1k_2^2+c_2=0.$$]{}
Let $(M_0,M_1,M_2,M_{5})$ be the canonical trivialization of $\mathcal{T}^3(\gamma)$ and $S_{\gamma}:I\to \mathcal{N}^3(\gamma)$ be the cross section defined as in (\[SS1\]). We set $$B_3=S_{\gamma}/\sqrt{\langle S_{\gamma},S_{\gamma}\rangle}$$ and we pick the unique unit cross section $B_4:I\to \mathcal{N}^4(\gamma)$ such that $\mathrm{det}(M_0,M_1,M_2,M_3,B_4,M_5)>0$. Putting $$B_0=M_0,\quad B_1=M_1,\quad B_2=M_2,\quad B_5=M_5,$$ the map $B=(B_0,\dots, B_5)$ is a second-order frame along $\gamma$. In view of (\[LS\]) we have $$\label{FR1}\dot{B}_0=B_1,\quad \dot{B}_1=-k_1B_0-B_5,\quad \dot{B}_5=k_1B_1+B_2$$ where[^6] $k_1=h_1$. Since $D$ is a metric covariant derivative, $DB_3$ is a multiple of $B_4$ and hence $DB_3=k_3B_4$, for some non-zero function $k_3$. From (\[LS\]) we have $$\label{FR2}\dot{B}_2=B_0+k_2B_3,$$ where $k_2=\sqrt{\langle S_{\gamma},S_{\gamma}\rangle}>0$. Keeping in mind (\[EL3\]) we get $$\label{FR3}\dot{B}_3=-k_2B_2+k_3B_4.$$ Furthermore, since $B$ takes values in $\mathrm{A}^{\uparrow}_{+}(2,4)$ and using (\[FR1\])-(\[FR3\]) we obtain $$\label{FR4}\dot{B}_4=-k_3B_3.$$ Combining (\[FR1\])-(\[FR4\]) we deduce that $B$ satisfies (\[mcc\]).
Suppose that the functions $k_1,k_2$ and $k_3$ are non-constant. The first and second covariant derivatives of $S_{\gamma}$ can be written as $$DS_{\gamma}=\dot{k}_2B_3+k_2k_3B_4,\quad D^2S_{\gamma}=(\ddot{k}_2-k_2k_3^2)B_3+(2\dot{k}_2k_3+k_3\dot{k}_2)B_4.$$ Then, from variational equation $W_{\gamma}=0$ we get $$\label{EM2}
\ddot{k}_2=k_2(k_3^2+k_2^2-2k_1),\quad k_2\dot{k}_3+2k_3\dot{k}_2=0\quad \dot{k}_1=3k_2\dot{k}_2.$$ The third and the second equations in (\[EM2\]) imply the existence of two constants $c_1$ and $c_3\neq 0$ such that $$\label{EM3}k_1=\frac{3}{2}k_2^2+c_1,\quad k_3=c_3k_2^{-2}.$$ Substituting (\[EM3\]) into (\[EM2\]) we find $$\ddot{k}_2=-2k_2^{3}+c_3^2k^{-3}-4k_2-2c_1k_2.$$ Then there exist a constant $c_2$ such that $$\dot{k}_2^2+k_2^4+c_3^2k_{2}^{-2}+2c_1k_2^2+c_2=0.$$
[We call $B$ the [*canonical conformal frame*]{}. The functions $k_1,k_2,k_3$ are said the [*conformal curvatures*]{}. The constants $c_1,c_2,c_3$ are referred to as the [*characters*]{} of the world-line.]{}
\[remarkuniq\][Conversely, if $k_1,k_2,k_3:I\to \R$ are smooth functions satisfying (\[eqmrk4\]), by solving the linear system (\[mcc\]) with initial condition $\mathrm{B}(u_0)\in \mathrm{A}^{\uparrow}_+(2,4)$, we get a smooth map $\mathrm{B}:I\to \mathrm{A}^{\uparrow}_+(2,4)$. Then, $\gamma=[B_0]$ is a linearly full conformal world-line with conformal curvatures $k_1,k_2,k_3$ and canonical conformal frame $B$. Any other world-line $\widetilde{\gamma}$ with the same curvatures is congruent to $\gamma$ with respect to the restricted conformal group, ie there exist $\mathbf{X}\in \mathrm{A}^{\uparrow}_+(2,4)$ such that $\widetilde{\gamma}=\mathbf{X}\cdot \gamma$.]{}
[The sign ambiguity of the third curvature can be removed with the following reasoning : if the third conformal curvature of $\gamma$ is $k_3$ and if $\mathbf{X}$ is an orientation-preserving and time-reversing conformal transformation, then $\mathbf{X}\cdot \gamma(-u)$ is a conformal world-line with conformal curvatures $k_1(-u),k_2(-u)$ and $-k_3(-u)$. Therefore, up to a time-reversing conformal transformation, the curvature $k_3$ is positive.]{}
[The sign of the third curvature is said the [*conformal helicity*]{} of the world-line. By the previous remark we see that is not restrictive to consider conformal world-lines with positive helicity. From now on this additional assumption is implicitly assumed.]{}
[If $k_1,k_2$ and $k_3$ are constant, then the variational equations imply $k_1=(k_2^2+k_3^2)/2$. In this case $\mathcal{K}$ is a fixed element of the Lie algebra $\mathfrak{a}(2,4)$ and $\gamma$ is congruent to the orbit through $[^t(1,0,0,0,0)]$ of the $1$-parameter group of conformal transformations $u\in \R\to \mathrm{Exp}(u\mathcal{K})\in A^{\uparrow}_+(2,4)$. Therefore, from a theoretical point of view the determination of the conformal world-lines with constant curvatures is reduced to the calculation of the exponentials of the matrices $u\mathcal{K}$. From a computational point of view this is an elementary but not completely straightforward matter and requires a detailed analysis of the possible orbit types of the infinitesimal generator $\mathcal{K}$.]{}
Conformal curvatures in terms of Jacobi’s elliptic functions
------------------------------------------------------------
From now on we suppose that the conformal curvatures are non-constant.
[If $c_1,c_2,c_3$ are the characters of $\gamma$ then the third-order polynomial $Q_1(t)=t^3+2c_1t^2+c_2t+c_3^2$ has three distinct real roots $e_1,e_2,e_3$ such that $e_1<0<e_2<e_3$.]{}
[Let $e_1,e_2,e_3$ be the roots of $Q_1$. Then, $$\dot{k}_2^2+k_2^4+c_3^2k_{2}^{-2}+2c_1k_2^2+c_2=0$$ implies $$\label{eee1}
(k_2\dot{k}_2)^2=-(k_2^2-e_1)(k_2^2-e_2)(k_2^2-e_3)$$ and $$\label{eee2}
c_1=-\frac{1}{2}(e_1+e_2+e_3),\quad c_2=e_1e_2+e_1e_3+e_2e_3,\quad c_3^2=-e_1e_2e_3.$$ If two roots are complex conjugate each other, say $e_2$ and $e_3$, then the third equation of (\[eee2\]) implies $e_1<0$. Hence the right hand side of (\[eee1\]) is strictly negative. This contradicts the fact that $k_2$ is non constant. If $Q_1$ has a double root, say $e_2=e_3$, then the third equation of (\[eee2\]) implies $e_1<0$. So, as in the previous case, the right hand side of (\[eee1\]) is strictly negative. But this can’t occur. Therefore, the roots are real and distinct. We choose the ordering $e_1<e_2<e_3$. Using again the third equation of (\[eee2\]) we see that two possibilities may occur : either $e_1<e_2<e_3<0$ or else $e_1<0<e_2<e_3$. In the first case the right hand side of (\[eee1\]) is negative. But, as in the previous cases, this conclusion is impossible. Thus the Lemma is proved.]{}
[We say that $e_1,e_2,e_3$ are the [*phase parameters*]{} of the world-line. Note that, according to (\[eee2\]), the characters can be read off from the phase parameters.]{}
We put $$\label{ell}\ell_1=e_2\ell_3,\quad \ell_2=e_1\ell_4,\quad \ell_3=e_3-e_1,\quad \ell_4=e_3-e_2,\quad m=\ell_4/\ell_3,$$ and we denote by $K(m)$ and $\mathrm{sn}(-,m)$ the complete integral of the first kind and the Jacobi’s $sn$-function with parameter $m$[^7].
\[ConfCurv\][Let $\gamma$ be a conformal world-line with phase parameters $e_1<0<e_2<e_3$, then $$\label{k21}k_2(u)=\sqrt{\frac{\ell_1-\ell_2\mathrm{sn}^2(\sqrt{\ell_3}u+u_0,m)}
{\ell_3-\ell_4\mathrm{sn}^2(\sqrt{\ell_3}u+u_0,m)}},$$ and $$\label{k13}k_1=\frac{3}{2}k_2^2-\frac{1}{2}(e_1+e_2+e_3),\quad k_3= \sqrt{-e_1e_2e_3}k_2^{-2},$$ where $u_0$ is a constant. ]{}
[Put $f(u)=k_2(u/\sqrt{\ell_3})$ then, using (\[eee1\]), we get $$\dot{f}^2=-\frac{4}{e_3-e_1}(f-e_1)(f-e_2)(f-e_3).$$ If $e_1<0<e_2<e_3$, the general solution of the equation above (cfr. [@BF] pag. $77$) is $$f(u)=\frac{\ell_1-\ell_2\mathrm{sn}^2(u+u_0,m)}{\ell_3-\ell_4\mathrm{sn}^2(u+u_0,m)},$$ where $u_0$ is a constant. This implies (\[k21\]). We conclude the proof by observing that (\[k13\]) is an immediate consequence of (\[EM3\]) and (\[eee2\]). ]{}
This proposition has two consequences : $k_2$ is a strictly positive, even periodic function with period $$\label{period}\omega=2K(m)/\sqrt{\ell_3}$$ and the parameterizations by the conformal parameter of a world-line are defined on the whole real line. In addition, with a shift of the independent variable the constant $u_0$ in (\[k21\]) can be put equal to zero. This means that a conformal world line admits a [*canonical parametrization*]{} by conformal parameter such that $k_2(0)=\sqrt{\ell_1/\ell_3}$.
[We say that $\gamma$ is the [*standard configuration*]{} of a world-line if $\gamma$ is parameterized by the canonical parameter and if, in addition, its canonical frame satisfies the initial condition $B|_0=\mathrm{Id}|_{6\times 6}$. Clearly, each world-line is conformally equivalent to a unique standard configuration.]{}
To summarize what has been said, we state the following Proposition :
\[PHRS\][The standard configurations of world-lines are in one-to-one correspondence with the point of the domain $$\mathcal{F}=\{\mathbf{e}=(e_1,e_2,e_3)\in \R^3 : e_1<0<e_2<e_3\}\subset \R^3.$$ In other words, for each $\mathbf{e}=(e_1,e_2,e_3)\in \mathcal{F}$ there exist a unique world-line $\gamma$ in its standard configuration with phase parameters $(e_1,e_3,e_3)$.]{}
The momentum operator
---------------------
Let $\gamma$ be a standard configuration of a conformal world-line with curvatures $k_1,k_2,k_3$. We denote by $\mathcal{H}: Y\to \mathfrak{a}(2,4)$ the map $$\label{obs}\mathcal{H}=
\left(
\begin{array}{cccccc}
0 & -1 & -(k_2^2-k_1) & \dot{k}_2 & k_2k_3 & 0 \\
0 & 0 & 0 & -k_2 & 0 & 1 \\
-1 & 0 & 0 & 0 & 0 & -(k_2^2-k_1) \\
0 & -k_2 & 0 & 0 & 0 & \dot{k}_2 \\
0 & 0 & 0 & 0 & 0 & k_2k_3 \\
0 & 0 & -1 & 0 & 0 & 0 \\
\end{array}
\right).$$ Then, (\[k21\]) and (\[k13\]) imply $$\label{lax} \dot{\mathcal{H}}=[\mathcal{H},\mathcal{K}].$$ This equation together with $B'=B\cdot \mathcal{K}$ implies $$\label{momentum}B|_u\cdot \mathcal{H}|_u\cdot B^{-1}|_u = \mathcal{H}|_0,\quad \forall u\in \R.$$
[We put $\mathfrak{m}=\mathcal{H}|_0$ and we say that $\mathfrak{m}$ is the [*momentum operator*]{} of $\gamma$]{}
\[ThEx\][We give a brief explanation of the conceptual origin of the momentum operator. The first step is the construction of the [*momentum space*]{} and of the [*Euler-Lagrange exterior differential system*]{} [@Gr; @GM]. In our specific situation, the momentum space is the $19$-dimensional manifold $Z=A^{\uparrow}_+(2,4)\times \mathfrak{K}$, where $$\mathfrak{K}=\{\mathfrak{k}=(\mathfrak{k}_1,\mathfrak{k}_2,\mathfrak{k}_3,\dot{\mathfrak{k}}_2)\in \R^4, \mathfrak{k}_2>0, \mathfrak{k}_3>0\}\subset \R^4.$$ The restricted conformal group acts freely on the left of $Z$ by $$L_{Y}(X,\mathfrak{k})= (Y\cdot X,\mathfrak{k}),\quad \forall X,Y\in A^{\uparrow}_+(2,4), \forall \mathfrak{k}\in \mathfrak{K}.$$ The Euler-Lagrange differential system is the $A^{\uparrow}_+(2,4)$-invariant Pfaffian differential ideal $\mathcal{I}\subset \Omega^*(Z)$ generated by the $1$-forms $$\mu^2_0,\quad \mu^3_0,\quad \mu^4_0,\quad \mu^2_1,\quad \mu^3_1,\quad \mu^4_1,\quad \mu^3_5,\quad \mu^4_5,\quad \mu^4_2,\quad \mu^0_0$$ $$\mu^2_5-\mu^1_0,\quad \mu^1_5-\mathfrak{k}_1\mu^1_0,\quad \mu^3_2-\mathfrak{k}_2\mu^1_0,\quad
\mu^4_3-\mathfrak{k}_3\mu^1_0$$ and $$d\dot{\mathfrak{k}}_2-\mathfrak{k}_2(\mathfrak{k}_3^2+\mathfrak{k}_2^2-2\mathfrak{k}_1)\mu^1,\quad
d\mathfrak{k}_2-\dot{\mathfrak{k}}_2\mu^1_0,\quad d\mathfrak{k}_1-3\mathfrak{k}_2\dot{k}_2\mu^1_0,\quad
\mathfrak{k}_2d\mathfrak{k}_3+2\mathfrak{k}_2\dot{\mathfrak{k}}_2\mu^1_0.$$ The independence condition of the system is the invariant $1$-form $\mu^1_0$. The integral curves of $(\mathcal{I},\mu^1_0)$ can be build as follows : let $\gamma$ be a conformal world-line[^8] with curvatures $k_1,k_2,k_3$ and canonical frame $B$, the lifting $\mathfrak{b}=(B,k_1,k_2,k_3,\dot{k}_2):\R\to Z$ of $\gamma$ to $Z$ is said the [*extended frame*]{} along $\gamma$. Proposition \[propuniq\] tell us that the integral curves of $(\mathcal{I},\mu^1_0)$ are the extended frames of the world-lines. Using the Maurer-Cartan equations (\[MC\]) we see that $$\label{Liouville}\zeta = \frac{1}{2}(\mu^1_0+\mu^2_5+(\mathfrak{k}_2^2-\mathfrak{k}_1)\mu^2_0-\dot{\mathfrak{k}}_2\mu^3_0-
\mathfrak{k}_2\mathfrak{k}_3\mu^4_0+\mathfrak{k}_2\mu^3_1)$$ is an invariant contact $1$-form such that the integral curves of its characteristic vector field[^9] $X_{\zeta}$ are the extended frames of the world-lines. We can think of $\zeta$ as a map into the dual space $\mathfrak{a}(2,4)^*$ of the conformal Lie-algebra. Denote by $\natural:\mathfrak{a}^*(4,2)\to \mathfrak{a}(2,4)$ the pairing defined by the Killing form, then $\zeta^{\natural}=\natural \circ \zeta : Z\to \mathfrak{a}(2,4)$ is an equivariant map that, in our context, plays the role of the Legendre transformation. If $\mathfrak{b}$ is the extended frame of $\gamma$, then $\mathcal{H}$ coincides with $\zeta^{\natural}\circ \mathfrak{b}$. The momentum map of action of $A^{\uparrow}_+(2,3)$ on $(Z,\zeta)$ is given by $$\widehat{\mathfrak{m}}:(X,\mathfrak{k})\in Z\to ad^*_X(\zeta|_{(B,\mathfrak{k})})\in \mathfrak{a}^*(2,4).$$ The [*Nöther conservations theorem*]{} for a Hamiltonian action on a contact manifold says that $\widehat{\mathfrak{m}}$ is constant along the characteristic curves. Thus, if $\mathfrak{b}$ is the extended frame of a world-line $\gamma$ then $\widehat{\mathfrak{m}}_{\gamma}=\widehat{\mathfrak{m}}\circ \mathfrak{b}$ is constant and $\widehat{\mathfrak{m}}_{\gamma}^{\natural}$ is the momentum operator of $\gamma$. This explains the geometrical origin of the momentum operator of a world-line.]{}
Hereafter we will adopt the following notations:
- $P_{\mathfrak{m}}$ is the characteristic polynomial of $\mathfrak{m}$;
- $\mathcal{S}_{\mathfrak{m}}$ is the spectrum of $\mathfrak{m}$, viewed as an endomorphism of $\C^6$;
- for each $\lambda\in \mathcal{S}_{\mathfrak{m}}$, $n_1(\lambda)$ is the multiplicity of $\lambda$ as root of $P_\mathfrak{m}$ and $n_2(\lambda)$ is the complex dimension of the $\mathfrak{m}$-eigenspace $\mathbb{V}_{\lambda}$ of $\lambda$.
[A conformal world-line is said to be [*regular, exceptional or singular*]{} depending on whether $\mathfrak{m}$ is a regular, exceptional or singular element of the Lie algebra $\mathfrak{a}(2,4)$. In other words, $\gamma$ is regular if $\mathcal{S}_{\mathfrak{m}}$ consists of six elements, is exceptional if $\mathcal{S}_{\mathfrak{m}}$ has less that six elements and $n_2(\lambda)=1$, for every $\lambda \in \mathcal{S}_{\mathfrak{m}}$ and is singular if $n_2(\lambda)>1$, for some $\lambda$.]{}
From (\[momentum\]) it follows that $P_\mathfrak{m}$ coincides with the characteristic polynomial of $\mathcal{H}|_u$, for every $u\in \R$. Keeping in mind (\[eqmrk4\]) we get $$\label{chpolm}P_\mathfrak{m}(t)=t^6+2c_1t^4+(c_2+1)t^2+c_3^2 = Q_1(t^2)+t^2.$$ Let $Q_2(t)$ be the third-order polynomial $Q_1(t)+t^2$. Since the roots of $Q_1(t)$ are the phase parameters $e_1<0<e_2<e_3$, we infer that three possibilities may occur :
- $Q_2$ has three distinct real roots $\rho_1,\rho_2,\rho_3$ such that $e_1<\rho_1<0<e_2<\rho_2<\rho_3<e_3$.
- $Q_2$ has one negative real root $\rho_1$ with $e_1<\rho_1$ and two complex conjugate roots $\rho_2=\mu+i\nu$, $\nu>0$ and $\rho_3=\mu-i\nu$.
- $Q_2$ has one simple real root $\rho_1$ such that $e_1<\rho_1<0$ and a double real root $\rho_2=\rho_3$, $e_2<\rho<e_3$.
We put[^10] $$\label{eigenvl}\begin{split}
\lambda_0 &=i\sqrt{|\rho_1|},\quad \lambda_1=-i\sqrt{|\rho_1|},\quad \lambda _2=\sqrt{\rho_2},\\ \lambda_3&=-\sqrt{\rho_2},\quad \lambda_4=\sqrt{\rho_3},\quad \lambda_5=-\sqrt{\rho_3}.
\end{split}$$ In the first two cases the eigenvalues of $P_\mathfrak{m}$ are simple and $$\label{sptr1}\mathcal{S}_{\mathfrak{m}}=\{\lambda_0,\dots, \lambda_5\}.$$ While, in the third case we have $$\label{sptr1}\mathcal{S}_{\mathfrak{m}}=\{\lambda_0,\dots, \lambda_3\}$$ and the two real eigenvalues $\lambda_2$ and $\lambda_3$ are the double roots of $P_{\mathfrak{m}}$.
\[CT\][A conformal world-line is either regular or exceptional.]{}
For every $\lambda\in \mathcal{S}_{\mathfrak{m}}$ we define $L_{\lambda}=(L_{\lambda}^0,\dots L_{\lambda}^5):\R\to \C^6$ by $$\label{L}\begin{cases}
L^0_{\lambda} =\lambda(\lambda^2-k_2^2)(\lambda^2+k_1-k_2^2),\\
L^1_{\lambda}=\lambda(\lambda-k_2\dot{k}_2),\\
L_{\lambda}^2=-\lambda^2(\lambda^2-k_2^2),\\
L^3_{\lambda}=\lambda(\lambda \dot{k}_2-k_2),\\
L^4_{\lambda}=k_2k_3(\lambda^2-k_2^2),\\
L_{\lambda}^5=\lambda(\lambda^2-k_2^2).
\end{cases}$$
The map $L_{\lambda}$ is real-analytic and periodic, with period $\omega$. Let $D_{\lambda}$ be its zero set. If $\lambda\notin \R$ then $D_{\lambda}=\emptyset$ and, if $\lambda\in \R$, we have $$\label{zeroes}
D_{\lambda}=\begin{cases} D^+_{\lambda}=\{n\omega+p_{\lambda}, n\in \mathbb{Z}\},\quad \lambda >0,\\
D^-_{\lambda}=\{n\omega-p_{\lambda}, n\in \mathbb{Z}\},\quad \lambda <0,
\end{cases}$$ where $p_{\lambda}\in (0,\omega)$ is given by $$\label{palpha}
p_{\lambda}=\mathrm{sn}^{-1}\left(\frac{1}{\alpha},m\right),\quad \alpha=\sqrt{\frac{\ell_2-\lambda^2\ell_4}{\ell_1-\lambda^2\ell_3}}.$$ The $\lambda$-eigenspace $\mathcal{V}_{\lambda}|_u$ of $\mathcal{H}|_u$ is spanned by $L_{\lambda}|_u$, for every $u\notin D_{\lambda}$. Since $D_{\lambda}$ is a discrete set, this implies that $\mathrm{dim}(\mathcal{V}_j|_u)=1$, for every $u\in \R$. On the other hand, $\mathfrak{m}$ and $\mathcal{H}|_u$ belongs to the same adjoint orbit, for every $u$. Thus, also the eigenspaces of the momentum are $1$-dimensional. This yields the result.
![The graphs of the functions $s_{\lambda}$ (on the left) and $r_{\lambda}$ (on the right), $\lambda\in \R$.[]{data-label="FIG1"}](FIG1I.pdf "fig:"){height="5cm" width="6cm"} ![The graphs of the functions $s_{\lambda}$ (on the left) and $r_{\lambda}$ (on the right), $\lambda\in \R$.[]{data-label="FIG1"}](FIG1II.pdf "fig:"){height="5cm" width="6cm"}
Integrating factors and principal vectors
-----------------------------------------
Let $\gamma$ be the standard configuration of a linearly full conformal string with non-constant curvatures. Denote by $e_1<0<e_2<e_3$ its phase parameters and by $B:\R\to A^{\uparrow}_+(2,4)$ its canonical frame. For each $\lambda\in \mathcal{S}_{\mathfrak{m}}$ we consider the functions $$\label{if2}r_{\lambda}=\frac{k_2\dot{k}_2+\lambda}{k_2^2-\lambda^2},\quad
s_{\lambda}=\frac{\lambda^2+k_2^2-2\lambda k_2\dot{k}_2}{(\lambda^2-k_2^2)^2}.$$ If $\lambda\notin \R$, the functions $r_{\lambda}$ and $s_{\lambda}$ are periodic, complex-valued and real-analytic; if $\lambda\in \R$, $r_{\lambda}$ and $s_{\lambda}$ are periodic, real-valued and real-analytic on the complement of the discrete set $\widetilde{D}_{\lambda}=D^+_{\lambda}\cup D^-_{\lambda}$; their absolute values tend to infinity when $u$ approaches one of the points of $\widetilde{D}_{\lambda}$ (see Figure \[FIG1\]). For notational consistency, we put $\widetilde{D}_{\lambda}=\emptyset$ when $\lambda\notin \R$.
A primitive $\delta_{\lambda}:\R-\widetilde{D}_{\lambda}\to \C$ of $r_{\lambda}$ is said an [*integrating factor of the first kind*]{} for the eigenvalue $\lambda$ if
- $\delta_{\lambda}|_0=0$;
- $e^{-\delta_{\lambda}}L_{\lambda}:\R-\widetilde{D}_{\lambda}\to \C^6$ extends to a real-analytic map $\R\to \C^6$.
![The graphs of the functions $\|L_{\lambda}\|^2$ (on the left) and $e^{-\delta_{\lambda}}$ (on the right), $\lambda\in \R$.[]{data-label="FIG2"}](FIG2I.pdf "fig:"){height="5cm" width="6cm"} ![The graphs of the functions $\|L_{\lambda}\|^2$ (on the left) and $e^{-\delta_{\lambda}}$ (on the right), $\lambda\in \R$.[]{data-label="FIG2"}](FIG2II.pdf "fig:"){height="5cm" width="6cm"}
[The integrating factors of the first kind are quasi-periodic functions, with quasi-period $2\omega$. If $\lambda\notin \R$, the function $\delta_{\lambda}$ is a regular, complex-valued function. If $\lambda\in \R$, the integrating factor $\delta_{\lambda}$ is real-analytic on the complement of the discrete set $D_{\lambda}$ and its imaginary part is locally constant. The function $e^{\delta_{\lambda}}$ is real-valued, with singularities at the points of $D_{\lambda}$ (see Figure \[FIG2\]). The singularities of $e^{-\delta_{\lambda}}$ compensate the zeroes of the functions $L^j_{\lambda}$ so that the products $e^{-\delta_{\lambda}}L^j_{\lambda}$ are regular, real-analytic maps (see Figure \[FIG3\]). The evaluation of the integrating factors in terms of elliptic integrals and theta functions is analyzed in the appendix. The explicit expression of the integrating factor of the first kind for a non-real eigenvalue is given in (\[ifc\]) while the integrating factor of a real eigenvalue can be found in (\[IIFR\]).]{}
![The graphs of the functions $e^{-\delta_{\lambda}}L^0_{\lambda}$ (on the left) and $e^{-\delta_{\lambda}}L^3_{\lambda}$ (on the right), $\lambda\in \R$.[]{data-label="FIG3"}](FIG3I.pdf "fig:"){height="5cm" width="6cm"} ![The graphs of the functions $e^{-\delta_{\lambda}}L^0_{\lambda}$ (on the left) and $e^{-\delta_{\lambda}}L^3_{\lambda}$ (on the right), $\lambda\in \R$.[]{data-label="FIG3"}](FIG3II.pdf "fig:"){height="5cm" width="6cm"}
Let $\lambda$ be a multiple root of $P_\mathfrak{m}$. We set $$\widehat{D}_{\lambda}=\{-\mathrm{sign}(\lambda)p+n\omega : n\in \mathbb{Z}\}$$ and we define $T_{\lambda}=(T_{\lambda}^0,\dots T_{\lambda}^5):\R-\widehat{D}_{\lambda}\to \R^6$ by $$\label{T}
\begin{cases}
T^0_{\lambda}= \frac{1}{2}(\lambda^2-k_2^2)(6\lambda^2+2c_1+k_2^2),\\ T^1_{\lambda}=\frac{k_2(k_2^2\dot{k}_2+\lambda^2\dot{k}_2-2\lambda k_2)}{\lambda^2-k_2^2},\\
T^2_{\lambda}=-2\lambda(\lambda^2-k_2^2),\\
T^3_{\lambda}=\frac{k_2(\lambda^2+k_2^2-2\lambda k_2 \dot{k}_2)}{\lambda^2-k_2^2},\\
T^4_{\lambda}=0,\\
T_{\lambda}^5=(\lambda^2-k_2^2).
\end{cases}$$
[The map $T_{\lambda}$ is periodic with period $\omega$, is real-analytic on the complement of $\widehat{D}_{\lambda}$ and tends to $\pm \infty$ when $u$ tends to a point of $\widehat{D}_{\lambda}$. It vanishes at the point of $D_{\lambda}$ (see Figure \[FIG4\])]{}
![The graphs of the functions $\|T_{\lambda}\|^2$ (on the left) and $\eta_{\lambda}$ (on the right) when $\lambda\in \R$ is a multiple root of $P_\mathfrak{m}$.[]{data-label="FIG4"}](FIG4I.pdf "fig:"){height="5cm" width="6cm"} ![The graphs of the functions $\|T_{\lambda}\|^2$ (on the left) and $\eta_{\lambda}$ (on the right) when $\lambda\in \R$ is a multiple root of $P_\mathfrak{m}$.[]{data-label="FIG4"}](FIG4II.pdf "fig:"){height="5cm" width="6cm"}
A primitive $\eta_{\lambda}:\R-\widetilde{D}_{\lambda}\to \R$ of $s_{\lambda}$ is said an [*integrating factor of the second kind*]{} for the multiple eigenvalue $\lambda$ if
- $\eta_{\lambda}|_0=0$;
- $e^{-\delta_{\lambda}}(T_{\lambda}-\eta_{\lambda}L_{\lambda}):\R-\widetilde{D}_{\lambda}\to \C^6$ extends to a real-analytic map $\R\to \C^6$.
[The integrating factor of the second kind vanishes at the points of $D_{\lambda}$ and tends to $\pm \infty$ when $u$ tends to a point of $\widehat{D}_{\lambda}$ (see Figure \[FIG4\]). The functions $\eta_{\lambda}L_{\lambda}$ and $T_{\lambda}-\eta_{\lambda}L_{\lambda}$ behave in a similar way. Multiplying $T_{\lambda}-\eta_{\lambda}L_{\lambda}$ with $e^{-\delta_{\lambda}}$, the zeroes of one factor compensate the singularities of the other so that the product is a regular analytic function (see Figures \[FIG5\]). The formula expressing the integrating factor of the second kind is given in (\[IFIIK\]). Despite the apparent opacity, the formulas of the integrating factors can be easily made operative using standard programs of symbolic computation such as [*Mathematica 11*]{}.]{}
![The graphs of the functions $e^{-\delta_{\lambda}}(T^0_{\lambda}-\eta_{\lambda}L^0_{\lambda})$ (on the left) and $e^{-\delta_{\lambda}}(T^3_{\lambda}-\eta_{\lambda}L^3_{\lambda})$ (on the right) when $\lambda\in \R$ is a multiple root of $P_\mathfrak{m}$.[]{data-label="FIG5"}](FIG5I.pdf "fig:"){height="5cm" width="6cm"} ![The graphs of the functions $e^{-\delta_{\lambda}}(T^0_{\lambda}-\eta_{\lambda}L^0_{\lambda})$ (on the left) and $e^{-\delta_{\lambda}}(T^3_{\lambda}-\eta_{\lambda}L^3_{\lambda})$ (on the right) when $\lambda\in \R$ is a multiple root of $P_\mathfrak{m}$.[]{data-label="FIG5"}](FIG5II.pdf "fig:"){height="5cm" width="6cm"}
\[PV\][Let $\lambda\in \mathcal{S}_{\mathfrak{m}}$ be an eigenvalue of the momentum, then $$\label{prax1}e^{-\delta_{\lambda}}\sum_{j=0}^{5}L^j_{\lambda}B_j=\mathbf{A}_{\lambda},$$ where $\mathbf{A}_{\lambda}\in \C^6$ is an $\mathfrak{m}$-eigenvector of the eigenvalue $\lambda$. If $\lambda$ is a multiple root of $P_{\mathfrak{m}}$, then $$\label{prax2}e^{-\delta_{\lambda}}\sum_{j=0}^{5}(T^j_{\lambda}-\eta_{\lambda}L^j_{\lambda})B_j
=\mathbf{C}_{\lambda},$$ where $\mathbf{C}_{\lambda}\in \C^6$ is a non-zero vector such that $$\mathbf{C}_{\lambda}\wedge \mathbf{A}_{\lambda}\neq 0,\quad \mathfrak{m}(\mathbf{C}_{\lambda})=\lambda \mathbf{C}_{\lambda}+\mathbf{A}_{\lambda}.$$ We call $\mathbf{A}_{\lambda}$ the [principal vector]{} of the eigenvalue $\lambda$ and $\mathbf{C}_{\lambda}$ the [secondary principal vector]{} of the multiple eigenvalue $\lambda$.]{}
\[PFIF\]
Denote by $\mathbf{L}_{\lambda}:\R\to \C^6$ the map $$\label{eigen}\mathbf{L}_{\lambda}=\sum_{i=0}^{5}L_{\lambda}^i(u)B_i|_u.$$ Let $\mathbb{V}_{\lambda}$ be the $1$-dimensional $\mathfrak{m}$-eigenspace of the eigenvalue $\lambda$. Then, (\[momentum\]) implies that $\mathbf{L}_{\lambda}|_u\in \mathbb{V}_{\lambda}$, for every $u\in \R$. Therefore, there exist a unique real-analytic map $\widetilde{r}_{\lambda}:\R-D_{\lambda}\to \C$ such that $$\label{if1} \dot{\mathbf{L}}_{\lambda}|_u=\widetilde{r}_{\lambda}|_u\mathbf{L}_{\lambda}|_u\quad \forall u\in \R-D_{\lambda}.$$ From $\dot{B}=B\cdot \mathcal{K}$ and keeping in mind (\[mcc2\]) we have $$\label{if11}\dot{\mathbf{L}}_{\lambda}\equiv \frac{k_2\dot{k}_2+\lambda}{k_2^2-\lambda_j^2}\mathbf{L}_{\lambda}\quad \mathrm{mod}(B_0,B_1,B_2,B_3,B_4).$$ From (\[if1\]) and (\[if11\]) we have $r_{\lambda}=\widetilde{r}_{\lambda}$. Using (\[if1\]), we deduce that $e^{-\delta_{\lambda}}\mathbf{L}_{\lambda}$ is constant on $\R-D_{\lambda}$. On the other hand, $e^{-\delta_{\lambda}}L_{\lambda}$ extends smoothly across $D_{\lambda}$ and hence also $e^{-\delta_{\lambda}}\mathbf{L}_{\lambda}$ extends to a real-analytic map $\R\to \C^6$. This implies that $e^{-\delta_{\lambda}}\mathbf{L}_{\lambda}=\mathbf{A}_{\lambda}$, for some $\mathbf{A}_{\lambda}\in \mathbb{V}_{\lambda}$. This proves the first part of the statement.
Let $\lambda$ be a multiple root of $P_{\mathfrak{m}}$. The map $T_{\lambda}$ satisfies $$\label{PT} \mathcal{H}|_u\cdot T_{\lambda}|_u=\lambda T_{\lambda}|_u+L_{\lambda}|_u,\quad
T_{\lambda}|_u\wedge L_{\lambda}|_u\neq 0,\quad \forall u\in \R-D_{\lambda}.$$ We put $$\mathbf{T}_{\lambda}=e^{-\delta_{\lambda}}\sum_{j=0}^{5} T^j_{\lambda}B_j : \R-D_{\lambda}\to \C^6.$$ From (\[PT\]) we obtain $$\label{PPT}\mathfrak{m}(\mathbf{T}_{\lambda})=\lambda \mathbf{T}_{\lambda} +\mathbf{A}_{\lambda},\quad
\mathbf{T}_{\lambda}\wedge \mathbf{A}_{\lambda}\neq 0.$$ Differentiating the first equation in (\[PPT\]) we get $\mathfrak{m}(\mathbf{T}_{\lambda}')=\lambda \mathbf{T}_{\lambda}'$. Thus, there exist a unique real-analytic function $\widetilde{s}_{\lambda}:\R-D_{\lambda}\to \C$ such that $\mathbf{T}_{\lambda}'=\widetilde{s}_{\lambda}\mathbf{A}_{\lambda}$. Using $\dot{B}=B\cdot \mathcal{K}$ we obtain $$\dot{\mathbf{T}}_{\lambda} \equiv \frac{\lambda^2+k_2^2-2\lambda k_2\dot{k}_2}{(\lambda-k_2^2)^2} \mathbf{A}_{\lambda},\quad \mathrm{mod}(B_0,B_1,B_2,B_3,B_5).$$ Then, $\widetilde{s}_{\lambda}=s_{\lambda}$. This implies that $\mathbf{T}_{\lambda}-\eta_{\lambda}\mathbf{A}_{\lambda}$ is constant on $\R-D_{\lambda}$. Since $e^{-\delta_{\lambda}}(T_{\lambda}-\eta_{\lambda}L_{\lambda})$ extends smoothly across $\widetilde{D}_{\lambda}$, also $\mathbf{T}_{\lambda}-\eta_{\lambda}\mathbf{A}_{\lambda}$ extends to a smooth (real-analytic) map $\R\to \C^6$. Hence there exist $\mathbf{C}_{\lambda}\in \C^6$ such that $\mathbf{T}_{\lambda}-\eta_{\lambda}A_{\lambda}=\mathbf{C}_{\lambda}$. Using (\[PPT\]) it follows that $\mathfrak{m}(\mathbf{C}_{\lambda})=\lambda \mathbf{C}_{\lambda} + \mathbf{A}_{\lambda}$ and that $\mathbf{C}_{\lambda}\wedge \mathbf{A}_{\lambda}\neq 0$. This yields the result.
[Since $\gamma$ is a standard configuration of a world-line then $B|_0=\mathrm{Id}_{6\times 6}$. Hence $$\mathbf{A}_{\lambda}=L_{\lambda}|_0\quad \mathbf{C}_{\lambda}= T_{\lambda}|_0.$$ Therefore, the principal vectors can be computed explicitly in terms of the phase parameters $e_1,e_2$ and $e_3$.]{}
Integrability by quadratures
============================
Integrability by quadratures of the regular conformal world-lines
-----------------------------------------------------------------
Let $\gamma:\R\to \mathcal{E}^{1,3}$ be the standard configuration of a regular linearly full conformal world-line. Its momentum operator has six simple roots $\lambda_0,\dots ,\lambda_5$, ordered as in (\[eigenvl\]). Let $\delta_j$ be the integrating factor of the first kind and $\mathbf{A}_j$ be the principal vector of $\lambda_j$ respectively. We denote by $\mathbf{A}\in \C(6,6)$ the matrix with column vectors $\mathbf{A}_0,\dots , \mathbf{A}_5$ and we define [^11] the real-analytic maps $$\Delta,\Lambda : \R-D_{\lambda}\to \C(6,6),\quad \mathbf{X}:\R\to \C(6,6)$$ by $$\begin{cases}
\Delta = (e^{-\delta_0}\varepsilon_0,\dots e^{-\delta_5}\varepsilon_5),\\
\Lambda = (L_{\lambda_0},\dots, L_{\lambda_5}),\\
\mathbf{X}=\Delta\cdot \Lambda.\end{cases}$$
\[ThmC\][Let $\gamma$ be as above, then $$\gamma =\mathtt{m}\cdot ^t(\mathbf{A}^{-1})\cdot \mathbf{X} \cdot \mathtt{m}\cdot E_0,$$ where $\mathtt{m}$ is the matrix representing the scalar product $\langle-,-\rangle$ with respect to the standard light-cone basis $(E_0,\dots, E_5)$ of $\R^{2,4}$.]{}
[Since $\mathbf{A}_j=e^{-\delta_j}B\cdot L_{\lambda_j}$, $j=0,\dots, 5$, we have $\mathbf{A}=B\cdot \Lambda\cdot \Delta$, that is $$\label{FFF1}B=\mathbf{A}\cdot \Delta^{-1}\cdot \Lambda^{-1}.$$ From (\[FFF1\]) and keeping in mind that $^tB\cdot \mathtt{m}\cdot B = \mathtt{m}$ we get $$\label{FFF2}\Lambda=\mathtt{m}\cdot \mathbf{X}^{-1}\cdot ^t\mathbf{A}\cdot \mathtt{m}\cdot \mathbf{A}\cdot \Delta^{-1}.$$ Inserting (\[FFF2\]) in (\[FFF1\]) we have $$B=\mathtt{m}\cdot ^t(\mathbf{A}^{-1})\cdot \mathbf{X}\cdot \mathtt{m}.$$ This yields the result. ]{}
Integrability by quadratures of the exceptional conformal world-lines
---------------------------------------------------------------------
Let $\gamma:\R\to \mathcal{E}^{1,3}$ be the standard configuration of an exceptional linearly full conformal world-line. Its momentum operator has four distinct roots $\lambda_0,\dots ,\lambda_3$, ordered as in (\[eigenvl\]). Then, $\lambda_0$ and $\lambda_1$ are simple purely imaginary roots and $\lambda_2,\lambda_3$ are real double roots of $P_{\mathfrak{m}}$. For each eigenvalue $\lambda_j$, $j=0,\dots, 3$, let $\delta_j$ be its integrating factor of the first kind and $\mathbf{A}_j$ be the corresponding principal vector. For each double root $\lambda_j$, $j=2,3$, let $\eta_j$ be the integrating factor of the second kind and $\mathbf{C}_j$ be the secondary principal vector. We denote by $\widetilde{\mathbf{A}}$ the matrix $\widetilde{\mathbf{A}}=
(\mathbf{A}_0,\mathbf{A}_1,\mathbf{A}_2,\mathbf{A}_3,\mathbf{C}_2,\mathbf{C}_3)$. Let $$\widetilde{\Delta},\widetilde{\Lambda}:\R-D_{\lambda}\to \C(6,6),\quad
\widetilde{\mathbf{X}}:\R\to \C(6,6)$$ be the real-analytic maps defined by $$\begin{cases}
\widetilde{\Delta}=(e^{-\delta_0}\varepsilon_0,e^{-\delta_1}\varepsilon_1,
e^{-\delta_2}\varepsilon_2,e^{-\delta_3}\varepsilon_3,e^{-\delta_2}\varepsilon_4,
e^{-\delta_3}\varepsilon_5),\\
\widetilde{\Lambda}=(L_{\lambda_0},L_{\lambda_1},L_{\lambda_2},L_{\lambda_3},
T_{\lambda_2}-\eta_2L_{\lambda_2},T_{\lambda_3}-\eta_3 L{\lambda_3}),\\
\widetilde{\mathbf{X}}=\widetilde{\Delta}\cdot \widetilde{\Lambda}
\end{cases}$$
\[ThmD\][Let $\gamma$ be as above, then $$\gamma =\mathtt{m}\cdot ^t(\widetilde{\mathbf{A}}^{-1})\cdot \widetilde{\mathbf{X}}\cdot \mathtt{m}\cdot E_0.$$]{}
[From Proposition \[PFIF\] it follows that the canonical frame, the principal vectors and the integrating factors satisfy $$e^{-\delta_0}B\cdot L_{\lambda_0}=\mathbf{A}_0,\quad e^{-\delta_1}B\cdot L_{\lambda_1}=\mathbf{A}_1,\quad e^{-\delta_2}B\cdot L_{\lambda_2}=\mathbf{A}_2,\quad e^{-\delta_3}B\cdot L_3=\mathbf{A}_3$$ and $$e^{-\delta_2}B\cdot (T_{\lambda_2}-\eta_2L_{\lambda_w})=\mathbf{C}_2\quad e^{-\delta_3}B\cdot (T_{\lambda_3}-\eta_3L_{\lambda_3})=\mathbf{C}_3.$$ We then have $$\label{FFFF1}B=\widetilde{\mathbf{A}}\cdot \widetilde{\Delta}^{-1}\cdot \widetilde{\Lambda}^{-1}.$$ Combining (\[FFFF1\]) with $^tB\cdot \mathtt{m}\cdot B= \mathtt{m}$ we obtain $$\label{FFFF2}\widetilde{\Lambda}^{-1}=\widetilde{\Delta}\cdot \widetilde{\mathbf{A}}^{-1}\cdot \mathtt{m}\cdot ^t(\widetilde{\mathbf{A}}^{-1})\cdot \widetilde{\mathbf{X}}\cdot \mathtt{m}.$$ Substituting (\[FFFF2\]) into (\[FFFF1\]) we find $B=\mathtt{m}\cdot ^t(\widetilde{\mathbf{A}}^{-1})\cdot \widetilde{\mathbf{X}}\cdot \mathtt{m}$. We have thus proved the result.]{}
Final comments
--------------
The theoretical explanation of the integrability by quadratures lies in the Arnold-Liouville integrability of the Euler-Lagrange differential system. With this we mean the following : let $\mathfrak{m}\in \mathfrak{a}(2,4)$ be the momentum of a linearly full world-line $\gamma$ with non-constant curvatures, then $\mathfrak{m}$ is either regular or exceptional. The stabilizer $A^{\uparrow}_+(2,4)_{\mathfrak{m}}$ of $\mathfrak{m}$ is a $3$-dimensional closed subgroup, diffeomorphic to $S^1\times \R^2$. The inverse image of $\mathfrak{m}^{\natural}$ by the momentum map is a four-dimensional sub-manifold $Z_{\mathfrak{m}}\subset Z$ and the characteristic vector field $X_{\zeta}$ is tangent to $Z_{\mathfrak{m}}$. The world-lines with momentum $\mathfrak{m}$ are originated by the integral curves of $X_{\zeta}|_{Z_{\mathfrak{m}}}$. The stabilizer $A^{\uparrow}_+(2,4)_{\mathfrak{m}}$ acts freely on $Z_{\mathfrak{m}}$ and the quotient space $Z_{\mathfrak{m}}/A^{\uparrow}_+(2,4)_{\mathfrak{m}}$ is a circle. This implies that $Z_{\mathfrak{m}}\subset Z$ is diffeomorphic to the Cartesian product of $\R^2$ with a $2$-dimensional torus $T^2$. In principle, the integration by quadratures can be achieved by a diffeomorphism $\Psi_{\mathfrak{m}} : Z_{\mathfrak{m}}\to \R^2\times T^2$ such that $\Psi_*(X_{\zeta})$ is a linear vector field. Since the stabilizer of the momentum operator is non compact then the trajectory of $\gamma$ can’t be closed. Instead, if $\gamma$ is trapped in a $3$-dimensional Einstein universe, the stabilizer of the momentum operator can be a maximal compact abelian subgroup of $A^{\uparrow}_+(2,3)$. Thus, in this case, there are countably many closed world-lines with non-constant curvatures, as it has been shown in [@DMN].
The periodicity of the conformal curvatures implies that its trajectory of $\gamma$ is left unchanged by the action of the infinite cyclic subgroup generated by $B(\omega)\cdot B(0)^{-1}\in A^{\uparrow}_+(2,4)$.
Lastly, we note that if we know the phase parameters $e_1,e_2,e_3$ all the steps of the integration procedure can be implemented in [*Mathematica*]{} and are fully operative from a computational viewpoint.
Appendix : the integrating factors
==================================
Integrating factors of the first kind
-------------------------------------
Given $\lambda\in \mathcal{S}_{\mathfrak{m}}$ we put $$\label{abc}a=\ell_1-\lambda^2\ell_3,\quad b=\ell_2-\lambda\ell_4,\quad c=\frac{\lambda\ell_4}{\ell_2-\lambda^2\ell_4}$$ and $$\label{d}d=\frac{\lambda (\ell_2\ell_3 -\ell_1 \ell_4)}
{(\ell_2-\lambda^2\ell_4)(\ell_1-\lambda^2\ell_3)},$$ where $\ell_1,\ell_2,\ell_3$ and $\ell_4$ are the constants defined as in (\[ell\]). From (\[k21\]) and (\[if2\]) we obtain $$\label{r}r_{\lambda}(u)=\frac{1}{2}\frac{d}{du}\left(\ln \frac{a-b\mathrm{sn}^2(\sqrt{\ell_3}u,m)}{\ell_3-\ell_4 \mathrm{sn}^2(\sqrt{\ell_3}u,m)}\right)+c+\frac{d}{1-\alpha^2\mathrm{sn}^2(\sqrt{\ell_3}u,m)},$$ where the parameter $m$ is as in (\[ell\]) and $\alpha$ is as in (\[palpha\]).
### The integrating factor of the first kind of a non-real eigenvalue
Let $\lambda$ be a non-real eigenvalue, then $$1-\alpha^2\mathrm{sn}^2(\sqrt{\ell_3}u,m)\neq 0,\quad \forall u\in \R$$ and $$\int \frac{du}{1-\alpha^2\mathrm{sn}^2(u,m)}=\Pi(\alpha^2,\mathrm{am}_m(u),m),$$ where $\Pi(n,\phi,m)$ is the incomplete integral of the third kind and $\mathrm{am}_m(-)$ is the Jacobi’s amplitude with parameter $m$. Note that in this case, the restriction of the incomplete integral of the third kind on the real axis is a regular real-analytic function. Since $$\frac{a-b\mathrm{sn}^2(\sqrt{\ell_3}u,m)}{\ell_3-\ell_4 \mathrm{sn}^2(\sqrt{\ell_3}u,m)}\notin \R_-,\quad \forall u\in \R$$ we can evaluate the logarithm[^12] of the function on the left hand side in the above formula. Thus, the integrating factor of the eigenvalue is given by $$\label{ifc}
\delta_{\lambda}(u)=\frac{1}{2}\ln\left(\frac{a-b\mathrm{sn}^2(\sqrt{\ell_3}u,m)}{\ell_3-\ell_4 \mathrm{sn}^2(\sqrt{\ell_3}u,m)}\right)+c u+\frac{d}{\sqrt{\ell_3}} \Pi(\alpha^2,\mathrm{am}_m(\sqrt{\ell_3} u),m).$$
### The integrating factor of the first kind of a real eigenvalue {#IFFR}
If $\lambda$ is a real eigenvalue, the function $r_{\lambda}$ is singular and the evaluation of the integrating factor requires some caution. Let $w$ and $v$ be the real constants $$w=\frac{\alpha}{\sqrt{(\alpha^2-m)(\alpha^2-1)}},$$ and $$v=\frac{E(m)}{K(m)}-E(p,m)-\mathrm{cs}(p,m)\mathrm{dn}(p,m)-
\frac{\sqrt{(\alpha^2-m)(\alpha^2-1)}}{\alpha},$$ where $p=p_{\lambda}$ is as in (\[palpha\]) and $E(m), E(-,m)$ are the complete and incomplete elliptic integrals of the second kind respectively. Let $f_{\lambda}$ be the periodic extension with period $2\omega$ of the locally constant function $$f_{\lambda}(u)=\begin{cases}-\frac{\pi}{2},\quad u\in [p-\omega,p),\\
-\frac{3\pi}{2},\quad u\in [p,\omega-p),\quad \quad \lambda <0 \\
\quad \frac{\pi}{2},\quad u\in [\omega-p,\omega+p),\end{cases}$$ and $$f_{\lambda}(u)=\begin{cases}-\frac{\pi}{2},\quad u\in [p-\omega,p),\\
\quad \frac{\pi}{2},\quad u\in [p,\omega+p),\end{cases}$$ if $\lambda>0$. Denote by $\vartheta_1(-,q_m)$ the first Jacobi’s theta function with nome $$q_m=\mathrm{exp}(-\pi K(1-m)/K(m)).$$ Proceeding as in [@La], p.$71$, we see that $$\label{PTR1}
g_{\lambda,1}(u)=\frac{w}{2\sqrt{\ell_3}}\ln
\left(\frac{\vartheta_1(\frac{\pi}{2K(m)}(p-
\frac{u}{\sqrt{\ell_3}}),q_m)}{\vartheta_1(\frac{\pi}{2K(m)}(p+\frac{u}{\sqrt{\ell_3}}),q_m)} \right)+wvu$$ is a real-valued primitive of $(1-\alpha^2\mathrm{sn}^2(\sqrt{\ell_3}u,m))^{-1}$. We take $$\label{PTR2}
\begin{split}
g_{\lambda,2}(u) & =\frac{1}{2}\ln(a)+cu-\frac{1}{2}\ln(\ell_3-
\ell_4\mathrm{sn}(\sqrt{\ell_3}u,m)+\\
& + \frac{1}{2}\ln(1-\alpha^2\mathrm{sn}^2(\sqrt{\ell_3}u,m))
\end{split}$$ as a real-valued primitive of $$\frac{1}{2}\frac{d}{du}\left(\ln \frac{a-b\mathrm{sn}^2(\sqrt{\ell_3}u,m)}{\ell_3-\ell_4 \mathrm{sn}^2(\sqrt{\ell_3}u,m)}\right)+c.$$ Then, $$\label{IIFR}\delta_{\lambda}= dg_{\lambda,1}+g_{\lambda,2}+if_{\lambda},$$ is the integrating factor for the real eigenvalue $\lambda$.
Integrating factors of the second kind
--------------------------------------
Let $\lambda$ be a multiple root of $P_{\mathfrak{m}}$. Note that $\lambda$ is necessarily real. From $$s_{\lambda}=\frac{\lambda^2+k_2^2-2\lambda k_2\dot{k}_2}{(\lambda^2-k_2^2)^2}
=-2\lambda \frac{k_2\dot{k}_2}{(\lambda^2-k_2^2)^2}+\frac{\lambda^2+k_2^2}{k_2^2-\lambda^2}$$ we can write $$\int s_{\lambda}du = \eta_{1,\lambda}+\eta_{2,\lambda}$$ where $$\label{eta1}
\begin{split}\eta_{1,\lambda}&=-2 \lambda\int \frac{k_2\dot{k}_2}{(\lambda^2-k_2^2)^2}du=
\frac{\lambda}{k_2^2-\lambda^2}=\\
&= \lambda \frac{\ell_3-\ell_4\mathrm{sn}^2(\sqrt{\ell_3}u,m)}{(\ell_1-\lambda^2\ell_3)-
(\ell_2-\lambda^2\ell_4)\mathrm{sn}^2(\sqrt{\ell_3}u,m)}=\\
&=\frac{\lambda}{\ell_1-\lambda^2\ell_3}\frac{\ell_3-\ell_4\mathrm{sn}^2(\sqrt{\ell_3}u,m)}
{1-\alpha^2\mathrm{sn}^2(\sqrt{\ell_3}u,m)}
\end{split}$$ and $$\label{FTR}
\begin{split}\eta_{2,\lambda}& =\int\frac{\lambda^2+k_2^2}{(\lambda^2-k_2^2)^2}du = \textsf{A} \int \frac{\mathrm{cn}^2(\sqrt{\ell_3}u,m)}{1-\alpha^2 \mathrm{sn}^2(\sqrt{\ell_3}u,m)}du+\\
&+\textsf{B} \int \frac{\mathrm{cn}^2(\sqrt{\ell_3}u,m)\mathrm{sn}^2(\sqrt{\ell_3}u,m)}{1-\alpha^2 \mathrm{sn}^2(\sqrt{\ell_3}u,m)}du + \textsf{C} \int \frac{\mathrm{sn}^2(\sqrt{\ell_3}u,m)}{1-\alpha^2 \mathrm{sn}^2(\sqrt{\ell_3}u,m)}du,
\end{split}$$ where $$\textsf{A}=\frac{\ell_3(\ell_1+\lambda^2\ell_3)}{(\ell_1-\lambda^2\ell_3)^2},\quad
\textsf{B}=-\frac{\ell_4(\ell_2+\lambda_2\ell_4)}{(\ell_1-\lambda^2\ell_3)^2},$$ and $$\textsf{C}=\frac{(\ell_3-\ell_4)((\ell_1-\ell_2)+\lambda^2(\ell_3-\ell_4))}{(\ell_1-\lambda^2\ell_3)^2}.$$ The integrals in the right hand side of (\[FTR\]) can be evaluated as in [@BF] p. $218$, and, as a result, we obtain $$\label{eta2}\begin{split}
\eta_{2,\lambda}&=\frac{\textsf{M}}{\sqrt{\ell_3}}E(\sqrt{\ell_3}u,m)+\textsf{N}u+
\textsf{P}(g_{\lambda,1}-i\frac{\pi}{2}\widehat{f}_{\lambda})+\\
&+\frac{\textsf{Q}}{\sqrt{\ell_3}}
\frac{\mathrm{sn}(\sqrt{\ell_3}u,m)\mathrm{cn}(\sqrt{\ell_3}u,m)\mathrm{dn}(\sqrt{\ell_3}u,m)}
{1-\alpha^2\mathrm{sn}^2(\sqrt{\ell_3}u,m)}
\end{split}$$ where $g_{\lambda,1}$ is as in (\[PTR1\]) and $\widehat{f}_{\lambda}$ is the periodic extension with period $\omega$ of the locally constant function $$f(u)=\begin{cases}
0,\quad u\in [0,p),\\
1,\quad u\in [p,\omega)\end{cases}$$ and $\textsf{M},\textsf{N},\textsf{P},\textsf{Q}$ are the constants $$\begin{cases}
\textsf{M}=\frac{\alpha^2}{2\alpha^2(m-\alpha^2)}\left(\textsf{A}+\frac{1}{\alpha^2}\textsf{B}+
\frac{1}{\alpha^2-1}\textsf{C}\right),\\
\textsf{N}=\frac{1}{2\alpha^2}\left(\textsf{A}+\textsf{C}-\frac{1}{\alpha^2}\textsf{B}\right),\\
\textsf{P}=\frac{1}{2\alpha^2(m-\alpha^2)}\left((2m\alpha^2-\alpha^4-m)\textsf{A}+(\alpha^4-m)\textsf{C}+
(\alpha^4-2\alpha^2+m)\textsf{B}\right),\\
\textsf{Q}=-\frac{\alpha^2}{2(m-\alpha^2)}\left(\textsf{A}+\frac{1}{\alpha^2-1}\textsf{C}+
\frac{1}{\alpha^2}\textsf{B}\right).
\end{cases}$$ Summing up : the integrating factor of the second kind of a multiple root is given by $$\label{IFIIK}\eta_{\lambda}=\eta_{1,\lambda}+\eta_{2,\lambda}$$ where $\eta_{1,\lambda}$ and $\eta_{2,\lambda}$ are the functions defined as in (\[eta1\]) and (\[eta2\]).
[AA]{}
M. Akivis and V. Goldberg, [*Conformal Differential Geometry and its Generalizations*]{}, John Wiley and Sons Inc. New York, 1996.
, [A primier on the (2+1)-Einstein universe]{}, in [*Recent developments in pseudo-Riemannian geometry*]{}, 179–229, ESI Lect. Math. Phys., Eur. Math. Soc., Zürich, 2008; [arXiv:0706.3055\[math.DG\]]{}
P.F. Byrd and M.D. Friedman, [*Handbook of Elliptic Integrals for Engineers and Scientist*]{}, Grundlehren der mathematischen Wissenschaften, 67, Springer-Verlag, New York, Heidelberg, Berlin, 1971.
, [*Vorlesungen über Differentialgeometrie. III: Differentialgeometrie der [K]{}reise und Kugeln*]{}, Grundlehren der mathematischen Wissenschaften, 29, Springer, Berlin, 1929.
E. Cartan, Les espaces $\grave{a}$ connexion conforme, [*Ann. Soc. pol. Math.*]{} **2** (1923), 171–221.
T.E. Cecil, [*Lie sphere geometry[:]{} with applications to submanifolds*]{}, Springer-Verlag, New York, 1992.
P. T. Chruściel, G. J. Galloway, D. Pollack, Mathematical general relativity: a sampler, [*Bull. Amer. Math. Soc. (N.S.)*]{} **47** (2010), 567–638.
, *Modern geometry–methods and applications. Part I*, 2nd edn., GTM 93. Springer-Verlag, New York, 1992.
Conformal geometry of timelike curves in the (1+2)-Einstein universe, *Nonlinear Analysis* (to appear); [arXiv:1603:01035v1\[math.DG\]]{}
, The conformal theory of curves, *Trans. Amer. Math. Soc.* [**51**]{} (1942), 435–501.
, *Géometrie et dynamique lorentziennes conformes*, Thése, E.N.S. Lyon (2002).
, Möbius energy of knots and unknots, [*Ann. of Math. (2)*]{} [**139**]{} (1994), no. 1, 1–50.
, Coisotropic variational problems, [*J. Geom. Phys.*]{} [**50**]{} (2004), 303–338.
, *Variations on a theme by Kepler*, AMS Colloquium Publications, **285**, Providence, RI, 1990.
, *Exterior differential systems and the calculus of variations*, Progress in Mathematics, 25, Birkhäuser, Boston, 1982.
, Conformal differential geometry. Curves in conformal euclidean spaces, *Proc. Nederl. Akad. Wetensch.* **44** (1941), 814–824.
, *The large scale structure of space-time*, Cambridge Monographs on Mathematical Physics, no. 1, Cambridge University Press, London-New York, 1973.
S. Kobayashi, *Transformation Groups in Differential Geometry*, Ergebnisse der Mathematik, 70, Springer-Verlag, New York, Heidelberg, Berlin, 1971.
G. Jensen, E. Musso and L. Nicolodi, *Surfaces in Classical Geometries*, Universitext, Springer International Publishing, Switzerland, 2016.
D. F. Lawden, *Elliptic Functions and Applications*, Applied Mathematics Series 80, Springer-Verlag, New York, Heidelberg, Berlin, 1989.
, Conformal arc-length as $\frac{1}{2}$-dimensional length of the set of osculating circles, *Comment. Math. Helv.* **85** (2010), no. 2, 273–312.
, The geometry of canal surfaces and the length of curves in de Sitter space, [*Adv. Geom.*]{} **11** (2011), no. 4, 58–601.
, Beiträge zur Inversionsgeometrie der Kurven, Münchener Berichte, 1923.
, On the geometry of curves and conformal geodesics in the Möbius space, *Ann. Global Anal. Geom.* **40** (2011), 133-165.
, Conformal curvatures of curves in $\R^{n+1}$, *Indag. Math. (N.S.)* **12** (2001), 369–382.
, The conformal arclength functional, *Math. Nachr.* **165** (1994), 107–131.
, Closed trajectories of the conformal arclength functional, [*Journal of Physics: Conference Series*]{} **410** (2013), 012031.
, Quantization of the conformal arclength functional on space curves, *Comm. Anal. Geom.* (to appear); arXiv:1501.04101\[math.DG\].
, *Cycles of time: An extraordinary new view of the universe*, Alfred A. Knopf, Inc., New York, 2010.
, On the gravitization of quantum mechanics 2: Conformal cyclic cosmology, *Found. Phys.* **44** (2014), 873-890.
, *Spinors and space-time*, Camridge monographs on Mathematical Physics, Vol. I and II, Cambridge University Press, 1986.
, Submanifolds of the Möbius space, *Math. Nachr.* **96** (1980), 165–183.
, Submanifolds of the Möbius space II, Frenet formula and curves of constant curvatures, *Math. Nachr.* **100** (1981), 235–257.
, *Differentialgeometrien in den Kugelräumen. Band I. Konforme Differentialkugelgeometrie von Liouville und Möbius*, Maruzen, Tokyo, 1938.
, Über konforme Geometrie I: Grundlagen der konformen Flächentheorie, [*Hamb. Math. Abh.*]{} **3** (1923), 31–56.
, Penrose’s Weyl curvature hypothesis and conformally-cyclic cosmology, [*Jornal of Physics: Conference Series*]{} **229** (2010), 1–5.
, Contribution à la géométrie conforme. Enveloppes de sphères et courbes gauches, *J. École Polytechnique* **25** (1925), 43–91.
, *Time, Space, Matter* Dover Publications, Inc., Mineola, N.Y., 1952.
[^1]: Authors partially supported by PRIN 2010-2011 “Varietà reali e complesse: geometria, topologia e analisi armonica” and by GNSAGA of INDAM.
[^2]: ie, $\widetilde{\gamma}=B\cdot (\gamma\circ h)$, where $B\in \mathrm{A}^{\uparrow}_+(1,n+1)$ and $h$ is a change of parameter.
[^3]: $\mathrm{supp}(\mathbf{g})$ denotes the support of the variation, ie the closure of the set of all $t\in I$ such that $\gamma_{\tau}(t)\neq \gamma(t)$ for some $\tau\in (-\epsilon,\epsilon)$.
[^4]: $pr|_u$ is the orthogonal projection of $\mathbb{M}^{1,n-1}$ onto $N(\alpha)|_u$
[^5]: $\mathbf{J}$ is the faithful representation of $P^{\uparrow}_+(1,n)$ into the conformal group.
[^6]: cfr (\[cinv1\]) for the definition of $h_1$
[^7]: The parameter $m$ is the square of the modulus $k$ of the elliptic function. The reader pay attention to the fact that in the literature is also used the notation $\mathrm{sn}(-,k)$ to denote the $\mathrm{sn}$-function with modulus $k$.
[^8]: Remember that $\gamma$ is linearly full, with positive helicity and parameterized by the conformal parameter
[^9]: ie the vector field defined by $\zeta(X_{\zeta})=1$, $\iota_{X_{\zeta}}d\zeta=0$.
[^10]: If $z\in \C-\R^+$, then $\sqrt{z}$ is the determination of the square root with positive imaginary part.
[^11]: $\varepsilon_j=^t(\delta^0_j,\dots, \delta^5_j)$ are the column vectors of the canonical basis of $\R^6$.
[^12]: We use the standard determination of the natural logarithm, with a branch cut discontinuity in the complex plane running from $-\infty$ to $0$.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'M. Jardine'
- 'A. Collier Cameron'
date: 'Received; accepted 2008'
title: 'Radio emission from exoplanets: the role of the stellar coronal density and magnetic field strength'
---
Introduction
============
Searches for the radio signatures of extra-solar planets have not yet yielded any detections. The motivation for these searches was the well-known decametre emission from Jupiter which can outshine the radio emission from the quiet Sun at these wavelengths by some four orders of magnitude [@farrell_radio_planets_99]. This emission is believed to originate from non-thermal electrons travelling down the converging field lines at the planet’s magnetic poles [@dulk85; @zarka_98]. There are several possible sources for these electrons, including currents generated by corotation breakdown inside the middle magnetosphere, the Io-Jupiter interaction and the reconnection of Jupiter’s magnetic field with the solar wind magnetic field [@cowley_01; @saur_04]. Here we focus on the third possibility since it is common to all the magnetised planets. If these electrons are accelerated to energies greater than 1-10 keV, they can be unstable to the electron cyclotron maser instability and as a result they emit at the local electron gyrofrequency $f \mbox{[MHz] = 2.8B(G)}$. The observed peak of Jupiter’s emission at 39.5MHz suggests a field strength of 14.5G at the site of the emission close to Jupiter’s surface [@connerney_98]. The proposal that such a situation might occur in some of the known “Hot Jupiters" led to several searches but none of these has revealed the expected emission [@ryabov_04]. A comprehensive review of the current status of theory and observation is given in @zarka_07.
One of the most promising reasons for this lack of detections is that tides raised on planets orbiting close to their host stars rapidly synchronise the axial spin with the orbital motion. In this case, the planet may be rotating slowly, producing only a weak magnetic field. Since the emission frequency is proportional to the magnetic field strength, this could result in emission at unobservably low frequencies. The other synchronisation that can take place is between the spin of the star and the orbit of the planet. This timescale is set by torques due to the tidal bulges raised on the star by the planet (see, e.g. @zahn77) and is generally expected to be longer than the star’s main-sequence lifetime even for close-orbiting gas-giant planets. The observed spin rates of planet-host stars are consistent with this expectation. A notable exception is the late F star tau Boo, which rotates synchronously with the 3.3-day orbit of its planet, whose mass is sufficient to lock the star’s rotation within its main-sequence lifetime [@marcy_51peg_97; @lubow_spindown_97]. More recently @dobbs_dixon_planets_04 have considered the combined effects of tidal torques and angular momentum loss from the star in the form of a wind. They suggest that synchronisation is more likely among F stars than among the lower mass G and K stars.
The nature of the stellar wind may also have an impact on the expected radio emission, since the wind may compress the stellar magnetosphere and also may strip particles out of the planetary exosphere [@griessmeier_planet_tides_04; @griessmeier_planet_winds_05; @jaritz_planet_roche_05; @lipatov_3Dhybrid_05; @stevens_planet_winds_05]. Fluctuations in the stellar wind pressure and speed that occur as a result of the stellar equivalent of solar coronal mass ejections may, however, enhance any planetary radio emission [@khodachenko_CME_07]. A recent search for radio emission from systems thought to be embedded in strong stellar winds has yielded no detections, but tight upper limits [@stevens_planet_winds_05; @george_radio_planets_07].
It is clear from this modeling that conditions both exterior to the planet (such as the speed, density and magnetic energy of the stellar wind) and within the planet itself (such as the nature of the dynamo-generated magnetic field) may have an impact on the detectability of exoplanetary radio emission. There is another factor, however, which is the process by which the power carried in the stellar wind is converted into radio emission. Within our own solar system, there is a clear scaling between the magnetic or kinetic power in the solar wind incident on a planet and the observed output radio power [@farrell_radio_planets_99; @zarka_01; @zarka_07]. This “radio-magnetic Bode’s law” can be used to predict the radio emission from a variety of stellar wind or planetary magnetosphere conditions.
In this paper we take a complementary approach. Rather than focussing on the physics of the stellar wind or the planet, we choose simple models for these and focus instead on the process of energy conversion. Our aim is to examine how the radio power scales with different stellar parameters, and to determine if this scaling is the same for all exoplanets.
The model
=========
We begin by considering the interaction between the planet and its immediate environment. In the case of the outer planets, this takes the form of a stellar wind impacting the planetary magnetosphere. This wind carries both magnetic and kinetic energy, but it is the magnetic component that we are particularly interested in. Following previous work (e.g. @zarka_01 [@zarka_07]), we can calculate the rate at which the stellar wind delivers magnetic energy to the planetary magnetosphere by calculating the Poynting flux $(\bf{E} \times {\bf B}/ \mu)$ that is incident on the cross-sectional area $A$ of the planetary magnetosphere. We can approximate this maximum available power as $P = AvB^2/\mu$. This depends on the cross-section of the planetary magnetosphere $A = \pi R_m^2$ the wind velocity $v$ and the stellar field strength $B$ as measured at the planetary magnetosphere. This maximum available power sets an upper limit on the power that could be detected in radio emission.
The situation for many “Hot Jupiters" is, however, somewhat different. For these planets, the orbital radius is extremely small - so much so that the planet may in fact be inside the stellar magnetosphere, where the stellar magnetic field lines are closed and confine the coronal gas. The maximum extent of the stellar magnetosphere is expected to increase with stellar activity (and hence with stellar rotation rate). Thus from a value of 2.5R$_\odot$ seen in eclipse images of the Sun [@altschuler69; @badalian_Kcorona_86], it may increase to $\approx$19R$_\odot$ or 0.09AU for a star with rotation period about 6 days that is 10 times more active than the Sun in terms of surface magnetic flux density or soft X-ray luminosity [@schrijver01; @schrijver_asterosphere_03]. Given the youth of some exoplanetary systems, many planet host stars may have magnetic activity strong enough to support magnetospheres extending to 0.1 AU or so. Hot Jupiters with orbital distance of order 0.05 AU will thus orbit within their host stars’ magnetospheres. In this case, the velocity with which new magnetic field lines are carried into the interaction region is the relative velocity between the stellar and planetary magnetospheres. In the frame of the planet, however, this would appear to be similar to the effect of a stellar wind.
The stellar wind therefore has two main effects on the planetary magnetosphere. The wind ram pressure compresses the planetary magnetosphere and elongates it as shown in Figs (\[cartoon\_1\]) and (\[cartoon\_2\]). The magnetic field that is carried in the wind is also pushed against the planetary magnetic field forming a narrow interaction region or [*current sheet*]{}. The magnetic field lines will pile up in this region until reconnection allows the field to dissipate. This process is observed both on the dayside and nightside of the Earth’s magnetosphere and is related to the occurrence of aurorae as accelerated electrons stream along the Earth’s magnetic field lines.
The interaction of the planetary and stellar magnetic fields also generates an electric field at the interaction (or reconnection) site. If this electric field has a component along the magnetic field lines, it will accelerate electrons along the magnetic field lines. These electrons will mostly be slowed down by collisions, but some fraction may escape and be freely accelerated. It is this small fraction of the electron population that may provide the pool of energetic electrons needed to supply the electron cyclotron maser instability that would result in observable radio emission.
In those rare cases, such as the anomalously massive $\tau$ Boo b, where the stellar rotation is synchronised with the orbital motion, then there will be no relative motion between the stellar field and the planetary field and so although reconnection may happen sporadically due to the small velocity differences due to stellar differential rotation or resistive instabilities such as perhaps the tearing mode [@Furth63; @cuntz_exoplanets_00], it is unlikely to happen in the quasi-continuous manner that would be needed to power an observable radio signal. In the majority of observed cases, however, the star’s rotation is not synchronous. New magnetic flux will then be brought into the interaction region at a rate determined by the velocity difference between the stellar and planetary fields. In this case the geometry of the interaction is altered and the elongation of the planetary magnetosphere will not be directed away from the star, but will trail the orbital motion ( see Figs. \[cartoon\_1\] and \[cartoon\_2\]).
Magnetospheric radius
=====================
The first step in calculating the radio power emitted by exoplanets is to calculate the cross-section of the planetary magnetosphere that intercepts the stellar wind. The magnetospheric radius is determined by pressure balance between the external medium and the planetary magnetosphere: $${B_p^2} = {B_e^2} +\mu \rho_e v^2.
\label{press_bal}$$ here $B_p$ is the magnetic field of the planet which we take to be 14.5G. We assume that this is a dipole, and so for a planet of radius $R_p$ this scales with distance as $(R/R_p)^{-3}$. The field in the external region $B_e$ is the stellar field. Close to the star, where the field lines are still closed, we assume that this field is also a dipole and so for a star of radius $R_\star$ this field scales with distance as $(R/R_\star)^{-3}$. At distance greater than that of the stellar magnetosphere, however, where the field lines have been opened up by the stellar wind, the field scales as $(R/R_\star)^{-2}$. For the inner planets, the velocity in (\[press\_bal\]) is the relative velocity between the star and planet: $$\begin{aligned}
v & = & v_p - v_\star \\
& = & 3 \times 10^4 \left[ \frac{R_{orb}}{0.03AU} \right]
\left[ \frac{11d}{p_\star} \right]
\left[ \frac{p_\star}{p_{orb}} - 1 \right] [\mbox{ms}^{-1}]
%\label{delta_v} \end{aligned}$$ where we take the stellar rotation period to be 11 days (appropriate for HD189733 [@bouchy_hd189733_05]). For the outer planets, this velocity is the wind velocity. This is an extremely difficult parameter to determine observationally, though use of hydrogen-wall absorption to deduce the wind ram pressure promises to be a useful tool [@wood_asterospheres_04]. Models of these winds also give very conflicting results [@stevens_planet_winds_05; @holzwarth_coolwinds_07] and so, for the moment, we simply assume that the winds have reached their terminal velocity and so the wind density falls off with distance as $(R/R_\star)^{-2}$. We scale the stellar wind velocity to solar values, assuming a nominal solar wind speed of 500kms$^{-1}$ and a density at 1AU of 1.7$\times 10^{-20}$kgm$^{-3}$ [@cox91]. This gives a number density $$\begin{aligned}
N & = & N_w(1AU)\left[ \frac{R_\star}{R_{orb}}\right]^2 \\
& = &10^7 \left[\frac{R_\star}{R_\odot} \right]^2
\left[\frac{R_{orb}}{1AU} \right] ^{-2}[\mbox{m}^{-3}].
\label{density_m}\end{aligned}$$
For close-orbiting planets, however, the external density is not that of a wind, but that of the stellar magnetosphere. This is likely to be a function not only of the orbital radius but also the magnetospheric temperature and the stellar rotation rate (which influences the level of magnetic activity). We assume that the stellar magnetosphere is in isothermal, hydrostatic balance and so obtain (see Fig. \[density\]) $$N = N_0 \rm{exp} \left(
-\frac{\mu m_H}{kT} \frac{GM_\star}{R_\star}
\left[ 1-\frac{R_\star}{R_{orb}}
\right]
\right)[\mbox{m}^{-3}]
\label{density_w}$$ with the number density at the base of the corona $N_0 = 4 \times 10^{14}$m$^{-3}$, the temperature T=$1.4\times 10^6$K and the mean particle mass $\mu=0.6$ [@badalian_Kcorona_86]. We note that this gives a different scaling of the density $N$ with orbital radius in the two cases; for the outer planets, $N \propto R^{-2}$, whereas for the inner planets, $N\propto R^{-1}$.
Using these figures, we find, in common with previous studies [@cuntz_exoplanets_00; @zarka_01; @zarka_07] that for the more distant planets, the contribution of the external field to pressure balance is negligible. For the inner planets the opposite is true, and we may neglect the ram pressure associated with the relative motion of the planet and the star. This then gives, for the inner planets: $$\frac{R_m}{R_p} =
15.7 \left[ \frac{R_{orb}}{0.03AU} \right] \left[ \frac{B_p}{14.5G} \right]^{1/3}
\left[ \frac{R_\star}{R_\odot} \right] ^{-1}\left[ \frac{B_\star}{1G} \right]^{-1/3},
\label{Rm_inner}$$ and for the outer planets: $$\frac{R_m}{R_p} =
46 \left[ \frac{R_{orb}}{5AU} \right]^{1/3} \left[ \frac{B_p}{14.5G} \right]^{1/3}
\left[ \frac{v_e}{500kms^{-1}} \right]^{-1/3}.
\label{Rm_outer}$$ We show in Fig. (\[size\]) the corresponding sizes of the planetary magnetospheres (assuming that all planets have a magnetic field similar to Jupiter’s). While for planets far from their parent stars the magnetospheric size is independent of the the stellar field strength (being determined mainly by the wind ram pressure), for the close-in planets, the pressure of the stellar magnetic field can have a significant effect in compressing the planetary magnetosphere. If the stellar field strength is too high, the planetary magnetosphere may be totally crushed, but for intermediate stellar field strengths or stronger planetary fields the planet may still retain a magnetosphere even in a very close orbit.
The magnetic interaction between the star and the planet
========================================================
The interaction of the planetary magnetosphere with the external medium not only determines the size of the planetary magnetosphere but it also generates an electric field at the site where the magnetic fields of the planet and the star meet given by $${\bf E} = -{\bf v}\times{\bf B} + {\bf j}/H$$ where the current density $\bf{j} = \bf{\nabla}\times\bf{B}/\mu$. Here we have scaled the electric field to its value far from the interaction site $E_0=-v_0B_0$ where $v_0$ and $B_0$ are the velocity and magnetic field strength. The current density is scaled to $B_0/\mu L$ where $L$ is the typical lengthscale for the interaction which we take to be 0.1$R_J$. The results are in fact fairly insensitive to the exact value of $L$ chosen. The magnetic Reynolds number $H=v_0 L/\eta$ where the Spitzer diffusivity is given by $\eta = 10^9 T^{-3/2}$m$^2$s$^{-1}$ [@priest84].
Thus, well away from the interaction site, the electric field is simply determined by the velocity and magnetic field strength and we have $\bf{E} = -\bf{v}\times\bf{B}$, while at the centre of the interaction region where $v=0$, the electric field is determined by the magnetic Reynolds number and we have ${\bf E}={\bf j}/H$.
It can be shown (see Appendix for details) that there is a component of this electric field that is parallel to the magnetic field and so is capable of accelerating electrons along the magnetic field lines. As shown in the Appendix, this electric field ($E_{||}$) has a maximum at the interaction site between the fields of the star and the planet. Its magnitude falls away with increasing distance $z$ from this site as $${\bf E}_{||} = {\bf E}\cdot{\bf B}
= {E}_{||}(0) e^{-z^2H}.$$ It is the magnitude of this parallel electric field that is important in accelerating electrons in the interaction region. Its value is given simply by $${E}_{||}(0) = -\frac{ 2\sqrt{2\pi}}
{\Gamma(1/4)} H^{-1/4} v_0 B_0.
\label{E_parallel}$$ Hence, although there is a weak dependence on the magnetic Reynolds number (and hence on the temperature of the local plasma), the magnitude of this accelerating electric field is determined primarily by the magnetic field strength and the velocity at some distance from the interaction site.
Power in accelerated electrons
==============================
Once the electrons are accelerated by this field however, they may be slowed down by collisions with a collision frequency $$\nu_c = 7.5\times 10^{-5} T^{-3/2} n$$ that depends on the local temperature $T$ and density $n$, here given in SI units [@kruskal64; @tandberghanssen88]. However, if the electric field is stronger than a critical value known as the [*Dreicer*]{} field $E_D$, electrons may be freely accelerated and so will “run away” out of the thermal distribution. The Dreicer field is also a function of the local temperature and density and is given by $$E_D = 18\times 10^{-12} n T^{-1} [{\rm V m}^{-1}].
\label{dreicer}$$ The number density of runaway electrons is then given by $$\begin{aligned}
N_{\rm run} & = & 0.35 n \nu_c f(E/E_D) [{\rm m}^{-3}] \\
& = & 2.6 \times 10^{-5} n^2T^{-3/2}f(E/E_D) [{\rm m}^{-3}]
\label{N_acc}\end{aligned}$$ where the function $f(E/E_D)$, shown in Fig. \[f(ER)\], is given by $$f\left(\frac{E}{E_D}\right) = \left(\frac{E_D}{E}\right)^{3/8} {\rm exp} \left[-\left(\frac{2E_D}{E}\right)^{1/2} - \frac{E_D}{4E}\right].
\label{runaway}$$ The behaviour of this function governs the way in which the electric field generated at the interaction site can affect the number density of electrons that run away and hence can power the radio emission. Once this imposed electric field is greater than the Dreicer field, there is a steep rise in $f(E/E_D)$ with increasing $E$. Once the ratio $E/E_D$ is sufficiently large, however, $f$ becomes insensitive to further increases in $E/E_D$ and saturates at a value of about 0.3. We can see from Eqs. \[E\_parallel\] and \[dreicer\] that for a given temperature, $$\frac{E}{E_D} \propto \frac{v_0 B_0}{n}.
\label{E_ratio}$$ The local temperature at the acceleration site $n$ will depend on the details of the reconnection process and its determination would require a full analysis of the local dynamics. It is a reasonable assumption however that it scales with the density of the plasma flowing into the reconnection site such that $n \propto N$. We determine the constant of proportionality empirically by fitting the predicted radio power to that observed for the solar system planets, and find it to be of order 7. Hence, as shown in Fig \[E\_ED\], the ratio $E/E_D$ is the same for all the solar system planets since the solar wind velocity is a constant and both the density and field strength vary as $R^{-2}_{\rm orb}$. This is not the case for close-in exoplanets, where the velocity, density and field strength vary differently with orbital radius.
Electrons that are accelerated over a lengthscale $R_m$ to an energy $K = qER_m$ carry a power $$P = \pi R_m^2 v N_{run} K [\mbox{W}]$$ where the velocity $v$ is either the wind velocity (for the outer planets) or the relative velocity of the star and planet (for the inner planets). Hence $P \propto vN^2 R_m^2$ and so the dominant factors determining the output power are the size of the planetary magnetosphere and the density. Using expressions \[density\_m\] to \[Rm\_outer\], we can show that for given planetary parameters, this power scales as $$P \propto N_c^2 vB_\star^{-2/3}$$ for the inner planets and $$P \propto N_w^2 vR_\star^4 R_{orb}^{-10/3}$$ for the outer planets. Both of these cases are shown in Fig. \[power\_all\]. For those planets embedded in the stellar wind, the fall-off in power with increasing orbital radius is very pronounced. For the inner planets, however, the power is almost independent of orbital radius. As the orbital radius decreases, the number of electrons available to be accelerated rises (since $N_{run} \propto n^2 \propto R_{orb}^{-2}$). This is balanced however by the shrinkage of the collecting region of the magnetosphere that is apparent in Fig. \[size\] since $R_m^2 \propto R_{orb}^{2}$.
The power carried by the accelerated electrons for these inner planets is therefore largely independent of the planetary orbital radius (albeit with a weak variation due to the relative velocity $v$). The magnitude of this power is however affected by the stellar parameters we have chosen. For a stellar field strength of only 1G, not only is the typical electron energy of the runaway electrons reduced to a few keV, (and so we set the characteristic electron energy to 1eV) but also since the electric field generated is in many cases less than the Dreicer field, fewer electrons are accelerated. A higher stellar field strength would of course provide a larger electric field and hence accelerate electrons to higher energies, but it would lead to a greater compression of the planetary magnetosphere and hence reduce the upper limit for the lengthscale over which these electrons could be accelerated.
We can quantify the effect of the magnetic field by determining the minimum stellar field strength needed to ensure that the distance $L_{\rm min}$ required for electrons to be accelerated above (say) 10keV is less than the maximum available lengthscale (R$_m$). Thus $$B_{\star,min} \propto \frac{R_{orb}^{3/2}}{R_\star^3}$$ for close-in planets, while for the outer planets $$B_{\star,min} \propto \frac{R_{orb}^{5/3}}{R_\star^3}.$$ We show this minimum stellar field strength in Fig. (\[B\_min\]) where we assume a stellar orbital period equal to that of HD189733 for close-in planets and a stellar wind speed of 500kms$^{-1}$ for the outer planets. Clearly, only a small increase in the stellar field strength above 1G would be adequate to ensure that a pool of electrons of energies greater than 10keV would be available to power the electron-cyclotron maser instability for most exoplanetary systems. Even below this limit, energies of greater than 1keV are easily achievable.
Comparison with the solar system
================================
While Fig. \[power\_all\] illustrates the scaling of the expected power with orbital radius, the level of this power depends on our assumptions for the stellar field strength and densities. The fraction of this power that may be converted into observable radio emission also depends on the details of the electron distribution and the electron cyclotron maser instability (see e.g. review by [@wu_review_ecmi_85]). Values of $1 - 20 \%$ have been claimed [@louarn_92; @mackinnon_ECM_92; @cairns_CMI_02; @vorgul_CMI_05] and so we assume that $10\%$ of the power in accelerated electrons is converted into radio power. This allows us to compare our predicted radio power with that observed from the planets in our own solar system, which is directly proportional to the input magnetic power carried by the solar wind (see Fig. (\[Pout\_Pin\])). This “radio-magnetic Bode’s law” has a constant of proportionality or efficiency of $1-10 \times 10^{-3}$ relating the magnetic and radio powers [@zarka_07]. We can write this efficiency $\epsilon = P_{\rm radio}/P_{\rm magnetic}$ as $$\epsilon = 0.1 \frac{\mu N_{\rm run} K}{B^2}$$ where $K$ is the typical energy acquired by the accelerated electrons. The factor of 10$\%$ is an estimate of the fraction of the power in accelerated electrons that can be converted to radio emission. Since we expect that $N_{\rm run} \propto N^2$, this gives $$\epsilon \propto \left[\frac{N}{B}\right]^2.$$ For the outer planets (such as those in our own solar system) we can easily reproduce this constant efficiency $\epsilon$ by noting that both $N$ and $B$ scale as $(R_{orb}/R_\star)^{-2}$ This gives a constant efficiency whose magnitude depends principally on the local field strength and density. If we scale these to their values at the Earth’s orbit (a density of $1.7\times10^{-20}$kgm$^{-3}$ and a field strength of 3.5nT) then we can reproduce the emission from the solar system planets shown in Fig. \[Pout\_Pin\].
Discussion
==========
For the exoplanets that are far from their parent stars, a universal “radio-magnetic Bode’s law” might therefore be expected, but with a slope that is determined by the ratio of $N/B$ in the stellar wind. These two factors may well not be independent and indeed may vary with both spectral type and stellar age. @griessmeier_planet_winds_05 have considered the effect on the detectability of radio emission of changes in stellar winds as stars age. They conclude that young systems with more powerful winds are the most likely candidates.
The inner planets are, however, much more promising candidates for detection because of their larger predicted radio powers. The magnitude of this power also depends on their coronal density and to a lesser extent the stellar field strength. Increasing B$_\star$ increases the input power and so moves all the points in Fig. \[Pout\_Pin\] to the right. The output power, however, is not significantly affected. Thus, for example, if we consider HD189733 we can convert the radio power to a flux
$\Phi$ received at Earth following @farrell_radio_planets_99 and @griessmeier_planet_winds_05 [@griessmeier_winds_07] to obtain $$\Phi = \frac{P}{\Omega d^2 \delta f}.$$ Here d is the distance to the star and we assume planetary properties similar to Jupiter [@zarka_conf_04]. Hence $\delta f$ is the bandwidth, taken as $\delta f = f_c$ where $f_c = 40$MHz is the maximum cyclotron frequency and we assume that the emission is beamed into a solid angle $\Omega$ = 1.6 sr. While for a stellar field of 1G (as shown in Fig. \[Flux\_planet\]), the flux from HD189733 is only 14mJy, if we use the observed stellar field strength of 40G, the electric field increases so that the typical electron energy rises, but the magnetosphere shrinks, with the result that the flux changes to only 15mJy. This assumes, however, that the coronal density is equal to that of the Sun. If we assume scalings in the range $N \propto \Omega^{0.6}$ to $N \propto \Omega$ [@unruh97loops; @ivanova_taam_braking_03] then for the rotation rate of 11.73 days appropriate for HD189733, the flux rises to from 15mJy to 39 - 72 mJy respectively.
This example points to the importance of the stellar coronal density in determining the possible radio power. It suggest that those stars with high X-ray flux (which like the radio flux also scale as $N^2$) are the most likely candidates. Although rapid rotation may therefore lead to a greater radio power, most radial-velocity searches for exoplanets have tended to exclude rapidly-rotating objects from their survey samples, because their high rotational broadening degrades the precision with with radial velocities can be determined, and their surface activity causes radial-velocity “jitter”. Stars that are magnetically active tend to have significant radio emission of their own, which could mask any planetary signature [@benz_guedel_94]. Indeed, we might expect planet-induced emission due to the electron-cyclotron maser process at the star. This would appear at the cyclotron frequency of the stellar magnetic field $f [MHz]= 2.8B_\star[G] $ [@dulk85]. Thus, for HD189733 with an observed field strength of 40G [@moutou_HD189733_07] we might expect emission at 112MHz. In order for this emission to escape from the stellar magnetosphere, however, the local cyclotron frequency must be greater than the plasma frequency [@dulk85; @bingham_horseshoe_00; @zarka_07; @griessmeier_07]. This means that at the site of emission, $$B > 3 \times 10^{-6} N_e[m^{-3}] ^{1/2} [\mbox{G}].$$ For emission from the planet, this condition is easily met, requiring only a planetary field strength of 0.1G for a density of 10$^9$m$^{-3}$. If we are considering emission from the star, however, this condition requires that $B_\star > 60$G for a density of $4\times 10^{14}$m$^{-3}$.
Stars that have higher field strengths will have a higher emission frequency and so may be easier to observe. In principle, this could be distinguished from the intrinsic emission of the star [@kellet_magnetic_trap_02] by its modulation due to the orbit of the planet, but this would require observation over many orbits to average out the effect of transient stellar emission due to stellar coronal activity. This process may also lead to chromospheric signatures of the interaction as electrons from the reconnection site are accelerated along the stellar field lines and impact the stellar surface. @shkolnik03 [@shkolnik_05; @shkolnik_spi_08] have reported planet-induced chromospheric activity on HD 179949 and $\upsilon$ And. @mcivor_planet_06 modelled this process and showed that this interaction could explain the chromospheric activity, but not the phase lag between the planet and the chromospheric hotspot for $\upsilon$ And. @preusse_05 [@preusse_06] have, however, reproduced this phase lag by considering the propagation of Alfvén waves generated by the passage of an exoplanet through a stellar magnetic field.
The level of magnetic activity may be important in determining the rate at which the stellar field changes may also be important. These changes are most likely to be driven by new flux emergence through the stellar surface, or by surface transport processes: diffusion, differential rotation and meridional flows. While there is now a general acceptance that surface differential rotation decreases with decreasing mass [@barnes_diffrot_05], the nature of the surface diffusion or meridional flows is as yet uncertain.
The stellar magnetic field also has a direct effect on the emission process through its control of the size of the planetary magnetosphere. Indeed, if the stellar field strength is too large, the planetary magnetosphere may be totally crushed. We can calculate the minimum stellar field strength required to do this by setting the size of the planetary magnetosphere equal to the planetary radius in the condition for pressure balance (\[press\_bal\]), and setting the ram pressure term to zero. This gives $$B_{\star,\rm min} = 3.8 \times 10^3 \left(\frac{R_{orb}}{0.03 AU}\right)^3 \left(\frac{R_\odot}{R_\star}\right)^3 [\mbox{G}].$$
For the subset of planets that transit their host stars, there is the possibility of observing radio occultations as the planet passes behind the star. With the observation of repeated occultations any time-dependence in the emission due to variations in the emission process itself could be averaged out. Such observations might provide information about the structure and orientation of the planetary magnetic field, since most of the emission is expected to originate in the region of the magnetic pole close to the planetary surface. Even if this emission is present, however, it may be beamed out of the line of sight. Recent claims of electron-cyclotron emission from low-mass stars with kG magnetic fields are, however, extremely encouraging and suggest that occultations are indeed possible as the source region rotates out of view behind the star [@hallinan_08; @antonova_hallinan_08]. The emission is at much higher frequencies than would be expected from a planetary magnetic field however.
Conclusions
===========
In studying the possible radio emission from exoplanets, we have focussed on the electric field that is generated when the planetary and stellar magnetic fields reconnect. This electric field can accelerate electrons to energies sufficient to power the electron-cyclotron maser instability which is believed to be responsible for much of the radio emission from the planets in our solar system. We have assumed that the mass, radius and magnetic field strength of all exoplanets is the same as Jupiter, but used the observed stellar values. By modeling the electric field developed in this interaction, we have calculated the number of accelerated electrons produced. We find that it scales as the square of the density $N$. Thus the output radio power scales as $N^2$ while the input power from the stellar magnetic field scales as $B^2$. Since for planets immersed in a stellar wind (as is the case for the planets in our own solar system) both the density and the field strength scale as $R_{\rm orb}^2$, this naturally predicts that the output and input power should be proportional, as is observed for the solar system.
We find, however, that the output power from the inner planets behaves differently. The planets lie inside the closed magnetic field (or magnetosphere) of the star. The magnetospheres of these planets are confined not by the ram pressure of the stellar wind, but by the magnetic pressure of the stellar magnetosphere. The density and magnetic pressure within which they are immersed do not vary with orbital radius in the same way as for the outer planets, and so we would not expect the output and input powers to be proportional. In fact, we find that as the planetary orbital radius decreases, the density rises and so the pool of available electrons increases, but this is exactly balanced by the shrinking of the magnetosphere which reduces the cross-section of the planet. The result is that while the input power might rise, the output power saturates at a value determined by the ratio $$P_{\rm out} \propto \left[ \frac{N_c}{B_\star^{1/3}}\right]^2$$ where $N_c$ is the density at the base of the stellar corona and $B_\star$ is the surface field strength. We conclude, therefore, that the radio power emitted by “hot Jupiters" is determined by the way that coronal density scales with stellar field strength as a function of stellar rotation rate and effective temperature.
[59]{} natexlab\#1[\#1]{}
, M. D. & [Newkirk, Jr.]{}, G. 1969, Solar Physics, 9, 131
, A., [Doyle]{}, J. G., [Hallinan]{}, G., [Bourke]{}, S., & [Golden]{}, A. 2008, ArXiv e-prints, 805
, O. G. & [Livshits]{}, M. A. 1986, Solar Physics, 103, 385
, J. R., [Cameron]{}, A. C., [Donati]{}, J.-F., [et al.]{} 2005, MN, 357, L1
, A. O. & [Guedel]{}, M. 1994, A&A, 285, 621
, R. & [Cairns]{}, R. A. 2000, Physics of Plasmas, 7, 3089
, F., [Udry]{}, S., [Mayor]{}, M., [et al.]{} 2005, A&A, 444, L15
, A. & [Bingham]{}, R. 2002, APS Meeting Abstracts, 1064P
, J. E. P., [Acu[ñ]{}a]{}, M. H., [Ness]{}, N. F., & [Satoh]{}, T. 1998, JGR, 103, 11929
, S. W. H. & [Bunce]{}, E. J. 2001, Planet. Space. Sci., 49, 1067
, A. N., [Livingston]{}, W. C., & [Matthews]{}, M. S. 1991, [Solar interior and atmosphere]{} (Tucson, AZ, University of Arizona Press, 1991, 1429 p. For individual items see A92-36202 to A92-36240.)
, M., [Saar]{}, S. H., & [Musielak]{}, Z. E. 2000, ApJL, 533, L151
, I., [Lin]{}, D. N. C., & [Mardling]{}, R. A. 2004, ApJ, 610, 464
, G. A. 1985, An. Rev. Astr. Astrophys., 23, 169
, W. M., [Desch]{}, M. D., & [Zarka]{}, P. 1999, JGR, 104, 14025
, H., [Killeen]{}, J., & [Rosenbluth]{}, M. 1963, Phys. Fluids, 6, 459
, S. & [Stevens]{}, I. 2007, ArXiv e-prints, 708
, J.-M., [Motschmann]{}, U., [Mann]{}, G., & [Rucker]{}, H. O. 2005, A&A, 437, 717
, J.-M., [Preusse]{}, S., [Khodachenko]{}, M., [et al.]{} 2007, Planet. Space. Sci., 55, 618
, J.-M., [Stadelmann]{}, A., [Penz]{}, T., [et al.]{} 2004, A&A, 425, 753
, J.-M., [Zarka]{}, P., & [Spreeuw]{}, H. 2007, A&A, 475, 359
, G., [Antonova]{}, A., [Doyle]{}, J. G., [et al.]{} 2008, ArXiv e-prints, 805
, V. & [Jardine]{}, M. 2007, A&A, 463, 11
, N. & [Taam]{}, R. E. 2003, ApJ, 599, 516
, M., [Allen]{}, H., & [Grundy]{}, R. 1993, JGR, 98, 19409
, M., [Allen]{}, H., & [Pollock]{}, A. A&A, 1996, 314, 594
, G. F., [Endler]{}, S., [Langmayr]{}, D., [et al.]{} 2005, A&A, 439, 771
, B. J., [Bingham]{}, R., [Cairns]{}, R. A., & [Tsikoudi]{}, V. 2002, MN, 329, 102
, M. L., [Ribas]{}, I., [Lammer]{}, H., [et al.]{} 2007, Astrobiology, 7, 167
, M. & [Bernstein]{}, I. 1964, Phys. Fluids, 7, 407
, A. S., [Motschmann]{}, U., [Bagdonat]{}, T., & [Grie[ß]{}meier]{}, J.-M. 2005, Planet. Space. Sci., 53, 423
, P. 1992, Advances in Space Research, 12, 121
, S. H., [Tout]{}, C. A., & [Livio]{}, M. 1997, ApJ, 484, 866
, A., [Vlahos]{}, L., & [Vilmer]{}, N. 1992, A&A, 256, 613
, G. W., [Butler]{}, R. P., [Williams]{}, E., [et al.]{} 1997, ApJ, 481, 926
, T., [Jardine]{}, M., & [Holzwarth]{}, V. 2006, MN, 367, L1
, C., [Donati]{}, J. ., [Savalle]{}, R., [et al.]{} 2007, ArXiv e-prints, 707
, S., [Kopp]{}, A., [B[ü]{}chner]{}, J., & [Motschmann]{}, U. 2005, A&A, 434, 1191
, S., [Kopp]{}, A., [B[ü]{}chner]{}, J., & [Motschmann]{}, U. 2006, A&A, 460, 317
, E. 1984, Solar Magnetohydrodynamics (Dordrecht: D. Reidel)
, V. B., [Zarka]{}, P., & [Ryabov]{}, B. P. 2004, Planet. Space. Sci., 52, 1479
, J., [Neubauer]{}, F. M., [Connerney]{}, J. E. P., [Zarka]{}, P., & [Kivelson]{}, M. G. 2004, [Plasma interaction of Io with its plasma torus]{} (Jupiter. The Planet, Satellites and Magnetosphere), 537–560
, C. J., [DeRosa]{}, M. L., & [Title]{}, A. M. 2003, ApJ, 590, 493
, C. J. & [Title]{}, A. M. 2001, ApJ, 551, 1099
, E., [Bohlender]{}, D. A., [Walker]{}, G. A. H., & [Collier Cameron]{}, A. 2008, ApJ, 676, 628
, E., [Walker]{}, G. A. H., & [Bohlender]{}, D. A. 2003, ApJ, 597, 1092
, E., [Walker]{}, G. A. H., [Bohlender]{}, D. A., [Gu]{}, P.-G., & [K[ü]{}rster]{}, M. 2005, ApJ, 622, 1075
, B. & [Priest]{}, E. 1975, J. Plasma Phys. 14, 283
, I. R. 2005, MN, 356, 1053
, E. & [Emslie]{}, A. 1988, The physics of solar flares (Cambridge)
, Y. & [Jardine]{}, M. 1997, A&A 321, 177
, I., [Cairns]{}, R. A., & [Bingham]{}, R. 2005, Physics of Plasmas, 12, 2903
, B. E., [M[" u]{}ller]{}, H.-R., [Zank]{}, G. P., [Izmodenov]{}, V. V., & [Linsky]{}, J. L. 2004, Advances in Space Research, 34, 66
, C. S. 1985, Space Science Reviews, 41, 215
, J.-P. 1977, A&A, 57, 383
, P. 1998, JGR, 103, 20159
, P. 2004, in Astronomical Society of the Pacific Conference Series, Vol. 321, Extrasolar Planets: Today and Tomorrow, ed. J. [Beaulieu]{}, A. [Lecavelier Des Etangs]{}, & C. [Terquem]{}, 160–+
, P. 2007, Planet. Space Sci., 55, 598
, P., [Treumann]{}, R. A., [Ryabov]{}, B. P., & [Ryabov]{}, V. B. 2001, Astr. Space Sci., 277, 293
Modelling the reconnection of the stellar and planetary magnetic fields
=======================================================================
We model the interaction of the stellar and planetary magnetic fields in the manner similar to that used for colliding wind systems [@jardine96particle]. Because of the large length scales involved we neglect the radius of curvature of the reconnecting field lines. In this case, the equation of motion of the flow is just $$\label{eqn:of:motion}
\rho({\bf v}\cdot{\bf \nabla}){\bf v} = - {\bf \nabla}
\left[
p +B^2/2\mu
\right]
+ \rho\nu\nabla^2{\bf v},$$ where $\nu$ is the kinematic viscosity and the other symbols have their usual meaning. For this local analysis, we use a Cartesian coordinate system where the $z$-axis lies along the line of centres of the star-planet system (see Figure \[coords\]). As the magnetic fields of the star and planet are pushed together by the local flow, a current sheet is generated. Within this current sheet, components of the magnetic field that are antiparallel will annihilate, while components that are parallel will combine. We therefore chose to align the $x$- and $y$-axes such that the $x$-components of the magnetic field on either side of the current sheet are exactly antiparallel and hence cancel out within the current sheet, while the $y$-components are parallel. As a result, at the center of the current sheet the magnetic field has only a $y$-component. The magnetic field is then of the form $${\bf B} = (B_x(z),B_y(z),0)$$ and the current density $\bf{j} = \bf{\nabla}\times\bf{B}/\mu$ is $$\mu{\bf j} = (-B'_y, B'_x, 0 ).$$ The variation with distance from the centre of the current sheet of the velocity and magnetic field for colliding magnetofluids has been determined for arbitrary vorticities by @jardine93,with a numerical solution that shows that the magnetic field strength at the centre of the current sheet varies only slowly with vorticity. We therefore consider only the case of a zero-vorticity flow for which an exact solution for the magnetic field profile can be found [@sonnerup75]. In this case, the velocity is of the form $${\bf \overline{v}} = (\overline{x},\overline{y},-2\overline{z})$$ where all distances are scaled to some global lengthscale $L$ so that $\overline{x}=x/L$ etc and $\overline{v}=v/v_0$ where $v_0$ is the velocity with which field lines are being carried towards the current sheet. Now, by our assumption of a steady state, $\bf{\nabla}\times \bf{E} = 0$ where the electric field is given by a simple Ohm’s law $$\bf{E} = -\bf{v}\times\bf{B} + \bf{j}/\sigma.$$ This can be written as $$\label{B}
\frac{\overline{B}''}{H}
+ 2\overline{z}\overline{B}'
+ \overline{B} =0,$$ where $H=v_0 L/\eta$ is the magnetic Reynolds number and $\overline{B}=B/B_0$ where $B_0 = B(\overline{z} = 1)$. The electric field $\overline{E} = E/v_0 B_0$ has components $${\bf \overline{E}} = (-2 \overline{z} \overline{B}_y -
\overline{B}'_y/H,
2 \overline{z} \overline{B}_x +
\overline{B}'_x/H,
- \overline{x} \overline{B}_y + \overline{y}
\overline{B}_x).$$ If we take the component of $\bf{E}$ along $\bf{B}$ to be $\bf{E}_{||}$, it is straightforward to show that $${\bf E}_{||} = {\bf E}\cdot {\bf B}
= {E}_{||}(0) e^{-z^2H}$$ and so we would expect that the centre of the current sheet is the site of the most efficient particle acceleration. We therefore concentrate our attention on the particles accelerated there. The degree to which the magnetic field rotates as it comes into the current sheet then depends on the relative magnitudes of the x- and y-components outside the current sheet. We specify this as one of the boundary conditions, in fact making them equal at $z=R_m$ so that the field rotates through $\pi/4$ radians. With boundary conditions $$\label{mag_bcs}
\overline{B}_{\rm{x}}(0)=0, \makebox[0.25cm]{}
\overline{B}_{\rm{x}}(1)=1, \makebox[0.25cm]{}
\overline{B}'_{\rm{y}}(0)=0, \makebox[0.25cm]{}
\overline{B}_{\rm{y}}(1)=1$$ the solution to (\[B\]) can be expressed in terms of modified Bessel functions as $$\overline{B}_{\rm{x}} = \frac{z^{1/2}e^{(1-z^2)H/2}}{I_{1/4}(H/2)}
I_{1/4}(z^2 H/2)$$ and $$\overline{B}_{\rm{y}} = \frac{z^{1/2}e^{(1-z^2)H/2}}{I_{-1/4}(H/2)}
I_{-1/4}(z^2 H/2)$$ For $H>>1$ we can write $I_{\pm1/4}(H/2) \approx \left( e^{H}/ \pi H \right)^{1/2}$ so that at the centre of the current sheet we have $$\overline{B}_{\rm{y}}(0) = \frac{\sqrt{2\pi}}
{\Gamma(3/4)} H^{1/4}
\label{odd}$$ $$\overline{B}'_{\rm{x}}(0) = \frac{ 2\sqrt{2\pi}}
{\Gamma(1/4)}H^{3/4}$$ and $$\overline{E}(0) = -\frac{ 2\sqrt{2\pi}}
{\Gamma(1/4)} H^{-1/4}$$ or $${E}(0) = -\frac{ 2\sqrt{2\pi}}
{\Gamma(1/4)} H^{-1/4} v_0 B_0.$$
|
{
"pile_set_name": "ArXiv"
}
|
=1
Ryohei Miyadera\
Kwansei Gakuin, Hyogo, Japan.\
*[email protected]*
**Abstract**
The author introduces the Wythoff game with a pass, a variant of the classical Wythoff game. The latter game is played with two piles of stones, and two players take turns removing stones from one or both piles; when removing stones from both piles, the numbers of stones removed from each pile must be equal. The player who removes the last stone or stones is the winner.
An equivalent description of the game is that a single chess queen is placed somewhere on a large grid of squares, and each player can move the queen towards the upper left corner of the grid: leftward, upward, or to the upper left, any number of steps. The winner is the player who moves the queen into the upper left corner.
In Wythoff’s game with a pass, we modify the standard rules of the game so as to allow for a one-time pass, i.e., a pass move which may be used at most once in a game, and not from a terminal position. Once the pass has been used by either player, it is no longer available.
The author discovered that, when a pass move is available, the set of $\mathcal{P}$-positions (previsous player’s position) of the Wythoff’s game with a pass is equal to $ T_0 \cup T_1\cup B -A$, where $T_0 = \{(x,y): \mathcal{G}(x,y)=0 \}$, $T_0 = \{(x,y): \mathcal{G}(x,y)=1 \}$, $B$ is a finite subset of $\{(x,y): \mathcal{G}(x,y)=4 \}$, $A$ is a finite subset of $T_0$ and $\mathcal{G}$ is the Grundy number of the classical Wythoff game. Here the size of $B$ is $6$, and the size of $A$ is $7$.
According to the research of U.Blass and A.S.Fraenkel, for a position $(x,y) \in T_1 $ there is always a position $(x^{\prime},y^{\prime}) \in T_0$ such that $\mid x- x^{\prime} \mid \leq 2$ and $\mid y - y^{\prime} \mid \leq 4$, and hence the graph of the set of $\mathcal{P}$-positions of the Wythoff’s game with a pass is very similar to the graph of the set of $\mathcal{P}$-positions of the classical Wythoff’s game.
Wythoff’s game
==============
Let $Z_{\ge 0}$ and $N$ be the sets of non-nagative number and natural numbers. Wythoff’s game is a played with two piles of stones. Two players take turns removing stones from one or both piles; when removing stones from both piles, the numbers of stones removed from each pile must be equal. The player who removes the last stone or stones is the winner.
An equivalent description of the game is that a single chess queen is placed somewhere on a large grid of squares, and each player can move the queen towards the upper left corner of the grid: leftward, upward, or to the upper left, any number of steps. The winner is the player who moves the queen into the corner.
Figure \[chessboard\] is the grid of squares, and Figure \[moveofqueen\] shows the move of a queen.
{width="\columnwidth"} \[chessboard\]
{width="\columnwidth"} \[moveofqueen\]
\[movewythoff\] We define $\textit{move}(x,y)$ of Wythoff’s game for $x,y \in Z_{\ge 0}$ with $1 \leq x+y$ by
& *move*(x,y) =M\_1(x,y) M\_2(x,y) M\_3(x,y), & &\
& M\_1(x,y)= {(u,y):u<x}, M\_2(x,y)={(x,v):v<y} u,vZ\_[0]{}&\
& &\
& M\_3(x,y)={(x-t,y-t): 1 t (x,y)} x,y 1.&\
\[explainm1m2m3\] $M_1(x,y), M_2(x,y), M_3(x,y)$ are the sets made by the leftward, the upward and the upperleft move. $M_3(x,y)$ is an empty set if $x = 0$ or $y = 0$.
When we study the Wythoff’s games, there are two important positions of the queen.
\[NPpositions\] $(a)$ $\mathcal{N}$-positions, from which the next player can force a win, as long as he plays correctly at every stage.\
$(b)$ $\mathcal{P}$-positions, from which the previous player $($the player who will play after the next player$)$ can force a win, as long as he plays correctly at every stage.
\[defofmexgrundy\] $(i)$ The *minimum excluded value* $(\textit{mex})$ of a set, $S$, of non-negative integers is the least non-negative integer that is not in S.\
$(ii)$ Each position $\mathbf{p}=(x,y)$ of an impartial game G has an associated Grundy number, and we denote it by $\mathcal{G}(\mathbf{p})$.\
The Grundy number is calculated recursively: $\mathcal{G}(\mathbf{p}) = \textit{mex}\{\mathcal{G}(\mathbf{h}): \mathbf{h} \in \textit{move}(\mathbf{p})\}.$
\[theoremofsumg\] Let $\mathcal{G}$ be the $Grundy$ number. Then, $\mathbf{h}$ is a $\mathcal{P}$-position if and only if $\mathcal{G}(\mathbf{h})=0$．
For the proof of this theorem, see $\cite{lesson}$.
\[defnoft0t1\] Let $T_i = \{(x,y):\mathcal{G}(x,y)=i\}$ for any natural number $i$.
\[pofqueen\] $T_0 = \{ (\lfloor n \phi \rfloor, \lfloor n \phi \rfloor + n ):n \in Z_{\ge 0} \}$ $\cup \{ ( \lfloor n \phi \rfloor + n, \lfloor n \phi \rfloor ):n \in Z_{\ge 0} \}$
, where $ \phi = \frac{1+\sqrt{5}}{2}$.
This is a well known result. See [@wythoffpaper].
\[lemmafort1ti\] $(a)$ If $(x,y) \in T_0$, then $move(x,y) \cap T_0 = \emptyset$.\
$(b)$ If $(x,y) \in T_1$, then $move(x,y) \cap T_1 = \emptyset$.\
$(c)$ If $(x,y) \in T_i$ with $i \neq 0$, then $move(x,y) \cap T_0 \neq \emptyset$.\
$(d)$ If $(x,y) \in T_i$ with $i \neq 0,1$, then $move(x,y) \cap T_1 \neq \emptyset$.
This is direct from the definition of Grundy number.
\[p1ofqueen\] Let $ \{(a_n,b_n)\}=T_1$, where $a_n$ is increasing. Then $\mid b_n - (\lfloor n \phi \rfloor + n) \mid \leq 4$ and $ \lfloor n \phi \rfloor -1 \leq a_n \leq \lfloor n \phi \rfloor +2$.v
This result is proved in pp.330 in [@fraenkelpaper] as Corollary 5.14.
\[positionsofgrundy1\] By Lemma \[pofqueen\] and Lemma \[p1ofqueen\], for each position $(x,y) \in T_1$ there exists a $(x^{\prime},y^{\prime}) \in T_0$ such that $\mid x- x^{\prime} \mid \leq 2$ and $\mid y - y^{\prime} \mid \leq 4$.
\[hvdpath\] $(a)$ For any $n \in Z_{\ge 0}$, let $V(n,0) = \{(n,y):y \in Z_{\ge 0} \}$. We call this a vertical path.\
$(b)$ For any $n \in Z_{\ge 0}$, let $H(0,n) = \{(x,n):x \in Z_{\ge 0} \}$ We call this a horizontal path.\
$(c)$ For any $n \in Z_{\ge 0}$, let $D(n,0) = \{(n+y,y):y \in Z_{\ge 0} \}$ and $D(0,n) = \{(x,n+x):x \in Z_{\ge 0} \}$. We call these diagonal paths.
\[defnofab\] Let $A = \{(0,1),(1,0),(2,2),(3,6),(6,3),(5,7),(7,5) \}$ and $B= $\
$(0,0),(1,3),(3,1),(2,5),(5,2),(6,7),(7,6) \}$.
\[lemmapequql1\] $(i)$ For any $(x,y) \in T_1 \cup B -A$, $move(x,y) \cap (T_1\cup B -A) = \emptyset $.\
$(ii)$ For any $(x,y) \notin T_0 \cup (T_1\cup B -A)$, $move(x,y) \cap (T_1\cup B -A) \neq \emptyset $.
We use a table of Grundy number. Figure \[grundyqueentable\] is a table of Grundy numbers of Wythoff’s game. This is mathematically the same as “The Grundy function of Wythoff’s game” of Table 1 in [@gabriel]. In Figure \[grundyqueentable\] the positions in $ (T_1\cup B -A)$ are in blue squares (or in gray square when printed in black and white.), and the positions in $A$ are in purple squares (or in light gray square when printed in black and white.)
{height="0.45\columnwidth"}
$(i)$ Let $(x,y) \in T_1\cup B -A$.\
$(a)$ Suppose that $(x,y) \in T_1-A$.\
$(a.1)$ By $(b)$ of Lemma \[lemmafort1ti\], $\textit{move}(x,y) \cap T_1 = \emptyset$.\
$(a.2)$ Next, we prove that $\textit{move}(x,y) \cap B = \emptyset$.\
$(a.2.1)$ We prove that we cannot reach from $(x,y) \in T_1-A$ to any of positon of $B$ by moving leftward or upward. $A = \{(0, 1), (1, 0), (2, 2), (3, 6), (6, 3), (5, 7), (7, 5) \}$\
$ \subset T_1$ and $(0,1) \in V(0,0) \cap A \ne \emptyset$, $(1,0) \in V(1,0) \cap A \ne \emptyset$, $(2,2) \in V(2,0) \cap A \ne \emptyset$, $(3,6) \in V(3,0) \cap A \ne \emptyset$, $(5,7) \in V(5,0) \cap A \ne \emptyset$, $(6,3) \in V(6,0) \cap A \ne \emptyset$ and $(7,5) \in V(7,0) \cap A \ne \emptyset$, where the vertical path is defined in Definition \[hvdpath\]. We also have $(1,0) \in H(0,0) \cap A \ne \emptyset$, $(0,1) \in H(0,1) \cap A \ne \emptyset$, $(2,2) \in H(0,2) \cap A \ne \emptyset$, $(6,3) \in H(0,3) \cap A \ne \emptyset$, $(7,5) \in H(0,5) \cap A \ne \emptyset$, $(3,6) \in H(0,6) \cap A \ne \emptyset$ and $(5,7) \in H(0,7) \cap A \ne \emptyset$, where the horizontal path is defined in Definition \[hvdpath\]. By $(b)$ of Lemma \[lemmafort1ti\] and Definition \[hvdpath\] there is only one position of $T_1$ in each vertical and horizontal path, and hence we have $V(0,0) \cap T_1-A = \emptyset$, $V(1,0) \cap T_1-A = \emptyset$, $V(2,0) \cap T_1-A = \emptyset$, $V(3,0) \cap T_1-A = \emptyset$, $V(5,0) \cap T_1-A = \emptyset$, $V(6,0) \cap T_1-A = \emptyset$, $V(7,0) \cap T_1-A = \emptyset$, $H(0,0) \cap T_1-A = \emptyset$, $H(0,1) \cap T_1-A = \emptyset$, $H(0,2) \cap T_1-A = \emptyset$, $H(0,3) \cap T_1-A = \emptyset$, $H(0,5) \cap T_1-A = \emptyset$, $H(0,6) \cap T_1-A = \emptyset$ and $H(0,7) \cap T_1-A = \emptyset$. Since $B \subset V(0,0) \cup V(1,0) \cup V(2,0) \cup V(3,0) \cup V(5,0) \cup V(6,0) \cup V(7,0)$ and $B \subset H(0,0) \cup H(0,1) \cup H(0,2) \cup H(0,3) \cup H(0,5) \cup H(0,6) \cup H(0,7)$, there does not exist $(x,y) \in T_1-A$ from which we can reach the set $B$ by the leftward move or the upward move.\
$(a.2.2)$ Next, we prove that we cannot reach from $(x,y) \in T_1-A$ to any positon of $B$ by moving to the upper left.\
$(a.2.2.1)$ The Diagonal path $D(0,1)$ contains the position $(0,1) \in T_1 \cap A$. Since $D(0,1)$ contains only one position of $T_1$, and hence $(T_1-A) \cap D(0,1) = \emptyset$. Therefore there is no position of $T_1-A$ from which we can reach the position $(0,1) \in D(0,1) $ by diagonal move. We prove almost the same thing for other positions in $B$.\
$(a.2.2.2)$ The position $(1,0) \in T_1 \cap D(1,0) \cap A$, and hence $(T_1-A) \cap D(1,0) = \emptyset$. Therefore there is no position of $T_1-A$ from which we can reach the position $(7,6) \in D(1,0) $ by the move to the upper left.\
$(a.2.2.3)$ The position $(2,2) \in T_1 \cap D(0,0) \cap A$. Since the diagonal path $D(1,0)$ contains only one position of $T_1$, $(T_1-A) \cap D(0,0) = \emptyset$. Therefore there is no position of $T_1-A$ from which we can reach the position $(0,0) \in D(0,0) $ by the move to the upper left.\
$(a.2.2.4)$ The position $(3,6) \in T_1 \cap D(0,3) \cap A$, and hence $(T_1-A) \cap D(0,3) = \emptyset$. Therefore there is no position of $T_1-A$ from which we can reach the position $(2,5) \in D(0,3) $ by the move to the upper left.\
$(a.2.2.5)$ The position $(5,7) \in T_1 \cap D(0,2) \cap A$, and hence $(T_1-A) \cap D(0,2) = \emptyset$. Therefore there is no position of $T_1-A$ from which we can reach the position $(1,3) \in D(0,2) $ by by the move to the upper left.\
$(a.2.2.6)$ The position $(6,3) \in T_1 \cap D(3,0) \cap A$, and hence $(T_1-A) \cap D(3,0) = \emptyset$. Therefore there is no position of $T_1-A$ from which we can reach the position $(5,2) \in D(3,0) $ by the move to the upper left.\
$(a.2.2.7)$ The position $(7,5) \in T_1 \cap D(2,0) \cap A$, and hence $(T_1-A) \cap D(2,0) = \emptyset$. Therefore there is no position of $T_1-A$ from which we can reach the position $(3,1) \in D(2,0) $ by the move to the upper left.\
$(b)$ Suppose that $(x,y) \in B$. Since $B \subset \{(u,v):u \leq 7 \text{ and } v \leq 7 \}$ and $move(x,y) \subset \{(u,v):u \leq 7 \text{ and } v \leq 7 \}$ , by Figure \[grundyqueentable\] it is clear that $move(x,y) \cap (T_1\cup B -A) \neq \emptyset $.\
$(ii)$ Let $(x,y) \notin T_0 \cup (T_1\cup B -A)$. We prove that $move(x,y) \cap (T_1\cup B -A) \neq \emptyset $.\
$(c)$ Suppose that $(x,y) \in \{(u,v):u \leq 7 \text{ and } v \leq 7 \}$. Then, $move(x,y) \subset \{(u,v):u \leq 7 \text{ and } v \leq 7 \}$. By Figure \[grundyqueentable\] we prove $move(x,y) \cap (T_1\cup B -A) \neq \emptyset $.\
$(d)$ Next, we suppose that $$\label{conditionofarea}
(x,y) \notin \{(u,v):u \leq 7 \text{ and } v \leq 7 \}.$$ Then, $A \subset \{(u,v):u \leq 7 \text{ and } v \leq 7 \}$, and hence $(x,y) \notin T_{0} \cup T_{1}$. This implies that $\mathcal{G}(x,y) \neq 0,1$. Therefore, by $(d)$ of Lemma \[lemmafort1ti\] there exists $(s,t) \in T_1 \cap move(x,y) $.\
$(d.1)$ If $(s,t) \in T_1 -A$, then $move(x,y) \cap (T_1\cup B -A) \neq \emptyset $.\
$(d.2)$ Suppose that $(s,t) \in A$. We prove that we can choose $(s^{\prime}, t^{\prime}) \in move(x,y) \cap A.$\
$(d.2.1)$ Suppose that $(s,t) = (5,7)$. \
$(d.2.1.1)$ If we can reach $(5,7)$ from $(x,y)$ by the upward move, then we can reach $(5,2) \in B$ by the upward move.\
$(d.2.1.2)$ If we can reach $(5,7)$ from $(x,y)$ by the leftwar move, then $y = 7 \leq 7$. Then by (\[conditionofarea\]) $x \geq 8$, and hence we can reach $(6,7) \in B$ from $(x,y)$ by the leftward move.\
$(d.2.1.3)$ If we can reach $(5,7)$ from $(x,y)$ by the move to the upper left, then we can reach $(1,3) \in B$ by the upward move.\
$(d.2.2)$ Suppose that $(s,t) = (7,5)$. Then we can use the method that is used in $(d.2.1)$.\
$(d.2.3)$ Suppose that $(s,t) = (3,6)$.\
$(d.2.3.1)$ If we can reach $(3,6)$ from $(x,y)$ by the upward move, then we can reach $(3,1) \in B$ by the upward move.\
$(d.2.3.2)$ If we can reach $(3,6)$ from $(x,y)$ by the leftwar move, then we can reach $(7,6) \in B$ by the leftwar move. Note that $x \geq 8$ by (\[conditionofarea\]).\
$(d.2.3.3)$ If we can reach $(3,6)$ from $(x,y)$ by the move to the upper left, then we can reach $(2,5) \in B$ by the move to the upper left.\
$(d.2.4)$ Suppose that $(s,t) = (6,3)$. Then we can use the method that is used in $(d.2.3)$.\
$(d.2.5)$ Suppose that $(s,t) = (2,2)$.\
$(d.2.5.1)$ If we can reach $(2,2)$ from $(x,y)$ by the upward move, then we can reach $(2,5) \in B$ by the upward move. Note that $y \geq 8$ by (\[conditionofarea\]).\
$(d.2.5.2)$ If we can reach $(2,2)$ from $(x,y)$ by the leftwar move, then we can reach $(5,2) \in B$ by the leftwar move. Note that $x \geq 8$ by (\[conditionofarea\]).\
$(d.2.5.3)$ If we can reach $(2,2)$ from $(x,y)$ by the move to the upper left, then we can reach $(0,0) \in B$ by the move to the upper left.\
$(d.2.6)$ Suppose that $(s,t) = (0,1)$.\
$(d.2.6.1)$ If we can reach $(0,1)$ from $(x,y)$ by the upward move, then we can reach $(0,0) \in B$ by the upward move.\
$(d.2.6.2)$ If we can reach $(0,1)$ from $(x,y)$ by the leftwar move, then we can reach $(3,1) \in B$ by the leftwar move. Note that $x \geq 8$ by (\[conditionofarea\]).\
$(d.2.6.3)$ If we can reach $(0,1)$ from $(x,y)$ by the move to the upper left, then $y=x+1$. Then, by (\[conditionofarea\]) $x \geq 7$ and $y \geq 8$. Therefore, we can reach $(6,7) \in B$ by the move to the upper left.\
$(d.2.7)$ Suppose that $(s,t) = (1,0)$. Then we can use the method that is used in $(d.2.6)$.\
Wythoff Game with a Pass {#queenwithapass}
========================
Throughout the remainder of this paper, we modify the standard rules of Wythoff’s game so as to allow for a one-time pass, i.e., a pass move which may be used at most once in a game, and not from a terminal position. Once the pass has been used by either player, it is no longer available. We call the Wythoff’s game without a pass “the classical Wythoff’s game”.
In this section, we denote the position of the queen with three coordinates $\{x,y,p\}$, where $x,y$ define the position of the queen on the chess board, and $p=1$ if the pass is still available and $p=0$ if not.
When $p=0$, then the game is mathematically the same as the classical Wythoff’s game.
we define the move of a queen in the Wythoff’s game with a pass in Definition \[defmovepass\].
\[defmovepass\]
& MP\_1(x,y,p)= {(s,t,p):(s,t) *move*(x,y) } x,y Z\_[0]{} p = 1,2 &\
& &\
& MP\_2(x,y,p)= {(x,y,0) } x,y Z\_[0]{} 1 x+y p=1,&
where $\textit{move}(x,y)$ is defined in Definition \[movewythoff\].
We define $\textit{move}_{pass}(x,y,p) =MP_1(x,y,p) \cup MP_2(x,y,p).$
$MP_2(x,y,p)$ is the set of the pass move, and it is an empty set if $x = 0$ or $y = 0$ or $p=0$.
\[defofgrundypass\] Each position $(x,y,p)$ of the the Wythoff’s game has an associated Grundy number, and we denote it by $\mathcal{G}_{pass}(x,y,p)$.\
The Grundy number is calculated recursively: $\mathcal{G}_{pass}((x,y,p) = \textit{mex}\{\mathcal{G}_{pass}(\mathbf{h}): \mathbf{h} \in \textit{move}_{pass}(x,y,p)\}.$
\[defnoft0t1pass\] Let $\mathcal{T}_0 = \{(x,y,0):(x,y) \in T_0 \}$ and $\mathcal{T}_1 = \{(x,y,0): (x,y) \in T_1\}$, where $T_i$ is defined in Definition \[defnoft0t1\].
\[lemmapass2\] $\mathcal{T}_0 = \{(x,y,0):\mathcal{G}(x,y,0) = 0 \}$.
This is a direct result of Definition \[defmovepass\], Definition \[defofgrundypass\] and Definition \[defofgrundypass\].
Therefore, the set $T_0$ that is the set of $\mathcal{P}$-positions in the classical Wythoff’s game is the set of $\mathcal{P}$-positions of the Wythoff’s game with a pass when $p=0$.
\[defnofAB\] Let $\mathcal{A} = \{ (s,t,1):(s,t) \in A\} = \{(0,1,1),(1,0,1),(2,2,1),(3,6,1),$\
$(6,3,1),(5,7,1),(7,5,1) \} $ and $\mathcal{B}= \{ (s,t,1):(s,t) \in B\}= \{(0,0,1),(1,3,1),(3,1,1),$\
$(2,5,1),(5,2,1),(6,7,1),(7,6,1) \} $, where the sets $A,B$ are defined in Definition \[defnofab\].
\[lemmaforpass1\] $(i)$ For any $(x,y,1) \in \mathcal{T}_1 \cup \mathcal{B} -\mathcal{A} $, $\textit{move}_{pass}(x,y,1) \cap (\mathcal{T}_1\cup \mathcal{B} -\mathcal{A}) = \emptyset $.\
$(ii)$ For any $(x,y,1) \notin \{(x,y,1):(x,y) \in T_0 \} \cup (\mathcal{T}_1\cup \mathcal{B} -\mathcal{A}) $, $\textit{move}_{pass}(x,y,1) \cap (\mathcal{T}_1\cup \mathcal{B} -\mathcal{A}) \neq \emptyset $.
This lemma is a direct from Lemma \[lemmapequql1\] and Definition \[defmovepass\].
\[thforqueenpass\] The set of $\mathcal{P}$-positions of Wythoff’s game with a pass is $\mathcal{P} = \mathcal{T}_0 \cup \mathcal{T}_1\cup B -A$.
$(a)$ Let $(x,y,p) \in \mathcal{P}$. We prove that $\textit{move}_{pass}(x,y,p) \cap \mathcal{P} = \emptyset$.\
$(a,1)$ If $p=0$, then $(x,y,0) \in \mathcal{T}_0$. Since $\textit{move}_{pass}(x,y,0) \subset \{(s,t,u):s=0 \}$, by $(a)$ of Lemma \[lemmafort1ti\] and Definition \[defmovepass\] $\textit{move}_{pass}(x,y,0) \cap \mathcal{P} $ $ = \textit{move}_{pass}(x,y,0) \cap \mathcal{T}_0 = \emptyset$.\
$(a.2)$ Suppose that $p=1$. If $(x,y,p) \in \mathcal{T}_1\cup \mathcal{B} -\mathcal{A}$, by Lemma \[lemmaforpass1\] and Definition \[defnoft0t1pass\] $\textit{move}_{pass}(x,y,1) \cap (\mathcal{T}_1\cup \mathcal{B} -\mathcal{A})= \emptyset$.\
$(b)$ Let $(x,y,p) \notin \mathcal{P}$.\
$(b,1)$ Suppose that $p=0$, then $(x,y,0) \notin \mathcal{T}_0$. Then $(x,y) \notin T_0$, and by $(c)$ of Lemma \[lemmafort1ti\] $\textit{move}(x,y) \cap T_0 \ne \emptyset $. Therefore $\textit{move}_{pass}(x,y,0) \cap \mathcal{T}_0 \ne \emptyset $\
$(b,2)$ Suppose that $p=1$, then $(x,y,1) \notin \mathcal{T}_1\cup \mathcal{B} -\mathcal{A}$.\
$(b.2.1)$ If $(x,y,1) \notin \{(x,y,1):(x,y) \in T_0 \}$, then by $(ii)$ of Lemma \[lemmaforpass1\] $\textit{move}_{pass}(x,y,1) \cap (\mathcal{T}_1\cup \mathcal{B} -\mathcal{A}) \neq \emptyset $.\
$(b.2.2)$ If $(x,y,1) \in \{(x,y,1):(x,y) \in T_0 \}$, then $(x,y,0) \in \mathcal{T}_0 $.\
If you start with a position $ \notin \mathcal{P} $, then by $(a)$ of this lemma
By Lemma \[p1ofqueen\] and Remark \[positionsofgrundy1\], the positions in $ \mathcal{T}_1$ are very near to the positions in $ \mathcal{T}_0$, and hence by Theorem \[thforqueenpass\] the graph of the set of $\mathcal{P}$-positions of the classical Wythoff’s game with a pass is very similar to the graph of the set of $\mathcal{P}$-positions of the classical Wythoff’s game. See Graphs in Figures \[wythoffclassic\] and \[wythoffpass\].
![The graph of the set of $\mathcal{P}$-positions of the classical Wythoff’s game[]{data-label="wythoffclassic"}](wythoffclassic.pdf){width="0.6\columnwidth"}
![The set of $\mathcal{P}$-positions of the Wythoff’s game with a pass[]{data-label="wythoffpass"}](wythoffclpass.pdf){width="0.6\columnwidth"}
[111]{} G. Nivasch:More on the Sprague Grundy function for Wythoff’s game, Games of No Chance 3 MSRI Publications Volume [**56**]{}, 2009. U.Blass and A.S.Fraenkel: The Sprague-Grundy Function for Wythoff’s Game, [*Theoretical Computer Science*]{} [**7**]{} , 311-333, (1990). W.A.Wythoff: A modification of the game of Nim, [*Nieuw Arch. Wiskd.*]{} [**7**]{} ,199-202, 1907. M.H. Albert, R.J. Nowakowski, and D. Wolfe, [*Lessons In Play*]{}, A K Peters/CRC Press, 139, 2007. A.N. Siegel, [*Combinatorial Game Theory (Graduate Studies in Mathematics)*]{}, American Mathematical Society, 2013.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We propose a novel formulation of robotic pick and place as a deep reinforcement learning (RL) problem. Whereas most deep RL approaches to robotic manipulation frame the problem in terms of low level states and actions, we propose a more abstract formulation. In this formulation, actions are target reach poses for the hand and states are a history of such reaches. We show this approach can solve a challenging class of pick-place and regrasping problems where the exact geometry of the objects to be handled is unknown. The only information our method requires is: 1) the sensor perception available to the robot at test time; 2) prior knowledge of the general class of objects for which the system was trained. We evaluate our method using objects belonging to two different categories, mugs and bottles, both in simulation and on real hardware. Results show a major improvement relative to a shape primitives baseline.'
author:
- 'Marcus Gualtieri, Andreas ten Pas, and Robert Platt[^1]'
bibliography:
- 'References.bib'
title: '**Pick and Place Without Geometric Object Models**'
---
Introduction {#sec:introduction}
============
Traditional approaches to pick-place and regrasping require precise estimates of the shape and pose of all relevant objects [@LozanoPerez1986; @Mason1985]. For example, consider the task of placing a mug on a saucer. To solve this problem using traditional techniques, it is necessary to plan a path in the combined space of the mug pose, the saucer pose, and the manipulator configuration. This requires the pose and shape of the mug to be fully observable. Unfortunately, even when the exact shape of the mug is known in advance, it can be hard to estimate the mug’s pose precisely and track it during task execution. The problem is more difficult in the more realistic scenario where the exact shape of the mug is unknown.
Approaches based on deep RL are an alternative to the model based approach described above [@Mnih2015]. Recent work has shown that deep RL has the potential to alleviate some of the perceptual challenges in manipulation. For example, Levine *et al.* showed deep learning in conjunction with policy gradient RL can learn a control policy expressed directly in terms of sensed RGB images [@Levine2016A]. Not only does this eliminate the need to develop a separate perceptual process for estimating state, but it also simplifies the perceptual problem by enabling the system to focus on only the perceptual information relevant to the specific manipulation task to be solved. This, along with encoding actions using low level robot commands (such as motor torque or Cartesian motion commands [@Levine2016B; @Levine2016A]), means the approach is quite flexible: a variety of different manipulation behaviors can be learned by varying only the reward function.
Unfortunately, deep RL approaches to robotics have an important weakness. While the convolutional layers of a deep network facilitate generalization over horizontal and vertical position in an image, they do not facilitate generalization over depth or in/out of plane orientation, i.e., the full 6-DOF pose space in which robots operate. This is a significant problem for robotics because deep RL methods must learn policies for many different relative poses between the robot and the objects in the world. Not only is this inefficient, but it detracts from the ability of the deep network to learn other things like policies that generalize well to novel object geometries.
We propose a new method of structuring robotic pick-place and regrasping tasks as a deep RL problem, i.e., as a Markov decision process (MDP). Our key idea is to formulate the problem using reach actions where the set of target poses that can be reached using these actions is sampled on each time step. Each reach action is represented by a descriptor that encodes the volumetric appearance of the scene in the local vicinity of the sampled reach target. In order to formulate the MDP, we note our problem is actually a partially observable MDP (POMDP) where object shape and pose are hidden state and the images or point clouds produced by the sensors are the observations. In order to solve this problem as an MDP, we encode belief state as a short history of recently taken reach actions expressed using the volumetric descriptors used to encode the reach action.
As a result of these innovations, our method is able to learn policies that work for novel objects. For example, we show that our system can learn to grasp novel mugs (for which prior geometric models are not available) from a pile of clutter and place them upright on a shelf as in Figure \[fig:illustratePickPlace\]. The same system can be trained to perform a similar task for other classes of objects, such as bottles, simply by retraining. Our system can also learn policies for performing complex regrasping operations in order to achieve a desired object pose. As far as we know, this is the first system described in the literature that has been demonstrated to accomplish the above without constructing or matching against geometric models of the specific objects involved.
Related Work {#sec:relatedWork}
============
One early approach to manipulation of unknown objects is based on shape primitives. Miller *et al.* explored this in the context of grasp synthesis [@Miller2003]. Others have extended these ideas to segmentation and manipulation problems [@Rusu2009; @Morwald2010; @Harada2013]. These methods have difficulty when the objects are not approximately cylindrical or cuboid and when the objects cannot be easily segmented. Our method performs much better than a cylinder-based shape primitives method, even when the objects involved (bottles and mugs) are nearly cylindrical.
Another approach to manipulating unknown objects is to estimate the object shape from a recent history of sensor feedback. For example, Dragiev and Toussaint explore an approach that models the implicit shape of an object as a Gaussian process [@Dragiev2011]. Mahler *et al.* do something similar for the purposes of grasping while incorporating tactile feedback [@Mahler2015]. These methods can run into trouble when there is not enough data to fit the implicit shape with high confidence. Both of the above approaches can be viewed as ways of estimating object shape and pose in order to facilitate traditional configuration space planning. The problem of object pose and shape estimation given various amounts of prior data remains an active area of research [@Hinterstoisser2012; @Pauwels2015; @Wohlhart2015].
Recently, there has been much advancement in grasping novel objects. Bohg *et al.* provide a survey [@Bohg2014]. Most of these methods are trained in a supervised fashion to predict whether a grasp is stable or not. The present paper can be viewed as extending our prior work in grasp detection [@Gualtieri2016; @tenPas2017] to pick-and-place and regrasping.
The approach nearest to ours is by Jiang *et al.* who propose a method for placing new objects in new place areas without explicitly estimating the object’s pose [@Jiang2012]. Their placements are sampled instead of, as in our case, fixed. However, they do not jointly reason about the grasp and the place – the grasp is predefined. This is an important drawback because the type of placement that is desired often has implications on how the grasp should be performed. Their learning method also relies on segmenting the object, segmenting the place area, hand-picked features, and human annotation for place appropriateness.
RL has long been studied for use in robot control. Kober *et al.* survey robotics applications that use RL [@Kober2013]. Since this survey, deep RL has become prominent in robotic manipulation [@Levine2016A; @Levine2016B; @Viereck2017]. These methods operate on the motor torque or Cartesian motion command level of the robot controller whereas ours operates at a higher level.
Problem Statement {#sec:problemStatement}
=================
We consider the problem of grasping, regrasping, and placing a novel object in a desired pose using a robotic arm equipped with a simple gripper. We assume this is a first-order kinematic system such that state is fully described by the geometry of the environment. Also, we assume the agent can act only by executing parameterized reach actions.
The problem can be expressed as an MDP as follows. Let $\Gamma \subseteq \mathbb{R}^3$ denote the portion of work space that is free of obstacles. For simplicity of exposition, suppose that it is known that the free space contains a finite set of $N$ rigid body objects, $O$. Let $\Lambda$ denote a parameter space that describes the space of all possible object shapes. Let $\xi(o) \in \Lambda \times SE(3)$ denote the shape and pose of object $o \in O$. Let $H \in SE(3)$ denote the current pose of the robot hand. The state of the system is fully described by the pose of the hand and the shape and pose of all $N$ objects: $s = (H, \xi(o^1), \dots, \xi(o^N)) \in S = SE(3) \times \{\Lambda \times SE(3)\}^N$.
We will assume the robot can act only by executing the following parameterized, pre-programmed actions: $\textsc{reach-grasp}(T)$ where $T \in SE(3)$ and $\textsc{reach-place}(t)$ where $t \in \textsc{place} \subset \mathbb{N}$ belongs to a discrete set of pre-programmed reach poses expressed relative to the robot base frame. The total set of available actions is then $A = SE(3) \cup \textsc{place}$.
Given a goal set $G \subset S$, we define a reward function to be $1$ when a goal state is reached and $0$ otherwise. The episode terminates either when a goal state is reached or after a maximum number of actions. Finally, we assume access to a simulator that models the effects of an action $a \in A$ taken from state $s \in S$. For stochastic systems, we assume the simulator draws a sample from the distribution of possible outcomes of an action. Given this formalization of the manipulation problem, we might express it as an MDP $\mathcal{M} = (S,A,\mathcal{T},r)$ with state-action space $S \times A$, unknown but stationary transition dynamics $\mathcal{T}$, and reward function $r(s,a) = 1$ if $s \in G$ and $r(s,a)=0$ otherwise.
A key assumption in this paper is object shape and pose are not observed directly and therefore the MDP defined above is not fully observed. Instead, it is only possible to observe a volumetric occupancy grid, $C(x) \in \{\textsc{occupied}, \textsc{free}, \textsc{unknown}\}$ where $x \in \bar{\Gamma} \subset \Gamma$ is a volumetric grid over $\Gamma$. We assume that $C$ is populated based on depth sensor data obtained during a single time step. (As such, there may be a large number of $\textsc{unknown}$ cells.) Given the above assumptions, the manipulation problem can be expressed as a POMDP $\mathcal{P} = (S,A,\mathcal{T},r,C,\mathcal{O})$, where $\mathcal{O}: S \times A \times C \rightarrow \mathbb{R}$ is the observation probabilities, assumed to be unknown. The goal of this paper is to find policies that maximize the expected sum of discounted rewards over the episode, i.e., reach the goal state in a minimum number of actions.
Approach {#sec:approach}
========
Solving the POMDP $\mathcal{P}$ using general purpose belief space solvers (e.g. [@Kurniawati2008]) is infeasible because the underlying MDP $\mathcal{M}$ is far too large to solve even if it were fully observed. Instead we propose what we call the *descriptor-based MDP* that encodes $\textsc{reach-grasp}$ actions using a special type of descriptor and encodes belief state implicitly as a short history of states and actions.
{width="0.750\columnwidth"}
\[fig:descriptor\]
The <span style="font-variant:small-caps;">reach-grasp</span> Descriptor {#sect:grasp_desc}
------------------------------------------------------------------------
The $\textsc{reach-grasp}$ descriptor is a key element of our state and action representation, based on the grasp descriptor developed in our prior work [@Gualtieri2016; @tenPas2017]. It encodes the relative pose between a robot hand and an object in terms of the portion of the volumetric occupancy grid in the vicinity of a prospective grasp. Let $\mathcal{C} = \{x \in \bar{\Gamma} | C(x) = \textsc{occupied}\}$ denote the voxelized point cloud corresponding to the occupancy grid $C$. Then, the $\textsc{reach-grasp}$ descriptor at pose $T \in SE(3)$ is $D(\mathcal{C},T) = trunc_{\gamma}(T \mathcal{C})$, where $T \mathcal{C}$ is the point cloud in the grasp reference frame, and where $trunc_{\gamma}(X)$ denotes the elements of $X$ that lie within a cuboid centered at the origin with dimensions $\gamma = (\gamma_x, \gamma_y, \gamma_z)$. This is illustrated in Figure \[fig:descriptor\]. The middle column shows the $\textsc{reach-grasp}$ descriptors corresponding to the three grasps of the object shown on the left. A $\textsc{reach-grasp}$ descriptor is encoded to the deep network as an image where the points are projected onto planes orthogonal to three different viewing directions and compiled into a single stacked image, $I(D)$, as described in [@Gualtieri2016; @tenPas2017].
The Descriptor-Based MDP {#sect:descMDP}
------------------------
Our key idea is to find goal-reaching solutions to the POMDP $\mathcal{P}$ by reformulating it as an MDP with descriptor-based states and actions. Specifically, we: 1) reparameterize the $\textsc{reach-grasp}$ action using $\textsc{reach-grasp}$ descriptors rather than 6-DOF poses; 2) redefine state as a history of the last two actions visited.
**Action representation:** Recall that the underlying MDP defines two types of actions: $\textsc{reach-grasp}(T)$ where $T$ denotes the pose of the grasp and $\textsc{reach-place}(t)$ where $t \in \textsc{place}$ and $\textsc{place}$ denotes a finite set of place poses. Since RL in a continuous action space can be challenging, we approximate the parameter space of $\textsc{reach-grasp}$ by sampling. That is, we sample a set of $m$ candidate poses for $\textsc{reach-grasp}$: $T_1, \dots, T_m \in SE(3)$. In principle, we can use any sampling method. However, since $\textsc{reach-grasp}$ is intended to reach toward *grasp* configurations, we use grasp pose detection (GPD) [@Gualtieri2016; @tenPas2017] to generate the samples. Each of the candidate poses generated by GPD is predicted to be a pose from which closing the robot hand is expected to result in a grasp (although the grasp could be of any object).
Since we are sampling candidate parameters for $\textsc{reach-grasp}$, we need a way to encode these choices to the action-value function. Normally, in RL, the agent has access to a fixed set of action choices where each choice always results in the same distribution of outcomes. However, since we are now sampling actions, this is no longer the case, and we need to encode actions to the action-value function differently. In this paper, we encode each target pose candidate for $\textsc{reach-grasp}$ by the corresponding <span style="font-variant:small-caps;">reach-grasp</span> descriptor, $D_i = D(\mathcal{C},T_i), i \in [1,m]$. Essentially, the descriptor encodes each target pose candidate by the image describing what the point cloud nearby the target pose looks like. The total action set consists of the set of descriptors corresponding to sampled reach-grasps, $\textsc{reach-grasp}(D(\mathcal{C},T_i)), i \in [1,m]$, and the discrete set of reach-places adopted from the underlying POMDP, $\textsc{reach-place}(i), i \in \textsc{place}$: $A = [1,m] \cup \textsc{place}$.
{width="0.750\columnwidth"}
\[fig:mdp\]
**State representation:** We encode state as the history of the $M$ most recent reach actions where $\textsc{reach-grasp}$ actions are represented using the corresponding descriptors. In all of our experiments, $M \leq 2$. Figure \[fig:mdp\] illustrates the resulting state-action space. The set of blue circles on the right labeled “Space of all grasp descriptors” denotes the set of states where an object has been grasped. This is a continuous space of states equal to the set of $\textsc{reach-grasp}$ descriptors resulting from the most recent $\textsc{reach-grasp}$ action, $\{trunc_\gamma(\mathcal{C}) | \mathcal{C} \subset \bar{\Gamma} \}$. The set of blue circles on the left labeled “Space of object placements” represents the set of states where an object has been placed somewhere in the environment. These states are encoded as the history of the two most recent reach actions: the $\textsc{reach-place}$ action taken on the last time step and the descriptor that encodes the $\textsc{reach-grasp}$ action taken two time steps ago. All together, a state in this new MDP is a point in $S = \{trunc_\gamma(\mathcal{C}) | \mathcal{C} \subset \bar{\Gamma} \} \times \textsc{place}$. The state labeled “Goal!” in Figure \[fig:mdp\] denotes an absorbing state where the object has been placed correctly, and the state labeled “Fell over!” denotes an absorbing state where the object has been placed incorrectly. When the agent reaches either of these states, it obtains a final reward and the episode ends.
**Reward:** Our agent obtains a reward of $+1$ when it reaches a placement state that satisfies the desired problem constraints, and otherwise, it obtains zero reward.
The Simulator
-------------
Deep RL requires such an enormous amount of experience that it is difficult to learn control policies on real robotic hardware without spending months or years in training [@Levine2016B; @Levine2016A]. As a result, learning in simulation is basically a requirement. Fortunately, our formulation of the manipulation problem in terms of pre-programmed, parameterized actions simplifies the simulations. Instead of needing to simulate *arbitrary* contact interactions, we only need a mechanism for simulating the grasp that results from executing $\textsc{reach-grasp}(T)$ and the object placement that results from executing $\textsc{reach-place}(t)$. The former can be simulated by evaluating standard grasp quality metrics [@Murray1994]. The later can be simulated by evaluating sufficient conditions to determine whether an object will fall over given the executed placement. Both are easy to evaluate in OpenRAVE [@Diankov2010], the simulator used in this work.
The Action-Value Function
-------------------------
We approximate the action-value function using the convolutional neural network (CNN) shown in Figure \[fig:cnn\]. The input is an encoding of the state and the action, and the output is a scalar, real value representing the value of that state-action pair. This structure is slightly different than that used in DQN [@Mnih2015] where the network has a number of outputs equal to the number of actions. Here, the fact that our MDP uses sampled reach actions means that we must take action as an input to the network. The action component of the input is comprised of the $\textsc{reach-grasp}$ descriptor (encoded as a $60 \times 60 \times 12$ stacked image as described in Section \[sect:grasp\_desc\]) denoting the $\textsc{reach-grasp}$ parameter and a one-hot vector denoting the $\textsc{reach-place}$ parameter. When the agent selects $\textsc{reach-grasp}$, the grasp descriptor is populated and the place vector is set to zero. When a $\textsc{reach-place}$ is selected, the grasp descriptor is set to zero and the place vector is populated.
The state component of the input is also comprised of a $\textsc{reach-grasp}$ descriptor and a place vector. However, here these two parameters encode the recent history of actions taken (Section \[sect:descMDP\]). After executing a grasp action, the grasp descriptor component of state is set to a stored version of the descriptor of the selected grasp and the place vector is set to zero. After executing a place action, the grasp descriptor retains the selected grasp and the place component is set to the just-executed place command, thereby implicitly encoding the resulting pose of the object following the place action. Each grasp image (both in the action input and the state input) is processed by a CNN similar to LeNet [@LeCun1998], except the output has 100 hidden nodes instead of 10. These outputs, together with the place information, are then concatenated and passed into two 60-unit fully connected, inner product (IP) layers, each followed by rectifier linear units (ReLU). After this there is one more inner product to produce the scalar output.
{width="0.750\columnwidth"}
\[fig:cnn\]
Learning Algorithm
------------------
Our learning algorithm is shown in Algorithm \[alg:batchSarsa\]. This is similar to standard DQN [@Mnih2015] with a couple of differences. First, we use a variant of Sarsa [@Rummery1994] rather than Q-learning because the large action branching factor makes the $\max_{a \in A} Q(s,a)$ in Q-learning expensive to evaluate and because Sarsa is known to perform slightly better on non-Markov problems. Second, we do not run a single stochastic gradient descent (SGD) step after each experience. Instead, we collect *nEpisodes* of experience before labeling the experience replay database using the most recent neural network weights. Every *nEpisodes* additional experiences, we run *nIterations* of SGD using Caffe [@Jia2014]. For the experiments in this paper, the learning algorithm is run only in simulation; although it could be used to fine-tune the network weights on the actual hardware.
[latex@errorgobble]{}
\[alg:training\]
Experiments in Simulation {#sec:simulationResults}
=========================
We performed experiments in simulation to evaluate how well our approach performs on pick-place and regrasping problems with novel objects. To do so, we obtained objects belonging to two different categories for experimentation: a set of 73 bottles and a set of 75 mugs – both in the form of mesh models from 3DNet [@Wohlkinger2012]. Both object sets were partitioned into a 75%/25% train/test split.
Experimental Scenarios
----------------------
There were three different experimental scenarios, *two-step-isolation*, *two-step-clutter*, and *multi-step-isolation*. In *two-step-isolation*, an object was selected at random from the training set and placed in a random pose in isolation on a tabletop. The goal condition was a right-side-up placement in a particular position on a table. In this scenario, the agent was only allowed to execute one grasp action followed by one place action (hence the “two-step” label). *Two-step-clutter* was the same as *two-step-isolation* except a set of seven objects was selected at random from the same object category and placed in random poses on a tabletop as if they had been physically dumped onto the table.
The *multi-step-isolation* scenario was like *two-step-isolation* except multiple picks/places were allowed for up to 10 actions (i.e., *maxTime*=10). Also, the goal condition was more restricted: the object needed to be placed upright, inside of a box rather than on a tabletop. Because the target pose was in a box, it became impossible to successfully reach it without grasping the object from the top before performing the final place (see Figure \[fig:robotExperimentIllustration\], bottom). Because the object could not always be grasped in the desired way initially, this additional constraint on the goal state sometimes forced the system to perform a regrasp in order to achieve the desired pose.
In all scenarios, point clouds were registered composites of two clouds taken from views above the object and $90^\circ$ apart: a single point cloud performs worse, presumably because features relevant for determining object pose are unobserved. In simulation, we assumed picks always succeed, because the grasp detector was already trained to recognize stable grasps with high probability [@Gualtieri2016; @tenPas2017] [^2]. A place was considered successful only if the object was placed within 3 cm of the table and 20 degrees of the vertical in the desired pose.
Algorithm Variations
--------------------
The algorithm was parameterized as follows. We used $70$ training rounds ($\textit{nTrainingRounds}=70$ in Algorithm \[alg:batchSarsa\]) for the two-step scenarios and $150$ for the multi-step scenario. We used $1,000$ episodes per training round (*nEpisodes* $=1,000$). For each training round we updated the CNN with $5,000$ iterations of SGD with a batch size of 32. *maxExperiences* was $25,000$ for the two-step scenarios and $50,000$ for the multi-step scenario. For each episode, bottles were randomly scaled in height between 10 and 20 cm. Mugs were randomly scaled in height between 6 and 12 cm. We linearly decreased the exploration factor $\epsilon$ from 100% down to 10% over the first $18$ training rounds.
We compared the performance of Algorithm \[alg:batchSarsa\] on two different types of <span style="font-variant:small-caps;">reach-grasp</span> descriptors. In the *standard* variation, we used descriptors of the standard size ($10 \times 10 \times 20$ cm). In the *large-volume* (LV) variation, we used descriptors evaluated over a larger volume ($20 \times 20 \times 40$ cm) but with the same image resolution.
We also compared with two baselines. The first was the *random baseline*, where grasp and place actions were chosen uniformly at random. The second was the *shape primitives baseline*, where object pose was approximated by segmenting the point cloud and fitting a cylinder. Although it is generally challenging to fit a shape when the precise geometry of the object to be grasped is unknown, we hypothesized that it could be possible to obtain good pick-place success rates by fitting a cylinder and using simple heuristics to decide which end should be up. We implemented this as follows. First, we segment the scene into k clusters, using k-means ($k=1$ for isolation and $k=7$ for clutter). Then we fit a cylinder to the most isolated cluster using MLESAC [@Torr2000]. We select the grasp most closely aligned with and nearest to the center of the fitted cylinder. The height of the placement action is determined based on the length of the fitted cylinder. The grasp up direction is chosen to be aligned with the cylinder half which contains fewer points. In order to get the shape primitive baseline to work, we had to remove points on the table plane from the point cloud. Although our learning methods do not require this and work nearly as well either way, we removed the table plane in all simulation experiments for consistency.
Results for the Two-Step Scenarios
----------------------------------
Figure \[fig:learningCurves\] shows learning curves for the two-step-isolation and two-step-clutter contingencies for bottles (left) and mugs (center) averaged over 10 runs. Table \[tab:twoStepResults\] shows place success rates when the test objects were used.
Several results are worth highlighting. First, our algorithm does very well with respect to the baselines. The random baseline (last row in Table \[tab:twoStepResults\]) succeeds only 2% of the time – suggesting that the problem is indeed challenging. The shape primitives baseline (where we localize objects by fitting cylinders) also does relatively poorly: it succeeds at most only 43% of the time for bottles and only 12% of the time for mugs. Second, place success rates are lower when objects are presented in clutter compared to isolation: 100% success versus 87% success rates for bottles; 84% versus 75% success for mugs. Also, if evaluation is to be in clutter (resp. isolation), then it helps to train in clutter (resp. isolation) as well: if trained only in isolation, then clutter success rates for bottles drops from 87% to 67%; clutter success rates for mugs drops from 75% to 60%. Also, using the LV descriptor can improve success rates in isolation (an increase of 84% to 91% for mugs), but hurts when evaluated in clutter: a decrease from 87% to 80% for bottles; a decrease from 75% to 70% for mugs. We suspect that this drop in performance reflects the fact that in clutter, the large receptive field of the LV descriptor encompasses “distracting” information created by other objects nearby the target object (remember we do not use segmentation) [@Mnih2014].
~Trained\ With~ / ^Tested\ With^ Bottle in Iso. Bottles in Clut.
---------------------------------- ---------------- ------------------
Isolation 1.00 0.67
Clutter 0.78 0.87
Isolation LV 0.99 0.47
Clutter LV 0.96 0.80
Shape Primitives Baseline 0.43 0.24
Random Baseline 0.02 0.02
~Trained\ With~ / ^Tested\ With^ Mug in Iso. Mugs in Clut.
Isolation 0.84 0.60
Clutter 0.74 0.75
Isolation LV 0.91 0.40
Clutter LV 0.67 0.70
Shape Primitives Baseline 0.08 0.12
Random Baseline 0.02 0.02
: Average correct placements over 300 episodes for bottles (top) and mugs (bottom) using test set, after training.[]{data-label="tab:twoStepResults"}
Results for the Multi-Step Scenario
-----------------------------------
Training for the multi-step-isolation scenario is the same as it was in the two-step scenario except we increased the number of training rounds because the longer policies took longer to learn. We only performed this experiment using mugs (not bottles) because it was difficult for our system to grasp many of the bottles in our dataset from the top. Figure \[fig:learningCurves\] shows the number of successful non-goal and goal placements as a function of training round [^3]. Initially, the system does not make much use of its ability to perform intermediate placements in order to achieve the desired goal placement, i.e., to pick up the mug, put it down, and then pick it up a second time in a different way. This is evidenced by the low values for non-goal placements (the blue line) prior to round 60. However, after round 60, the system learns the value of the non-goal placement, thereby enabling it to increase its final placement success rate to is maximum value (around 90%). Essentially, the agent learns to perform a non-goal placement when the mug cannot immediately be grasped from the top or if the orientation of the mug cannot be determined from the sensor perception. After learning is complete, we obtain an 84% pick and place success rate averaged over 300 test set trials.
Experiments on a Real Robot {#sect:real-robot}
===========================
We evaluated the same three scenarios on a real robot: *two-step-isolation*, *two-step-clutter*, and *multi-step-isolation*. As before, the two step scenarios were evaluated for both bottles and mugs, and the multi-step scenario was evaluated for only mugs. All training was done in simulation, and fixed CNN weights were used on the real robot.
The experiments were performed by a UR5 robot with 6 DOFs, equipped with a Robotiq parallel-jaw gripper and a wrist-mounted Structure depth sensor (Figure \[fig:robotExperimentIllustration\]). Two sensor views were always taken from fixed poses, $90^\circ$ apart. The object set included 7 bottles and 6 mugs, as shown in Figure \[fig:bottlesmugs\]. We used only objects that fit into the gripper, would not shatter when dropped, and had a non-reflective surface visible to our depth sensor. Some of the lighter bottles were partially filled so small disturbances (e.g., sticking to fingers) would not cause a failure. Figure \[fig:robotExperimentIllustration\] shows several examples of our two-step scenario for bottles presented in clutter.
Unlike in simulation, the UR5 requires an IK solution and motion plan for any grasp or place pose it plans to reach to. For grasps, GPD returns many grasp choices. We sort these by their pick-place Q-values in descending order and select the first reachable grasp. For places, the horizontal position on the shelf and orientation about the vertical (gravity) axis do not affect object uprightness or the height of the object. Thus, these variables were chosen to suit reachability.
\
After testing some trials on the UR5, we found we needed to adjust a couple of training/simulation parameters. First, we changed the conditions for a successful place in simulation because, during our initial experiments, we found the policy sometimes selected placements that caused the objects to fall over. As a result, we adjusted the maximum place height in simulation from 3 cm to 2 cm and changed the reward function to fall off exponentially from $+1$ for altitudes higher than 2 cm. Second, we raised the acceptance threshold [^4] used by our grasp detector, GPD [@Gualtieri2016; @tenPas2017].
1 Bottle 7 Bottles 1 Mug 6 Mugs Regrasp
------------ ---------- ----------- ------- -------- ---------
Grasp 0.99 0.97 0.96 0.93 0.94
FinalPlace 0.98 0.94 0.93 0.87 1.00
TempPlace - - - - 1.00
EntireTask 0.97 0.92 0.90 0.80 0.68
$n$ Trials 112 107 96 96 72
UpsideDown 0 4 5 10 0
Sideways 0 0 0 2 0
FellOver 2 2 1 0 0
$t>10$ - - - - 12
: (Top) Success rates for grasp, temporary place, final place, and entire task. (Bottom) Placement error counts by type. Results are averaged over the number of trials (middle).[]{data-label="tab:robotResultsSummary"}
Table \[tab:robotResultsSummary\] summarizes the results from our robot experiments. We performed 483 pick and place trials over five different scenarios. Column one of Table \[tab:robotResultsSummary\] shows results for pick and place for a single bottle presented in isolation averaged over all bottles in the seven-bottle set. Out of 112 trials, 99% of the grasps were successful and 98% of the placements were successful, resulting in a complete task pick/place success rate of 97%. Column two shows similar results for the bottles-in-clutter scenario, and columns three and four include results for the same experiments with mugs. Finally, column five summarizes results from the multi-step-isolation scenario for mugs: overall, our method succeeded in placing the mug upright into the box 68% of the time. The temporary place success is perfect because a temporary placement only fails if the mug is so high it rolls away after dropped or too low it is pushed into the table, neither of which ever happened after 72 trials. The final placement is perfect because it always did get the orientation right (for all 72 trials that got far enough to reach the final placement), and it is hard for the mug to fall over in the box. The multi-step scenario has low task success rate because 12 trials failed to perform the final place after $10$ time steps. Perhaps this is due to lower Q-function values on the real system (due to domain transfer issues), causing the robot to never become confident enough with its given state information to perform the final place.
Our experimental results are interesting for several reasons beyond demonstrating that the method can work. First, we noticed consistently lower place performance for the mug category relative to the bottle category. The reason for this is there is more perceptual ambiguity involved in determining the orientation of a mug compared to that of a bottle. In order to decide which end of a mug is “up”, it is necessary for the sensor to view into at least one end of the mug. Second, the robot had trouble completing the multi-step task in a reasonable number of steps with the real hardware compared with simulation. This may be because fewer grasps are available on the real robot versus the simulated robot due to collision modelling. Another unexpected result was our learned policies typically prefer particular types of grasps, e.g., to grasp bottles near the bottom (see Figure \[fig:robotExperimentIllustration\]). We suspect this is a result of the link between the location of a selected grasp and the grasp descriptor used to represent state. In order to increase the likelihood that the agent will make high-reward decisions in the future, it selects a grasp descriptor that enables it to easily determine the pose of the object. In the case of bottles, descriptors near the base of the bottle best enable it to determine which end is “up”.
Conclusion and Future Work
==========================
This paper proposes a new way of structuring robotic pick-place and regrasping tasks as a deep RL problem. Importantly, our learned policies can place objects very accurately without using shape primitives or attempting to model object geometry in any way. Our key insight is to encode a sampled set of end-to-end reaching actions using descriptors that encode the geometry of the reach target pose. We encode state as a history of recent actions and observations. The resulting policies, which are learned in simulation, simultaneously perceive relevant features in the environment and plan the appropriate grasp and place actions in order to achieve task goals. Our experiments show that the method consistently outperforms a baseline method based on shape primitives.
For future work, we plan to generalize the descriptor-based MDP in two ways. First, place poses could be sampled from a continuous, 6-DOF space, as grasps are. To do this we would develop a special purpose place detector in the same way GPD is a grasp detector. Second, the system should be able to work with a more diverse set of objects, e.g., kitchen items. This may require a CNN with more capacity and longer training time, motivating innovations to speed up the learning in simulation.
[^1]: College of Computer and Information Science, Northeastern University, Boston, Massachusetts, USA. This work has been supported in part by NSF through IIS-1427081 and ONR through N000141410047.
[^2]: It is possible to train grasping from the same reward signal, but this would require longer simulations. Empirically, this assumption did not lead to many grasp failures on the real robot (see Section \[sect:real-robot\]).
[^3]: Non-goal placements were considered successful if the object was 3 cm or less above the table. Any orientation was allowed. Unsuccessful non-goal placements terminate the episode.
[^4]: GPD outputs a machine-learned probability of a stable (i.e., force closure) grasp. The *threshold* is the grasp stability probability above which grasps are accepted.
|
{
"pile_set_name": "ArXiv"
}
|
= 0.6cm
0.8cm
[**Bulk-Boundary Propagator in Liouville Theory on a Disc**]{}
0.6cm
Kazuo Hosomichi[^1]\
=3.5ex
[**Abstract**]{}
We study Liouville theory on worldsheets with boundary using the solutions of Knizhnik-Zamolodchikov equation involving a degenerate representation of the Virasoro algebra. The expression for bulk-boundary propagator on a disc is proposed.
August 2001
= 0.6cm
Liouville theory is an interacting two-dimensional conformal field theory that has many applications in string theory. It was originally studied as a theory of two-dimensional gravity, and recently it has also applied to the study of strings propagating near the singularities of manifolds. It plays a key role in understanding the holography in string theory, since most of the known backgrounds realizing the holography have a radial direction which is described by a linear-dilaton like theory. To study the property of D-branes on such backgrounds from the CFT viewpoint, it is therefore necessary to understand the Liouville theory on open worldsheets.
Liouville theory on open worldsheets was analyzed in [@FZZ; @ZZ] where some basic correlators were derived. They were based on bootstraps and free field techniques which were first developed in [@DO; @ZZ2; @T] for the case without boundary. See also [@PT; @T2; @T3]. Our analysis in this letter is along these path.
Liouville theory on a worldsheet with boundary is defined by the action $$\begin{aligned}
I&=&\frac{1}{8\pi}\int_\Sigma d^2\sigma \sqrt{g}
\left[g^{mn}\partial_m\phi\partial_n\phi
+\sqrt{2}QR\phi+8\pi\mu e^{\sqrt{2}b\phi}\right]
\nonumber \\ &&
+\frac{1}{4\pi}\int_{\partial\Sigma}d\xi g^{1/4}
\left[\sqrt{2}QK\phi+4\pi\mu_B e^{b\phi/\sqrt{2}}\right] ,\end{aligned}$$ where $K$ is the curvature of the boundary and $\mu_B$ is called the boundary cosmological constant. $\mu_B$ can take different values for each connected component of the boundary and, what is more, $\mu_B$ can be different for each boundary segment bounded by two boundary vertex operators.
When $\mu$ and $\mu_B$ vanish, the theory reduces to a free theory with the stress tensor $$T = -\frac{1}{2}(\partial\phi\partial\phi-\sqrt{2}Q\partial^2\phi)$$ and the center $c=1+6Q^2$.
We have two kinds of primary fields in Liouville theory with boundary. Bulk primaries $V_\alpha(z)=e^{\sqrt{2}\alpha\phi(z)}$ correspond to closed string modes while boundary primaries $B_\alpha(x)=e^{\alpha\phi(x)/\sqrt{2}}$ correspond to open string modes. They are of conformal weight $h_\alpha=\alpha(Q-\alpha)$. We mainly use coordinates $z,w,\ldots$ for positions of bulk fields and $x,y,\cdots$ for boundary fields.
As was studied in [@FZZ; @ZZ], it is natural to regard $\mu_B$ as labeling D-branes or Cardy states in Liouville theory. The classification of Cardy states has been done in [@ZZ] using the modular property of Virasoro characters. Consistent Cardy states are expressed as linear combination of Ishibashi states, and the coefficients are proportional to the one-point function of a bulk field ${\left<{V_\alpha(z)}\right>}$ on a disc. This was obtained in [@FZZ]: $$\begin{aligned}
{\left<{V_\alpha(z)}\right>}_{\mu_B} &=& U(\alpha;s)|z-\bar{z}|^{-2h_\alpha},
\nonumber \\
U(\alpha;s) &=& 2 \cosh\left[2\pi s(2\alpha-Q)\right]
\{\mu\pi\gamma(b^2)\}^{(Q-2\alpha)/2b}
(2\alpha-Q)\Gamma[(2\alpha-Q)b]\Gamma[(2\alpha-Q)/b],
\label{U}\end{aligned}$$ where $\gamma(x)=\Gamma(x)/\Gamma(1-x)$ and we introduced a new parameter $s$ for labeling the Cardy states ${\left|{s}\right>_{\rm \!c}}$. It is related to $\mu_B$ via $$\mu_B = \cosh(2\pi sb)\left(\frac{\mu}{\sin\pi b^2}\right)^{1/2}.
\label{bcc-p}$$
Our goal is to obtain the expression for bulk-boundary propagator ${\left<{V_\alpha(z)B_\beta(x)}\right>}_s$ on a disc. What we have to determine is the structure constant $R(\alpha,\beta;s)$ defined by $${\left<{V_\alpha(z)B_\beta(x)}\right>}_{s}=
|z-\bar{z}|^{h_\beta-2h_\alpha}|z-x|^{-2h_\beta}R(\alpha,\beta;s).$$ To obtain this we use the ($1+2$)-point function with an auxiliary insertion of a degenerate operator $B_{-b/2}(y)$. Conformal invariance restricts the correlator to take the form $$\begin{aligned}
\lefteqn{
{\left<{V_\alpha(z)B_\beta(x)B_{-b/2}(y)}\right>}^{\rm (upper~half~plane)}_{s_L,s_R}
} \nonumber \\
&=& (x-z)^{-h_\beta-h_{-b/2}}(x-\bar{z})^{-h_\beta+h_{-b/2}}
|z-\bar{z}|^{h_\beta-2h_\alpha+h_{-b/2}}(y-\bar{z})^{-2h_{-b/2}}
\nonumber \\
& &\times {\left<{V_\alpha(0)B_\beta(1)B_{-b/2}(\eta)}\right>}^{\rm (disc)}_{s_L,s_R},
~~~~\eta={{\textstyle\frac{(y-z)(x-\bar{z})}{(y-\bar{z})(x-z)}}}\end{aligned}$$ This correlator has two boundary segments since there are two boundary operator insertions. The suffices $L,R$ mean the left or right of $B_{-b/2}$ on the real axis. The two boundary cosmological constants must satisfy $s_L\pm s_R = \pm' ib/2$ so as to be consistent with the fusion rule[@ZZ]. Otherwise the null vector actually does not decouple.
The $(1+2)$-point function can be expressed in a number of ways: $$\begin{aligned}
\lefteqn{
{\left<{V_\alpha(0)B_\beta(1)B_{-b/2}(\eta)}\right>}^{\rm (disc)}_{s_L,s_R}
}\nonumber \\
&=& c_+(\beta;s_L,s_R)R(\alpha,\beta-{{\textstyle\frac{b}{2}}};s_R)
G_{\alpha,\beta}(\eta)e^{i\pi b\beta/2}
\nonumber \\ && \!\!\!\!\!\!
+~ c_-(\beta;s_L,s_R)R(\alpha,\beta+{{\textstyle\frac{b}{2}}};s_R)
G_{\alpha,Q-\beta}(\eta)e^{i\pi b(Q-\beta)/2}
\nonumber \\
&=& c_+(\beta;s_R,s_L)R(\alpha,\beta-{{\textstyle\frac{b}{2}}};s_L)
G_{\alpha,\beta}(\eta e^{-2\pi i})e^{-i\pi b\beta/2}
\nonumber \\ &&\!\!\!\!\!\!
+~ c_-(\beta;s_R,s_L)R(\alpha,\beta+{{\textstyle\frac{b}{2}}};s_L)
G_{\alpha,Q-\beta}(\eta e^{-2\pi i})e^{-i\pi b(Q-\beta)/2},\end{aligned}$$ where the equality should hold at least up to overall factors. Here $G_\beta(\eta)$ is the following solution of KZ equation: $$G_{\alpha,\beta}(\eta) =\eta^{b\alpha}(1-\eta)^{b\beta}
F\left(b(2\alpha+\beta-Q-{{\textstyle\frac{b}{2}}}), b(\beta-{{\textstyle\frac{b}{2}}}),
b(2\beta-b);1-\eta\right).$$ We assume that the correlator is analytic on the complex $\eta$-plane except on the positive half of the real axis. The expression $\eta e^{-2\pi i}$ indicates going clockwise around $\eta=0$ once. $c_\pm(\beta;s_L,s_R)$ are the OPE coefficients: $$B_\beta(x)B_{-b/2}(y)_{s_L,s_R} \stackrel{x<y}{\sim}
|y-x|^{b\beta}c_+(\beta;s_L,s_R)B_{\beta-b/2}(x)
+ |y-x|^{b(Q-\beta)}c_-(\beta;s_L,s_R)B_{\beta+b/2}(x)$$ They are calculated as the free field correlators with an appropriate insertion of boundary screening operator $S_B = \int_{\partial\Sigma}B_b$: $$\begin{aligned}
c_+(\beta;s_L,s_R)
&=& {\left<{B_\beta B_{-b/2}B_{Q-\beta+b/2}}\right>}_{\rm free}
~=~ 1,
\nonumber \\
c_-(\beta;s_L,s_R)
&=& {\left<{(-\mu_BS_B)B_\beta B_{-b/2}B_{Q-\beta+b/2}}\right>}_{\rm free}
\nonumber \\
&=& \frac{2b^2}{\pi}\Gamma(1-2b\beta)\Gamma(2b\beta-bQ)
\cos\pi(b\beta-{{\textstyle\frac{bQ}{2}}})(\mu\pi\gamma(b^2))^{1/2}
\cos\pi(b\beta-{{\textstyle\frac{bQ}{2}}}\mp 2ibs_R)
\nonumber \\
& & ~~~~(s_R = s_L \pm {{\textstyle\frac{ib}{2}}}).\end{aligned}$$
We can derive some recursion relations between $R(\alpha,\beta;s)$ from the transformation property and monodromy of hypergeometric functions. We shall do this by rewriting the $(1+2)$-point function as a linear sum of the solutions $H_{\alpha,\beta}(\eta)$ and $H_{Q-\alpha,\beta}(\eta)$ that diagonalize the monodromy around $\eta=0$. They are defined as $$H_{\alpha,\beta}(\eta)=\eta^{b\alpha}(1-\eta)^{b\beta}
F\left(b(2\alpha+\beta-Q-{{\textstyle\frac{b}{2}}}),b(\beta-{{\textstyle\frac{b}{2}}}),
b(2\alpha-b);\eta\right)$$ and are related to $G_{\alpha,\beta}$ via $$\begin{aligned}
(G_{\alpha,\beta},G_{\alpha,Q-\beta})
&=&(H_{\alpha,\beta},H_{Q-\alpha,\beta})
{\left[\!\!\begin{array}{cc}x_{\beta,\alpha}&x_{Q-\beta,\alpha}\\#3&x_{Q-\beta,Q-\alpha}\end{array}\!\!\right]}, \nonumber \\
x_{\beta,\alpha} &=&
\frac{\Gamma(2b\beta-b^2)\Gamma(1+b^2-2b\alpha)}
{\Gamma(1+\frac{b^2}{2}-2b\alpha+b\beta)
\Gamma(b\beta-\frac{b^2}{2})}.\end{aligned}$$ There are subtle phase factors arising in deriving the recursion relation, which should be handled carefully so that the resultant relation is symmetric under the exchange of $s_L$ and $s_R$. We must also pay attention to the relation $${\left<{V_\alpha(0)B_\beta(1)B_{-b/2}(\eta)}\right>}^{\rm (disc)}_{s_L,s_R}
=\eta^{-2h_{-b/2}}
{\left<{V_\alpha(0)B_\beta(1)B_{-b/2}(\eta^{-1})}\right>}^{\rm (disc)}_{s_R,s_L},$$ which merely corresponds to flipping the real axis of the upper half-plane. We finally find the following relation: $$iX^+{\left[\!\!\begin{array}{c}c_+(\beta;s_R,s_L)R(\alpha,\beta-{{\textstyle\frac{b}{2}}};s_L)\\#2\end{array}\!\!\right]}
= X^-{\left[\!\!\begin{array}{c}c_+(\beta;s_L,s_R)R(\alpha,\beta-{{\textstyle\frac{b}{2}}};s_R)\\#2\end{array}\!\!\right]}
\label{recur}$$ where the components of $X^\pm$ are given by $$x^\pm_{\beta,\alpha}=
e^{\pm i\pi b(\alpha+\frac{\beta}{2}-\frac{3Q}{4})}x_{\beta,\alpha}.$$ The phases are fixed so that we can find a matrix $Y$ satisfying $YX^\mp = \pm iX^\pm$, ensuring that the relation is symmetric under the exchange of $s_L$ and $s_R$.
Let us pose a while and see the consistency of the relation (\[recur\]) with the reflection symmetry of the vertices. The structure constant $R(\alpha,\beta;s)$ should be invariant under the reflection of the bulk vertex operator $V_\alpha \Leftrightarrow V_{Q-\alpha}$: $$V_\alpha=D(\alpha)V_{Q-\alpha},~~~
D(\alpha)= -[\mu\pi\gamma(b^2)]^{(Q-2\alpha)/b}
\frac{\Gamma((2\alpha-Q)b)\Gamma((2\alpha-Q)/b)}
{\Gamma(-(2\alpha-Q)b)\Gamma(-(2\alpha-Q)/b)}.$$ and the reflection of the boundary operator $B_\beta \Leftrightarrow B_{Q-\beta}$: $$\begin{aligned}
B_\beta &=& d(\beta;s,s) B_{Q-\beta},\nonumber \\
d(\beta;s_L,s_R) &=&
\frac{{{\bf G}}(Q-2\beta){{\bf G}}(2\beta-Q)^{-1}
\{\mu\pi\gamma(b^2)b^{2-2b^2}\}^{(Q-2\beta)/2b}}
{{{\bf S}}(\beta+is_L+is_R){{\bf S}}(\beta-is_L-is_R)
{{\bf S}}(\beta+is_L-is_R){{\bf S}}(\beta-is_L+is_R)}.
\label{bref}\end{aligned}$$ Here we used some of the functions $\Upsilon$, ${{\bf S}}$, ${{\bf G}}$ introduced in [@DO; @ZZ2; @FZZ]. They are characterized by the shift relations $$\Upsilon(x+b)=b^{1-2bx}\gamma(bx)\Upsilon(x),~~~~
{{\bf S}}(x+b)=2\sin(\pi bx){{\bf S}}(x), ~~~~
{{\bf G}}(x+b)=\frac{b^{\frac{1}{2}-bx}}{\sqrt{2\pi}}\Gamma(bx){{\bf G}}(x),$$ and those with $b$ replaced with $1/b$. We can check that if we can find a solution $R(\alpha,\beta;s)$ of the recursion relation, then its reflections $$D(\alpha)R(Q-\alpha,\beta;s),~~~
d(\beta;s,s)R(\alpha,Q-\beta;s)$$ are also solutions of the same recursion relation.
By inserting $s_R=s_L+{{\textstyle\frac{ib}{2}}}$ into (\[recur\]) we obtain partial difference equations for $R(\alpha,\beta;s)$. To write down the solution, it seems that the most efficient way is to take the Fourier transform: $$\tilde{R}(\alpha,\beta;p)
= {{\textstyle\frac{1}{2}}}\int_{-\infty}^{\infty} ds e^{4\pi sp}R(\alpha,\beta;s),~~~~
R(\alpha,\beta,s)
= -i\int_{-i\infty}^{i\infty}dp e^{-4\pi sp}\tilde{R}(\alpha,\beta,p),$$ in terms of which the difference equation becomes simply $$\begin{aligned}
&&
\tilde{R}(\alpha,\beta;p+b)
\sin\pi b(p+\alpha-{{\textstyle\frac{\beta}{2}}}+{{\textstyle\frac{Q}{2}}})
\sin\pi b(p-\alpha-{{\textstyle\frac{\beta}{2}}}+{{\textstyle\frac{3Q}{2}}})
\nonumber \\ &=&
\tilde{R}(\alpha,\beta;p)
\sin\pi b(p+\alpha+{{\textstyle\frac{\beta}{2}}}-{{\textstyle\frac{Q}{2}}})
\sin\pi b(p-\alpha+{{\textstyle\frac{\beta}{2}}}+{{\textstyle\frac{Q}{2}}})
\nonumber \\ &=&
\tilde{R}(\alpha,\beta+b;p+{{\textstyle\frac{b}{2}}})
\frac{\pi^2 b^2(\mu\pi\gamma(b^2))^{1/2}
\Gamma(1-2b\beta)\Gamma(1-b^2-2b\beta)}
{\Gamma(b\beta)\Gamma(1-b\beta)^3
\Gamma(1-b\beta-2b\alpha+bQ)\Gamma(1-b\beta+2b\alpha-bQ)}.\end{aligned}$$ A solution is given by $$\begin{aligned}
\tilde{R}(\alpha,\beta;p) &=&
2\pi(\mu\pi\gamma(b^2)b^{2-2b^2})^{(Q-2\alpha-\beta)/2b}
\frac{{{\bf G}}(Q){{\bf G}}(\beta){{\bf G}}(Q-2\beta)\Upsilon(2\alpha)}
{{{\bf G}}(Q-\beta)^3{{\bf G}}(2Q-2\alpha-\beta){{\bf G}}(2\alpha-\beta)}
\nonumber \\ && \times
{{\bf S}}( p+{{\textstyle\frac{\beta+2\alpha-Q}{2}}})
{{\bf S}}( p+{{\textstyle\frac{\beta-2\alpha+Q}{2}}})
{{\bf S}}(-p+{{\textstyle\frac{\beta+2\alpha-Q}{2}}})
{{\bf S}}(-p+{{\textstyle\frac{\beta-2\alpha+Q}{2}}}).
\label{Rtil}\end{aligned}$$ The normalization was fixed so that we have $R(\alpha,0;s)=U(\alpha;s)$.
We have to analyze the transformation property of the above solution under the reflection (\[bref\]). To do this we check whether or not the value of $R(\alpha,\beta;s)$ at $\beta=Q$ is as required from the reflection symmetry: $$d(\beta;s,s)\times(-i)\int dp e^{-4\pi sp}\tilde{R}(\alpha,Q-\beta;p)
\stackrel{\beta\to 0}{\rightarrow} U(\alpha;s).$$ In doing this we have to be careful for the fact that when a factor like $e^{4\pi sp_0}$ is multiplied onto the inverse Fourier transform ($p$-integration), it shifts the contour of $p$-integration and possibly picks up some poles. A calculation shows that our structure constant satisfies the above equation, ensuring the reflection symmetry at $\beta=0$.
As we have seen above, the solution was obtained by solving the recursion relation (\[recur\]) and imposing the reflection symmetry. The normalization was fixed by matching with $U(\alpha;s)$ in the limit $\beta\rightarrow 0$. Any solution satisfying these conditions should be equal to the above $R(\alpha,\beta;s)$ for all $\beta\in\{b{\bf Z}+b^{-1}{\bf Z}\}$ due to recursion relation (\[recur\]) and that associated with the degenerate field $B_{-1/2b}$. In this sense the solution is unique.
It is known from the perturbative analysis based on path-integral formalism that the correlators diverge when the non-conservation of Liouville momentum can be screened by a non-negative integer number of screening operators. Our $R(\alpha,\beta;s)$ indeed has the corresponding poles at $2\alpha+\beta=Q-nb$. For the first few of them we can also check that the residue is precisely equal to the free-field correlators as expected.
If a bulk operator $V_\alpha$ approaches the boundary, it can be expanded as a sum over boundary fields $B_\beta$. One naively expects that the expansion obeys the following formula $$\begin{aligned}
V_{\alpha}(z)&\stackrel{z\rightarrow x}{\sim}&
{{\textstyle\frac{1}{2}}} \int d\beta
|z-\bar{z}|^{h_\beta-2h_\alpha} R(\alpha,\beta;s)B_{Q-\beta}(x)
\nonumber \\ &=&
{{\textstyle\frac{1}{2i}}}\int d\beta dp
|z-\bar{z}|^{h_\beta-2h_\alpha} e^{-4\pi sp}
\tilde{R}(\alpha,\beta;p)B_{Q-\beta}(x).
\label{VBexp}\end{aligned}$$ where the $\beta$-integration should be done over $\beta\in \frac{Q}{2}+i{\bf R}$ for $\alpha\in \frac{Q}{2}+i{\bf R}$, and suitably deformed for generic $\alpha$ in order to ensure the analyticity. Similar deformation of contour is also assumed for $p$-integration. As a special case, take as $\alpha$ those values corresponding to degenerate representation of the Virasoro algebra: $$2\alpha_{k,l}=Q-kb-lb^{-1},~~~(k,l\in {\bf Z}_{\ge 1}).$$ We then expect that $V_\alpha$ is expanded into boundary degenerate operators $B_{\beta_{m,n}}$ with $$2\beta_{m,n}=Q-mb-nb^{-1},~~~
(m = 1,3,\cdots, 2k-1,~~~
n = 1,3,\cdots, 2l-1).
\label{degB}$$ Our $R(\alpha,\beta;s)$ agrees with this expectation. To see this, note first that the integral over $\beta$ and $p$ has only contribution from poles due to the vanishing factor $\Upsilon(2\alpha)$ in the integrand. In order to cancel this factor, we actually need the degeneration of two poles that pinch the integration contour. By analyzing the location and the degeneracy of the poles of $$\begin{aligned}
\lefteqn{ \tilde{R}(\alpha,\beta;s)d(Q-\beta;s,s) } \nonumber \\
&\sim& \int dp
\frac{{{\bf G}}(2\beta-Q)\Upsilon(2\alpha)
{{\bf S}}( p+{{\textstyle\frac{\beta+2\alpha-Q}{2}}})
{{\bf S}}( p+{{\textstyle\frac{\beta-2\alpha+Q}{2}}})
{{\bf S}}(-p+{{\textstyle\frac{\beta+2\alpha-Q}{2}}})
{{\bf S}}(-p+{{\textstyle\frac{\beta-2\alpha+Q}{2}}})}
{{{\bf G}}(\beta){{\bf G}}(Q-\beta){{\bf G}}(2Q-2\alpha-\beta){{\bf G}}(2\alpha-\beta)},\end{aligned}$$ we find that the degenerate bulk field $V_{\alpha_{k,l}}$ is expanded into degenerate boundary fields $B_{\beta_{m,n}}$ with $(m,n)$ precisely given in (\[degB\]). Note that the reflection coefficient $d(\beta;s,s)$ becomes singular when $\beta$ belongs to a degenerate representation. Therefore, only $B_{\beta_{m,n}}$ (and not their reflection transform) can appear with finite coefficient in the expansion of $V_{\alpha_{k,l}}$.
The Fourier transformed structure constant $\tilde{R}(\alpha,\beta;p)$ may well be thought of as the fusion coefficient of an Ishibashi state ${|{p}\rangle\!\rangle}$ with a boundary primary $B_\beta$. It seems therefore reasonable that the bulk-boundary structure constant becomes simpler if we take Ishibashi states rather than Cardy states. In [@ZZ] the Cardy states ${\left|{m,n}\right>_{\rm \!c}}$ corresponding to degenerate representations of Virasoro algebra have also been constructed and analyzed in some detail. From the comparison of the wave functions for Cardy states ${\left|{s}\right>_{\rm \!c}}$ and ${\left|{m,n}\right>_{\rm \!c}}$, it is expected that the bulk-boundary structure constant for degenerate Cardy states might be given by $$R(\alpha,\beta)_{m,n} = 2i\int_{-i\infty}^{i\infty}dp
\sin(2\pi mpb)\sin(2\pi np/b)\tilde{R}(\alpha,\beta;p),$$ with $\tilde{R}(\alpha,\beta;p)$ given in (\[Rtil\]). On the other hand, it follows from the fusion rule that only degenerate boundary operators $B_{\beta_{r,s}}$ with $$r=1,3,\cdots, 2m-1,~~~~ s=1,3,\cdots, 2n-1$$ can appear on a degenerate Cardy state ${\left|{m,n}\right>_{\rm \!c}}$. To prove the consistency of all these we need further study of boundary degenerate operators and degenerate Cardy states.
Of all the basic structure constants in Liouville theory on a disc, it remains to calculate the three-point function of boundary operators. The boundary degenerate operator $B_{-b/2}$ and the associated KZ equation will be useful also there.
The author thanks the organizer and participants of the Workshop on Field Theory at Otaru (Japan) where he initiated this work. The work of the author is supported in part by JSPS Research Fellowships for Young Scientists.
[References]{}
40004000 ‘=1000=
V. Fateev, A. Zamolodchikov and Al. Zamolodchikov, [*“Boundary Liouville Field Theory I. Boundary State and Boundary Two-point Function”*]{}, [hep-th/0001012]{}. A. Zamolodchikov and Al. Zamolodchikov, [*“Liouville field theory on a pseudosphere”*]{}, [hep-th/0101152]{}. H. Dorn and H. J. Otto, [*“Two and three point functions in Liouville theory”*]{}, Nucl. Phys. B [**429**]{}, 375 (1994), [hep-th/9403141]{}. A. Zamolodchikov and Al. Zamolodchikov, [*“Structure constants and conformal bootstrap in Liouville field theory”*]{}, Nucl. Phys. B [**477**]{}, 577 (1996), [hep-th/9506136]{}. J. Teschner, [*“On the Liouville three point function”*]{}, Phys. Lett. B [**363**]{}, 65 (1995), [hep-th/9507109]{}. B. Ponsot and J. Teschner, [*“Liouville bootstrap via harmonic analysis on a noncompact quantum group”*]{}, [hep-th/9911110]{}. J. Teschner, [*“Remarks on Liouville theory with boundary”,*]{} [hep-th/0009138]{}. J. Teschner, [*“Liouville theory revisited”,*]{} [hep-th/0104158]{}.
[^1]: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
[**The set of autotopisms of partial Latin squares.**]{}
[Falcón, R. M.]{}
[Department of Applied Mathematics I.\
School of Building Engineering. University of Seville. Avda. Reina Mercedes 4 A, 41012 - Seville, Spain. E-mail: [*[email protected]*]{}]{}
[**Abstract**]{}
Symmetries of a partial Latin square are determined by its autotopism group. Analogously to the case of Latin squares, given an isotopism $\Theta$, the cardinality of the set $\mathcal{PLS}_{\Theta}$ of partial Latin squares which are invariant under $\Theta$ only depends on the conjugacy class of the latter, or, equivalently, on its cycle structure. In the current paper, the cycle structures of the set of autotopisms of partial Latin squares are characterized and several related properties studied. It is also seen that the cycle structure of $\Theta$ determines the possible sizes of the elements of $\mathcal{PLS}_{\Theta}$ and the number of those partial Latin squares of this set with a given size. Finally, it is generalized the traditional notion of partial Latin square completable to a Latin square.
[**MSC 2000:**]{} 05B15, 20N05, 20D45.
[**Keywords:**]{} Partial Latin square, autotopism group, cycle structure.
Introduction.
=============
Every permutation $\pi$ of the symmetric group $S_n$ can be univocally decomposed into product of disjoint cycles. Let $n_{\pi}$ be the number of these cycles. The numbers $\lambda_i^{\pi}$ of cycles of length $i$ in this decomposition determine its [*cycle structure*]{} as the expression $z_{\pi}=n^{\lambda_n^{\pi}}\cdot...\cdot 2^{\lambda_2^{\pi}}\cdot 1^{\lambda_1^{\pi}}$, where any term of the form $i^0$ is omitted and any term of the form $i^1$ is replaced by $i$. The cardinality of the set $\mathcal{CS}_n$ of possible cycle structures of $S_n$ is equal to the number $p(n)$ of partitions of $n$. Two permutations are conjugate if and only if they have the same cycle structure. Given $\pi\in S_n$, let $\lambda_{\pi}$ and $\pi_{\infty}$ be respectively its length and the union of its $1$-cycles written in natural order. Hereafter, we will suppose $\pi$ to be represented by following their univocal decomposition into a product $\pi_1\pi_2...\pi_{n_{\pi}}$ of disjoint cycles in order of decreasing length, where each cycle $\pi_i$ is written as $(p_{i,1}p_{i,2}...p_{i,\lambda_{\pi_i}})$, with $p_{i,1}=\min_j\{p_{i,j}\}$ and where $p_{i,1}<p_{j,1}$ whenever $i<j$ and $\lambda_{\pi_i}=\lambda_{\pi_j}$. Finally, given $a\in [n]=\{1,2,...,n\}$, we will write $a\in\pi_i$ if there exists $j\in [\lambda_{\pi_i}]$ such that $a=p_{i,j}$. Analogously, $a\in\pi_{\infty}$ will mean $\pi(a)=a$.
A [*Latin square*]{} of order $n$ is an $n \times n$ array with elements chosen from a set of $n$ distinct symbols such that each symbol occurs precisely once in each row and each column. Hereafter, $[n]$ will be assumed to be this set of symbols and $\mathcal{LS}_n$ will denote the set of Latin squares of order $n$. Given $L=(l_{rc})\in \mathcal{LS}_n$, its [*orthogonal representation*]{} $O(L)$ is the set of $n^2$ triples $\{(r,c,l_{rc})\,\mid\, r,c\in [n]\}$ defined by the rows $r$, columns $c$ and symbols $l_{rc}$ of $L$. This set verifies the [*Latin square condition*]{}, i.e., given two triples of $O(L)$ which coincide in two components, then the third component is also the same. Given $\pi\in S_3$, it is defined the Latin square $L^{\pi}$ such that $O(L^{\pi})=\{(l_{\pi(1)},l_{\pi(2)},l_{\pi(3)})\,\mid\, (l_1,l_2,l_3)\in O(L)\}$, which is said to be [*parastrophic*]{} to $L$. Permutations of rows, columns and symbols also give rise to new Latin squares. Specifically, given three permutations $\alpha,\beta,\gamma$ of the symmetric group $S_n$, the triple $\Theta=(\alpha,\beta,\gamma)\in \mathfrak{I}_n=S_n^3$ is an [*isotopism*]{} of Latin squares and $L^{\Theta}$ is said to be [*isotopic*]{} to $L$, where $O(L^{\Theta})=\{(\alpha(r),\beta(c),\gamma(s))\,\mid\, (r, c, s)\in O(L)\}$. To be isotopic is an equivalence relation, which will be denoted by $\sim$, and the set of Latin squares isotopic to $L$ is its [*isotopism class*]{} $[L]$. The number of Latin squares and isotopism classes of $\mathcal{LS}_n$ are known for $n\leq 11$ [@McKay05; @Hulpke11]. A list of representatives of isotopism classes for $n\leq 8$ is given in [@McKay]. The [*cycle structure*]{} of $\Theta$ is the triple $z_{\Theta}=(z_{\alpha},z_{\beta},z_{\gamma})$. Hereafter, given a subset $S\subseteq \mathfrak{I}_n$, $\mathcal{CS}_S$ will denote the set of cycle structures of the elements of $S$. Given $z=(z_1,z_2,z_3)\in \mathcal{CS}_S$, where $z_i=n^{z_{in}}\cdot...\cdot 2^{z_{i2}}\cdot 1^{z_{i1}}$, then $n_{z_i}$ will denote the number of cycles of $z_i$, i.e., $n_{z_i}=\sum_{j\in [n]} z_{ij}$. Finally, the [*parastrophic class*]{} of $z$ is the set $[z]=\{z^{\pi}=(z_{\pi(1)},z_{\pi(2)},z_{\pi(3)}\,\mid\, \pi\in S_3\}$.
If $L^{\Theta}=L$, then $\Theta$ is said to be an [*autotopism*]{} of $L$. If $\alpha=\beta=\gamma$, then $\Theta$ is an [*automorphism*]{} of $L$ and $\Theta=\alpha$ is written instead of $(\alpha,\alpha,\alpha)$. Let $\mathcal{LS}_\Theta, \Delta(\Theta)$, $\mathfrak{A}_n$ and $\mathcal{A}_n$ denote respectively the set of Latin squares which have $\Theta$ as an autotopism, its cardinality and the sets of autotopisms and automorphisms of at least one Latin square of order $n$. Necessary conditions for an isotopism to be an autotopism have been given in [@Sade68; @McKay05; @FalconAC; @Stones11] and $\mathcal{A}_n$ has been studied in [@Wanless04; @Bryant09; @Kerby10a; @Kerby10b; @Stones11]. If $\pi\in S_3$ and $L\in\mathcal{LS}_{\Theta}$, then $L^{\pi}\in \mathcal{LS}_{\Theta^{\pi}}$, so permutations on the components of $\Theta$ preserve $\Delta(\Theta)$. Moreover, this cardinality only depends on the conjugacy class of $\Theta$ [@FalconAC] or, equivalently, on its cycle structure, so we will also denote it by $\Delta(z_{\Theta})$. A classification of ${\mathcal{CS}_{\mathfrak{A}}}_n$ is known for all $n\leq 17$ [@FalconAC; @Stones11]. Given $z\in{\mathcal{CS}_{\mathfrak{A}}}_n$, let $\mathfrak{I}_z=\{\Theta\in\mathfrak{I}_n\,\mid\, z_{\Theta}=z\}$ and $\mathcal{LS}_z=\bigcup_{\Theta\in\mathfrak{I}_z}\mathcal{LS}_{\Theta}$.
An [*incidence structure*]{} is a triple $(P,B,I)$, where $P$ and $B$ respectively are finite sets of [*points*]{} and [*blocks*]{} and $I\subseteq P\times B$ is an [*incidence relation*]{}. It is [*r-uniform*]{} if every block contains $r$ points and it is [*s-regular*]{} if every point is exactly on the same number of blocks. Two blocks are [*equivalent*]{} if they contain the same set of points and the [*multiplicity*]{} of a block is the size of its equivalence class. In the study of Latin squares, it can be defined a natural incidence relation $I_n$ between $\mathcal{LS}_n$ and $\mathfrak{I}_n$, where, given $L\in \mathcal{LS}_n$ and $\Theta\in \mathfrak{I}_n$, then $(L,\Theta)\in I_n$ if and only if $L\in \mathcal{LS}_{\Theta}$. So, given $z\in {\mathcal{CS}_{\mathfrak{A}}}_n$, the triple $(\mathcal{LS}_z,\mathfrak{I}_z,I_n)$ is a $\Delta(z)$-uniform incidence structure such that every block have the same multiplicity [@FalconHMAMS]. Moreover, given $L\in\mathcal{LS}_n$, the triple $([L],\mathfrak{I}_z,I_n)$ is a uniform and regular incidence structure, where every block contains $\Delta_{[L]}(z)$ elements, whose exact value is known for order up to $6$.
Although a general expression for the values of $\Delta(z)$ and $\Delta_{[L]}(z)$ remains unknown, some general and explicit formulas have been given for the former [@Laywine81; @Laywine85; @Falcon06; @Stones10] and Gröbner bases have been used to know its exact value for all autotopisms of Latin squares of order up to $7$ [@FalconMartinJSC07]. For higher orders, Gröbner bases have problem with the exponential growth of data storage and the time of computation, for which the use of new combinatorial tools seems to be the key. So, for example, given $\Theta=(\alpha,\beta,\gamma)\in \mathfrak{I}_z$, Gröbner bases were used in [@FalconMartinEACA08] to obtain the value of $\Delta(z)$ for the majority of the cycle structures of autotopisms of $\mathfrak{A}_8$ and $\mathfrak{A}_9$, by solving the linear equation system formed after adding the constraints $x_{rcs}=x_{\alpha(r)\beta(c)\gamma(s)}$, for all $r, c, s\in [n]$, to those related to the [*planar 3-index assignment problem*]{} [@Euler86]: $$\begin{array}{l}\min \sum_{r, c, s\in [n]} w_{rcs}\cdot x_{rcs},\\
\begin{array}{ll}\text{subject to } & \sum_{r\in [n]}x_{rcs}=1, \forall c, s\in [n],\\
\ & \sum_{c\in [n]}x_{rcs}=1, \forall r, s\in [n],\\
\ & \sum_{s\in [n]}x_{rcs}=1, \forall r, c\in [n],\\
\ & x_{rcs}\in \{0,1\}, \forall r, c, s \in [n],\end{array}\end{array} \hspace{2cm} (3PAP_n)$$ where $w_{rcs}$ are real weights for all $r, c, s \in [n]$ and whose set of feasible solutions are in $1-1$ correspondence with $\mathcal{LS}_n$ if we define the Latin square $L=(l_{rc})$ such that $l_{rc}=s$ if and only if $x_{rcs}=1$.
All the previous concepts can be naturally extended to [*partial Latin squares*]{}, i.e., square arrays with elements chosen from a set of $n$ symbols, such that each symbol occurs at most once in each row and in each column. The [*size*]{} of a partial Latin square $P$ is the number of its non-blank cells and will be denoted by $|P|$. Let $\mathcal{PLS}_n$ and $\mathcal{PLS}_{n,s}$ denote respectively the set of non-empty partial Latin squares of order $n$ and its subset of arrays of size $s$. An upper bound of the elements of $\mathcal{PLS}_{n,s}$ is given in [@Ghandehari05]. The [*orthogonal representation*]{} of $P\in\mathcal{PLS}_n$ is the set $O(P)$ of $|P|$ triples related to the non-blank cells of $P$. Parastrophic partial Latin squares have therefore the same size. Given $\Theta=(\alpha,\beta,\gamma)\in\mathfrak{I}_n$, it is defined the partial Latin square $P^{\Theta}$ such that $O(P^{\Theta})=\{(\alpha(r),\beta(c),\gamma(s))\,\mid\, (r,c,s)\in O(P)\}$, which is said to be [*isotopic*]{} to $P$ and $[P]$ will denote its [*isotopism class*]{}. So, $|P^{\Theta}|=|P|$. $\Theta$ is said to be an [*autotopism*]{} of $P$ if $P^{\Theta}=P$. Let ${\mathfrak{A}_{\mathcal{P}}}_n$ and $\mathcal{PLS}_{\Theta}$ denote respectively the set of autotopisms of $\mathcal{PLS}_n$ and that of non-empty partial Latin squares which have $\Theta$ as an autotopism. Besides, given $z\in {\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$, $\mathcal{PLS}_z$ will denote the set $\bigcup_{\Theta\in\mathfrak{I}_z} \mathcal{PLS}_{\Theta}$.
A partial Latin square $P\in\mathcal{PLS}_n$ can be [*completed*]{} to a Latin square $L\in \mathcal{LS}_n$ if $O(P)\subseteq O(L)$. Given $\Theta\in{\mathcal{CS}_{\mathfrak{A}}}_n$, the subset of $\mathcal{LS}_{\Theta}$ of Latin squares to which $P$ can be completed is denoted by $\mathcal{LS}_{\Theta,P}$. The computation of $\Delta(z)$ can be then simplified [@FalconMartinJSC07] if a multiplicative factor $c_P\in\mathbb{N}$ is found such that $\Delta(z)=c_P\cdot |\mathcal{LS}_{\Theta,P}|$. Although this factor, which is called [*$P$-coefficient of symmetry of $\Theta$*]{}, becomes crucial in the processing of high orders, none exhaustive study has been developed in this regard. Indeed, a comprehensive analysis of ${\mathfrak{A}_{\mathcal{P}}}_n$ and $\mathcal{PLS}_{\Theta}$ has not been properly done until now.
The present paper deals with this last question. It is organized as follows: In Section 2, the set ${\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$ will be characterized and several related results exposed. In Section 3, given $z\in {\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$, it will be dealt with the possible sizes of a partial Latin square $P\in \mathcal{PLS}_z$. In Section 4, analogously to the case of Latin squares, it will be proven that the number of partial Latin squares related to an autotopism only depends on the cycle structure of the latter, in such a way that the elements of $\mathcal{PLS}_n$ and $\mathfrak{I}_n$ can be respectively considered as points and blocks of incidence structures whose uniformity and regularity will be studied. Moreover, new constraints will be imposed to the $3PAP_n$ in order to obtain the set $\mathcal{PLS}_{z,s}$ of partial Latin squares of size $s\in [n]$ related to an autotopism of cycle structure $z\in {\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$. Besides, by using Gröbner bases, its cardinality $\Delta_s(z)$ will be obtained for $n\leq 4$. Finally, in Section 5, a theoretical ground for the coefficient of symmetry will be exposed. Specifically, given $\Theta\in\mathfrak{I}_n$, it will be studied the set of partial Latin squares of $\mathcal{PLS}_{\Theta}$ which can be completed to Latin squares of $\mathcal{LS}_{\Theta}$.
The set ${\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$.
========================================================
Autotopisms of partial Latin squares are univocally determined by their cycle structures:
\[lem0a\] $\Theta\in {\mathfrak{A}_{\mathcal{P}}}_n$ if and only if $z_{\Theta}\in {\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$.
[**Proof.**]{} The necessary condition holds by definition of ${\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$. Now, if $z\in{\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$, then there must exist $\Theta_0\in \mathfrak{I}_z$ and $P\in \mathcal{PLS}_{\Theta_0}$. So, given $\Theta\in\mathfrak{I}_z$, $\Theta$ and $\Theta_0$ are conjugate and, therefore, there exists $\Theta'\in\mathfrak{I}_n$ such that $\Theta=\Theta'\Theta_0\Theta'^{-1}$. As a consequence, $P^{\Theta'}\in \mathcal{PLS}_{\Theta}$ and $\Theta\in{\mathfrak{A}_{\mathcal{P}}}_n$. $\Box$
Let us define the set: $$\mathrm{LCM}_n=\{(i,j,k)\in [n]^3\,\mid\,\mathrm{lcm}(i,j)=\mathrm{lcm}(i,k)=\mathrm{lcm}(j,k)=\mathrm{lcm}(i,j,k)\}.$$
The next result characterizes the set ${\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$ and can be considered as an immediate generalization for partial Latin squares of the necessary condition given by Stones, Vojtěchovský and Wanless in [@Stones11] for membership in $\mathfrak{A}_n$:
\[lem0\] Given $z=(z_1,z_2,z_3)\in \mathcal{CS}_{\mathfrak{I}_n}$, it is $z\in {\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$ if and only if there exists $(i,j,k)\in \mathrm{LCM}_n$ such that $z_{1i}\cdot z_{2j}\cdot z_{3k}>0$.
[**Proof.**]{} If $z\in {\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$, then there must exist $\Theta=(\alpha,\beta,\gamma)\in\mathfrak{I}_z$ and $P\in\mathcal{PLS}_{\Theta}$. Given $(r,c,s)\in O(P)$, let $(u,v,w)\in [n_{\alpha}]\times[n_{\beta}]\times[n_\gamma]$ be such that $r\in \alpha_u$, $c\in\beta_v$ and $s\in\gamma_w$. Since $\Theta$ is an autotopism of $P$, it must be $(\alpha_u^t(r),\beta_v^t(c),\gamma_w^t(s))\in O(P)$, for all $t\in\mathbb{N}$. The necessary condition is then a consequence of the Latin square condition, by considering $i,j,k$ to be, respectively, the lengths of $\alpha_u$, $\beta_v$ and $\gamma_w$.
To prove the converse, let $\Theta=(\alpha,\beta,\gamma)\in\mathfrak{I}_z$ and let $\alpha_u, \beta_v$ and $\gamma_w$ be, respectively, $i$-, $j$- and $k$-cycles of $\alpha$, $\beta$ and $\gamma$. Let $r, c, s$ be, respectively, elements of $\alpha_u, \beta_v$ and $\gamma_w$. The set of triples $\{(\alpha_u^t(r),\beta_v^t(c),\gamma_w^t(s))\,\mid\, t\in [\mathrm{lcm}(i,j,k)]\}$ verifies the Latin square condition because of being $(i,j,k)\in \mathrm{LCM}_n$ and, therefore, it is the orthogonal representation of a partial Latin square $P\in \mathcal{PLS}_{\Theta}$. $\Box$
Given $n>1$, Lemma \[lem0\] implies ${\mathfrak{A}_{\mathcal{P}}}_n$ to be a proper subset of $\mathfrak{I}_n$, because, for instance, $(1^2, 1^2, n^1)\not\in {\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$. Analogously, $\mathfrak{A}_n$ is a proper subset of ${\mathfrak{A}_{\mathcal{P}}}_n$, because, for example, $(2,2,2)\in{\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_2$ and $(2\cdot 1^{n-2},2\cdot 1^{n-2},1^n)\in{\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$, for $n>2$, but neither of them are cycle structures of an autotopism of Latin square. Thus, the next claim is verified:
\[prp0\] ${{\mathfrak{A}}}_n\subset {{\mathfrak{A}_{\mathcal{P}}}}_n\subset {{\mathfrak{I}}}_n$, $\forall n>1$. $\Box$
Moreover, ${{\mathfrak{A}_{\mathcal{P}}}}_n$ and ${{\mathfrak{I}}}_n$ can be identified when $n$ tends to infinity. To see it, it is enough to assure that the cardinalities of the sets of their cycle structures coincide in the limit, which will be proven in Theorem \[thm0\]. Previously, although ${\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$ can be explicitly obtained for any order $n\in\mathbb{N}$ by implementing Lemma \[lem0\] in a computer procedure, a lower bound of its cardinality will be determined by studying the following sets which partition $\mathcal{CS}_n$: $$\mathcal{CS}_{n,m}=\{n^{z_n}\cdot ...\cdot 2^{z_2}\cdot 1^{z_1}\in \mathcal{CS}_n\,\mid\, z_m>0 \text{ and } z_i=0, \forall i\in [m-1]\},$$ where $m\in[n]$. The following results hold:
\[lemCS\] $|\mathcal{CS}_{n,m}|=\begin{cases} 1, \text{ if } m=n,\\
0, \text{ if } m\in \{\lceil\frac n2 \rceil,...,n-1\},\\
p(n-m) - \sum_{i=1}^{m-1} |\mathcal{CS}_{n-m,i}|, \text{ otherwise}.\end{cases}$
[**Proof.**]{} The cases $m\geq \lceil\frac n2 \rceil$ are straightforward verified. Let $m\leq \lfloor\frac n2 \rfloor$. Given $z\in \mathcal{CS}_{n,m}$, it is $z_m>0$ and, therefore, it must be $z_{n-m+i}=0$, for all $i\in [m]$. So, we can define $z'\in \mathcal{CS}_{n-m}$ such that $z'_i=z$, for all $i\in [n-m]\setminus\{m\}$ and $z'_m=z_m-1$. Specifically, since $z\in \mathcal{CS}_{n,m}$, it must be $z'_i=0$, for all $i\in [m-1]$. Thus, $z'\in \bigcup_{i=m}^{n-m}\mathcal{CS}_{n-m,i}=\mathcal{CS}_{n-m}\setminus \bigcup_{i=1}^{m-1}\mathcal{CS}_{n-m,i}$ and the claim is verified. $\Box$
\[prpCS\] $|{\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n|\geq \sum_{(i,j,k)\in \mathrm{LCM}_n} |\mathcal{CS}_{n,i}|\cdot |\mathcal{CS}_{n,j}|\cdot |\mathcal{CS}_{n,k}|$.
[**Proof.**]{} Since the sets $\mathcal{CS}_{n,m}$ constitute a partition of $\mathcal{CS}_n$, the result is consequence of Lemma \[lem0\].$\Box$
\[thm0\] $\lim_{n\rightarrow\infty} \frac {|{\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n|}{|{\mathcal{CS}_{\mathfrak{I}}}_n|}=1$.
[**Proof.**]{} Since $(1,1,1)\in \mathrm{LCM}_n$, Proposition \[prpCS\] implies that $|{\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n|\geq |\mathcal{CS}_{n,1}|^3$ and, therefore, from Lemma \[lemCS\], $|{\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n|\geq p(n-1)^3$. Now, since $p(n)$ is equivalent to $\frac {e^{\pi\sqrt{2n/3}}}{4n\sqrt{3}}$ when $n$ tends to infinity [@Hardy18] and $|{\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n|\leq |{\mathcal{CS}_{\mathfrak{I}}}_n|=p(n)^3$, then: $$1\geq\lim_{n\rightarrow\infty} \frac {|{\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n|}{|{\mathcal{CS}_{\mathfrak{I}}}_n|} \geq \lim_{n\rightarrow\infty} \frac {p(n-1)^3}{p(n)^3}=1.$$ $\Box$
For $n\leq 17$, Table 1 shows the values $|\mathcal{CS}_{n,m}|$ and $|{\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n|$, where $m\leq \lfloor \frac n2\rfloor$, in comparison with those of $|{\mathcal{CS}_{\mathfrak{A}}}_n|$, which can be obtained by using the classification given in [@FalconAC; @Stones11]. The number $|[{\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n]|$ of parastrophic classes of ${\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$ is also shown.
\[t1\]
---- ----- ----- ---- ---- ----- ------ ------- -------- -------- ---------- ---------
1 2 3 4 5 6 7 8
1 1 1 1
2 4 1 5 3
3 6 2 15 7
4 19 3 1 65 22
5 8 5 1 223 60
6 45 7 2 1 869 197
7 12 11 2 1 2535 526
8 87 15 4 1 1 7663 1492
9 43 22 4 2 1 21156 3937
10 89 30 7 2 1 1 60264 10850
11 21 42 8 3 1 1 150953 26628
12 407 56 12 4 2 1 1 385538 66984
13 27 77 14 5 2 1 1 915452 157398
14 141 101 21 6 3 1 1 1 2193225 374127
15 150 135 24 9 3 2 1 1 4928696 836154
16 503 176 34 10 5 2 1 1 1 11209311 1893607
17 40 231 41 13 5 3 1 1 1 24406191 4110132
---- ----- ----- ---- ---- ----- ------ ------- -------- -------- ---------- ---------
: Cardinality of the sets of cycle structures, for $n\leq 17$ and $m\leq \lfloor \frac n2\rfloor$.
The size of a partial Latin square related to an autotopism.
============================================================
Given $z=(z_1,z_2,z_3)\in {\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$ and $\Theta=(\alpha,\beta,\gamma)\in\mathfrak{I}_z$, any partial Latin square $P\in\mathcal{PLS}_{\Theta}$ can be decomposed into $n_{z_1}\cdot n_{z_2}$ blocks $P_{ij}$ whose rows and columns are respectively determined by the elements of the cycle $\alpha_i$ of $\alpha$ and the cycle $\beta_j$ of $\beta$, i.e., $O(P_{ij})=\{(r,c,s)\in O(P)\,\mid\, r\in\alpha_i \text{ and } c\in\beta_j\}$. It will be called the [*$\Theta$-decomposition of $P$*]{}. Specifically, $z$ determines not only the number of these blocks, but also their possible sizes and, consequently, a pair of bounds for the size of $P$. To see it, let us define the set: $$\mathrm{LCM}_z=\{(i,j)\in [n]^2\,\mid\, \exists k\in[n] \text{ s.t. } (i,j,k)\in \mathrm{LCM}_n \text{ and } z_{1i}\cdot z_{2j}\cdot z_{3k}>0\}.$$
The following results hold:
\[lemR1\] Given $z\in {\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$, $\Theta\in \mathfrak{I}_z$, $P\in\mathcal{PLS}_{\Theta}$ and an $i\times j$-block $B$ of the $\Theta$-decomposition of $P$, there exists $\omega_B\in [\mathrm{gcd}(i,j)]\cup\{0\}$ such that $|B|=\omega_B\cdot \mathrm{lcm}(i,j)$. Specifically, $\omega_B=0$ if $(i,j)\not\in \mathrm{LCM}_z$.
[**Proof.**]{} Analogously to the proof of Lemma \[lem0\], the Latin square condition implies $|B|=0$, whenever $(i,j)\not\in \mathrm{LCM}_z$. Besides, given $(r,c,s)\in O(B)$, its orbit by the action of $\Theta=(\alpha,\beta,\gamma)$ is the set of triples $(\alpha^t(r),\beta^t(c),$ $\gamma^t(s))\in O(B)$, for all $t\in [\mathrm{lcm}(i,j)]$. So, the Latin square condition implies $|B|$ to be a multiple of $\mathrm{lcm}(i,j)$. Finally, since there are $i\cdot j$ cells in $B$, the multiplicative factor must be at most $\mathrm{gcd}(i,j)$. $\Box$
\[prpR1\] Given $z=(z_1,z_2,z_3)\in {\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$ and $P\in\mathcal{PLS}_z$, it is $\mathfrak{l}_z\leq |P|\leq \mathfrak{u}_z$, where: $$\mathfrak{l}_z=\min_{(i,j)\in \mathrm{LCM}_z}\{\mathrm{lcm}(i,j)\},$$ [$$\mathfrak{u}_z=\min\{\sum_{(i,j)\in \mathrm{LCM}_z} z_{1i}\cdot z_{2j} \cdot i\cdot j, \sum_{(i,k)\in \mathrm{LCM}_{z^{(23)}}} z_{1i}\cdot z_{3k} \cdot i\cdot k, \sum_{(k,j)\in \mathrm{LCM}_{z^{(13)}}} z_{2j}\cdot z_{3k} \cdot j\cdot k\}.$$]{}
[**Proof.**]{} Let $\Theta\in \mathfrak{I}_z$ be such that $P\in\mathcal{PLS}_{\Theta}$ and let $B$ be a block of the $\Theta$-decomposition of $P$ such that $|B|>0$. From Lemma \[lemR1\], if $B$ is an $i\times j$-block, where $(i,j)\in\mathrm{LCM}_z$, then $\mathrm{lcm}(i,j)\leq |B|\leq i\cdot j$ and so, $\mathfrak{l}_z\leq |P|\leq \sum_{(i,j)\in \mathrm{LCM}_z} z_{1i}\cdot z_{2j} \cdot i\cdot j$. Since the size of a partial Latin square is invariant by parastrophism and $P^{\pi}\in \mathcal{PLS}_{\Theta^{\pi}}$ for all $\pi\in S_3$, then $\mathfrak{u}_z$ is an upper bound of $|P|$. $\Box$
From the previous results, it is deduced that the possible sizes of the elements of $\mathcal{PLS}_z$ must be in the set: $$\mathrm{Sizes}(z)=\left\{\sum_{(i,j)\in \mathrm{LCM}_z} \omega_{ij} \cdot \mathrm{lcm}(i,j)\leq\mathfrak{u}_z\,\mid\, \omega_{ij}\in [z_{1i}\cdot z_{2j}\cdot \mathrm{gcd}(i,j)]\right\}.$$
As an example, let us consider $z=(6, 3\cdot 2\cdot 1, 4 \cdot 2)\in{\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_6$ and $\Theta=((123456),(123)(45)(6),(1234)(56))\in\mathcal{PLS}_z$. The $\Theta$-decomposition of any partial Latin square $P\in \mathcal{PLS}_{\Theta}$ is then formed by three blocks, $P_{11}, P_{12}$ and $P_{13}$, whose cells are respectively indicated by the symbols $\cdot$, $*$ and $\circ$ in the following diagram: [$$\left(\begin{array}{cccccc} \cdot & \cdot & \cdot & * &* &\circ\\
\cdot & \cdot & \cdot & * &* &\circ\\
\cdot & \cdot & \cdot &* &* &\circ\\
\cdot & \cdot & \cdot &* &* &\circ\\
\cdot & \cdot & \cdot &* &* &\circ\\
\cdot & \cdot & \cdot &* &* &\circ\\
\end{array}\right).$$]{}
Besides, $\mathrm{LCM}_z=\{(6,3)\}$, $\mathrm{LCM}_{z^{(23)}}=\{(6,2)\}$ and $\mathrm{LCM}_{z^{(13)}}=\{(2,3)\}$. So, from Proposition \[prpR1\], it must be $6\leq |P|\leq \min\{18,12,6\}=6$. Thus, $\mathrm{Sizes}(z)=\{6\}$ and $|P|=6$. Specifically, there are six possibilities for $P$: [$$\left(\begin{array}{cccccc} 5 & \cdot & \cdot &\cdot &\cdot &\cdot\\
\cdot & 6 & \cdot &\cdot &\cdot &\cdot\\
\cdot & \cdot & 5 &\cdot &\cdot &\cdot\\
6 & \cdot & \cdot &\cdot &\cdot &\cdot\\
\cdot & 5 & \cdot &\cdot &\cdot &\cdot\\
\cdot & \cdot & 6 &\cdot &\cdot &\cdot\\
\end{array}\right), \hspace{1cm} \left(\begin{array}{cccccc} 6 & \cdot & \cdot &\cdot &\cdot &\cdot\\
\cdot & 5 & \cdot &\cdot &\cdot &\cdot\\
\cdot & \cdot & 6 &\cdot &\cdot &\cdot\\
5 & \cdot & \cdot &\cdot &\cdot &\cdot\\
\cdot & 6 & \cdot &\cdot &\cdot &\cdot\\
\cdot & \cdot & 5 &\cdot &\cdot &\cdot\\
\end{array}\right), \hspace{1cm} \left(\begin{array}{cccccc} \cdot & 5 & \cdot &\cdot &\cdot &\cdot\\
\cdot & \cdot & 6 &\cdot &\cdot &\cdot\\
5 & \cdot & \cdot &\cdot &\cdot &\cdot\\
\cdot & 6 & \cdot &\cdot &\cdot &\cdot\\
\cdot & \cdot & 5 &\cdot &\cdot &\cdot\\
6 & \cdot & \cdot &\cdot &\cdot &\cdot\\
\end{array}\right),$$ $$\left(\begin{array}{cccccc} \cdot & 6 & \cdot &\cdot &\cdot &\cdot\\
\cdot & \cdot & 5 &\cdot &\cdot &\cdot\\
6 & \cdot & \cdot &\cdot &\cdot &\cdot\\
\cdot & 5 & \cdot &\cdot &\cdot &\cdot\\
\cdot & \cdot & 6 &\cdot &\cdot &\cdot\\
5 & \cdot & \cdot &\cdot &\cdot &\cdot\\
\end{array}\right),\hspace{1cm} \left(\begin{array}{cccccc} \cdot & \cdot & 5 &\cdot &\cdot &\cdot\\
6 & \cdot & \cdot &\cdot &\cdot &\cdot\\
\cdot & 5 & \cdot &\cdot &\cdot &\cdot\\
\cdot & \cdot & 6 &\cdot &\cdot &\cdot\\
5 & \cdot & \cdot &\cdot &\cdot &\cdot\\
\cdot & 6 & \cdot &\cdot &\cdot &\cdot\\
\end{array}\right),\hspace{1cm} \left(\begin{array}{cccccc} \cdot & \cdot & 6 &\cdot &\cdot &\cdot\\
5 & \cdot & \cdot &\cdot &\cdot &\cdot\\
\cdot & 6 & \cdot &\cdot &\cdot &\cdot\\
\cdot & \cdot & 5 &\cdot &\cdot &\cdot\\
6 & \cdot & \cdot &\cdot &\cdot &\cdot\\
\cdot & 5 & \cdot &\cdot &\cdot &\cdot\\
\end{array}\right).$$]{}
The number of partial Latin squares related to an autotopism.
=============================================================
Given $z\in {\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$, $\Theta\in\mathfrak{I}_z$, $P\in\mathcal{PLS}_{\Theta}$ and $s\in [n]$, let us define the sets: $$\mathcal{PLS}_{\Theta,[P]}=\mathcal{PLS}_{\Theta}\cap [P], \hspace{1.5cm} \mathcal{PLS}_{\Theta,s}=\mathcal{PLS}_{\Theta}\cap \mathcal{PLS}_{n,s}.$$
In the current section, given $\Theta\in\mathfrak{I}_n$, the cardinality of the set $\mathcal{PLS}_{\Theta}$ will be studied. The following result implies that it only depends on the cycle structure of $\Theta$:
\[lemD1\] The number of isotopic partial Latin squares related to an autotopism only depends on the parastrophic class of the cycle structure of the latter.
[**Proof.**]{} Let $\Theta_1, \Theta_2\in \mathfrak{I}_n$ and $\pi\in S_3$ be such that $z_{\Theta_1}=z_{\Theta_2^{\pi}}$ and let $P\in\mathcal{PLS}_n$. Since $\Theta_1$ and $\Theta_2^{\pi}$ are conjugate, there exists $\Theta\in \mathfrak{I}_n$ such that $\Theta_2^{\pi}=\Theta\Theta_1\Theta^{-1}$. Now, given $Q\in\mathcal{PLS}_{\Theta_1,[P]}$, it is $Q^{\Theta}\in \mathcal{PLS}_{\Theta_2^{\pi},[P]}$ and, therefore, $\left(Q^{\Theta}\right)^{\pi^{-1}}\in \mathcal{PLS}_{\Theta_2,[P]}$. So, $|\mathcal{PLS}_{\Theta_1,[P]}|\leq|\mathcal{PLS}_{\Theta_2,[P]}|$. The opposite inequality is similarly proven. $\Box$
Since the size of a partial Latin square is preserved by isotopism, Lemma \[lemD1\] implies the following cardinalities to be well-defined: $$\Delta_{[P]}(z)=\left|\mathcal{PLS}_{\Theta, [P]}\right|,$$ $$\Delta_s(z)=\left|\mathcal{PLS}_{\Theta,s}\right|=\sum_{\scriptsize \begin{array}{c}[Q]\in\mathcal{PLS}_{\Theta}/\sim\\s.t.\ |Q|=s\end{array}}\Delta_{[Q]}(z),$$ $$\Delta_{\mathcal{P}}(z)=\left| \mathcal{PLS}_{\Theta}\right|=\sum_{\scriptsize [Q]\in\mathcal{PLS}_{\Theta}/\sim}\Delta_{[Q]}(z)=\sum_{s\in \mathrm{Sizes}(z)}\Delta_s(z).$$
It can be defined a natural incidence relation $I_{\mathcal{P}_n}$ between $\mathcal{PLS}_n$ and $\mathfrak{I}_n$, where, given $P\in \mathcal{PLS}_n$ and $\Theta\in \mathfrak{I}_n$, then $(P,\Theta)\in I_n$ if and only if $P\in \mathcal{PLS}_{\Theta}$. Besides, let us denote by $\mathfrak{A}_P$ the set of autotopisms of $P$. The following results are then proven:
\[prpD1\] Let $P\in \mathcal{PLS}_n$ and $z\in {\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$. The triples $([P],\mathfrak{I}_z,I_{\mathcal{P}_n})$, $(\mathcal{PLS}_{n,s},\mathfrak{I}_z,I_{\mathcal{P}_n})$ and $(\mathcal{PLS}_n,\mathfrak{I}_z,I_{\mathcal{P}_n})$ are, respectively, $\Delta_{[P]}(z)$-, $\Delta_s(z)$- and $\Delta_{\mathcal{P}}(z)$-uniform incidence structures and all its blocks have the same multiplicity. Moreover, the former incidence structure is regular.
[**Proof.**]{} From Lemma \[lemD1\], it is enough to study the uniformity and multiplicity of $([P],\mathfrak{I}_z,I_{\mathcal{P}_n})$. Indeed, the uniformity is an immediate consequence of that lemma. Now, in order to see that all the blocks have the same multiplicity, let $\Theta_1,\Theta'_1\in\mathfrak{I}_z$ be such that $\mathcal{PLS}_{\Theta_1,[P]}=\mathcal{PLS}_{\Theta'_1,[P]}$ and let us consider $\Theta_2\in\mathfrak{I}_z$. Let $\Theta,\Theta'\in\mathfrak{I}_n$ be such that $\Theta_1=\Theta\Theta_2\Theta^{-1}$ and $\Theta'_1=\Theta'\Theta_1\Theta'^{-1}$. Then, $\mathcal{PLS}_{\Theta_2,[P]}=\mathcal{PLS}_{\Theta^{-1}\Theta'\Theta,[P]}$, because $Q\in\mathcal{PLS}_{\Theta_2,[P]}\Leftrightarrow Q^{\Theta}\in\mathcal{PLS}_{\Theta_1,[P]}\Leftrightarrow Q^{\Theta'\Theta}\in\mathcal{PLS}_{\Theta'_1,[P]}=\mathcal{PLS}_{\Theta_1,[P]}\Leftrightarrow Q^{\Theta^{-1}\Theta'\Theta}\in\mathcal{PLS}_{\Theta_2,[P]}$. Moreover, $\Theta^{-1}\Theta'\Theta=\Theta_2 \Leftrightarrow \Theta'=\Theta^{-1}\Theta_2\Theta=\Theta_1 \Leftrightarrow \Theta'_1=\Theta_1$. Thus, the arbitrariness of $\Theta_1,\Theta'_1$ and $\Theta_2$ implies the claim about the multiplicity.
Finally, in order to see that $([P],\mathfrak{I}_z,I_{\mathcal{P}_n})$ is a regular incidence structure, let us consider $Q_1, Q_2\in [P]$ and let $\Theta\in\mathfrak{A}_{Q_1}\cap \mathfrak{I}_z$. Since $Q_1$ and $Q_2$ are isotopic, there must exist $\Theta'\in \mathfrak{I}_n$ such that $Q_1^{\Theta'}=Q_2$. So, $\Theta'\Theta\Theta'^{-1}\in \mathfrak{A}_{Q_2}\cap\mathfrak{I}_z$ and, therefore, $|\mathfrak{A}_{Q_1}\cap\mathfrak{I}_z|\leq |\mathfrak{A}_{Q_2}\cap\mathfrak{I}_z|$. The regularity holds because the opposite inequality is analogously proven. $\Box$
\[thmD2\] Let $P\in\mathcal{PLS}_n$. If $Q\in [P]$, then $|\mathfrak{A}_Q|=|\mathfrak{A}_P|$ and it coincides with the cardinality of the set $\mathfrak{I}_{P,Q}$ of isotopisms from $P$ to $Q$. Moreover, given $\Theta\in\mathfrak{A}_P$, it is $\mathfrak{A}_Q=\{\Theta'\Theta{\Theta'}^{-1}\,\mid\, \Theta'\in \mathfrak{I}_{P,Q}\}$.
[**Proof.**]{} From Proposition \[prpD1\], it is verified that $|\mathfrak{A}_Q|=\sum_{z\in {\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n}|\mathfrak{A}_Q\cap\mathfrak{I}_z|=\sum_{z\in {\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n}|\mathfrak{A}_P\cap\mathfrak{I}_z|=|\mathfrak{A}_P|$. Besides, given $\Theta\in\mathfrak{A}_P$, it is $\Theta'\Theta{\Theta'}^{-1}\in\mathfrak{A}_Q$, for all $\Theta'\in\mathfrak{I}_{P,Q}$ and, therefore, $|\mathfrak{I}_{P,Q}|\leq |\mathfrak{A}_Q|$. The opposite inequality is also verified, because, given $\Theta\in\mathfrak{I}_{P,Q}$, it is $\Theta'\Theta\in\mathfrak{I}_{P,Q}$, for all $\Theta'\in\mathfrak{A}_Q$. The equality also implies that the last assertion of the result is then an immediate consequence of the first part of the reasoning. $\Box$
Hereafter, we focus our study on the values $\Delta_s(z)$. The values $\Delta_{[P]}(z)$ needs a comprehensive analysis of the isotopic classes of partial Latin squares and will be considered in a further study. Firstly, it raises the natural question of whether it is possible to obtain some general expression which determines these values for some specific size or cycle structure. So, for instance, it is immediate to see that $\Delta_s((1^n,1^n,1^n))=|\mathcal{PLS}_{n,s}|$ and, since $\mathcal{PLS}_{n,n^2}=\mathcal{LS}_n$, it is also clear that $\Delta_{n^2}(z)=\Delta(z)$. In this regard, let us study some cases in which a general formula is given:
\[prpD2\] Let $s\in [n^2]$. It is verified that: $$\Delta_s((n,n,1^n))=\begin{cases}\frac {{n!}^2}{k! \cdot {(n-k)!}^2}, \text { if } \exists k\in [n] \text{ s.t. } s=k\cdot n,\\0, \text{ otherwise}.\end{cases}$$
[**Proof.**]{} Let $\Theta=(\alpha,\beta,\mathrm{Id})\in\mathfrak{I}_{(n,n,1^n)}$, where ${\mathrm{Id}}$ denotes the trivial permutation, and $P\in\mathcal{PLS}_{\Theta,s}$. Since the $\Theta$-decomposition of $P$ is only formed by $P$ itself and $LMC_z=\{(n,n)\}$, then Lemma \[lemR1\] implies $s=k\cdot n$, for some $k\in [n]$. Thus, $O(P)$ is decomposed under the action of $\Theta$ into $k$ orbits of length $n$. Specifically, there exist exactly $k$ distinct columns $c_1,c_2,...,c_k\in [n]$ and $k$ distinct symbols $s_1,s_2,...,s_k\in [n]$, such that $(1,c_i,s_i)\in O(P)$, for all $i\in [k]$. The $k$ orbits of $O(P)$ under $\Theta$ are then the sets $\{(\alpha^t(1),\beta^t(c_i),s_i)\,\mid\, t\in [n]\}$, with $i\in [k]$.
Every element of $\mathcal{PLS}_{\Theta,k\cdot n}$ is therefore univocally determined by the choice of the columns $c_i$ and symbols $s_i$. Namely, there exist $\left(\begin{array}{c}n\\k\end{array}\right)$ possible ways of choosing the $k$ columns and, once they have been selected, there exist $\frac {n!}{(n-k)!}$ different ways of assigning $k$ symbols to the cells $(1,c_1),...,(1,c_n)$. So, $\Delta_s(z)=\left(\begin{array}{c}n\\k\end{array}\right)\cdot \frac {n!}{(n-k)!}=\frac {{n!}^2}{k! \cdot {(n-k)!}^2}$.$\Box$
\[prpD3\] It is verified that $\Delta_n((n,n,n))=n^2$. Besides, if $n>2$: $$\Delta_{2n}((n,n,n))=\frac {n^2\cdot (n-1)\cdot (n-2)}2.$$
[**Proof.**]{} Let $\Theta=(\alpha,\beta,\gamma)\in\mathfrak{I}_{(n,n,n)}$. Similarly to the proof of Proposition \[prpD2\], every partial Latin square of $\mathcal{PLS}_{\Theta,n}$ is univocally determined by the only non-empty cell $(1,c)$ of its first row. Specifically, $c$ can be selected from $n$ different columns and the symbol of $(1,c)$ will be able to be chosen between $n$ candidates. So, $\Delta_n((n,n,n))=n^2$.
Analogously, if $n>2$, every partial Latin square of $\mathcal{PLS}_{\Theta,2n}$ is univocally determined by the only two non-empty cells of its first row. Their corresponding columns $c_1, c_2\in [n]$ can be selected of $\left(\begin{array}{c}n\\2\end{array}\right)$ different ways. Now, the symbol $s_1$ of the cell $(1,c_1)$ can be chosen from $n$ possibilities, but, once it has been chosen, there exist only $n-2$ candidates for the symbol $s_2$ of the cell $(1,c_2)$. To see it, let $t\in [n-1]$ be such that $\beta^t(c_1)=c_2$. So, the triple $(\alpha^t(1),c_2,\gamma^t(s_1))$ belongs to $O(P)$. Moreover, since $z_{\gamma}=n$, then $s_1\neq \gamma^t(s_1)$ and the Latin square condition implies $s_2\not\in\{s_1,\gamma^t(s_1)\}$. Thus, $\Delta_{2n}((n,n,n))=\left(\begin{array}{c}n\\2\end{array}\right)\cdot n\cdot (n-2)=\frac {n^2\cdot (n-1)\cdot (n-2)}2$. $\Box$
\[thmD1\] Given $z=(z_1,z_2,z_3)\in {\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$: $$\Delta_{\mathfrak{l}_z}(z)=\sum_{\scriptsize \begin{array}{c}(i,j)\in \mathrm{LCM}_z\\\text{s.t. } \mathrm{lcm}(i,j)=\mathfrak{l}_z\end{array}} z_{1i}\cdot z_{2j}\cdot \mathrm{gcd}(i,j)\cdot \sum_{\scriptsize \begin{array}{c}k\in [n]\\\text{s.t. } (i,j,k)\in \mathrm{LCM}_n\end{array}}k\cdot z_{3k}.$$
[**Proof.**]{} Given $\Theta=(\alpha,\beta,\gamma)\in\mathfrak{I}_z$, let $P\in \mathcal{PLS}_{\Theta}$ be such that $|P|=\mathfrak{l}_z$. From Lemma \[lemR1\] and Proposition \[prpR1\], there must exist only one non-empty block $B$ in the $\Theta$-decomposition of $P$. Specifically, $B$ must be an $i\times j$-block of size $\mathrm{lcm}(i,j)=\mathfrak{l}_z$, where $(i,j)\in \mathrm{LCM}_z$. There exist $z_{1i}\cdot z_{2j}$ possible blocks in this way.
Moreover, $O(B)$ must be composed by all the triples of one of the $\mathrm{gcd}(i,j)$ orbits induced on $B$ by the action of $\Theta$. If $(r,c,s)\in [n]^3$ is one of these triples, then the symbol $s$ must be one of the $k\cdot z_{3k}$ elements of a $k$-cycle of $\gamma$ such that $(i,j,k)\in \mathrm{LCM}_n$. The result follows then by considering all the previous possibilities. $\Box$
\[corD1\] Let $P\in \mathcal{PLS}_{n,1}$. Given $z=(z_1,z_2,$ $z_3)\in {\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$: $$\Delta_{[P]}(z)=\Delta_1(z)=z_{11}\cdot z_{21}\cdot z_{31}.$$
[**Proof.**]{} Since there exists only one isotopic class of partial Latin squares of size 1, it is $\Delta_{[P]}(z)=\Delta_1(z)$. Now, if $\mathfrak{l}_z>1$, then $(1,1)\not\in \mathrm{LCM}_z$. So, $z_{11}\cdot z_{21}\cdot z_{31}=0$ and the result holds. Finally, if $\mathfrak{l}_z=1$, then it is enough to observe that it must be $(i,j,k)=(1,1,1)$ in the formula of Theorem \[thmD1\]. $\Box$
The number $\Delta_s(z)$ can also obviously be obtained if the set $\mathcal{PLS}_{\Theta,n}$ is known for some $\Theta\in\mathfrak{I}_z$. In order to determine this set, let us observe that, analogously to the case of Latin squares [@Euler86], $\mathcal{PLS}_n$ can be identified [@Kumar99] with the set $\mathcal{FS}_{\mathcal{P}_n}$ of feasible solutions of the integer program: $$\begin{array}{l}\min \sum_{r, c, s\in [n]} w_{rcs}\cdot x_{rcs},\\
\begin{array}{ll}\text{subject to } & \sum_{r\in [n]}x_{rcs}\leq 1, \forall c, s\in [n],\\
\ & \sum_{c\in [n]}x_{rcs}\leq 1, \forall r, s\in [n],\\
\ & \sum_{s\in [n]}x_{rcs}\leq 1, \forall r, c\in [n],\\
\ & x_{rcs}\in \{0,1\}, \forall r, c, s \in [n],\end{array}\end{array} \hspace{2cm} (1)$$ where $w_{rcs}$ are real weights for all $r, c, s \in [n]$. Specifically, it is enough to define the map $\varphi_n:\mathcal{PLS}_n\rightarrow \mathcal{FS}_{\mathcal{P}_n}$, such that, given $P\in\mathcal{PLS}_n$, it is $\varphi_n(P)=(x^P_{111},...,x^P_{11n},x^P_{121},...,x^P_{nnn})$, where, $x^P_{rcs}=1$ if $(r,c,s)\in O(P)$ and $0$, otherwise. The restriction of $\varphi_n$ to $\mathcal{PLS}_{\Theta}$ and $\mathcal{PLS}_{\Theta,m}$ assures the truthfulness of the following result:
\[prpP1\] Given $\Theta=(\alpha,\beta,\gamma)\in \mathfrak{I}_n$, there exists a bijection between $\mathcal{PLS}_{\Theta}$ and the set of feasible solutions of the equation system which results after adding to $(1)$ the constraints: $$x_{rcs}=x_{\alpha(r)\beta(c)\gamma(s)},\forall r, c, s \in [n].$$ Moreover, given $m\in [n^2]$, if the equation: $$\sum_{r,c,s\in [n]} x_{rcs} = m$$ is also added, then there exists a bijection between $\mathcal{PLS}_{\Theta,m}$ and the set of feasible solutions of the resulting equation system. $\Box$
Proposition \[prpP1\] implies $\mathcal{PLS}_{\Theta,m}$ to be determined by $2n^3+3n^2+1$ polynomial equations of degree $1$ and $2$ in $n^3$ variables:
\[corP1\] Given $\Theta=(\alpha,\beta,\gamma)\in \mathfrak{I}_n$ and $m\in[n^2]$, $\mathcal{PLS}_{\Theta,m}$ is the set of zeros of the ideal $I=\langle\,(\sum_{r\in [n]}x_{rcs})\cdot (1-\sum_{r\in [n]}x_{rcs})=0\,\mid\, c, s\in [n]\,\rangle + \langle\,(\sum_{c\in [n]}x_{rcs})\cdot (1-\sum_{c\in [n]}x_{rcs})\,\mid\, r, s\in [n]\,\rangle + \langle\,(\sum_{s\in [n]}x_{rcs})\cdot (1-\sum_{s\in [n]}x_{rcs})\,\mid\, r, c\in [n]\,\rangle + \langle\,x_{rcs}\cdot\left(1-x_{rcs}\right)\,\mid\, r, c, s \in [n]\,\rangle + \langle\,x_{rcs}-x_{\alpha(r)\beta(c)\gamma(s)}\,\mid\, r, c, s \in [n] \,\rangle + \langle\,m-\sum_{r,c,s\in [n]} x_{rcs}\,\rangle\,\subseteq \mathbb{Q}[{\bf x_n}] = \mathbb{Q}[x_{111},...,$ $x_{nnn}]$. $\Box$
The ideal $I$ of Corollary \[corP1\] is [*zero-dimensional*]{}, i.e., there exists only a finite number of solutions of the corresponding system of polynomial equations. Moreover, $I\cap \mathbb{Q}[x_{rcs}]=\langle\,x_{rcs}\cdot(1-x_{rcs})\,\rangle\subseteq I$, for all $r,c,s\in [n]$, and, therefore, Proposition 2.7 of [@Cox98] implies $I$ to be [*radical*]{}, i.e., any polynomial $p({\bf x_n})$ belongs to $I$ whenever there exists $t\in \mathbb{N}$ such that $p({\bf x_n})^t\in I$. Since the affine variety defined by $I$ is $V(I)=\mathcal{PLS}_{\Theta,m}$, then Theorem 2.10 of [@Cox98] assures $\Delta_m(z_{\Theta})=|V(I)|=\mathrm{dim}_{\mathbb{Q}}(\mathbb{Q}[{\bf x_n}]/I)$, which can be computed from any Gröbner basis of $I$, with respect to any term ordering. So, for instance, [Singular]{} [@Decker11] has been used in order to obtain the values of $\Delta_s(z)$ and $\Delta_{\mathcal{P}}(z)$ for each parastrophic class of ${\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$, where $n\leq 4$. These values are shown in Tables 2 and 3, where the blank cells correspond to those $s\not\in\mathrm{Sizes}(z)$.
--- --------------------------------- ---- ----- ------ ------ ------ ------ ----- ----- ---- -------
1 2 3 4 5 6 7 8 9
1 $(1,1,1)$ 1 1
(2,2,2) 4 0 4
$(2,2,1^2)$ 4 2 6
$(1^2,1^2,1^2)$ 8 16 8 2 34
(3,3,3) 9 9 3 21
(3,3,2$\cdot$1) 3 3
$(3,3,1^3)$ 9 18 6 33
(2$\cdot$1,2$\cdot$1,2$\cdot$1) 1 10 10 24 24 20 20 4 4 117
(2$\cdot$1,2$\cdot$1,$1^3$) 3 6 18 6 18 51
(2$\cdot$1,$1^3$,$1^3$) 9 18 6 33
$(1^3,1^3,1^3)$ 27 270 1278 3078 3834 2412 756 108 12 11775
--- --------------------------------- ---- ----- ------ ------ ------ ------ ----- ----- ---- -------
: $\Delta_s(z)$ and $\Delta_{\mathcal{P}}(z)$ for each parastrophic class of ${\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$, where $n\leq 3$.
--------------------------------------- ---- ------ ------- -------- --------- --------- ---------- ---------- ---------- ---------- ---------- --------- -------- -------- ------ ----- ----------- --
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
(4,4,4) 16 48 32 0 96
(4,4,3$\cdot$1) 4 4
(4,4,$2^2$) 16 56 32 8 112
(4,4,2$\cdot 1^2$) 16 64 64 8 152
(4,4,$1^4$) 16 72 96 24 208
(3$\cdot$1,3$\cdot$1,3$\cdot$1) 1 18 18 90 90 165 165 99 99 9 9 763
(3$\cdot$1,3$\cdot$1,2$\cdot1^2$) 2 6 12 6 12 38
(3$\cdot$1,3$\cdot$1,$1^4$) 4 12 48 36 144 24 96 364
(3$\cdot$1,$2^2$,$2^2$) 8 8 16
(3$\cdot$1,$2^2$,2$\cdot1^2$) 4 4
(3$\cdot$1,2$\cdot1^2$,2$\cdot1^2$) 4 4 8 4 20
(3$\cdot$1,2$\cdot1^2$,$1^4$) 8 12 20
(3$\cdot$1,$1^4$,$1^4$) 16 72 96 24 208
($2^2$,$2^2$,$2^2$) 32 352 1664 3552 3328 1408 256 32 10624
($2^2$,$2^2$,2$\cdot1^2$) 32 360 1792 4152 4416 2048 384 32 13216
($2^2$,$2^2$,$1^4$) 32 368 1920 4800 5760 3264 768 96 17008
($2^2$,2$\cdot1^2$,2$\cdot1^2$) 24 192 640 880 416 32 2184
($2^2$,2$\cdot1^2$,$1^4$) 16 72 96 24 208
(2$\cdot1^2$,2$\cdot1^2$,2$\cdot1^2$) 8 32 136 336 752 1440 1904 2856 2400 2608 1504 1056 448 224 64 16 15784
(2$\cdot1^2$,2$\cdot1^2$,$1^4$) 16 88 272 736 1344 1632 1728 1008 6824
(2$\cdot1^2$,$1^4$,$1^4$) 28 352 2208 6504 9792 7104 2112 216 28352
($1^4$,$1^4$,$1^4$) 64 1728 25920 239760 1437696 5728896 15326208 27534816 32971008 25941504 13153536 4215744 847872 110592 9216 576 127545136
--------------------------------------- ---- ------ ------- -------- --------- --------- ---------- ---------- ---------- ---------- ---------- --------- -------- -------- ------ ----- ----------- --
$\Theta$-completable partial Latin squares.
===========================================
In the Introduction, given $z\in {\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$ and $\Theta\in\mathfrak{I}_z$, it has been indicated that a partial Latin square $P\in\mathcal{PLS}_{\Theta}$ can be used in the computation of $\Delta(z)$, if a multiplicative factor ($P$-coefficient of symmetry of $\Theta$ [@FalconMartinJSC07]) $c_P\in\mathbb{N}$ is found such that $\Delta(z)=c_P\cdot |\mathcal{LS}_{\Theta,P}|$. In this regard, let us finish the present study with a theoretical basis for this concept of coefficient of symmetry. To do it, it is necessary to generalize the traditional concept of completability. Specifically, $P$ will be said to be [*$\Theta$-completable*]{} if $\mathcal{LS}_{\Theta,P}\neq\emptyset$. So, the traditional completability corresponds to the trivial isotopism $\Theta=(\mathrm{Id},\mathrm{Id},\mathrm{Id})$. Moreover, let us observe that, if a partial Latin square is $\Theta$-completable, then it is also completable in the traditional way.
It can be easily checked that every partial Latin square of order $n\leq 2$ is $\Theta$-completable whenever $\Theta$ is non-trivial. An example of non-trivial isotopism for which there exists a related partial Latin square which is neither $\Theta$-completable nor traditionally completable is $\Theta=((12)(3),(12)(3),(12)(3))\in\mathfrak{A}_3$. A partial Latin square in such conditions is:[$$\left(\begin{array}{ccc}3 & \cdot & 2\\
\cdot & 3 & 1\\
2 & 1 & \cdot\end{array}\right).$$]{}
An example where it is possible to observe the difference between both concepts is given if $\Theta=((12)(34),(12)(34),(12)(3)(4))\in\mathfrak{A}_4$. In this case, the following partial Latin square is not $\Theta$-completable, but it is completable in the traditional way:[$$\left(\begin{array}{cccc}3 & 4 & \cdot & \cdot\\
4 & 3 & \cdot & \cdot\\
\cdot & \cdot & \cdot & \cdot\\
\cdot & \cdot & \cdot & \cdot\end{array}\right).$$]{}
Let $C_{\Theta}$ denote the set of $\Theta$-completable partial Latin squares and let $C_{\Theta,s}=C_{\Theta}\cap\mathcal{PLS}_{\Theta,s}$. The cardinalities of these sets only depends on the parastrophic class of the cycle structure of $\Theta$:
\[lemT1\] Let $\Theta_1,\Theta_2\in \mathfrak{I}_n$ be such that $[z_{\Theta_1}]=[z_{\Theta_2}]$. Then, $|C_{\Theta_1,s}|=|C_{\Theta_2,s}|$, for all $s\in[n^2]$. As a consequence, $|C_{\Theta_1}|=|C_{\Theta_2}|$.
[**Proof.**]{} Let $\pi\in S_3$ be such that $z_{\Theta_1}=z_{\Theta^{\pi}_2}$. Given $s\in [n^2]$ and $P\in C_{\Theta_1,s}$, there exists $L\in \mathcal{LS}_{\Theta_1}$ such that $O(P)\subseteq O(L)$. Besides, since $\Theta_1$ and $\Theta^{\pi}_2$ are conjugate, then there exists $\Theta\in\mathfrak{I}_n$ such that $\Theta^{\pi}_2=\Theta\Theta_1\Theta^{-1}$. Thus, $(P^{\Theta})^{\pi^{-1}}\in\mathcal{PLS}_{\Theta_2}$, $(L^{\Theta})^{\pi^{-1}}\in\mathcal{LS}_{\Theta_2}$ and $O((P^{\Theta})^{\pi^{-1}})\subseteq O((L^{\Theta})^{\pi^{-1}})$. Since $|(P^{\Theta})^{\pi^{-1}}|=|P|$, then $|C_{\Theta_1,s}|\leq |C_{\Theta_2,s}|$. The opposite inequality is analogously proven and the consequence is immediate, because $|C_{\Theta_1}|=\sum_{s\in [n^2]}|C_{\Theta_1,s}|=\sum_{s\in [n^2]}|C_{\Theta_2,s}|=|C_{\Theta_2}|$. $\Box$
From the previous result, it is natural to define the numbers $\mathfrak{c}_z$ and $\mathfrak{c}_{z,s}$ as the respective cardinalities of $C_{\Theta}$ and $C_{\Theta,s}$, for any $\Theta\in\mathfrak{I}_z$. The following result holds:
\[thmT1\] Let $z\in {\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$ and $\Theta\in\mathfrak{I}_z$. It is verified that $P\in C_{\Theta}$ if and only if $\mathcal{PLS}_{\Theta,[P]}\subseteq C_{\Theta,|P|}$. As a consequence: $$\mathfrak{c}_{z,s}=\sum_{\scriptsize \begin{array}{c}[P]\in \mathcal{PLS}_{\Theta,s}/\sim\\ \text{s.t. }[P]\cap C_{\Theta}\neq \emptyset\end{array}}
\Delta_{[P]}(z).$$
[**Proof.**]{} The sufficient condition is immediate. So, let us consider $P\in C_{\Theta}$ and $Q\in \mathcal{PLS}_{\Theta,[P]}$. Since $Q\in [P]$ and $\Theta\in\mathfrak{A}_P\cap\mathfrak{A}_Q$, Theorem \[thmD2\] implies that there exists $\Theta'\in\mathfrak{I}_{P,Q}$ such that $\Theta=\Theta'\Theta\Theta'^{-1}$. Thus, $\Theta\Theta'=\Theta'\Theta$. Now, let $L\in\mathcal{LS}_{\Theta}$ be such that $O(P)\subseteq O(L)$. It must be then $L^{\Theta'}\in\mathcal{LS}_{\Theta}$, because $(L^{\Theta'})^{\Theta}=(L^{\Theta})^{\Theta'}=L^{\Theta'}$. Since $|Q|=|P|$ and $O(Q)=O(P^{\Theta'})\subseteq O(L^{\Theta'})$, then $Q\in C_{\Theta,|P|}$ and the first claim is verified. The consequence is then immediate. $\Box$
The previous theorem implies that it is enough to check the completability of one element of each isotopic class of partial Latin squares. Moreover, it is convenient to do it in increasing order of the size, because, given $\Theta\in\mathfrak{I}_n$ and $P,Q\in\mathcal{PLS}_{\Theta}$ such that $O(P)\subseteq O(Q)$, if $P$ is not $\Theta$-completable, neither is $Q$. Taking into account this strategy, the numbers $\mathfrak{c}_{z,s}$ and $\mathfrak{c}_z$ have been obtained (Table 5) for each non-trivial parastrophic class of $\mathcal{CS}_{{\mathfrak{A}}_n}$, where $n\leq 4$.
--- --------------------------------------- --- ---- ----- ----- ----- ------ ------ ------ ------ ------ ------ ------ ----- ----- ---- ---- -------
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
1 $(1,1,1)$ 1 1
2 $(2,2,1^2)$ 4 2 6
(3,3,3) 9 9 3 21
$(3,3,1^3)$ 9 18 6 33
(2$\cdot$1,2$\cdot$1,2$\cdot$1) 1 10 10 24 24 16 16 4 4 109
(4,4,$2^2$) 16 40 32 8 96
(4,4,2$\cdot 1^2$) 16 40 32 8 96
(4,4,$1^4$) 16 72 96 24 208
(3$\cdot$1,3$\cdot$1,3$\cdot$1) 1 18 18 90 90 90 90 45 45 9 9 505
($2^2$,$2^2$,$2^2$) 32 352 1408 2144 1792 896 256 32 6912
($2^2$,$2^2$,2$\cdot1^2$) 32 336 1344 2144 1792 896 256 32 6832
($2^2$,$2^2$,$1^4$) 32 368 1728 3792 4224 2496 768 96 13504
(2$\cdot1^2$,2$\cdot1^2$,2$\cdot1^2$) 8 32 136 200 528 784 1328 1560 1760 1568 1248 800 448 192 64 16 10672
--- --------------------------------------- --- ---- ----- ----- ----- ------ ------ ------ ------ ------ ------ ------ ----- ----- ---- ---- -------
: $\mathfrak{c}_{z,s}$ and $\mathfrak{c}_{z}$ for each non-trivial parastrophic class of ${\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$, where $n\leq 4$.
Given $z\in \mathcal{CS}_{\mathfrak{A}_{\mathcal{P}_n}}$ and $\Theta\in\mathfrak{I}_z$, a set $\{P_1,P_2,...,P_m\}$ of $\Theta$-completable partial Latin squares will be said to be a [*basis*]{} of $\mathcal{LS}_{\Theta}$ if $\bigcup_{i\in [m]} \mathcal{LS}_{\Theta,P_i} = \mathcal{LS}_{\Theta}$ and $\mathcal{LS}_{\Theta,P_i}\cap\mathcal{LS}_{\Theta,P_j}=\emptyset$, whenever $i\neq j$. In this case, $\Delta(z)=\sum_{i\in [m]}|\mathcal{LS}_{\Theta,P_i}|$. Let us observe that, from a computational point of view, it is interesting to determine a basis of $\mathcal{LS}_{\Theta}$ such that the sizes of its elements are as great as it is possible, because then, for each $P_i$, it would be feasible to add to the constraints of Proposition \[prpP1\], all those of the form $x_{rcs}=1$, if $(r,c,s)\in O(P_i)$. The calculus of the corresponding Gröbner basis would be then more efficient and it would allow to obtain new values $\Delta(z)$. The following result is proven:
\[lemT2\] Let $S\subseteq [n]^2$ and $\Theta=(\alpha,\beta,\gamma)\in\mathfrak{A}_n$. Each of the following sets is non-empty if and only if it is a basis of $\mathcal{LS}_{\Theta}$: $$S_{RC}=\{P\in C_{\Theta}\,\mid\, (r,c,s)\in O(P) \Leftrightarrow (r,c)\in S\},$$ $$S_{RS}=\{P\in C_{\Theta}\,\mid\, (r,c,s)\in O(P) \Leftrightarrow (r,s)\in S\},$$ $$S_{CS}=\{P\in C_{\Theta}\,\mid\, (r,c,s)\in O(P) \Leftrightarrow (c,s)\in S\}.$$ $\Box$
[**Proof.**]{} The sufficient condition is immediate. In order to see the necessary condition, let us prove that $S_{RC}$ is a basis of $\mathcal{LS}_{\Theta}$; the other cases are similar. Since $S_{RC}$ is non-empty, there exists $P_0\in C_{\Theta}$ such that $(r,c,s)\in O(P_0) \Leftrightarrow (r,c)\in S$. Given $L\in \mathcal{LS}_{\Theta}$, let $P\in\mathcal{PLS}_n$ be such that $O(P)=\{(r,c,s)\in O(L)\,\mid\, (r,c)\in S\}$ and let us consider $(r,c,s)\in O(P)$. It must be $(r,c)\in S$ and so, there must exist $s_0\in [n]$ such that $(r,c,s_0)\in O(P_0)$. Since $P_0\in \mathcal{PLS}_{\Theta}$, it must be $(\alpha(r),\beta(c),\gamma(s_0))\in O(P_0)$ and therefore, $(\alpha(r),\beta(c))\in S$. Hence, $(\alpha(r),\beta(c),\gamma(s))\in O(P)$ and thus, $P\in\mathcal{PLS}_{\Theta}$. As a consequence, $P$ is $\Theta$-completable and then, $\mathcal{LS}_{\Theta}\subseteq \bigcup_{Q\in S_{RC}} \mathcal{LS}_{\Theta,Q}$. Indeed, both sets are equal because $\mathcal{LS}_{\Theta,Q}\subseteq\mathcal{LS}_{\Theta}$, for all $Q\in S_{RC}$. Finally, given two distinct elements $Q,Q'\in S_{RC}$, it must exist $(r,c)\in S$ and $s\in [n]$ such that $(r,c,s)\in O(Q)\setminus O(Q')$. It implies that $\mathcal{LS}_{\Theta,Q}\cap\mathcal{LS}_{\Theta,Q'}=\emptyset$ and therefore, $S_{RC}$ is a basis of $\mathcal{LS}_{\Theta}$. $\Box$
A special case appears when $|\mathcal{LS}_{\Theta,P_i}|=|\mathcal{LS}_{\Theta,P_j}|$, for all $i,j\in [m]$. Such a basis will be called [*homogeneous*]{} and it will be verified that $\Delta(z)=m\cdot |\mathcal{LS}_{\Theta,P_i}|$, for all $i\in[m]$. The cardinality $m$ of the homogeneous basis would be therefore the $P_i$-coefficient of symmetry of $\Theta$, for all $i\in [m]$. Although a comprehensive study must be developed in this regard, let us finish the current paper with a result with gives a theoretical support to the majority of the coefficients of symmetry which were used in [@FalconMartinJSC07]:
\[thmT2\] Let $z=(z_1,z_2,z_3)\in \mathcal{CS}_{\mathfrak{A}_n}$ be such that $z_{11}\cdot z_{21}\cdot z_{31}\neq 0$. Let $\Theta=(\alpha,\beta,\gamma)\in\mathfrak{I}_z$ and $S=\{(i,j)\in [n]^2\,\mid\, i\in\alpha_{\infty}, j\in \beta_{\infty}\}$. It is verified that $S_{RC}$ is an homogeneous basis of $\mathcal{LS}_{\Theta}$ of cardinality $|\mathcal{LS}_{z_{11}}|$.
[**Proof.**]{} From the hypothesis, it must be $z_1=z_2=z_3$ ([@McKay05], Theorem 1). Furthermore, given $P\in\mathcal{PLS}_{\Theta}$, the corresponding block $P_{\infty\infty}$ of the $\Theta$-decomposition of $P$ is a $z_{11}\times z_{11}$-array, such that each of its non-filled cells must contain one of the $z_{11}$ fixed symbols of $\gamma$, i.e., it is a Latin subsquare of $P$ of order $z_{11}$. Thus, since $\Theta\in \mathfrak{A}_n$, Lemma \[lemT2\] implies the set $S_{RC}$ to be a basis of $\mathcal{LS}_{\Theta}$ of $|\mathcal{LS}_{z_{11}}|$ elements. Now, let us consider two distinct elements $Q,Q'\in S_{RC}$. Given $L\in\mathcal{LS}_{\Theta,Q}$, let us define the Latin square $L'\in\mathcal{LS}_n$ such that $O(L')=\{(r,c,s)\in [n]^3\,\mid\, (r,c,s)\in O(Q') \text{ if } (r,c)\in S, \text{ or } (r,c,s)\in O(L), \text{ otherwise}\}$, i.e., the only difference of $L'$ with respect to $L$ is the block $L'_{\infty\infty}$, which is $Q'$ instead of $Q$. Since $L\in \mathcal{LS}_{\Theta}$ and $Q'\in\mathcal{PLS}_{\Theta}$, it must be $L'\in\mathcal{LS}_{\Theta}$. Hence, $|\mathcal{LS}_{\Theta,Q}|\leq |\mathcal{LS}_{\Theta,Q'}|$. The opposite inequality is analogously proven and, therefore, $S_{RC}$ is homogeneous. $\Box$
Final remarks and further work.
===============================
In the current paper, it has been dealt with the set of autotopisms of partial Latin squares in order to develop further techniques which allow to improve some results about the set of autotopisms of Latin squares, such as those related with the obtention of the values $\Delta(z)$. In Section 2, the cardinality of ${\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n$ has been studied and a lower bound has been determined. Although it can be obtained by an exhaustive search once Lemma \[lem0\] is implemented in a computer procedure, it raises the question of whether it is possible to obtain a general formula for $|{\mathcal{CS}_{\mathfrak{A}_{\mathcal{P}}}}_n|$. A similar question appears in Section 4 with the values $\Delta_{[P]}(z)$, for which a comprehensive study of isotopic classes of $\mathcal{PLS}_n$ would be necessary. It would also be useful in order to improve the computation and increase the order $n\leq 4$ which have been used in the examples of the present paper. Finally, once a theoretical basis has been exposed in Section 5 for the concept of coefficient of symmetry of an autotopism, it seems that an exhaustive study in this regard would be necessary to solve some of the problems of computation related to the calculus of the values $\Delta(z)$.
[00]{}
D. Bryant, M. Buchanan and I. M. Wanless, The spectrum for quasigroups with cyclic automorphisms and additional symmetries, Discrete Math. 304 (2009), 821-833.
D. A. Cox, J. B. Little and D. O’Shea, Using Algebraic Geometry, Springer-Verlag, New York, 1998.
W. Decker, G.-M. Greuel, G. Pfister and H. Schönemann, [Singular]{} 3-1-3. A computer algebra system for polynomial computations, 2011. http://www.singular.uni-kl.de.
R. Euler, R. E. Burkard and R. Grommes, On Latin squares and the facial structure of related polytopes, Discrete Math. 62 (1986), 155-181.
R. M. Falcón, Latin squares associated to principal autotopisms of long cycles. Application in Cryptography, Proceedings of Transgressive Computing 2006: a conference in honor of Jean Della Dora (2006), 213-230.
R. M. Falcón, Cycle structures of autotopisms of the Latin squares of order up to 11, [*Ars Combinatoria*]{} (in press). Available from http://arxiv.org/abs/0709.2973.
R. M. Falcón, Designs based on the cycle structure of a Latin square autotopism, Proceedings of 1st Hispano-Moroccan Days on Applied Mathematics and Statistics (2008), 479-484.
R. M. Falcón and J. Martín-Morales, Gröbner bases and the number of Latin squares related to autotopisms of order $\leq$ 7, J. Symbolic Comput. 42 (2007), 1142-1154.
R. M. Falcón and J. Martín-Morales, The 3-dimensional planar assignment problem and the number of Latin squares related to an autotopism, Proceedings of XI Spanish Meeting on Computational Algebra and Applications (2008), 89-92.
M. Ghandehari, H. Hatami, E. S. Mahmoodian, On the size of the minimum critical set of a Latin square, Discrete Math. 293 (2005) 121-127.
G. H. Hardy, S. and Ramanujan, Asymptotic Formulae in Combinatory Analysis, Proc. London Math. Soc. 17 (1918), 75-115.
A. Hulpke, P. Kaski and P. R. J. Österg[å]{}rd, The number of Latin squares of order 11, Math. Comp. 80 (2011), 1197-1219.
B. Kerby and J. D. H. Smith, Quasigroup automorphisms and symmetric group characters, Comment. Math. Univ. Carol. 51 (2010), 279-286.
B. Kerby and J. D. H. Smith, Quasigroup automorphisms and the Norton-Stein complex, Proc. Amer. Math. Soc. 138, No.9 (2010), 3079-3088.
S. R. Kumar, A. Russell and R. Sundaram, Approximating Latin square extensions, Algorithmica 24 (1999), 128–138.
C. Laywine, An expression for the number of equivalence classes of Latin squares under row and column permutations, J. Combin. Theory Ser. A 30 (1981), 317-321.
C. Laywine and G. L. Mullen, Latin cubes and hypercubes of prime order, Fibonacci Quart. 23 (1985), 139-145.
B. D. McKay, A. Meynert and W. Myrvold. Small Latin Squares, Quasigroups and Loops, J. Comb. Designs, 15, No.2 (2007), 98-119.
B. D. McKay. http://cs.anu.edu.au/$\sim$bdm/data/latin.html.
A. A. Sade, Autotopies des quasigroupes et des systèmes associatives, Arch. Math. 4, No. 1 (1968), 1-23.
D. S. Stones, The parity of the number of quasigroups, Discrete Math. 310 (2010), 3033-3039.
D. S. Stones, P. Vojtěchovský and I. Wanless, Cycle structure of autotopisms of quasigroups and Latin squares. Preprint avaliable from http://www.du.edu/media/documents/nsm/mathematics/preprints/ m1101.pdf
I. M. Wanless, Diagonally ciclic Latin squares, European J. Combin 25 (2004), 393-413.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The possibility to modify the strength of the Casimir effect by tailoring the dielectric functions of the interacting surfaces is regarded as a unique opportunity in the development of Micro- and NanoElectroMechanical Systems. In air, however, one expects that, unless noble metals are used, the electrostatic force arising from trapped charges overcomes the Casimir attraction, leaving no room for exploitation of Casimir force engineering at ambient conditions. Here we show that, in the presence of a conductive oxide, the Casimir force can be the dominant interaction even in air, and that the use of conductive oxides allows one to reduce the Casimir force up to a factor of 2 when compared to noble metals.'
author:
- 'S. de Man'
- 'K. Heeck'
- 'R. J. Wijngaarden'
- 'D. Iannuzzi'
title: Halving the Casimir force with conductive oxides
---
The mechanical parts of Micro- and NanoElectroMechanical Systems (MEMS and NEMS) are often designed to work at separations where the Casimir effect [@casimir] might play a relevant role [@vdwmems; @chanscience; @channonlinear; @memsstiction]. It is thus commonly believed that, if one could suitably engineer the strength of the Casimir force, unprecedented opportunities would come available for the development of conceptually new MEMS and NEMS [@ieeecapasso; @repulsivevacuum; @torque; @capassonature2008]. The most simple approach to tailor the Casimir force is to properly choose the materials of which the interacting surfaces are made. According to the Lifshitz theory [@lifshitz], the interaction between two objects depends on their dielectric functions. Transparent dielectrics, for example, attract less than reflective mirrors. This property may be used to reduce the Casimir attraction whenever the design requires a smaller short range interaction. It is however fair to say that, for the vast majority of applications, MEMS and NEMS operate in air, where surfaces tend to accumulate trapped charges. Those charges give rise to a strong electrostatic interaction that cannot be compensated by a counterbias voltage and that typically overcomes the Casimir force. It is thus difficult to imagine that the Casimir force can play an important role in MEMS and NEMS operating in air, unless all surfaces are coated with noble metals to reduce the forces due to surface charges to negligible levels. In that case, however, there is not much room to tune the strength of the Casimir interaction because the diversity in the dielectric functions of different metals is simply not large enough [@deccaCu; @pnas; @light; @skin; @mohideenSi]. As a matter of fact, to date, there is no experiment that shows that, in air, the Casimir force can still be tuned significantly while remaining the dominant interaction mechanism.
In this paper we present a precise measurement of the Casimir force between a gold coated sphere and a glass plate coated with either a thick gold layer or a highly conductive, transparent oxide film. The experiment was performed in air, and no electrostatic force due to residual trapped charges was observed over several weeks of measurements in either case. The decrease of the Casimir force due to the different dielectric properties of the reflective gold layer and the transparent oxide film resulted to be as high as $\simeq 40\% - 50\%$ at all separations (from $\simeq 50$ to $\simeq
150$ nm). Our experiment shows that, in the presence of a conductive oxide layer, the Casimir force can still be the dominant interaction mechanism even in air, and indicates that, whenever the design might require it, it is possible to tune the Casimir attraction by a factor of 2.
Our experimental set-up is designed to perform precise measurements of surface forces between a 100 $\mu$m radius sphere and a plate as a function of their separation (see Fig. \[fig:schematic\] and [@pra]). The sphere is glued onto the hanging end of a micromachined cantilever (spring constant $\approx 1$ N/m, resonance frequency (with the sphere attached) $\approx 1.9$ kHz). The plate is mounted on a capacitive feedback controlled piezoelectric transducer that allows one to accurately vary the separation between the sphere and the plate in discrete steps. Any force acting between the two surfaces results in a bending of the cantilever that is detected by the optical lever of a commercial Atomic Force Microscope (AFM) head [@afm]. The set-up is kept at a fixed temperature to reduce mechanical drifts and is placed on an active anti-vibration stage inside an acoustic isolation box to decouple the force sensor from external vibrations.
Precise Casimir force measurements typically require careful analysis of three crucial issues. First, even when both the sphere and the plate are coated with metallic films, there might still exist an electrostatic potential difference $V_0$ between the two surfaces that gives rise to a residual electrostatic force. This force must be actively canceled by counter biasing $V_0$ with an externally applied voltage. Second, although the relative displacements of the piezoelectric transducer that moves the plate, $d_{pz}$, are precisely controlled, the separation between the two interacting surfaces at the start of the measurement, $d_0$, is *a priori* unknown (see Fig. \[fig:schematic\]). Therefore, the absolute separation $d=d_0-d_{pz}$ has to be obtained from a calibration procedure. Third, the electronic signal coming out of the AFM head must be converted into force. It is thus necessary to calibrate the instrument with a controlled force. To address these problems, we have designed a measurement technique that allows one to simultaneously: (i) compensate for the residual potential, (ii) calibrate the set-up, and (iii) measure the Casimir force [@pra; @inprep]. In a nutshell, in each calibration/measurement run the plate is moved in discrete steps from $d_0$ to a minimum value of $d$ (just before contact with the sphere). A calibrated AC electrostatic potential is applied between the sphere and the plate at a frequency $\omega_1$ much lower than the resonance frequency of the force sensor ($\frac{\omega_1}{2\pi}=72.2$ Hz). This AC excitation is used to drive a feedback circuit that compensates for the residual voltage $V_0$, and, simultaneously, generates an electrostatic force that makes the cantilever oscillate at $2\omega_1$. The amplitude of the oscillations, measured with a lock-in amplifier (calibration lock-in in Fig. \[fig:schematic\]b), are recorded as a function of $d_{pz}$, and are then used to calibrate the instrument and extract $d_0$. At the same time, a transducer mechanically coupled to the piezoelectric translator makes the plate move around $d_{pz}$ with an amplitude of $3.85 \pm
0.08$ nm at a frequency $\omega_2$, which is again much lower than the resonance frequency of the force sensor ($\frac{\omega_2}{2\pi}=119$ Hz) [@carugnoruoso]. In the presence of a force that depends on separation (e.g., the Casimir force), the cantilever bends in phase with the modulation of $d$. The amplitude of the in-phase oscillation, measured with another lock-in amplifier (measurement lock-in in Fig. \[fig:schematic\]b), is proportional to the derivative of the force with respect to $d$. Furthermore, the presence of the cushion of air between the two surfaces gives rise to a hydrodynamic force that depends on the velocity with which the plate moves. The signal produced by this force manifests itself at the same frequency at which $d$ is modulated, but with a phase rotated by 90 degrees. This contribution does not influence the output of the in-phase component and can be measured independently with the same lock-in amplifier. The integration times of the lock-in amplifiers are 8 s for every value of $d_{pz}$. A typical measurement run consists of $\simeq 50$ $d_{pz}$ set-points in the measurement range $50 < d < 1100$ nm, and takes roughly 7 minutes. The cantilever responses to the modulations at $\omega_1$ and $\omega_2$ are $<80$ pm (root-mean-square) during the entire experiment. All force measurements are performed in air at atmospheric pressure, temperature 300 K, and 29% relative humidity.
In this paper we present two experiments performed with the same sphere and two different plates. The sphere is a polystyrene sphere with nominal radius 100 $\mu$m coated with a Ti adhesion layer followed by a 100 nm Au film (surface roughness 3.8 nm RMS). The plate used in the first experiment is a polished sapphire substrate coated with a metallic film similar to the one deposited on the sphere (surface rougness 0.8 nm RMS). The plate used in the second experiment is a float glass substrate with a 190 nm Indium Tin Oxide (ITO, In$_2$O$_3$:Sn) sputtered thin film on top (PGO CEC010S, typically $8.5~\Omega/\Box$, or, equivalently, $\rho =
1.6~10^{-4}~\Omega$cm, total surface roughness 4 nm RMS). We have measured the reflection and transmission spectra of both plates in the wavelength range $180~\mathrm{nm} < \lambda < 2.5~\mu$m, and observed that the optical properties of our films are in agreement with the literature [@palik; @fujiwara].
Fig. \[fig:forceresults\]a presents measurements of the force between the sphere and the plates coated with either Au or ITO [@noteond0]. The experimental data represent the spatial derivative of the total force (normalized by the sphere radius $R$), which is the sum of the Casimir interaction, a Coulomb interaction induced by the presence of trapped charges (if any), and an electrostatic attraction due to the AC calibration potential. The strength of the latter can be estimated from the simultaneous calibration procedure [@inprep]. From Fig. \[fig:forceresults\], it is evident that this electrostatic contribution, which is anyway equal in both experiments (within 2%), is small compared to the total force signal. The black lines in Fig. \[fig:forceresults\]a are computations of the Casimir force using the Lifshitz theory [@parsegian] with dielectric functions calculated as in [@fujiwara] (for ITO) and [@deman] (for Au); the electrostatic force due to the calibration potential is added to the theory in order to compare with the raw data. The calculation of the Casimir force should only be considered approximate, because the dielectric functions of the samples are not known precisely [@svetovoy], and no surface roughness corrections are applied [@noteond0]. Still, the agreement between the calculation and the data shows that the Casimir effect largely dominates any Coulomb interaction that would have been otherwise observed in the presence of a significant amount of trapped charges. Fig. \[fig:forceresults\]b shows the data and theory on a double logarithmic scale, where we have subtracted the electrostatic background due to the AC calibration potential using the simultaneously obtained calibration data. At small separations $d<60$ nm, both data sets curve upwards, which might be a sign of surface roughness effects. At separations $d>120$ nm, the experimental data for the Au-Au Casimir interaction start to deviate significantly from the theory because of an artefact caused by reflections from the optical lever light by the sample. This reflected light reaches the photodetector and causes a background signal that is not related to any force. This artefact is common to all optical lever based AFM techniques, and the related signal is typically assumed to be linear in the piezo extension and subtracted from the data accordingly [@mohideenPRLold]. Because the reflectivities of our two samples are so different, we prefer to refrain from such a procedure, and present the data as is. In Figs. \[fig:forceresults\]c and d we show histograms of all the obtained measurements for the derivative of the Casimir force for two specific separations ($d=120$ nm and $d=80$ nm). The histograms at $d=80$ nm can be described by Gaussians with a standard deviation of roughly 5%, which means that our method provides a precision in the mean measured Casimir force derivative of 0.2%. It is evident from the histograms that the spatial derivative of the Casimir force between a Au and an ITO surface is roughly $\simeq 40\% - 50\%$ smaller than between two Au surfaces. In our geometry (i.e., for separations much smaller than the radius of the sphere), the spatial derivative of the force is proportional to the pressure between two parallel plates [@parsegian]. We can thus conclude that the Casimir pressure that one would measure between a Au plate kept parallel to an ITO plate would be roughly $\simeq 40\% - 50\%$ smaller than in the case of two Au plates.
Even though the agreement between the theoretical prediction and the measurement of the Casimir force in both situations is good, one might still argue that the observed decrease could be mimicked by drifts in $d_0$. For both measurements series, we gathered 580 data sets continuously, which allows us to directly assess the run-to-run drift of $d_0$. Due to the temperature stabilization of our setup, the mechanical drift is very small at $\simeq 0.1$ nm and $\simeq
0.2$ nm per run for the Au-Au and the Au-ITO experiments, respectively. We conclude that the decrease of the Casimir force cannot be ascribed to drifts in $d_0$. We have also verified that the electrostatic force used to calibrate the instrument and extract $d_0$ follows what expected from elementary electrostatic arguments, as suggested in [@roberto] and discussed in [@pra]. Concerning the compensation voltage, we observed that $V_0$ varies approximately $1$ mV and $3$ mV over the complete measurement range in the Au-Au and Au-ITO cases, respectively [@note_on_v0]. These slight variations of $V_0$ do not compromise the measurement of the Casimir force at the current level of sensitivity.
Finally, the different surface roughnesses of the sphere and the plates also influence the strength of the Casimir effect. Since both experiments are conducted with the same sphere, the difference in the observed Casimir force can never be due to the surface roughness of the sphere. Second, we recall that the surface roughness of the ITO sample is larger than that of the Au substrate. Since surface roughness enhances the Casimir force [@maradudin] we note that, if it played a role in the probed separation range, it would lead to a stronger interaction between Au and ITO than between two Au surfaces, contrary to the measurements presented in Fig. \[fig:forceresults\].
To make our claim even more robust, we can now compare the hydrodynamic force observed during the two experiments. Because the geometrical configuration of the experiment is equal in both cases, we expect to measure the same hydrodynamic force. The results are shown in Fig. \[fig:hydroresults\]. It is clear that the hydrodynamic force is very similar in both cases, although there exists a slight discrepancy between the two curves ($\simeq 2\%$). Since both curves are parallel on a double logarithmic scale, we conclude that this discrepancy cannot be ascribed to a difference in the calibration of $d_0$. Therefore, we rule out that the large difference in the Casimir force reported in Fig. \[fig:forceresults\] be due to an error in the determination of $d_0$.
In conclusion, we have demonstrated that an ITO coating of one of the two surfaces is sufficient to readily create situations where the Casimir force is still the dominant interaction mechanism regardless the presence of air in the surroundings. Since ITO is transparent over a wide range of frequencies, the Casimir attraction is up to a factor of 2 smaller when compared to the case of the Au-Au interaction, leaving ample room for Casimir force engineering even at ambient conditions, where MEMS and NEMS typically operate.
The authors thank A. Baldi, F. Mul, U. Mohideen, J. H. Rector, B. Dam, and R. Griessen for useful discussions. This work was supported by the Netherlands Organisation for Scientific Research (NWO), under the Innovational Research Incentives Scheme VIDI-680-47-209. D.I. acknowledges financial support from the European Research Council under the European Community’s Seventh Framework Programme (FP7/2007-2013)/ERC grant agreement 201739.
[99]{}
H. B. G. Casimir, Proc. K. Ned. Akad. Wet. **60**, 793 (1948). F. W. Delrio *et al.*, Nature Materials **4**, 629 (2005). H. B. Chan *et al.*, Science **291**, 1941 (2001). H. B. Chan *et al.*, Phys. Rev. Lett. **87**, 211801 (2001). F. M. Serry, D. Walliser, and G. J. Maclay, J. Appl. Phys. **84**, 2501 (1997). F. Capasso *et al.*, IEEE J. Select. Topics Quant. Electr. **13**, 400 (2007). O. Kenneth *et al.*, Phys. Rev. Lett. **89**, 033001 (2002). J. N. Munday *et al.*, Phys. Rev. A **71**, 042102 (2005). J. N. Munday, F. Capasso, and V. A. Parsegian, Nature **457**, 170 (2009). E. M. Lifshitz, Sov. Phys. JETP **2**, 73 (1956). R. S. Decca *et al.*, Phys. Rev. Lett. **91**, 050402 (2003). An unsuccesful attempt to change the Casimir force in gas is reported in D. Iannuzzi, M. Lisanti, and F. Capasso, Proc. Natl. Acad. Sci. USA **101**, 4019 (2004). $\simeq 1\%$ tunability of the Casimir force in vacuum was shown in F. Chen *et al.*, Phys. Rev. B **76**, 035338 (2007). $\simeq 20\%$ tunability of the Casimir force was observed in M. Lisanti, D. Iannuzzi, and F. Capasso, Proc. Natl. Acad. Sci. USA **102**, 11989 (2005). That method is not practical, because the interaction strength strongly varies on the local thickness of the film (see G. Jourdan, A. Lambrecht, F. Comin, and J. Chevrier, Europhys. Lett. **85** 175501 (2009)) $\simeq 30\%$ variations of the Casimir force were observed in U. Mohideen *et al.*, Phys. Rev. A **72**, 020101(R) (2005). Also this strategy is not practical for MEMS in air, because the passivation layer is not stable under atmospheric conditions. S. de Man, K. Heeck, and D. Iannuzzi, Phys. Rev. A **79**, 024102 (2009). Measurements of surface forces between a sphere and a plate with an AFM were first reported in W. A. Ducker, T. J. Senden, and R. M. Pashley, Nature **353** 239 (1991). Its utilisation for large range forces between metals was first introduced in [@mohideenPRLold]. U. Mohideen and A. Roy, Phys. Rev. Lett. **81** 4549 (1998). S. de Man *et al.*, in preparation. A similar technique was introduced in G. Bressi *et al.*, Class. Quantum Grav. **18**, 3943 (2001). E. D. Palik, *Handbook of Optical Constants of Solids* (Academic Press, San Diego, 1998). H. Fujiwara and M. Kondo, Phys. Rev. B **71**, 075109 (2005), with $N_{\mathrm{Hall}}=1.2~10^{21}~\mathrm{cm}^{-3}$. Our estimate of $d_0$ relies on the *proximity force approximation* [@parsegian], which is only valid for $d \ll R$. This assumption is not entirely correct in the probed separation range [@pra] and results in a systematic error in $d_0$ of $\simeq 1.4$ nm. Still, the corresponding underestimate of the separation is equal for both Au-Au and Au-ITO, and can thus be neglected in the comparison. V. A. Parsegian, *Van der Waals forces* (Cambridge University Press, New York, 2006). S. de Man and D. Iannuzzi, New J. Phys. **8**, 235 (2006). V. B. Svetovoy *et al.*, Phys. Rev. B **77**, 035439 (2008). W. J. Kim *et al.*, Phys. Rev. A **78**, 020101(R) (2008). The value of $V_0$ at $d=100$ nm during the entire experiment drifts from $-106$ mV to $-103$ mV for Au and from $72$ mV to $50$ mV for ITO. A. A. Maradudin and P. Mazur, Phys. Rev. B **22** 1677 (1980).
FIG. 1. **a,** Drawing of the experimental set-up. **b,** Schematic representation of the working principle of the experimental technique. **c,** Definition of $d_0$ (initial separation), $d_{pz}$ (movement of the piezoelectric stage), and $d$ (separation between the two surfaces).
FIG. 2. **a,** Spatial derivative of the total force as a function of absolute surface separation for the Au-Au (green squares) and Au-ITO (red squares) interactions for randomly chosen subsets of the data (150 out of 580 for both cases). The blue line represents the derivative of the electrostatic force caused by the simultaneous calibration procedure (common to both the gold and ITO measurements). The black lines indicate the calculated Casimir forces with the electrostatic background added. **b,** Spatial derivative of the Casimir force, with the electrostatic background subtracted from the data. The black lines correspond to the calculations of the Casimir force. **c,** Histograms of 580 force measurements for Au-Au and Au-ITO at $d=120$ nm. **d,** Same as c, but for $d=80$ nm.
FIG. 3. Hydrodynamic force acting on the sphere as a function of the absolute separation $d$ for Au-Au (green squares) and Au-ITO (red squares), for the data sets shown in Figs. \[fig:forceresults\]a and b.
![[]{data-label="fig:schematic"}](fig1){width="13cm"}
![[]{data-label="fig:forceresults"}](fig2){width="13cm"}
![[]{data-label="fig:hydroresults"}](fig3){width="13cm"}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Synchronization induced by long-range hydrodynamic interactions is attracting attention as a candidate mechanism behind coordinated beating of cilia and flagella. Here we consider a minimal model of hydrodynamic synchronization in the low Reynolds number limit. The model consists of rotors, each of which assumed to be a rigid bead making a fixed trajectory under periodically varying driving force. By a linear analysis, we derive the necessary and sufficient conditions for a pair of rotors to synchronize in phase. We also derive a non-linear evolution equation for their phase difference, which is reduced to minimization of an effective potential. The effective potential is calculated for a variety of trajectory shapes and geometries (either bulk or substrated), for which the stable and metastable states of the system are identified. Finite size of the trajectory induces asymmetry of the potential, which also depends sensitively on the tilt of the trajectory. Our results show that flexibility of cilia or flagella is [*not*]{} a requisite for their synchronized motion, in contrast to previous expectations. We discuss the possibility to directly implement the model and verify our results by optically driven colloids.'
author:
- Nariya Uchida
- Ramin Golestanian
title: Hydrodynamic Synchronization between Objects with Cyclic Rigid Trajectories
---
Introduction {#sec:intro}
============
Coordinated cyclic beating of elastic organelles such as cilia and eukaryotic flagella serve a multitude of functions in living organisms, ranging from motility to fluid transport and polarity symmetry breaking in developing embryos [@Gray; @bray; @hamada]. It has long been known that the beating cycle of cilia has a characteristic asymmetry, with two distinct parts described as power stroke and recovery stroke [@Blake-Sleigh], and that the cyclic pattern could lead to metachronal waves (of varying kinds) [@Knight-Jones] in dense arrays of cilia [@Blake72; @GL97; @GL99]. While the necessity of this asymmetry for generating symmetry breaking fluid flow or propulsion could be easily understood from the time-reversal symmetry properties of the Stokes equation for viscous fluid flow, what exactly constitute the minimal conditions for synchronization and coordination between two or more of such cyclically beating organelles is a subject of current investigation [@GYU2011].
There have been a number of systematic experimental studies in a variety of systems to probe whether viscous hydrodynamic interaction alone can lead to synchronization as Taylor [@Taylor] originally proposed. The experiments, which all verify the existence of hydrodynamic synchronization, range from studying macroscopic model flagella in highly viscous silicone oil (such that the low Reynolds number condition was maintained) [@Kim-PNAS-03; @QJ09] to probing the relative phase dynamics in pairs of beating eukaryotic flagella [@Goldstein-Science-09; @GP09], to tracking colloidal linear oscillators using optical tweezers equipped with feedback control [@Kotar-10] and light driven asymmetrically micro-fabricated rotors [@DiLeonardo-1]. Experiments on carpets of bacteria with active flagella [@Berg] and arrays of artificial magnetically actuated cilia [@Vilfan-10; @Shields-10; @Bartolo-1] have revealed collective effects mediated by hydrodynamic interactions, such as complex flow patterns and collective phase shifts.
Theoretical studies of metachronal coordination and synchronization of cyclically beating organelles have been performed using models and descriptions of varying levels of complexity, ranging from simple models of coupled oscillators to actuated beads and more elaborate elastic filament models [@Blake72; @GL97; @GL99; @LB02; @LJ68; @KP04; @RS05; @KN06; @VJ06; @RL06; @GJ07; @NEL08; @EL09; @UG10a; @UG10b; @UG11; @OV11]. While the more realistic beating elastic filament descriptions are crucial for understanding detailed properties of metachronal waves, the simpler actuated bead models that typically have a minimal number of degrees of freedom could be useful in understanding what key ingredients are needed for hydrodynamic synchronization to occur, and under what conditions such dynamical states could be stable. In the course the studies of actuated bead models, one of the questions that have been discussed in whether or not beads following rigid trajectories could lead to synchronization. The discussion started when it was shown that two rigid helices that are rotating under constant torque cannot synchronize [@KP04] while with an added a small flexibility, say to the axis of rotation, the system can synchronize [@RS05]. Further studies of the actuated bead model followed the prescription of always having a flexible element, and somehow this was later on erroneously interpreted by many authors as a necessary condition for synchronization. In an earlier publication [@UG11], we showed that flexibility is not a necessity, and that beads following rigid trajectories could lead to synchronization provided the shape of the trajectories and the beating force profile satisfy certain conditions. The aim of this paper is to present a thorough discussion of how synchronization could be achieved for rigid trajectories in a variety of cases.
The rest of the paper is organized as follows. In Section II, we introduce the model and derive the coupled-oscillator equation. In Section III, generic conditions for synchronization are derived by linear stability analysis, and then applied to some specific trajectories and force profiles. In Section IV, we discuss flow properties, especially the net flow and energy dissipation rate in the synchronized state. In Section V, we describe the nonlinear time-evolution equation for the phase difference by an effective potential, which is then used to determine the stable and metastable stationary states for various trajectories and force profiles. In Section VI, the effect of flexibility is taken into account in a model of beads driven by moving harmonic traps. Finally in Section VII, before conclusion, we discuss the implications of our results to biological systems and their direct verification by optically driven colloids.
Model
=====
Dynamical Equations
-------------------
We consider a pair of rotors (indexed by $i=1,2$) and assume that each is a spherical bead of radius $a$ that follows a fixed periodic trajectory ${\bf r}_i = {\bf r}_i(\phi_i)$, where $\phi_i = \phi_i(t)$ is the phase variable with the period $2\pi$ \[see Fig. \[fig:twobeads\]\]. The bead is driven by an active force $F_i = F_i(\phi_i)$ that is tangential to the orbit and is an arbitrary function of the phase. We assume that the two rotors are situated in parallel to each other, and that the center points of the trajectories are at height $h$ from a flat substrate. The $xy$-plane is taken along the substrate, with the $x$-axis parallel to the line connecting the center points, and the $z$-coordinate is taken vertically to the substrate.
![ A pair of rotors with the trajectory shape specified by ${\bf R}(\phi_i)$ ($i=1,2$). Each bead is driven by the tangential force $F(\phi_i)$. The centers of the trajectories are both on the $x$-axis. \[fig:twobeads\] ](fig1.eps){width="0.79\columnwidth"}
The hydrodynamic drag force acting on the $i$-th bead is written in the form $\bg_i = \bzeta \cdot [\bv(\br_i) - \dot{\br}_i]$. The friction coefficient tensor $\bzeta$ depends on the height $z$ of the bead from the substrate. However, the dependence is $O(a/z)$ and we neglect it by assuming $a \ll z$. Then the friction coefficient tensor is expressed by the friction coefficient $\zeta_0 = 6\pi \eta a$ as $\bzeta = \zeta_0 \bI$. The tangential component of the drag force is balanced by the driving force acting on each rotor, namely, $F_i + {\bf t}_i \cdot {\bf g}_i = 0,$ where ${\bf t}_i$ is the tangential unit vector of the orbit given by ${\bf t}_i = {{\bf r}'_i}/{|{\bf r}'_i|}$ with ${\bf r}'_i = {d{\bf r}_i}/{d\phi_i}$. Substituting the expression for the drag force with $\dot{\bf r}_i = {\bf r}'_i \dot\phi_i$ into the force balance equation, we obtain the phase velocity as $\dot\phi_i
= \omega_i
+ {{\bf t}_i \cdot {\bf v}({\bf r}_i)}/{|{\bf r}'_i|},$ where $\omega_i(\phi_i)
= {F_i(\phi_i)}/{\zeta_0 |{\bf r}_i'|}$ is the intrinsic phase velocity. The reaction force $- {\bf g}_i$ exerted by the bead on the fluid generates the flow field $$\begin{aligned}
{\bf v}({\bf r}) = - \sum_j {\bf G}({\bf r}, {\bf r}_j) \cdot {\bf g}_j
\simeq
\sum_{j} \zeta_0{\bf G}({\bf r}, {\bf r}_j) \cdot {\bf r}'_j \omega_j.
\label{vflow}\end{aligned}$$ Here, ${\bf G}({\bf r}, {\bf r}_j)$ is the Green function of the Stokes equation with the no-slip boundary condition at the substrate (Blake tensor). We will give its expression in the next subsection. On the RHS of Eq. (\[vflow\]), we assumed $|{\bf r} - {\bf r}_j| \gg a$ and retained the leading order term with respect to $\zeta_0{\bf G}({\bf r}, {\bf r}_j) = O(a/|{\bf r}-{\bf r}_j|)$ [@Oseen]. Substituting this into the above expression for the phase velocity, we arrive at the coupled phase oscillator equation $$\begin{aligned}
\dot\phi_i
= \omega_i +
\sum_{j\neq i}
\left(
\frac{{\bf t}_i}{|{\bf r}'_i|}
\cdot \zeta_0{\bf G}_{ij} \cdot |{\bf r}'_j| {\bf t}_j
\right)
\omega_j,
\label{dotphi1}\end{aligned}$$ where ${\bf G}_{ij} = {\bf G}({\bf r}_i, {\bf r}_j)$.
Blake Tensor
------------
The Blake tensor $\bG(\br_1, \br_2)$ is given by [@Blake] G\_(\_1, \_2) &=& G\_\^S(\_1 - \_2) - G\_\^S(\_1 - ) + 2 z\_2\^2 G\_\^D(\_1 - ) - 2 z\_2 G\_\^[SD]{}(\_1 - ), \[Blake\] where $\br_i = (x_i, y_i, z_i)$ $(i=1,2)$, and $\overline{\br_2} = (x_2, y_2, -z_2)$ is the point of reflection with respect to the substrate, and G\_\^S() &=& ( + ),\
\
G\_\^D() &=& (1-2\_[z]{}) ()\
&=& (1-2\_[z]{}) ( - )\
&=& (
[ccc]{} & &\
\
& &\
\
& &
) , \[GD\]\
\
G\_\^[SD]{}() &=& (1-2\_[z]{}) G\^S\_[z]{}()\
&=& (1-2\_[z]{}) ( - )\
&=& (
[ccc]{} & &\
\
& &\
\
- &- &
) , \[GSD\] where $\mu,\nu = x,y,z$ with summation over repeated indices assumed, are the fields of a Stokeslet, source doublet and a Stokeslet doublet, respectively. To $O(z_1^2 z_2, z_1 z_2^2)$, we have (\_1, \_2) &=& (x\_1,y\_1,z\_1, x\_2,y\_2,z\_2)\
&& (
[ccc]{} (x\_1-x\_2)\^2 & (x\_1-x\_2) (y\_1-y\_2) & -(x\_1-x\_2) z\_2\
(x\_1-x\_2) (y\_1-y\_2) & (y\_1-y\_2)\^2 & -(y\_1-y\_2) z\_2\
(x\_1-x\_2) z\_1 & (y\_1-y\_2) z\_1 & 0
), \[Gz3\] where $\br_i^{\perp} = (x_i, y_i, 0)$ is the horizontal component of the position vector. Note that $\bG(\br_1,\br_2) \neq \bG(\br_2,\br_1)$ because of the cubic terms ($G_{3\nu}$, $G_{\mu3}$). It is convenient to decompose the Blake tensor into the symmetric and asymmetric part as \_[12,s]{} &=& 12 = (
[ccc]{} 2 x\_[12]{}\^2 & 2 x\_[12]{} y\_[12]{} & x\_[12]{} z\_[12]{}\
2 x\_[12]{} y\_[12]{} & 2 y\_[12]{}\^2 & y\_[12]{} z\_[12]{}\
x\_[12]{} z\_[12]{} & y\_[12]{} z\_[12]{} & 0
). \[G12s\]\
\_[12,a]{} &=& 12 = (
[ccc]{} 0 & 0 & -x\_[12]{} w\_[12]{}\
0 & 0 & -y\_[12]{} w\_[12]{}\
x\_[12]{} w\_[12]{} & y\_[12]{} w\_[12]{} & 0
), \[G12a\] where $\br_{12} = (x_{12}, y_{12}, z_{12}) = \br_1 - \br_2$ and $w_{12} = z_1 + z_2$.
The friction coefficient tensor $\bzeta$ of the bead at height $z$ from the substrate is given by [@HappelBrennerBook] \_(z) = \_0 up to $O(a/z)$. As we stated before, we assume that the relation $a \ll z$ always holds and neglect the correction terms.
Geometric Factor
----------------
In the following, we will assume that two rotors have the same trajectories shape and the force profiles. We can write each trajectory as ${\bf r}_i(\phi) = {\bf r}_{i0} + {\bf R}(\phi)$, where ${\bf r}_{i0}$ is the position of the center, and ${\bf R}(\phi)$ describes the shape of the trajectory. We also assume that the center positions are lying along the $x$-axis at height $h$ from the substrate, and are separated by distance $d$ ($\gg a$) from each other, and that their coordinates are given by \_[10]{} = (0,0,h), \_[20]{} = (d,0,h). We will also denote the typical size of the trajectory by $b$ and the typical magnitude of the driving force by $F_0$: |()| \~b, F() \~F\_0 Note that ${\bf r}'_i(\phi) = |{\bf R}'(\phi)| {\bf t}(\phi)$ where ${\bf t}(\phi) = {\bf R}'(\phi)/|{\bf R}'(\phi)|$ unit tangential vector of the trajectory. The intrinsic frequency $\omega(\phi)$ is given by the force profile $F(\phi)$ as $$\begin{aligned}
\omega(\phi) = \frac{F(\phi)}{\zeta_0|{\bf R}'(\phi)|}.\end{aligned}$$ It is useful to rewrite Eq.(\[dotphi1\]) \_i &=& (\_i) ( 1 + \_[ji]{} H\_[ij]{}(\_i, \_j) ), \[dotphi2\] where H\_[ij]{}(\_i, \_j) &=& (\_i) \_0(\_i(\_i), \_j(\_j)) (\_j) \[Hdef\] is a dimensionless quantity of $O(a h^2/d^3)$, and is determined solely by the geometric configuration of the trajectories (i.e., shape, orientation, distance between each other, and height from the substrate). Hereafter we will call $H_{ij}$ the geometric factor. Note that the symmetry relation H\_[12]{}(\_1, \_2) = H\_[21]{}(\_2, \_1). \[Hsymmetry1\] holds because $\bG(\br_1, \br_2)$ is identical to the transposed matrix $\bG^{t}(\br_2, \br_1)$.
Linear Stability Analysis
=========================
Generic Conditions for Synchronization
--------------------------------------
Let us now examine the stability of the synchronized state by linearizing the evolution equation of the phase difference $\delta = \phi_1 - \phi_2$, which reads $$\begin{aligned}
\dot{\phi}_1 - \dot{\phi}_2 &=& \omega(\phi_1) - \omega(\phi_2)
+
\left[
\omega(\phi_1)
\frac{F(\phi_2)}{F(\phi_1)}
-
\omega(\phi_2)
\frac{F(\phi_1)}{F(\phi_2)}
\right]
H_{12}(\phi_1, \phi_2).
\label{dotdelta}\end{aligned}$$ Here, we used the relation (\[Hsymmetry1\]). Setting $\phi_1 = \phi(t) + \delta(t)$, $\phi_2 = \phi(t)$ and linearizing Eq. (\[dotdelta\]) with respect to $\delta$, we obtain the linear growth rate $$\begin{aligned}
\frac{\dot\delta}{\delta}
=
\omega'(\phi) +
\left[
\omega'(\phi) - \frac{2 F'(\phi)}{F(\phi)} \omega(\phi)
\right]
H_{12}(\phi, \phi).
\label{growthrate}\end{aligned}$$ Integrating (\[growthrate\]) over the period $T = \int_0^{2\pi} d\phi/\dot\phi$ in the limit $\delta \to 0$, we obtain the cycle-averaged growth rate as $$\begin{aligned}
\Gamma
&=& \frac{1}{T} \int_0^{2\pi} d\phi \,
\frac
{\omega'(\phi) [1+ H_{12}(\phi,\phi)] -2 [\ln F(\phi)]'\omega(\phi) H_{12}(\phi,\phi)}
{ \omega(\phi) [1+ H_{12}(\phi,\phi)]}
\nonumber\\
&\simeq& -\frac{2}{T_0}
\int_0^{2\pi} d\phi \, [\ln F(\phi)]' H_{12}(\phi, \phi),
\label{stabcond}\end{aligned}$$ where the approximation is taken to the lowest order in the coupling $H_{12}$, and $T_0$ is the intrinsic period defined by T\_0 = \_0\^[2]{} .
The synchronized state is stable when $\Gamma<0$. Equation (\[stabcond\]) shows that a necessary condition for synchronization is that both the force profile $F(\phi)$ and the geometric factor $H_{12}(\phi,\phi)$ are not constant. However, the latter is constant only for linear and parallel trajectories, as we shall see below. For other trajectory shapes, the necessary condition for synchronization is the non-constantness of the driving force. Equation (\[stabcond\]) guarantees that, if a force profile $F(\phi)$ makes $\Gamma$ positive for a specific trajectory, then the force profile proportional to $1/F(\phi)$ makes $\Gamma$ negative for the same trajectory. In this sense, we can say that roughly half of the possible force profiles in the functional space are capable of inducing in-phase synchronization. We can also state that, for any given trajectory shape $\bR(\phi)$ except for the linear one [@note_linear_trajectory], there exists a force profile $F(\phi)$ that leads to synchronization. For example, the force profile F() = F\_0 , where $\overline{H_{12}}$ is the period-average of $H_{12}(\phi,\phi)$, makes $\Gamma$ negative-definite and hence stabilizes the synchronized state.
Far-Field Limit
---------------
Let us now consider the far-field limit in which the distance between the rotors is much larger than the typical size of the trajectory. Also, the height from the substrate is assumed to be much larger or smaller than the distance: 1, {
[ll]{} 1\
1
. In these limits, the Blake tensor can be approximated by the sum of isotropic (I) and dyadic (D) parts as $$\begin{aligned}
\zeta_0{\bf G}_{12}
&\simeq&
G_I(d) {\bf I} + G_D(d) \be_x \be_x
\label{G12FF}\end{aligned}$$ where we have used $\br_1 - \br_2 = -d \be_x$, and the dimensionless factors $G_I(d)$ and $G_D(d)$ are given by G\_I(d) = G\_D(d) = \[Gbulk\] in the bulk geometry ($h/d \gg 1$) and G\_I(d) = 0, G\_D(d) = \[Gsub\] in the near-substrate geometry ($h/d \ll 1$). With this approximation, the geometric factor (\[Hdef\]) reads $$\begin{aligned}
H_{12}(\phi_1,\phi_2) &=&
G_I(d) +
G_D(d) \, t_x(\phi_1) t_x(\phi_2).
\label{H0th}\end{aligned}$$ Note that the first term is a constant and drops off from the integral (\[stabcond\]). Therefore, only the non-diagonal part of the hydrodynamic interaction controls the stability of the synchronized state in the far-field limit. Let us examine some specific trajectory/force profiles in this limit.
### Circular Trajectories
![ Examples of the force profiles that act to synchronize two beads on circular trajectories aligned on the $x$-axis. (a) $F(\phi)= F_0 [1 - \frac12 \sin (2\phi)]$. (b) $F(\phi)= F_0 \left[1 + \frac12
\sin \left(\phi + \frac{\pi}{4}\right)\right]$. \[fig:circles\] ](fig2.eps){width="0.74\columnwidth"}
As the first example, let us consider the circular trajectory \[see Fig. \[fig:twobeads\](b)\] $$\begin{aligned}
{\bf R}(\phi) = b (\cos\phi, \sin\phi, 0).
\label{circle}\end{aligned}$$ For this trajectory, we have $|{\bf R}'(\phi)| = b$ and ${\bf t}(\phi) = (-\sin\phi, \cos\phi, 0)$, which gives H\_[12]{}(,) = G\_D \^2= - G\_D (2) + [const]{}. and = \_0\^[2]{} d \[F()\]’ (2). Note that the factor $\cos 2\phi$ represents the second-rank tensorial nature of the hydrodynamic kernel. The synchronized state is linearly stable if and only if the Fourier representation of $\ln F (\phi)$ contains a negative coefficient for $\sin 2\phi$. Let the Fourier representation of the force profile be F() = F\_0 \[Ffourier\] where we assume $0 < A_n < 1$ to avoid singularity of $\ln F$. Up to $O(A_n^2)$, we obtain the growth rate as &=& . \[Gammacircle\] The only harmonic mode that contributes to $\Gamma$ at $O(A_n)$ is $n=2$, for which $\Gamma$ is most negative at $\delta_2 = \pi$. In this sense, the force profile that is most efficient in inducing synchronization is given by F()= F\_0 \[1 - A\_2 (2)\], 0<A\_2 < 1. \[Fsin2phi\] Another harmonic mode that contributes to $\Gamma$ by itself is $n=1$, for which $\Gamma$ is most negative at $\delta_1 = {\pi}/{4}, {5\pi}/{4}$. Thus we obtain the candidate force profile F()= F\_0 , -1 < A\_1 < 1. \[Fsinphi\] These two force profiles are illustrated in Fig. \[fig:circles\]. Note that higher harmonic modes ($n\ge 3$) can stabilize synchronization only when they are mixed with the other modes.
Next, we consider rotated circular trajectories. The trajectory (\[circle\]) rotated around the $y$- axis by angle $\alpha$, $
\bR(\phi) = b(\cos\alpha \cos \phi, \sin\phi, \sin\alpha \cos\phi)
$, gives the additional factor $\cos^2 \alpha$ to the linear growth rate via Eq.(\[H0th\]). Note that if the trajectory planes are perpendicular to the $x$-axis ($\alpha=\pi/2$), we have $\Gamma=0$ and the synchronized state is only marginally stable. On the other hand, rotation around the $x$-axis does not change the linear growth rate, because the hydrodynamic kernel (\[G12FF\]) is invariant for the rotation. However, near-field corrections will introduce an important dependence, as we shall see in the next section. We do not consider rotation around the $z$-axis, which is equivalent to shift of the phase by a constant.
### Linear Trajectories
The linear trajectory () = R() [**e**]{}\_x, \[linear\] gives ${\bf t}(\phi) = {\rm sgn} [R'(\phi)] {\bf e}_x$, which makes the geometric factor (\[H0th\]) constant. Thus, at the level of linear stability analysis, the synchronized state is neither stabilized nor destabilized for any force profile. However, nonlinear stability analysis shows that the stability is weakly affected by force modulation, as we shall see in the next section.
### Elliptic Trajectories
For the elliptic trajectory () = (b\_x , b\_y ), b\_x,b\_y>0, \[ellipse\] the $x$-component of the tangential vector $t_x(\phi) = b_x\cos\phi/\sqrt{b_x^2 \cos^2 \phi
+ b_y^2 \sin^2\phi}$ contains all the harmonic modes with odd $n$ if $b_x \neq b_y$. As a result, force modulations containing any harmonic mode with even $n$ can induce synchronization at $O(A_n)$, if the Fourier coefficients are suitably chosen. For example, when $b_x > b_y$, the force profile F() = F\_0 , 0 < A\_4 < 1 \[Fsin4phi\] gives the negative growth rate up to $O(A_4)$, = - (1+)(2+), = . Force modulations with odd harmonics can also induce synchronization, because they give rise to even harmonic modes in $\ln F(\phi)$, but only at $O(A_n^2)$.
Flow Rate and Energy Dissipation
================================
Now let us see how synchronization affects flow properties in the substrated geometry. We define the volume flow rate $Q$ as the flux through a half-plane in the “down stream” ($x\to-\infty$): Q(t) &=& - \_[x-]{} \_[-]{}\^ dy \_0\^dz v\_x(, t)\
&=& \_[x-]{} \_[-]{}\^ dy \_0\^dz \_i G\_[x]{}(, \_i) \_i. \[flowrate1\] In the second line we used the expression for the flow field (\[vflow\]). Note that, due to volume conservation, the integral in (\[flowrate1\]) does not depend on the $x$-position of the half-plane. However, it is easier to calculate it in the limit $x\to-\infty$, where we can use the $O(r^{-2})$ approximation for the Blake tensor, G\_(, \_i) = (
[ccc]{} x\^2 &xy &xz\
xy &y\^2 &yz\
0 &0 &0
), which gives Q(t) &=& \_i \[h + R\_z(\_i)\] g\_[ix]{}. \[flowrate2\] Because we are interested in the change in the flow rate due to hydrodynamic interaction between the rotors, we retain the first order term with respect to $\bG$ in calculating the drag force, which reads, \_i &=& \_0 \[(\_i) - \_i\]\
&=& \_0 . \[gexplicit\] In the in-phase synchronized state $\phi_i = \phi_j = \phi$, the two rotors have the same period $T$, and the cycle-averaged flow rate is calculated using (\[flowrate2\]) and (\[gexplicit\]) as &=& \_0\^T dt Q(t)\
&=& - \_i \_0\^[2]{} d. \[meanflowrate1\] We shall use the far-field approximation (\[G12FF\]) and (\[Gsub\]), to obtain = \_0\^[2]{} d R\_z() R\_[x]{}’(). \[meanflowrate2\] Note that the flow rate is zero for planar geometry ($R_z(\phi)=0$). The hydrodynamic interaction modifies the flow rate not only through the prefactor $1 - G_D$ but also through the period $T$, which is given by T = \_0\^[2]{} \_0\^[2]{} . \[period\]
We can also calculate the power needed to drive the beads, which is given by P(t) = \_i \_i (-\_i). \[power\] Its cycle-average in the synchronized state is calculated to the first order of $\bG$ as &=& \_i \_0\^[2]{} d\
&=& \_0\^[2]{} d () { |’()|\^2 - G\_D R\_x’()\^2 }. \[meanpower2\]
For example, let us compute the flow rate and power for the vertical circular trajectory () = b(, 0, ). For this trajectory, the integrals in (\[meanflowrate2\]) and (\[meanpower2\]) give $\pi b^2$ and $2\pi F_0 b/\zeta_0$, respectively, where $F_0$ is the cycle-average of the driving force. It yields the mean flow rate = \[meanflowrate3\] and the mean power = . \[meanpower3\] The period $T$ depends on the force profile, and is given by T &=& (\_0 - \_1 G\_D),\
\_0 &=& \_0\^[2]{} d ,\
\_1 &=& \_0\^[2]{} d . \[tau\] The dimensionless coefficients $\tau_0$ and $\tau_1$ are positive for any force profile (with $F(\phi)>0$), which means that the period decreases by the hydrodynamic interaction. Furthermore, we have $\tau_1/\tau_0 < 1$ for any force profile, which means that the mean flow rate also decreases by the hydrodynamic interaction.
![ The coefficients $\tau_0$ and $\tau_1$ in Eq.(\[tau\]) for the force profiles (\[Fsin2phi\]) and (\[Fsinphi\]), as functions of $A_1$ and $A_2$ (resp.). For the two force profiles, the curves are identical and $\tau_1/\tau_0$ is equal to $1/2$. \[fig:tau\] ](fig3.eps){width="0.49\columnwidth"}
In Fig. \[fig:tau\], we plot $\tau_0$ and $\tau_1$ for the force profiles (\[Fsin2phi\]) and (\[Fsinphi\]) as functions of the amplitude $A_2$ and $A_1$, respectively. The dependencies on $A_1$ and $A_2$ turn out to be identical for each of $\tau_0$ and $\tau_1$, with the ratio $\tau_1/\tau_0$ equal to $0.5$. Both of the coefficients, and hence the period, diverge as $A_{1,2}$ are taken to unity. When $A_{1,2}=1$, there are stall points (where $F(\phi)=0$) on the trajectory, and it takes infinite time for the bead to pass the points.
Nonlinear Analysis
==================
In this section, we analyze the fully nonlinear evolution equation for the phase difference (\[dotdelta\]), which allows us to explore various dynamical states and their stability. By using the full Blake tensor, we will also discuss the near-field effects due to finite size and height of the trajectories.
Effective Potential
-------------------
The difference in the phase velocities (\[dotdelta\]) consists of the intrinsic phase velocities (the first and second terms in the RHS) and the interaction term (the third term in the RHS). In order to focus on the latter, we exploit the gauge invariance of Eq.(\[dotdelta\]), i.e. the invariance under the transformation $\phi \to \Phi(\phi)$ where $\Phi$ is a new phase variable (or a “gauge”) satisfying $\Phi(\phi+2\pi) = \Phi(\phi) + 2\pi$. We choose the specific gauge $\Phi$ that gives a constant intrinsic phase velocity, which we will call the canonical gauge. It satisfies $\dot\Phi = 2\pi/T_0 = \Omega$ in the absence of hydrodynamic interaction, and is obtained from the original gauge $\phi$ via the relation = = . \[phitoPhi\] In the canonical gauge, the intrinsic terms in Eq.(\[dotdelta\]) cancel out, and the phase difference $\Delta = \Phi_1 - \Phi_2$ obeys = \_[12]{}(\_1, \_2), \[dotDelta1\] where the force profile $\tilde{F}(\Phi)$ and the geometric factor $\tilde{H}_{12}(\Phi_1, \Phi_2)$ are related to those in the original gauge via $\tilde{F}(\Phi) = F(\phi)$ and $\tilde{H}_{12}(\Phi_1, \Phi_2) = H_{12}(\phi_1, \phi_2)$. Note also that $\tilde{\bR}(\Phi) = \bR(\phi)$ and $\tilde{\bR}'(\Phi) = \frac{d\phi}{d\Phi} \bR'(\phi)
= \frac{1}{\Omega\zeta_0} F(\phi) \bt(\phi)$. We rewrite (\[dotDelta1\]) in terms of $\Delta$ and the phase sum $\Sigma = \Phi_1 + \Phi_2$, as = \_[12]{} [ (,) ]{} = W(, ). \[Udef\] Note that $\dot{\Delta}/\Omega = O(G_D) \ll 1$, while $\dot{\Sigma}/\Omega$ is an $O(1)$ quantity. Therefore, we can approximate $\Delta$ to be constant over one period where $\Sigma$ increases by $4\pi$. With the approximation, we take the average of (\[Udef\]) over one period $0<t<T$, to obtain = \_0\^[4]{} d W(, ) = (), \[Wdef\] which defines the effective force $\overline{W}(\Delta)$. We introduce the effective potential $V(\Delta)$ by V() = - \_0\^d’ (’) \[Vdef\] with which the dynamics reduces to minimizing the potential: = -. \[dotDelta3\] Thus we have eliminated the fast variable $\Sigma$ and describe the slow dynamics only by $\Delta$. This approximation is shown to be to correct to the lowest order in the interaction [@Kuramoto].
Note that the factor $
{\tilde{F}\left(\frac{\Sigma-\Delta}{2}\right)}/
{\tilde{F}\left(\frac{\Sigma+\Delta}{2}\right)}
-
{\tilde{F}\left(\frac{\Sigma+\Delta}{2}\right)}/
{\tilde{F}\left(\frac{\Sigma-\Delta}{2}\right)}
$ in Eq. (\[Udef\]) is an odd function of $\Delta$ by construction. The other factor $
\tilde{H}_{12}
{
\left(\frac{\Sigma+\Delta}{2},\frac{\Sigma-\Delta}{2}
\right)
}
$ is an even function of $\Delta$ if the following identity holds: \_[12]{}(\_1,\_2) = \_[12]{}(\_2,\_1). \[Hsymmetry2\] This is the case for the far-field limit ($b/d \to 0$), as we see from Eqs.(\[Hdef\]) and (\[G12FF\]). In that case, $W(\Sigma, \Delta)$ and $\overline{W}(\Delta)$ are odd functions of $\Delta$ and hence the effective potential is an even function: $V(\Delta) = V(-\Delta)$.
Far-Field Limit
---------------
First, let us derive the effective potential in the far-field limit ($b/d \to 0$), for trajectories in the bulk ($h/d \to \infty$) and/or near substrate ($h/d \ll 1$).
### Circular Trajectories
First let us consider circular trajectory (\[circle\]) with the force profile (\[Fsin2phi\]). The phase in the canonical gauge is obtained via (\[phitoPhi\]), as () &=& 2, K() = \_0\^. \[Phi\_sin2phi\] Accordingly, the intrinsic phase velocity in the canonical gauge is given by &=& , \_0 = . \[Omega\_sin2phi\] For $A_2\ll1$, we can approximate $\phi(\Phi)$ and $\tilde{F}(\Phi)$ as $\phi = \Phi + \frac{A_2}{2} \cos 2\Phi$ and $\tilde{F}(\Phi) = F(\Phi)$ to $O(A_2)$, which gives V() V\_0() = \_0 G\_D A\_2 (1-). \[Vsin2phi\] For not small values of $A_2$, we compute the integrals in (\[Wdef\]) and (\[Phi\_sin2phi\]) numerically. We plot $V(\Delta)$ in Fig.\[fig:A2dep-0\], which shows that the approximation (\[Vsin2phi\]) is very good. Even for $A_2=0.99$, the deviation $[V(\Delta)-V_0(\Delta)]/V_0(\Delta)$ falls within $11$ % for any value of $\Delta$. Also note that $V(\Delta)$ is an odd function of $A_2$. For $A_2 > 0$, it is minimized at the in-phase synchronized state ($\Delta=0$), while for $A_2 < 0$, it is minimized at the anti-phase synchronized state ($\Delta= \pi$).
![ The effective potential $V(\Delta)$ in the far-field limit $b/d \to 0$ for the circular trajectory (\[circle\]), either in the bulk ($h/d \gg 1$) or near the substrate ($h/d \ll 1$), for (a) the force profile (\[Fsin2phi\]), and (b) the force profile (\[Fsinphi\]). []{data-label="fig:A2dep-0"}](fig4.eps){width="99.00000%"}
For the force profile (\[Fsinphi\]), the effective potential can be calculated in a similar way as above, and is plotted in Fig. \[fig:A2dep-0\]. A perturbative calculation to $O(A_1^2)$ gives V() V\_0() = (1-). \[Vsinphi\] Again, the approximation (\[Vsinphi\]) is good for moderate values of $A_1$. The deviation $[V(\Delta) - V_0(\Delta)]/V_0(\Delta)$ falls within $0.02$ for $A_1=0.5$. Although the deviation increases to $0.57$ at $A_1=0.99$, the shapes of $V(\Delta)$ and $V_0(\Delta)$ are quite similar.
### Linear Trajectories
![ The effective potential $V(\Delta)$ in the far-field limit $b/d \to 0$ for the linear trajectory (\[circle\]) near the substrate ($h/d \ll 1$), for (a) the force profile (\[Fsin2phi\]) and (b) the force profile (\[Fsinphi\]). (c) Non-analytic behavior $V'(\Delta) = -\overline{W}(\Delta) \propto \Delta^{3/2}$ for the case (a). []{data-label="fig:linear"}](fig5.eps){width="48.00000%"}
Next let us consider the linear trajectory (\[linear\]) with $R(\phi) = b \cos \phi$ using the near-substrate approximation. In Fig. \[fig:linear\], we plot the effective potential for (a) the force profile (\[Fsin2phi\]) and (b) the force profile (\[Fsinphi\]). In both cases the potential is minimized at $\Delta=0$. This result is not expected from the linear stability analysis, which showed that the in-phase synchronized state is only marginally stable for any force profile. It suggests that the potential scales with $V(\Delta) \propto |\Delta|^{\kappa}$ with the exponent $\kappa > 2$ near $\Delta=0$. In Fig.\[fig:linear\](c), we plot $V'(\Delta)=-\overline{W}(\Delta)$ for the case (a). It indicates the non-analytical behavior $\kappa=5/2$, which we will prove in the following paragraph.
The gauge condition (\[phitoPhi\]) gives () &=& 2, K() = \_0\^d’ \[Phiphi\] We see that $\Phi(\pi + \phi) = \pi + \Phi(\phi)$, and especially $\Phi(\pi) = \pi$. For $|\phi| \ll 1$, we have $K(\phi) \approx {\rm sgn(\phi)} \cdot \phi^2/2$ and hence $\Phi(\phi) \approx [\pi/K(2\pi)] \,{\rm sgn(\phi)} \cdot \phi^2$.
We also expand the factor in (\[Udef\]) in powers of $\Delta$ as - = - + O(\^3) \[FoverFtaylor\] while the other factor behaves like a step function: \_[12]{}(,) &=& G\_D [sgn]{} . For small and positive value of $\Delta$, the latter equals $-G_D$ when $2n \pi - \Delta < \Sigma < 2n\pi + \Delta$ ($n$: integer) and equals $G_D$ otherwise. These give the effective force to $O(\Delta^2)$ as () - , where we used $\ln \tilde{F}(\Phi) = \ln F(\phi)
\approx \ln F_0 - 2 A \phi
\approx \ln F_0 - {\rm sgn}(\Phi) \cdot A \sqrt{[K(2\pi)/\pi] |\Phi|}$, which is an approximation for $|\Phi| \ll 1$. Thus we obtained the non-analytic behavior $\overline{W}(\Delta) \propto -{\rm sgn}(\Delta) \cdot |\Delta|^{3/2}$, or $V(\Delta) \propto {\rm sgn}(\Delta) \cdot |\Delta|^{5/2}$.
### Elliptic Trajectories
![ The effective potential $V(\Delta)$ in the far-field limit $b/d \to 0$ for the elliptic trajectory (\[ellipse\]) near the substrate ($h/d \ll 1$), for (a,b) the force profile (\[Fsin2phi\]) with $A_2=0.5$, (c,d) the force profile (\[Fsinphi\]) with $A_1=0.5$, and (e,f) the force profile (\[Fsin4phi\]) with $A_4=0.5$. The long-axis of the ellipse is along the $x$-direction in (a,c,e), and along the $y$-direction in (b,d,f). \[fig:ellipse\] ](fig6.eps){width="99.00000%"}
Next we consider the elliptic trajectory (\[ellipse\]) in the near-substrate approximation. For elliptic trajectories, the tangential vector $\bt(\phi) = \bR'(\phi)/|\bR'(\phi)|$ and hence the geometric factor (\[H0th\]) contain various harmonic modes, which produce richer behaviors than the circular trajectories. In Fig. \[fig:ellipse\], we show the effective potential for (a,b) the force profile (\[Fsin2phi\]) with $A_2=0.5$, (c,d) the force profile (\[Fsinphi\]) with $A_1=0.5$, and (e,f) the force profile (\[Fsin4phi\]) with $A_4=0.5$. In (a,c,e), we show the potential curves for $b_x \ge b_y$, while in (b,d,f), the potential curves for $b_x \le b_y$ are scaled by $(b_x/b_y)^2$. (Note that the potential converges to zero in the limit $b_x/b_y \to 0$.)
In (a), the potential has a single minimum at $\Delta=0$. As $b_x/b_y \to 0$, the scaled potential $V(\Delta)/(b_x/b_y)^2$ converges to a V-shape curve.
In (b), a local minimum at $\Delta=\pm\pi$ appears for $b_x/b_y<1$ in addition to the minimum at $\Delta=0$. For $b_x/b_y \simle 0.6$, the anti-phase synchronized state becomes stable, while the in-phase state becomes metastable. For $\delta_2=3\pi/4$, the sign of the potential has an opposite sign, and we obtain bistable minima at $\Delta=\pm \Delta_0$, with $\Delta_0 \simeq \pi/2$ for $b_x/b_y=0.4$ and $\Delta_0 \to \pi$ as $b_x/b_y$ is increased to unity.
In (c), we have bistable minima at $\Delta \simeq \pm \Delta_0$ with $\pi/2<\Delta_0<\pi$ for $b_x/b_y<1$, and $0<\Delta_0<\pi/2$ for $b_x/b_y>1$. A metastable minimum is located at $\Delta = 0$ and $\Delta=\pm\pi$, for $b_x/b_y<1$ and $b_x/b_y>1$, respectively. For $b_x/b_y=1$, no phase locking occurs because the potential is constant. In (b) and (c) also, the scaled potential $V(\Delta)/(b_x/b_y)^2$ converges to a master curve with sharp peaks and valleys in the limit $b_x/b_y \to 0$ (not shown).
![ (a) The phase function $\Phi(\phi)$ for the elliptic trajectory with $b_y/b_x = 0.5$ and the force profile (\[Fsin4phi\]) with $A_4=0.5$. (b) Temporal oscillation of $\delta$ in the phase-locked state $\Delta=\Delta_0=0.992$. \[fig:delta\] ](fig7.eps){width="99.00000%"}
Now let us consider the meaning of the minimum at non-zero $\Delta$. In the phase-locked state, the phase difference $\Delta$ in the canonical gauge is constant, but it generally means an oscillation of the phase difference $\delta$ in the original gauge, because it is a function of both $\Delta$ and $\Sigma \simeq 2\Omega t$. Let us take for example, the elliptic trajectory with $b_y/b_x = 0.5$ and the force profile (\[Fsin4phi\]) with $A_4=0.5$.
In Fig. \[fig:delta\](a) we show the phase function $\Phi(\phi)$, which has the period $\pi$. The figure also illustrates the relation between $\delta=\phi_1-\phi_2$ and $\Delta=\Phi_1-\Phi_2$. The effective potential, shown in Fig.\[fig:ellipse\](e), has double minima at $\Delta=\pm \Delta_0$ with $\Delta_0 = 0.992$.
Also, it has a metastable minimum at $\Delta=\pi$. In Fig. \[fig:delta\](b), we show $\delta$ in the phase-locked state $\Delta=\Delta_0$ as a function of time (using the relation $\Omega t = \Sigma/2$, where the origin of time is chosen arbitrarily). It oscillates with the period of $\Phi(\phi)$. The amplitude of oscillation is as large as $1.25$, and is larger for stronger modulation of the force profile (that is, for larger amplitude $A_4$). On the other hand, $\delta$ remains constant in the anti-phase synchronized state $\Delta=\pi$, because it matches the period of the phase function.
Near-Field Corrections
----------------------
Next we consider the near-field effects arising from finite size of the trajectory, by using the full Blake tensor $\bG_{12}$ given by (\[Blake\]). The finite size of the trajectory introduces dependences of $\bG_{12}
=\bG_{12}(\br_{10} + \tilde{\bR}(\Phi_1), \br_{20} + \tilde{\bR}(\Phi_2))$ on $\Phi_1$ and $\Phi_2$. In general, $ \bG_{12}$ is not symmetric with respect to the exchange of $\Phi_1$ and $\Phi_2$, which leads to asymmetry of the effective potential $V(\Delta)$. To see this explicitly, we expand the Blake tensor to the first order with respect to \_i = , i=1,2, which is assumed to be small. We also assume that the height of the trajectory is of the same order as its size, and introduce the dimensionless height, = . Then we can use the $O(\hat{h}^3)$ approximation (\[Gz3\]) as a starting point for the expansion. Substituting $x_{12} = d(-1 + \rho_{1x} - \rho_{2x}), \,
y_{12} = d (\rho_{1y} - \rho_{2y}),\,
z_{i} = d (\hat{h} + \rho_{iz}), \,
z_{12} = d(\rho_{1z} - \rho_{2z}),$ and $w_{12} = d(2 \hat{h} + \rho_{1z} + \rho_{2z})$ into (\[G12s\],\[G12a\]) and retaining $O(\brho, \hat{h})$ terms, we have \_0 \_[12,s]{}\^[(2)]{} && C\_D (1+ ) (1+ ) (
[ccc]{} 1+3(\_[1x]{} - \_[2x]{}) & -(\_[1y]{} - \_[2y]{}) & 0\
-(\_[1y]{} - \_[2y]{}) & 0 & 0\
0 & 0 & 0
)\
&& + (1+) (1+) (
[ccc]{} 0 & 0 & -\
0 & 0 & 0\
- & 0 & 0
), \[G12s1\] and \_0 \_[12,a]{} && (1+) (1+) (
[ccc]{} 0 & 0 & -(2 + + )\
0 & 0 & 0\
2 + + & 0 & 0
). \[G12a1\] When we exchange $\phi_1$ and $\phi_2$ or $\brho_1$ and $\brho_2$, the sign of $\bG_{12,s}$ is reversed, while $\bG_{12,a}$ remains unchanged to this order.
In the bulk geometry $h/d \to \infty$, the effects of finite trajectory size can be examined more simply, by expanding the Oseen tensor to the first order in $\brho$. It gives \_0 (\_1, \_2) && C\_[bulk]{} (
[ccc]{} 2(1+ \_[1x]{} - \_[2x]{}) & -(\_[1y]{} - \_[2y]{}) & -(\_[1z]{} - \_[2z]{})\
-(\_[1y]{} - \_[2y]{}) & 1+\_[1x]{} - \_[2x]{} & 0\
-(\_[1z]{} - \_[2z]{}) & 0 & 1+\_[1x]{} - \_[2z]{}
) \[G12b1\] with $C_{bulk} = 3a/4d$. Therefore all the terms change their signs upon exchanging $\phi_1$ and $\phi_2$.
In the following, the effective potential is calculated for the circular trajectory (\[circle\]) rotated around the $x$-axis by the angle $\beta$, or () = b (, , -). The aspect ratio $b/d$ will be fixed to $0.05$ unless otherwise stated. We use both the full Blake tensor (\[Blake\]) and its first-order approximation \[(\[G12s1\]), (\[G12a1\]), and (\[G12b1\])\], and compare with the zeroth-order (far-field) results.
### Force profile (\[Fsin2phi\]) with $A_2=0.5$
The effective potential is plotted in Fig. \[fig:A2proximity\] for trajectories lying (a) in the bulk ($h/d\to \infty$), (b) in a plane horizontal to the substrate ($h/d = 0.1$, $\beta=0$), and (c) vertical to the substrate ($h/d=0.1$, $\beta=\pi/2$). In (a) and (b), the potential is an even function of $\Delta$ and is well approximated by the zeroth order result (\[Vsinphi\]), while in (c), $V(\Delta)$ has a negative average gradient with $V(\pi) < V(-\pi)$.
![ The effective potential $V(\Delta)$ for the circular trajectory (\[circle\]) with $b/d=0.05$ and the force profile (\[Fsin2phi\]) with $A_2=0.5$. The trajectories are either (a) in the bulk ($h/d \to \infty$), (b) horizontal to the substrate ($h/d=0.1$, $\beta=0$), or (c) vertical to the substrate ($h/d=0.1$, $\beta=\pi/2$). Shown are results from the full Blake tensor as well as its zeroth-order and first-order approximations in terms of $b/d$. []{data-label="fig:A2proximity"}](fig8.eps){width="48.00000%"}
### Force profile (\[Fsinphi\]) with $A_1=0.5$
![ The effective potential $V(\Delta)$ for the circular trajectory (\[circle\]) with $b/d=0.05$ and the force profile (\[Fsinphi\]) with $A_1=0.5$. The trajectories are either (a) in the bulk ($h/d \to \infty$), (b) horizontal to the substrate ($h/d=0.1$, $\beta=0$), or (c) vertical to the substrate ($h/d=0.1$, $\beta=\pi/2$). Shown are results from the full Blake tensor as well as its zeroth-order and first-order approximations in terms of $b/d$. []{data-label="fig:A1proximity"}](fig9.eps){width="48.00000%"}
![ The effective potential $V(\Delta)$ for the circular trajectory (\[circle\]) with $b/d=0.05$ and the force profile (\[Fsinphi\]) with $A_1=0.5$. Shown are dependencies on (a) $b/d$ in the bulk geometry ($h/d \to \infty$), (b) $h/d$, and (c) the tilt angle $\beta$ of the trajectories. []{data-label="fig:A1proximity2"}](fig10.eps){width="48.00000%"}
The effective potential is plotted in Fig. \[fig:A1proximity\]. The trajectories are lying either (a) in the bulk ($h/d\to \infty$), (b) in a plane horizontal to the substrate ($h/d = 0.1$, $\beta=0$), or (c) in a plane vertical to the substrate ($h/d = 0.1, \beta=\pi/2$). In (a) and (b), the potential curves have negative average gradient with $V(\pi) < V(-\pi)$, and have local minima at $\Delta=0$. Note that the asymmetry of the potential curve is larger in the bulk case. In (c), the potential has a positive average gradient with $V(\pi) > V(-\pi)$, and has two metastable minima. There is a saddle point at $\Delta=0$. Note that all these features are already seen in the first-order approximation.
In Fig. \[fig:A1proximity2\](a), we plot the potential for different trajectory size in the bulk geometry. The average gradient of the potential is enhanced with $b/d$. In Fig. \[fig:A1proximity2\](b), we plot the potential for the vertical trajectory ($\beta=\pi/2$) and with different height. Note that the average gradient of the potential changes its sign from positive to negative at intermediate height. In Fig. \[fig:A1proximity2\](c), we show the dependence on the tilt angle $\beta$. The average gradient of the potential changes its sign around $\beta = \pi/6$. We thus find that the asymmetry of the potential sensitively depends on the size, the height, and the tilt of the trajectories.
Effect of Flexibility
=====================
Hitherto we have only considered rotors with rigid trajectories, but in a real system the trajectory could be affected by hydrodynamic flow due to flexibility or compliance of the rotor. As an example, let us consider a bead driven by optical tweezers, whose focus moves along a prescribed trajectory. By controlling the distance between the focal point and the bead, one can tune the tangential driving force [@Bruot]. We approximate the potential created by the laser beam by the harmonic potential $U(\bS) = \frac{k}{2} S^2$, where $\bS$ is the displacement of the bead from the focal point. The bead of the $i$-th rotor is thus positioned at $
\br_i(t) = \br_{i0} + \bR(\phi_i(t)) + \bS_i(t),
$ and its velocity is $\dot\br_i = \bR'(\phi_i) \dot\phi_i + \dot{\bS}_i$. The traction force due to the laser beam is balanced with the viscous drag force as k \_i = \_i = \_0 \[(\_i) - \_i\], \[Ftweezer2\] while its tangential component is prescribed as $F(\phi_i) = - k \bS_i \cdot \bt(\phi_i)$, or F\_i = - \_i \_i, \[Ftweezer1\] where abbreviations $F_i = F(\phi_i)$ and $\bt_i = \bt(\phi_i)$ are used as before. In the limit $k \to \infty$, we restore the model of rigid rotors developed in the previous sections. Let us derive the phase evolution equation by expansion in powers of $k^{-1}$.
First, the intrinsic phase velocity $\omega_i = \omega(\phi_i)$ is determined by setting $\bv(\br)=0$. Equation (\[Ftweezer2\]) gives \_i = - (\_i’ \_i + \_i) - (\_i’\_i) - (F\_i \_i)’ \_i \[Ftweezer3\] up to $O(k^{-1})$. Using this in (\[Ftweezer1\]), we obtain the intrinsic frequency to $O(k^{-1})$ as \_i ( 1 + ). The hydrodynamic interaction is incorporated by substituting (\_i) = - \_j \_[ij]{} \_j - \_j \_0 \_[ij]{} (’\_j \_j + \_j) into (\[Ftweezer2\]). For simplicity let us assume the far-field limit, where $\bG_{ij}$ is given by the constant symmetric tensor (\[G12FF\]). After some calculation, we obtain the phase evolution equation in the form \_i &=& \_i ( 1+ \_[ji]{} J\_[ij]{} ) + \_[j i]{} H\_[ij]{} \_j, \[Ftweezer4\] where the function $H_{ij} = H_{ij}(\phi_i,\phi_j)$ now includes an $O(1/k)$ correction as H\_[ij]{} &=& \_i \_0 \_[ij]{} , and $J_{ij} = J_{ij}(\phi_i,\phi_j)$ is defined by J\_[ij]{} &=& \_0 \_[ij]{} F\_j \_j. The stability of the synchronized state is examined by setting $\phi_1 = \phi + \delta$ and $\phi_2 = \phi$ and linearizing the evolution equation of $\delta$ as before. After some straightforward calculation, we obtain the cycle-averaged growth rate of the phase difference $\delta$ as = \_0\^[2]{} d{ -2 ’ H\_[12]{} + ()’ J\_[12]{} + H’\_[12]{} + J’\_[12]{} }, \[Ftweezer6\] where the functions in the integrand are to be evaluated at $\phi_1 = \phi_2 = \phi$ and we define A’ &=& . ( - ) |\_[\_1 = \_2 = ]{} \[Ftweezer5\] for any two-variable function $A(\phi_1,\phi_2)$.
For example, let us consider the circular trajectory $\bR(\phi) = b(\cos\phi, \sin\phi, 0) = b \bn(\phi)$. Using $|\bR'(\phi)| = b$, $\bn'(\phi) = \bt(\phi)$, $\bt'(\phi) = -\bn(\phi)$, and the Fourier representation (\[Ffourier\]) of the force profile, we obtain the growth rate up to $O(A_n)$ as = . \[Gammaflexcircle\] This should be compared to the result (\[Gammacircle\]) for the rigid rotors. We see that the flexibility tends to enhance synchronization due to the last term on the RHS. Note also that the small parameter representing the flexibility is $F_0/kb$. If the displacement from the focal point (which has the typical magnitude $S_0 \sim F_0/k$) is much smaller than the size of the trajectory, which is the case in the optical tweezer experiment [@Bruot], the flexibility has only a weak effect in inducing synchronization. These results qualitatively agree with the findings of the previous study [@NEL08] that assumed constant driving force and radial displacement from a circular trajectory. In the paper, the model parameters are estimated for cilia, which give the dimensionless coupling (that corresponds to our $F_0/kb$) to be on the order of $10^{-2}-10^{-3}$. On the other hand, we can expect $O(1)$ modulation of the driving force from the effective and recovery strokes of cilia. Therefore, we conjecture that the force modulation plays dominant roles in establishing the coordinated ciliary beating. Finally, we mention that our model of flexibility can be also modified for rotors allowing tangential displacement, such as a bead attached to the tip of an elastic rod.
Concluding Remarks
==================
By linear and nonlinear analysis of the coupled oscillator equation, we have fully characterized the dynamical states of a pair of rotors making rigid trajectories. In particular, we obtained the necessary and sufficient conditions for in-phase synchronization, which show that a wide variety of beating patterns induce synchronization for an arbitrary trajectory shape. Even for parallel linear trajectories, which predict only marginal stability in the linear analysis, the effective potential has a global minimum at the in-phase state if we choose a suitable force profile.
The results confirm and strengthen our previous finding [@UG11] that flexibility of the rotors is [*not*]{} a requisite for synchronization, although it has been highlighted in many other studies. In the present paper, we incorporated flexibility into our model and explicitly compared its effect to the effect of force modulation. If the disturbance of the trajectory due to hydrodynamic interaction is small compared to the size of trajectory, the flexibility has only weak effect in establishing synchronization. For cilia, sizable modulation of the driving force is expected from their effective and recovery strokes, and it should play a dominant role in coordinating their beating. Recently, another mechanism for driving synchronization between flagella of a swimming [*Chlamydomonas*]{} has been proposed using a simple three-sphere model [@FJ12; @BG12]. In these studies, the phases of the two flagella are predominantly coupled via translation and rotation of the cell body. This type of coupling originates from the condition that the net force and torque acting on the cell vanish, and is specific to rotors attached to a freely-suspended body. The condition for synchronization with this type of coupling is different from hydrodynamic synchronization (for example even constant forcing profile and circular trajectory could lead to synchronization under certain circumstances). Also, the coupling is weaker than the hydrodynamic one if the cell body is much larger than the distance between neighboring flagella or cilia, which is the case in densely flagellated/ciliated cells such as [*Volvox*]{} and [*Paramecium*]{}.
The effective potential that governs the nonlinear dynamics of the rotors have a number of remarkable features. First, it allows us to locate all the stable and metastable states of the system at a glance. In the far-field limit, the potential is symmetric. For circular trajectories with simple force modulation (consisting of a single harmonic mode), the potential has only one minimum that describes either in-phase or anti-phase synchronization. Bistable and metastable states appear for more complex trajectory shapes such as ellipses. When the system is trapped in an out-of-phase stable/metastable state, the phase difference (in a natural gauge) oscillates as a function of time. We have incorporated near-field corrections due to finite trajectory size, and found that the overall shape of the potential, especially its average gradient, sensitively depends on the size/height/tilt of the trajectory. When the potential has a non-vanishing average gradient, each of its local minimum corresponds to a metastable state. In the presence of strong noise, we may observe phase slippage in a specific direction. We note that experiments on the flagellar beating of a mutant [*Chlamydomonas*]{} have recently shown anti-phase synchronization [@Goldstein2]. It will be interesting to probe the differences between this mutant strain and the wild-type [*Chlamydomonas*]{} in terms of the beating pattern of the flagella, and examine whether the phenomenon can be quantified within the framework of our model.
A more direct experimental test of our findings could be pursued in a simpler system that does not have the complexity of the living organisms, such as optically driven colloids. Optical tweezers with moving focus can drive the colloidal particle on a prescribed trajectory, and by controlling the distance between the focal point and the bead, one can also prescribe the force profile. Experiments are currently underway in the group of Pietro Cicuta at the Cavendish Laboratory along these lines [@Bruot]. Also, optical vortices [@Gahagan96; @Curtis03] have been used to trap colloidal particles on a ring and drive them in one direction. The driving torque could be modulated by tailoring the helical structure of the laser beam to give a prescribed force profile.
In forthcoming papers, we plan to discuss the collective dynamics of arrayed rotors, and in particular the formation of traveling waves. Such a study should become a first step towards understanding the relation between the beating pattern of cilia and the metachronal waves they form. We will also consider a pair of rotors with different intrinsic frequencies, which will induce phase slips similar to those observed in [*Chlamydomonas*]{} [@Goldstein-Science-09; @GP09]. The present paper assumes spherical beads, but the analysis could be extended to non-spherical bodies such as rods or helices, which are closer to the shapes of biological filaments.
NU acknowledges the support by JSPS KAKENHI (Grant Number 23740286) and the JSPS Core-to-Core Program “International research network for non-equilibrium dynamics of soft matter”.
[99]{}
J. Gray, [*Ciliary Movements*]{} (Cambridge University Press, Cambridge, 1928).
D. Bray, [*Cell Movements: From Molecules to Motility*]{} 2nd ed. (Garland, New York, 2001).
S. Nonaka, S. Yoshiba, D. Watanabe, S. Ikeuchi, T. Goto, W. F. Marshall, and H. Hamada, PLoS Biol. [**3**]{}, (8):e268 (2005).
J. R. Blake and M. A. Sleigh, Biol. Rev. [**49**]{}, 85 (1974).
E. W. Knight-Jones, Q. J. Microsc. Sci. [**95**]{}, 503 (1954).
J. R. Blake, J. Fluid Mech [**55**]{} 1 (1972).
S. Gueron, K. Levit-Gurevich, N. Liron and J.J. Blum, Proc. Natl. Acad. Sci. (USA) [**94**]{}, 6001 (1997).
S. Gueron and K. Levit-Gurevich, Proc. Natl. Acad. Sci. (USA) [**96**]{}, 12240 (1999).
R. Golestanian, J. M. Yeomans, and N. Uchida, Soft Matter [**7**]{}, 3074 (2011).
G. I. Taylor, Proc. R. Soc. A [**209**]{}, 447 (1951).
M. J. Kim, J. C. Bird, A. J. Van Parys, K. S. Breuer, and T. R. Powers, Proc. Natl. Acad. Sci. (USA) [**100**]{}, 15481 (2003).
B. Qian, H. Jiang, D. A. Gagnon, K. S. Breuer, and T. R. Powers, Phys. Rev. E [**80**]{}, 061919 (2009).
M. Polin, I. Tuval, K. Drescher, J. P. Gollub, and R. E. Goldstein, Science [**325**]{}, 487 (2009).
R. E. Goldstein, M. Polin, and I. Tuval, Phys. Rev. Lett. [**103**]{}, 168103 (2009).
J. Kotar, M. Leoni, B. Bassetti, M. C. Lagomarsino, and P. Cicuta, Proc. Natl. Acad. Sci. (USA) [**107**]{}, 7669 (2010).
R. Di Leonardo, A. Buzas, L. Kelemen, G. Vizsnyiczai, L. Oroszi, P. Ormos, Phys. Rev. Lett. [**109**]{}, 034104 (2012).
N. Darnton, L. Turner, K. Breuer, and H. C. Berg, Biophys. J. [**86**]{}, 1863 (2004).
M. Vilfan, A. Potocvnik, B. Kavcic, N. Osterman, I. Poberaj, A. Vilfan, and D. Babic, Proc. Natl. Acad. Sci. (USA) [**107**]{}, 1844 (2010).
A. R. Shields, B. L. Fiser, B. A. Evans, M. R. Falvo, S. Washburn, and R. Superfine, Proc. Natl. Acad. Sci. (USA) [**107**]{}, 15670 (2010).
N. Coq, A. Bricard, F.-D. Delapierre, L. Malaquin, O. du Roure, M. Fermigier and D. Bartolo, Phys. Rev. Lett. [**107**]{}, 014501 (2011)
M. Cosentino Lagomarsino, B. Bassetti and P. Jona, Eur. Phys. J. B [**26**]{}, 81 (2002).
M. Cosentino Lagomarsino, P. Jona and B. Bassetti, Phys. Rev. E [**68**]{}, 021908 (2003).
M. Kim and T.R. Powers, Phys. Rev. E [**69**]{}, 061910 (2004).
M. Reichert and H. Stark, Eur. Phys. J. E [**17**]{}, 493 (2005).
Y. W. Kim and R. R. Netz, Phys. Rev. Lett. [**96**]{}, 158101 (2006).
A. Vilfan and F. Jülicher, Phys. Rev. Lett. [**96**]{}, 058102 (2006).
A. Ryskin and P. Lenz, Phys. Biol. [**3**]{}, 285 (2006).
B. Guirao and J.-F. Joanny, Biophys. J. [**92**]{}, 1900 (2007).
T. Niedermayer, B. Eckhardt, and P. Lenz, Chaos [**18**]{}, 037128 (2008).
G. J. Elfring and E. Lauga, Phys. Rev. Lett. [**103**]{}, 088101 (2009).
N. Uchida and R. Golestanian, Phys. Rev. Lett. [**104**]{}, 178103 (2010).
N. Uchida and R. Golestanian, Europhys. Lett. [**89**]{}, 50011 (2010).
N. Uchida and R. Golestanian, Phys. Rev. Lett. [**106**]{}, 058104 (2011).
N. Osterman and A. Vilfan, Proc. Natl. Acad. Sci. (USA) [**108**]{}, 15727 (2011).
C. W. Oseen, [*Neuere Methoden und Ergebnisse in der Hydrodynamik*]{} (Akademishe Verlagsgesellschaft, Leipzig, 1927).
J. R. Blake, Proc. Camb. Phil. Soc. [**70**]{}, 303 (1971).
J. Happel, H. Brenner, [*Low Reynolds Number Hydrodynamics*]{} (M. Nijhoff, The Hague, 1983).
By considering nonlinear effect, we show in Section IV that linear trajectories are also capable of synchronization.
Y. Kuramoto, [*Chemical Oscillations, Waves, and Turbulence*]{} (Springer, New York, 1984).
N. Bruot, private communication.
R. E. Goldstein, private communication.
B. M. Friedrich and F. Julicher, Phys. Rev. Lett. [**109**]{}, 138102 (2012).
R. R. Bennett and R. Golestanian, arXiv:1211.3272.
K. T. Gahagan and G. A. Swartzlander, Opt. Lett. [**21**]{}, 827 (1996).
J. E. Curtis and D. G. Grier, Phys. Rev. Lett. [**90**]{}, 133901 (2003).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We investigate the $L^p$ boundedness of the lacunary maximal function $A_rf$ associated to the spherical means on the Heisenberg group. By suitable adaptation of an approach of M. Lacey in the Euclidean case, we obtain sparse bounds for these maximal functions, which lead to new weighted estimates. In order to prove the result, several properties of the spherical means have to be accomplished, namely, the $L^p$ improving property of the operator $A_rf$ and a continuity property of the difference $A_rf-\tau_y A_rf$, where $\tau_yf(x)=f(xy^{-1})$ is the right translation operator.'
address:
- 'Stat-Math Unit, Indian Statistical Institute, Kolkata, India.'
- |
Department of Mathematics\
Indian Institute of Science\
560 012 Bangalore, India
- |
BCAM - Basque Center for Applied Mathematics\
48009 Bilbao, Spain and Ikerbasque, Basque Foundation for Science, 48011 Bilbao, Spain
- |
Department of Mathematics\
Indian Institute of Science\
560 012 Bangalore, India\
and BCAM - Basque Center for Applied Mathematics\
48009 Bilbao, Spain
author:
- Sayan Bagchi
- Sourav Hait
- Luz Roncal
- Sundaram Thangavelu
title: |
On the maximal function associated to the\
lacunary spherical means on the Heisenberg group
---
Introduction and main results
=============================
A celebrated theorem of Stein [@Stein] proved in 1976 says that the spherical maximal function $ M $ defined by $$Mf(x) = \sup_{r>0} |f\ast \sigma_r(x)| = \sup_{r>0}\Big| \int_{|y|=r} f(x-y) d\sigma_r(y)\Big|$$ is bounded on $ L^p({\mathbb R}^n)$, $n \ge 2$, if and only if $ p > n/(n-1).$ Here $ \sigma_r $ stands for the normalised surface measure on the sphere $ S_r = \{ x\in {\mathbb R}^n: |x|=r\} $ in $ {\mathbb R}^n.$ The case $ n =2 $ was proved later by Bourgain [@Bourgain]. As opposed to this, in 1979, C. P. Calderón [@C] proved that the lacunary maximal function $$M_{\text{lac}}f(x) = \sup_{ j \in {\mathbb Z}} \Big|\int_{|y|=2^j} f(x-y) d\sigma_{2^j}(y)\Big|$$ is bounded on $ L^p({\mathbb R}^n) $ for all $ 1 < p < \infty $ for any $ n \ge 1.$ In a recent article, Lacey [@Lacey] has revisited the spherical maximal function. Using a new approach he has managed to prove certain sparse bounds for these maximal functions which has led him to obtain new weighted norm inequalities. Our main goal in this paper is to adapt the method of Lacey to prove an analogue of Calderón’s theorem in the context of certain spherical means on the Heisenberg group, and deduce weighted inequalities as immediate consequences.
Let ${\mathbb H}^n={{\mathbb C}}^n\times {\mathbb R}$ be the $(2n+1)$-dimensional Heisenberg group with the group law $$(z,t)(w,s)=\Big(z+w,t+s+\frac12{\operatorname{Im}}z\cdot \overline{w}\Big).$$ Given a function $f$ on ${\mathbb H}^n$, consider the spherical means $$\label{eq:defin}
A_rf(z,t):=f\ast \mu_r(z,t)=\int_{|w|=r}f\Big(z-w,t-\frac12{\operatorname{Im}}z\cdot \overline{w}\Big)\,d\mu_r(w)$$ where $\mu_r$ is the normalised surface measure on the sphere $S_r=\{(z,0):|z|=r\}$ in ${\mathbb H}^n$. The maximal function associated to these spherical means was first studied by Nevo and Thangavelu in [@NeT]. Later, improving the results in [@NeT], Narayanan and Thangavelu [@NaT] and Müller and Seeger [@MS], independently, proved the following sharp maximal theorem: the full maximal function $ Mf(z,t) = \sup_{r>0} |A_rf(z,t)| $ is bounded on $ L^p({\mathbb H}^n), n \geq 2 $ if and only if $ p > (2n)/(2n-1).$ In this work we consider the lacunary maximal function $$M_{\operatorname{lac}}f(z,t)=\sup_{j\in {\mathbb Z}}|A_{\delta^j}f(z,t)|, \quad \delta>0,$$ associated to the spherical means and prove the following result.
\[thm:spherical\] Assume that $ n \geq 2.$ Then for any $ 0 < \delta <\frac{1}{96}$ the associated lacunary maximal funcion $ M_{\operatorname{lac}}$ is bounded on $ L^p({\mathbb H}^n) $ for any $ 1 < p < \infty.$
Actually we can take any $\delta$ in Theorem \[thm:spherical\]. For example, we can take $\delta=2$. In our result we are taking $\delta< \frac{1}{96}$ not because the proof requires the restriction, but because we want to keep the proof simple, see more explanation after the statement of Lemma \[lem:Hn\].
We remark that another kind of spherical maximal function on the Heisenberg group has been considered by Cowling. In [@Cowling] he has studied the maximal function associated to the spherical means taken over genuine Heisenberg spheres, i.e., averages taken over spheres defined in terms of a homogeneous norm on $ {\mathbb H}^n.$ It would be interesting to see if lacunary maximal functions asociated these spherical means also have better mapping properties. We remark in passing that the spherical means studied in [@NeT; @NaT; @MS] (and hence in this paper) are more singular than the one studied in [@Cowling] as these means are supported on codimension two submanifolds.
Theorem \[thm:spherical\], as well as certain weighted versions, are easy consequences of the sparse bound in Theorem \[thm:sparse\], which is the main result of this paper. Before stating the result let us set up the notation. As in the case of $ {\mathbb R}^n $, there is a notion of dyadic grids on $ {\mathbb H}^n $, the members of which are called (dyadic) cubes. A collection of cubes $\mathcal{S}$ in ${\mathbb H}^n $ is said to be $ \eta$-sparse if there are sets $\{E_S \subset S:S\in \mathcal{S}\}$ which are pairwise disjoint and satisfy $|E_S|>\eta|S|$ for all $S\in \mathcal{S}$. For any cube $Q$ and $1<p<\infty$, we define $$\langle f\rangle_{Q,p}:=\bigg(\frac{1}{|Q|}\int_Q|f(x)|^pdx\bigg)^{1/p}, \qquad \langle f\rangle_{Q}:=\frac{1}{|Q|}\int_Q|f(x)|dx.$$ In the above $ x =(z,t) \in {\mathbb H}^n $ and $ dx = dz dt $ is the Lebesgue measure on $ {{\mathbb C}}^n \times {\mathbb R}$ which incidentally is the Haar measure on the Heisenberg group. Following Lacey [@Lacey], by the term $(p,q)$-sparse form we mean the following: $$\Lambda_{\mathcal{S},p,q}(f,g)=\sum_{S\in\mathcal{S}}|S|\langle f\rangle_{S,p}\langle g\rangle_{S,q}.$$
\[thm:sparse\] Assume $ n \geq 2$ and fix $ 0 < \delta < \frac{1}{96}.$ Let $ 1 < p, q < \infty $ be such that $ (\frac{1}{p},\frac{1}{q}) $ belongs to the interior of the triangle joining the points $ (0,1), (1,0) $ and $ (\frac{n}{n+1},\frac{n}{n+1}).$ Then for any pair of compactly supported bounded functions $ (f,g) $ there exists a $ (p,q)$-sparse form such that $ \langle M_{\operatorname{lac}}f, g\rangle \leq C \Lambda_{\mathcal{S},p,q}(f,g).$
In proving the corresponding result for the spherical means on $ {\mathbb R}^n $, Lacey [@Lacey] has made use of two important properties of the spherical means, namely, the $ L^p $ improving property of the operator $ S_r f = f \ast \sigma_r $ for a fixed $ r $, and the continuity property of the difference $ S_r f- \tau_y S_r f $ where $ \tau_y f(x) = f(x-y) $ is the translation operator.
A remark is in order. In order to keep the shape of our main results analogous to the ones related to the lacunary maximal function in ${\mathbb R}^n$, we decided to restrict the range of $(p,q)$ to the same regions as in the Euclidean case. Nevertheless, enhanced results for the lacunary maximal function in ${\mathbb H}^n$ are obtained (although we cannot say anything about the sharpness of such results) and we will also state them throughout the paper, see Subsections \[rem:sharp\], \[sub:sharpenc\], \[sub:sharpens\] and \[sub:sharpenw\]. In particular, a sharpened version of Theorem \[thm:sparse\] is given in Theorem \[thm:mainSH\].
In the next section we establish $ L^p-L^q $ estimates for our spherical means $ A_r f $ on the Heisenberg group. In Section \[sec:continuity\] we prove the continuity property of $ A_r f- A_r\tau_y f$, where now $ \tau_yf(x) =f(xy^{-1}) $ is the right translation operator. In Section \[sec:sparse\] we establish the sparse bound and finally in the last section we deduce weighted boundedness properties of the lacunary maximal function.
The $L^p$ improving property of the spherical mean value operator {#sec:Lp}
=================================================================
The observation that the spherical mean value operator $ S_r f := f\ast \sigma_r $ on $ {\mathbb R}^n $ is a Fourier multiplier plays an important role in every work dealing with the spherical maximal function. In fact, we know that $$\label{eq:ArEuc}
f \ast \sigma_r(x) = (2\pi)^{-n/2} \int_{{\mathbb R}^n} e^{i x \cdot \xi} \widehat{f}(\xi) \frac{ J_{n/2-1}(r|\xi|)} {(r|\xi|)^{n/2-1}} d\xi$$ where $ J_{n/2-1} $ is the Bessel function of order $ n/2-1.$ As Bessel functions $ J_\alpha $ are defined even for complex values of $ \alpha $ the above allows one to embed $ S_r f$ in an analytic family of operators and Stein’s analytic interpolation theorem comes in handy in studying the spherical maximal function. The same technique was employed by Strichartz [@S] who studied the $ L^p $ improving properties of $ S_r .$ For the spherical means on the Heisenberg group we do have such a representation if we replace the Euclidean Fourier transform by the group Fourier transform on $ {\mathbb H}^n.$
For the group $ {\mathbb H}^n $ we have a family of irreducible unitary representations $ \pi_\lambda $ indexed by non-zero reals $ \lambda $ and realised on $ L^2({\mathbb R}^n)$. The action of $\pi_{\lambda}(z,t)$ on $L^2({\mathbb R}^n)$ is explicitly given by $$\label{eq:pilambda}
\pi_{\lambda}(z,t)\varphi(\xi)=e^{i\lambda t}e^{i\lambda(x\cdot \xi+\frac12 x\cdot y)}\varphi(\xi+y)$$ where $\varphi\in L^2({\mathbb R}^n)$ and $z=x+iy$. By the theorem of Stone and von Neumann, which classifies all the irreducible unitary representations of $ {\mathbb H}^n $, combined with the fact that the Plancherel measure for $ {\mathbb H}^n $ is supported only on the infinite dimensional representations, it is enough to consider the following operator valued function known as the group Fourier transform of a given function $ f $ on $ {\mathbb H}^n$: $$\widehat{f}(\lambda) = \int_{{\mathbb H}^n} f(z,t) \pi_\lambda(z,t) \,dz \,dt.$$ The above is well defined, e.g., when $ f \in L^1({\mathbb H}^n) $ and for each $ \lambda \neq 0,$ $ \widehat{f}(\lambda) $ is a bounded linear operator on $ L^2({\mathbb R}^n).$ Observe that the above definition makes sense even if we replace $ f $ by a finite Borel measure $ \mu.$ In particular, $ \widehat{\mu}_r(\lambda) $ are well defined bounded operators on $ L^2({\mathbb R}^n)$ which can be described explicitly. Combined with the fact that $ \widehat{ f\ast g}(\lambda) = \widehat{f}(\lambda) \widehat{g}(\lambda) $ we obtain $ \widehat{A_rf}(\lambda) = \widehat{f}(\lambda)\widehat{\mu}_r(\lambda).$
The operators $ \widehat{\mu}_r(\lambda) $ turn out to be functions of the Hermite operator $H(\lambda) = -\Delta+\lambda^2 |x|^2.$ Indeed, if the spectral decomposition of $ H(\lambda) $ is written as $$H(\lambda) = \sum_{k=0}^\infty (2k+n)|\lambda| P_k(\lambda)$$ where $ P_k(\lambda) $ are the Hermite projection operators, then (see [@TRMI Proposition 4.1]) $$\widehat{\mu}_r(\lambda) = \sum_{k=0}^\infty \psi_k^{n-1}(\sqrt{|\lambda|}r) P_k(\lambda),$$ where for any $ \delta > -1 $ the normalised Laguerre functions are defined by $$\label{eq:LagF}
\psi_k^{\delta}(r)=\frac{\Gamma(k+1)\Gamma(\delta+1)}{\Gamma(k+\delta+1)}L_k^{\delta}\Big(\frac12r^2\Big)e^{-\frac14r^2}.$$ In the above definition $ L_k^\delta(r) $ stands for the Laguerre polynomials of type $ \delta.$ Thus we have the relation $$\label{ArHe}
\widehat{A_rf}(\lambda) = \widehat{f}(\lambda) \sum_{k=0}^\infty \psi_k^{n-1}(\sqrt{|\lambda|}r) P_k(\lambda),$$ which is the analogue of in our situation. Thus, as in the Euclidean case, the spherical mean value operators $ A_r $ are (right) Fourier multipliers on the Heisenberg group. We now proceed to rewrite in terms of Laguerre expansions, which is more suitable for defining an analytic family of operators containing the spherical means.
The irreducible unitary representations $ \pi_\lambda $ admit the factorisation $ \pi_\lambda(z,t) = e^{i\lambda t} \pi_\lambda(z,0) $ and hence we can write the Fourier transform as $$\widehat{f}(\lambda) = \int_{{{\mathbb C}}^n} f^\lambda(z) \pi_\lambda(z,0)\, dz,$$ where for a function $f$ on ${\mathbb H}^n$, $f^{\lambda}(z)$ stands for the partial inverse Fourier transform $$f^{\lambda}(z)=\int_{-\infty}^{\infty}e^{i\lambda t}f(z,t)\,dt.$$ We now make use of the special Hermite expansion of the function $ f^\lambda $, which can be put in a compact form as follows. Let $ \varphi_k^{\lambda}(z)=L_k^{n-1}\big(\frac12|\lambda||z|^2\big)e^{-\frac14|\lambda||z|^2}$ stand for the Laguerre functions of type $ (n-1) $ on $ {{\mathbb C}}^n.$ The $\lambda$-twisted convolution $f^{\lambda}\ast_{\lambda}\varphi_k^{\lambda}(z)$ is then defined by $$f^{\lambda}\ast_{\lambda}\varphi_k^{\lambda}(z)=\int_{{{\mathbb C}}^n}f^{\lambda}(z-w)\varphi_k^{\lambda}(w)e^{i\frac{\lambda}{2}{\operatorname{Im}}z\cdot \overline{w}}\,dw.$$ It is well known that one has the expansion (see [@STH Chapter 3, proof of Theorem 3.5.6]) $$f^\lambda(z) = (2\pi)^{-n} |\lambda|^n \sum_{k=0}^\infty f^{\lambda}\ast_{\lambda}\varphi_k^{\lambda}(z),$$ which leads to the formula (see [@STH Theorem 2.1.1]) $$f(z,t) = (2\pi)^{-n-1} \int_{-\infty}^\infty e^{-i\lambda t} \Big( \sum_{k=0}^\infty f^{\lambda}\ast_{\lambda}\varphi_k^{\lambda}(z)\Big) |\lambda|^n d\lambda$$ Applying this to $f\ast \mu_r$ we have the formula $$f\ast \mu_r(z,t)=\frac{1}{2\pi}\int_{-\infty}^{\infty}e^{-i\lambda t}f^{\lambda}\ast_{\lambda}\mu_r(z) \,d\lambda$$ where we used the fact that $(f\ast \mu_r)^{\lambda}(z)=f^{\lambda}\ast_{\lambda}\mu_r(z).$ It can be shown that [@TRMI Theorem 4.1], [@NeT Proof of Proposition 6.1], $$f^{\lambda}\ast_{\lambda}\mu_r(z)= (2\pi)^{-n} |\lambda|^n \sum_{k=0}^{\infty}\frac{k!(n-1)!}{(k+n-1)!}\varphi_k^{\lambda}(r)f^{\lambda}\ast_{\lambda}\varphi_k^{\lambda}(z),$$ leading to the expansion (see [@NeT; @NaT]) $$\label{eq:expression}
A_rf(z,t)=(2\pi)^{-n-1}\int_{-\infty}^{\infty}e^{-i\lambda t}\Big(\sum_{k=0}^{\infty}\psi_k^{n-1}(\sqrt{|\lambda|}r)f^{\lambda}\ast_{\lambda}\varphi_k^{\lambda}(z)\Big)|\lambda|^n\,d\lambda.$$ By replacing $ \psi_k^{n-1} $ by $ \psi_k^\delta $ we get the family of operators taking $ f $ into $$(2\pi)^{-n-1}\sum_{k=0}^{\infty}\int_{-\infty}^{\infty}e^{-i\lambda t}\psi_k^{\delta}(\sqrt{|\lambda|}r)f^{\lambda}\ast_{\lambda}\varphi_k^{\lambda}(z)|\lambda|^n\, d\lambda.$$ We make use of these operators in studying the $ L^p $ improving properties of the spherical mean value operator.
In what follows we require sharp estimates on the normalised Laguerre functions given in . It is convenient to express $\psi_k^{\delta}(r)$ in terms of the standard Laguerre functions $$\mathcal{L}_k^{\delta}(r)=\Big(\frac{\Gamma(k+1)\Gamma(\delta+1)}{\Gamma(k+\delta+1)}\Big)^{\frac12}L_k^{\delta}(r)e^{-\frac12r}r^{\delta/2}$$ which form an orthonormal system in $L^2((0,\infty),dr)$. In terms of $\mathcal{L}_k^{\delta}(r),$ we have $$\psi_k^{\delta}(r)=2^{\delta}\Big(\frac{\Gamma(k+1)\Gamma(\delta+1)}{\Gamma(k+\delta+1)}\Big)^{\frac12}r^{-\delta}\mathcal{L}_k^{\delta}\Big(\frac12r^2\Big).$$ Asymptotic properties of $\mathcal{L}_k^{\delta}(r)$ are well known in the literature and we have the following result, see [@T Lemma 1.5.3] (actually, the estimates in Lemma \[lem:T\] below are sharp, see [@M Section 2] and [@Mu Section 7]).
\[lem:T\] For $\delta>-1$, we have the following: $$|\mathcal{L}_k^{\delta}(r)|\le C\begin{cases}(kr)^{\delta/2}, &0\le r\le \frac{1}{k}\\
(kr)^{-\frac14}, &\frac{1}{k}\le r\le \frac{k}{2}\\
k^{-\frac14}(k^{\frac13}+|k-r|)^{-\frac14}, &\frac{k}{2}\le r\le \frac{3k}{2}\\
e^{-\gamma r}, & r\ge \frac{3k}{2},
\end{cases}$$ where $\gamma>0$ is a fixed constant.
We can now rewrite the above estimates of $\mathcal{L}_k^{\delta}$ in terms of estimates for the normalised Laguerre functions $\psi_k^{\delta}$.
\[lem:uniform\] For any $\delta\ge-\frac13,$ we have the uniform estimates $$\sup_k|\psi_k^{\delta}(\sqrt{|\lambda|})|\le C\begin{cases}
1, &\text{ if } |\lambda|\le 1\\
|\lambda|^{-\delta-\frac13}, &\text{ if } |\lambda|> 1.
\end{cases}$$
Since $\frac{\Gamma(k+1)\Gamma(\delta+1)}{\Gamma(k+\delta+1)}\le Ck^{-\delta}$ we need to bound $(k|\lambda|)^{-\delta/2}\mathcal{L}_k^{\delta}\big(\frac12|\lambda|\big)$ for $|\lambda|\ge 1$. When $1\le \frac12|\lambda|\le \frac{k}{2}$ we have the estimate $$|\psi_k^{\delta}(\sqrt{|\lambda|})|\le C(k|\lambda|)^{-\delta/2-1/4}.$$ From here, since $\delta+\frac12\ge 0$, $\lambda^2\le k|\lambda|$, we get $$|\psi_k^{\delta}(\sqrt{|\lambda|})|\le C|\lambda|^{-\delta-1/2}.$$ When $\frac{k}{2}\le \frac12|\lambda|\le \frac{3k}{2}$, $|\lambda|$ is comparable to $k$ and hence we have $$|\psi_k^{\delta}(\sqrt{|\lambda|})|\le C(k|\lambda|)^{-\delta/2}k^{-\frac14}k^{-\frac{1}{12}}\le C|\lambda|^{-\delta-\frac13}.$$ On the region $|\lambda|\ge \frac{3k}{2}$ we have exponential decay. Finally, the estimate $\sup_k|\psi_k^{\delta}(\sqrt{|\lambda|})|\le C$ for $0\le |\lambda|\le1$ is immediate, in view of Lemma \[lem:T\]. With this we prove the lemma.
\[lem:uniform2\] For any $\delta\ge \frac12$ and $|\lambda|\ge1$ we have $$\sup_k(k|\lambda|)^{\frac12}|\psi_k^{\delta}(\sqrt{|\lambda|})|\le C|\lambda|^{-\delta+\frac23}.$$
As in the proof of Lemma \[lem:uniform\], in the range $1\le \frac12|\lambda|\le \frac{k}{2}$, $$(k|\lambda|)^{\frac12}|\psi_k^{\delta}(\sqrt{|\lambda|})|\le C(k|\lambda|)^{-\delta/2+1/4}\le C|\lambda|^{-\delta+\frac12}$$ as $\delta\ge\frac12$. When $\frac{k}{2}\le \frac12|\lambda|\le \frac{3k}{2}$, as before $$(k|\lambda|)^{\frac12}|\psi_k^{\delta}(\sqrt{|\lambda|})|\le C|\lambda|^{-\delta+1-\frac13}=C|\lambda|^{-\delta+\frac23}.$$
The Laguerre functions $\psi_k^{\delta}$ can be defined for all values of $\delta>-1$, even for complex $\delta$ with ${\operatorname{Re}}\delta>-1$ and we would like to use this fact to embed $A_1$ into an analytic family of operators. With the analytic interpolation in mind we define $$\label{eq:Abeta}
\mathcal{A}^{\beta}f(z,t)=(2\pi)^{-n-1}\int_{-\infty}^{\infty}e^{-i\lambda t}\Big(\sum_{k=0}^{\infty}\psi_k^{\beta+n-1}(\sqrt{|\lambda|})f^{\lambda}\ast_{\lambda}\varphi_k^{\lambda}(z)\Big)|\lambda|^n\,d\lambda,$$ for ${\operatorname{Re}}(\beta+n-1)>-1$. Note that for $\beta=0$ we recover $A_1$, thus $A_1=\mathcal{A}^0$. We will use the following relation between Laguerre polynomials of different types in order to express $\mathcal{A}^{\beta}$ in terms of $A_1$ (see [@PBM (2.19.2.12)]) $$\label{eq:connection}
L_k^{\mu+\nu}(r)=\frac{\Gamma(k+\mu+\nu+1)}{\Gamma(\nu)\Gamma(k+\mu+1)}\int_0^1t^{\mu}(1-t)^{\nu-1}L_k^{\mu}(rt)\,dt,$$ valid for ${\operatorname{Re}}\mu>-1$ and ${\operatorname{Re}}\nu>0$. We define $$\label{eq:Poisson}
P_rf(z,t)=\frac{1}{2\pi}\int_{-\infty}^{\infty}e^{-i\lambda t}e^{-\frac14|\lambda|r}f^{\lambda}(z)\,d\lambda$$ to be the Poisson integral of $f$ in the $t$-variable. We see that for ${\operatorname{Re}}\beta>0$, $\mathcal{A}^{\beta}$ is given by the following representation.
\[lem:family\] Let ${\operatorname{Re}}\beta>0$. The operator $\mathcal{A}^{\beta}$ is given by the formula $$\mathcal{A}^{\beta}f(z,t)=2\frac{\Gamma(\beta+n)}{\Gamma(\beta)\Gamma(n)}\int_0^1r^{2n-1}(1-r^2)^{\beta-1}P_{1-r^2}f\ast \mu_r(z,t)\,dr.$$
In view of , it is enough to verify $$\begin{gathered}
2\frac{\Gamma(\beta+n)}{\Gamma(\beta)\Gamma(n)}\int_0^1r^{2n-1}(1-r^2)^{\beta-1}P_{1-r^2}f\ast \mu_r(z,t)\,dr\\
=(2\pi)^{-n-1}\int_{-\infty}^{\infty}e^{-i\lambda t}\Big(\sum_{k=0}^{\infty}\psi_k^{\beta+n-1}(\sqrt{|\lambda|})f^{\lambda}\ast_{\lambda}\varphi_k^{\lambda}(z)\Big)|\lambda|^n\,d\lambda.\end{gathered}$$ Note that the left hand side of the above equation is well defined only for ${\operatorname{Re}}\beta>0$ whereas the right hand side makes sense for all ${\operatorname{Re}}\beta>-n$. We can thus think of the right hand side as an analytic continuation of the left hand side. In view of , the Poisson integral $P_rf$ of $f$ in the $t$-variable can be written as $$(P_rf)^{\lambda}(z)=e^{-\frac14|\lambda|r}f^{\lambda}(z).$$ Then, by we consider the equation $$P_{1-r^2}f\ast \mu_r(z,t)=(2\pi)^{-n-1}\sum_{k=0}^{\infty}\int_{-\infty}^{\infty}e^{-i\lambda t}\psi_k^{n-1}(\sqrt{|\lambda|}r)e^{-\frac14|\lambda|(1-r^2)}f^{\lambda}\ast_{\lambda}\varphi_k^{\lambda}(z)|\lambda|^n\,d\lambda.$$ Integrating the above equation against $r^{2n-1}(1-r^2)^{\beta-1}\,dr$, we obtain $$\int_0^1r^{2n-1}(1-r^2)^{\beta-1}P_{1-r^2}f\ast \mu_r(z,t)\,dr=(2\pi)^{-n-1}\sum_{k=0}^{\infty}\int_{-\infty}^{\infty}e^{-i\lambda t}\rho_k(\sqrt{|\lambda|})f^{\lambda}\ast_{\lambda}\varphi_k^{\lambda}(z)|\lambda|^n\,d\lambda,$$ where $$\label{eq:rho}
\rho_k(\sqrt{|\lambda|})=\int_0^1r^{2n-1}(1-r^2)^{\beta-1}\psi_k^{n-1}(\sqrt{|\lambda|}r)e^{-\frac14|\lambda|(1-r^2)}\,dr.$$ Recalling the definition of $\psi_k^{n-1}$ given in we have $$\rho_k(\sqrt{|\lambda|})=\frac{\Gamma(k+1)\Gamma(n)}{\Gamma(k+n)}\int_0^1r^{2n-1}(1-r^2)^{\beta-1}L_k^{n-1}\Big(\frac12r^2|\lambda|\Big)e^{-\frac14|\lambda|}\,dr.$$ We now use the formula . First we make a change of variables $t\to s^2$ and then choose $\mu=n-1$ and $\nu=\beta$, so that $$\label{eq:rho2}
\rho_k(\sqrt{|\lambda|})=\frac{\Gamma(k+1)\Gamma(n)}{\Gamma(k+n)}\frac12\frac{\Gamma(\beta)\Gamma(k+n)}{\Gamma(k+n+\beta)}e^{-\frac14|\lambda|}L_k^{n+\beta-1}\Big(\frac{|\lambda|}{2}\Big)=\frac12\frac{\Gamma(\beta)\Gamma(n)}{\Gamma(\beta+n)}\psi_k^{n+\beta-1}(\sqrt{|\lambda|}).$$ The proof is complete.
In particular, from the computations in the proof of Lemma \[lem:family\], we infer the following identity.
\[cor:ident\] Let ${\operatorname{Re}}\beta>0$ and $\alpha>-1$. Then, for $t>0$, $$\psi_k^{\alpha+\beta}(t)=2\frac{\Gamma(\beta+\alpha+1)}{\Gamma(\beta)\Gamma(\alpha+1)}\int_0^1s^{\alpha}(1-s)^{\beta-1}\psi_k^{\alpha}(t\sqrt{s})e^{-\frac14t^2(1-s)}\,ds.$$
The identity follows from and , after a change of variable.
We slightly modify the family in Lemma \[lem:family\] and define a new family $T_{\beta}$. The modification becomes necessary since we want our family to have some $ L^p $ improving property for large values of $ \beta.$ The original operator $ \mathcal{A}^\beta $ remains as convolution with a distribution supported on $ {{\mathbb C}}^n \times \{0\} $ however large $ \beta$ is. This is in sharp contrast with the Euclidean case, see [@S]. As we will see below the modified family of operators $ T_\beta$ has a better behaviour for $ \beta \ge 1.$ Let $k_\beta(t)=\frac{1}{\Gamma(\beta)}t_+^{\beta-1}e^{-t}$, ${\operatorname{Re}}\beta>0$, which defines a family of distributions on ${\mathbb R}$ and $\lim_{\beta\to 0} k_\beta(t)=\delta_0$, the Dirac distribution at $0$. Given a function $f$ on ${\mathbb H}^n$ and $\varphi$ on ${\mathbb R}$ we use the notation $f\ast_3 \varphi$ to stand for the convolution in the central variable: $$f\ast_3 \varphi(z,t)=\int_{-\infty}^{\infty}f(z,t-s)\varphi(s)\,ds.$$ Thus we note that $P_{1-r^2}f(z,t)=f\ast_3 p_{1-r^2}(z,t)$ where $p_{1-r^2}$ is the usual Poisson kernel in the one dimensional variable $t$, associated to $P_{1-r^2}$. In fact, $p_r(t) $ is defined by the relation $ \int_{-\infty}^\infty e^{i \lambda t} p_r(t) dt = e^{-\frac{1}{4}r|\lambda|} $ and it is explicitly known: $ p_r(t) = c r(r^2+16 t^2)^{-1}$ for some constant $ c>0,$ see for example [@SW]. With the above notation we define the new family by $$T_{\beta}f(z,t)=\frac{\Gamma(\beta+n)}{\Gamma(\beta)\Gamma(n)}\int_0^1r^{2n-1}(1-r^2)^{\beta-1}P_{1-r^2}(f\ast_3 k_\beta)\ast \mu_r(z,t)\,dr.$$ In other words $$T_{\beta}f=\mathcal{A}^{\beta}(f\ast_3 k_\beta).$$
The operator $T_{\beta}f$ has the explicit expansion $$T_{\beta}f(z,t)=(2\pi)^{-n-1}\int_{-\infty}^{\infty}e^{-i\lambda t}(1-i\lambda)^{-\beta}\Big(\sum_{k=0}^{\infty}\psi_k^{\beta+n-1}(\sqrt{|\lambda|})f^{\lambda}\ast_{\lambda}\varphi_k^{\lambda}(z)\Big)|\lambda|^n\,d\lambda.$$
The statement follows from Lemma \[lem:family\], , and from the fact $$\int_{-\infty}^{\infty}e^{i\lambda t}k_\beta(t)\,dt=\frac{1}{\Gamma(\beta)}\int_0^{\infty}e^{i\lambda t}t^{\beta-1}e^{-t}\,dt=(1-i\lambda)^{-\beta}.$$ This can be verified by looking at the function $$F(\beta,z)=\frac{1}{\Gamma(\beta)}\int_0^{\infty}t^{\beta-1}e^{-tz}\,dt$$ defined and holomorphic for ${\operatorname{Re}}\beta>0$, ${\operatorname{Re}}z>0$. Indeed, when $z$, with ${\operatorname{Re}}z>0$, is fixed, we have the relation $F(\beta,z)=zF(\beta+1,z)$ which allows us to holomorphically extend $F(\beta,z)$ in the $\beta$ variable. It is clear that when $z>0$, $F(\beta,z)=z^{-\beta}$, which allows us to conclude that the Fourier transform of $ k_\beta$ at $\lambda$ is given by $(1-i\lambda)^{-\beta}$, as claimed.
We will show that when $\beta=1+i\gamma$, $T_{\beta}$ is bounded from $L^p({\mathbb H}^n)$ into $L^{\infty}({\mathbb H}^n)$ for any $p>1$, and for certain negative values of $\beta$, $T_{\beta}$ is bounded on $L^2({\mathbb H}^n)$. We can then use analytic interpolation to obtain a result for $T_0=\mathcal{A}^0=A_1$.
\[prop:Linfty\] For any $\delta>0$, $\gamma\in {\mathbb R}$ $$\|T_{1+i\gamma}f\|_{\infty}\le C_1(\gamma)\|f\|_{1+\delta},$$ where $C_1(\gamma)$ is of admissible growth.
Without loss of generality we can assume that $f\ge0$. For $\beta=1+i\gamma$ it follows that $$|T_{1+i\gamma}f(z,t)|\le \frac{|\Gamma(1+i\gamma+n)|}{|\Gamma(1+i\gamma)|^2\Gamma(n)}\int_0^1r^{2n-1}P_{1-r^2}(f\ast_3 \varphi)\ast \mu_r(z,t)\,dr$$ where $\varphi(t)=e^{-t}\chi_{(0,\infty)}(t)$. Since $\varphi\ge0$ it follows that $$P_{1-r^2}(f\ast_3 \varphi)=\varphi\ast_3p_{1-r^2}\ast_3f\le \varphi\ast_3\Lambda f$$ where $\Lambda f$ is the Hardy–Littlewood maximal function in the $t$-variable. In proving the above we have used the well known fact that $\sup_{r>0}|P_rg(t)|\le C\Lambda g(t)$ for any $g$ on ${\mathbb R}$. Thus we have the estimate $$|T_{1+i\gamma}f(z,t)|\le C_1(\gamma)\int_0^1 (\Lambda f\ast_3 \varphi)\ast \mu_r(z,t)r^{2n-1}\,dr.$$ Now we make the following observation: Suppose $K(z,t)=k(|z|)\varphi(t)$. Then $$f\ast K(z,t)=\int_0^{\infty}(f\ast_3 \varphi)\ast \mu_r(z,t)k(r)r^{2n-1}\,dr,$$ which can be verified by recalling the definition of the spherical means $f\ast\mu_r(z,t)$ in and integrating in polar coordinates. This gives us $$|T_{1+i\gamma}f(z,t)|\le C_1(\gamma)\Lambda f\ast K(z,t)$$ where $K(z,t)=\chi_{|z|\le 1}(z)\varphi(t)$. As $\Lambda f\in L^{1+\delta}({\mathbb H}^n)$ and $K\in L^q({\mathbb H}^n)$ for any $q\ge 1$, by Hölder we get $$\|T_{1+i\gamma}f\|_{\infty}\le C_1(\gamma)\|\Lambda f\|_{1+\delta}\le C_1(\gamma)\|f\|_{1+\delta}.$$
In the next proposition we show that $T_{\beta}$ is bounded on $L^2({\mathbb H}^n)$ for some $\beta<0$. It is possible to sharpen the following result, see Subsection \[rem:sharp\], but for the sake of simplicity (and to mimic the corresponding Euclidean result), we consider only the case ${\operatorname{Re}}\beta \ge -\frac{(n-1)}{2}$.
\[prop:L2\] Assume that $n\ge1$ and $\beta\ge -\frac{(n-1)}{2}$. Then for any $\gamma\in {\mathbb R}$ $$\|T_{\beta+i\gamma}f\|_{2}\le C_2(\gamma)\|f\|_{2}.$$
We only have to check that the functions $$(1+\lambda^2)^{-\beta/2}|\psi_k^{\beta+i\gamma+n-1}(\sqrt{|\lambda|})|\le C_2(\gamma)$$ where $C_2(\gamma)$ is independent of $K$ and $\lambda$. When $\gamma=0$, it follows from the estimates of Lemma \[lem:uniform\] that $$(1+\lambda^2)^{-\beta/2}|\psi_k^{\beta+n-1}(\sqrt{|\lambda|})|\le C|\lambda|^{-\beta}|\lambda|^{-\beta-(n-1)-\frac13}$$ for $|\lambda|\ge 1$, which is clearly bounded for $\beta\ge -\frac{n-1}{2}$ (actually, it is bounded for $\beta\ge -\frac{n}{2}+\frac13$, so it is for $\beta\ge -\frac{n}{2}+\frac12$). For $\gamma \neq 0$ we can express $\psi_k^{\beta+i\gamma+n-1}(\sqrt{|\lambda|})$ in terms of $\psi_k^{\beta-\varepsilon+n-1}(\sqrt{|\lambda|})$ for a small enough $\varepsilon>0$ and obtain the same estimate. Indeed, by Corollary \[cor:ident\] and using the asymptotic formula $|\Gamma(\mu+iv)|\sim \sqrt{2\pi}|v|^{\mu-1/2}e^{-\pi|v|/2}$, as $v\to \infty$ (see for instance [@SW p. 281 bottom note]) $$\begin{aligned}
|\psi_k^{\beta+i\gamma+n-1}(\sqrt{|\lambda|})|&=\Big|2\frac{\Gamma(\beta+i\gamma+n)}{\Gamma(\varepsilon+i\gamma)\Gamma(\beta-\varepsilon+n)}\\
&\quad \times \int_0^1s^{\beta-\varepsilon+n-1}(1-s)^{\varepsilon+i\gamma-1}\psi_k^{\beta-\varepsilon+n-1}(\sqrt{|\lambda|s})e^{-\frac14|\lambda|(1-s)}\,ds\Big|\\
&\lesssim \frac{|\gamma|^{\beta+n-1/2}}{|\gamma|^{\varepsilon-1/2}}\Big|\int_0^1s^{\beta-\varepsilon+n-1}(1-s)^{\varepsilon+i\gamma-1}\psi_k^{\beta-\varepsilon+n-1}(\sqrt{|\lambda|s})e^{-\frac14|\lambda|(1-s)}\,ds\Big|,\end{aligned}$$ where the constant depends on $\beta$. Now, by the estimate for $\psi_k^{\delta}$ in Lemma \[lem:uniform\] and the integrability of the function $s^{\beta-\varepsilon+n-1}(1-s)^{\varepsilon+i\gamma-1}$ we have $$(1+\lambda^2)^{-\beta/2}|\psi_k^{\beta+i\gamma+n-1}(\sqrt{|\lambda|})|\le C|\lambda|^{-\beta}|\gamma|^{\beta+n-1-\varepsilon}|\lambda|^{-(\beta+n-1-\varepsilon)-1/3}.$$ For $|\lambda|\ge 1$, the above is bounded for $\beta\ge -\frac{n-1}{2}$ (actually, it is bounded for $\beta-\varepsilon\ge -\frac{n}{2}+\frac13$ with $\varepsilon$ small enough, so it is for $\beta\ge -\frac{n}{2}+\frac12$). The proof is complete.
\[thm:Lp\] Assume that $n\ge 2$. Then $A_1:L^p({\mathbb H}^n)\to L^{n+1}({\mathbb H}^n)$ for any $\frac{n+1}{n} < p < (n+1)$.
For $ \frac{n+1}{n} < p < (n+1)$ choose $\delta>0$ such that $p=\frac{(n+1)(1+\delta)}{n+\delta} $, which is possible as $ \frac{1}{n} < \frac{1+\delta}{n+\delta} < 1.$ By considering the analytic family $T_{\alpha(z)}$ where $\alpha(z)=\frac{n-1}{2}(z-1)+z$ with $z=u+iv$, in view of Propositions \[prop:Linfty\] and \[prop:L2\], and interpolation between the endpoints ${\operatorname{Re}}z=0$ and ${\operatorname{Re}}z=1$ we obtain $$T_{\alpha(u)}:L^{p_u}({\mathbb H}^n)\to L^{q_u}({\mathbb H}^n)$$ where $\frac{1}{p_u}=\frac{1-u}{2}+\frac{u}{1+\delta}$ and $\frac{1}{q_u}=\frac{1-u}{2}$. The choice $u=\frac{n-1}{n+1}$ gives $q_u=n+1$ and $p_u=\frac{(n+1)(1+\delta)}{n+\delta}=p$. Since $\alpha\big(\frac{n-1}{n+1}\big)=0$ we obtain the result.
Observe the restriction on the dimension in Theorem \[thm:Lp\], that comes into play due to the restriction (that we imposed, a bit artificially, for cosmetic reasons) on the parameter $\beta$ in Proposition \[prop:L2\]. This is the only place in the $L^p$-improving estimates where the dimensional restriction arises, but we insist that we imposed that. Actually, the results we can obtain concerning $L^p$-improving estimates are sharp and valid for all the dimensions, see Subsection \[rem:sharp\].
\[cor:LpLq\] Assume that $n\ge 2$. Then $$A_1:L^{p}({\mathbb H}^n)\to L^{q}({\mathbb H}^n)$$ whenever $\big(\frac{1}{p},\frac{1}{q}\big)$ lies in the interior of the triangle joining the points $(0,0), (1,1)$ and $\big(\frac{n}{n+1},\frac{1}{n+1}\big)$, as well as the straight line segment joining the points $(0,0), (1,1)$, see $\mathbf{L}_n'$ in Figure \[figurea\].
The result follows from Theorem \[thm:Lp\] after applying Marcinkiewicz interpolation theorem with the obvious estimates $\|A_1f\|_1\le \|f\|_1$ and $\|A_1f\|_{\infty}\le \|f\|_{\infty}$.
![Triangle $\mathbf{L}_n'$ shows the region for $L^p-L^q$ estimates for $A_1$. The dual triangle $\mathbf{L}_n$ is on the left.[]{data-label="figurea"}](FiguresaDual.pdf "fig:") ![Triangle $\mathbf{L}_n'$ shows the region for $L^p-L^q$ estimates for $A_1$. The dual triangle $\mathbf{L}_n$ is on the left.[]{data-label="figurea"}](Figuresa.pdf "fig:")
A sharpened result {#rem:sharp}
------------------
As indicated in the proof of Proposition \[prop:L2\], we could state an enhanced result as follows.
\[prop:L2s\] Assume that $n\ge1$ and $\beta> -\frac{n}{2}+\frac13$. Then for any $\gamma\in {\mathbb R}$ $$\|T_{\beta+i\gamma}f\|_{2}\le C_2(\gamma)\|f\|_{2}.$$
On the other hand, let us consider the following holomorphic function $\alpha(z)$ on the strip $\{z:0\le {\operatorname{Re}}z\le 1\}$, given by $\alpha(z)=\big(\frac{n}{2}-\frac13\big)(z-1+\varepsilon)+z$. We have $\alpha(0)=\big(-\frac{n}{2}+\frac13\big)(1-\varepsilon)$ and $\alpha(1)=1$. Then, $T_{\alpha(z)}$ is an analytic family of linear operators and it was already shown that $T_{1+i\gamma}$ is bounded from $L^{1+\delta}({\mathbb H}^n)$ to $L^{\infty}({\mathbb H}^n)$. Therefore, we can apply Stein’s interpolation theorem. Letting $z=u+iv$, we have $$\alpha(z)=0 \Longleftrightarrow \Big(\frac{n}{2}-\frac13\Big)(u-1+\varepsilon)+u=0 \Longleftrightarrow u=\frac{3n-2}{3n+4}(1-\varepsilon).$$ Since $\varepsilon>0$ is arbitrary, we obtain $$T_{\alpha(u)}:L^{p_u}({\mathbb H}^n)\to L^{q_u}({\mathbb H}^n)$$ where $$\frac{1}{p_u}=\frac{3n+1}{3n+4}-\varepsilon\frac{3n-2}{2(3n+4)},\quad\frac{1}{q_u}=\frac{3+\frac12(3n-2)\varepsilon}{3n+4}.$$ This leads to the following result, the enhanced version of Theorem \[thm:Lp\].
\[prop:L2en\] Assume that $n\ge 1$ and $\varepsilon>0$. Then $A_1:L^{p}({\mathbb H}^n)\to L^{q}({\mathbb H}^n)$ for any $p,q$ such that $$\frac{1}{p}=\frac{3n+1}{3n+4}-\varepsilon\frac{3n-2}{2(3n+4)},\quad\frac{1}{q}=\frac{3+\frac12(3n-2)\varepsilon}{3n+4}.$$
And we deduce the following corollary.
\[cor:LPs\] Assume that $n\ge 1$. Then $$A_1:L^{p}({\mathbb H}^n)\to L^{q}({\mathbb H}^n)$$ whenever $\big(\frac{1}{p},\frac{1}{q}\big)$ lies in the interior of the triangle joining the points $(0,0), (1,1)$ and $\big(\frac{3n+1}{3n+4},\frac{3}{3n+4}\big)$, as well as the straight line segment joining the points $(0,0), (1,1)$, see $\mathbf{S}_n'$ in Figure \[figureaSH\].
![Triangle $\mathbf{S}_n'$ shows the region for sharpened $L^p-L^q$ estimates for $A_1$. The dual triangle $\mathbf{S}_n$ is on the left.[]{data-label="figureaSH"}](FiguresaSHDual.pdf "fig:") ![Triangle $\mathbf{S}_n'$ shows the region for sharpened $L^p-L^q$ estimates for $A_1$. The dual triangle $\mathbf{S}_n$ is on the left.[]{data-label="figureaSH"}](FiguresaSH.pdf "fig:")
The continuity property of the spherical mean value operator {#sec:continuity}
============================================================
In the work of Lacey [@Lacey] dealing with the lacunary spherical maximal function on $ {\mathbb R}^n $, the continuity property of the spherical mean value operator plays a crucial role. In the case of the Heisenberg group we require the following continuity property.
\[prop:continuity\] Assume that $ n \geq 2.$ Then for $ y \in {\mathbb H}^n, |y| \leq 1, $ we have $$\| A_1 - A_1 \tau_y \|_{L^2\to L^2} \leq C |y|$$ where $ \tau_y f(x) = f(xy^{-1}) $ is the right translation operator.
For $ f \in L^2({\mathbb H}^n) $ we estimate the $ L^2 $ norm of $ A_1f -A_1(\tau_yf) $ using Plancherel theorem for the Fourier transform on $ {\mathbb H}^n.$ Recall that $ A_1f(x) = f \ast \mu_1(x) $ so that $ \widehat{A_1f}(\lambda) = \widehat{f}(\lambda)\widehat{\mu_1}(\lambda)$, where $ \widehat{\mu_1}(\lambda) $ is explicitly given by $$\widehat{\mu_1}(\lambda) = \sum_{k=0}^\infty \psi_k^{n-1}(\sqrt{|\lambda|}) P_k(\lambda) .$$ We also have $$\widehat{\tau_yf}(\lambda) = \int_{{\mathbb H}^n} f(xy^{-1}) \pi_\lambda(x) dx = \widehat{f}(\lambda)\pi_\lambda(y).$$ Thus by the Plancherel theorem for the Fourier transform we have $$\| A_1f -A_1(\tau_yf)\|_2^2 = c_n \int_{-\infty}^\infty \| \widehat{f}(\lambda)(I-\pi_\lambda(y))\widehat{\mu_1}(\lambda)\|_{\operatorname{HS}}^2 |\lambda |^n d\lambda.$$ Since the space of all Hilbert-Schmidt operators is a two sided ideal in the space of all bounded linear operators, it is enough to estimate the operator norm of $ (I-\pi_\lambda(y))\widehat{\mu_1}(\lambda).$ (For more about Hilbert-Schmidt operators see V. S. Sunder [@VSS].) Again, $ \widehat{\mu_1}(\lambda) $ is self adjoint and $ \pi_\lambda(y)^\ast = \pi_\lambda(y^{-1})$ and so we will estimate $ \widehat{\mu_1}(\lambda) (I-\pi_\lambda(y)).$
We make use of the fact that for every $ \sigma \in U(n) $ there is a unitary operator $ \mu_\lambda(\sigma) $ acting on $ L^2({\mathbb R}^n) $ such that $ \pi_\lambda(\sigma z, t) = \mu_\lambda(\sigma)^\ast \pi_\lambda(z,t) \mu_\lambda(\sigma) $ for all $ (z,t) \in {\mathbb H}^n.$ This follows from the well known Stone–von Neumann theorem which says that any irreducible unitary representation of the Heisenberg group which acts like $ e^{i\lambda t}I $ when restricted to the center is unitarily equivalent to $ \pi_\lambda,$ see [@F]. Actually, $ \mu_\lambda $ has an extension to a double cover of the symplectic group as a unitary representation and is called the metaplectic representation. Given $ y = (z, t) \in {\mathbb H}^n $ we can choose $ \sigma \in U(n) $ such that $ y =(|z|\sigma e_1,t) $ where $ e_1 = (1,0,....,0).$ Thus $$\pi_\lambda(y) = \mu_\lambda(\sigma)^\ast \pi_\lambda(|z|e_1,t)\mu_\lambda(\sigma).$$ Also, it is well known that $ \mu_\lambda(\sigma) $ commutes with functions of the Hermite operator $H(\lambda)$ given by $H(\lambda)=-\Delta+\lambda^2|x|^2$. Since $ \widehat{\mu_1}(\lambda) $ is a function of $ H(\lambda) $ it follows that $$\widehat{\mu_1}(\lambda) (I-\pi_\lambda(z,t)) =\mu_\lambda(\sigma)^\ast \widehat{\mu_1}(\lambda) (I-\pi_\lambda(|z|e_1,t))\mu_\lambda(\sigma).$$ Thus it is enough to estimate the operator norm of $ \widehat{\mu_1}(\lambda) (I-\pi_\lambda(|z|e_1,t)).$ In view of the factorisation $ \pi_\lambda(|z|e_1,t) = \pi_\lambda(|z|e_1,0) \pi_\lambda(0,t) $ we have that $$I-\pi_\lambda(|z|e_1,t)=I- \pi_\lambda(|z|e_1,0) \pi_\lambda(0,t)=(I- \pi_\lambda(0,t))+(I-\pi_\lambda(|z|e_1,0)) \pi_\lambda(0,t)$$ so it suffices to estimate the norms of $ \widehat{\mu_1}(\lambda) (I-\pi_\lambda(0,t)) $ and $\widehat{\mu_1}(\lambda) (I-\pi_\lambda(|z|e_1,0)) \pi_\lambda(0,t) $ separately. Moreover, we only have to estimate them for $ |\lambda| \geq 1$ as they are uniformly bounded for $ |\lambda| \leq 1.$
Assuming $ |\lambda| \geq 1 $ we have, in view of , $$\widehat{\mu_1}(\lambda) (I-\pi_\lambda(0,t)) \varphi(\xi) = (1-e^{i\lambda t}) \widehat{\mu_1}(\lambda)\varphi(\xi), \quad \varphi \in L^2({\mathbb R}^n).$$ By mean value theorem, the operator norm of $ (1-e^{i\lambda t}) \widehat{\mu_1}(\lambda) $ is bounded by $$C |t| |\lambda| \sup_{k} |\psi_k^{n-1}(\sqrt{|\lambda|})| \leq C |t| |\lambda|^{-(n-1)+2/3}$$ where we have used the estimate in Lemma \[lem:uniform\]. Thus for $ n \geq 2 ,$ $$\|\widehat{\mu_1}(\lambda) (I-\pi_\lambda(0,t)) \|_{L^2\to L^2} \leq C |t| \leq C |(z,t)|^2,$$ where $ |x| = |(z,t)| = (|z|^4+t^2)^{1/4}$ is the Koranyi norm on $ {\mathbb H}^n $. In order to estimate $\widehat{\mu_1}(\lambda) (I-\pi_\lambda(|z|e_1,0))$ we recall that $$\pi_\lambda(|z|e_1,0)\varphi(\xi) = e^{i\lambda |z| \xi_1} \varphi(\xi), \quad \varphi \in L^2({\mathbb R}^n).$$ Since we can write $$(1-e^{i\lambda |z| \xi_1}) = -i\lambda |z| \xi_1 \int_0^1 e^{it\lambda |z| \xi_1} dt = \lambda |z| \xi_1 m_\lambda(|z|, \xi)$$ with a bounded function $ m_\lambda(|z|,\xi), $ it is enough to estimate the norm of the operator $ |z| \widehat{\mu_1}(\lambda) M_\lambda $ where $ M_\lambda \varphi (\xi) = \lambda \xi_1 \varphi(\xi).$
Let $ A(\lambda) = \frac{\partial}{\partial \xi_1}+|\lambda| \xi_1 $ and $ A(\lambda)^\ast = -\frac{\partial}{\partial \xi_1}+|\lambda| \xi_1 $ be the annihilation and creation operators, so that we can express $ M_\lambda $ as $ M_\lambda = \frac{1}{2} ( A(\lambda) + A(\lambda)^\ast) .$ Thus it is enough to consider $ |z| \widehat{\mu_1}(\lambda) A(\lambda) $ and $ |z| \widehat{\mu_1}(\lambda) A(\lambda)^\ast. $ Moreover as the Riesz transforms $ H(\lambda)^{-1/2}A(\lambda) $ and $ H(\lambda)^{-1/2}A(\lambda)^\ast $ are bounded on $ L^2({\mathbb R}^n) $ we only need to consider $ |z|\widehat{\mu_1}(\lambda) H(\lambda)^{1/2}.$ But the operator norm of $
\widehat{\mu_1}(\lambda) H(\lambda)^{1/2} $ is given by $\sup_{k} ((2k+n)|\lambda|)^{1/2} |\psi_k^{n-1}(\sqrt{|\lambda|})|$ which, in view of Lemma \[lem:uniform2\], is bounded by $ C |\lambda|^{-(n-1)+2/3}.$ Thus for $ n \geq 2 $ we obtain $$\|\widehat{\mu_1}(\lambda) (I-\pi_\lambda(|z|e_1,0))\|_{L^2\to L^2} \leq C |z| \leq C |(z,t)|.$$ This completes the proof of the proposition.
Observe that the result above is restricted to the case $n\ge2$, and this due to the restriction on the available (and sharp!) estimates for the Laguerre functions in Lemmas \[lem:T\], \[lem:uniform\] and \[lem:uniform2\]. We do not know whether there is a way to reach $n=1$ with our approach.
\[cor:A1\] Assume that $n\ge 2$. Then for $ y \in {\mathbb H}^n$, $|y| \leq 1$, and for $\big(\frac1p,\frac1q\big)$ in the interior of the triangle joining the points $(0,0), (1,1)$ and $\big(\frac{n}{n+1},\frac{1}{n+1}\big)$, we have the inequalities $$\| A_1 - A_1 \tau_y \|_{L^p\to L^q} \leq C |y|^{\eta}$$ for some $0<\eta<1$, where $ \tau_y f(x) = f(xy^{-1}) $ is the right translation operator.
The result follows by Riesz-Thorin interpolation theorem, taking into account Corollary \[cor:LpLq\] and Proposition \[prop:continuity\].
We need a version of the inequality in Corollary \[cor:A1\] when $ A_1 $ is replaced by $A_r.$ This can be easily achieved by making use of the following lemma which expresses $ A_r $ in terms of $ A_1.$ Let $\delta_r\varphi(w,t)=\varphi(rw,r^2t)$ stand for the non-isotropic dilation on $ {\mathbb H}^n.$
\[lem:dilat\] For any $r> 0$ we have $A_rf=\delta_r^{-1}A_1\delta_rf.$
This is just an easy verification. Starting from the expression in we have $$\begin{aligned}
A_rf(z,t)&=(2\pi)^{-n-1}\int_{-\infty}^{\infty}e^{-i\lambda t}\Big(\sum_{k=0}^{\infty}\psi_k^{n-1}(\sqrt{|\lambda|}r)f^{\lambda}\ast_{\lambda}\varphi_k^{\lambda}(z)\Big)|\lambda|^n\,d\lambda\\
&=(2\pi)^{-n-1}r^{-2n-2}\int_{-\infty}^{\infty}e^{-i\frac{\lambda}{r^2} t}\Big(\sum_{k=0}^{\infty}\psi_k^{n-1}(\sqrt{|\lambda|})f^{\lambda/r^2}\ast_{\lambda/r^2}\varphi_k^{\lambda/r^2}(z)\Big) |\lambda|^n\,d\lambda.\end{aligned}$$ In view of the relation $$f^{\lambda/r^2}(rw)=\int_{-\infty}^{\infty}f(rw,t)e^{i\lambda/r^2t}\,dt=r^2\int_{-\infty}^{\infty}f(rw,r^2t)e^{i\lambda t}\,dt$$ we make the following simple computation: $$\begin{aligned}
f^{\lambda/r^2}\ast_{\lambda/r^2}\varphi_k^{\lambda/r^2}(z)&=\int_{{{\mathbb C}}^n}f^{\lambda/r^2}(w)\varphi_k^{\lambda/r^2}(z-w)e^{-i\frac{\lambda}{r^2} {\operatorname{Im}}z\cdot \bar{w}}\,dw\\
&=\int_{{{\mathbb C}}^n}f^{\lambda/r^2}(rw)\varphi_k^{\lambda}(z/r-w)e^{-i\frac{\lambda}{r^2} {\operatorname{Im}}\frac{z}{r}\cdot \bar{w}}r^{2n}\,dw\\
&=r^{2+2n}\int_{{{\mathbb C}}^n}(\delta_rf)^{\lambda}(w)\varphi_k^{\lambda}(z/r-w)e^{-i\frac{\lambda}{r^2} {\operatorname{Im}}\frac{z}{r}\cdot \bar{w}}r^{2n}\,dw\\
&=r^{2+2n}(\delta_rf\ast_{\lambda}\varphi_k^{\lambda})(z/r).\end{aligned}$$ Therefore, we have $$\begin{aligned}
A_rf(z,t)&=(2\pi)^{-n-1}\int_{-\infty}^{\infty}e^{-i\frac{\lambda}{r^2} t}\Big(\sum_{k=0}^{\infty}\psi_k^{n-1}(\sqrt{|\lambda|})(\delta_rf\ast_{\lambda}\varphi_k^{\lambda})(z/r)\Big)|\lambda|^nd\lambda=A_1(\delta_rf)\Big(\frac{z}{r},\frac{t}{r^2}\Big),\end{aligned}$$ which proves the stated result.
\[cor:dilat2\] Assume that $n\ge 2$. Then for $ y \in {\mathbb H}^n$, $|y| \leq 1$, and for $\big(\frac1p,\frac1q\big)$ in the interior of the triangle joining the points $(0,0), (1,1)$ and $\big(\frac{n}{n+1},\frac{1}{n+1}\big),$ we have the inequality $$\| A_r - A_r \tau_y \|_{L^p\to L^q} \leq C r^{-\eta} |y|^{\eta} r^{(2n+2)(\frac{1}{q}-\frac{1}{p})}$$ for some $ \eta >0$.
Observe that $ \delta_r( \tau_y f) = \tau_{\delta_r^{-1}y}(\delta_r f )$, which follows from the fact that $ \delta_r : {\mathbb H}^n \rightarrow {\mathbb H}^n $ is an automorphism. The corollary follows from Corollary \[cor:A1\], Lemma \[lem:dilat\] and the fact that $ \|\delta_r f\|_p = r^{-\frac{(2n+2)}{p}}$ for any $ 1 \leq p < \infty.$
A sharpened continuity property {#sub:sharpenc}
-------------------------------
By using Corollary \[cor:LPs\] instead of Corollary \[cor:LpLq\], we could obtain a sharpened version of Corollary \[cor:A1\], so that we indeed can obtain the following.
\[cor:dilat2s\] Assume that $n\ge 2$. Then for $ y \in {\mathbb H}^n, |y| \leq 1$, and for $\big(\frac1p,\frac1q\big)$ in the interior of the triangle joining the points $(0,0), (1,1)$ and $\big(\frac{3n+1}{3n+4},\frac{3}{3n+4}\big),$ we have the inequality $$\| A_r - A_r \tau_y \|_{L^p\to L^q} \leq C r^{-\eta} |y|^{\eta} r^{(2n+2)(\frac{1}{q}-\frac{1}{p})}$$ for some $ \eta >0$.
Sparse bounds {#sec:sparse}
=============
Our aim in this section is to prove the sparse bounds for the lacunary spherical maximal function stated in Theorem \[thm:sparse\]. In doing so we closely follow [@Lacey] with suitable modifications that are necessary since we are dealing with a non-commutative set up. We can equip $ {\mathbb H}^n $ with a metric induced by the Koranyi norm which makes it a homogeneous space. On such spaces there is a well defined notion of dyadic cubes and grids with properties similar to their counter parts in the Euclidean setting. However, we need to be careful with the metric we choose since the group is non-commutative.
Recall that the Koranyi norm on $ {\mathbb H}^n $ is defined by $ |x| = |(z,t)| = (|z|^4+t^2)^{1/4}$ which is homogeneous of degree one with respect to the non-isotropic dilations. Since we are considering $ f \ast \mu_r $ it is necessary to work with the left invariant metric $ d_L(x,y) = |x^{-1}y| = d_L(0, x^{-1}y)$ instead of the standard metric $ d(x,y) = |xy^{-1}| = d(0, xy^{-1})$, which is right invariant. The balls and cubes are then defined using $ d_L $. Thus $ B(a,r) = \{ x \in {\mathbb H}^n: |a^{-1}x| < r \}.$ With this definition we note that $ B(a,r) =a\cdot B(0,r) $, a fact which is crucial. This allows us to conclude that when $ f $ is supported in $ B(a,r) $ then $ f \ast \mu_s $ is supported in $ B(a,r+s).$ Indeed, as support of $ \mu_s $ is contained in $ B(0,s) $ we see that $ f \ast \mu_s $ is supported in $ B(a,r)\cdot B(0,s) \subset a\cdot B(0,r)\cdot B(0,s) \subset B(a,r+s).$
Let $\delta\in (0,1)$ with $\delta\le 1/96$. Then there exists a countable set of points $\{z_{\nu}^{k,\alpha} : \nu\in \mathscr{A}_k\}$, $k\in {\mathbb Z}$, $\alpha=1,2,\ldots,N$ and a finite number of dyadic systems $\mathcal{D}^{\alpha}:=\cup_{k\in {\mathbb Z}}\mathcal{D}_k^{\alpha}$, $\mathcal{D}_k^{\alpha}:=\{Q_{\nu}^{k,\alpha}:\nu\in \mathscr{A}_k\}$ such that
1. For every $\alpha\in \{1,2,\ldots, N\}$ and $k\in {\mathbb Z}$ we have
- ${\mathbb H}^n=\cup_{Q\in \mathcal{D}_k^{\alpha}}Q$ (disjoint union).
- $Q,P\in \mathcal{D}^{\alpha}\Rightarrow Q\cap P\in \{\emptyset, P, Q\}$.
- $Q_{\nu}^{k,\alpha}\in \mathcal{D}^{\alpha}\Rightarrow B\big(z_{\nu}^{k,\alpha}, \frac{1}{12}\delta^k\big)\subseteq Q_{\nu}^{k,\alpha}\subseteq B\big(z_{\nu}^{k,\alpha}, 4\delta^k\big)$. In this situation $z_\nu^{k,\alpha} $ is called the center of the cube and the side length $\ell{ (Q_\nu^{k,\alpha})}$ is defined to be $ \delta^k.$
2. For every ball $B=B(x,r)$, there exists a cube $Q_B\in \cup_{\alpha}\mathcal{D}^{\alpha}$ such that $B\subseteq Q_B$ and $\ell(Q_B)=\delta^{k-1}$, where $k$ is the unique integer such that $\delta^{k+1}<r\le \delta^k.$
It follows from Theorem 4.1, the proof of Lemma 4.12, Remark 4.13 and Theorem 2.2 in [@HK], where the choices $c_0=1/4$ and $C_0=2$ in [@HK Theorem 2.2] are made so that the property $(2)$ holds (see [@HK Lemma 4.10]).
We will first prove a lemma that is the analogue of [@Lacey Lemma 2.3].
\[lem:anal23\] Let $f$ and $g$ be supported on a cube $Q$ and let $\ell(Q)=r$. For $\big(\frac1p,\frac1q\big)$ in the interior of the triangle joining the points $(0,1), (1,0)$ and $\big(\frac{n}{n+1},\frac{n}{n+1}\big)$, there holds $$|\langle A_rf-A_r\tau_yf,g\rangle|\lesssim |y/r|^{\eta}|Q|\langle f\rangle_{Q,p}\langle g\rangle_{Q,q}, \quad |y|<r.$$
Observe that continuity property holds for the pair $\big(\frac1p,\frac1{q'}\big)$. By Hölder’s inequality and Corollary \[cor:dilat2\] we have, for $|y|<r$, $$\begin{aligned}
|\langle A_rf-A_r\tau_yf,g\rangle|&\le \|A_rf-A_r\tau_yf\|_{q'}\|g\|_{q}\\
&\le Cr^{(2n+2)(\frac{1}{q'}-\frac{1}{p})}r^{-\eta} |y|^{\eta}\|f\|_p\|g\|_{q}\\
&=Cr^{(2n+2)(\frac{1}{q'}-\frac{1}{p})}r^{-\eta} |y|^{\eta}|Q|^{\frac{1}{p}+\frac{1}{q}}\langle f\rangle_{Q,p}\langle g\rangle_{Q,q}\\
&\lesssim |Q|^{\frac{1}{q'}-\frac{1}{p}}|Q|^{\frac{1}{p}+\frac{1}{q}}r^{-\eta} |y|^{\eta}\langle f\rangle_{Q,p}\langle g\rangle_{Q,q}\\
&\lesssim |Q|r^{-\eta} |y|^{\eta}\langle g\rangle_{Q,p}\langle g\rangle_{Q,q},\end{aligned}$$ as $ |Q|$ is comparable to $ r^{2n+2}$.
\[lem:Hn\] For $Q$ with $\ell(Q)=\delta^k$ we consider $$\mathbb{V}_{Q}=\{P\in \mathcal{D}^1_{k+3}: B(z_{P},\delta^{k+1})\subseteq Q\}.$$ and define $$A_{Q}f=A_{\delta^{k+2}}(f{\bf{1}}_{V_Q})$$ where $ V_{Q}=\cup_{P\in \mathbb{V}_{Q}}P.$ Then for any $ f $ supported in $ Q $ the support of $A_{Q}f $ is also contained in $ Q.$ Moreover, $$A_{\delta^{k+2}}f\le \sum_{\alpha=1}^N\sum_{Q\in \mathcal{D}_k^{\alpha}}A_Q(f).$$
We emphasize that we can take any $\delta$ in Lemma \[lem:Hn\] (and in the rest of the paper), in particular we could take $\delta=2$. In that case we have to do some modifications in defining $A_Q f$, where one has to use the fact that if $\delta<\frac{1}{96}$ then the number of points of the form $2^m$, $m\in {\mathbb Z}$, liying between $\delta^j$ and $\delta^{j+1}$, $j\in {\mathbb Z}$, does not depend on $j$.
Observe that for any $ x \in {\mathbb H}^n$ there exists $P\in \mathcal{D}^1_{k+3}$ such that $ x \in P \subseteq B(z_P, 4\delta^{k+3}).$ Then $ P \subseteq B(z_P,\delta^{k+1})\subseteq Q$ for some $Q$ in $\mathcal{D}_k^{\alpha}$, for some $\alpha$. Therefore $P\in V_{Q}$ and hence $ x \in V_{Q}$. This proves that $ {\mathbb H}^n=\bigcup_{\alpha=1}^N\bigcup_{Q\in \mathcal{D}_k^{\alpha}}\mathbb{V}_Q $ and hence we have $f\le \sum_{\alpha=1}^N\sum_{Q\in \mathcal{D}_k^{\alpha}}f{\bf{1}}_{V_Q}$ and consequently, $A_{\delta^{k+2}}f\le \sum_{\alpha=1}^N\sum_{Q\in \mathcal{D}_k^{\alpha}}A_Qf$. It remains to be proved that $ A_Qf $ is supported in $ Q.$ Now assume that ${\operatorname{supp}}f\subseteq Q$ and recall $A_{\delta^{k+2}}f(x)=f\ast \mu_{\delta^{k+2}}(x)$. Then it is enough to show that ${\operatorname{supp}}A_{\delta^{k+2}}(f{\bf{1}}_P) \subseteq B(z_P, \delta^{k+1})$ for every $P \in \mathbb{V}_Q$. Indeed, $${\operatorname{supp}}(f {\bf{1}}_P) \ast \mu_{\delta^{k+2}}\subseteq ({\operatorname{supp}}(f {\bf{1}}_P))\cdot({\operatorname{supp}}\mu_{\delta^{k+2}})\subseteq z_P\cdot B(0, \delta^{k+2})\cdot B(0, \delta^{k+2})$$ which is contained in $ B(z_P, \delta^{k+1}) \subseteq Q $ by the definition of $ V_Q$. Observe that the above argument fails if we use balls defined by the standard right invariant metric. The lemma is proved.
In view of Lemma \[lem:Hn\] it suffices to prove the sparse bound for each $M_{\mathcal{D}^{\alpha}}f=\sup_{Q\in \mathcal{D}^{\alpha}}A_Qf$ for $\alpha=1,2,\ldots,N$. Let us fix then $\mathcal{D}=\mathcal{D}^{\alpha}$. We will linearise the supremum. Let $\mathcal{Q}$ be the collection of all dyadic subcubes of $Q_0\in \mathcal{D}$. Let us define $$E_Q:=\big\{x\in Q : A_Qf(x)\ge \frac{1}{2} \sup_{P\in \mathcal{Q}} A_Pf(x)\big\}$$ for $Q\in \mathcal{Q}$. Note that for any $ x \in {\mathbb H}^n $ there exists a $Q \in \mathcal{Q}$ such that $$A_Qf(x)\ge \frac{1}{2} \sup_{P\in \mathcal{Q}} A_Pf(x)$$ and hence $ x \in E_Q.$ If we define $B_Q=E_Q\setminus \cup_{Q'\supseteq Q}E_{Q'}$, then $\{B_Q: Q\in \mathcal{Q}\}$ are disjoint and also, $\cup_{Q\in \mathcal{Q}}B_Q=\cup_{Q\in \mathcal{Q}}E_Q$. Let $f,g>0$. Then $$\begin{aligned}
\notag\langle\sup_{Q\in \mathcal{Q}}A_Qf,g\rangle&=\sum_{Q\in \mathcal{Q}}\int_{B_Q}\sup_{P\in \mathcal{Q}}A_Pf(x)g(x)\,dx\\
\notag&\le 2 \sum_{Q\in \mathcal{Q}}\int_{B_Q}A_Qf(x)g(x)\,dx\\
\notag&\le 2 \sum_{Q\in \mathcal{Q}}\int_{{\mathbb H}^n}A_Qf(x)g(x)\mathbf{1}_{B_Q}(x)\,dx\\
&\le 2 \sum_{Q\in \mathcal{Q}}\langle A_Qf,g\mathbf{1}_{B_Q}\rangle.\end{aligned}$$ Defining $g_Q=g\mathbf{1}_{B_Q}$ we will deal with $\sum_{Q\in \mathcal{Q}}\langle A_Qf,g_Q\rangle$.
\[lem:key\] Let $1<p,q<\infty$ be such that $\big(\frac1p,\frac1q\big)$ in the interior of the triangle joining the points $(0,1), (1,0)$ and $\big(\frac{n}{n+1},\frac{n}{n+1}\big)$. Let $f=\mathbf{1}_{F}$ and let $ g $ be any bounded function supported in $ Q_0$. Let $C_0>1$ be a constant and let $\mathcal{Q}$ be a collection of dyadic subcubes of $Q_0\in \mathcal{D}$ for which the following holds $$\label{eq:keyCondf1}
\sup_{Q'\in \mathcal{Q}}\sup_{Q: Q'\subset Q\subset Q_0}\frac{\langle f\rangle_{Q,p}}{\langle f\rangle_{Q_0,p}}<C_0.$$ Then there holds $$\sum_{Q\in \mathcal{Q}}\langle A_Qf,g_Q\rangle\lesssim |Q_0|\langle f\rangle_{Q_0,p}\langle g\rangle_{Q_0,q}.$$
We perform a Calderón–Zygmund decomposition of $f$ at height $2C_0\langle f \rangle_{Q_0,p}$. Let us denote by $\mathcal{B}$ the resulting collection of (maximal) dyadic subcubes of $Q_0$ so that $$\label{eq:stopp}
\langle f\rangle_{Q,p}>2C_0\langle f \rangle_{Q_0,p}.$$ Set $f=g_1+b_1$, where $$b_1=\sum_{P\in \mathcal{B}}(f-\langle f\rangle_{P})\mathbf{1}_P=\sum_{k=s_0+1}^\infty \sum_{P\in \mathcal{B}(k)}(f-\langle f\rangle_{P})\mathbf{1}_P=:\sum_{k=s_0+1}^{\infty}B_{1,k},$$ where $\ell(Q_0)=\delta^{s_0}$ and $\mathcal{B}(k)=\{P\in \mathcal{B}: \ell(P)=\delta^k\}$. Now $$\big|\sum_{Q\in \mathcal{Q}}\langle A_Qf,g_Q\rangle\big|\le \sum_{Q\in \mathcal{Q}}|\langle A_Qg_1,g_Q\rangle|+\sum_{Q\in \mathcal{Q}}|\langle A_Qb_1,g_Q\rangle|.$$ Since $g_1$ is a bounded function then $\|A_Qg_1\|_{\infty}<\infty$. Hence $$\sum_{Q\in \mathcal{Q}}|\langle A_Qg_1,g_Q\rangle|\lesssim \sum_{Q\in \mathcal{Q}}\|g\mathbf{1}_{B_Q}\|_1\lesssim |Q_0|.$$ We now make the following useful observation. For all $Q\in \mathcal{Q}$ and $P\in \mathcal{B}$, if $P\cap Q\neq \emptyset$ then $P$ is properly contained in $ Q$. For otherwise, $Q\subseteq P$ and by the assumption on $ \mathcal{Q}$, we get $\langle f\rangle_{P,p}<C_0\langle f \rangle_{Q_0,p}$. But this contradicts the Calderón–Zygmund decomposition since $ \langle f\rangle_{P,p}>2C_0\langle f \rangle_{Q_0,p}$. Therefore, for any $Q\in \mathcal{Q}$ with $\ell(Q)=\delta^s$ we have $$\langle A_Qb_1,g_Q\rangle=\sum_{k>s}\langle A_QB_{1,k},g_Q\rangle=\sum_{k=1}^{\infty}\langle A_QB_{1,s+k},g_Q\rangle$$ and so $$\big|\sum_{Q\in \mathcal{Q}}\langle A_Qb_1,g_Q\rangle\big|\le \sum_{k=1}^{\infty}\sum_{Q\in \mathcal{Q}}|\langle A_QB_{1,s+k},g_Q\rangle|.$$ By making use of the mean zero property of $b_1$, we see that $$\begin{aligned}
|\langle A_QB_{1,s+k},g_Q\rangle|&=|\langle B_{1,s+k},A_Q^*g_Q\rangle|\\
&= \sum_{P\in B(s+k)}\big|\int_P A_Q^*g_Q(x)B_{1,s+k}(x)\,dx\big|\\
&\le \sum_{P\in B(s+k)}\frac{1}{|P|}\Big|\int_P\int_P\big[A_Q^*g_Q(x)-A_Q^*g_Q(x')\big]B_{1,s+k}(x)\,dx\,dx'\Big|.\end{aligned}$$ In the integral with respect to $ x'$ we make the change of variables $x'=xy^{-1}$ and note that $ P^{-1}x \subset P^{-1}P.$ Since $ P \subset B(z_P, 4\delta^{s+k}) = z_P\cdot B(0,4\delta^{s+k})$ it follows that $ P^{-1} \subset B(0, 4\delta^{s+k})z_P^{-1} $ and hence $ P^{-1}P \subset P_0 = B(0, 8\delta^{s+k}) \subset B(0,\delta^{s+k-1})$ (observe that for the above argument it is important that the balls are defined using the left invariant metric). Thus we have $$\begin{aligned}
|\langle A_QB_{1,s+k},g_Q\rangle|&\le \sum_{P\in B(s+k)}\frac{1}{|P|}\Big|\int_{P^{-1}P}\int_P\big[A_Q^*g_Q(x)-\tau_yA_Q^*g_Q(x)\big]B_{1,s+k}(x)\,dx\,dy\Big|\\
&\lesssim \frac{1}{|P_0|}\int_{P_0}\Big|\int_Qg_Q(x)(A_Q-A_Q\tau_{y^{-1}})B_{1,s+k}(x)\,dx\Big|\,dy\\
&\lesssim \frac{1}{|P_0|}\int_{P_0}\Big|\frac{y}{\ell(Q)}\Big|^{\eta}|Q|\langle B_{1,s+k}\mathbf{1}_Q\rangle_{Q,p}\langle g_Q\rangle_{Q,q}\,dy\\
&\lesssim \frac{\delta^{(q+k-1)\eta}}{\delta^{q\eta}}|Q|\langle B_{1,s+k}\mathbf{1}_Q\rangle_{Q,p}\langle g_Q\rangle_{Q,q}\\
&\lesssim \delta^{k\eta}|Q|\langle B_{1,s+k}\mathbf{1}_Q\rangle_{Q,p}\langle g_Q\rangle_{Q,q},\end{aligned}$$ where we used Lemma \[lem:anal23\] in the third inequality.
Now we will prove $$\label{eq:claimII}
\sum_{Q\in \mathcal{Q}}|Q|\langle B_{1,s+k}\mathbf{1}_Q\rangle_{Q,p}\langle g\mathbf{1}_{B_Q}\rangle_{Q,q}\lesssim |Q_0|\langle f\rangle_{Q_0,p}\langle g\rangle_{Q_0,q},$$ for all $k\ge1$ and for all $1<p,q<\infty$ such that $\big(\frac1p,\frac1q\big)$ are in the interior of the triangle joining the points $(0,1), (1,0)$ and $(1,1)$ (including the segment joining $(0,1)$ and $(1,0)$, excluding the endpoints).
Let us fix the integer $k$. From the definition and it follows that we can dominate $$|B_{1,s+k}|\lesssim \langle f\rangle_{Q_0,p}\mathbf{1}_{E_s}+\mathbf{1}_{F_{1,s}},$$ where $E_s=E_{s,k}$ are pairwise disjoint sets in $Q_0$ as $s$ varies, and $F_{1,s}=F_{1,s,k}$ are pairwise disjoint sets in $F_1$. This produces two terms to control. For the first one, we will show that $$\label{eq:first}
\langle f\rangle_{Q_0,p}\sum_{Q\in \mathcal{Q}}|Q|\langle \mathbf{1}_{E_s}\rangle_{Q,p}\langle g\mathbf{1}_{B_Q}\rangle_{Q,q}\lesssim |Q_0|\langle f\rangle_{Q_0,p}\langle g\rangle_{Q_0,q}.$$
First we consider the case when $1/p+1/q=1,$ i.e. $ p =q'$, for $1<p<\infty$. $$\begin{aligned}
\sum_{Q\in \mathcal{Q}}|Q|\langle \mathbf{1}_{E_s}\rangle_{Q,p}\langle g\mathbf{1}_{B_Q}\rangle_{Q,p'}&=\sum_{Q\in \mathcal{Q}}\Big(\int_Q\mathbf{1}_{E_s} \,dx\Big)^{1/p}\Big(\int_{Q}|g(x)|^{p'}\mathbf{1}_{B_Q}\,dx\Big)^{1/p'}\\
& \le \Big(\sum_{Q\in \mathcal{Q}}\int_Q\mathbf{1}_{E_s} \,dx\Big)^{1/p}\Big(\sum_{Q\in \mathcal{Q}}\int_{Q}|g(x)|^{p'}\mathbf{1}_{B_Q}\,dx\Big)^{1/p'}.\end{aligned}$$ On one hand, from the disjointness of $B_Q$, $$\sum_{Q\in \mathcal{Q}}\int_{Q}|g(x)|^{p'}\mathbf{1}_{B_Q}\,dx = \int_{\cup B_Q}|g(x)|^{p'}\,dx\leq \Big( \frac{1}{|Q_0|}\int_{Q_0}|g(x)|^{p'}\,dx \Big) |Q_0| = |Q_0|\langle g\rangle_{Q_0,p'}^{p'}.$$ On the other hand, as $E_s \cap Q $ are disjoing subsets of $ Q_0 $ we finally obtain $$\sum_{Q\in \mathcal{Q}}\int_Q\mathbf{1}_{E_s} \,dx= \sum_{Q\in \mathcal{Q}}|E_s \cap Q| \leq |Q_0|.$$ Thus the required inequality is proved in the case $ 1/p +1/q =1$. In the case $1/p +1/q =1+\tau>1$, set $1/\widetilde{p}=1/p-\tau$. Then, $1/\widetilde{p}+1/q=1$, and $p<\widetilde{p}$, so that $$\langle \mathbf{1}_{E_s}\rangle_{Q,p}\langle g\mathbf{1}_{B_Q}\rangle_{Q,q}\lesssim \langle \mathbf{1}_{E_s}\rangle_{Q,\widetilde{p}}\langle g\mathbf{1}_{B_Q}\rangle_{Q,q}.$$ Then, follows from the previous case since $1/\widetilde{p}+1/q'=1$.
Concerning the second term, we will show that $$\label{eq:second}
\sum_{Q\in \mathcal{Q}}|Q|\langle \mathbf{1}_{F_{1,s}}\rangle_{Q,p}\langle g\mathbf{1}_{B_Q}\rangle_{Q,q}\lesssim |Q_0|\langle f\rangle_{Q_0,p}\langle g\rangle_{Q_0,q}.$$ Again, the inequality holds in the case of $1/p+1/q=1$. For $1/p+1/q=1+\tau>1$, we define $\widetilde{p}$ as above. By using the stopping condition we have then $$\langle \mathbf{1}_{F_{1,s}}\rangle_{Q,p}\langle g\mathbf{1}_{B_Q}\rangle_{Q,q}\lesssim\langle \mathbf{1}_{F_{1}}\rangle_{Q_0}^{\tau}\langle \mathbf{1}_{F_{1,s}}\rangle_{Q,\widetilde{p}}\langle g\mathbf{1}_{B_Q}\rangle_{Q,q}.$$ From this and by using the previous case, since $1/\widetilde{p}+1/q=1$, we can conclude , and therefore . The proof is complete.
Let us proceed to prove Theorem \[thm:sparse\]. We will state it also here, for the sake of the reading.
\[thm:main\] Assume $ n \geq 2$ and fix $ 0 < \delta < \frac{1}{96}.$ Let $ 1 < p, q < \infty $ be such that $ (\frac{1}{p},\frac{1}{q}) $ belongs to the interior of the triangle joining the points $ (0,1), (1,0) $ and $ (\frac{n}{n+1},\frac{n}{n+1}).$ Then for any pair of compactly supported bounded functions $ (f,g) $ there exists a $ (p,q)$-sparse form such that $ \langle M_{\operatorname{lac}}f, g\rangle \leq C \Lambda_{\mathcal{S},p,q}(f,g).$
Fix a dyadic grid $\mathcal{D}$ and consider the maximal function $$M_{\mathcal{D}}f(x)=\sup_{Q\in \mathcal{D}}|A_{Q}f(x)|.$$ We can assume that $f\ge0$ and supported in $Q_0$ so that $A_Qf=0$ for all large enough cubes. According to this, we will therefore prove the sparse bound for the maximal function $$M_{\mathcal{D}\cap Q_0}f(x)=\sup_{Q\in \mathcal{D}}|A_{Q}f(x)|.$$ From this, it follows that $M_{\operatorname{lac}}$ is bounded by the sum of a finite number of sparse forms. But it is known that there exists one universal dominating sparse form (see for instance [@LM Lemma 4.7] and [@CaR Proposition 2.1]). Namely, given $f,g$, there is a constant $C>1$ and sparse family of dyadic cubes $\mathcal{S}_0$ so that $\sup_{\mathcal{S}}\Lambda_{\mathcal{S},p,q}(f,g)\le C \Lambda_{\mathcal{S}_0,p,q}(f,g)$. This fact, proved in the Euclidean setting, is also valid in our case and we will not enter into details. Therefore, the claimed sparse bound holds.
As explained above, by linearising the supremum it is enough to prove the sparse bound for the sum $$\label{eq:linear}
\sum_{Q\in\mathcal{D}\cap Q_0}\langle A_Qf,g\mathbf{1}_{B_Q}\rangle$$ for the collection of pairwise disjoint $B_{Q}\subset Q$ described just before Lemma \[lem:key\].
Given $1<p,q<\infty$ so that the $L^p$ improving and continuity properties of the spherical means hold for $\big(\frac1p,\frac1q\big)$ (i.e., Corollaries \[cor:LpLq\] and \[cor:dilat2\] hold), we have to produce a sparse family $\mathcal{S}$ of subcubes of $Q_0$ such that $$\langle M_{\mathcal{D}\cap Q_0}f, g\rangle \le 2 \sum_{Q\in\mathcal{D}\cap Q_0}\langle A_Qf,g\mathbf{1}_{B_Q}\rangle \le C\sum_{S\in \mathcal{S}}|S|\langle f\rangle_{S,p}\langle g \rangle_{S,q}$$ where for each $S\in \mathcal{S}$, there exists $F_S\subset S$ with $|F_S|\ge \frac12 |S|$.
We first prove when $f$ is the characteristic function of a set $F\subset Q_0$. Consider the collection $\mathcal{E}_{Q_0}$ of maximal children $P\subset Q_0$ for which $$\langle f\rangle_{P,p}>2\langle f\rangle_{Q_0,p}.$$ Let $E_{Q_0}=\cup_{P\in \mathcal{E}_{Q_0}}$. For a suitable choice of $c_n>1$ we can arrange $|E_{Q_0}|<\frac12|Q_0|$. We let $F_{Q_0}=Q_0\setminus E_{Q_0}$ so that $|F_{Q_0}|\ge \frac12|Q_0|$. We define $$\label{eq:zero}
\mathcal{Q}_0=\{Q\in \mathcal{D}\cap Q_0: Q\cap E_{Q_0}=\emptyset\}.$$ Note that when $Q\in \mathcal{Q}_0$ then $\langle f\rangle_{Q,p}\le 2\langle f\rangle_{Q_0,p}$. For otherwise, if $\langle f\rangle_{Q,p}>2\langle f\rangle_{Q_0,p}$ then there exists $P\in \mathcal{E}_{Q_0}$ such that $P\supset Q$, which is a contradiction. For the same reason, if $Q'\in \mathcal{Q}_0$ and $Q'\subset Q\subset Q_0$ then $\langle f\rangle_{Q,p}\le 2\langle f\rangle_{Q_0,p}$. Thus $$\sup_{Q'\in \mathcal{Q}_0}\sup_{Q:Q'\subset Q\subset Q_0}\langle f\rangle_{Q,p}\le 2\langle f\rangle_{Q_0,p}.$$ Note that for any $Q\in \mathcal{D}\cap Q_0$, either $Q\in \mathcal{Q}_0$ or $Q\subset P$ for some $P\in \mathcal{E}_{Q_0}$. Thus $$\sum_{Q\in \mathcal{D}\cap Q_0}\langle A_Q f,g\mathbf{1}_{B_Q}\rangle=\sum_{Q\in Q_0}\langle A_Q f,g\mathbf{1}_{B_Q}\rangle+\sum_{P\in \mathcal{E}_{Q_0}}\sum_{Q\subset P}\langle A_Q f,g\mathbf{1}_{B_Q}\rangle$$ for any $Q\in \mathcal{Q}_0$, $Q\subset F_{Q_0}$ and hence $$\sum_{Q\in \mathcal{Q}_0}\langle A_Q f,g\mathbf{1}_{B_Q}\rangle=\sum_{Q\in \mathcal{Q}_0}\langle A_Q f,g\mathbf{1}_{F_{Q_0}}\mathbf{1}_{B_Q}\rangle.$$ Applying Lemma \[lem:key\] we obtain $$\sum_{Q\in \mathcal{Q}_0}\langle A_Q f,g\mathbf{1}_{B_Q}\rangle\le C|Q_0|\langle f\rangle_{Q_0,p} \langle g\mathbf{1}_{F_{Q_0}}\rangle_{Q_0,q}.$$
Let $\{P_j\}$ be an enumeration of the cubes in $\mathcal{E}_{Q_0}$. Then the second sum above is given by $$\sum_{j=1}^{\infty}\sum_{Q\in P_j\cap \mathcal{D}}\langle A_Q f,g\mathbf{1}_{B_Q}\rangle.$$ For each $j$ we can repeat the above argument recursively. Putting everything together we get a sparse collection $\mathcal{S}$ for which $$\label{eq:toget}
\sum_{Q\in \mathcal{D}\cap Q_0}\langle A_Q f,g\mathbf{1}_{B_Q}\rangle\le C\sum_{S\in \mathcal{S}}|S| |\langle f\rangle_{S,p} \langle g\mathbf{1}_{F_{S}}\rangle_{S,q}.$$ This proves the result when $f=\mathbf{1}_F$. We pause for a moment to remark that we have actually proved a sparse domination stronger than the one stated in the theorem. However, we are not able to prove such a result for general $ f.$
Now we prove the theorem for any bounded $f\ge0$ supported in $Q_0$. We start as in the case of $f=\mathbf{1}_F$ but now we define $\mathcal{Q}_0$ using stopping conditions on both $f$ and $g$. Thus we let $\mathcal{E}_{Q_0}$ stand for the collection of maximal subcubes $P$ of $Q_0$ for which either $\langle f\rangle_{P,p}>2\langle f\rangle_{Q_0,p}$ or $\langle g\rangle_{P,q}>2\langle g\rangle_{Q_0,q}$. As before, we define $E_{Q_0}=\cup_{P\in \mathcal{E}_{Q_0}}$ and $F_{Q_0}=Q_0\setminus E_{Q_0}$ so that $|F_{Q_0}|\ge \frac12|Q_0|$. We let $$\mathcal{Q}_0=\{Q\in \mathcal{D}\cap Q_0: Q\cap E_{Q_0}=\emptyset\}.$$ Then it follows that $$\sup_{Q'\in \mathcal{Q}_0}\sup_{Q:Q'\subset Q\subset Q_0}\langle f\rangle_{Q,p}\le 2\langle f\rangle_{Q_0,p}$$ and $$\sup_{Q'\in \mathcal{Q}_0}\sup_{Q:Q'\subset Q\subset Q_0}\langle g\rangle_{Q,q}\le 2\langle g\rangle_{Q_0,q}.$$ If we can show that $$\label{eq:star}
\sum_{Q\in \mathcal{Q}_0}\langle A_Q f,g\mathbf{1}_{B_Q}\rangle\le C|Q_0|\langle f\rangle_{Q_0,\rho} \langle g\rangle_{Q_0,q}$$ for some $\rho>p$, then we can proceed as in the case of $f=\mathbf{1}_F$ to get the sparse domination $$\langle M_{\mathcal{D}}f,g\rangle\le C\sum_{S\in \mathcal{S}}|S| |\langle f\rangle_{S,\rho} \langle g\rangle_{S,q}.$$
In order to prove $\eqref{eq:star}$ we make use of the sparse domination already proved for $f=\mathbf{1}_F$. Defining $E_m=\{x\in Q_0: 2^m\le f(x)\le 2^{m+1}\}$ and $f_m=f\mathbf{1}_{E_m}$ we have the decomposition $f=\sum_mf_m$ (since $f$ is bounded it follows that $E_m=\emptyset$ for all $m\ge m_0$ for some $m_0\in {\mathbb Z}$). By applying the sparse domination to $\mathbf{1}_{E_m}$ we obtain the following: $$\begin{aligned}
\sum_{Q\in \mathcal{Q}_0}\langle A_Q f_m,g\mathbf{1}_{B_Q}\rangle&\le 2^{m+1}\sum_{Q\in \mathcal{Q}_0}\langle A_Q\mathbf{1}_{E_m},g\mathbf{1}_{B_Q}\rangle\\
&=2^{m+1}\sum_{Q\in \mathcal{Q}_0}\langle A_Q\mathbf{1}_{E_m},g\mathbf{1}_{F_{Q_0}}\mathbf{1}_{B_Q}\rangle\\
&\le2^{m+1}\sum_{Q\in Q_0\cap \mathcal{D}}\langle A_Q\mathbf{1}_{E_m},g\mathbf{1}_{F_{Q_0}}\mathbf{1}_{B_Q}\rangle\\
&\le C2^{m+1}\sum_{S\in \mathcal{S}_m}|S|\langle \mathbf{1}_{E_m}\rangle_{S,p}\langle g\mathbf{1}_{F_{Q_0}}\rangle_{S,q},\end{aligned}$$ where in the last three lines we used that for any $Q\in \mathcal{Q}_0$, $Q\subset F_{Q_0}$, and . In the above sum, $\langle g\mathbf{1}_{F_{Q_0}}\rangle_{S,q}=0$ unless $S\cap F_{Q_0}\neq \emptyset$. If $S\subset F_{Q_0}$ then by the definition of $\mathcal{Q}_0$ in it follows that $S\in \mathcal{Q}_0$ and $$\langle g\mathbf{1}_{F_{Q_0}}\rangle_{S,q}\le \langle g\rangle_{S,q}\le c_n \langle g\rangle_{Q_0,q}.$$ If $S\cap F_{Q_0}\neq \emptyset$ as well as $S\cap E_{Q_0}\neq \emptyset$ then for some $P\in \mathcal{E}_{{Q}_0}$, $P\subset S$. But then by the maximality of $P$ we have $$\langle g\mathbf{1}_{F_{Q_0}}\rangle_{S,q}\le \langle g\rangle_{S,q}\le 2 \langle g\rangle_{Q_0,q}.$$ Using this we obtain $$\sum_{Q\in \mathcal{Q}_0}\langle A_Q f_m,g\mathbf{1}_{B_Q}\rangle\le C2^{m+1}\langle g\rangle_{Q_0,q}\sum_{S\in \mathcal{S}_m}|S|\langle \mathbf{1}_{E_m}\rangle_{S,p}.$$ By Lemma \[lem:carleson\] we get $$\sum_{Q\in \mathcal{Q}_0}\langle A_Q f_m,g\mathbf{1}_{B_Q}\rangle\le C2^{m+1}\langle g\rangle_{Q_0,q}\langle \mathbf{1}_{E_m}\rangle_{Q_0,\rho_1}|Q_0|$$ for some $\rho_1>p$. As $f=\sum_m f_m$ it follows that $$\sum_{Q\in \mathcal{Q}_0}\langle A_Q f,g\mathbf{1}_{B_Q}\rangle\le C\langle g\rangle_{Q_0,q}|Q_0|\sum_m2^m\langle \mathbf{1}_{E_m}\rangle_{Q_0,\rho_1}.$$ We now claim that (see Lemma \[lem:lorentz\] below) $$\label{eq:twostar}
\sum_m2^m\langle \mathbf{1}_{E_m}\rangle_{Q_0,\rho_1}\le C\|f\|_{L^{\rho_1,1}(Q_0,d\mu)}$$ where $L^{\rho_1,1}(Q_0,d\mu)$ stands for the Lorentz space defined on the measure space $(Q_0,d\mu)$, $d\mu=\frac{1}{|Q_0|}dx.$ We also know that on a probability space, the $L^{\rho_1,1}(Q_0,d\mu)$ norm is dominated by the $L^{\rho}(Q_0,d\mu)$ norm for any $\rho>\rho_1$ (Lemma \[lem:proba\]). Using these two results we see that $$\sum_{Q\in \mathcal{Q}_0}\langle A_Q f,g\mathbf{1}_{B_Q}\rangle\le C\langle g\rangle_{Q_0,q}|Q_0|\langle f\rangle_{Q_0,\rho}.$$ Hence is proved and thus completes the proof of Theorem \[thm:main\].
It remains to prove Lemma \[lem:proba\] and the claim . The first one is a well known fact which we include here for the sake of completeness.
\[lem:proba\] On a probability space $ (X,d\mu)$, $L^p(X,d\mu)\subset L^{r,1}(X,d\mu) \text{ for } p>r.$
Recall that the Lorentz spaces $ L^{p,q}(X,d\mu) $ are defined in terms of the Lorentz norms (see [@GM1]) $$\| f\|_{p,q} =
\begin{cases}
\Big(\int_0^{\infty}\big(t^{\frac{1}{p}}f^*(t)\big)^q\frac{dt}{t}\Big)^{\frac{1}{q}} \quad &\text{ if } q<\infty,\\
\sup_{t>0}t^{\frac{1}{p}}f^*(t)\quad &\text{ if } q=\infty,
\end{cases}$$ where $ f^*(t) $ stands for the non-decreasing rearrangement of $ f.$ When $f\in L^p(X,d\mu)$, as $d\mu$ is a probability measure, we know that the distribution function $df(s)$ of $f$ is bounded by $1$ and hence $f^*(t)=0$ for $t\ge 1$. Now $$\|f\|_{L^{r,1}(X,d\mu)}=\int_0^{\infty}t^{\frac{1}{r}-1}f^*(t)\,dt=\int_0^{1}t^{-\frac{1}{r'}}f^*(t)\,dt.$$ By Hölder’s inequality $$\|f\|_{L^{r,1}(X,d\mu)}\le \Big(\int_0^1 t^{-\frac{p'}{r'}} dt\Big)^{1/p'} \Big(\int_0^1f^*(t)^p\,dt\Big)^{1/p} =C_{r,p}\Big(\int_0^1f^*(t)^p\,dt\Big)^{1/p}$$ where $C_{r,p}<\infty$ since $p'<r'$. This proves the claim since $$\Big(\int_0^1f^*(t)^p\,dt\Big)^{\frac1p}=\|f\|_{L^p(X,d\mu)}.$$
The claim is the content of the next lemma.
\[lem:lorentz\] Let $f=\sum_mf_m$, $f_m=f\mathbf{1}_{E_m}$ where $E_m=\{x\in Q_0: 2^m\le |f(x)|\le 2^{m+1}\}.$ We consider the probability measure $ d\mu = |Q_0|^{-1} dx $ on $ X = Q_0.$ Then for any $ r >1 $ we have $$\sum_m2^m\langle \mathbf{1}_{E_m}\rangle_{Q_0,r}\le C\|f\|_{L^{r,1}(Q_0,d\mu)}.$$
We make use the following definition of the Lorentz norm in terms of $df(s)$: $$\|f\|_{L^{r,1}(X,d\mu)}=\int_0^{\infty}df(s)^{\frac{1}{r}}\,ds.$$ As $df(s)$ is a decreasing function of $s$ we have $$\begin{aligned}
\|f\|_{L^{r,1}(X,d\mu)}&= \sum_m\int_{2^m}^{2^{m+1}}df(s)^{\frac{1}{r}}\,ds\\
&\ge \sum_mdf(2^{m})^{\frac1r}(2^{m+1}-2^m)\\
&=\frac12\sum_mdf(2^m)^{\frac1r}2^m.\end{aligned}$$ As $f_m=f\mathbf{1}_{E_m}$, it follows that $
\mu(E_m) =df(2^m)-df(2^{m+1})\le df(2^m)$ and consequently, $$\sum_m \mu(E_m)^{\frac1r}2^m\le \sum_mdf(2^m)^{\frac1r}2^m\le 2\|f\|_{L^{r,1}(X,d\mu)}.$$ This proves the lemma.
In proving Theorem \[thm:main\] we have made use of the following lemma, which is proved in [@Lacey Proposition 2.19]. We include a proof here for the convenience of the reader.
\[lem:carleson\] Let $\mathcal{S}$ be a collection of sparse subcubes of a fixed dyadic cube $Q_0$ and let $1\le s<t<\infty$. Then, for a bounded function $\phi$, $$\sum_{Q\in \mathcal{S}}\langle \phi\rangle_{Q,s}|Q|\lesssim \langle \phi\rangle_{Q_0,t}|Q_0|.$$
By sparsity, $$\begin{aligned}
\sum_{Q\in \mathcal{S}}\langle \phi\rangle_{Q,s}|Q|&=\sum_{Q\in \mathcal{S}}\langle \phi\rangle_{Q,s}|Q|^{1/t+1/t'}\\
&\le \Big(\sum_{Q\in \mathcal{S}}\langle \phi\rangle_{Q,s}^t|Q|\Big)^{1/t} \Big(\sum_{Q\in \mathcal{S}}|Q|\Big)^{1/t'}\\
&\lesssim \Big(\sum_{Q\in \mathcal{S}}\langle |\phi|^s\rangle_{Q}^{t/s}|Q|\Big)^{1/t}|Q_0|^{1/t'}\\
&\lesssim \|\phi\mathbf{1}_{Q_0}\|_t|Q_0|^{1/t'}.\end{aligned}$$
A sharpened sparse domination {#sub:sharpens}
-----------------------------
Although we have stated Theorem \[thm:main\] for a slightly more restricted region $\mathbf{L}_n$, indeed the sparse domination holds for $\big(\frac1p,\frac1q\big)$ in the interior of the triangle $\mathbf{S}_n$ (of course Lemma \[lem:key\] holds also for the enlarged triangle since Lemma \[lem:anal23\] does and so on). This means, in particular, that Theorem \[thm:main\] is true for the closed triangle $\mathbf{L}_n$.
\[thm:mainSH\] Assume $ n \geq 2$ and fix $ 0 < \delta < \frac{1}{96}.$ Let $ 1 < p, q < \infty $ be such that $ (\frac{1}{p},\frac{1}{q}) $ belongs to the interior of the triangle joining the points $ (0,1), (1,0) $ and $ (\frac{3n+1}{3n+4},\frac{3n+1}{3n+4}).$ Then for any pair of compactly supported bounded functions $ (f,g) $ there exists a $ (p,q)$-sparse form such that $ \langle M_{\operatorname{lac}}f, g\rangle \leq C \Lambda_{\mathcal{S},p,q}(f,g).$
Boundedness properties
======================
Consequences inferred from sparse domination are well-known and have been studied in the literature. We refer to [@BC Section 4] for an account of the same. In particular, sparse domination provides unweighted and weighted inequalities for the operators under consideration.
The strong boundedness is a result by now standard, see [@CUMP], also [@Lacey Proposition 6.1]. Our Theorem \[thm:spherical\] follows from Theorem \[thm:mainSH\] (or just Theorem \[thm:sparse\]) and Proposition \[prop:un\].
\[prop:un\] Let $1\le r<s'\le\infty$. Then, $$\Lambda_{r,s}(f,g)\lesssim \|f\|_{L^p}\|g\|_{L^{p'}}, \quad r<p<s'.$$
Once again for the sake of completeness we reproduce the proof which is quite simple: as the collection $ \mathcal{S} $ is sparse, we have $$\Lambda_{r,s}(f,g) \leq C \sum_{S \in \mathcal{S}} \int_{E_S} \langle f\rangle_{S,r} \langle g \rangle_{S,s} \mathbf{1}_{E_S} dx$$ where $ E_S \subset S $ are disjoint with the property that $ |E_S| \geq \eta |S|.$ The above leads to the estimate $$\Lambda_{r,s}(f,g) \leq C \int_{{\mathbb H}^n} \big(\Lambda |f|^r(x)\big)^{1/r} \big(\Lambda |g|^s(x)\big)^{1/s} dx$$ where $ \Lambda h $ stands for the Hardy-Littlewood maximal function of $ h.$ In view of the boundedness of $\Lambda$, an application of Hölder’s inequality completes the proof of the proposition.
A weight $w$ is a nonnegative locally integrable function defined on ${\mathbb H}^n$. Given $1<p<\infty$, the Muckhenhoupt class of weights $A_p$ consists of all $w$ satisfying $$[w]_{A_p}:=\sup_Q \langle w\rangle_{Q}\langle \sigma \rangle_{Q}^{p-1}<\infty,\quad \sigma:= w^{1-p'}$$ where the supremum is taken over all cubes $Q$ in ${\mathbb H}^n$. On the other hand, a weight $w$ is in the reverse Hölder class $\operatorname{RH}_p$, $1\le p<\infty$, if $$[w]_{\operatorname{RH}_p}=\sup_Q\langle w\rangle_Q^{-1}\langle w\rangle_{Q,p}<\infty,$$ again the supremum taken over all cubes in ${\mathbb H}^n$.
The following theorem was shown in [@BFP Section 6].
\[thm:BFP\] Let $1\le p_0<q_0'\le\infty$. Then, $$\Lambda_{p_0,q_0}(f,g)\le \{[w]_{A_{p/p_0}}\cdot[w]_{\operatorname{RH}_{(q_0'/p)'}}\}^{\alpha}\|f\|_{L^p(w)}\|g\|_{L^{p'}(\sigma)}, \quad p_0<p<q_0',$$ with $\alpha=\max\Big\{\frac{1}{p-1},\frac{q_0'-1}{q_0'-p}\Big\}$.
In view of Theorem \[thm:BFP\] and with the sharpened sparse domination in Theorem \[thm:mainSH\] at hand, but restricting ourselves to values of $(1/p,1/q)$ on $\mathbf{L}_n$, we can obtain the following corollary: it provides unprecedented weighted estimates for the lacunary maximal spherical means in ${\mathbb H}^n$.
\[cor:weight\] Let $n\ge2$ and define $$\frac{1}{\phi(1/p_0)}=\begin{cases}1-\frac{1}{np_0}, \quad 0<\frac1p_0\le \frac{n}{n+1},\\
n\Big(1-\frac1p_0\Big), \quad \frac{n}{n+1}<\frac1p_0<1.
\end{cases}$$ Then $M_{\operatorname{lac}}$ is bounded on $L^p(w)$ for $w\in A_{p/p_0}\cap \operatorname{RH}_{(\phi(1/p_0)'/p)'}$ and all $1<p_0<p<(\phi(1/p_0))'$.
Quantitative weighted estimates could have been stated in Corollary \[cor:weight\], because by Theorem \[thm:mainSH\] we have the sparse domination in the closed triangle $\textbf{L}_n$.
A sharpened weighted inequality {#sub:sharpenw}
-------------------------------
Finally, we remark that an enhanced version of Corollary \[cor:weight\], with the range of $(1/p,1/q)$ in the interior of $\mathbf{S}_n$, might be also stated (see [@Lacey Corollary 6.3] and [@BC Corollary 4.2] for similar discussions).
\[cor:weightSH\] Let $n\ge2$ and define $$\frac{1}{\phi(1/p_0)}=\begin{cases}1-\frac{1}{p_0}\frac{3}{3n+1}, \quad 0<\frac1p_0\le \frac{3n+1}{3n+4},\\
\frac{3n+1}{3}\Big(1-\frac1p_0\Big), \quad \frac{3n+1}{3n+4}<\frac1p_0<1.
\end{cases}$$ Then $M_{\operatorname{lac}}$ is bounded on $L^p(w)$ for $w\in A_{p/p_0}\cap \operatorname{RH}_{(\phi(1/p_0)'/p)'}$ and all $1<p_0<p<(\phi(1/p_0))'$.
[**Acknowledgments**]{}
This work was mainly carried out when the first, second and the fourth author were visiting the third author in Bilbao. They wish to thank BCAM in general and Luz Roncal in particular for the warm hospitality they enjoyed during their visit. The last author fondly remembers the daily shots of cortado as well as the changing colours of The Puppy!
All the four authors are thankful to Michael Lacey for answering several queries and offering clarifications regarding his work [@Lacey]. They also acknowledge helpful discussions with David Beltran and Kangwei Li.
[10]{}
D. Beltran and L. Cladek, Sparse bounds for pseudodifferential operators, arXiv:1711.02339, to appear in *J. Anal. Math.*.
F. Bernicot, D. Frey, and S. Petermichl, , 9 (2016), no. 5, 1079–1113.
J. Bourgain, [Averages in the plane over convex curves and maximal operators]{}, *J. Anal. Math.* **47** (1986), 69–85.
C. P. Calderón, [Lacunary spherical means]{}, *Illinois J. Math.* **23** (1979), 476–484.
J. M. Conde–Alonso and G. Rey, [A pointwise estimate for positive dyadic shifts and some applications]{}, *Math. Ann.* **365** (2017), 1111–1135.
M. Cowling, [On Littlewood–Paley–Stein theory]{}, *Rend. Circ. Mat. Palermo (2)* (1981), suppl. 1, 21–55.
D. Cruz–Uribe, J. M. Martell and C. Pérez, [Sharp weighted estimates for classical operators]{}, *Adv. Math.* **229** (2012), 408–441.
G. B. Folland, *Harmonic analysis in phase phase*, Annals of Mathematics Studies, **122**. Princeton University Press, Princeton, NJ, 1989.
L. Grafakos, *Classical Fourier analysis. Third edition.* Graduate Texts in Mathematics, **250**. Springer, New York, 2014.
T. Hytönen and A. Kairema, [Systems of dyadic cubes in a doubling metric space]{}, *Colloq. Math.* **126** (2012), 1–33.
M. T. Lacey, Sparse bounds for spherical maximal functions, arXiv:1702.08594v6 (2017), to appear in *J. Anal. Math.*
M. T. Lacey and D. Mena, [The sparse $T1$ theorem]{}, *Houston J. Math.* **43** (2017), 111–127.
C. Markett, [Mean Cesàro summability of Laguerre expansions and norm estimates with shifted parameter]{}, *Anal. Math.* **8** (1982), no. 1, 19–37.
B. Muckenhoupt, [Mean convergence of Hermite and Laguerre series. II]{} *Trans. Amer. Math. Soc.* **147** (1970), 433–460
D. Müller and A. Seeger, [Singular spherical maximal operators on a class of two step nilpotent Lie groups]{}, *Israel J. Math.* **141** (2004), 315–340.
E. K. Narayanan and S. Thangavelu, [An optimal theorem for the spherical maximal operator on the Heisenberg group]{}, *Israel J. Math.* **144** (2004), 211–219.
A. Nevo and S. Thangavelu, [Pointwise ergodic theorems for radial averages on the Heisenberg group,]{} *Adv. Math.* **127** (1997), 307–339.
A. P. Prudnikov, A. Y. Brychkov, and O. I. Marichev, *Integrals and Series. Vol. 2. Special Functions*, Gordon and Breach Science Publishers, New York, 1990. 2005.
E. M. Stein, Maximal functions. I. Spherical means, *Proc. Nat. Acad. Sci. U.S.A.* **73** (1976), 2174–2175.
E. M. Stein, G. Weiss, *Introduction to Fourier analysis in Euclidean spaces,* Princeton University Press, Princeton, N. J. 1971.
R. S. Strichartz, [Convolutions with kernels having singularities on a sphere]{}, *Trans. Amer. Math. Soc.* **148** (1970), 461–471.
V. S. Sunder, *Operators on Hilbert space*, Texts and Readings in Mathematics, [**71**]{}. Hindustan Book Agency, New Delhi, 2015.
S. Thangavelu, [Spherical means on the Heisenberg group and a restriction theorem for the symplectic Fourier transform]{}, *Rev. Mat. Iberoamericana* **7** (1991), 135–155.
S. Thangavelu, *Lectures on Hermite and Laguerre expansions*, Mathematical Notes [**42**]{}. Princeton University Press, Princeton, NJ, 1993.
S. Thangavelu, *Harmonic Analysis on the Heisenberg group*, Progress in Mathematics [**159**]{}. Birkhäuser, Boston, MA, 1998.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Invertible local transformations of a multipartite system are used to define equivalence classes in the set of entangled states. This classification concerns the entanglement properties of a single copy of the state. Accordingly, we say that two states have the same kind of entanglement if both of them can be obtained from the other by means of local operations and classical communication (LOCC) with nonzero probability. When applied to pure states of a three-qubit system, this approach reveals the existence of two inequivalent kinds of genuine tripartite entanglement, for which the GHZ state and a W state appear as remarkable representatives. In particular, we show that the W state retains maximally bipartite entanglement when any one of the three qubits is traced out. We generalize our results both to the case of higher dimensional subsystems and also to more than three subsystems, for all of which we show that, typically, two randomly chosen pure states cannot be converted into each other by means of LOCC, not even with a small probability of success.'
address: 'Institut für Theoretische Physik, Universität Innsbruck, A-6020 Innsbruck, Austria'
author:
- 'W. Dür, G. Vidal and J. I. Cirac'
title: Three qubits can be entangled in two inequivalent ways
---
[C]{} [H]{}
= 10000
Introduction
============
The understanding of entanglement is at the very heart of Quantum Information Theory (QIT). In recent years, there has been an ongoing effort to characterize qualitatively and quantitatively the entanglement properties of multiparticle systems. A situation of particular interest in QIT consists of several parties that are spatially separated from each other and share a composite system in an entangled state. This setting conditionates the parties —which are typically allowed to communicate through a classical channel— to only act locally on their subsystems. But even restricted to local operations assisted with classical communication (LOCC), the parties can still modify the entanglement properties of the system and in particular they can try to convert an entangled state into another. This possibility leads to natural ways of defining equivalence relations in the set of entangled states —where equivalent states are then said to contain the same kind of entanglement—, also of establishing hierarchies between the resulting classes.
For instance, we could agree in identifying any two states which can be obtained from each other with certainty by means of LOCC. Clearly, this criterion is interesting in QIT because the parties can use these two states indistintively for exactly the same tasks. It is a celebrated result [@Be95] that, when applied to many copies of a state, this criterion leads to identifying all bipartite pure-state entanglement with that of the EPR state $1/\sqrt{2}({\mbox{$| 00 \rangle$}}+{\mbox{$| 11 \rangle$}}) $[@Ei35]. That is, the entanglement of any pure state ${\mbox{$| \psi \rangle$}}_{AB}$ is asymptotically equivalent, under deterministic LOCC, to that of the EPR state, the entropy of entanglement $E(\psi_{AB})$ —the entropy of the reduced density matrix of either system $A$ or $B$— quantifying the amount of EPR entanglement contained asymptotically in ${\mbox{$| \psi \rangle$}}_{AB}$. In contrast, recent contributions have shown that in systems shared by three or more parties there are several inequivalent forms of entanglement under asymptotic LOCC [@asy; @Be99].
This paper is essentially concerned with the entanglement properties of a single copy of a state, and thus asymptotic results do not apply here. For single copies it is known that two pure states ${\mbox{$| \psi \rangle$}}$ and ${\mbox{$| \phi \rangle$}}$ can be obtained with certainty from each other by means of LOCC if and only if they are related by local unitaries LU [@Vi00J; @Be99]. But even in the simplest bipartite systems, ${\mbox{$| \psi \rangle$}}$ and ${\mbox{$| \phi \rangle$}}$ are typically not related by LU, and continuous parameters are needed to label all equivalence classes [@Li97; @Sc; @Su00; @Ca00; @Ke99]. That is, one has to deal with infinitely many kinds of entanglement. In this context an alternative, simpler classification would be advisable.
One such classification is possible if we just demand that the conversion of the states is through stochastic local operations and classical communication (SLOCC) [@Be99]; that is, through LOCC but without imposing that it has to be achieved with certainty. In that case we can establish an equivalence relation stating that two states ${\mbox{$| \psi \rangle$}}$ and ${\mbox{$| \phi \rangle$}}$ are equivalent if the parties have a non-vanishing probability of success when trying to convert ${\mbox{$| \psi \rangle$}}$ into ${\mbox{$| \phi \rangle$}}$, and also ${\mbox{$| \phi \rangle$}}$ into ${\mbox{$| \psi \rangle$}}$ [@comment]. This relation has been termed stochastic equivalence in Ref. [@Be99]. Their equivalence under SLOCC indicates that both states are again suited to implement the same tasks of QIT, although this time the probability of a successful performance of the task may differ from ${\mbox{$| \phi \rangle$}}$ to ${\mbox{$| \psi \rangle$}}$. Notice in addition that since LU are a particular case of SLOCC, states equivalent under LU are also equivalent under SLOCC, the new classification being a coarse graining of the previous one.
The main aim of this work is to identify and characterize all possible kinds of pure-state entanglement of three qubits under SLOCC. Unentangled states, and also those which are product in one party while entangled with respect to the remaining two, appear as expected, trivial cases. More surprising is the fact that there are two different kinds of genuine tripartite entanglement. Indeed, we will show that any (non-trivial) tripartite entangled state can be converted, by means of SLOCC, into one of two standard forms, namely either the GHZ state [@Gr89] =1/(|000+|111) \[GHZ\], or else a second state |W=1/(|001+|010+|100) \[W\], and that this splits the set of genuinely trifold entangled states into two sets which are unrelated under LOCC. That is, we will see that if ${\mbox{$| \psi \rangle$}}$ can be converted into the state ${\mbox{$| GHZ \rangle$}}$ in (\[GHZ\]) and ${\mbox{$| \phi \rangle$}}$ can be converted into the state ${\mbox{$| W \rangle$}}$ in (\[W\]), then it is not possible to transform, not even with only a very small probability of success, ${\mbox{$| \psi \rangle$}}$ into ${\mbox{$| \phi \rangle$}}$ nor the other way round.
The previous result is based on the fact that, unlike the GHZ state, not all entangled states of three qubits can be expressed as a linear combination of only two product states. Remarkably enough, the inequivalence under SLOCC of the states ${\mbox{$| GHZ \rangle$}}$ and ${\mbox{$| W \rangle$}}$ can alternatively be shown from the fact that the 3-tangle (residual tangle), a measure of tripartite correlations introduced by Coffman et. al. [@Wo99], does not increase on average under LOCC, as we will prove here.
We will then move to the second main goal of this work, namely the analysis of the state ${\mbox{$| W \rangle$}}$. It can not be obtained from a state ${\mbox{$| GHZ \rangle$}}$ by means of LOCC and thus one could expect, in principle, that it has some interesting, characteristic properties. Recall that in several aspects the GHZ state can be regarded as the maximally entangled state of three qubits. However, if one of the three qubits is traced out, the remaining state is completely unentangled. Thus, the entanglement properties of the state ${\mbox{$| GHZ \rangle$}}$ are very fragile under particle losses. We will prove that, oppositely, the entanglement of ${\mbox{$| W \rangle$}}$ is maximally robust under disposal of any one of the three qubits, in the sense that the remaining reduced density matrices[^1] $\rho_{AB}$, $\rho_{BC}$ and $\rho_{AC}$ retain, according to several criteria, the greatest possible amount of entanglement, compared to any other state of three qubits, either pure or mixed.
We will finally analyze entanglement under SLOCC in more general multipartite systems. We will show that, for most of these systems, there is typically no chance at all to transform locally a given state into some other if they are chosen randomly, because the space of entangled pure states depends on more parameters than those that can be modified by acting locally on the subsystems.
The paper is organized as follows. In section II we characterize mathematically the equivalence relation established by stochastic conversions under LOCC, and illustrate its performance by applying it to the well-known bipartite case. In section III we move to consider a system of three qubits, for which we prove the existence of 6 classes of states under SLOCC —including the 2 genuinely tripartite ones—. Section IV is devoted to study the endurance of the entanglement of the state ${\mbox{$| W \rangle$}}$ against particle losses. In section V more general multipartite systems are considered. Section VI contains some conclusions. Finally, appendix A to C prove, respectively, some needed results related to SLOCC, the monotonicity of the 3-tangle under LOCC and the fact that ${\mbox{$| W \rangle$}}$ retains optimally bipartite entanglement when one qubit is traced out.
Kinds of entanglement under Stochastic LOCC
===========================================
In this work we define as equivalent the entanglement of two states ${\mbox{$| \psi \rangle$}}$ and ${\mbox{$| \phi \rangle$}}$ of a multipartite system iff local protocols exist that allow the parties to convert each of the two states into the other one with some a priori probability of success. In this approach, we follow the definition for stochastic equivalence as given in [@Be99][^2]. The underlying motivation for this definition is that, if the entanglement of is ${\mbox{$| \psi \rangle$}}$ and ${\mbox{$| \phi \rangle$}}$ is equivalent, then the two states can be used to perform the same tasks, although the probability of a successful performance of the task may depend on the state that is being used.
Invertible local operators
--------------------------
Sensible enough, this classification would remain useless if in practice we would not be able to find out which states are related by SLOCC. Let us recall that, all in all, no practical criterion is known so far that determines whether a generic transformation can be implemented by means of LOCC. However, we can think of any local protocol as a series of rounds of operations, where in each round a given party manipulates locally its subsystem and communicates classically the result of its operation (if it included a measurement) to the rest of parties. Subsequent operations can be made dependent on previous results and the protocol splits into several branches. This picture is useful because for our purposes we need only focus on one of these branches. Suppose that state ${\mbox{$| \psi \rangle$}}$ can be locally converted into state ${\mbox{$| \phi \rangle$}}$ with non-zero probability. This means that at least one branch of the protocol does the job. Since we are concerned only with pure states we can always characterize mathematically this branch as an operator which factors out as the tensor product of a local operator for each party. For instance, in a three-qubit case we would have that ${\mbox{$| \psi \rangle$}}$ can be locally converted into ${\mbox{$| \phi \rangle$}}$ with some finite probability iff an operator $A\otimes B \otimes C$ exists such that = AB C , \[phipsi\] where operator $A$ contains contributions coming from any round in which party A acted on its subsystem, and similarly for operators $B$ and $C$ [^3]. Carrying on with the 3-qubit example, let us now consider for simplicity that both states ${\mbox{$| \psi \rangle$}}$ and ${\mbox{$| \phi \rangle$}}$ have rank 2 reduced density matrices $\rho_A \equiv $ tr$_{BC}(|\psi\rangle\langle\psi|), \rho_B$ and $\rho_C$. Then clearly the rank of operators $A$, $B$ and $C$ need to be 2 (see appendix A). In other words, each of these operators is necessarily invertible, and in particular = A\^[-1]{}B\^[-1]{}C\^[-1]{} . We see thus that, under the assumption of maximal rank for the reduced density matrices, two-way convertibility implies the existence of invertible operators $A$, $B$ and $C$ as in (\[phipsi\]) \[actually, one-way convertibility alone has already implied that an invertible local operator (ILO) $A\otimes B\otimes C$ exists\]. Obviously, the converse also holds, namely that if an ILO $A\otimes B\otimes C$ exists then for each direction of the conversion a local protocol can be build that succeeds with non-zero probability. As explained in appendix A in detail, we can get rid of the previous assumption on the ranks and announce with generality,
[**Result:**]{} States ${\mbox{$| \psi \rangle$}}$ and ${\mbox{$| \phi \rangle$}}$ are equivalent under stochastic local operations and classical communication —SLOCC— iff an invertible local operator —ILO— relating them \[as in, for instance, equation (\[phipsi\])\] exists.
Bipartite entanglement under SLOCC
----------------------------------
What does this classification implies in the well-known case [@Lo97; @Vi99; @Ni99] of bipartite systems? Since LU are included in SLOCC, we can take the Schmidt decomposition of a pure state ${\mbox{$| \psi \rangle$}}\in\C^n\otimes\C^m$, $n\leq m$, as the starting point for our analysis. Thus, \_[i=1]{}\^[n\_]{} = U\_AU\_B; \_i >0, n\_ n,\[Schmidt1\] where $U_A$ and $U_B$ are some proper local unitaries, the coefficients $\lambda_i$ decrease with $i$, and $n_{\psi}$ is the number of non-vanishing terms in the Schmidt decomposition. Clearly, the ILO (\_[i=1]{}\^[n\_]{} + \_[i=n\_+1]{}\^n )1\_B transforms (\[Schmidt1\]) into a maximally entangled state \_i\^[n\_]{} , \[maxi\] which depends only on the Schmidt number $n_{\psi}$. Since SLOCC cannot modify the rank of the reduced density matrices $\rho_A$ and $\rho_B$, which is given by $n_{\psi}$, we conclude that in $\C^n\otimes\C^m$, $n\leq m$, there are $n$ different kinds of entangled states, corresponding to $n$ different classes under SLOCC. Each of these classes is characterized by a given Schmidt number, and we can choose as their representatives the state (\[maxi\]) with $n_{\psi}=1,...,n$. Clearly $n_{\psi}=1$ corresponds to states that are less entangled than the rest (they are, after all, unentangled). This hierarchical relation can be extended to the rest of classes by noting that none-invertible local operators can project out some of the Schmidt terms and thus diminish the Schmidt number of a state. Therefore the state ${\mbox{$| \psi \rangle$}}$ can be locally converted into the state ${\mbox{$| \phi \rangle$}}$ with some finite probability iff $n_{\psi}\geq n_{\phi}$, or in terms of kinds of entanglement, we can say that the entanglement of the class characterized by a given Schmidt number is more powerful than that of a class with a smaller Schmidt number.
For later reference we also note that in a two-qubit system, $\H=\C^2\otimes\C^2$, we can write any state, after using a convenient LU, uniquely as = c\_ + s\_ ; c\_ s\_0, where $c_{\delta}, s_{\delta}$ stand for cos$\delta$ and sin$\delta$. This is either a product (unentangled) state ${\mbox{$| \psi_{A-B} \rangle$}}={\mbox{$| 0 \rangle$}}\otimes {\mbox{$| 0 \rangle$}}$ for $c_{\delta}=1$ or else an entangled state that can be converted into the EPR state, (+), \[EPR\] with probability $p=E_{2}(\psi)$, where $E_2(\psi)\equiv \lambda_2$ is the entanglement monotone that provides a quantitative description of the non-local resources contained in a single copy of a two-qubit pure state [@monotones]. Any state ${\mbox{$| \psi \rangle$}}$ can be obtained from state (\[EPR\]) with certainty, this contributing to the fact that the EPR state is considered the maximally entangled state of two qubits.
Entanglement of pure states of three qubits
===========================================
In this section we analyze a system of three qubits. We show that SLOCC split the set of pure states into 6 inequivalent classes, which further structure themselves into a three-grade hierarchy when non-invertible local operations are used to relate them. At the top of the hierarchy we find two inequivalent classes of true tripartite entanglement, which we name GHZ-class and W-class after our choice of corresponding representatives. The three possible classes of bipartite entanglement are accessible (with some non-vanishing probability) from [*any*]{} state of the W and GHZ classes by means of a non-invertible local operator. Finally, at the bottom of the hierarchy we find non-entangled states.
The ranks r$(\rho_A)$, r$(\rho_B)$ and r$(\rho_C)$ of the reduced density matrices, together with the range $R(\rho_{BC})$ of $\rho_{BC}$, will be the main mathematical tools used through the first part of this section. By analysing them we will be able to make an exhaustive classification of three-qubit entanglement. Later on we will rephrase some of these results in terms of well-known measures of entanglement. In particular, we will see that the existence of two inequivalent kinds of true tripartite entanglement under SLOCC is very much related to the fact that the 3-tangle, a measure of tripartite entanglement introduced in [@Wo99], is an entanglement monotone (see appendix B).
At the end of the section also a practical way to identify the class an arbitrary state belongs to will be discussed.
Non-entangled states and bipartite entanglement.
------------------------------------------------
If at least one of the local ranks r$(\rho_A)$, r$(\rho_B)$ or r$(\rho_C)$ is 1, then the pure state of the three qubits factors out as the tensor product of two pure states, and this implies that at least one of the qubits is uncorrelated with the other two. SLOCC distinguish states with this feature depending on which qubits are uncorrelated from the rest.
[**Class A-B-C (product states)**]{}\
This class corresponds to states with $r(\rho_A)=r(\rho_B)=r(\rho_C)=1$. They can be taken, after using some convenient LU, into the form = , where we have already relaxed the notation for ${\mbox{$| 0 \rangle$}}\otimes{\mbox{$| 0 \rangle$}}\otimes{\mbox{$| 0 \rangle$}}$.
[**Classes A-BC, AB-C and C-AB\
(bipartite entanglement)**]{}\
These three classes of states contain only bipartite entanglement between two of the qubits, one of the reduced density matrices having rank 1 and the other two having rank 2. For example, the states in class $A-BC$ possess entanglement between the systems $B$ and $C$ ($r(\rho_B)=r(\rho_C)=2$) and are product with respect to system $A$ ($r(\rho_A)=1$). LU allow us to write uniquely states of the class $A-BC$ as |\_[A-BC]{}=|0(c\_|0|0+s\_|1|1), c\_ s\_>0, and similary for $|\psi_{B-AC}\rangle$ and $|\psi_{C-AB}\rangle$. We choose the maximally entangled state |0(|0|0+|1|1) \[repreA-BC\] as representative of the class $A-BC$. Any other state within this class can be obtained from (\[repreA-BC\]) with certainty by means of LOCC.
The proof that these four marginal classes are inequivalent under SLOCC is very simple. We only need to recall that the local ranks are invariant under ILO (see appendix A). In what follows we will analyze the more interesting case of $r(\rho_{\kappa})=2,~\kappa = A,B,C$. To see that there are two inequivalent classes fulfilling this condition we will have to study possible product decompositions of pure states.
True three-qubit entanglement.
------------------------------
There turns out to be a close connection between convertibility under SLOCC and the way entangled states can be expressed minimally as a linear combination of product states. For instance, as we shall prove later on, the GHZ and W states have a different number of terms in their minimal product decompositions (\[GHZ\]) and (\[W\]), namely 2 and 3 product terms respectively, and this readily implies that there is no way to convert one state into the other by means of an ILO $A\otimes B\otimes C$. Indeed, let us consider, e.g., the most general pure state that can be obtained reversibly from a ${\mbox{$| GHZ \rangle$}}$. It reads ABC=( +), \[new\] where ${\mbox{$| A0 \rangle$}}$ and ${\mbox{$| A1 \rangle$}}$ are linearly independent vectors (since $A$ is invertible) and similarly for the other two qubits. That is, the minimal number of terms in a product decomposition for the state (\[new\]) is also 2. Actually, we have that also for a general multipartite system,
[**Observation:**]{} The minimal number of product terms for any given state remains unchanged under SLOCC.
This simple observation tells us already that in three qubits there are at least two inequivalent kinds of genuine tripartite entanglement under SLOCC, that of ${\mbox{$| GHZ \rangle$}}$ and that of ${\mbox{$| W \rangle$}}$.
However, we still have to prove that the state ${\mbox{$| W \rangle$}}$ cannot be expressed as a linear combination of just two product vectors. In order to complete our classification we also have to show that any pure state of three qubits with maximal local ranks can be reversibly converted into either the state ${\mbox{$| GHZ \rangle$}}$ or the state ${\mbox{$| W \rangle$}}$. We start with an obvious lemma regarding product decompositions:
[**Lemma:**]{} Let $\sum_{i=1}^l {\mbox{$| e_i \rangle$}}{\mbox{$| f_i \rangle$}}$ be a product decomposition for the state ${\mbox{$| \eta \rangle$}}\in\H_{E}\otimes\H_{F}$. Then the set of states $\{{\mbox{$| e_i \rangle$}}\}_{i=1}^l$ span the range of $\rho_E\equiv$Tr$_F {{\mbox{$| \eta \rangle$}}\!{\mbox{$\langle \eta |$}}}$.
[**Proof:**]{} We have that $\rho_E = \sum_{i,j=1}^l {\mbox{$\langle f_i | f_j \rangle$}}
{\mbox{$| e_j \rangle$}}{\mbox{$\langle e_i |$}}$. On the other hand ${\mbox{$| \nu \rangle$}}$ is in the range of $\rho_E$ iff a state ${\mbox{$| \mu \rangle$}}$ exists such that ${\mbox{$| \nu \rangle$}}=\rho_E{\mbox{$| \mu \rangle$}}$, that is ${\mbox{$| \nu \rangle$}} = \sum_{i,j=1}^l {\mbox{$\langle f_i | f_j \rangle$}}{\mbox{$\langle e_i | \mu \rangle$}} {\mbox{$| e_j \rangle$}}$. $\Box$
In particular, $r(\rho_A)=2$ implies that at least two product terms are needed to expand ${\mbox{$| \psi \rangle$}}\in \C^2\otimes\C^2\otimes\C^2$. Let us suppose that a product decomposition with only two terms is possible, namely = + . \[2deco\] Then, also according to the previous lemma, ${\mbox{$| b_1 \rangle$}}{\mbox{$| c_1 \rangle$}}$ and ${\mbox{$| b_2 \rangle$}}{\mbox{$| c_2 \rangle$}}$ have to span the range of $\rho_{BC}$, R$(\rho_{BC})$.
But R$(\rho_{BC})$ is a two dimensional subspace of $\C^2\otimes\C^2$. Therefore it always contains either only one or only two product states [@STV] \[unless R$(\rho_{BC})$ was supported in $\C\otimes\C^2$ or $\C^2\otimes\C$, but we already excluded this possibility because we are considering r$(\rho_B)=$r$(\rho_C)=2$\]. Notice that a two-term decomposition (\[2deco\]) requires that R$(\rho_{BC})$ contains at least two product vectors. Only one product vector in R$(\rho_{BC})$, and thus the impossibility of decomposition (\[2deco\]), is going to be precisely the trait of the states in the W-class.
[**GHZ-class**]{}\
Let us suppose first that R$(\rho_{BC})$ contains two product vectors, ${\mbox{$| b_1 \rangle$}}{\mbox{$| c_1 \rangle$}}$ and ${\mbox{$| b_2 \rangle$}}{\mbox{$| c_2 \rangle$}}$. Then decomposition (\[2deco\]) is possible, and actually unique, with ${\mbox{$| a_i \rangle$}}= {\mbox{$\langle \xi_i | \psi \rangle$}}$, $i=1,2$, where ${\mbox{$| \xi_i \rangle$}}$ are the two vectors supported in $R(\rho_{BC})$ that are biorthonormal to the ${\mbox{$| b_i \rangle$}}{\mbox{$| c_i \rangle$}}$. In this case we can use LU in order to take ${\mbox{$| \psi \rangle$}}$ into the useful standard product form (see also [@Ac00]) |\_[GHZ]{}=(c\_|0|0|0+s\_e\^[i]{}|\_A|\_B|\_C), \[GHZclass\] where =c\_+s\_\
=c\_+s\_\
=c\_+s\_ and $K=(1+2 c_\delta s_\delta c_\alpha c_\beta c_\gamma c_\varphi)^{-1} \in
(1/2,\infty)$ is a normalization factor. The ranges for the five parameters are $\delta \in (0,\pi/4], \alpha,\beta,\gamma \in (0,\pi/2]$ and $\varphi \in
[0,2\pi)$.
All these states are in the same equivalence class as the ${\mbox{$| GHZ \rangle$}}$ (\[GHZ\]) under SLOCC. Indeed, the ILO (
[ll]{} c\_& s\_c\_e\^[i]{}\
0 & s\_s\_e\^[i ]{}
) (
[ll]{} 1 & c\_\
0 & s\_
) (
[ll]{} 1 & c\_\
0 & s\_
), applied to ${\mbox{$| GHZ \rangle$}}$ produces the state (\[GHZclass\]).
The GHZ state is a remarkable representative of this class. It is maximally entangled in several senses [@Gi98]. For instance, it maximally violates Bell-type inequalities, the mutual information of measurement outcomes is maximal, it is maximally stable against (white) noise and one can locally obtain from a GHZ state with unit propability an EPR state shared between any two of the three parties. Another relevant feature is that when any one of the three qubits is traced out, the remaining two are in a separable —and therefore unentangled— state.
[**W-class**]{}\
Let us move to analyze the case where R$(\rho_{BC})$ contains only one product vector. We already argued that decomposition (\[2deco\]) is now not possible. Instead we can (uniquely) write = + , \[const\] where ${\mbox{$| \phi_{BC} \rangle$}}$ is the vector of $R(\rho_{BC})$ which is orthogonal to ${\mbox{$| b_1 \rangle$}}{\mbox{$| c_1 \rangle$}}$, and ${\mbox{$| a_1 \rangle$}}$ and ${\mbox{$| a_2 \rangle$}}$ are given by ${\mbox{$\langle b_1|\langle c_1 | \psi \rangle$}}$ and ${\mbox{$\langle \phi_{BC} | \psi \rangle$}}$. By means of LU (\[const\]) can be always rewritten as = (+ )&&\
+&&( + ). \[Wclass2\] Indeed, we first take ${\mbox{$| b_1 \rangle$}}{\mbox{$| c_1 \rangle$}}$ into ${\mbox{$| 0 \rangle$}}{\mbox{$| 0 \rangle$}}$. Then, since ${\mbox{$| \phi_{BC} \rangle$}}$ has been chosen orthogonal to ${\mbox{$| b_1 \rangle$}}{\mbox{$| c_1 \rangle$}}$, it must become $x{\mbox{$| 01 \rangle$}} +y{\mbox{$| 10 \rangle$}}+z {\mbox{$| 11 \rangle$}}$. By requiring that a linear combination of these two vectors has no second product vector we obtain that $z=0$ [@product]. In addittion the coefficients $ \sqrt{a}\equiv x,~\sqrt{b}\equiv
y,~\sqrt{c}$ and $\sqrt{d}$ can be made positive by absorbing the three relative phases into the definition of state ${\mbox{$| 1 \rangle$}}$ of subsystems $A$, $B$ and $C$. Thus case (i) has been taken into the form (\[Wclass2\]) by just using LU. Before we showed that 2 terms could not suffice in a product decomposition of the state. Now we see that 3 product terms always do the job, for instance $(\sqrt{c}{\mbox{$| 1 \rangle$}}+ \sqrt{d}{\mbox{$| 0 \rangle$}}){\mbox{$| 00 \rangle$}}$, $\sqrt{a}{\mbox{$| 0 \rangle$}}{\mbox{$| 01 \rangle$}}$ and $\sqrt{b}{\mbox{$| 0 \rangle$}}{\mbox{$| 10 \rangle$}}$ once we took the original state into the standard, unique form = + + + , \[Wclass\] where $a, b, c >0$, and $d \equiv 1- (a + b+ c) \geq 0$.
The parties can locally obtain the state (\[Wclass\]) from the state ${\mbox{$| W \rangle$}}$ in (\[W\]), which we choose as a representative of the class —and whose study we postpone for later on—, by application of an ILO of the form (
[ll]{} &\
0 &
) (
[ll]{} & 0\
0 &
) (
[ll]{} 1 & 0\
0 & 1
).
Before moving to relate these classes by means of non–invertible local operators, we note that states within the GHZ-class and the W-class depend, respectively, on 5 and 3 parameters that cannot be changed by means of LU. Previous works [@Li97; @Sc; @Ac00; @Ca00] have shown that a generic state of three qubits depends, up to LU, on 5 parameters. This means that states tipically belong to the GHZ-class, or equivalently, that a [*generic*]{} pure state of three qubits can be locally transformed into a GHZ with finite probability of success (see also [@Co00]). The W-class is of zero measure compared to the GHZ-class. This does not mean, however, that it is irrelevant. In a similar way as separable mixed states are not of zero measure with respect to entangled states, even though product states are, it is in principle conceivable that mixed states having only W-class entanglement are also not of zero measure in the set of mixed states.
Relating SLOCC–classes by means of non–invertible operators
-----------------------------------------------------------
In this subsection, we investigate the hierarchical relation of the 6 SLOCC-equivalence classes under non–invertible operators, i.e. under general LOCC.
A non–invertible local operator transforms ${\mbox{$| \psi \rangle$}}$ into ${\mbox{$| \phi \rangle$}}$ according to (\[phipsi\]), but with at least one of the local operators $A$, $B$ and $C$ having rank 1. This means that the local ranks of the pure states can be diminished. For instance, if the initial state ${\mbox{$| \psi \rangle$}}$ belongs either to the GHZ or W class, then a non-invertible operator will diminish at least one of the local ranks. That is, ${\mbox{$| \phi \rangle$}}$ belongs necessarily to one of the bipartite classes $\kappa-\mu\nu~ (\kappa\not=\mu\not=\nu \in\{A,B,C\})$ or else is a product state $A-B-C$.
Thus we have that the classes GHZ and W are also inequivalent even under most general LOCC, whereas e.g. a measurement of the projector $P=|+\rangle\langle+|$ with $|+\rangle=1/\sqrt{2}(|0\rangle+|1\rangle)$ in party $A$ maps states within the classes $W$ (\[Wclass\]) and $GHZ$ (\[GHZclass\]) to states within the class $A-BC$. In a similar way, non–invertible local operators (local, standard measurements) can convert states within one of the classes $\kappa-\mu\nu$ to states within the class $A-B-C$. Note that in all cases described above, the inverse transformations, e.g. from the class $A-B-C$ to one of the classes $\kappa-\mu\nu$ are impossible as they would imply an increase of the rank of at least one of the reduced density operators $\rho_A,\rho_B,\rho_C$. These results are summarized in Fig. \[Fig1\].
Measures of entanglement and classes under SLOCC
------------------------------------------------
Several measures have been introduced so far in the literature in order to quantify entanglement. Although this section is mainly concerned with qualitative aspects of multipartite quantum correlations, we would like to relate some of these measures, namely some bipartite ones and the tripartite 3-tangle [@Wo99](see appendix B), to our classification. Remarkably, the existence of two kinds of genuine tripartite entanglement in a three-qubit system, as well as the inequivalence between bipartite and tripartite entanglement, can be easily understood from the non-increasing character of these measures under LOCC. In addittion, the 3-tangle allows for a systematic and practical identification of which class under SLOCC any pure state belongs to.
For each $\kappa=A,B$ and $C$ we can regard the three-qubit system as a bipartite system, with qubit $\kappa$, say $A$ for concreteness, being one part of the system and the remaining two qubits, $B$ and $C$, being the other. Correspondingly, a state ${\mbox{$| \psi \rangle$}}$ of the three qubits can be viewed as a bipartite state ${\mbox{$| \psi_{A(BC)} \rangle$}}$. For bipartite states several measures are known, which are entanglement monotones [@Vi00J]; that is, which cannot be increased, on average, under LOCC. For instance, we already mentioned the entropy of entanglement $E(\psi)$ for asymptotic conversions –given by the entropy $S_A$ of the eigenvalues of $\rho_A$— and the monotone $E_2(\psi)$ for the single copy case —which is given by the smallest eigenvalue $\lambda_2$ of $\rho_A$. They all satisfy that vanish for product states (corresponding to $\rho_A$ with rank 1) while having a positive value for any other state (corresponding to $\rho_A$ with rank 2). Thus we can interpret the inequivalence under SLOCC of states whose reduced density matrices differ in rank also in terms of the impossibility of creating any of the bipartite measures. For instance, a state in the $A-BC$ class has $S_A=0$, and thus cannot be transformed with any finite probability into a state of the $AB-C$ class, because this would have $S_A>0$. We conclude that the monotonicity of these measures readily split the set of pure states of three qubits into five subsets which are inequivalent under SLOCC, namely unentagled states $A-B-C$, the three classes $A-BC$, $AB-C$ and $C-AB$ containing only bipartite entanglement, and a fifth subset of entangled states with $S_A,S_B,S_C \neq 0$ (i.e. r$(\rho_A)=$r$(\rho_B)=$r$(\rho_C)=2$). Bipartite measures cannot, however, determine the inequivalence of the GHZ and W classes.
Is there any known measure of tripartite entanglement which can distinguish between these two classes? The 3-tangle does. Indeed, it can be computed from the product decompositions (\[GHZclass\]) and (\[Wclass\]) (see [@Wo99] for details), and reads (\_[GHZ]{})=(2Ks\_s\_s\_s\_c\_)\^2 \[tangle\] =0\[tauGHZ\] for any state in the GHZ class, while it vanishes for any state in the W class. In the appendix B we prove that the 3-tangle is an entanglement monotone, a very desirable property for any quantity aiming at measuring entanglement. Consequently, a state in the W class cannot be transformed by means of LOCC (and in particular SLOCC) to a state in the GHZ class, which is an independent proof of the fact that the two kinds of true tripartite entanglement are indeed inequivalent under SLOCC.
Practical identification
------------------------
Given an arbitrary state ${\mbox{$| \psi \rangle$}}$ of three qubits, expressed in any basis, it may be interesting to know, for instance, whether it can be converted by means of LOCC into a GHZ or a W state, if any, or into a EPR state shared between two of the parties. In our original analysis of the classes we already have provided a constructive method, based on the analysis of r$(\rho_{\kappa})$ and R$(\rho_{BC})$, to determine the class of ${\mbox{$| \psi \rangle$}}$ under SLOCC. Analysing the R$(\rho_{BC})$ may, however, not be the most practical way to proceed. Here we suggest to proceed instead according to the following two steps:
- compute $\rho_{\kappa}$, $\kappa=A,B$ and $C$, and check whether they have a vanishing determinant. \[note that det$\rho_{\kappa}=0 \Leftrightarrow
S_{\kappa} = 0 \Leftrightarrow $r$(\rho_{\kappa})=1$\]
- If none of the previous determinants vanish \[that is, ${\mbox{$| \psi \rangle$}}$ has true tripartite entanglement\], then compute the $3$-tangle using the recipe introduced in [@Wo99].
Then Table I, which sumarizes the relation between classes under SLOCC and measures of entanglement, can be used to catalogue state ${\mbox{$| \psi \rangle$}}$.
State ${\mbox{$| W \rangle$}}$ and residual bipartite entanglement.
===================================================================
As mentioned in the previous section, in several aspects the state ${\mbox{$| GHZ \rangle$}}$ is the maximally entangled state of three qubits. It also has the feature that when one of the qubits is traced out, then the remaining two are completely unentangled. This means, in particular, that if one of the three parties sharing the system decides not to cooperate with the other two, then they can not use at all the entanglement resources of the state. The same happens if for some reason the information about one of the qubits —namely the identity of the corresponding states ${\mbox{$| 0 \rangle$}}$ and ${\mbox{$| 1 \rangle$}}$ in (\[GHZ\])— is lost.
Here we would like to investigate the robustness of the entanglement of a three-qubit state ${\mbox{$| \psi \rangle$}}$ against disposal of one of the qubits [@Br00]. The residual, two-qubit states $\rho_{AB}$, $\rho_{AC}$ and $\rho_{BC}$ are in general mixed states. There are several measures of entanglement of mixed states and therefore multiple ways of quantifying how much (mixed-state) bipartite entanglement the state ${\mbox{$| \psi \rangle$}}$ turns into when one of the qubits is traced out. Nevertheless, most of the criteria we have examined coincide in pointing out the state ${\mbox{$| W \rangle$}}$ as the one that maximally retains bipartite entanglement. Note that the reduced density matrix of $|W\rangle$ is identical for any two subsystems and is e.g. given by \_[AB]{}=|\^+\^+|+|0000|,\[reducedW\] with $|\Psi^+\rangle=1/\sqrt{2}(|01\rangle+|10\rangle)$ being a maximally entangled state of two qubits. Note that one can obtain from a single copy of (\[reducedW\]) a state which is arbitrarily close to the state $|\Psi^+\rangle$ by means of a filtering measurement [@Gi96].
Average residual entanglement
-----------------------------
Let us consider first which is the amount of bipartite entanglement, according to some measure $\E(\rho)$, that the two remaining qubits retain on average when a third one is traced out, that is, |() ((\_[AB]{})+ (\_[AC]{}) + (\_[BC]{})). \[average\] In general, computing the amount of entanglement $\E(\rho)$ for bipartite mixed states is a difficult problem. However numerical results have shown that ${\mbox{$| W \rangle$}}$ maximizes the average entanglement of formation, that is the choice $\E(\rho) = E_f(\rho)$, where $E_f(\rho)$ [^4] is the minimal amount of bipartite pure-state entanglement \[as quantified by means of the entropy of entanglement\] required to prepare locally one single copy of the state $\rho$ [@Vi00].
In addition, we have managed to show analytically (see appendix C) for the particular choice $\E(\rho) = \c(\rho)^2$, where $\c(\rho)$ is the concurrence (for a definition of the concurrence see e.g. [@Wo99]), the state ${\mbox{$| W \rangle$}}$ reaches the maximal average value $\bar{\c^2}(W)=4/9$, which no other state can match.
Least entangled pair
--------------------
Another way of quantifying how resistent the entanglement of a tripartite state ${\mbox{$| \psi \rangle$}}$ is to dismissal of one part of the system consists in looking at the least entangled of the three possible remaining parts, namely at the function \_() ((\_[AB]{}), (\_[AC]{}), (\_[BC]{})). \[worstcase\] For this “worst case scenario” we have been able to prove analytically (see appendix C) that the maximal value of $\E_{\min}(\psi)$ is obtained by the state ${\mbox{$| W \rangle$}}$ for any bipartite measure $\E(\rho)$ which is monotonic with the concurrence, $\c(\rho)$, such as the entanglement of formation $E_f(\rho)$ and the monotone $E_2(\rho)$ [^5], which denotes the minimal amount of bipartite pure-state entanglement \[quantified by means of $E_2(\psi)$\] required to prepare locally one single copy of the state $\rho$.
We conclude that the state ${\mbox{$| W \rangle$}}$ is the state of three-qubits whose entanglement has the highest degree of endurance against loss of one of the three qubits. We conceive this property as important in any situation where one of the three parties sharing the system, say Alice, may suddenly decide not to cooperate with the other two. Notice that even in the case that Alice would decide to try to destroy the entanglement between Bob and Claire, this would not be possible, since any local action on A cannot prevent Bob and Claire from sharing, at least, the entanglement contained in $\rho_{BC}$ (for instance, by simply ignoring Alice’s actions). Therefore, although essentially tripartite, the entanglement of the state ${\mbox{$| W \rangle$}}$ is also readily bipartite, in contrast to that of the state ${\mbox{$| GHZ \rangle$}}$, which only after some local manipulation can be brought into a bipartite form.
Generalization to $N$ parties
=============================
In this last section we would like to apply the same techniques to analyze the entanglement of more general multipartite systems. We will learn that the set of entangled states is a rather inaccessible jungle for the local explorer, for two pure states ${\mbox{$| \psi \rangle$}}$ and ${\mbox{$| \phi \rangle$}}$ are typically not connected by means of LOCC, so that the parties are usually unable to convert states locally. We will also study generalizations to $N$ qubits of the state ${\mbox{$| W \rangle$}}$.
Local inaccessibility of states in general multipartite systems
---------------------------------------------------------------
Let us consider first $N$ parties each possessing a qubit. The Hilbert space of the system is \^[(N)]{}=\_[N]{}, and therefore up to a global, physically irrelevant complex constant, a generic vector depends on $2(2^N-1)$ real parameters. On the other hand we want to identify vectors which are related by means of a ILO. A general one-party, invertible operator $A$ must have non-vanishing determinant, which we can fix to one, det$A=1$, because the operator $kA$ only differs in that it introduces in the transformed states an extra constant factor $k\in\C$, which we have already addressed. That is, $A\in SL_{2}(\C)$, and it depends on $6$ real parameters. Therefore the set of equivalence classes under SLOCC, , depends [*at least*]{} on $2(2^N-1) - 6N$ parameters. This lower bound allows for a finite number of classes for $N=3$, but shows that for any larger number $N$ of qubits there are infinitely many classes, labeled by at least one continuous parameter. The reason is that the number of parameters from a state ${\mbox{$| \psi \rangle$}}$ which the parties can modify by means of a general ILO $A\otimes
B\otimes...\otimes N$ grows linearly with $N$ ($6N$ for the multi-qubit case), whereas the number of parameters required to specify ${\mbox{$| \psi \rangle$}}$ grows exponentially with $N$.
More generally, if the Hilbert space is given by $\H=\C^{n_1}\otimes ... \otimes
\C^{n_N}$, then the set of equivalence classes under SLOCC, , depends at least on $2(n_1n_2...n_N-1) - 2\sum_{i=1}^N(n_i^2-1))$. This shows that only for $N=3$ there are still some systems with (potentially) only a finite number of classe under SLOCC, namely those with Hilbert space $\C^2\otimes\C^{n_2}\otimes\C^{n_3}$, that is, having a qubit as at least one of the subsystems. In all other cases, one finds an infinite number of classes.
We notice that even allowing for non-invertible local operations the amount of parameters that can be changed by local manipulations is typically smaller than that the state depends on. That is, the subset of states that can be reached locally from a given state ${\mbox{$| \psi \rangle$}}$ is of zero measure in the set of states of the multipartite system. Recall that in the bipartite scenario, $\H=\C^n\otimes\C^m$, there is always a maximally entangled state from which all the other states can be locally prepared with certainty of success. We see now that, in constrast, there is typically in a multipartite system no state from which all the others can be prepared, not even with some probability of success. Of course, the parties can always resort to, say, using a sufficient amount of EPR states distributed among them to prepare any multipartite state by standard teleportation. This implies, however, using an initial state (that of many EPR states) which belongs to a Hilbert space much larger than the Hilbert space of the state the parties are trying to create, and thus does not change the previous conclusion.
State in multi-qubit systems
----------------------------
Let us have a look at the generalized form $|W_N\rangle$ of the state $|W\rangle$ (\[W\]). We define the state |W\_N1/|N-1,1, where $|N-1,1\rangle$ denotes the totally symmetric state including $N-1$ zeros and $1$ ones. For example, we obtain for $N=4$ |W\_4=1/(|0001+|0010+|0100+|1000). One immediately observes that the entanglement of this state is again very robust against particle losses, i.e. the state $|W_N\rangle$ remains entangled even if any $N\!-\!2$ parties lose the information about their particle. This means that any two out of $N$ parties possess an entangled state, independently of whether the remaining $(N-2)$ parties decide to cooperate with them or not. This can be seen by computing the reduced density operator $\rho_{AB}$ of $|W_N\rangle$, i.e. by tracing out all but the first and the second systems. By symmetry of the state $|W_N\rangle$, we have that all reduced density operators $\rho_{\kappa\mu}$ are identical and we obtain \_=(2 |\^+\^+|+(N-2)|0000|). The concurrence can easily determined to be \_(W\_N)=, which shows that $\rho_{\kappa\mu}$ is entangled, even distillable. We conjecture that the average value of the square of the concurrence for ${\mbox{$| W_N \rangle$}}$, \_ \_ [C]{}\^2\_(W\_N)=, is again the maximal value achievable for any state of $N$ qubits.
Summary and conclusions
=======================
In this work, we investigated equivalence classes of multipartite states specified by stochastic local operations and classical communication. We showed that for pure states of three qubits there are 6 different classes of this kind. In particular, we found that there are two inequivalent types of genuine tripartite entanglement, represented by the GHZ state and the state W. We showed that the state W is the state of three qubits that retains a maximal amount of bipartite entanglement when any one of the three qubits is traced out. For multipartite ($N\geq 4$) and multilevel systems, we showed that there exist infinitely many inequivalent kinds of entanglement (i.e. classes under SLOCC).
Acknowledgments {#acknowledgments .unnumbered}
===============
This work was supported by the Austrian Science Foundation under the SFB “control and measurement of coherent quantum systems” (Project 11), the European Community under the TMR network ERB–FMRX–CT96–0087, the European Science Foundation and the Institute for Quantum Information GmbH. G.V also acknowledges a Marie Curie Fellowship HPMF-CT-1999-00200 (European Community).
Appendix A: SLOCC and local ranks {#appendix-a-slocc-and-local-ranks .unnumbered}
=================================
In this appendix we show that states ${\mbox{$| \psi \rangle$}}$ and ${\mbox{$| \phi \rangle$}}$ belong to the same class under SLOCC iff they are related by means of a invertible local operator (ILO). From this connection it follows easily that the local ranks of a pure state, r($\rho_{\kappa})$, $\kappa=A,B,...$, are invariant under SLOCC, whereas under LOCC they can only decrease.
[**Lemma:**]{} If the bipartite vectors ${\mbox{$| \psi \rangle$}}$ and ${\mbox{$| \phi \rangle$}} \in \C^n\otimes\C^m$ fulfill = A1\_B , then the ranks of the corresponding reduced density matrices satisfy r$(\rho_A^{\psi}) \geq $ r$(\rho_A^{\phi})$ and r$(\rho_B^{\psi}) \geq
$r$(\rho_B^{\phi})$.
[**Proof:**]{} We consider the Schmidt decomposition of ${\mbox{$| \psi \rangle$}}$, = \_[i=1]{}\^[n\_]{} , \^\_i > 0, n\_(n,m), and write the operator $A$ as A = \_[i=1]{}\^n , \[operator\] where ${\mbox{$| \mu_i \rangle$}} \in \C^n$ do not need to be normalized nor linearly independent. Then we have that $\rho_A^{\psi} =
\sum_{i=1}^{n_{\psi}}{{\mbox{$| i \rangle$}}\!{\mbox{$\langle i |$}}}$ and $\rho_A^{\phi} = A\rho_A^{\psi}A^{\dagger} =
\sum_{i=1}^{n_{\psi}}{{\mbox{$| \mu_i \rangle$}}\!{\mbox{$\langle \mu_i |$}}}$, so that r$(\rho_A^{\phi})\leq n_{\psi}$. The second inequality of the Lemma follows from the fact that for any bipartite vector r$(\rho_A)=$ r$(\rho_B)$. $\Box$
[**Corollary:**]{} If the vectors ${\mbox{$| \psi \rangle$}},{\mbox{$| \phi \rangle$}}\in\H_A\otimes\H_B\otimes...\otimes\H_N$ are connected by a local operator as ${\mbox{$| \phi \rangle$}}=A\otimes B\otimes...\otimes N {\mbox{$| \psi \rangle$}}$, then the local ranks satisfy r$(\rho_{\kappa}^{\psi}) \geq $ r$(\rho_{\kappa}^{\phi})$, $\kappa=A, B,..., N$.
[**Proof:**]{} Indeed, for each of the parties, say Alice for concreteness, we can view the operator $A\otimes B\otimes...\otimes N$ as the composition of two local operators, $A\otimes
1_{B...N}$ and $1_A\otimes (B\otimes...\otimes N)$, and the Hilbert space as $\H_A\otimes \H_{B...N}$. Then, because of the previous lemma, application of the first operator cannot increase r($\rho_A$), and the same happens with the second operator, which cannot increase r$(\rho_{B...N})$ \[recall that for any pure state r$(\rho_A)=$ r$(\rho_{B...N})$\]. $\Box$
[**Theorem:**]{} Two pure states of a multipartite system are equivalent under SLOCC iff they are related by a local invertible operator.
[**Proof:**]{} If =AB...N , \[localoperator\] then a local protocol exists for the parties to transform ${\mbox{$| \psi \rangle$}}$ into ${\mbox{$| \phi \rangle$}}$ with a finite probability of success. Indeed, each party needs simply perform a local POVM including a normalized version of the corresponding local operator in (\[localoperator\]). For instance, Alice has to apply a POVM defined by operators $\sqrt{p_A}A$ and $\sqrt{1_A - p_AA^{\dagger}A}$, where $p_A\leq 1$ is a positive weight such that $p_AA^{\dagger}A \leq 1_A$, and similarly for the rest of the parties. Then such a local protocol converts ${\mbox{$| \psi \rangle$}}$ succesfully into ${\mbox{$| \phi \rangle$}}$ with probability $p_Ap_B...p_N$. If, in addition, $A,B,...,N$ are invertible operators, then obviously =A\^[-1]{}B\^[-1]{}...N\^[-1]{} and the conversion can be reversed locally. Let us then move to prove the converse. We already argued (section II.A) that if ${\mbox{$| \psi \rangle$}}$ can be converted into ${\mbox{$| \phi \rangle$}}$ by LOCC, then a local operator relate them. We want to prove now that equivalence of ${\mbox{$| \psi \rangle$}}$ and ${\mbox{$| \phi \rangle$}}$ under SLOCC implies that this operator can always be chosen to be invertible. For simplicity, we will assume that ${\mbox{$| \psi \rangle$}}$ and ${\mbox{$| \phi \rangle$}}$ are related by a local operator acting non-trivially only in Alice’s part, = A1\_[B...N]{} . \[onlyA\] We can then consider the Schmidt decomposition of the states with respect to part $A$ and part $B...N$ = \_[i=1]{}\^[n\_]{} , \^\_i > 0\
= \_[i=1]{}\^[n\_]{} (U\_A), \^\_i > 0 where the local unitary $U_A$ relate the two local Schmidt basis in Alice’s part, $\{{\mbox{$| i \rangle$}}\}_{i=1}^n \in \H_A = \C^n$, ${\mbox{$| \tau_i \rangle$}}\in
\H_B\otimes...\otimes\H_N$, and $n_{\psi}=n_{\phi}$ because of the previous corollary. Now, operator $A$ in equation (\[onlyA\]) must be of the form (up to some irrelevant permutations in the Schmidt basis) A = U\_A (A\_1 + A\_2)\
A\_1 \_[i=1]{}\^[n\_]{} ,\
A\_2 \_[i=n\_+1]{}\^n where ${\mbox{$| \mu_i \rangle$}}$ are arbitrary unnormalized vectors. Notice that vectors ${\mbox{$| \mu_i \rangle$}}$ play no role in equation (\[onlyA\]) since $A_2\otimes1_{B...N}{\mbox{$| \psi \rangle$}}=0$. Therefore we can redefine A\_2 \_[i=n\_ +1]{}\^n , which implies that $A$ is an invertible operator.$\Box$
Appendix B: $\tau$ is an entanglement monotone {#appendix-b-tau-is-an-entanglement-monotone .unnumbered}
==============================================
In this appendix, we show that the 3-tangle $\tau$ is an entanglement monotone, i.e. decreasing on average under LOCC in all the three parties. We first note that any local protocol can be decomposed into POVM’s such that only one party performs operations on the system. This, together with the invariance of the 3-tangle $\tau$ under permutations of the parties, ensures that it is sufficient to consider a local POVM in $A$ only. Furthermore, we can restrict ourselves to two–outcome POVM’s due to the fact that a genarlized (local) POVM can be implemented by a sequence of two outcome POVM’s. Let $A_1,A_2$ be the two POVM elements such that $A_1^\dagger A_1+A_2^\dagger A_2 = {\mbox{$1 \hspace{-1.0mm} {\bf l}$}}$. We can write $A_i=U_iD_iV$, where $U_i$, $V$ are unitary matrices and $D_i$ are diagonal matrices with entries $(a,b)$ $[((1-a^2)^{\frac{1}{2}},(1-b^2)^{\frac{1}{2}}$)\] respectively. Note that we used the singular value decomposition for $A_i$, and we have that the restriction that $A_1,A_2$ constitute a POVM immediately implies that the unitary operation $V$ can be chosen to be the same in both cases. We consider an initial state $|\psi\rangle$ with 3-tangle $\tau(\psi)$. Let $|\tilde\phi_i\rangle=
A_i|\psi\rangle$ be the (unnormalized) states after the application of the POVM. Normalizing them, we obtain $|\phi_i\rangle=|\tilde\phi_i\rangle/\sqrt{p_i}$ with $p_i=\langle\tilde\phi_i|\tilde\phi_i\rangle$ and $p_1+p_2=1$. We want to show that $\tau^\eta$, $0<\eta\leq1$ is, on average, always decreasing and thus an entanglement monotone, i.e for = p\_1\^(\_1)+ p\_2 \^(\_2)\[monotontau\] we have that \^() \[monoton\] is fulfilled for all possible choices of the POVM $\{A_1,A_2\}$. Using that $\tau$ is invariant under local unitaries, we do not have to consider the unitary operations $U_i$ in our calculations, i.e. $\tau(U_iD_iV\psi)=\tau(D_iV\psi)$. Taking this simplification into account, a straightforward calculation shows that (\_1)= () , (\_2)= (), where we used that $\tau(\epsilon \tilde{\phi_i})=\epsilon^4
\tau(\tilde{\phi_i})$, which can be checked by noting that $\tau$ is a quartic function with respect to its coefficients in the standard basis[@Wo99]. Note that the dependence of $\tau(\phi_i)$ on the unitary operation $V$ is hidden in $p_i$. For $\eta=1/2$, one obtains for example $\tau^{\frac{1}{2}}(\phi_1)=ab/p_1\tau^{\frac{1}{2}}(\psi)$. Substituting in (\[monotontau\]), we find =(ab+) \^(). \[tau12\] In this case, one can easily check that ($\ref{tau12}) \leq \tau^{\frac{1}{2}}$ by noting that (\[tau12\]) is maximized for $a=b$. We thus have that $\tau^{\frac{1}{2}}$ is, on average, always decreasing and thus an entanglement monotone. In a similar way, one can check for $0< \eta \leq 1$ that $\tau^\eta$ is an entanglement monotone. However, for $\eta \not= 1/2$, the derivation is a bit more involved due to the fact that in this case the propabilities $p_i$ in the expression for $<\tau^\eta>$ do no longer cancel and have to be calculated explicitly.
Appendix C: $|W\rangle$ maximizes residual bipartite entanglement {#appendix-c-wrangle-maximizes-residual-bipartite-entanglement .unnumbered}
=================================================================
Here we show that for all tripartite pure states, except the state $|W\rangle$ the following inequality holds E\_\^2\_[AB]{}+[C]{}\^2\_[AC]{}+[C]{}\^2\_[BC]{} < ,\[inequ\] while the state $|W\rangle$ reaches the value $E_\tau=4/3$. Note that we used the shorthand notation ${\cal C}_{AB}$ for the concurrence of the reduced density operator $\rho_{AB}, {\cal C}(\rho_{AB})$, and similary for ${\cal
C}_{AC}$,${\cal C}_{BC}$.
Inequality (\[inequ\]) already implies that the state $|W\rangle$ reaches the maximum average value $\bar{\E}(\psi)$ of Equ. (\[average\]) for the choice of $\E(\rho) =
\c(\rho)^2$, namely $\bar{\E}(W)=4/9$.
At the same time, inequality (\[inequ\]) also shows that the state $|W\rangle$ maximizes the function $\E_{\min}(\psi)$ (\[worstcase\]) for the choice of $\E(\rho) = \c(\rho)^2$, since (\[inequ\]) implies that \^2\_() ([C]{}\^2\_[AB]{},[C]{}\^2\_[AC]{},[C]{}\^2\_[BC]{}) < 4/9 \[mini\] for all states except the state $|W\rangle$, for which the value $4/9$ is reached. From (\[mini\]) follows that for any bipartite measure of entanglement $\E(\rho)$ which is monotonically increasing with the square of the concurrence (and hence with the concurrence itself), the state $|W\rangle$ maximizes the function $\E_{\min}(\psi)$ (\[worstcase\]), i.e. \_() < \_(W) = ([C]{}\^2=4/9). Assume that this is not the case, i.e. there exist a state $\psi$ for which $\E_{\min}(\psi) > \E_{\min}(W)$. Since by assumption $\E$ is monotonically increasing with the concurrence, this would imply that ${\cal C}^2_{\min}(\psi)
>4/9$, which contradicts Equ. (\[mini\]) and is hence impossible.
Note in addition that any good measure of entanglement should be a convex function [@Vi00J], as $\c(\rho), E_f(\rho)$ and $E_2(\rho)$ are. This implies, when applied to (\[average\]) and (\[worstcase\]) that the optimal values for $\bar{\E}$ and $\E_{\min}$ are achieved for pure states.
Ther remainder of this appendix is devoted to prove inequality (\[inequ\]). Using the definition of the 3-tangle, $\tau\equiv\tau_{ABC}={\cal C}^2_{A(BC)}-{\cal
C}^2_{AB}-{\cal C}^2_{AC}$ [@Wo99]and the invariance of the 3-tangle under permutations of the parties, we can rewrite $E_\tau$ as $1/2({\cal C}^2_{A(BC)}+{\cal
C}^2_{B(AC)}+{\cal C}^2_{C(AB)}-3\tau)$. Using that ${\cal
C}^2_{\kappa(\mu\nu)}=4{\rm det}\rho_{\kappa}$, we can evaluate $E_\tau$ for the different classes.
Starting with the class $A-B-C$, we immeadetly obtain that $E_\tau(\Psi_{A-B-C})=0$. For the class $A-BC$, we have that $\tau=0$ and ${\cal C}^2_{A(BC)}=0$. Since ${\cal C}^2_{B(AC)},{\cal C}^2_{C(AB)} \leq 1$, we have that $E_\tau(\Psi_{A-BC}) \leq 1$ in this case (and similary for the classes $B-AC,
C-AB$).
Now we consider the class $W$, specified by equ. (\[Wclass\]). Again, we have that $\tau=0$. We find that $E_\tau(\Psi_W)=4(ab+ac+bc)$ (which does not depend on $d$). Notice that $E_\tau$ is maximized for $a=b=c=1/3$ - which corresponds to the state $|W\rangle$ - and leads to $E_\tau=4/3$. For all other values of $a,b,c,d$, we have that $E_\tau < 4/3$.
Let us now turn to the class GHZ, specified in eq. (\[GHZclass\]). Using that $\tau(\Psi_{GHZ})$ is given in eq. (\[tauGHZ\]) and $\det\rho_A=K^2c_\delta^2s_\delta^2s_\alpha^2(1-c_\beta^2c_\gamma^2)$ (and similary for $\det\rho_{B,C}$), we obtain E\_= \[Etau\] One readily checks that (\[Etau\]) is maximized for $\delta=\pi/4$ and $\varphi=\pi$ (which corresponds to $c_\delta=s_\delta=1/\sqrt{2}$ and $c_\varphi=-1$), independent of the values of $\alpha,\beta,\gamma \in
(0,\pi/2]$. Thus we have that $E_\tau \leq E_\tau(\delta=\pi/4,\varphi=\pi)$ and after some algebra we obtain E\_ \[Etau1\] We want to show that the (rhs) of eq. \[Etau1\] $< 4/3$. Let us call $x\equiv
c_\alpha,y\equiv c_\beta,z\equiv c_\gamma$ with $0\leq x,y,z<1$. We thus have to show that f(x,y,z)&&3(x\^2+y\^2+z\^2)-6(x\^2y\^2+x\^2z\^2+y\^2z\^2)\
&+&5(x\^2y\^2z\^2)-4+8xyz < 0 Let us calculate the maximum of $f(x,y,z)$. We therefore take the derivatives of $f(x,y,z)$ with respect to $x,y,z$ respectively (which we denote by $f_x,f_y,f_z$) and set them to zero. One immeadetly observes (by considering linear combination of the resulting equations, e.g. $xf_x-yf_y$, where one e.g. obtains $(x^2-y^2)(1-2z^2)=0$), that for a maximum we must have $x=y=z$. The possible solutions of the resulting polynomial of degree 5 can be checked to lie outside the intervall $[0,1)$, i.e. outside the range of $x,y,z$ except for $x=y=z=0$. It can however be easily verified that this solution give rise to a minimum of $f(x,y,z)$, namely $f(0,0,0)=-4$. Thus the maximum of $f(x,y,z)$ is obtained at the border of the range for $x,y,z$, which corresponds to the surfaces of a cube. Due to the fact that $f(x,y,z)$ is invariant under permutations of the variables, we only have to check two of the surfaces, e.g. the surfaces specified by $x=0$ and $x=1$ (actually $x=1-\epsilon$, where $\epsilon$ is an infinitesimally small positive number) and we find (i) $f(0,x,y)=3(y^2+z^2)-6y^2z^2-4 \leq -1$ (the maximum in this case is e.g. obtained for $y=0,z=1-\epsilon$)) and (ii) $f(1,y,z)=8yz-3(y^2+z^2)-y^2z^2-1 <
0$. In (ii), it can be checked that a necessary condition for a maximum is $y=z$ and that $f(1,y,y)$ is monotonically increasing in $[0,1)$ and is thus maximized for $y=z=(1-\epsilon)$. One obtains $f(x,y,z) \leq f(1,1-\epsilon,1-\epsilon) <
0$ as desired.
So we managed to show that the state $|W\rangle$ is the only state which fulfills $E_\tau=4/3$, and for all other tripartite pure states we have that $E_\tau < 4/3$.
C. H. Bennett, H. J. Bernstein, S. Popescu, B. Schumacher, quant-ph/9511030
Einstein, Podolsky & Rosen, Phys. Rev. [**47**]{}, 777-780 (1935).
N. Linden, S. Popescu, B. Schumacher and M. Westmoreland, quant-ph/9912039; G. Vidal , W. Dür and J. I. Cirac, quant-ph/0004009; S. Wu and Y. Zhang, quant-ph/0004020.
C. H. Bennett, S. Popescu, D. Rohrlich, J.A. Smolin and A.V. Thapliyal, quant-ph/9908073;
G. Vidal, Journ. of Mod. Opt. [**47**]{}, 355 (2000);
N. Linden and S. Popescu, Fortsch.Phys. [**46**]{}, 567 (1998);
J. Schlienz, Ph.D. thesis
A. Sudbery, quant-ph/0001116;
H. A. Carteret and A. Sudbery, quant-ph/0001091;
J. Kempe, Phys. Rev. A[**60**]{} 910-916 (1999);
Different classifications based on LU invariants were recently proposed in [@Ac00] and [@Ca00]. Depending on the values of these invariants, several classes of states in one case and types of entanglement in the other are identified. We want to remark that our classification is based on probabilistic conversions under LOCC, and therefore, as expectable, our results are not fully compatible with those of these approaches.
D. M. Greenberger, M. Horne, A. Zeilinger, [*Bell’s theorem, Quantum Theory, and Conceptions of the Universe,*]{} ed. M. Kafatos, Kluwer, Dordrecht 69 (1989); D. Bouwmeester et al., Phys. Rev. Lett. [**82** ]{}, 1345 (1999).
V. Coffman, J. Kundu, and W. K. Wootters, Phys. Rev. A [**61**]{}, 052306 (2000); see also quant-ph/9907047
H. K. Lo and S. Popescu, quant-ph/9707038;
G. Vidal, Phys. Rev. Lett. [**83** ]{}, 1046 (1999);
M. Nielsen, Phys. Rev. Lett. [**83** ]{}, 436 (1999);
Recall that the $n$ entanglement monotones $E_k(\psi)\equiv\sum_{i=k}^n \lambda_i$, where $\lambda_i$ come from the Schmidt decomposition of ${\mbox{$| \psi \rangle$}} \in \C^n\otimes\C^n$, = \_[i=1]{}\^n , \_i\_[i+1]{}0, are the measures that provide a quantitative description of the entanglement resources of a single copy of pure states, in that, for instance, they give the optimal probability $P(\psi\rightarrow\phi)$ of conversion of the state ${\mbox{$| \psi \rangle$}}$ into the state ${\mbox{$| \phi \rangle$}}$ under LOCC [@Vi99], namely P()= {,...,}. In a two-qubit system we have only two non-trivial monotones, $E_1(\psi)=\lambda_1+\lambda_2 =1$ and $E_2=\lambda_2\leq 1/2.$ In a similar way as $E_2(\psi)>E_2(\phi)$ determines that ${\mbox{$| \psi \rangle$}}$ can be transformed into ${\mbox{$| \phi \rangle$}}$ with certainty by means of LOCC [@Ni99], the extension to mixed states $\rho$ of two qubits, $E_2(\rho)$, also says when the entanglement of the pure state ${\mbox{$| \psi \rangle$}}$ suffices to locally prepare $\rho$ with certainty [@Vi00].
A. Sampera, R. Tarrach, G. Vidal, Phys. Rev. A [**58**]{} (1998) 826-830.
A. Acín, A. Andrianov, L. Costa, E. Jané, J. I. Latorre and R. Tarrach, quant-ph/0003050;
N. Gisin and H. Bechmann-Pasquinucci, quant-ph 9804045
A state $t{\mbox{$| 00 \rangle$}}+x{\mbox{$| 01 \rangle$}}+y{\mbox{$| 10 \rangle$}}+z{\mbox{$| 11 \rangle$}}$ is product iff it can be written as $(a{\mbox{$| 0 \rangle$}}+b{\mbox{$| 1 \rangle$}})(c{\mbox{$| 0 \rangle$}}+d{\mbox{$| 1 \rangle$}})$, that is iff tz=xy. A complex linear combination $\lambda_1{\mbox{$| 00 \rangle$}}+\lambda_2(x{\mbox{$| 01 \rangle$}}+y{\mbox{$| 10 \rangle$}}+z{\mbox{$| 11 \rangle$}})$ with $\lambda_2\neq
0$ cannot be product for any $\lambda_1$ iff $z=0$.
O. Cohen, T. A. Brun, quant-ph/0001084
A. Higuchi, A. Sudbery, quant-ph/0005013; H. J. Briegel, R. Raussendorf, quant-ph/0004051
N. Gisin, Phys. Lett. A [**210**]{} 151 (1996).
G. Vidal, quant-ph/0003002;
Class $S_A$ $S_B$ $S_C$ $\tau$
------- ------- ------- ------- --------
A-B-C 0 0 0 0
A-BC 0 $>0$ $>0$ 0
B-AC $>0$ 0 $>0$ 0
C-AB $>0$ $>0$ 0 0
W $>0$ $>0$ $>0$ 0
GHZ $>0$ $>0$ $>0$ $>0$
: Values of the local entropies $S_A, S_B, S_C$ and the 3-tangle $\tau$ for the different classes[]{data-label="Table1"}
(230,220) (5,55)
[^1]: The reduced density matrix $\rho_{AB}$ of a pure tripartite state $|\psi\rangle$ is defined as $\rho_{AB}\equiv tr_C(|\psi\rangle\langle\psi|)$.
[^2]: Stochastic transformations under LOCC had been previously analyzed in [@Lo97; @Vi99].
[^3]: In practice the constraints $A^{\dagger}A,~B^{\dagger}B,~C^{\dagger}C \leq 1$ should be fulfilled if the invertible operators $A,B,C$ are to come from local POVMs. In this work we do not normalize them in order to avoid introducing unimportant constants to the equations. Instead, both the initial and final states are normalized.
[^4]: The entanglement of formation is given by $E_f(\rho)=h(\frac{1}{2}+\frac{1}{2}\sqrt{1-{\cal C}^2})$, where ${\cal C}$ is the concurrence and $h$ is the binary entropy function $h(x)=-x{\rm log}_2x-(1-x){\rm log}_2(1-x)$.
[^5]: The entanglement monotone $E_2$, expressed in terms of the concurrence $\c$ is given by $E_2(\rho)=\frac{1}{2}-\frac{1}{2}\sqrt{1-{\cal C}^2}$.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- Chang Sun
- Travis Askham
- 'J. Nathan Kutz'
bibliography:
- 'refs.bib'
title: 'Stability and Dynamics of Microring Combs: Elliptic function solutions of the Lugiato-Lefever equation'
---
Introduction
============
Frequency comb generation in microresonators has become a critically enabling technology for applications in metrology, high-resolution spectroscopy and microwave photonics [@app2; @kippenberg2011microresonator; @del2011octave; @papp2011spectral; @ferdous2011spectral; @app1; @liang2015high; @suh2016microresonator]. A clear goal in such microresonators is the generation of octave-spanning combs, which is often achieved by the generation of a single soliton in a high-Q microresonator cavity [@herr2014temporal; @yi2015soliton]. Much like the multi-pulsing instability (MPI) in mode-locked laser cavities [@kutz1; @kutz2; @kutz3], microresonators are prone to generating multiple pulses in the cavity [@kip1; @kip2], thus compromising the performance of the frequency comb generation. Consequently, the dynamics and stability of pulse generation in the microresonator is of significant interest. In this manuscript, we explore analytically tractable solutions of the Lugiato-Lefever equation (LLE) [@lugiato1987spatial], which is the governing equation for the microresonator dynamics [@PhysRevA.87.053852]. While solitons have been observed in a number of experimental architectures, the deterministic manipulation of states with multiple solitons in microresonators has only been recently explored with the goal of prediction and control [@kip1]. We develop a perturbation theory for periodic pulse train solutions, known as Jacobi elliptic functions, which characterize the underlying solutions in the microresonator cavity. Our work provides a theoretically rigorous complement to recent experimental observations for the transitions between $N$ to $N+1$ (or vice versa) pulses in a microresonator. We further show how cavity perturbations, due to, for instance, the Raman effect or spontaneous emission noise, affect the resulting combline stability and robustness.
Soliton perturbation theory has been one of the most successful theoretical tools developed for characterizing the underlying physics in optical communication systems [@karpman1977perturbation; @kodama1981perturbations; @kaup1990perturbation; @elgin1993perturbations] and mode-locked lasers [@kartner1996soliton; @kapitula2002stability; @kapitula2004evans; @Bale:08]. In this work, we develop a LLE combline perturbation theory. The theory relies on an analytic solution, the Galilean invariant one-soliton solution, of the nonlinear Schrödinger equation. Jacobi elliptic functions are a generalization of soliton solutions of the LLE equation, capable of representing both single localized pulse solutions and periodic pulse trains. Much like solitons, the solutions are parameterized by a number of free parameters whose slow evolution under perturbation characterizes the stability of the solution. A linear stability analysis of the Jacobi elliptic solutions is capable of revealing key properties of the combline properties under perturbation. Specifically, our analysis characterizes the stability of $N$ pulses per round trip in the laser cavity. Much like MPI in mode-locked lasers [@kutz1; @kutz2; @kutz3], an initial cavity cold start will jump to the most energetically favorable configuration. However, our analysis shows how one can manipulate the number of pulses per round trip by simply manipulating the microresonator detuning, confirming experimental findings.
From a technical point of view, our stability analysis follows closely the rigorous theory of soliton perturbation theory. For the LLE [Jacobi elliptic solutions]{}, the linearized operator contains four zero modes which correspond to invariances of the solutions. The effect of perturbations on these zero modes is quantified and shows how various perturbations can either destabilize the solutions or force solutions to translate or bifurcate to a higher or lower number of [pulses]{} per round trip. Additionally, we show that the remainder of the continuous spectrum remains bounded under perturbation. We demonstrate the application of the theory on two canonical examples: (i) the LLE modified to include Raman dynamics, and (ii) the LLE under the influence of white noise (stimulated emission noise) perturbations. In both cases, we show that the comblines remain stable while the solitons undergo translation. Given the tremendous impact that soliton perturbation theory has had on theoretical understanding of light-wave transmission systems, our goal is to provide a similar LLE combline perturbation theory for theoretical characterization of microresonators.
The paper is outlined as follows: In Sec. 2, the LLE is introduced along with the scalings to be used in our perturbation theory. Section 3 gives a brief overview of the perturbation theory to be used for modeling the microresonator. Section 4 and 5 present the Jacobi elliptic function solutions that satisfy the LLE equation and their detailed linear stability analysis respectively. The effects of two canonical perturbations due to Raman and simulated emission are considered in Sec. 6. Section 7 provides a brief summary and outlook for the theoretical method developed.
Lugiato Lefever equation
========================
The Lugiato-Lefever equation (LLE), which was originally derived in the context of detuned cavity resonators [@lugiato1987spatial], has been shown to describe the evolution of the electromagnetic field in microresonators [@PhysRevA.87.053852]. The LLE is a modification of the nonlinear Schrödinger equation (NLSE) which includes damping, detuning and a driving/pumping term. In dimensionless form, the LLE is given by the partial differential equation (PDE) $$\frac{\partial u}{\partial t}=-(\epsilon+i\alpha)u+i|u|^{2}u-i\frac{\beta}{2}\frac{\partial ^{2}u}{\partial x^{2}}+\epsilon F + \epsilon G(u,x,t),$$ where $u(x,t)$ the complex envelope of the total intracavity electric field, $\beta$ determines the microring dispersion ($\beta>0$ is normal group-velocity dispersion while $\beta<0$ is anomalous group-velocity dispersion), $\alpha$ is the cavity detuning parameter, $F$ characterizes the external cavity pumping, and $x\in [-\pi,\pi)$ since the microresonator enforces periodic boundary conditions [@PhysRevA.87.053852]. In our specific scaling, the parameter $\epsilon\ll 1$ is used to model the effects of linear cavity attenuation and small perturbations of the form $G(u,x,t)$ to the dominant balance dynamics of dispersion, Kerr self-phase modulation, and detuning.
In our scalings, the LLE can be written as a perturbed version of the detuned NLSE so that $$i\frac{\partial u}{\partial t} - \frac{\beta}{2}\frac{\partial ^{2}u}{\partial x^{2}} +|u|^{2}u -\alpha u=i\epsilon(F-u + G(u,x,t)).
\label{eq:lle}$$ This scaling allows us to develop a systematic perturbation analysis of previously unconsidered periodic, Jacobi elliptic solutions of the LLE. This complements the detailed stability analysis of Godey [*et al*]{}. [@PhysRevA.89.063814] which details the onset of a myriad of spatio-temporal patterns in the LLE model. Specifically, they show that the steady-state solutions of the LLE (with all temporal and spatial derivatives set to zero) lead to a host of pattern-forming instabilities [@cross1993pattern] that are ultimately responsible for the generation of strongly nonlinear periodic waveforms. In our analysis, we consider the stability of Jacobi elliptic solutions which are strongly nonlinear solutions whose dominant balance includes temporal and spatial derivative terms [@carr2000stationary; @bronski2001bose; @bronski1; @bronski2].
Background: Perturbation Theory
===============================
Our stability analysis determines the spectrum of the resulting linearized operator along with the effects of perturbations on the evolution of the solution parameters. In its most general form, we can consider the one dimensional PDE $$\frac{\partial u}{\partial t} = N(u, u_x, u_{xx}, \cdots, \mu)
+\epsilon G(u,x,t) \; ,$$ where $N(\cdot)$ represents some nonlinear dynamics (for which an analytical solution is known), $\epsilon G(u,x,t)$ is a perturbation to these dynamics, and $\mu$ is a (bifurcation) parameter. A multi-scale perturbation expansion [@bender2013advanced; @kevorkian2013perturbation] is a representation of the solution of the form $$u(x,t)=u_{0}(x,t,\tau)+\epsilon u_{1}(x,t) + \epsilon^2 u_2 (x,t) + \cdots
\; ,
\label{eq:perturb}$$ where $\tau=\epsilon t$ corresponds to a slow variable dependence [@kodama1981perturbations; @weinstein1985modulational].
Collecting terms at each order of $\epsilon$ gives nonlinear dynamics for the leading order term and forced, linear dynamics for all other orders, i.e.
&=& N (u\_0, [u\_0]{}\_x, [u\_0]{}\_[xx]{}, , ) ,\
&=& L\_1 (u\_0) u\_1 + F\_1 (u\_0) ,\
&=& L\_2 (u\_0) u\_2 + F\_2 (u\_0,u\_1) ,\
&&
where the first equation is the $O(1)$ balance, the second equation is the $O(\epsilon)$ balance and the third equations is the $O(\epsilon^2)$ balance. As in the approach of Weinstein [@weinstein1985modulational], we consider a solution of the leading order problem with slow-time modulation. Let $u_0(x,t)$ be given by $$u_0(x,t)= \Phi(x,t,A_1,A_2, \cdots) \; ,$$ where the parameters $A_i(\tau)$ vary with the slow time scale $\tau$. Applying the Fredholm alternative to the forced, linear PDE for $u_1$ requires that the forcing term $F_1$ be orthogonal to the generalized null space of the adjoint operator $L_1^\dag$, i.e. if $(L_1^\dag)^m v =0$ for some $m > 0$, then $$\langle v, F_1 \rangle = 0 \; ,
\label{eq:fredholm}$$ where $\langle u,v \rangle=\int_D u v^* \, dx$ is the inner product over the domain $D$. For a given perturbation, this constraint will result in equations for the slow evolution of the parameters $A_i$ of the form $$\frac{\partial {A_i}}{\partial \tau} = f_i(A_1,A_2,\cdots) \; .$$
Remarkably, in Weinstein’s analysis of the NLSE [@weinstein1985modulational], these constraints are all that needs to be satisfied to show that $\epsilon u_1$ is small for small values of $\epsilon$ up to times of order $1/\epsilon$. Similar results hold for elliptic function solutions of the NLSE, which we outline in the following. We will show that the additional terms in the LLE, when viewed as a perturbation of the NLSE, have a stabilizing effect on the parameters of $\operatorname{dn}$ type solutions. Further, we provide expressions for the evolution of the parameters for two particular cavity perturbations of the LLE: the Raman effect and spontaneous emission noise.
Jacobi Elliptic Functions for the NLSE
======================================
The Jacobi elliptic functions are periodic wavefunctions that satisfy the NLSE with detuning [@carr2000stationary; @bronski2001bose; @bronski1; @bronski2], i.e. the leading order dynamics as described by . The three basic functions are denoted $\operatorname{sn}(x|k)$, $\operatorname{cn}(x|k)$, and $\operatorname{dn}(x|k)$, where the elliptic modulus, $k$, parameterizes the solutions. The value of $k$ is constrained such that $k \in [0,1)$; we note that the reader may be more familiar with the parameter $m=k^2$, which is commonly used in software for evaluating the Jacobi elliptic functions.
The stability of these solutions is well-studied. For the defocusing case, the $\operatorname{sn}$ solutions are known to be modulationally stable [@Bernard]. For the focusing case, the $\operatorname{cn}$ and $\operatorname{dn}$ solutions are modulationally unstable [@deconinck2017stability]. Recent research has shown the spectral stability of the $\operatorname{dn}$ solution under perturbations with a period equal to the fundamental period, but not under perturbations with a period equal to a multiple of the fundamental period [@stabilityDn]. [ Spectral stability of the $\operatorname{cn}(x|k)$ solution only holds when $k \in (0, k_{c})$ under perturbations with a period equal to the fundamental period, with $k_{c} \approx 0.908$ [@stabilityDn].]{} In-depth discussion of the stability properties of Jacobi elliptic function solutions of the NLSE can be found in [@Bernard; @deconinck2017stability; @stabilityDn].
![ Numerical simulation of the (a) $\operatorname{cn}$ and (b) $\operatorname{sn}$ solutions [ of with $|\beta| = 0.01$, $\epsilon=0.1$, $G=0$, and the detuning $\alpha$ set to (a) $\alpha = 1.8732$ and (b) $\alpha = 3.7464$ (these values of the detuning are chosen so that $k^{2}=1-10^{-12}\approx 1$ in the analog of for these solutions).]{} The solutions were seeded with a white noise perturbation to induce instability in the evolution. Both solutions are unstable, even in the limit $k\rightarrow 1$ where the linear stability analysis shows the eigenvalues to shrink to the real axis. Note that the $\operatorname{cn}$ solution collapses from an $N=4$ solution to a stable $N=2$ $\operatorname{dn}$ solution.[]{data-label="fig:cn_sn"}](Wave_Cn_Sn_Epsilon=01_2){width="\linewidth"}
With the addition of the LLE terms, i.e. the damping and forcing of the microresonator, the $\operatorname{cn}$ and $\operatorname{sn}$ solutions are unstable in their respective regimes. In Figure \[fig:cn\_sn\], we plot a numerical simulation of the evolution of $\operatorname{cn}$ and $\operatorname{sn}$ wave forms (with four [pulses]{}) governed by the LLE. The $\operatorname{sn}$ wave form quickly decays and the $\operatorname{cn}$ wave form evolves into a solution of $\operatorname{dn}$ type (with two [pulses]{}). It appears that the LLE does not support [pulses]{} that are separated by a node, i.e. those with a $\pi$ phase change between neighboring pulses. In contrast with its instability as a solution of the NLSE, the $\operatorname{dn}$ type solutions of the LLE are in fact stable, even with multiple [pulses]{} in the cavity. We will show that this stability can be understood analytically and we will focus on the $\operatorname{dn}$ type solutions for the remainder of the manuscript.
Solutions of $\operatorname{dn}$ type: anomalous dispersion {#solutions-of-operatornamedn-type-anomalous-dispersion .unnumbered}
-----------------------------------------------------------
The $\operatorname{dn}$ solution is of the most practical importance, as it is the only stable solution we find for the LLE in the anomalous dispersion regime [($\beta < 0$)]{}. For this solution, we assume the general form $$u_{0}(x,t)={{\hat u}_{0}}e^{i\psi}=A\operatorname{dn}(B(x-x_{0})|k)e^{i[\xi(x-x_{0})+\sigma-\sigma_{0}]},
\label{eq:dn}$$ where $A^{2}=-\beta B^{2}$, and
&& =-,\
&& =--B\^[2]{}(2-k\^[2]{})-\^[2]{}.
Since the wavefunctions of the LLE should be ${2\pi}/{N}$ periodic, where $N$ is a positive integer, the value of $B$ determines the value of $k$ and vice-versa. Specifically, the period of the Jacobi elliptic function $y=\operatorname{dn}(x|k)$ is $2K$, where $K(k)$ is the elliptic integral of the first kind. So the period of $\hat{u}_{0}=\operatorname{dn}(Bx|k)$ should be $T={2K}/{B}$. If $T={2\pi}/{N}$, then we have ${2K}/{B}={2\pi}/{N}$, thus $B={KN}/{\pi}$. Note that $N$ is the number of localized ([pulses]{}) per round trip in the microresonator.
[.23]{}
[4Dn\_kSmall.pdf]{} (17,14)[(a)]{}
[.23]{}
[4Dn\_kLarge.pdf]{} (17,14)[(b)]{}
Figure \[fig:dn\] shows the $\operatorname{dn}$ solution for two values of the parameter $k$, where $k \in [0,1)$. These figures are illustrated with $N=4$ so that four [pulses]{} are shown around the cavity. In the limit $k\rightarrow 1$, the function $\operatorname{dn}(x|k) \rightarrow \operatorname{sech}(x)$, which is the standard hyperbolic secant soliton solution generated by the dominant NLSE terms. When $k\rightarrow 0$, the function $\operatorname{dn}(x|k)\rightarrow 1$, which is a continuous wave solution of the LLE. The figure illustrates the $k^2=0.9$ and $k^{2}=1-10^{-12}\approx 1$ solutions of the LLE.
[.23]{}
[1Dn\_kLarge.pdf]{}
[.23]{}
[2Dn\_kLarge.pdf]{}
[.23]{}
[3Dn\_kLarge.pdf]{} (17,76)[(a)]{} (17,14)[(c)]{}
\[fig:sfig3\]
[.23]{}
[4Dn\_kLarge.pdf]{} (17,76)[(b)]{} (17,14)[(d)]{}
\[fig:sfig4\]
Figure \[fig:dn2\] shows the $\operatorname{dn}$ solutions as the parameter $N$ is varied from one to four. [Based on the observed behavior of these solutions of the LLE in numerical simulations, we consider solutions about the center frequency, $\xi=0$, and with a fixed phase term, which can be obtained by setting]{} $${\alpha = -{\beta}B^{2}(2-k^{2})/2.} \label{eq:alpha}$$ With these choices, the value of the detuning must be increased in order to accommodate more pulses per round trip, which is consistent with experimental findings.
[Dn\_EnergyVsAlpha]{} (92,12.5)[$N\!=\!1$]{} (92,18)[$N\!=\!2$]{} (94,38) (92,63)[$N\!=\!10$]{}
Importantly, we can compute the cavity energy $e_c$ versus detuning frequency $\alpha$ for the $\operatorname{dn}$ solutions by the definition of the cavity energy as $$e_c=\int_{-\pi}^{\pi}|u_{0}|^{2}dx=-\beta B^{2}\int_{-\pi}^{\pi}\operatorname{dn}^{2}(By|k)dy \, .$$ The energy of each solution branch can then be computed for different $N$ values as shown in Fig. \[fig:dn3\]. The stability of each branch will be discussed in what follows, but the energy versus detuning shows the important trends to be considered. For $k \rightarrow 1$, the function $\operatorname{dn}(x|k) \rightarrow \operatorname{sech}(x)$ so that the energy integral can be approximated explicitly $${e_c \approx -\beta B\int_{-B\pi}^{B\pi}\operatorname{sech}^{2}zdz= -2\beta B \; .}$$ [Given that $\alpha \approx -\beta B^{2}/2$, we can then]{} simplify the relationship between the detuning and cavity energy, i.e. $|e_c/\beta|\approx 2\sqrt{2}\sqrt{|\alpha/\beta|}$. This value is for only a single [pulse]{}. If there are $N$ [pulses]{}, we obtain $$|e_c/\beta|=2\sqrt{2}N\sqrt{|\alpha/\beta|}.$$ This gives a simple quantization of the energy as a function of the number of pulses in the limit $k\rightarrow 1$. We will show in what follows that the $k\rightarrow 1$ limit is where solutions to the LLE are stable, thus the energy quantization formula is a good approximation for the LLE microresonator dynamics. [Note that in Fig. \[fig:dn3\], since $|\alpha/\beta|=B^{2}/2$, we have $|\alpha/\beta| \rightarrow \infty$ when $k \rightarrow 1$ and $|\alpha/\beta| \rightarrow 0$ when $k \rightarrow 0$.]{}
Stability Analysis of the LLE
=============================
The stability of the Jacobi elliptic function solutions to the LLE can be characterized using a linear stability analysis. Let $u_1 = e^{i\psi} w_1$. Following the perturbation expansion of , we find at leading order the Jacobi elliptic solutions and at $O(\epsilon)$ the linearized evolution $$\begin{aligned}
\hat{F} &= i\frac{\partial w_1}{\partial t}-\alpha w_1
+2|u_{0}|^{2}w_1 -\frac{\beta}{2}\frac{\partial ^{2}w_1}{\partial x^{2}}
+|u_{0}|^{2}w_1^{*} \; ,\end{aligned}$$ where $\hat{F}=i( e^{-i\psi} F + e^{-i\psi}G(u_0,x,t) - \hat{u}_{0}
- e^{-i\psi}u_{0\tau})$.
We can decompose the linearized evolution into real and imaginary components by letting $w_1=R+iI$ ($w_1^{*}=R-iI$) so that in matrix notation it takes the form $$\begin{aligned}
\left[\!\!
\begin{array}{c}
R_{t}\\
I_{t}
\end{array}
\!\!\right] \!\!&=\!\!
\left[\!\!
\begin{array}{cc}
0 &
\frac{\beta}{2}\partial^2_{x} -{{\hat u}_{0}}^{2}
+ \alpha\\
-\frac{\beta}{2}\partial^2_{x} +3{{\hat u}_{0}}^{2}
-\alpha &
0
\end{array}
\!\!\right]
\!\!
\left[\!\!
\begin{array}{c}
R\\
I
\end{array}
\!\!\right]
\!+\!\left[ \!\!
\begin{array}{c}
\operatorname{Im}\hat{F}\\
\operatorname{Re}\hat{F}
\end{array}
\!\!\right]
\label{eq:spec} \; ,\end{aligned}$$ where $\partial^2_x$ denotes the second order derivative. [The eigenvalue spectrum of the matrix in yields the spectral stability of $\operatorname{dn}$ solutions, generally. Note that for ${\hat u}_{0}$ given by the $\operatorname{dn}$ solution, $\alpha=-{\beta}B^{2}(2-k^{2})/2$.]{}
![Eigenvalue spectrum of the matrix ${\mathbf{L}}$ defined in for the $\operatorname{dn}$ solution with (a) $k^2=0.9$ and (b) $k^{2}=1-10^{-12}\approx 1$. Although the eigenvalues shrink to the imaginary axis as $k\rightarrow 1$, the solutions are known to be unstable under generic perturbations.[]{data-label="fig:dn_spec"}](Eigen_Dn){width="\linewidth"}
Figure \[fig:dn\_spec\] shows the computed spectrum of the linearized operator in for the $\operatorname{dn}$ solution with $N=4$. The operator was numerically evaluated using a spectrally accurate method with 1024 grid points (a fast Fourier transform was used to evaluate the second derivatives) and a standard matrix eigenvalue solver. The eigenvalues corresponding to both $k^2=0.9$ and $k^2=1-10^{-12} \approx 1$ are evaluated. [Note for the case $N=4$, the fundamental period $T=\frac{2\pi}{4}$, thus $[-\pi, \pi]$ is a multiple of the fundamental period, so we expect instability [@stabilityDn].]{} For $k^2=0.9$, [the $\operatorname{dn}$]{} solution clearly has unstable eigenvalues, i.e. eigenvalues with large positive real part. As $k\rightarrow 1$, the real part of the eigenvalues [of $\operatorname{dn}$]{} shrink to the imaginary axis, suggesting that the $k\rightarrow 1$ solutions [will be better behaved, even if they are technically unstable [@stabilityDn].]{} Thus a critical part of the analysis is to determine if the addition of the LLE term $F$ stabilizes such microresonator solutions subject to slow-time modulation of the parameters.
Discrete Spectrum of $\operatorname{dn}$ Solutions
--------------------------------------------------
As with soliton perturbation theory, the generalized null space of the adjoint of the linearized operator is critical for determining stability. Specifically, the Fredholm-alternative theorem stated in requires that perturbations be orthogonal to the null space of the adjoint linear operator (this removes so-called “secular” modes which have polynomial growth in time [@weinstein1985modulational]). For the $\operatorname{dn}$ solution, the linear operator reduces to $${\bf L}=-\frac{\beta}{2}B^2 \left[ \begin{array}{cc} 0 & L_- \\ -L_+ & 0 \end{array} \right],
\label{eq:Lmatrix}$$ with the self-adjoint operators
&& L\_- = - -2 \^2 z - k\^[2]{} + 2 ,\
&& L\_+ = - - 6 \^2 z - k\^[2]{} + 2 ,
where a change of variables to $z=B(x-x_0)$ has been made and the dependence of $\operatorname{dn}$ on $k$ has been dropped for notational convenience.
In the following, we will denote the generalized nullspace of a linear operator $L$ by $\ker_g(L)$, i.e.
$$\ker_g(L) = \bigcup_{m=1}^\infty \ker(L^m) \; .$$
We also require the space [$H^m_{{\textnormal{per}}}[a,b)$]{}, which denotes a periodic Sobolev space on [$[a,b)$]{}. This space may be characterized by the Fourier coefficients of a given function. Let $f$ be a function [defined on $[a,b) = [-\pi,\pi)$]{} and let $c_j$ defined by
$$c_j = \frac{1}{2\pi} \int_{-\pi}^\pi f(x) e^{-i j x} \, dx \; .$$
Then the [$H^m_{{\textnormal{per}}}[-\pi,\pi)$]{} norm is defined by
$${ \| f \|^2_{H^m_{{\textnormal{per}}}} = \sum_j |c_j|^2 (1+|j|^2+\cdots |j|^{2m}) \; .}$$
Note that if [$\| f\|_{H^m_{{\textnormal{per}}}} < \infty$]{}, then the Fourier coefficients of the [$(m-1)st$ derivative of]{} $f$ are absolutely summable so that [$f^{(m-1)}$]{} is continuous as a periodic function on $[-\pi,\pi)$. The space [$H^m_{{\textnormal{per}}}$]{} can be defined for other intervals by appropriate scaling.
Let $w_1 = R + i I$ as above. As in [@weinstein1985modulational], we define the space $$\mathcal{M} = H^1_{{\textnormal{per}}}\times H^1_{{\textnormal{per}}}\bigcap \left (\ker_g( {\bf L}^\dag ) \right )^\perp \; ,$$ which is where we will constrain the evolution of $(R,I)^\intercal$. [Note that the domain for $z$ is $[-NK(k),NK(k))$]{}. We also define the periodic functions $\phi(z)$ and $\varphi(z)$ to be
&& (z)=(K(k)E(z,k)- E(k)z) z-k\^[2]{}K(k)zz,\
&& (z)=k\^[2]{} z z (K(k)E(z,k)-E(k)z)\
&& +(E(k)-K(k)) z+k\^[2]{}K(k)\^[2]{} zz ,
where [$E(z,k)$]{}$=\int_{0}^{z}\operatorname{dn}^{2} ydy$ is the incomplete elliptic integral of the second kind, $E(k) = E(K(k),k)$ is the complete elliptic integral of the second kind, and $K(k)$ is as above. For the sake of compactness, we will often drop the dependence of $E(k)$ and $K(k)$ on the modulus $k$ in the following. [Note that $E(z,k)$ is odd, $\phi(z)$ is odd, and $\varphi(z)$ is even — the parity of functions simplifies much of the following analysis.]{} A set of eigenfunctions that span $\ker_g({\bf L}^\dag)$ can then be computed from the following observations
&& L\_[-]{}\[z\]=0,\
&& L\_[+]{}\[zz\]=0,\
&&\
&&\
These results are used to derive some important properties of the operators ${\bf L}^\dag$, $L_+$, and $L_-$, which are summarized in propositions \[prop:lplmnull\] and \[prop:geneig\]. Proofs are included in the appendix.
\[prop:lplmnull\] [Assume $N\in \mathbb{N}$ and]{} $0<k<1$. The operator $L_-$ is non-negative and self-adjoint, with $\ker(L_-) = \operatorname{span}\{ \operatorname{dn}z \}$. The operator $L_+$ is self-adjoint, with $\ker(L_+) = \operatorname{span}\{ \operatorname{sn}z \operatorname{cn}z \}$.
\[prop:geneig\] [Assume $N\in \mathbb{N}$ and]{} $0<k<1$ and let $(f,g)^\intercal \in H^1_{{\textnormal{per}}}\times H^1_{{\textnormal{per}}}$. If the following orthogonality relations hold
f, z &= 0 ,\
f, [(z)]{} &= 0 ,\
g, z z &= 0 ,\
g, [(z)]{} &= 0 ,
then $(f,g)^\intercal \in \mathcal{M}$.
Bounding the evolution of $w_1$ [($N=1$)]{}
-------------------------------------------
Following the analysis of Weinstein [@weinstein1985modulational], the evolution of the term $w_1 = R + iI$ is bounded by considering the function $$Q(f,g) =-\frac{\beta}{2}B^{2}[\langle L_{+}f,f \rangle
+ \langle L_{-}g,g \rangle ] \; ,$$ which is a conserved quantity along the solution trajectory for $w_1$, i.e. $dQ(R,I)/dt=0$. For $(R,I)^\intercal \in \mathcal{M}$, we have the following bound.
\[prop:h1control\] [Assume $N=1$]{}. Let $w = R +iI \in \mathcal{M}$. Then there exist constants $C_1$ and $C_2$ such that $$\begin{aligned}
C_1 \left ( \|R\|^2_{H^1_{{\textnormal{per}}}} + \|I\|^2_{H^1_{{\textnormal{per}}}} \right) &\leq&
Q(R,I) \; ,\\
C_2 \left ( \|R\|^2_{H^1_{{\textnormal{per}}}} + \|I\|^2_{H^1_{{\textnormal{per}}}} \right) &\geq&
Q(R,I) \; .
\end{aligned}$$
This proposition is the primary result needed in our analysis: if the slow evolution of the parameters $B$, $\xi$, $x_0$, and $\sigma$ is such that $(R(t),I(t))^\intercal \in \mathcal{M}$, then for any $T_0$ we have [$\sup_{0 \leq t \leq T_0/\epsilon}$]{} $\| \epsilon w_1(t) \|_{H^1_{{\textnormal{per}}}}
\to 0$ as $\epsilon \to 0$. See [@weinstein1985modulational] for details. A proof of Proposition \[prop:h1control\] based on a variational formulation can be found in [@stabilityDn]. A more classical proof modeled after [@weinstein1985modulational] is provided in the appendix.
Modulation equations for the $\operatorname{dn}$ solution [($N=1$)]{}
---------------------------------------------------------------------
For the $\operatorname{dn}$ solution, $$\begin{aligned}
\partial_\tau \left (\hat{u}_0 \right) &= \sqrt{-\beta} B_{\tau}\operatorname{dn}z
+ A \frac{dk}{dB} B_\tau \operatorname{dn}_{k} z \nonumber\\
&\qquad+ A [B_{\tau}(x-x_{0})-Bx_{0\tau}] \operatorname{dn}_{z} z\; .\end{aligned}$$ We consider solutions with $k \rightarrow 1$, so that $dk/dB \approx 0$. Using this approximation, the above reduces to
$$\begin{aligned}
\partial_\tau \left (\hat{u}_0 \right) &= \sqrt{-\beta}B_{\tau} \operatorname{dn}z- Ak^{2}
\operatorname{sn}z \operatorname{cn}z \left (\frac{B_{\tau}}{B}z-Bx_{0\tau} \right ) \; ,\end{aligned}$$
so that
$$\begin{aligned}
\partial_\tau \left( u_0 \right ) &=
e^{i\psi} \sqrt{-\beta}B_{\tau} \operatorname{dn}z- e^{i\psi}Ak^{2}
\operatorname{sn}z \operatorname{cn}z \left (\frac{B_{\tau}}{B}z-Bx_{0\tau} \right )
\nonumber\\
&\qquad +i \left (\xi_\tau \frac{z}{B} - \xi x_{0\tau} - \sigma_{0\tau}
\right ) e^{i\psi} A \operatorname{dn}z\end{aligned}$$
This gives the following expression for the forcing term
$$\begin{aligned}
\operatorname{Im}\hat{F} &= F \cos \psi + \operatorname{Re}(e^{-i\psi}G)
-\sqrt{-\beta}B_{\tau} \operatorname{dn}z - A \operatorname{dn}z \nonumber \\
& \qquad +Ak^{2} \left (\frac{B_{\tau}}{B}z
-Bx_{0\tau} \right ) \operatorname{sn}z \operatorname{cn}z,\\
\operatorname{Re}\hat{F} &= F\sin \psi -\operatorname{Im}(e^{-i\psi}G)
+ \sqrt{-\beta}\xi_{\tau}z \operatorname{dn}z \nonumber \\
& \qquad -(\xi x_{0\tau}+\sigma_{0\tau})A \operatorname{dn}z \; .\end{aligned}$$
To constrain the forcing term $(\operatorname{Im}\hat{F}, \operatorname{Re}\hat{F})^\intercal$ to be in $\mathcal{M}$, Proposition \[prop:geneig\] implies the following constraints
&& ,z= 0,\
&& ,[(z)]{} = 0,\
&& ,zz= 0,\
&& ,[(z)]{} = 0.
These constraints require the slow evolution of the parameters to satisfy the following system of differential equations
&&= ,\
&& ,\
&&=- ,\
&& , \[eq:pert2\]
where $\psi= {\xi}z/B+\sigma-\sigma_{0}$. We can further simplify the equations above by applying trigonometric identities, we have
$$\begin{aligned}
&\frac{dB}{d\tau}=\frac{\langle\operatorname{Re}(e^{-i\psi}G),\operatorname{dn}z\rangle+F\cos(\sigma-\sigma_{0})p_{1}(\xi)}{\sqrt{|\beta|}(q_1(k)-k^2q_2(k))} \nonumber\\
& \qquad\qquad -\frac{\sqrt{|\beta|}B\langle\operatorname{dn}z, \operatorname{dn}z \rangle}{\sqrt{|\beta|}(q_1(k)-k^2q_2(k)} \; , \slabel{eq:pert2a} \\
&{\frac{dx_{0}}{d\tau}=\frac{\langle \operatorname{Re}(e^{-i\psi}G), \phi(z) \rangle -F\sin(\sigma-\sigma_{0})p_{2}(\xi)}{\sqrt{|\beta|}B^{2}k^2q_3(k)}} \; ,\\
&\frac{d\xi}{d\tau}=\frac{\langle \operatorname{Im}(e^{-i\psi}G), \operatorname{sn}z\operatorname{cn}z\rangle-F\cos(\sigma-\sigma_{0})p_{3}(\xi)}{\sqrt{|\beta|}q_2(k)} \; , \\
&{\frac{d\sigma_{0}}{d\tau}+\xi\frac{dx_{0}}{d\tau} =\frac{F\sin(\sigma-\sigma_{0})p_{4}(\xi)-\langle \operatorname{Im}(e^{-i\psi}G), \varphi(z)\rangle}{\sqrt{|\beta|}Bq_4(k)}} \, , \nonumber \\
\label{eq:pert2}\end{aligned}$$
where
\[eq:xiterms\] $$\begin{aligned}
p_1(\xi) &= \langle \cos(\xi z/B), \operatorname{dn}z \rangle \; ,\\
p_2(\xi) &= {\langle \sin(\xi z/B), \phi(z) \rangle} \; ,\\
p_3(\xi) &= \langle \sin(\xi z/B), \operatorname{sn}z \operatorname{cn}z \rangle \; ,\\
p_4(\xi) &= {\langle \cos(\xi z/B), \varphi(z) \rangle} \; ,\end{aligned}$$
and
$$\begin{aligned}
q_1(k) &= \langle \operatorname{dn}z,\operatorname{dn}z\rangle \\
q_2(k) &= \langle z \operatorname{dn}z,\operatorname{sn}z \operatorname{cn}z\rangle \\
q_3(k) &= {\langle \operatorname{sn}z\operatorname{cn}z, \phi(z)\rangle} \\
q_4(k) &= {\langle \operatorname{dn}z, \varphi(z)\rangle} \; .\end{aligned}$$
In addition to these constraints, $\xi$ should be an integer so that $u_0$ remains in $H^1_{{\textnormal{per}}}$. Therefore the analysis is only rigorous when applied to perturbations for which $d\xi/d\tau = 0$, but we have found that the analysis provides insight in other cases.
Consider the stability of this system of differential equations around the center frequency, i.e. $\xi=0$, so that $\psi= \sigma-\sigma_{0}$ and $p_{2}(\xi)=p_{3}(\xi)=0$. For $k \approx 1$, we can approximate many of the inner products in the above expressions using the limiting forms of the Jacobi elliptic functions. We obtain the approximations $p_1(0) \approx \pi$, $p_4(0) \approx 0$, $q_1(k) \approx 2$, $q_2(k) \approx 1$, [$q_3(k) \approx -1$, and $q_4(k) \approx 1$.]{} With these approximations, the evolution equations simplify to
&&=((e\^[-i]{}G),z+F(-\_[0]{}))-2B ,\
&&=- ,\
&&=(e\^[-i]{}G),zz ,\
&&+=[-]{} . \[eq:pert2\]
These slow evolution equations approximate the effect of a perturbation $G$ on the microresonator comb.
Stability of $\operatorname{dn}$ solutions of the LLE [($N=1$)]{}
-----------------------------------------------------------------
When the perturbation $G=0$, the parameter evolution constraints yield the following set of equations
&& =-2B ,\
&& =0 ,\
&& =0 ,\
&&+=0 ,
which gives the solution with $B=F\pi \cos(\sigma-\sigma_{0})/(2 \sqrt{|\beta|})$ as a steady-state attractor to the dynamics. Specifically, values of $B$ larger and smaller than this exponentially decay back to the steady-state value. In addition, the fast time scale dynamics give
&& =-,\
&& =--B\^[2]{}(2-k\^[2]{})-\^[2]{}.
Integrating the constant $\sigma_0$ into the second equation and setting $\xi=0$, i.e. we are working around the center frequency, then $$\frac{d(\sigma-\sigma_{0})}{dt}=-\alpha-\frac{\beta}{2}B^{2}(2-k^{2}).$$ Since the first solvability condition gives the steady-state $B={F}\pi \cos(\sigma-\sigma_{0})/({2\sqrt{-\beta}})$, then $$\frac{d(\sigma-\sigma_{0})}{dt}={-\alpha+\frac{F^{2}\pi^{2}(2-k^{2})}{8}\cos^{2}(\sigma-\sigma_{0})} \; ,$$ and $$\label{eq:phase_predict}
{\cos(\sigma-\sigma_{0}) = \frac{2B\sqrt{-\beta}}{F\pi}} \; ,$$ which gives the time-independent phase of the microresonator comb. Specifically, the real part of the solution is $u_{0}=A\operatorname{dn}(B(x-x_{0}),k)\cos(\sigma-\sigma_{0})$ and the imaginary part is $u_{0}=A\operatorname{dn}(B(x-x_{0}),k)\sin(\sigma-\sigma_{0})$.
These asymptotic results show that $B\neq 0$ provided $F>0$. Moreover, the stable microresonator solution has a fixed phase relation which does not evolve in time. Simulations show that these two predictions are accurate representations of the dynamics. More than that, the prediction here shows them to be attractors for general initial conditions, which is again borne out by simulation.
The case $N>1$ in practice
--------------------------
Proposition \[prop:lplmnull\] states that the nullspaces of $L_+$ and $L_-$ are each spanned by one function, for any positive number of pulses $N$. However, in simulations of finite precision, this mathematical truth is not observed for $k\approx 1$. Indeed, the discretizations of these operators are observed to have $N$ eigenfunctions corresponding to a zero (to numerical precision) eigenvalue when $k\approx 1$. Intuitively, this results from the fact that the $\operatorname{dn}$ function is nearly zero between pulses for $k\approx 1$ so that the pulses are essentially decoupled. Indeed, the set of $N$ shifted copies of the eigenfunctions for $N=1$, i.e. individual pulses in each of $[-K,K),[K,3K),\ldots,[(2N-3)K,(2N-1)K)$, is seen to give a basis for these nullspaces, again to numerical precision.
Counterintuitively, it is this failure of Proposition \[prop:lplmnull\] in practice which explains the predictive power of the modulation equations of the previous sections — which hold mathematically only when $N=1$ — for numerical simulations with $N>1$. In particular, for the LLE type perturbation alone ($G=0$), we observe that the stabilizing effect on the amplitude of the comb as predicted by and the generation of a time independent phase as predicted by for the $N=1$ case also hold for $N>1$ when $k\approx 1$. See Figures \[fig:dn\_evolve1\] and \[fig:dn\_evolve0\] for a comparison of the stability of a $\operatorname{dn}$ initial condition with and without the LLE terms, which we discuss in more detail in the next section. For nonzero $G$ perturbations, as in the Raman effect and spontaneous emission noise examples below, the more qualitative $N=1$ predictions are also observed numerically when $N>1$. The fact that the behaviors of the pulses have decoupled is only apparent for the spontaneous emission noise example, as the perturbations acting on each pulse are identical in the other examples.
![Numerical simulation [of the ($N=4$) $\operatorname{dn}$ solution of with $\epsilon=0.1$, $G=0$, and the detuning $\alpha$ set to (a) $\alpha = 0.0593$ and (b) $\alpha = 1.8732$ so that the appropriate parameter for the initial waveform is (a) $k^2=0.9$ and (b) $k^{2}=1-10^{-12}\approx 1$, according to ]{}. The [initial waveform was corrupted with white noise]{} to induce instability in the evolution. The $k^2=0.9$ solution is shown to be unstable whereas the $k^2=1-10^{-12}$ solution is stable. This is consistent with our linear stability analysis and Fig. \[fig:dn3\].[]{data-label="fig:dn_evolve1"}](Wave_Dn_Epsilon=01_Noise=3){width="\linewidth"}
![[Numerical simulation of the ($N=4$) $\operatorname{dn}$ solution of with $\epsilon=0$, $G=0$, and the detuning $\alpha$ set to (a) $\alpha = 0.0593$ and (b) $\alpha = 1.8732$ so that the appropriate parameter for the initial waveform is (a) $k^2=0.9$ and (b) $k^{2}=1-10^{-12}\approx 1$, according to ]{}. [The initial waveform was corrupted with the same white noise as Fig. \[fig:dn\_evolve1\] to induce instability in the evolution. Both the $k^2=0.9$ solution and the $k^2=1-10^{-12}$ solution are unstable.]{} [Comparing with Fig. \[fig:dn\_evolve1\], we observe that the $k^{2}=1-10^{-12}$ solution is stabilized by the LLE type perturbation.]{}[]{data-label="fig:dn_evolve0"}](Wave_Dn_Epsilon=0_Noise=3){width="\linewidth"}
![Stable numerical solutions of [ with $G=0$ and $\alpha=1.7793$ for various values of $\epsilon$.]{} [The initial waveform was set as a ($N=3$) $\operatorname{dn}$ solution with $k^2=1-10^{-16}$.]{} As $\epsilon$ is increased from $\epsilon\ll 1$, the solutions deform from the asymptotic $\operatorname{dn}$ form to a localized structure that sits atop a shelf. Importantly, like the $\operatorname{dn}$ solution, the resulting evolution produces solutions which have no nodal separation between neighboring pulses. []{data-label="fig:dn_evolve2"}](3Dn_kLarge_epsilonChanging){width="\linewidth"}
[.25]{}
[3Dn\_k=12\_rho=095\_epsilon=01]{} (19,13)[(a) $\epsilon=0.1$]{}
[.25]{}
[3Dn\_k=12\_rho=095\_epsilon=1]{} (19,13)[(b) $\epsilon=1$]{}
Numerical Simulations
=====================
[In this section, we compare numerical simulations of with predictions made by the theory outlined above. In all simulations the value of $\beta$ is fixed, with $\beta = -0.01$. When $\epsilon \neq 0$, we set $F=(\rho(1+(\rho-\alpha)^2))^{1/2}$ with $\rho=0.95$ to remain in the right parameter space for the generation of frequency combs. The initial waveforms ($u$ at time zero) are set according to with $\xi=\sigma=\sigma_0=x_0=0$ and the value of $k$ determined by the detuning $\alpha$ as in .]{} First, we simulate the LLE to show that the $\operatorname{dn}$ solution is stable, as opposed to the observed instability of the $\operatorname{cn}$ and $\operatorname{sn}$ solutions in Figure \[fig:cn\_sn\]. Figure \[fig:dn\_evolve1\] shows the evolution of the $\operatorname{dn}$ solution for $\epsilon=0.1$ [and the detuning $\alpha$ chosen so that]{} $k^2=0.9$ and $k^2=1-10^{-12}\approx 1$. Recall that for $k^2=0.9$ the linear stability analysis showed strong instability and for $k^2=1-10^{-12}$ the analysis showed weaker instability, see Figure \[fig:dn\_spec\]. Further, recall that for both values of the parameter, the solution should be unstable for generic perturbations of the equation. [In both simulations, the initial waveforms are corrupted]{} with white noise in order to induce instability if it exists. For $k^2 = 1-10^{-12}$, the pumping and damping terms of the LLE, i.e. the LLE-specific perturbations, have stabilized the $\operatorname{dn}$ solution. The $k^2 = 0.9$ solution is still unstable with this perturbation.
In Fig. \[fig:dn\_evolve0\], we repeat these calculations without the LLE perturbation, i.e. setting $\epsilon=0$. The $k^2=1-10^{-12}$ solution is seen to be less stable than that in Fig. \[fig:dn\_evolve1\].
In Fig. \[fig:dn\_evolve2\], we plot equilibrated solutions of as $\epsilon$ is increased. In this example, the ($N=3$) $\operatorname{dn}$-type solutions for $k^2=1-10^{-16}$ remain stable, even for large values of $\epsilon$. Note that the solutions deform away from the original $\operatorname{dn}$ waveform and develop a pedestal as $\epsilon$ is increased.
[Finally, Fig. \[fig:dn\_evolve3\] contains plots of the predicted time-independent]{} phase, determined by , and the phase of a simulated microresonator solution with a $\operatorname{dn}$ initial waveform, showing good agreement between theory and simulation.
![[ Top view of a numerical simulation of with $\alpha=1.8732$ and the addition of the Raman effect of . The perturbation parameter $\epsilon$ is set to (a) $\epsilon=0.1$ and (b) $\epsilon=1$. As predicted and quantified by our perturbation theory, the $\operatorname{dn}$ solution remains stable despite the induced drift of the solution. The drift velocity is compared with that computed from our theory using self-frequency shift in both cases. The dotted lines represent the theoretically calculated trajectories of the drift of the solutions. The perturbation theory holds well when $\epsilon$ is small.]{}[]{data-label="fig:Raman"}](Raman_comparison){width="\linewidth"}
Raman term
----------
An important modification to the LLE equation is the addition of the Raman effect which is known to induce a self-frequency shift in the microresonator [@karpov2016raman; @milian2015solitons]. The Raman effect is included in the LLE as part of the perturbation term $G(u,x,t)$ in . [Letting $U$ denote the waveform and $\mathfrak{G}(U)$ denote the Raman perturbation in physical units, we have]{} [@karpov2016raman] $${\mathfrak{G}(U) = i \left[-f_R|U|^2 + f_{R}h_{R}\bigotimes |U|^{2} \right] U \approx -i \left[ f_{R}\tau_{R} \frac{\partial |U|^{2}}{\partial x} \right] U},
\label{eq:GRaman}$$ where the constants $f_{R}$ and $\tau_{R}$ are the Raman fraction and the Raman shock time, respectively, and $\bigotimes$ denotes a convolution. In simulations, the Raman response function $h_{R}$ is typically chosen to be [@blow1989theoretical] $$h_{R}(x)=\frac{\tau_{1}^{2}+\tau_{2}^{2}}{\tau_{1}\tau_{2}^{2}}e^{-x/\tau_{2}}\sin(x/\tau_{1}),$$ where $\tau_{1}=12.2$fs and $\tau_{2}=32$fs. [In our numerical simulation of the dimensionless LLE, , the Raman term becomes $G(u)=-iC \frac{\partial |u|^{2}}{\partial x}u$, where $C=0.001$.]{}
The effect of the Raman perturbation of can be substituted into the modulation constraints of to evaluate the impact on the comb dynamics. The symmetry properties of the perturbation play a large role in determining the resulting behavior. Specifically, symmetry considerations yield
&& =0 ,\
&& =0 ,
with the additional constraints that
&& =-2B ,\
&& = 0
This determines the self-frequency shift induced by the Raman term since the value of $\xi$ gives the shift from the center frequency used to derive the LLE. In addition to the self-frequency shift, it should be recalled that $$\frac{dx_{0}}{dt}=-\beta \xi \, .$$ As the term $\xi$ is slowly evolving, it can be thought of as a constant over short time intervals so that the self-frequency shift generates a linear translation of the solution with a group velocity determined by the Raman term. Importantly, the Raman term [*does not*]{} destabilize the comb, rather it simply shifts it in frequency and forces a translation.
[ In Fig. \[fig:Raman\], we plot simulations of the LLE with the addition of the Raman effect, i.e. $G(u)=-iC \frac{\partial |u|^{2}}{\partial x}u$, for both $\epsilon=0.1$ and $\epsilon=1$. The comb quickly forms and the induced translation is readily apparent. We also plot a line corresponding to the predicted drift velocity $dx_{0}/dt = -\beta \xi$. As noted above, only integer values of $\xi$ are allowed by the model. Nonetheless, the frequency shift that $\xi$ represents can be estimated from the simulation, and need not be integer valued. In particular, we take the empirical value of $\xi$ to be the center of mass of the Fourier coefficients of the simulated waveform (computed using the FFT). After the first few time steps, this value holds steady at approximately $\xi = 0.3890$ for the $\epsilon= 0.1$ simulation and $\xi = 0.2608$ for the $\epsilon = 1$ simulation. The theoretical drift velocity matches well with the observed drift velocity of the simulation when $\epsilon=0.1$, whereas, for $\epsilon=1$, the prediction is not quantitatively satisfactory but corresponds to the qualitative behavior of the simulation (note that $\epsilon=1$ is far from the asymptotic regime). ]{}
![(a) Top view of a numerical simulation of [ with $\epsilon = 0.1$, $\alpha=1.8732$, and]{} the addition of spontaneous emission noise [ as defined in]{} . As predicted and quantified by our perturbation theory, the $\operatorname{dn}$ solution remains stable despite the induced random walk (drift) of the individual [pulse]{} solutions. Much like the Gordon-Haus jitter, our perturbation theory captures the effect of the timing variance of individual pulses. To highlight the random walk of each pulse, panels (b) and (c) show a detail of the pulses near $x=-\pi/2$ and $x=\pi/2$ respectively. []{data-label="fig:Noise"}](ContPert_epsilon=01_noise=1){width="\linewidth"}
Spontaneous emission noise
--------------------------
Spontaneous emission noise from pumping/amplification has always been a significant source of performance limitations in optical systems. For instance, in optical communication systems, the noise from amplification results in the Gordon-Haus timing jitter [@gordon1986random] which imposes a fundamental limit on transmission lengths for a given bit-error-rate constraint in lightwave communication systems. Soliton perturbation theory provided the fundamental calculation of this limitation. It also provided a number of engineering design strategies for trying to overcome the Gordon-Haus limitations, including sliding filters [@mecozzi1991soliton; @mollenauer1992sliding] and dispersion management [@suzuki1995reduction; @smith1996reduced; @kutz1998gordon].
The LLE perturbation theory developed here can also be used to evaluate the effects of spontaneous emission noise in the microresonator, something that has only recently been studied experimentally [@herr2012universal; @liao2017dependence]. Specifically, for this case the perturbation in takes the form $$G(u,x,t) = S(x,t) \; ,
\label{eq:wn}$$ where $S(x,t)$ is a white noise process modeling the spontaneous emission [@weinstein1985modulational]. In this case, for a specific realization of noise, the effects on the LLE comb parameters can be evaluated using . Generically, the noise generates amplitude, phase, center-position and center-frequency jitter. But the most pronounced effect comes from the fast scale dependency of the center position on the center frequency. Thus the evolution $$\frac{d\xi}{d\tau}=\frac{1}{\sqrt{-\beta}}(\langle\operatorname{Re}(e^{-i\psi} S(x,t) ),\operatorname{sn}z\operatorname{cn}z\rangle)$$ produces a center frequency with mean $\langle \xi \rangle$ and variance $\langle \xi^2 \rangle$ which then drives the center position through the relation $dx_0/dt = -\beta \xi$. As with the Gordon-Haus jitter, this produces a jitter in the [pulse]{} position, leading to a degradation in performance. Figure \[fig:Noise\] provides a simulation of the LLE under the influence of white noise perturbations . Note that the comb is stable, with fluctuations induced in the various solution parameters. Most notably, the zoom in of the individual pulses shows the random-walk generated as a result of the noise. As with Gordon-Haus jitter, the statistics of this random walk could be evaluated with the LLE perturbation theory we have developed.
Conclusions
===========
In conclusion, we have shown that the LLE equation supports stable solutions of the Jacobi elliptic type. These solutions model periodic pulse trains of soliton-like solutions for which the pumping $F$ is critical for stabilization. Our rigorous stability analysis also results in a perturbation theory for characterizing the effects of higher-order terms in the microresonator, such as may arise from Raman scattering, higher-order dispersion and spontaneous emission noise. The historical success of soliton perturbation theory in describing, for instance, Gordon-Haus timing jitter and/or the soliton self-frequency shifts, was critical in characterizing lightwave transmission systems and mode-locked lasers. Similarly, the LLE perturbation theory presented here can be a critically enabling tool for characterizing a host of additional microresonator phenomenon and potentially engineering new resonator designs with improved performance metrics.
Our stability analysis helps confirm several experimental observations. Most notably, it supports the recent observations that soliton states in the microresonator are not detuning degenerate, and can be individually addressed by laser detuning. Indeed, the theory rigorously confirms that the detuning can be used to lock the microresonator to any target multiple-[pulse]{} state, where the stability of each multiple-[pulse]{} state is explicitly computed and its minimum detuning assessed. The theory additionally shows that the phase-locking of the $\operatorname{dn}$ comb solution is an attractor to the resonator. Moreover, only solutions with no nodal separation (a zero separating [pulses]{}) are stabilized. Finally, the application of our theory to Raman scattering and stimulated emission perturbations show that neither effects destabilizes the comb. Rather, they both generate a drift in the pulse train, one which is deterministic in nature (Raman) and one which produces a random walk (noise).
Acknowledgments {#acknowledgments .unnumbered}
===============
J. N. Kutz acknowledges support from the Air Force Office of Scientific Research (FA9550-17-1-0200). The authors would also like to thank the anonymous reviewers for their suggestions, which improved the quality of the analysis.
Appendix {#appendix .unnumbered}
========
In the following, we utilize some standard facts concerning the eigenvalues and zeros of Sturm-Liouville operators with periodic boundary conditions. See, for example, Theorem 3.1 of Chapter 8 in [@coddington]. These arguments are modeled after those in [@weinstein1985modulational].
Proof of Proposition \[prop:lplmnull\]
--------------------------------------
One can directly verify that $L_- [\operatorname{dn}z] = 0$. Because $\operatorname{dn}z$ has no zeros, $\lambda = 0$ is the first eigenvalue (listed in increasing order).
Again, one can verify that $L_+ [\operatorname{sn}z \operatorname{cn}z] = 0$.
There is at most one function (up to a constant multiple) in $\ker (L_+)$ which is linearly independent of $w(z) = \operatorname{sn}z \operatorname{cn}z$. Note that the natural domain for $L_+$ is $H^2_{{\textnormal{per}}}[0,2NK)$ and recall that functions in $H^2_{{\textnormal{per}}}$ are determined by their values on $[0,2NK)$ and periodicity. For integer $j$, we have that $w(jK)=0$ and $w'(jK) = (-1)^j$. Suppose that $v$ is another solution of $L_+[v]=0$. We have that $w(z)v'(z) - w'(z)v(z)$ is constant, so that $(v/w)' = d/w^2$ for some constant $d$ on any interval where $w \ne 0$. Consider an interval of the form $(jK,(j+1)K)$ and let $x_j = (j+1/2)K$. For $jK<z<(j+1)K$, we have
$$v(z) = c_j w(z) + d_j w(z) \int_{x_j}^z \frac{dy}{w^2(y)} \; .$$
Let
$$\tilde{w}_j(z) = w(z) \int_{x_j}^z \frac{dy}{w^2(y)}$$
be defined on each interval $(jK,(j+1)K)$. It can be verified that the limit of $\tilde{w}_j(z)$ exists as you approach either endpoint. In particular, we have
$$\begin{aligned}
\lim_{z\to 2jK^+} \tilde{w}_{2j}(z)
&= \lim_{z\to 2jK^-} \tilde{w}_{2j-1}(z) =& -1\; , \\
\lim_{z\to (2j+1)K^-} \tilde{w}_{2j}(z)
&= \lim_{z\to (2j+1)K^+} \tilde{w}_{2j+1}(z) =& \frac{1}{\sqrt{1-k^2}} \; .
\end{aligned}$$
Because $w$ is zero at all of these endpoints, we see that for $v$ to be continuous, the $d_j$ should all be equal. Without loss of generality, we set $d_j = 1$ for all $j$.
While the derivatives are still defined at the endpoints, they are not so well behaved. We have that
$$\begin{aligned}
j_1 &:= \lim_{z\to 2jK^+} \tilde{w}'_{2j}(z)
- \lim_{z\to 2jK^-} \tilde{w}'_{2j-1}(z) \\
&= \frac{2}{1-k^2} \left ( (1-k)^{3/2} -1
+ (2-k^2) E(K/2,k) - (1-k^2) K \right ) \; ,\\
j_2 &:= \lim_{z\to (2j+1)K^+} \tilde{w}'_{2j+1}(z)
- \lim_{z\to (2j+1)K^-} \tilde{w}'_{2j}(z) \\
&= \sqrt{1-k^2} \left ( j_1 - \frac{2(2-k^2)E-4(1-k^2)K}{1-k^2}
\right ) \; .
\end{aligned}$$
Note that, for $0<k<1$, $j_1 \ne j_2$. To enforce that $v$ has continuous derivatives, we then obtain the following system of equations
$$\begin{aligned}
c_1-c_0 &= -j_1 \\
c_2-c_1 &= j_2 \\
c_3-c_2 &= -j_1 \\
&\vdots \\
c_{2N-1}-c_{2N-2} &= -j_1 \\
c_0-c_{2N-1} &= j_2 \; .
\end{aligned}$$
By summing all of the equations, we obtain that $0 = N(j_2-j_1) \neq 0$, so that the equations are inconsistent. Therefore, there is no such $v$ with a continuous derivative, i.e. there is no such $v$ in $H^2_{{\textnormal{per}}}[0,2NK)$. Note that for the case $k=0$, we see that $j_1 = j_2$ so that such a $v$ does exist, as expected.
Proof of Proposition \[prop:geneig\]
------------------------------------
From Proposition \[prop:lplmnull\], we have that
$$\ker({\bf L}^\dag) = \operatorname{span}\{ (\operatorname{dn}z,0)^\intercal,
(0,\operatorname{sn}z \operatorname{cn}z)^\intercal\} \; .$$
Recall the definitions of $\phi$ and $\varphi$:
&& (z)=(K [ E(z,k)]{}- E z ) z-k\^[2]{}zz,\
&& (z)=k\^[2]{} z z (K[ E(z,k)]{}-Ez)+(E-K) z\
&& +k\^[2]{}K\^[2]{} zz.
It can be verified that
&& L\_[+]{} L\_[-]{}\[(z)\] = L\_[+]{} = 0,\
&& L\_[-]{} L\_[+]{}\[(z)\] = L\_[-]{} \[2((k\^[2]{} - 2)E-2(k\^[2]{}-1)K)z\]=0.
Therefore, $$\begin{aligned}
\ker( ({\bf L}^\dag)^2) &= \operatorname{span}\{ (\operatorname{dn}z, 0)^\intercal,
(0,\operatorname{sn}z \operatorname{cn}z)^\intercal, (\phi(z),0)^\intercal, \nonumber \\
& \qquad (0,\varphi(z))^\intercal \} \; .\end{aligned}$$ Suppose that $(f,g)^\intercal \in \ker( ({\bf L}^\dag)^3 )$. Then, formally,
$$\begin{aligned}
f &= c_1 L_-^{-1} \varphi(z)
+ c_2 \phi(z) + c_3 \operatorname{dn}z \; ,\\
g &= c_4 L_+^{-1} \phi(z)
+ c_5 \varphi(z) + c_6 \operatorname{sn}z \operatorname{cn}z \; ,\end{aligned}$$
where the inverses above denote a particular solution of the corresponding inhomogeneous ODE. Consider $L_-^{-1} \varphi(z)$. Note that the Fredholm alternative implies that
$$\begin{aligned}
0 &= \langle k^{2} \operatorname{cn}z \operatorname{sn}z (K{ E(z,k)}-Ez)+(E-K) \operatorname{dn}z, \operatorname{dn}z \rangle
\nonumber \\ &\qquad + \langle k^{2}K\operatorname{cn}^{2} z\operatorname{dn}z, \operatorname{dn}z \rangle \\
&= {N}(E^2 + (k^2-1)K^2) \; .\end{aligned}$$
For $0<k<1$, the expression $E^2 + (k^2-1)K^2 > 0$, a contradiction. Therefore, there is no such particular solution. Similarly, consider $L_+^{-1} \phi(z)$. The Fredholm alternative implies that
$$\begin{aligned}
0 &= \langle \left (K{ E(z,k)}- Ez \right)
\operatorname{dn}z-k^{2}K\operatorname{sn}z\operatorname{cn}z , \operatorname{sn}z \operatorname{cn}z \rangle \\
&= -\frac{{N}}{k^2} \left (E^2 + (k^2-1)K^2 \right ) \; ,\end{aligned}$$
again, a contradiction. Therefore,
$$\ker_g( {\bf L}^\dag ) = \ker( ({\bf L}^\dag)^2 ) \; .$$
Proof of Proposition \[prop:h1control\]
---------------------------------------
The existence of $C_2$ is simple to establish. To establish the existence of $C_1$, we require the following two lemmas. [Note that for the remainder of these statements, we assume that $N=1$.]{}
\[lem:lpbound\] Suppose that $\langle f, \operatorname{dn}z \rangle = 0$ and $\langle f, \phi(z) \rangle$. Then there exists a positive constant $C_1^+$ such that $$\langle L_+ f, f \rangle \geq C_1^+ \| f\|^2_{L^2} \; .$$
\[lem:lmbound\] Suppose that $\langle g, \operatorname{sn}z \operatorname{cn}z \rangle = 0$ and $\langle g, \varphi(z) \rangle = 0$. Then there exists a positive constant $C_1^-$ such that $$\langle L_- g, g \rangle \geq C_1^- \| g\|^2_{L^2} \; .$$
Suppose that $\langle f, \operatorname{dn}z \rangle = 0$, $\langle f, \phi(z) \rangle$, $\langle g, \operatorname{sn}z \operatorname{cn}z \rangle = 0$, and $\langle g, \varphi(z) \rangle = 0$. Let $C_1^+$ and $C_1^-$ be as in Lemmas \[lem:lpbound\] and \[lem:lmbound\], respectively. Then
$$\begin{aligned}
& \langle L_+ f,f \rangle + 6 \|f \|_{L^2}
+ \langle L_- g,g \rangle + 2 \| g \|_{L^2} \\
& \quad = \| \dfrac{d}{dz} f \|_{L^2}
+ 6 \langle (1-\operatorname{dn}^2 z)f,f \rangle
+ (2-k^2) \|f \|_{L^2} \\
& \quad \qquad + \| \dfrac{d}{dz} g \|_{L^2}
+ 2 \langle (1-\operatorname{dn}^2 z)g,g \rangle
+ (2-k^2) \|g \|_{L^2} \; , \\
& \quad \geq \| f\|_{H^1_{{\textnormal{per}}}} + \| g \|_{H^1_{{\textnormal{per}}}} \; .\end{aligned}$$
Therefore, the proposition holds with
$$C_1 = \min \left ( \frac{1}{1+\frac{6}{C_1^+}},
\frac{1}{1+\frac{2}{C_1^-}} \right ) \; .$$
### Proof of Lemma \[lem:lpbound\]
In the following, we repeat the argument of [@weinstein1985modulational], making appropriate changes to handle the periodic case. First, we note that by Theorem 3.1 of Chapter 8 in [@coddington], $L_+$ has one negative eigenvalue [(when $N=1$)]{} with a corresponding eigenfunction $f_0$, which we take to be nonnegative without loss of generality. Define
$$\gamma_1 = \min_f \langle L_+ f, f \rangle
\; , \mbox{ where } \|f \|_2 = 1, \;
\langle f, \operatorname{dn}z \rangle = 0 \; .$$
Then, by Lemma E.1 of [@weinstein1985modulational], we have that $\gamma_1 \geq 0$ if
$$\langle L_+^{-1} \operatorname{dn}z, \operatorname{dn}z \rangle
\leq 0 \; ,$$
which is straightforward to verify using arguments similar to those in the proof of Proposition \[prop:geneig\]. Therefore, $\gamma_1 \geq 0$. The lemma is then proved if we can show that $ \gamma_2 = \inf_{f} \langle L_+ f, f \rangle$ with $f$ restricted such that $\|f\|_{L^2} = 1$, $\langle f, \operatorname{dn}z \rangle = 0$, and $\langle f, \phi(z) \rangle = 0$ is non-zero, as $\gamma_2 \geq \gamma_1 \geq 0$.
Suppose that $\gamma_2 = 0$. Let $f_m$ be a minimizing sequence of $\langle L_+ f, f \rangle$ satisfying $\|f_m\|_{L^2} = 1$, $\langle f_m, \operatorname{dn}z\rangle = 0$, and $\langle f_m, \phi(z) \rangle = 0$. Given $\delta > 0$, there exists a $M(\delta)$ such that
$$\begin{aligned}
0 &< \int_{-K(k)}^{K(k)} \left (\frac{d}{dz} f_m \right )^2
\, dz
+ (2-k^2) \int_{-K(k)}^{K(k)} f_m^2 \, dz \\
& \leq 6 \int_{-K(k)}^{K(k)} \operatorname{dn}^2 z f_m^2 \, dz + \delta \; ,\end{aligned}$$
for all $m \geq M(\delta)$. In particular, the sequence $f_m$ is uniformly bounded in the $H^1_{{\textnormal{per}}}$ norm. Therefore, there is a subsequence of $f_m$ which converges weakly to an $H^1_{{\textnormal{per}}}$ function $f_*$. This function satisfies the constraints $\langle f_*, \operatorname{dn}z \rangle = 0$ and $\langle f_*, \phi(z) \rangle = 0$ by weak convergence. Because $H^1_{{\textnormal{per}}}$ is compactly embedded in $L^2$, there exists a further subsequence, which we denote by $f_{m_j}$, that converges in the $L^2$ norm (to $f_*$). Therefore, $ \| f_* \|_{L^2} = 1$.
Let $h$ be such that $\|h\|_2 = 1$. Note that $\langle h, f_*' \rangle = \lim \langle h, f_{m_j}' \rangle
\leq \liminf \| f_{m_j}' \|_{L^2}$ by weak convergence in $H^1_{{\textnormal{per}}}$. Taking the maximum over all such $h$ implies that $\| f_*' \|_{L^2} \leq \liminf \| f_{m_j}' \|_{L^2}$. Combining this with the $L^2$ convergence of $f_{m_j}$ gives that $\langle L_+ f_*, f_* \rangle \leq \liminf
\langle L_+ f_{m_j}, f_{m_j} \rangle = 0$, so that $\langle L_+ f_*, f_* \rangle = 0$.
Because $f_*$ attains the minimum and is admissible, there exists a critical point of the problem
$$\begin{aligned}
(L_+ - \lambda_1)f &= \lambda_2 \operatorname{dn}z + \lambda_3 \phi(z)
\; , \label{eq:lagmul1} \\
\|f \|_2 &= 1 \; , \\
\langle f, \operatorname{dn}z \rangle &= 0 \; ,\\
\langle f, \phi(z) \rangle &= 0 \; ,\end{aligned}$$
of the form $(f_*, \lambda_1, \lambda_2, \lambda_3)$. Taking the inner product of $f_*$ with , we obtain that $\lambda_1 = \langle L_+f_*,f_*\rangle = 0$. This implies that
$$L_+ f_* = \lambda_2 \operatorname{dn}z + \lambda_3 \phi(z) \; .
\label{eq:lagmul2}$$
Taking the inner product of $\operatorname{sn}z \operatorname{cn}z$ with , we obtain that $\lambda_3 = 0$. Following the arguments in the proof of Proposition \[prop:geneig\], this implies that
$$f_* = \frac{\lambda_2}{2((k^2-2)E-2(k^2-1)K)} \varphi(z)
+ \lambda_4 \operatorname{sn}z \operatorname{cn}z \; ,$$
for some $\lambda_4$. The constraint $\langle f_*, \phi(z) \rangle = 0$ implies that $\lambda_4 = 0$ and the constraint $\langle f_*, \operatorname{dn}z \rangle = 0$ implies that $\lambda_2 = 0$. We have that $f_* \equiv 0$, a contradiction. Therefore, $\gamma_2 > 0$, proving the lemma.
### Proof of Lemma \[lem:lmbound\]
This lemma can be proved using arguments similar to the above.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present the first interferometric detections of Fast Radio Bursts (FRBs), an enigmatic new class of astrophysical transient. In a 180-day survey of the Southern sky we discovered 3 FRBs at 843 MHz with the UTMOST array, as part of commissioning science during a major ongoing upgrade. The wide field of view of UTMOST ($\approx 9$ deg$^{2}$) is well suited to FRB searches. The primary beam is covered by 352 partially overlapping fan-beams, each of which is searched for FRBs in real time with pulse widths in the range 0.655 to 42 ms, and dispersion measures $\leq$2000 pc cm$^{-3}$. Detections of FRBs with the UTMOST array places a lower limit on their distances of $\approx 10^4$ km (limit of the telescope near-field) supporting the case for an astronomical origin. Repeating FRBs at UTMOST or an FRB detected simultaneously with the Parkes radio telescope and UTMOST, would allow a few arcsec localisation, thereby providing an excellent means of identifying FRB host galaxies, if present. Up to 100 hours of follow-up for each FRB has been carried out with the UTMOST, with no repeating bursts seen. From the detected position, we present 3$\sigma$ error ellipses of 15$\arcsec \times 8.4^{\circ}$ on the sky for the point of origin for the FRBs. We estimate an all-sky FRB rate at 843 MHz above a fluence $\cal F_\mathrm{lim}$ of 11 Jy ms of $\sim 78$ events sky$^{-1}$ d$^{-1}$ at the 95 percent confidence level. The measured rate of FRBs at 843 MHz is of order two times higher than we had expected, scaling from the FRB rate at the Parkes radio telescope, assuming that FRBs have a flat spectral index and a uniform distribution in Euclidean space [@Caleb]. We examine how this can be explained by FRBs having a steeper spectral index and/or a flatter log$N$-log$\mathcal{F}$ distribution than expected for a Euclidean Universe.'
author:
- |
M. Caleb$^{1,2,3}$[^1], C. Flynn$^{2,3}$, M. Bailes$^{2,3}$, E.D. Barr$^{2,3,4}$, T. Bateman$^{6}$, S. Bhandari$^{2,3}$, D. Campbell-Wilson$^{6,3}$ W. Farah$^{2}$, A.J. Green$^{6,3}$, R.W. Hunstead$^{6}$ A. Jameson$^{2,3}$, F. Jankowski$^{2,3}$, E.F. Keane$^{5}$, A. Parthasarathy$^{2,3}$, V. Ravi$^{2,3,7}$, P.A. Rosado$^{2,8}$, W. van Straten$^{2,9}$, V. Venkatraman Krishnan$^{2,3}$\
\
$^{1}$ Research School of Astronomy and Astrophysics, Australian National University, ACT, 2611, Australia\
$^{2}$ Centre for Astrophysics and Supercomputing, Swinburne University of Technology, P.O. Box 218, Hawthorn, VIC 3122, Australia\
$^{3}$ ARC Centre of Excellence for All-sky Astrophysics\
$^{4}$ Max-Planck-Institut f[ü]{}r Radioastronomie, Auf dem H[ü]{}gel 69, D-53121 Bonn, Germany\
$^{5}$ SKA Organisation, Jodrell Bank Observatory, Cheshire, SK11 9DL, UK\
$^{6}$ Sydney Institute for Astronomy (SIfA), School of Physics, The University of Sydney, NSW 2006, Australia\
$^{7}$ Cahill Center for Astronomy and Astrophysics, MC249-17, California Institute of Technology, Pasadena, CA 91125, USA\
$^{8}$ Monash Centre for Astrophysics, School of Physics and Astronomy, Monash University, VIC 3800, Australia\
$^{9}$ Institute for Radio Astronomy & Space Research, Auckland University of Technology, Private Bag 92006, Auckland 1142, New Zealand
bibliography:
- 'UTMOST\_FRBs.bib'
title: The first interferometric detections of Fast Radio Bursts
---
instrumentation: interferometers – intergalactic medium – surveys – radio continuum – methods: data analysis
Introduction
============
Fast radio bursts (FRBs) are a relatively new class of radio transient that are short, bright and highly dispersed. The pulses are typically of durations of a few milliseconds, and exhibit dispersion sweeps characteristic of propagation through a cold diffuse plasma [@Lorimer; @Thornton]. The dispersion measures (DMs) of these pulses are significantly higher than the contribution from the line-of-sight through the Galactic Interstellar Medium (ISM), suggestive of a cosmological origin in which the large DMs are due to passage through the Intergalactic Medium (IGM). If they are at cosmological distances, their inferred intrinsic energies ($> 10^{31}$ J) and brightness temperatures (${T}_\mathrm{b} >10^{33}$ K) necessitate a coherent emission mechanism, while the short durations of the pulses suggest a very compact source of origin [@Luan; @Dennison]. The 18 FRBs published to date (refer to the FRBCAT repository$^{\ref{frbcat}}$ for the complete list) have been discovered in either post-processing of archival surveys or, in real-time, using the Parkes radio telescope with the exception of two, detected at the Arecibo [@Spitler] and Green Bank telescopes All but one of the bursts have been found at 1.4 GHz, with the exception being the GBT burst, which was seen at 800 MHz.
The observed FRB all-sky rate is very high. [@Champion] derive a rate of $7^{+5}_{-3} \times 10^{3}$ events sky$^{-1}$ d$^{-1}$ at 1.4 GHz for bursts between 0.13 and 1.5 Jy ms in fluence and widths in the range 0.128 ms to 16 ms. The high FRB rate is a major constraint on theories for their origin. Until recently, such theories have generally assumed they are cataclysmic events, in which the progenitor is obliterated. However, one FRB is now known to repeat in a non-periodic manner [FRB 121102, @nat_spitler], opening up possibilities for other progenitor models. Following the discoveries reported in this paper, [@Chatterjee] have achieved sub-arsecond localisation of the FRB 121102 using radio interferometric observations from the Very Large Array. The source has been localised to a $m_{r^\prime} = 25.1$ AB mag low-metallicity, star-forming dwarf galaxy at $z = 0.19273(8)$ [@SriHarsh]. The precise localisation shows that the source is either co-located with a 180 $\upmu$Jy active galactic nucleus or an unresolved type of extragalactic source. However, the exact nature of the FRB progenitor is still unknown.
Despite concerted follow-up efforts for almost all FRBs, this remains the only FRB seen to repeat. These efforts have been quite substantial. For instance, $\approx80$ hrs of followup for the Lorimer burst [@Lorimer], $\approx80$ hrs for FRB 131104 [@Ravi] and $\approx110$ hrs of selected FRB positions [@Petroff_followup] at the Parkes radio telescope yielded no repeats. This suggests the possibility of there being two independent classes of FRBs – repeating and non-repeating – with two classes of possible progenitors [@nat_keane]. Progenitor theories include flaring magnetars [@Lyubarsky], giant pulses from pulsars [@Wasserman; @Connor], binary white dwarf mergers [@Kashiyama], neutron star mergers [@Totani] and collapsing supramassive neutron stars [@Falcke]. It is possible that the lack of repetition of pulses for the FRB discoveries at the Parkes radio telescope is merely due to limited sensitivity and follow-up time, and that all FRBs have a common origin [@Scholz]. FRB 010724 is an exception to this however: its extreme brightness ($\sim 30$ Jy) far outweighs the lower gain of Parkes relative to Arecibo, so that one cannot infer its lack of repeat bursts is due to limited sensitivity. Recently, [@RaviSci] have reported the detection of FRB 150708, which is of comparable brightness ($\sim 12$ Jy) to FRB 010724, and exhibits 100 percent polarisation and suggests weak turbulence in the ionised IGM. [@DeLaunay] have associated a $\gamma$-ray transient with the FRB 131104 discovered by [@Ravi]. However [@Ravi2016] in contrast, report on the discovery of a variable source (consistent with an AGN) temporally and spatially coincident with the FRB 131104 but not spatially coincident with the $\gamma$-ray burst, and rule out the association of the $\gamma$-ray burst with the FRB using probabilistic reasoning.
![Frequency vs time behaviour of FRBs 160317, 160410 and 160608 detected at UTMOST at the centre frequency of 834.765 MHz. The top panel in each case shows the frequency-averaged pulse profile. The bottom panel shows that narrow-band RFI has been excised and the effects of inter-channel dispersion have been removed assuming DMs of $1165 \pm11$, $278\pm3$ and $682\pm7$ pc cm$^{-3}$ respectively. The data are uncalibrated as the bandpass of the system varies as a function of meridian angle, and the flux densities are in arbitrary units. Note the different time range on the abscissa for FRB 160410.[]{data-label="fig:frbs1"}](Fig1a){width="50.00000%"}
![Frequency vs time behaviour of FRBs 160317, 160410 and 160608 detected at UTMOST at the centre frequency of 834.765 MHz. The top panel in each case shows the frequency-averaged pulse profile. The bottom panel shows that narrow-band RFI has been excised and the effects of inter-channel dispersion have been removed assuming DMs of $1165 \pm11$, $278\pm3$ and $682\pm7$ pc cm$^{-3}$ respectively. The data are uncalibrated as the bandpass of the system varies as a function of meridian angle, and the flux densities are in arbitrary units. Note the different time range on the abscissa for FRB 160410.[]{data-label="fig:frbs1"}](Fig1b){width="50.00000%"}
![Frequency vs time behaviour of FRBs 160317, 160410 and 160608 detected at UTMOST at the centre frequency of 834.765 MHz. The top panel in each case shows the frequency-averaged pulse profile. The bottom panel shows that narrow-band RFI has been excised and the effects of inter-channel dispersion have been removed assuming DMs of $1165 \pm11$, $278\pm3$ and $682\pm7$ pc cm$^{-3}$ respectively. The data are uncalibrated as the bandpass of the system varies as a function of meridian angle, and the flux densities are in arbitrary units. Note the different time range on the abscissa for FRB 160410.[]{data-label="fig:frbs1"}](Fig1c){width="50.00000%"}
Most published FRBs have been detected with single dish antennas, with relatively poor angular resolution, and we are unable to indisputably rule out a near-field or atmospheric origin for the one-off events until now. The FRB detections made with the multi-beam receiver at the Parkes radio telescope however, are likely to originate at $\gtrsim 20$ km [@Vedantham]. Also FRB 150418 has been proposed to be associated with a galaxy at $z \sim 0.5$. However this association has been called into question by [@Williams] and [@Vedantham], and other models like giant pulses from extragalactic pulsars which could account for the excess DM in the local environment, have been proposed [@Connor]. Better localisation during discovery in the radio requires an interferometric detection.
![The three panels display the total power pulse profiles for one polarisation in three adjacent fan-beams. FRBs 160317 and 160410 were also detected as sub-threshold events in neighbouring fan-beams (in addition to the high S/Ns in the primary detection fan-beams), indicating that they did not occur near the centres of the primary fan-beam. On the contrary, FRB 160608 was only detected in one fan-beam suggesting that it occurred close to the the centre of beam 208 (see bottom panel).[]{data-label="fig:frbs2"}](Fig2a.pdf){width="50.00000%"}
![The three panels display the total power pulse profiles for one polarisation in three adjacent fan-beams. FRBs 160317 and 160410 were also detected as sub-threshold events in neighbouring fan-beams (in addition to the high S/Ns in the primary detection fan-beams), indicating that they did not occur near the centres of the primary fan-beam. On the contrary, FRB 160608 was only detected in one fan-beam suggesting that it occurred close to the the centre of beam 208 (see bottom panel).[]{data-label="fig:frbs2"}](Fig2b.pdf){width="50.00000%"}
![The three panels display the total power pulse profiles for one polarisation in three adjacent fan-beams. FRBs 160317 and 160410 were also detected as sub-threshold events in neighbouring fan-beams (in addition to the high S/Ns in the primary detection fan-beams), indicating that they did not occur near the centres of the primary fan-beam. On the contrary, FRB 160608 was only detected in one fan-beam suggesting that it occurred close to the the centre of beam 208 (see bottom panel).[]{data-label="fig:frbs2"}](Fig2c.pdf){width="50.00000%"}
In a companion paper, we describe how the Molonglo Observatory Synthesis Telescope (sited near Canberra in Australia) is currently undergoing a major upgrade, with the addition of a state-of-the-art correlator to transform it into an FRB finding machine - the UTMOST (Bailes et al., submitted). Two FRB searches were performed with UTMOST in 2015 during the upgrade, when the system was operating at a small fraction of the final expected sensitivity, and only yielded an upper limit of the FRB rate [@Caleb].
We have now undertaken a third FRB survey at UTMOST and discovered 3 FRBs. These are the first FRBs observed with an interferometer, further strengthening the case for an astronomical origin in addition to the detections at other telescopes and in the expected number of beams at Parkes for far-field events, as detection with UTMOST implies the events are in the far-field region $\gtrsim 10^4$ km. Section \[sec:OA\] of this paper briefly outlines the telescope specifications, survey properties and the transient detection pipeline. We present the bursts’ properties and their follow-up observations and localisation areas in Section \[sec:results\]. The event rate estimates of the FRBs at 843 MHz based on the detections of the 3 FRBs and constraints on their spectral index are detailed in Section \[sec:disc\] followed by our conclusions in Section \[sec:conc\].
UTMOST specifications and survey properties {#sec:OA}
===========================================
The UTMOST consists of an East-West (E-W) aligned cylindrical paraboloid divided into two ‘arms’ (separated by a 15-m gap), each 11.6-m wide and 778-m long, with 7744 right circularly polarised ring antennas operating at 843 MHz on a line feed system at its focus. Groups of 22 consecutive ring antennas (these groups are termed ‘modules’) are phased to the physical centre of the module, forming 352 unique inputs (each with a beam $4.0^{\circ} \times 2.8^{\circ}$ FWHP) which are then beamformed (Bailes at al., submitted). We operate the telescope by tilting the arms North-South and steering the ring antennae East-West by differential rotation. UTMOST can access the sky South of $\delta=+18^{\circ}$ with the East-West steering limited to $\pm60^{\circ}$. The telescope’s field of view, sensitivity and high duty cycle make it a near ideal survey instrument for finding FRBs and other radio transients. Since late 2015, we have been using UTMOST to search for fast radio transients for an average of 18 hours a day, while simultaneously timing more than 300 pulsars weekly (Bailes et al., submitted, Jankowski et al., in prep).
In FRB search mode, the 4.0 degrees FWHP of the primary beam is tiled in the E-W direction by 352 elliptical, coherent, tied-array beams (called ‘fan-beams’, each 46$\arcsec$ wide), spaced 41$\arcsec$ apart and overlapping at very close to their half power points at 843 MHz. In the N-S direction the resolution of the fan-beams is the same as that of the primary beam ($\approx2.8$ degrees). The fan-beams are numbered from 1 to 352 running from East to West across the primary beam, with fan-beam 177 directly centred on boresight. The sensitivity of the telescope to bursts can be estimated using the radiometer equation:
$$S_\mathrm{min} = \beta \, \frac {(\mathrm{S/N_\mathrm{min}}) \, T_\mathrm{sys}} {G \, {\sqrt{\Delta\nu \, W \, N_\mathrm{p}}}}$$
{width="5.0"}
where ${S_\mathrm{min}}$ is the minimum detectable flux for a threshold signal-to-noise $\mathrm{S/N_{min}}$, $\beta$ is the digitisation factor, $\Delta\nu$ is the bandwidth in Hz, ${N_\mathrm{p}}$ is the number of polarisations (${N_\mathrm{p}}=1$ for UTMOST as it is right circularly polarised only), $W$ is the pulse width in ms, ${T_\mathrm{sys}}$ is the system temperatures in K respectively, and $G$ is the system gain in K Jy$^{-1}$. We define S/N as the ratio of the sum of the on-pulse flux to the product of the rms of the off-pulse flux and square root of number of on-pulse bins ($\mathrm{S/N} = \frac{\mathrm{I_{on}}}{\sqrt{\mathrm{nbin}}\,\mathrm{I_{off}}}$). For the fully upgraded instrument, we expect $S_\mathrm{min} = 1.6$ Jyms for a 10$\sigma$ 1-ms wide pulse, 3.5 K Jy$^{-1}$ gain, 100 K system temperature and 31.25 MHz bandwidth. The system bandwidth is however only about half of the initially anticipated 31.25 MHz bandwidth, as the ring antennas have a significant roll-off in sensitivity away from 843 MHz. This has been measured using integrated pulses from the pulsar J1644$-$4559. We find that on average $\sim 86$ percent of the total S/N is concentrated in the upper half of the band ($\sim836 - 850$) as the antennas are tuned to maximum sensitivity at 843 MHz. We adopt a bandwidth of 16 MHz for the sensitivity calculations in the paper, to be conservative.
During the upgrade, we characterise the system sensitivity by a fraction of the final expected gain $\upepsilon$. This factor encompasses systemic losses due to (1) pointing errors (from physical misalignment in the modules N-S, and phasing errors in the antenna system E-W), (2) self-generated radio frequency interference (RFI) mainly due to improperly shielded electronics in the receiver boxes near the telescope, (3) coherent noise in the receiver boxes, which affects some sets of adjacent modules, and other inefficiencies in the system performance that we are still characterising, such as systematic errors in the phase/delay solutions across the interferometer (Bailes et al. in submitted).
At present (October 2016), we estimate $\upepsilon \approx 0.14$, based on observations of strong calibrators of known flux densities and a number of high DM pulsars with relatively stable flux densities. This implies an effective $T_\mathrm{sys}$ of $400 \pm 100$ K. This is significantly higher than the system temperature seen on the best performing modules, which can be as low as 100 K. We note that $\upepsilon$ can vary from day to day as modules are either serviced in the field or have electronics maintenance in the workshops, and typically lie in the range $0.15 < \upepsilon <
0.20$. Occasionally, if only one arm is operational, we have the option to continue surveys at half sensitivity (i.e. $0.07 < \upepsilon < 0.10$). The telescope can access the Southern sky for $\delta < +18^{\circ}$, and for most parts of the sky we tend to observe reasonably close to the meridian, in order to maximise sensitivity. The sensitivity is reduced by projection effects away from the meridian.
{width="47.00000%"}{width="47.00000%"} {width="47.00000%"}
\[sec:pipeline\]
In November 2015, we commenced our third FRB survey “V3.0”. It ran for a total of 159.0 days on sky (between 01-11-2015 and 30-11-2016), at $\upepsilon \approx 0.14$ of the final target telescope sensitivity. Our fluence limit of the survey, that is the fluence of the narrowest detectable pulse $\cal{F}_\mathrm{lim}$ can be parametrized as,
$$\label{eq:lim_flue}
{\cal{F}_\mathrm{lim}} \approx 11 \Bigg(\frac{W}{\mathrm{ms}}\Bigg)^{1/2}\, \mathrm{Jy \, ms}$$
where, 11 Jy is the UTMOST flux limit for $\mathrm{S/N}=10$, $G=3.0$ K Jy$^{-1}$, $\Delta\nu=16$ MHz, $W=1$ ms, $N_\mathrm{p}=1$ and $T_\mathrm{sys}=400$ K. It should be noted that this is not the same as the fluence completeness limit $\cal{F}_\mathrm{complete}$. Between $\cal{F}_\mathrm{lim}$ and $\cal{F}_\mathrm{complete}$ we are incomplete and not all FRBs with fluences in this range are detectable. This incompleteness region corresponds to the pink shaded region in Figure \[fig:fluence\]. The two previous surveys (V1.0 and V2.0) reported in [@Caleb] yielded no FRB events. Relative to V3.0, V1.0 ran for 19.5 days at lower sensitivity ($\upepsilon = 0.07$), while V2.0 operated for 9.4 days at the same sensitivity ($\upepsilon =
0.14$). FRB survey V3.0 consists primarily of pointings taken commensally during pulsar timing observations. In this mode, the time series data from 352 fan-beams are searched for dispersed single pulses in real time, using a custom version of the heimdall software on 8 Nvidia GeForce GTX TITAN X (Maxwell) GPUs with a latency of 8-s. The resulting candidates were then processed offline, typically the following morning for overnight pulsar timing (RFI is much reduced at night, and the telescope is made available for maintenance on week days). On weekends, the telescope is usually operated continuously. The candidate processing pipeline used is described in detail in [@Caleb]. The process followed is:
1. obtain 352 data streams (8-bits/sample), one for each fan-beam, at 655.36-$\upmu$s sampling
2. search time series for single pulses with width, $0.65536 < W <
41.943$ ms ($W = 2^\mathrm{N} \times 0.65536$ ms, where N = 0,1,2...) and DMs in the range $100 < \mathrm{DM} < 2000$ pc cm$^{-3}$,
3. remove events occurring simultaneously in more than 3 fan-beams at a given instant in time,
4. classify only events with S/N $\geq10$, DM $\geq100$ pc cm$^{-3}$ and $W\leq41.943$ ms as potential FRB candidates. These then require human scrutiny of the diagnostic plots, to remove candidates that were RFI, almost always due to narrow-band mobile handset emissions in our operating passband and single pulses from known pulsars.
Results {#sec:results}
=======
The false positive rate at UTMOST is high due to RFI caused by mobile phone handsets, which produce narrow band (5-MHz) emission in our band, typically in $\approx 20$ ms pulses. These can be eliminated because celestial pulses are expected to be broadband, modulated by a frequency dependent response across the 31.25 MHz bandwidth. This process has been validated using individual pulses from about 20 bright pulsars seen to date. We are presently automating this process using machine learning algorithms, so that pulses can trigger a full voltage dump of the raw data while they are still in the $\approx30$ seconds of ring buffer storage, with alerts issued in near real-time. RFI occurs predominantly at low DM, but the rate is high enough to produce a few hundred spurious candidates above our DM limit of 100 pccm$^{-3}$ daily. Candidates were typically vetted each morning after data taking.
In 2016 March, April and June we made the first interferometric detections of FRBs at 843 MHz: FRB 160317, FRB 160410 and FRB 160608, as shown in Figure \[fig:frbs1\].
FRB 160317
----------
This was detected on 2016 March 17 at 09:00:36.530 UTC while observing an X-ray magnetar SGR 0755$-$23, in response to an Astronomers Telegram [@Barthelmy]. The burst occurred about 0.4 degrees East of the magnetar, and was detected $\sim$1$^{\circ}$ off the Galactic plane with a DM of 1165(11) pc cm$^{-3}$. The DM due to the ISM at this sight-line is $\sim320$ pc cm$^{-3}$ from the NE2001 model by [@Cordes] and $\sim395$ pc cm$^{-3}$ from the YMW16 model [@Yao]. The burst with S/N $ \sim 13$, occurred East of the centre of the primary fan-beam of detection (Beam 212) since it appeared weakly in the adjacent fan-beam with S/N $\sim 5$ (Beam 213) as shown in Figure \[fig:frbs2\].
FRB 160410
----------
Similarly to FRB 160317, this FRB was also detected in two adjacent fan-beams (Beam 085 with S/N $\sim 13$ and Beam 084 with S/N $\sim 4$) as seen in Figure \[fig:frbs2\]. A single dispersed pulse was discovered on 2016 April 04 at 08:33:39.680 UTC, in an observation of the pulsar J0837$+$0410 at the telescope’s boresight. This pulsar is so bright that individual pulses were seen from it as the FRB occurred, meaning the flux density scale and bandpass response of the observation were well understood. The FRB was seen $\sim 1^{\circ}$ away from boresight. This pulse was detected at Galactic latitude, $\sim27^{\circ}$ with the line-of-sight DM accounting for only $\sim58$ pc cm$^{-3}$ of the total observed DM from the NE2001 model. The YMW16 model estimates $\sim 63$ pc cm$^{-3}$. FRB 160410 has the lowest DM excess $\sim220$ pc cm$^{-3}$ of any published FRB, potentially making it the closest FRB discovered to date and an excellent candidate to search for repeat pulses.
FRB 160608
----------
The burst occurred in an observation of the pulsar J0738$-$4042 at $l = 254.11$ deg and $b=-9.54$ deg on 2016 June 06 at 03:53:01.088 UT with a total DM of $\sim682$ pc cm$^{-3}$ and $\sim238$ pc cm$^{-3}$ contribution from the Milky Way (NE2001). The YMW16 model’s estimate however is $\sim 310$ pc cm$^{-3}$. It was seen $\sim0.5^{\circ}$ from the boresight position. FRB 160608 was detected with S/N $\sim 12$, just above the detection threshold of 10 and it occurred towards the centre of the primary detection fan-beam (Beam 208). No pulse was detected in the adjacent fan-beams (see Figure \[fig:frbs2\]). This was initially of concern, but tests with the Vela pulsar placed sufficiently far South of the telescope boresight, to produce an individual pulse with the same S/N showed that detection in a single fan-beam occurred $\approx20\%$ of the time. The localisation of this FRB is thus slightly poorer ($21\arcsec \times 8.4^{\circ}$) than for the other two FRBs, for which 2 fan-beam detections allow more accurate positions.\
\[tab:specs\]
------------------------------------------------------------- ------------------------- ------------------------- -------------------------
Parameter FRB 160317 FRB 160410 FRB 160608
\[0.5ex\] UTC start 2016-03-17-08:30:58 2016-04-10-08:16:54 2016-06-08-03:52:24
UTC event 2016-03-17-09:00:36.530 2016-04-10-08:33:39.680 2016-06-08-03:53:01.088
RA J2000 (hh:mm:ss) 07:53:47 08:41:25 07:36:42
DEC J2000 (dd:mm:ss) $-$29:36:31 +06:05:05 $-$40:47:52
$l$ (deg) 246.05 220.36 254.11
$b$ (deg) $-$0.99 27.19 $-$9.54
Detection S/N 13 13 12
$S_\mathrm{peak, obs}$ (Jy) $>3.0$ $>7.0$ $>4.3$
Boresight Fluence (Jyms) $>69$ $>34$ $>37$
Isotropic energy, $E_{0}$ (J) $\sim 10^{34}$ $\sim 10^{32}$ $\sim 10^{33}$
Observed width, $W$ (ms) 21(7) 4(1) 9(6)
DM smearing, $\uptau_{843 \,\mathrm{MHz}}$ (ms) 12.6 3.0 7.4
$\mathrm{DM}_\mathrm{FRB}$ (pc cm$^{-3}$) 1165(11) 278(3) 682(7)
$\mathrm{DM}_{\mathrm{Gal},\mathrm{NE2001}}$ (pc cm$^{-3}$) 319.6 57.7 238.3
$\mathrm{DM}_{\mathrm{Gal},\mathrm{YMW16}}$ (pc cm$^{-3}$) 394.6 62.5 310.3
Inferred redshift, $z$ 0.7 0.2 0.4
Luminosity Distance (Gpc) 4.30 0.89 1.97
Co-moving Distance (Gpc) 2.52 0.75 1.44
Boresight source SGR 0755$-$2933 J0837$+$0610 J0738$-$4042
\[1ex\]
------------------------------------------------------------- ------------------------- ------------------------- -------------------------
The primary advantage of the array is that a pulse from a far-field point source is detected in a maximum of 3 adjacent fan-beams at any given time, confirmed by extensive pulsar observations. RFI is typically near-field, and predominantly appears in more than 3 adjacent fan-beams, meaning that it can be reliably excised to reduce false positive rates when searching for transients. Using the adjacent fan-beam detections of FRB 160317, we have modeled the point of separation between the near-field or Fresnel region and the far-field or Fraunhofer region of the telescope. Assuming a point source at $10^{6}$ km, we compute the S/N for a tied-array beam (e.g. fan-beam 212) phased at an offset of 0.3 from the centre of the beam to ensure a two fan-beam detection. We compute the path length to each module, the phase of the signal along the array and perpendicular to the array, and add all these as a vector sum weighted by the module performance, to get the “boresight” S/N. We see that in Figure \[fig:fresnel\] at a distance of $\gtrsim 10^4$ km, we achieve a two fan-beam detection with S/N $\sim 13$ in the primary detection beam and S/N $\sim 5$ in the secondary detection beam, similar to the FRB being modeled. Detections of FRBs in 1 or 2 fan-beams only, thus allows us to identify them as sources more distant than this, placing them well away from the Earth and hence effectively rule out sources of local origin.
The discovery observations containing the FRBs were carefully inspected to check for similar events at the same time and with the same DM as the FRB, in other fan-beams. No other broadband pulses were detected in any other fan-beams within approximately 60 seconds of the bursts. Moreover, in addition to all the tied array fan-beams, we form a single special fan-beam as the incoherent sum of all the other fan-beams. This “total power” fan-beam was also searched for events near the UTC of the 3 bursts. For the 3 FRBs, this fan-beam contained no unusual sources of RFI. Only twice during the 3 surveys did we find FRB-like candidates (i.e. appearing across the band) which were identified as RFI upon closer analysis. In each case, similar events could be found in dozens to hundreds of fan beams, and were thus obvious near-field RFI. These false candidates also had ‘patchy’ power across the observing band, indicative of RFI generated from different carrier handsets operating at the same time in our band.
![Simulations of the detections of FRB 160317 in adjacent fan-beams (FB) to determine the Fresnel limit of the telescope. At a distance of $\gtrsim 10^{4}$ km, the S/Ns of the modeled pulse in FB 212 and FB 213, match that of the observations with non-detections in the other fan-beams.[]{data-label="fig:fresnel"}](Fig5.pdf){width="3.5"}
{width="6.0"}
Two of the three FRBs have been discovered relatively close to the Galactic plane, with the locations marked as stars in Figure \[fig:skydist\]. All three have DMs significantly in excess of the Galactic contribution, suggesting an extragalactic or cosmological origin. Under this assumption, the contribution from the IGM to the DM can be used to infer a redshift, using the scaling relation in [@Ioka] and [@Inoue]. This places FRBs 160317, 160410 and 160608 at a redshift upper limit of 0.7, 0.2 and 0.4 respectively, assuming zero contribution from any potential host galaxy. Any contribution from a host galaxy or the immediate vicinity of an associated source, could be a significant fraction of the total DM depending on its orientation and location. The average DM for elliptical galaxies is 37 pc cm$^{-3}$ and for spiral galaxies is 45 pc cm$^{-3}$ based on the probability distribution of DMs computed for a range of host galaxies [@Xu]. For spirals, the weighted average over a range of inclination angles is estimated to be 142 pc cm$^{-3}$. However the host contribution to the DM from high redshift galaxies can be small due to cosmological time dilation and the corresponding redshifting of frequency [@Zhou]. It also does not account for any bias in FRB locations within galaxies.
The S/Ns, DMs and widths of all three FRBs have been computed using the <span style="font-variant:small-caps;">destroy</span> single pulse search software, <span style="font-variant:small-caps;">psrchive</span> with scripts made publicly available through the FRBCAT repository. The observed widths of all three FRBs are dominated by dispersion smearing as shown in Table \[tab:specs\]. This is due to our small bandwidth and limited number of channels (40 channels). We have now implemented a fine channel mode (320 channels) which will potentially increase our sensitivity and the FRB detection rate by a factor of $\sqrt{8}$. Our total bandwidth of only 31.25 MHz is too narrow to permit a measurement of dispersion index. Single pulses from the Vela pulsar were used to test our sensitivity to the DM index. The DM and the DM index $\delta$ where the dispersive delay is given by, $$\Delta t \propto \Delta\nu^{-\delta}$$ are found to be highly correlated, so that we can place no practical limit on $\delta$. We therefore set the DM index to $\delta = -2$.
{width="5.0"}
{width="5.0"}
Localisation
------------
All 3 FRBs were discovered while following up known sources, which were at “bore-sight”, and thus centred on Beam 177. We localise each FRB’s position on the sky, using the angular separation between the FRB’s position in fan-beam space from the bore-sight fan-beam. The 1$\sigma$ uncertainty in the direction of the semi-major axis is defined by the primary beam ($\sim 2.8^{\circ}$) while the uncertainty in the semi-minor axis is controlled by the fractional S/N measured from the adjacent fan-beam detections. For FRBs detected in adjacent fan-beams, we perform a simple linear interpolation based on the S/N to localise the event in the fan-beam grid. For the FRB with a single fan-beam detection only, we assume the centre of the fan-beam for the localisation. This allows us to construct a trace on the sky relative to the bore-sight (RAJ, DECJ), taking into account the meridian angle of the observation at the time of the FRB, and the known, slight skew and tilt of the East-West arms relative to true East-West and horizontal. The trace is a strong function of the hour angle of the observation, as one would expect in an East-West array. The uncertainty in the direction of the semi-minor axis was confirmed observationally using single pulse detections from bright pulsars. Single pulses that were detected in two adjacent fan-beams with S/N similar to our FRBs were chosen to estimate our localisation accuracy on the sky. These fan-beam localisations were then compared to the true position of the pulsar. The 1$\sigma$ scatter of the calculated position of the pulsar from individual pulses, compared to the known position of the pulsar, is $\sim 0.1$ fan-beams or $\sim 5\arcsec$, in the direction perpendicular to the fan-beam. For single fan-beam detections at low S/N, a similar analysis yielded a slightly poorer localisation precision of 7$\arcsec$. Two of the FRBs found with UTMOST have FWHP error ellipses of 5$\arcsec \times 2.8^{\circ}$ ($\sim 11$ arcmin$^2$) on the sky as seen, in Figure \[fig:frbspos\]. For comparison, single beam FRB detections (with FWHP beamwidth of 14.0$\arcsec$) at Parkes, are localised to $\sim160$ arcmin$^2$. The probability density of the localisation is shown in right ascension (RA) and declination (DEC), with the cross marking the most likely position of the burst for each FRB.
One of the advantages of UTMOST is that we can localise pulses to a few arcsec radius if the source is observed at different hour angles. The tilt of the error ellipses seen in Figure \[fig:frbspos\] demonstrates this and is a result of the geometry of the telescope. The most likely position of the FRB is marked by the cross. The fan-beams rotate in position according to the hour angle pointing of the telescope. A repeat FRB pulse, at a significantly different hour angle, allows us to localise the bursts to a few arcsec, depending on the S/N of the event. We have tested this scenario using individual pulses from four different pulsars, with results for four different hour angles (and offsets from boresight) for the Vela pulsar shown in Figure \[fig:vela\]. Localisations to about 5${\arcsec}$ accuracy, or 0.1 fan-beam widths, should be possible for a bright, repeating FRB. UTMOST is part of the shadowing campaign of the Survey for Pulsars and Extragalactic Radio Bursts (SUPERB) at the Parkes radio telescope (Keane et al. in prep). An FRB detected simultaneously with Parkes and UTMOST will yield a good localisation even if the burst does not repeat. The sky positions of the 3 FRBs were also re-observed at different HAs, to look for additional bursts. We spent 105 hours following FRB 160317, 43 hours on FRB 160410 and 35 hours on FRB 160608. The data were searched offline for pulses with S/N $\geq10$ and with $\pm20\%$ of the DM of the FRB, using the pipeline described in Section \[sec:pipeline\]. No repeat pulses were found from any of the FRB positions.
FRB event rate at UTMOST {#sec:disc}
========================
Event rate analysis {#sec:rate}
-------------------
[@Liam] present detailed analyses constraining FRB rates at various telescopes, scaling from a single FRB discovered at 800 MHz at the GBT (FRB 110523) [@Masui]. They estimate a rate of $4.2^{+19.6}_{-3.2} \times 10^{-1}$ d$^{-1}$ at UTMOST operating at its design sensitivity, based on comparisons between the sensitivities and areas surveyed by the two telescopes. This estimate is consistent with [@Caleb]. Using the same method as outlined in [@Liam] we calculate a rate of $0.014^{+0.05}_{-0.013}$ d$^{-1}$, which agrees with the rate we measure at UTMOST. We have spent a total of 180 days on sky and discovered 3 FRBs with a FoV of 8.8 deg$^{2}$. Based on this, we measure a detectable event rate of ($\mathcal{R}$),
$$\label{eq:rate}
\mathcal{R}\, (\mathcal{F} \gtrsim 11 \, \mathrm{Jy\, ms}) \gtrsim 0.78^{+1.24}_{-0.57} \, \times 10^{2} \, \mathrm{events \,\, sky^{-1} \, d^{-1}}$$
at the 95 percent confidence level [@Gehrels], above a full power boresight fluence of 11 Jy ms as parametrised by Equation \[eq:lim\_flue\], at the half-power FoV. The rate is given as a lower limit since all searches are incomplete in the fluence-width plane. Following [@KeanePetroff], our fluence complete rate is $$\mathcal{R}\, (\mathcal{F} \gtrsim 69 \, \mathrm{Jy\, ms}) \sim 5.0^{+18.7}_{-4.7} \, \mathrm{events \,\, sky^{-1} \, d^{-1}},
%{\color{red}\mathcal{R}\, (\mathcal{F} \gtrsim 69 \, \mathrm{Jy\, ms}) \sim 0.049^{+0.187}_{-0.047} \times 10^{2} \, \mathrm{events \,\, sky^{-1} \, d^{-1}}.}$$
as shown in Figure \[fig:fluence\].
In [@Caleb_sim; @Caleb] we have made estimates of the event rate expected at UTMOST, scaling from the event rate at Parkes, under assumptions about the integral source count distribution (log$N$-log$\mathcal{F}$ relation) and the spectral index of FRBs. To do this we assume that the spectral energy distribution is flat between the Parkes 1.4 GHz and UTMOST’s 843 MHz operating frequencies, and that the source count distribution scales as,
$$N(>\mathcal{F}_\mathrm{lim}) = \mathcal{F}_\mathrm{lim}^{\alpha}$$
where $\alpha=-3/2$ for events populated in a Euclidean Universe. Under these assumptions, [@Caleb] predict a rate of $\gtrsim0.008(0.004)$ events d$^{-1}$ for a 10$\sigma$, 1-ms wide FRB to a minimum detectable fluence of $\mathcal{F}_\mathrm{lim} = 11$ Jy ms at boresight (see Equation \[eq:lim\_flue\]). If we correct this rate for the measured primary beam dimensions being 10% larger (Bailes et al., submitted) than adopted in [@Caleb] we get a rate of $\gtrsim0.007(0.004)$ events d$^{-1}$ which translates to being able to detect 1.3 events in 180 days on sky. This is in mild tension with our discovery of 3 events in the survey. We quantify this tension by calculating the probability of observing 3 or more events to be 14.3%, assuming Poisson statistics with a mean of 1.3. Note that the predicted rate at UTMOST takes into account pulse-width broadening in the current implementation of the backend (channel widths $\sim780$ kHz) but does not account for possibly highly atypical scintillation properties along specific lines of sight to FRB events. The difference in the estimated and measured rates could be due to FRBs being brighter than expected at 843 MHz, and/or the slope of the source count distribution $\alpha$, being shallower than the assumed value. Simultaneous broad-band detection of an FRB (e.g. Parkes + UTMOST) would help constrain the spectral index and resolve the question. In [@Caleb_sim], we measured $\alpha \approx -0.9 \pm 0.3$, from 9 FRBs discovered in the high latitude sub-survey of the HTRU survey at Parkes. This is consistent with the events occurring at cosmological distances in a $\Lambda$CDM Universe, in which the Euclidean value for $\alpha$ does not hold. Assuming a flat spectral index for FRBs ($\gamma=0$), if we scale the rate at UTMOST from Equation \[eq:rate\], assuming $\alpha=-1.0$ for the slope of the log$N$-log$\cal{F}$ relation based on the best-fit from the bottom panel in Figure \[fig:spectral\_index\], we obtain a rate of $\gtrsim 2.1 \times 10^{3}$ events sky$^{-1}$ d$^{-1}$ at 1.4 GHz, consistent with the observed Parkes rate at the 2$\sigma$ level [@Champion].
Constraints on spectral and source count distribution indices
-------------------------------------------------------------
As discussed in the previous section, the observed FRB rate at UTMOST at 843 MHz can be brought to consistency with the rate found at 1.4 GHz at Parkes if we assume FRBs are flat spectrum sources on average, and that the log$N$-log$\mathcal{F}$ relation has slope $\alpha = -1.0$ (best-fit for $\alpha$ assuming $\gamma = 0$, in bottom panel of Figure \[fig:spectral\_index\]). Alternatively, we can relax the flat spectrum assumption, and ascribe the higher than expected rate to FRBs being brighter at 843 MHz than at 1.4 GHz. Assuming the log$N$-log$\mathcal{F}$ relation has slope $\alpha$ and that the FRBs have a power law spectral energy distribution with index $\gamma$ such that $S \propto \nu^{\gamma}$, we examine the following scenarios:
1. *$\alpha=-3/2$*: Based on the detection rates at the Parkes, GBT and UTMOST telescopes, we can constrain a spectral index for FRBs as shown in Figure \[fig:spectral\_index\]. The rate of $\sim 7000$ events sky$^{-1}$ d$^{-1}$ at 1.4 GHz at Parkes, above a fluence limit of 0.4 Jy ms [@Champion] scales to $\sim56$ events sky$^{-1}$ d$^{-1}$ above UTMOST’s fluence limit of 11 Jy ms, and assuming the spectral index to be flat ($\gamma=0$) (see Section \[sec:rate\]). Similarly, the rate of $\sim 2.7 \times 10^{4}$ events sky$^{-1}$ d$^{-1}$ at 800 MHz at the GBT, above a fluence threshold of 0.3 Jy ms [@Liam] and scaled to UTMOST’s fluence threshold of 11 Jy ms is $\sim116$ events sky$^{-1}$ d$^{-1}$. Using these values we fit for the FRB spectra to be $\gamma = -1.1(1.2)$ (top panel of Figure \[fig:spectral\_index\]). This is found to be consistent with the observed constraint in [@nat_keane] albeit for just one FRB.
2. *$\gamma=0$*: A similar analysis can be done to constrain the index $\alpha$ of the integral source count distribution assuming a flat spectral index. We constrain a value of $\alpha=-1.0(1.1)$ for $\gamma=0$ (bottom panel of Figure \[fig:spectral\_index\]). This value of $\alpha$ gives scaled rates of $\sim 270$ events sky$^{-1}$ d$^{-1}$ at 1.4 GHz at Parkes and $\sim 690$ events sky$^{-1}$ d$^{-1}$ at 800 MHz at GBT.
Another possible scenario is that FRBs are giant pulses from pulsars [@Wasserman]. The average observed spectral index for pulsars is $\gamma = -1.6$ (Jankowski et al., in submitted). If we assume this to be typical of FRBs, we fit a slope of $\alpha = -1.76$ for their log$N$-log$\cal{F}$ distribution. We note that the repeat FRB pulses from the Arecibo FRB 121102 exhibit a wide range of spectral indices [$\gamma \sim \, -10 \, \mathrm{to} \,+14$ ; @nat_spitler], similar to giant pulses from the Crab pulsar. For example giant pulses from the Crab pulsar exhibit spectral volatility in their broad range of spectral indices [$\gamma \sim \, -15 \, \mathrm{to} \, +10$ ; @Karuppusamy], therefore it will be difficult to estimate the mean of the spectral indices until the numbers are sufficiently high. From simultaneous observations of FRB 150418 with Parkes at 1.4 GHz and the Murchison Widefield Array (MWA) at 150 MHz [@nat_keane], the non-detection at the MWA places a limit of $\gamma > -3.0$. [@Vfastr] estimate a weak constraint of $-7.6 < \gamma < 5.8$ based on the detection sensitivity of [@Champion] which is consistent with our estimated values. These constraints are only valid if the spectral energy distribution (SED) is an unbroken broadband power law and insignificantly affected by scintillation. This remains to be observationally proven. Future broadband instruments like CHIME [@Bandura] should have high FRB discovery rates and spectral coverage to definitively test this. Using the method in [@Caleb] we scale the observed rate at UTMOST for a boresight fluence of 11 Jy ms, to estimate the rates at CHIME and HIRAX [@Newburgh] under a Euclidean Universe assumption. We expect CHIME to detect $\sim 70$ events beam$^{-1}$ d$^{-1}$ for $T_\mathrm{sys} = 50$ K, S/N $= 10$, $G = 1.38$ K Jy$^{-1}$, $N_\mathrm{p} = 2$ and FoV = 250 deg$^{2}$ [@Connor; @Ng]. Similarly, we expect 350 events beam$^{-1}$ d$^{-1}$ at HIRAX for $T_\mathrm{sys} = 50$ K, S/N $= 10$, $G = 10.5$ K Jy$^{-1}$, $N_\mathrm{p} = 2$ and FoV = 56 deg$^{2}$ [@Newburgh].
Conclusions {#sec:conc}
===========
In this paper we present the first interferometric detections of FRBs, found during 180 days on sky at UTMOST. The events are beyond the $\approx 10^4$ km near-field limit of the telescope, ruling out local (terrestrial) sources of interference as a possible origin. We demonstrate with pulsars that a repeating FRB seen at UTMOST has the potential to be localised to $\approx 15\arcsec$ diameter error circle, an exciting prospect for identifying the host.
An all-sky rate of $\mathcal{R}(\gtrsim 11$ Jyms) $\gtrsim
0.78^{+1.24}_{-0.57} \times 10^{2}\, \mathrm{events \,\, sky^{-1} \,
d^{-1}}$ at 843 MHz is calculated from our 3 events, at the boresight fluence out to the half-power FoV. Based on the time spent on sky and the number of detections made, we measure a rate of $0.017^{+0.03}_{-0.01}$ events beam$^{-1}$ d$^{-1}$ at UTMOST, for the sensitivity achieved during the upgrade. The rates estimated by [@Caleb] for the present sensitivity, is only 14.3% unlikely assuming Poisson statistics with a mean of 1.3. One possibility could be due to the log$N$-log$\cal F$ relation for events being flatter than for a Euclidean Universe, having a slope $\alpha \approx -1.0$, rather than $\alpha = -3/2$. In this case, searching for FRBs with a less sensitive, but wider field of view instrument, appears to be a competitive strategy, [e.g. @Vedantham]. Alternatively, FRBs may simply be brighter at 843 MHz on average than at 1.4 GHz, implying a steeper spectral index for FRBs. Assuming a Euclidean Universe scaling, we find a best fit spectral index of $\gamma = -1.1(1.2)$.
Our ongoing work, and the work of others at many other facilities, will settle these questions once sufficient numbers of FRBs are detected over a broad frequency range. Understanding the spectra and log$N$-log$\mathcal{F}$ distributions are vital in the quest to understand this enigmatic population.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors would like to thank the referee for their insightful suggestions. We would also like to thank Jr-Wei Tsai and Liam Connor for useful discussions. The Molonglo Observatory is owned and operated by the University of Sydney with support from the School of Physics. The upgrade to the observatory has been supported by the University of Sydney through the Federation Fellowship FF0561298 and the Science Leveraging Fund of the New South Wales Department of Trade and Investment. Support for operations and continuing development of the observatory is provided by the University of Sydney, Swinburne University of Technology, the Australian Research Council Centre for All-Sky Astrophysics (CAASTRO), through project number CE110001020, and the Laureate Fellowship FL150100148. The late Professor George Collins allocated strategic funds for the purchase of the supercomputer in use at the facility from Swinburne University and was a passionate advocate for the project.
Positional coordinates of FRBs 160317, 160410 and 160608
========================================================
The coordinates of the FRB localisation ellipses in Figure \[fig:frbspos\] are given in Table \[table:coords\].
[c c c | c c c | c c c ]{} & &\
\[0.5ex\] RA (hrs) & DEC (deg) & Prob. & RA (hrs) & DEC (deg) & Prob. & RA (hrs) & DEC (deg) & Prob.\
\
\[0.5ex\] 7.9356 & $-32.4208$ & 0.0027 & 8.6865 & 3.3129 & 0.0028 & 7.6415 & $-43.5684$ & 0.0031\
7.9332 & $-32.2575$ & 0.0038 & 8.6867 & 3.4761 & 0.0039 & 7.6396 & $-43.4052$ & 0.0043\
7.9308 & $-32.0943$ & 0.0053 & 8.6869 & 3.6394 & 0.0054 & 7.6378 & $-43.2419$ & 0.0058\
7.9284 & $-31.9310$ & 0.0072 & 8.6870 & 3.8027 & 0.0072 & 7.6359 & $-43.0786$ & 0.0077\
7.9260 & $-31.7677$ & 0.0095 & 8.6872 & 3.9659 & 0.0096 & 7.6341 & $-42.9154$ & 0.0101\
7.9237 & $-31.6045$ & 0.0124 & 8.6874 & 4.1292 & 0.0123 & 7.6323 & $-42.7521$ & 0.0129\
7.9213 & $-31.4412$ & 0.0159 & 8.6876 & 4.2924 & 0.0156 & 7.6305 & $-42.5888$ & 0.0163\
7.9190 & $-31.2780$ & 0.0199 & 8.6878 & 4.4557 & 0.0194 & 7.6287 & $-42.4256$ & 0.0200\
7.9167 & $-31.1147$ & 0.0243 & 8.6881 & 4.6190 & 0.0236 & 7.6269 & $-42.2623$ & 0.0242\
7.9145 & $-30.9514$ & 0.0291 & 8.6883 & 4.7822 & 0.0282 & 7.6252 & $-42.0991$ & 0.0287\
7.9122 & $-30.7882$ & 0.0342 & 8.6885 & 4.9455 & 0.0330 & 7.6234 & $-41.9358$ & 0.0334\
7.9100 & $-30.6249$ & 0.0393 & 8.6888 & 5.1088 & 0.0378 & 7.6217 & $-41.7725$ & 0.0381\
7.9078 & $-30.4616$ & 0.0442 & 8.6890 & 5.2720 & 0.0425 & 7.6200 & $-41.6093$ & 0.0426\
7.9056 & $-30.2984$ & 0.0486 & 8.6893 & 5.4353 & 0.0468 & 7.6183 & $-41.4460$ & 0.0467\
7.9034 & $-30.1351$ & 0.0524 & 8.6895 & 5.5986 & 0.0505 & 7.6166 & $-41.2827$ & 0.0502\
7.9012 & $-29.9718$ & 0.0554 & 8.6898 & 5.7618 & 0.0535 & 7.6149 & $-41.1195$ & 0.0530\
7.8991 & $-29.8086$ & 0.0572 & 8.6901 & 5.9251 & 0.0555 & 7.6133 & $-40.9562$ & 0.0548\
7.8970 & $-29.6453$ & 0.0579 & 8.6904 & 6.0884 & 0.0564 & 7.6116 & $-40.7929$ & 0.0556\
7.8949 & $-29.4820$ & 0.0574 & 8.6907 & 6.2516 & 0.0562 & 7.6100 & $-40.6297$ & 0.0553\
7.8928 & $-29.3188$ & 0.0557 & 8.6910 & 6.4149 & 0.0549 & 7.6084 & $-40.4664$ & 0.0540\
7.8907 & $-29.1555$ & 0.0529 & 8.6913 & 6.5782 & 0.0526 & 7.6068 & $-40.3031$ & 0.0517\
7.8887 & $-28.9922$ & 0.0492 & 8.6916 & 6.7414 & 0.0493 & 7.6052 & $-40.1399$ & 0.0485\
7.8866 & $-28.8290$ & 0.0448 & 8.6919 & 6.9047 & 0.0454 & 7.6036 & $-39.9766$ & 0.0447\
7.8846 & $-28.6657$ & 0.0400 & 8.6923 & 7.0680 & 0.0409 & 7.6021 & $-39.8133$ & 0.0404\
7.8826 & $-28.5024$ & 0.0349 & 8.6926 & 7.2312 & 0.0362 & 7.6005 & $-39.6501$ & 0.0357\
7.8806 & $-28.3392$ & 0.0299 & 8.6930 & 7.3945 & 0.0313 & 7.5990 & $-39.4868$ & 0.0310\
7.8787 & $-28.1759$ & 0.0250 & 8.6933 & 7.5578 & 0.0266 & 7.5975 & $-39.3235$ & 0.0265\
7.8767 & $-28.0126$ & 0.0205 & 8.6937 & 7.7210 & 0.0221 & 7.5960 & $-39.1603$ & 0.0221\
7.8748 & $-27.8494$ & 0.0165 & 8.6941 & 7.8843 & 0.0181 & 7.5945 & $-38.9970$ & 0.0181\
7.8729 & $-27.6861$ & 0.0130 & 8.6945 & 8.0476 & 0.0144 & 7.5930 & $-38.8337$ & 0.0146\
7.8710 & $-27.5229$ & 0.0100 & 8.6949 & 8.2108 & 0.0113 & 7.5915 & $-38.6705$ & 0.0115\
7.8691 & $-27.3596$ & 0.0075 & 8.6953 & 8.3741 & 0.0087 & 7.5900 & $-38.5072$ & 0.0089\
7.8672 & $-27.1963$ & 0.0056 & 8.6957 & 8.5373 & 0.0065 & 7.5886 & $-38.3440$ & 0.0067\
7.8654 & $-27.0331$ & 0.0040 & 8.6961 & 8.7006 & 0.0048 & 7.5871 & $-38.1807$ & 0.0050\
7.8635 & $-26.8698$ & 0.0029 & 8.6965 & 8.8639 & 0.0035 & 7.5857 & $-38.0174$ & 0.0037\
[^1]: Email: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
In this paper, we address low-rank matrix completion problems with fixed basis coefficients, which include the low-rank correlation matrix completion in various fields such as the financial market and the low-rank density matrix completion from the quantum state tomography. For this class of problems, the efficiency of the common nuclear norm penalized estimator for recovery may be challenged. Here, with a reasonable initial estimator, we propose a rank-corrected procedure to generate an estimator of high accuracy and low rank. For this new estimator, we establish a non-asymptotic recovery error bound and analyze the impact of adding the rank-correction term on improving the recoverability. We also provide necessary and sufficient conditions for rank consistency in the sense of Bach [@Bac08], in which the concept of constraint nondegeneracy in matrix optimization plays an important role. As a byproduct, our results provide a theoretical foundation for the majorized penalty method of Gao and Sun [@GaoS10] and Gao [@Gao10] for structured low-rank matrix optimization problems.\
[**Keywords:**]{} matrix completion, fixed basis coefficients, low-rank, convex optimization, rank consistency, constraint nondegeneracy.
author:
- 'Weimin Miao[^1], Shaohua Pan[^2] and Defeng Sun[^3]'
bibliography:
- 'rank\_correction.bib'
date: 'October 13, 2012'
title: 'A Rank-Corrected Procedure for Matrix Completion with Fixed Basis Coefficients'
---
Introduction
============
The low-rank matrix completion is referred to recover an unknown low-rank matrix, exactly or approximately, from the under-sampled observations with or without noises. This problem is of considerable interest in many application areas, from machine learning to quantum state tomography. A basic idea to address a low-rank matrix completion problem is to minimize the rank of a matrix subject to certain constraints involving observations. Given that the direct minimization of rank function is generally NP-hard, a widely-used convex relaxation approach is to replace the rank function with the nuclear norm as the latter is the convex envelope of the rank function over a unit ball of the spectral norm [@Faz02].
Nuclear norm minimization (NNM) has been observed to provide a low-rank solution in practice for a long time (see, e.g., [@MesP97; @Mes98; @Faz02]). The first theoretical characterization for the minimum rank solution of the NNM was given by Recht, Fazel and Parrilo [@RecFP10], with the help of the concept of Restricted Isometric Property (RIP). Recognizing that the matrix completion problem does not obey the RIP, Cand[è]{}s and Recht [@CanR09] introduced the concept of incoherence property and proved that most low-rank matrices can be exactly recovered from a surprisingly small number of noiseless observations of randomly sampled entries via the NNM. The bound of the number of sampled entries was later improved to be near-optimal by Cand[é]{}s and Tao [@CanT10] through a counting argument. Such a bound was also obtained by Keshavan et al. [@KesMO10] for their proposed OptSpace algorithm. Later, Gross [@Gro11] sharpened the bound by employing a novel technique from quantum information theory developed in [@GroLFBE10], with the extension of noiseless observations of entries to coefficients relative to any basis. This technique was also adapted by Recht [@Rec11]. All the above results focus on noiseless matrix completion. The matrix completion with noise was first addressed by Cand[é]{}s and Plan [@CanP10]. More recently, nuclear norm penalized estimators for matrix completion with noise have been well studied by Koltchinskii, Lounici and Tsybakov [@KolLT11], Negahban and Wainwright [@NegW12], and Klopp [@Klo12]. Besides the nuclear norm, several other penalties for matrix completion have also been studied in [@RohT11; @Klo11; @Kol12; @SreRJ05; @FoyS11].
The NNM has been demonstrated to be a successful approach to encourage a low-rank solution in many situations. However, the efficiency of the NNM may be challenged under general sampling schemes. For example, the conditions characterized by Bach [@Bac08] for rank consistency of the nuclear norm penalized least squares estimator may not be satisfied. In particular, for matrix completion problems, Salakhutdinov and Srebro [@SalS10] showed that when certain rows and/or columns are sampled with high probability, the NNM may fail in the sense that the number of observations required for recovery is much more than the setting of most matrix completion problems. Negahban and Wainwright [@NegW12] also pointed out the impact of such heavy sampling schemes on the recovery error bound. As a remedy for this, a weighted nuclear norm (trace norm), based on row- and column-marginals of the sampling distribution, was suggested in [@NegW12; @SalS10; @FoySSS11] if the prior information on sampling distribution is available.
When the true matrix possesses a symmetric/Hermitian positive semidefinite structure, the impact of general sampling schemes on the recoverability of the NNM is more remarkable. In this situation, the nuclear norm reduces to the trace and thus only depends on diagonal entries rather than all entries as the rank function does. As a result, if diagonal entries are heavily sampled, the ability of the NNM to promote a low-rank solution, as well as the recoverability, will be highly weakened. This phenomenon is fully reflected in the widely-used correlation matrix completion problem, for which the nuclear norm becomes a constant and completely loses effectiveness for matrix recovery. Another example of particular interest in quantum state tomography is to recover a density matrix of a quantum system from Pauli measurements (see, e.g., [@GroLFBE10; @FlaGLE12; @Wan12]). A density matrix is a Hermitian positive semidefinite matrix of trace one. Obviously, if the constraints of positive semidefiniteness and trace one are simultaneously imposed on the NNM, the nuclear norm completely fails in promoting a low-rank solution. Thus, one of the two constraints has to be abandoned in the NNM and then be restored in the post-processing stage. In fact, this idea has been explored in [@GroLFBE10; @FlaGLE12] and the numerical results there indicated its relative efficiency though it is at best sub-optimal.
In this paper, with a strong motivation to optimally address the difficulties in correlation and density matrix completion problems, we propose a low-rank matrix completion model with fixed basis coefficients. In our setting, for any given basis of the matrix space, a few basis coefficients of the true matrix are assumed to be fixed due to a certain structure or some prior information, and the rest are allowed to be observed with noises under general sampling schemes. Certainly, one can apply the nuclear norm penalized technique to our model. The challenge is that, as argued earlier, this may not yield a desired low-rank solution with small estimation errors. Here, we introduce a rank-correction step to address this critical issue provided that a reasonable initial estimator is available. A satisfactory choice of the initial estimator is the nuclear norm penalized estimator or one of its analogies. The rank-correction step solves a convex “nuclear norm $-$ rank-correction term $+$ proximal term” regularized least squares problem with fixed basis coefficients (and the possible positive semidefinite constraint). The rank-correction term is a linear term constructed from the initial estimator, and the proximal term is a quadratic term added to ensure the boundness of the solution to the convex problem. The resulting convex matrix optimization problem can be solved by the efficient algorithms recently developed in [@JiaST12; @JiaST12_1; @JiaST12_2] even for large-scale cases.
The idea of using a two-stage or even multi-stage procedure is not brand new for dealing with sparse recovery in the statistical and machine learning literature. The $l_1$-norm penalized least squares method, also known as the Lasso [@Tib96], is very attractive and popular for variable selection in statistics, thanks to the invention of the fast and efficient LARS algorithm [@EfrHJT04]. On the other hand, the $l_1$-norm penalty has long been known by statisticians to yield biased estimators and cannot attain the estimation optimality [@FanL01; @FanP04]. The issue of bias can be overcome by nonconvex penalization methods, see, e.g., [@LenLW06; @Fan97; @Zha10]. A multi-stage procedure naturally occurs if the nonconvex problem obtained is solved by an iterative algorithm [@ZouL08]. In particular, once a good initial estimator is used, a two-stage estimator is enough to achieve the desired asymptotic efficiency, e.g., the adaptive Lasso proposed by Zou [@Zou06]. There are also a number of important papers in this line on variable selection, including [@LenLW06; @MeiB06; @ZhaY07; @HuaMZ10; @ZhoVB09; @Mei07; @FanL08], to name only a few. For a broad overview, the interested readers are referred to the recent survey papers [@FanL10; @FanLQ11]. It is natural to extend the ideas from the vector case to the matrix case. Recently, Bach [@Bac08] made an important step in extending the adaptive Lasso of Zou [@Zou06] to the matrix case for seeking rank consistency under general sampling schemes. However, it is not clear how to apply Bach’s idea to our matrix completion model with fixed basis coefficients since the required rate of convergence of the initial estimator for achieving asymptotic properties is no longer valid as far as we can see. More critically, there are numerical difficulties in efficiently solving the resulting optimization problems. Such difficulties also occur when the reweighted nuclear norm proposed by Mohan and Fazel [@MohF10] is applied to the rectangular matrix completion problems.
The rank-correction step to be proposed in this paper is for the purpose to overcome the above difficulties. This approach is inspired by the majorized penalty method recently proposed by Gao and Sun [@GaoS10] for solving structured matrix optimization problems with a low-rank constraint. For our proposed rank-correction step, we provide a non-asymptotic recovery error bound in the Frobenius norm, following a similar argument adopted by Klopp in [@Klo12]. The obtained error bound indicates that adding the rank-correction term could help to substantially improve the recoverability. As the estimator is expected to be of low-rank, we also study the asymptotic property — rank consistency in the sense of Bach [@Bac08], under the setting that the matrix size is assumed to be fixed. This setting may not be ideal for analyzing asymptotic properties for matrix completion, but it does allow us to take the crucial first step to gain insights into the limitation of the nuclear norm penalization. Among others, the concept of constraint nondegeneracy for conic optimization problem plays a key role in our analysis. Interestingly, our results of recovery error bound and rank consistency suggest a consistent criterion for constructing a suitable rank-correction function. In particular, for the correlation and density matrix completion problems, we prove that the rank consistency automatically holds for a broad selection of rank-correction functions. To achieve better performance for recovery, the rank-correction step may be iteratively used for several times, especially when the sample ratio is relatively low. Finally, we remark that our results can also be used to provide a theoretical foundation for the majorized penalty method of Gao and Sun [@GaoS10] and Gao [@Gao10] for structured low-rank matrix optimization problems. This paper is organized as follows. In Section \[section2\], we introduce the observation model of matrix completion with fixed basis coefficients and the formulation of the rank-correction step. In Section \[section3\], we establish a non-asymptotic recovery error bound and discuss the impact of the rank-correction term on recovery. Section \[section4\] provides necessary and sufficient conditions for rank consistency. Section \[section5\] is devoted to the construction of the rank-correction function. In Section \[section6\], we report numerical results to validate the efficiency of our proposed rank-corrected procedure. We conclude this paper in Section \[section7\]. All proofs are left in the Appendix.
[**Notation.**]{} Here we provide a brief summary of the notation used in this paper.
- Let $\mathbb{R}^{n_1\times n_2}$ and $\mathbb{C}^{n_1\times n_2}$ denote the space of all $n_1\times n_2$ real and complex matrices, respectively. Let $\mathcal{S}^n(\mathcal{S}_{+}^n,\,\mathcal{S}_{++}^n)$ denote the set of all $n\times n$ real symmetric (positive semidefinite, positive definite) matrices and $\mathcal{H}^n(\mathcal{H}_{+}^n,\,\mathcal{H}_{++}^n)$ denote the set of all $n\times n$ Hermitian (positive semidefinite, positive definite) matrices. Let $\mathbb{S}^n\,(\mathbb{S}^n_+,\,\mathbb{S}^n_{++})$ represent $\mathcal{S}^n\,(\mathcal{S}^n_+,\,\mathcal{S}^n_{++})$ for the real case and $\mathcal{H}^n\,(\mathcal{H}^n_+,\,\mathcal{H}^n_{++})$ for the complex case.
- Let $\mathbb{V}^{n_1\times n_2}$ represent $\mathbb{R}^{n_1\times n_2}$, $\mathbb{C}^{n_1\times n_2}$, $\mathcal{S}^n$ or $\mathcal{H}^n$. We define $n:=\min(n_1,n_2)$ for the previous two cases and stipulate $n_1=n_2=n$ for the latter two cases. Let $\mathbb{V}^{n_1\times n_2}$ be endowed with the trace inner product $\langle \cdot, \cdot \rangle$ and its induced norm $\|\cdot\|_{F}$, i.e., $\langle X, Y \rangle:= \text{Re}\big(\text{Tr}(X^{\mathbb{T}}Y)\big)$ for $X, Y\in\mathbb{V}^{n_1\times n_2}$, where $``\text{Tr}"$ stands for the trace of a matrix and $``\text{Re}"$ means the real part of a complex number.
- For the real case, $\mathbb{O}^{n\times k}$ denotes the set of all $n\times k$ real matrices with orthonormal columns, and for the complex case, $\mathbb{O}^{n\times k}$ denotes the set of all $n\times k$ complex matrices with orthonormal columns. When $k=n$, we write $\mathbb{O}^{n\times k}$ as $\mathbb{O}^n$ for short.
- The notation $^{\mathbb{T}}$ denotes the transpose for the real case and the conjugate transpose for the complex case. The notation $^\ast$ means the adjoint of operator.
- For any given vector $x$, $\text{Diag}(x)$ denotes a rectangular diagonal matrix of suitable size with the $i$-th diagonal entry being $x_i$.
- For any $x \in \mathbb{R}^n$, let $\|x\|_2$ and $\|x\|_\infty$ denote the Euclidean norm and the maximum norm, respectively. For any $X\in \mathbb{V}^{n_1\times n_2}$, let $\|X\|$ and $\|X\|_*$ denote the spectral norm and the nuclear norm, respectively.
- The notations $\stackrel{a.s.}{\rightarrow}$, $\stackrel{p}{\rightarrow}$ and $\stackrel{d}{\rightarrow}$ mean almost sure convergence, convergence in probability and convergence in distribution, respectively. We write $x_m = O_p(1)$ if $x_m$ is bounded in probability.
- For any set $K$, let $|K|$ denote the cardinality of $K$ and let $\delta_{K}(x)$ denote the indicator function of $K$, i.e., $\delta_{K}(x)=0$ if $x\in K$, and $\delta_{K}(x)=+\infty$ otherwise. Let $I_n$ denote the $n\times n$ identity matrix.
Problem formulation {#section2}
===================
In this section, we formulate the model of the matrix completion problem with fixed basis coefficients, and then propose a rank-correction step for solving this class of problems.
The observation model {#subsecobs}
---------------------
Let $\{\Theta_1,\ldots,\Theta_d\}$ be a given orthonormal basis of the given real inner product space $\mathbb{V}^{n_1\times n_2}$. Then, any matrix $X\in\mathbb{V}^{n_1\times n_2}$ can be uniquely expressed in the form of $X=\sum_{k=1}^d \langle \Theta_k, X \rangle\Theta_k$, where $\langle \Theta_k, X\rangle$ is called the basis coefficient of $X$ relative to $\Theta_k$. Let $\overline{X} \in \mathbb{V}^{n_1\times n_2}$ be the unknown low-rank matrix to be recovered. In some practical applications, for example, the correlation and density matrix completion, a few basis coefficients of the unknown matrix $\overline{X}$ are fixed (or assumed to be fixed) due to a certain structure or reliable prior information. Throughout this paper, we let $\alpha\subseteq\{1,2,\ldots,d\}$ denote the set of the indices relative to which the basis coefficients are fixed, and $\beta$ denote the complement of $\alpha$ in $\{1,2,\ldots,d\}$, i.e., $\alpha\cap\beta=\emptyset$ and $\alpha\cup\beta = \{1,\ldots, d\}$. We define $d_1: = |\alpha|$ and $d_2 := |\beta|$.
When a few basis coefficients are fixed, one only needs to observe the rest for recovering the unknown matrix $\overline{X}$. Assume that we are given a collection of $m$ noisy observations of the basis coefficients relative to $\{\Theta_{k}:k\in\beta\}$ in the following form $$\label{eqnobsori}
y_i = \left\langle \Theta_{\omega_i}, \overline{X}\right\rangle + \nu \xi_i, \quad i = 1, \ldots, m,$$ where $\omega_i$ are the indices randomly sampled from the index set $\beta$, $\xi_i$ are the independent and identically distributed (i.i.d.) noises with $\mathbb{E}(\xi_i)=0$ and $\mathbb{E}(\xi^2_i)=1$, and $\nu>0$ controls the magnitude of noise. Unless otherwise stated, we assume a general weighted sampling (with replacement) scheme with the sampling distributions of $\omega_i$ as follows.
\[asmpprob\] The indices $\omega_1,\ldots, \omega_m$ are i.i.d. copies of a random variable $\omega$ that has a probability distribution $\Pi$ over $\{1,\ldots, d\}$ defined by $${\rm Pr}(\omega = k) =\left\{\begin{array}{ll}
0 & {\rm if}\ k\in \alpha,\\
p_k>0 & {\rm if}\ k\in \beta.
\end{array}\right.$$
Note that each $\Theta_k, k\in\beta$ is assumed to be sampled with a positive probability in this sampling scheme. In particular, when the sampling probability of all $k\in \beta$ are equal, i.e., $p_k =1/d_2 \ \forall\, k\in \beta$, we say that the observations are sampled uniformly at random.
Next, we present some examples of low-rank matrix completion problems in the above settings.
(1)
: [**Correlation matrix completion.**]{} A correlation matrix is an $n\times n$ real symmetric or Hermitian positive semidefinite matrix with all diagonal entries being ones. Let $e_i$ be the vector with the $i$-th entry being one and the others being zeros. Then, $\langle e_ie_i^{\mathbb{T}}, \overline{X}\rangle =\overline{X}_{ii}=1 \ \forall\, 1\leq i\leq n$. The recovery of a correlation matrix is based on the observations of entries. For the real case, $\mathbb{V}^{n_1\times n_2} = \mathcal{S}^n$, $d = n(n+1)/2$, $d_1=n$, $$\Theta_{\alpha} = \big\{e_ie_i^{\mathbb{T}} \ | \ 1\leq i \leq n\big\} \quad \text{and} \quad \Theta_\beta = \left\{\frac{1}{\sqrt{2}}(e_ie_j^{\mathbb{T}}+e_je_i^{\mathbb{T}}) \ \Big| \ 1\leq i < j \leq n\right\};$$ and for the complex case, $\mathbb{V}^{n_1\times n_2} = \mathcal{H}^n$, $d=n^2$, $d_1 =n$, $$\Theta_\alpha =\!\big\{e_ie_i^{\mathbb{T}} \ | \ 1\leq i \leq n\big\} \ \ \text{and} \ \
\Theta_\beta\!=\!\left\{\!\frac{1}{\sqrt{2}}(e_ie_j^{\mathbb{T}}\!+e_je_i^{\mathbb{T}}),\frac{\sqrt{-1}}{\sqrt{2}}(e_ie_j^{\mathbb{T}}\!-e_je_i^{\mathbb{T}})
\ \Big| \ i<j\!\right\}.$$ Here, $\sqrt{-1}$ represents the imaginary unit. Of course, one may fix some off-diagonal entries in specific applications.
(2)
: [**Density matrix completion.**]{} A density matrix of dimension $n=2^l$ for some positive integer $l$ is an $n\times n$ Hermitian positive semidefinite matrix with trace one. In quantum state tomography, one aims to recover a density matrix from Pauli measurements (observations of the coefficients relative to the Pauli basis) [@GroLFBE10; @FlaGLE12], given by $$\Theta_\alpha = \left\{\frac{1}{\sqrt{n}}I_n\right\} \ \text{and} \ \Theta_\beta = \left\{\frac{1}{\sqrt{n}} (\sigma_{s_1}\otimes \cdots \otimes \sigma_{s_l}) \ \Big | \ (s_1,\ldots,s_l) \in \{0,1,2,3\}^l\right\}\Big \backslash \Theta_\alpha,$$ where “$\otimes$” means the Kronecker product of two matrices and $$\sigma_0 = \begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix},\ \sigma_1 = \begin{pmatrix} 0 & 1 \\ 1 & 0 \end{pmatrix}, \ \sigma_2 = \begin{pmatrix} 0 & -\sqrt{-1} \\ \sqrt{-1} & 0 \end{pmatrix}, \ \sigma_3 = \begin{pmatrix} 1 & 0 \\ 0 & -1 \end{pmatrix}$$ are the Pauli matrices. In this setting, $\mathbb{V}^{n_1\times n_2} = \mathcal{H}^{n}$, $\text{Tr}(\overline{X}) = \langle I_n, \overline{X} \rangle = 1$, $d = n^2$, and $d_1 = 1$.
(3)
: [**Rectangular matrix completion.**]{} Assume that a few entries of a rectangular matrix are known and let $\mathcal{I}$ be the index set of these entries. One aims to recover this rectangular matrix from the observations of the rest entries. In the real case, $\mathbb{V}^{n_1\times n_2} = \mathbb{R}^{n_1\times n_2}$, $d= n_1n_2$, $d_1 = |\mathcal{I}|$, $$\Theta_\alpha = \big\{e_ie_j^{\mathbb{T}} \ | \ (i,j) \in \mathcal{I}\big\} \quad \text{and} \quad \Theta_\beta = \big\{e_ie_j^{\mathbb{T}} \ | \ (i,j) \notin \mathcal{I}\big\};$$ and in the complex case, $\mathbb{V}^{n_1\times n_2} = \mathbb{C}^{n_1\times n_2}$, $d= 2n_1n_2$, $d_1 = 2|\mathcal{I}|$, $$\Theta_\alpha = \big\{e_ie_j^{\mathbb{T}}, \sqrt{-1}e_ie_j^{\mathbb{T}} \ | \ (i,j) \in \mathcal{I}\big\} \quad \text{and} \quad \Theta_\beta = \big\{e_ie_j^{\mathbb{T}}, \sqrt{-1}e_ie_j^{\mathbb{T}} \ | \ (i,j) \notin \mathcal{I}\big\}.$$
Now we introduce some linear operators that are frequently used in the subsequent sections. For any given index set $\pi \subseteq \{1,\ldots,d\}$, say $\alpha$ and $\beta$, we define the linear operators $\mathcal{R}_\pi$: $\mathbb{V}^{n_1\times n_2} \rightarrow \mathbb{R}^{|\pi|}$ and $\mathcal{P}_\pi$: $\mathbb{V}^{n_1\times n_2} \rightarrow \mathbb{V}^{n_1\times n_2}$, respectively, by $$\label{operator-R-P}
\mathcal{R}_\pi(X):= \big(\langle \Theta_k, X \rangle \big)^{\mathbb{T}}_{k\in\pi}\ \ {\rm and}\ \
\mathcal{P}_\pi(X) := \sum_{k\in \pi} \langle \Theta_k, X \rangle \Theta_k, \quad\ X \in \mathbb{V}^{n_1\times n_2}.$$ It is easy to see that $\mathcal{P}_\pi=\mathcal{R}_\pi^*\mathcal{R}_\pi$. Define the self-adjoint operators $\mathcal{Q}_\beta: \mathbb{V}^{n_1\times n_2} \rightarrow \mathbb{V}^{n_1\times n_2}$ and $\mathcal{Q}_\beta^\dag: \mathbb{V}^{n_1\times n_2} \rightarrow \mathbb{V}^{n_1\times n_2}$ associated with the sampling probability, respectively, by $$\label{operator-Q}
\mathcal{Q}_\beta(X): = \sum_{k\in \beta} p_k \langle \Theta_k, X \rangle \Theta_k \quad \text{and} \quad \mathcal{Q}_\beta^\dag (X): = \sum_{k\in \beta} \frac{1}{p_k} \langle \Theta_k, X \rangle \Theta_k,
\quad X \in \mathbb{V}^{n_1\times n_2}.$$ One may easily verify that the operators $\mathcal{Q}_\beta$, $\mathcal{Q}_\beta^\dag$ and $\mathcal{P}_\beta$ satisfy the following relations $$\label{relation-operator}
\mathcal{Q}_\beta\mathcal{Q}_\beta^\dag = \mathcal{Q}_\beta^\dag \mathcal{Q}_\beta = \mathcal{P}_\beta,\ \
\mathcal{P}_\beta\mathcal{Q}_\beta =\mathcal{Q}_\beta\mathcal{P}_\beta= \mathcal{Q}_\beta,\ \
\mathcal{Q}_\beta^\dag\mathcal{R}_{\alpha}^* = 0.$$ Let $\Omega$ be the multiset of all the sampled indices from the index set $\beta$, i.e., $\Omega =\{\omega_1, \ldots, \omega_m\}$. With a slight abuse on notation, we define the sampling operator $\mathcal{R}_\Omega$: $\mathbb{V}^{n_1\times n_2} \rightarrow \mathbb{R}^m$ associated with $\Omega$ by $$\mathcal{R}_\Omega(X) := \big(\langle \Theta_{\omega_1}, X \rangle, \ldots, \langle \Theta_{\omega_m}, X \rangle \big)^{\mathbb{T}}, \quad X \in \mathbb{V}^{n_1\times n_2}.$$ Then, the observation model (\[eqnobsori\]) can be expressed in the following vector form $$\label{eqnobs}
y = \mathcal{R}_\Omega(\overline{X})+\nu\xi,$$ where $y = (y_1,\ldots,y_m)^{\mathbb{T}} \in \mathbb{R}^m$ and $\xi =\!(\xi_1,\ldots, \xi_m)^{\mathbb{T}} \in \mathbb{R}^m$ denote the observation vector and the noise vector, respectively.
The rank-correction step
------------------------
In many situations, the nuclear norm is able to encourage a low-rank solution for matrix recovery, but its efficiency may be challenged if the observations are sampled at random obeying a general distribution such as the one considered in [@SalS10]. The setting of fixed basis coefficients in our matrix completion model can be regarded to be under an extreme sampling scheme. In particular, for the correlation and density matrix completion, the nuclear norm completely loses its efficiency since in this case it reduces to a constant. In order to overcome the shortcomings of the nuclear norm penalization, we propose a rank-correction step to generate an estimator in pursuit of a better recovery performance.
For convenience of discussions, in the rest of this paper, for any given $X\in\mathbb{V}^{n_1\times n_2}$, we denote by $\sigma(X)=\big(\sigma_1(X), \ldots, \sigma_n(X)\big)^{\mathbb{T}}$ the singular value vector of $X$ arranged in the nonincreasing order and define $$\mathbb{O}^{n_1,n_2}(X) :=\big\{(U,V) \in \mathbb{O}^{n_1}\times \mathbb{O}^{n_2}\mid X = U\text{Diag}\big(\sigma(X)\big)V^\mathbb{T}\big\}.$$ In particular, when $\mathbb{V}^{n_1\times n_2} = \mathbb{S}^n$, we denote by $\lambda(X)=\big(\lambda_1(X), \ldots, \lambda_n(X)\big)^{\mathbb{T}} $ the eigenvalue vector of $X$ with $|\lambda_1(X)|\geq \ldots \geq |\lambda_n(X)|$ and define $$\mathbb{O}^n (X) :=\big\{P \in \mathbb{O}^n \mid X = P\text{Diag}(\lambda(X))P^\mathbb{T}\big\}.$$
Before stating our rank-correction step, we introduce the concept of spectral operator associated to a symmetric vector-valued function. A function $f: \mathbb{R}^n \rightarrow \mathbb{R}^n$ is said to be symmetric if $$f(x) = Q^{\mathbb{T}} f(Qx) \quad \forall \, \text{signed permutation matrix} \ Q \ \text{and} \ x \in \mathbb{R}^n,$$ where a signed permutation matrix is a real matrix that contains exactly one nonzero entry $1$ or $-1$ in each row and column and $0$ elsewhere. From this definition, we see that $$f_i(x) = 0 \quad \text{if} \ x_i=0.$$ The spectral operator $F: \mathbb{V}^{n_1\times n_2}\rightarrow \mathbb{V}^{n_1\times n_2}$ associated with the function $f$ is defined by $$\begin{aligned}
\label{Foperator}
F(X): = U \text{Diag}\big(f(\sigma(X))\big) V^\mathbb{T},
\end{aligned}$$ where $(U,V)\in \mathbb{O}^{n_1,n_2}(X)$ and $X \in \mathbb{V}^{n_1\times n_2}$. From [@Din12 Theorems 3.1 & 3.6], the symmetry of $f$ guarantees the well-definiteness of the spectral operator $F$, and the continuous differentiability of $f$ implies the continuous differentiability of $F$. When $\mathbb{V}^{n_1\times n_2}=\mathbb{S}^n$, we have that $$F(X) = P\text{Diag}\big(f(|\lambda(X)|)\big)\big(P\text{Diag}(s(X))\big)^\mathbb{T},$$ where $P \in \mathbb{O}^n(X)$ and $s(X)\in\mathbb{R}^n$ with the $i$-th component $s_i(X)=-1$ if $\lambda_i(X)<0$ and $s_i(X)=1$ otherwise. In particular for the positive semidefinite case, both $U$ and $V$ in (\[Foperator\]) reduce to $P$. For more details on spectral operators, the readerS may refer to the PhD thesis [@Din12].
Given a spectral operator $F\!:\mathbb{V}^{n_1\times n_2} \rightarrow \mathbb{V}^{n_1\times n_2}$ and an initial estimator $\widetilde{X}_m$ for the unknown matrix $\overline{X}$, say the nuclear norm penalized least squares estimator or one of its analogies, our rank-correction step is to solve the convex optimization problem $$\label{eqnrcs}
\begin{aligned}
\min_{X\in \mathbb{V}^{n_1\times n_2}}&\ \frac{1}{2m} \left\|y - \mathcal{R}_\Omega(X)\right\|_2^2 + \rho_m \left(\|X\|_* - \langle F(\widetilde{X}_m),X\rangle + \frac{\gamma_m}{2}\|X-\widetilde{X}_m\|_F^2\right)\\
\text{s.t.}\ \ \ & \mathcal{R}_\alpha(X) = \mathcal{R}_\alpha(\overline{X}),
\end{aligned}$$ where $\rho_m > 0$ and $\gamma_m \geq 0$ are the regularization parameters depending on the number of observations. The last quadratic proximal term is added to guarantee the boundness of the solution to (\[eqnrcs\]). If the function $\|X\|_* - \langle F(\widetilde{X}_m), X \rangle$ is level-bounded, one may simply set $\gamma_m = 0$. Clearly, when $F \equiv 0$ and $\gamma_m\!=0$, the problem (\[eqnrcs\]) reduces to the nuclear norm penalized least squares problem. In the sequel, we call $-\langle F(\widetilde{X}_m), X \rangle$ the rank-correction term. If the true matrix is known to be positive semidefinite, we add the constraint $X\in\mathbb{S}^n_+$ to (\[eqnrcs\]). Thus, the rank-correction step is to solve the convex conic optimization problem $$\label{eqnrcspos}
\begin{aligned}
\min_{X\in\mathbb{S}^n} & \ \frac{1}{2m} \left\|y - \mathcal{R}_\Omega(X)\right\|_2^2 + \rho_m \left(\langle I - F(\widetilde{X}_m),X\rangle + \frac{\gamma_m}{2}\|X-\widetilde{X}_m\|_F^2\right)\\
\text{s.t.} \ & \ \mathcal{R}_\alpha(X) = \mathcal{R}_\alpha(\overline{X}),\ \ X\in\mathbb{S}^n_+.
\end{aligned}$$ For this case, we assume that the initial estimator $\widetilde{X}_m$ belongs to $\mathbb{S}_+^n$ as the projection of any estimator onto $\mathbb{S}^n_+$ can approximate the true matrix $\overline{X}$ better.
The rank-correction step above is inspired by the majorized penalty approach recently proposed by Gao and Sun [@GaoS10] for solving the rank constrained matrix optimization problem: $$\label{major1}
\min_{X \in \mathcal{C}} \big\{h(X):\ {\rm rank}(X) \leq r\big\},$$ where $r \geq 1$, $h:\mathbb{V}^{n_1\times n_2} \rightarrow \mathbb{R}$ is a given continuous function and $\mathcal{C} \in \mathbb{V}^{n_1\times n_2}$ is a closed convex set. Note that for any $X \in \mathbb{V}^{n_1\times n_2}$, the constraint $\text{rank}(X)\leq r$ is equivalent to $$0=\sigma_{r+1}(X)+\cdots + \sigma_n(X) = \|X\|_* - \|X\|_{(r)},$$ where $\|X\|_{(r)}:=\sigma_1(X)+\cdots + \sigma_r(X)$ denotes the Ky Fan $r$-norm. The central idea of the majorized penalty approach is to solve the following penalized version of (\[major1\]): $$\min_{X \in \mathcal{C}}\ h(X) + \rho\big(\|X\|_*-\|X\|_{(r)}\big),$$ where $\rho>0$ is the penalty parameter. With the current iterate $X^k$, the majorized penalty approach yields the next iterate $X^{k+1}$ by solving the convex optimization problem $$\label{eqnmpa}
\min_{X\in\mathcal{C}}\ \widehat{h}^k(X) + \rho\Big(\|X\|_* - \langle G^k, X \rangle +\frac{\gamma_k}{2}\|X-X^k\|_F^2\Big),$$ where $\gamma_k \geq 0$, $G^k$ is a subgradient of the convex function $\|X\|_{(r)}$ at $X^k$, and $\widehat{h}^k$ is a convex majorization function of $h$ at $X^k$. Comparing with (\[eqnrcs\]), one may notice that our proposed rank-correction step is close to one step of the majorized penalty approach.
Due to the structured randomness of matrix completion, we expect that the estimator generated from the rank-correction step possesses some favorable properties for recovery. The key issue is how to construct the rank-correction function $F$ to make such improvements possible. In the next two sections, we provide theoretical supports to our proposed rank-correction step, from which some important guidelines on the construction of $F$ can be captured.
Henceforth, we let $\widehat{X}_m$ denote the estimator generated from the rank-correction step (\[eqnrcs\]) or (\[eqnrcspos\]) for the corresponding cases and let $r=\text{rank}(\overline{X})\geq 1$. For any $X\in\mathbb{V}^{n_1\times n_2}$ and any $(U,V)\in \mathbb{O}^{n_1,n_2}(X)$, we write $U = [U_1 \ \ U_2 ]$ and $V =[V_1 \ \ V_2 ]$ with $U_1 \in \mathbb{O}^{n_1\times r}$, $U_2 \in \mathbb{O}^{n_1\times (n_1-r)}$, $V_1 \in \mathbb{O}^{n_2\times r}$ and $V_2 \in \mathbb{O}^{n_2\times (n_2-r)}$. In particular, for any $X\in\mathbb{S}_+^n$ and any $P \in \mathbb{O}^n(X)$, we write $P = [P_1 \ \ P_2]$ with $P_1 \in \mathbb{O}^{n \times r}$ and $P_2 \in \mathbb{O}^{n\times (n-r)}$.
Error bounds {#section3}
============
In this section, we aim to derive a recovery error bound in the Frobenius norm for the rank-correction step and discuss the impact of the rank-correction term on the obtained bound. The following analysis focuses on the rectangular case. All the results obtained in this section are applicable to the positive semidefinite case since adding more prior information can only improve recoverability.
We first introduce the orthogonal decomposition $\mathbb{V}^{n_1\times n_2}=T\oplus T^\perp$ with $$\begin{gathered}
T:=\Big\{X\in \mathbb{V}^{n_1\times n_2} \ |\ X = X_1 + X_2\ {\rm with}\ {\rm col}(X_1)\subseteq {\rm col}(\overline{X}),
{\rm row}(X_2)\subseteq {\rm row}(\overline{X})\Big\},\\
T^{\bot}:=\Big\{X\in \mathbb{V}^{n_1\times n_2}\ |\ {\rm row}(X)\perp {\rm row}(\overline{X})\ {\rm and}\
{\rm col}(X)\perp {\rm col}(\overline{X})\Big\},
\end{gathered}$$ where ${\rm row}(X)$ and ${\rm col}(X)$ denote the row space and column space of the matrix $X$, respectively. Let $\mathcal{P}_{T}:\mathbb{V}^{n_1\times n_2} \rightarrow \mathbb{V}^{n_1\times n_2}$ and $\mathcal{P}_{T^\perp}: \mathbb{V}^{n_1\times n_2} \rightarrow \mathbb{V}^{n_1\times n_2}$ be the orthogonal projection operators onto the subspaces $T$ and $T^\perp$, respectively. It is not hard to verify that $$\label{Operator-PT}
\mathcal{P}_T(X) = \overline{U}_1\overline{U}_1^\mathbb{T}X + X \overline{V}_1\overline{V}_1^\mathbb{T} - \overline{U}_1\overline{U}_1^\mathbb{T}X \overline{V}_1 \overline{V}_1^\mathbb{T} \ \ \text{and}\ \
\mathcal{P}_{T^\perp}(X) = \overline{U}_2\overline{U}_2^\mathbb{T} X \overline{V}_2\overline{V}_2^\mathbb{T}$$ for any $X\in\mathbb{V}^{n_1\times n_2}$ and $(\overline{U},\overline{V})\in\mathbb{O}^{n_1,n_2}(\overline{X})$. Define $a_m$ and $b_m$, respectively, by $$\label{eqndelalpbet}
a_m : = \|\overline{U}_1\overline{V}_1^\mathbb{T} - \mathcal{P}_T(F(\widetilde{X}_m)+\gamma_m\widetilde{X}_m))\| \ \ \text{and}\ \
b_m: = 1- \|\mathcal{P}_{T^\perp} (F(\widetilde{X}_m)+\gamma_m\widetilde{X}_m))\|.$$
Note that the first term in the objective function of (\[eqnrcs\]) can be rewritten as $$\frac{1}{2m} \left\|y-\mathcal{R}_\Omega(X)\right\|_2^2 = \frac{1}{2m}\left\|\mathcal{R}_\Omega(X-\overline{X})\right\|_2^2 - \frac{\nu}{m}\left\langle \mathcal{R}_\Omega^*(\xi), X\right\rangle.$$ Using the optimality of $\widehat{X}_m$ to the problem (\[eqnrcs\]), we obtain the following result.
\[thmopbd\] Assume that $\|\mathcal{P}_{T^\perp}(F(\widetilde{X}_m)\!+\gamma_m\widetilde{X}_m)\|\!<1$. For any given $\kappa>1$, if $$\label{eqndefrho}
\rho_m \geq \frac{\kappa \nu}{b_m}\Big\|\frac{1}{m}\mathcal{R}_\Omega^*(\xi)\Big\|,$$ then the following inequality holds: $$\label{eqnopbd}
\frac{1}{2m} \big\|\mathcal{R}_\Omega(\widehat{X}_m\!-\overline{X})\big\|_2^2 \le \!\sqrt{2r} \Big(a_m\!+\!\frac{b_m}{\kappa}\Big)\rho_m\|\widehat{X}_m\!-\overline{X}\|_F +\frac{\rho_m\gamma_m}{2}
\left(\!\|\overline{X}\|_F^2 -\!\|\widehat{X}_m\|_F^2\right).$$
Theorem \[thmopbd\] shows that, to derive an error bound on $\|\widehat{X}_m-\overline{X}\|_{F}$, we only need to establish the relation between $\|\widehat{X}_m-\overline{X}\|_F^2$ and $\frac{1}{m}\|\mathcal{R}_\Omega(\widehat{X}_m-\overline{X})\|_{2}^2$. It is well known that the sampling operator $\mathcal{R}_\Omega$ does not satisfy the RIP, but it has a similar property with high probability under certain conditions (see, e.g., [@NegW12; @KolLT11; @Klo12; @Liu11]). For deriving such a property, here, we impose a bound restriction on the true matrix $\overline{X}$ in the form of $\|\mathcal{R}_\beta(\overline{X})\|_\infty\leq c$. This condition is very mild since a bound is often known in some applications such as in the correlation and density matrix completion. Correspondingly, we add the bound constraint $\|\mathcal{R}_\beta(X)\!\|_\infty\!\leq c$ to the problem (\[eqnrcs\]) in the rank-correction step. Since the feasible set is bounded in this case, we simply set $\gamma_m =0$ and let $\widehat{X}_m^{c}$ denote the estimator generated from the rank-correction step in this case.
The above boundedness setting is similar to the one adopted by Klopp [@Klo12] for the nuclear norm penalized least squares estimator. A slight difference is that the upper bound is imposed on the basis coefficients of $\overline{X}$ relative to $\{\Theta_{k}:k\in\beta\}$ rather than all the entries of $\overline{X}$. It is easy to see that if the bound is not too tight, the estimator $\widehat{X}_m^{c}$ is the same as $\widehat{X}_m$. Therefore, we next derive the recovery error bound of $\widehat{X}_m^c$ instead of $\widehat{X}_m$, by following Klopp’s arguments in [@Klo12], which are also in line with the work done by Negahban and Wainwright [@NegW12].
Let $\mu_1$ be a constant to control the smallest sampling probability for observations as $$\label{pk}
p_k \geq (\mu_1 d_2)^{-1}\quad \forall\, k\in\beta.$$ It follows from Assumption \[asmpprob\] that $\mu_1 \geq 1$ and in particular $\mu_1=1$ for the uniform sampling. Note that the magnitude of $\mu_1$ does not depend on $d_2$ or the matrix size. By the definition of $\mathcal{Q}_\beta$, we then have $$\label{eqndefiot}\langle \mathcal{Q}_\beta(\Delta), \Delta \rangle \geq (\mu_1 d_2)^{-1}\|\Delta\|_F^2 \quad \forall \, \Delta \in \{\Delta \in \mathbb{V}^{n_1\times n_2} \mid \mathcal{R}_\alpha(\Delta) = 0\}.$$ Let $\{\epsilon_1,\ldots, \epsilon_m\}$ be an i.i.d. Rademacher sequence, i.e., an i.i.d. sequence of Bernoulli random variables taking the values $1$ and $-1$ with probability $1/2$. Define $$\label{defvaritheta}
\vartheta_m := \mathbb{E} \Big\|\frac{1}{m}\mathcal{R}_\Omega^*(\epsilon)\Big\|\ \ {\rm with}\ \
\epsilon= (\epsilon_1,\ldots,\epsilon_m)^{\mathbb{T}}.$$ Then, we can obtain a similar result to [@Klo12 Lemma 12] by showing that the sampling operator $\mathcal{R}_\Omega$ satisfies some approximate RIP for the matrices in the following set $$\begin{aligned}
\mathcal{C}(r): = & \bigg\{\Delta \in \mathbb{V}^{n_1\times n_2}\mid\ \mathcal{R}_\alpha(\Delta)=0,\ \|\mathcal{R}_\beta(\Delta)\|_\infty=1,\ \|\Delta\|_* \leq \sqrt{r}\|\Delta\|_F,\\
& \qquad\qquad\qquad\qquad\qquad\qquad \langle \mathcal{Q}_\beta(\Delta), \Delta \rangle \geq \sqrt{\frac{64\log(n_1+n_2)}{\log(2)m}}\bigg\}.
\end{aligned}$$
\[lemiso\] For all matrices $\Delta \in \mathcal{C}(r)$, with probability at least $1\!-2/(n_1\!+n_2)$, we have $$\frac{1}{m}\|\mathcal{R}_\Omega(\Delta)\|_2^2 \geq \frac{1}{2} \langle \mathcal{Q}_\beta(\Delta),\Delta\rangle - 128\mu_1 d_2 r \vartheta_m^2.$$
Now, by combining Theorem \[thmopbd\] and Lemma \[lemiso\], we obtain the following result.
\[thmbdmid\] Assume that $\|\mathcal{P}_{T^\perp} (F(\widetilde{X}_m))\|<1$ and $\|\mathcal{R}_\beta(\overline{X})\|_\infty \le c$ for some constant $c$. If $\rho_m$ is chosen to satisfy (\[eqndefrho\]), then there exists a numerical constant $C$ such that $$\begin{aligned}
\frac{\|\widehat{X}_m^c \!- \!\overline{X}\|_F^2}{d_2}\! \leq \! C \max\!\Bigg\{\!\mu_1^2 d_2 r \!\left(\!\Big(a_m \!+ \! \frac{b_m}{\kappa}\Big)^2\!\rho_m^2\! +\! \frac{\kappa^2(a_m\!+b_m)^2}{(\kappa-1)^2b_m^2}c^2 \vartheta_m ^2 \!\right)\!,
c^2 \mu_1 \sqrt{\frac{\log(n_1\!+\!n_2)}{m}} \Bigg\}
\end{aligned}$$ with probability at least $1-{2}/{(n_1\!+n_2)}$.
In order to choose a parameter $\rho_m$ such that (\[eqndefrho\]) holds, we need to estimate $\|\frac{1}{m}\mathcal{R}_\Omega^*(\xi)\|$. For this purpose, we make the following assumption on the noises.
\[asmpnoi\] The i.i.d. noise variables $\xi_i$ are sub-exponential, i.e., there exist positive constants $c_1, c_2$ and $c_3$ such that for all $t>0$, ${\rm Pr}(|\xi_i| \geq t)\leq c_1\exp(-c_2t^{c_3}).$
The noncommutative Bernstein inequality is a useful tool for the study of matrix completion problems. It provides bounds of the probability that the sum of random matrices deviates from its mean in the operator norm (see, e.g., [@Rec11; @Tro11; @Gro11]). Recently, the noncommutative Bernstein inequality was extended by replacing bounds of the operator norm of matrices with bounds of the Orlicz norms (see [@Kol12; @KolLT11]). Given any $s\geq 1$, the $\psi_s$ Orlicz norm of a random variable $\theta$ is defined by $$\|\theta\|_{\psi_s}:=\inf\big\{t>0 \mid\ \mathbb{E}\exp(|\theta|^s / t^s) \leq 2\big\}.$$ The Orlicz norms are useful to characterize the tail behavior of random variables. The following noncommutative Bernstein inequality is taken from [@Kol11 Corollary 2.1].
\[propnbi\] Let $Z_1,\ldots, Z_m\in\mathbb{V}^{n_1\times n_2}$ be independent random matrices with mean zero. Suppose that $
\max\big\{\big\|\|Z_i\|\big\|_{\psi_s},2\mathbb{E}^{\frac{1}{2}}(\|Z_i\|^2)\big\} <\varpi_{s}
$ for some constant $\varpi_{s}$. Define $$\sigma_Z := \max\left\{\bigg\|\frac{1}{m} \sum_{i=1}^m \mathbb{E}(Z_iZ_i^\mathbb{T})\bigg\|^{1/2},\ \bigg\|\frac{1}{m} \sum_{i=1}^m \mathbb{E}(Z_i^\mathbb{T}Z_i)\bigg\|^{1/2}\right\}.$$ Then, there exists a constant $C$ such that for all $t>0$, with probability at least $1\!-\exp(-t)$, $$\bigg\|\frac{1}{m} \sum_{i=1}^m Z_i\bigg\| \leq C \max\left\{ \sigma_Z\sqrt{\frac{t+\log(n_1+n_2)}{m}}, \varpi_{s}\left(\log\frac{\varpi_{s}}{\sigma_Z}\right)^{1/s}\frac{t+\log(n_1\!+n_2)}{m}\right\}.$$
It is known that a random variable is sub-exponential if and only its $\psi_1$ Orlicz norm is finite [@MilS86]. To apply the noncommutative Bernstein inequality, we let $\mu_2$ be a constant such that $$\label{eqndefL}
\max\left\{\bigg\|\sum_{k\in\beta} p_k\Theta_k\Theta_k^\mathbb{T}\bigg\|,\ \bigg\|\sum_{k\in\beta} p_k\Theta_k^\mathbb{T}\Theta_k\bigg\|\right\} \leq \frac{\mu_2}{n}.$$ Notice that since $\text{Tr}\big(\sum_{k\in\beta} p_k\Theta_k\Theta_k^\mathbb{T}\big) = \text{Tr}\big(\sum_{k\in\beta} p_k\Theta_k^\mathbb{T}\Theta_k\big)=1$, the lower bound of the term on the left-hand side is $1/n$. This implies that $\mu_2 \geq 1$. In the following, we also assume that the magnitude of $\mu_2$ does not depend on the matrix size. For example, $\mu_2=1$ for the correlation matrix completion under uniform sampling and the density matrix completion described in Section \[section2\]. The following result extends [@KolLT11 Lemma 2] and [@Klo12 Lemmas 5 & 6] from the standard basis to an arbitrary orthonormal basis. A similar result can also be found in [@NegW12 Lemma 6].
\[lemben\] Under Assumption \[asmpnoi\], there exists a constant $C^*$ (only depending on the $\psi_1$ Orlicz norm of $\xi_k$) such that for all $t>0$, with probability at least $1-\exp(-t)$, $$\label{eqnnoibd}
\left\|\frac{1}{m}\mathcal{R}^*_\Omega(\xi)\right\| \leq C^*\max\left\{\sqrt{\frac{\mu_2(t+\log(n_1+n_2))}{mn}}, \frac{\log(n)(t+\log(n_1+n_2))}{m}\right\}.$$ In particular, when $m\geq n\log^3(n_1+n_2)/\mu_2$, we also have $$\label{eqnnoiexpd}
\mathbb{E}\bigg\|\frac{1}{m}\mathcal{R}^*_\Omega(\xi)\bigg\| \leq C^*\sqrt{\frac{2e\mu_2\log(n_1+n_2)}{mn}}.$$
Since Bernoulli random variables are sub-exponential, the right-hand side of (\[eqnnoiexpd\]) provides an upper bound of $\vartheta_m$ defined by (\[defvaritheta\]). Now, we choose $t=\log(n_1\!+n_2)$ in Lemma \[lemben\] for achieving an optimal order bound. With this choice, when $m \geq 2n\log^2(n_1\!+n_2)/\mu_2$, the first term in the maximum of (\[eqnnoibd\]) dominates the second term. Hence, for any given $\kappa>1$, by choosing $$\label{eqnrhoopt}
\rho_m = \frac{\kappa \nu}{b_m} C^*\sqrt{\frac{2\mu_2\log(n_1+n_2)}{mn}},$$ from Theorem \[thmbdmid\] and Lemma \[lemben\], we obtain the following main result for recovery error bound.
\[thmstobd\] Assume that $\|\mathcal{P}_{T^\perp} (F(\widetilde{X}_m))\|\!<1$, $\|\mathcal{R}_\beta(\overline{X})\|_\infty\!\leq c$ for some constant $c$, and Assumption \[asmpnoi\] holds. For any given $\kappa>1$, if $\rho_m$ is chosen according to (\[eqnrhoopt\]), then there exists a numerical constant $C'$ such that, when $m \geq n\log^3(n_1\!+n_2)/\mu_2$, $$\begin{aligned}
\frac{\|\widehat{X}_m^c\!-\overline{X}\|_F^2}{d_2}\! \leq \!C'\max\Bigg\{& \! \left[\Big(1\!+\!\kappa \frac{a_m}{b_m}\Big)^2\nu^2\!+\!\Big(\frac{\kappa}{\kappa\!-\!1}\Big)^2 \left(1\!+\!\frac{a_m}{b_m}\right)^2c^2\right]\frac{\mu_1^2\mu_2d_2r\log(n_1\!+\!n_2)}{mn},\nonumber \\
& \qquad\qquad\qquad\qquad\qquad\qquad\qquad c^2 \mu_1 \sqrt{\frac{\log(n_1+n_2)}{m}} \Bigg\}
\label{eqnstobd}
\end{aligned}$$ with probability at least $1-3/(n_1\!+n_2)$.
When the matrix size is large, the second term in the maximum of (\[eqnstobd\]) is negligible, compared with the first term. Thus, Theorem \[thmstobd\] indicates that for any rank-correction function such that $\|\mathcal{P}_{T^\perp} (F(\widetilde{X}_m))\|<1$, one only needs samples with size of order $d_2 r \log(n_1+n_2)/n$ to control the recovery error. Note that $d_2$ is of order $n_1n_2$ in general. Hence, the order of sample size needed is roughly the degree of freedom of a rank $r$ matrix up to a logarithmic factor in the matrix size. In addition, it is very interesting to notice that the value of $\kappa$ (or the value of $\rho_m$) has a substantial influence on the recovery error bound. The first term in the maximum of (\[eqnstobd\]) is a sum of two parts related to $\nu$ and $c$, respectively. The part related to $\nu$ will increase as $\kappa$ increases provided $a_m/b_m>0$, while the part related to $c$ will slightly decreases to its limit as $\kappa$ increases.
Theorem \[thmstobd\] also reveals the impact of the rank-correction term on recovery error. Note that the value of $a_m/b_m$ fully depends on the rank-correction function $F$ when an initial estimator $\widetilde{X}_m$ is given. A smaller value of $a_m/b_m$ brings a smaller error bound and potentially leads to a smaller recovery error for the rank-correction step. Note that for any given $\varepsilon_1\ge 0$ and $0\le \varepsilon_2<1$, we have $$\begin{aligned}
\frac{a_m}{b_m} \leq \frac{\varepsilon_1}{1-\varepsilon_2}\quad {\rm if}\quad
\big\|\mathcal{P}_T\big(F(\widetilde{X}_m)) - \overline{U}_1\overline{V}_1^\mathbb{T}\big\|\le\varepsilon_1\ {\rm and}\
\big\|\mathcal{P}_{T^\perp} \big(F(\widetilde{X}_m) \big)\big\|\leq \varepsilon_2.
\end{aligned}$$ In particular, if $F\equiv 0$, then the estimator of the rank-correction step reduces to the nuclear norm penalized least squares estimator with $a_m/b_m=1$. Thus, Theorem \[thmstobd\] shows that, with a suitable rank-correction function $F$, the estimator generated from the rank-correction step for recovery is very likely to perform better than the nuclear norm penalized least squares estimator. In addition, this observation also provides us clues on how to construct a good rank-correction function, to be discussed in Section \[section5\].
Rank consistency {#section4}
================
In this section we study the asymptotic behavior of the rank of the estimator $\widehat{X}_m$ for both the rectangular case and the positive semidefinite case. Theorem \[thmstobd\] shows that under mild conditions, the distribution of $\widehat{X}_m$ becomes more and more concentrated to the true matrix $\overline{X}$. Due to the low-rank structure of $\overline{X}$, we expect that the estimator $\widehat{X}_m$ has the same low-rank property as $\overline{X}$. For this purpose, we consider the rank consistency in the sense of Bach [@Bac08] under the setting that the matrix size is fixed.
An estimator $X_m$ of the true matrix $\overline{X}$ is said to be rank consistent if $$\lim\limits_{m\rightarrow \infty}{\rm Pr}\big({\rm rank}(X_m)= {\rm rank}(\overline{X})\big)=1.$$
Throughout this section we make the following assumptions:
\[asmpfun\] The spectral operator $F$ is continuous at $\overline{X}$.
\[asmpini\] The initial estimator $\widetilde{X}_m$ satisfies $\widetilde{X}_m \stackrel{p}{\rightarrow} \overline{X}$ as $m\rightarrow \infty$.
In addition, we also need the following properties of the operator $\mathcal{R}_\Omega$ and its adjoint $\mathcal{R}_\Omega^*$.
\[lemoper\] [**(i)**]{} For any given $X \in \mathbb{V}^{n_1\times n_2}$, the random matrix $\displaystyle{\frac{1}{m}} \mathcal{R}_\Omega^* \mathcal{R}_\Omega(X) \stackrel{a.s.}{\rightarrow} \mathcal{Q}_\beta(X)$.\
[**(ii)**]{} The random vector $\displaystyle{\frac{1}{\sqrt{m}}}\mathcal{R}_{\alpha\cup\beta}\mathcal{R}_\Omega^*(\xi) \stackrel{d}{\rightarrow} N\big(0, {\rm Diag}(p)\big)$, where $p=(p_1,\ldots,p_d)^{\mathbb{T}}$.
Epi-convergence in distribution is useful in proving the convergence in distribution of minimizers or $\varepsilon_m$-minimizers. The following epi-convergence result is taken from [@Kni99].
\[propepicon\] Let $\{\Phi_m\}$ be a sequence of random lower-semicontinuous functions that epi-converges in distribution to $\Phi$. Assume that
(i)
: $\widehat{x}_m$ is an $\varepsilon_m$-minimizer of $\Phi_m$, i.e., $\Phi_m(\widehat{x}_m)\leq \inf \Phi_m(x) +\varepsilon_m$, where $\varepsilon_m \stackrel{p}{\rightarrow} 0$;
(ii)
: $\widehat{x}_m= O_p(1)$;
(iii)
: the function $\Phi$ has a unique minimizer $\overline{x}$.
Then, $\widehat{x}_m \stackrel{d}{\rightarrow} \overline{x}$. In addition, if $\Phi$ is a deterministic function, then $\widehat{x}_m \stackrel{p}{\rightarrow} \overline{x}$.
We know fFrom [@Gey96] that $\widehat{x}_m$ is guaranteed to be $O_p(1)$ when all $\Phi_m$ are convex functions and $\Phi$ has a unique minimizer. For more details on epi-convergence in distribution, one may refer to King and Wets [@KinW91], Geyer [@Gey94], Pflug [@Pfl95] and Knight [@Kni99]. In order to apply the epi-convergence theorem to a constrained optimization problem, we need to transform the constrained optimization problem into an unconstrained one by using the indicator function of the feasible set. This leads to the epi-convergence issue of the sum of two sequences of functions. Thus, we need the following epi-convergence result stated in [@Pfl95 Lemma 1].
\[propepisum\] Let $\{\Phi_m\}$ be a sequence of random lower-semicontinuous functions and $\{\Psi_m\}$ be a sequence of deterministic lower-semicontinuous functions. If either of the following two assumptions holds:
(i)
: $\Phi_m$ epi-converges in distribution to $\Phi$ and $\Psi_m$ converges to $\Psi$ with respect to the topology of uniform convergence on compact sets;
(ii)
: $\Phi_m$ converges in distribution to $\Phi$ with respect to the topology of uniform convergence on compact sets and $\Psi_m$ epi-converges to $\Psi$,
then $\Phi_m + \Psi_m$ epi-converges in distribution to $\Phi+\Psi$.
Based on the above epi-convergence results, we can analyze the asymptotic behavior of optimal solutions of a sequence of constrained optimization problems. The following result is a direct consequence of the above epi-convergence theorems and Lemma \[lemoper\].
\[thmcons\] If $\rho_m \rightarrow 0$ and $\gamma_m = O_p(1)$, then $\widehat{X}_m \stackrel{p}{\rightarrow} \overline{X}$ as $m\rightarrow \infty$.
Then, according to Theorem \[thmcons\] and lower semi-continuity of the rank function, it is straightforward to obtain:
\[cororkrhs\] If $X_m \stackrel{p}{\rightarrow} \overline{X}$, then $\lim\limits_{m\rightarrow \infty}{\rm Pr}\big({\rm rank}(X_m) \geq {\rm rank}(\overline{X})\big)=1$.
In what follows, we focus on the characterization of necessary and sufficient conditions for rank consistency of $\widehat{X}_m$. The idea is similar to that of [@Bac08] for the nuclear norm penalized least squares estimator. Note that, unlike for the recover error bound, adding more constraints may break the rank consistency. Therefore, we separate the discussion into the rectangular case and the positive semidefinite case below.
The rectangular case
--------------------
Since we have established that $\widehat{X}_m \stackrel{p}{\rightarrow} \overline{X}$, we only need to focus on some neighborhood of $\overline{X}$ for the discussion about the rank consistency of $\widehat{X}_m$. First, we take a look at a local property of the rank function via the directional derivative of the singular value functions.
Let $\sigma'_i(X;\cdot)$ denote the directional derivative function of the $i$-th largest singular value function $\sigma_i(\cdot)$ at $X$. From [@Lew05 Section 5.1] and [@DinST10 Proposition 6], for $\mathbb{V}^{n_1\times n_2} \ni H \rightarrow 0$, $$\label{eqnlocsin}
\sigma_i(X + H)-\sigma_i(X)-\sigma'_i (X;H) = O(\|H\|_F^2), \quad i = 1,\ldots, n.$$ Recall that $r=\text{rank}(\overline{X})$. From [@DinST10 Proposition 6], we have $$\sigma_{r+1}'(\overline{X};H) = \|\overline{U}_2^\mathbb{T}H\overline{V}_2\|, \quad\ H\in\mathbb{V}^{n_1\times n_2}.$$ This leads to the following result for the perturbation of the rank function. A similar result can also be found in [@Bac08 Proposition 18], whose proof is more involved.
\[lemlocrk\] Let $\overline{\Delta} \in \mathbb{V}^{n_1\times n_2}$ satisfy $\overline{U}_2^\mathbb{T} \overline{\Delta}\,\overline{V}_2 \neq 0$. Then, for all $\rho\neq 0$ sufficiently small and $\Delta$ sufficiently close to $\overline{\Delta}$, ${\rm rank}(\overline{X}+\rho \Delta) > {\rm rank}(\overline{X})$.
To guarantee the efficiency of the rank-correction term on encouraging a low-rank solution, the parameter $\rho_m$ should not decay too fast. Define $\widehat{\Delta}_m: = \rho_m^{-1}(\widehat{X}_m-\overline{X})$. Then, for a slow decay on $\rho_m$, we can establish the following result.
\[propdellim\] If $\rho_m \rightarrow 0, \sqrt{m}\rho_m\rightarrow \infty$ and $\gamma_m = O_p(1)$, then $\widehat{\Delta}_m \stackrel{p}{\rightarrow} \widehat{\Delta}$, where $\widehat{\Delta}$ is the unique optimal solution to the following convex optimization problem $$\label{eqndellim}
\begin{aligned}
\min_{\Delta \in \mathbb{V}^{n_1\times n_2}} &\ {\displaystyle \frac{1}{2}} \langle \mathcal{Q}_\beta(\Delta), \Delta \rangle + \langle \overline{U}_1 \overline{V}_1^\mathbb{T} - F(\overline{X}), \Delta \rangle + \|\overline{U}_2^\mathbb{T} \Delta \overline{V}_2\|_*\\
{\rm s.t.}\ \ \ &\ \mathcal{R}_\alpha(\Delta) = 0.
\end{aligned}$$
Note that $\widehat{X}_m = \overline{X} + \rho_m \widehat{\Delta}_m$. From Corollary \[cororkrhs\], Lemma \[lemlocrk\] and Proposition \[propdellim\], we see that the condition $\overline{U}_2^\mathbb{T}\widehat{\Delta}\overline{V}_2=0$ is necessary for the rank consistency of $\widehat{X}_m$. From the following property of the unique solution $\widehat{\triangle}$ to (\[eqndellim\]), we can derive a more detailed necessary condition for rank consistency as stated in Theorem \[thmnes\] below.
\[lemdelcond\] Let $\widehat{\Delta}$ be the optimal solution to (\[eqndellim\]). Then $\overline{U}_2^\mathbb{T} \widehat{\Delta} \overline{V}_2 = 0$ if and only if the linear system $$\label{eqndelcond}
\overline{U}_2^\mathbb{T}\mathcal{Q}_\beta^\dag(\overline{U}_2 \Gamma\overline{V}_2^\mathbb{T})\overline{V}_2 = \overline{U}_2^\mathbb{T}\mathcal{Q}_\beta^\dag\big(\overline{U}_1 \overline{V}_1^\mathbb{T}-F(\overline{X})\big)\overline{V}_2$$ has a solution $\widehat{\Gamma} \in \mathbb{V}^{(n_1-r)\times (n_2-r)}$ with $\|\widehat{\Gamma}\| \leq 1$. Moreover, in this case, $$\label{eqndelvaldel}
\widehat{\Delta} = \mathcal{Q}_\beta^\dag\big(\overline{U}_2\widehat{\Gamma}\,\overline{V}_2^\mathbb{T}-\overline{U}_1\overline{V}_1^\mathbb{T} + F(\overline{X})\big).$$
\[thmnes\] If $\rho_m \rightarrow 0$, $\sqrt{m}\rho_m\rightarrow \infty$ and $\gamma_m = O_p(1)$, then a necessary condition for the rank consistency of $\widehat{X}_m$ is that the linear system (\[eqndelcond\]) has a solution $\widehat{\Gamma} \in \mathbb{V}^{(n_1-r)\times (n_2-r)}$ with $\|\widehat{\Gamma}\|\le 1$.
By making a slight modification for the necessary condition in Theorem \[thmnes\], we provide a sufficient condition for the rank consistency of the estimator $\widehat{X}_m$ as follows.
\[thmsuf\] If $\rho_m \rightarrow 0, \sqrt{m}\rho_m\rightarrow \infty$ and $\gamma_m = O_p(1)$, then a sufficient condition for the rank consistency of the estimator $\widehat{X}_m$ is that the linear system (\[eqndelcond\]) has a unique solution $\widehat{\Gamma} \in \mathbb{V}^{(n_1-r)\times (n_2-r)}$ with $\|\widehat{\Gamma}\| < 1$.
The positive semidefinite case
------------------------------
For the positive semidefinite case, we first need the following Slater condition.
\[asmpslater\] There exists some $X^0 \in \mathbb{S}^n_{++}$ such that $\mathcal{R}_\alpha(X^0) = \mathcal{R}_\alpha(\overline{X})$.
\[propdellimpos\] If $\rho_m \rightarrow 0, \sqrt{m}\rho_m\rightarrow \infty$ and $\gamma_m = O_p(1)$, then $\widehat{\Delta}_m \stackrel{p}{\rightarrow} \widehat{\Delta}$, where $\widehat{\Delta}$ is the unique optimal solution to the following convex optimization problem $$\label{eqndellimpos}
\begin{aligned}
\min_{\Delta \in \mathbb{S}^n} & \ \ {\displaystyle \frac{1}{2}} \langle \mathcal{Q}_\beta(\Delta), \Delta \rangle + \langle I_n - F(\overline{X}), \Delta \rangle\\
{\rm s.t.} & \ \ \mathcal{R}_\alpha(\Delta) = 0, \quad \overline{P}_2^\mathbb{T} \Delta \overline{P}_2 \in \mathbb{S}_+^{n-r}.
\end{aligned}$$
For the optimal solution $\widehat{\Delta}$ to (\[eqndellimpos\]), we also have the following further characterization.
\[lemdelcondpos\] Let $\widehat{\Delta}$ be the optimal solution to (\[eqndellimpos\]). Then $\overline{P}_2^\mathbb{T} \widehat{\Delta} \overline{P}_2 = 0$ if and only if the linear system $$\label{eqndelcondpos}
\overline{P}_2^\mathbb{T}\mathcal{Q}_\beta^\dag(\overline{P}_2 \Lambda \overline{P}_2^\mathbb{T})\overline{P}_2 = \overline{P}_2^\mathbb{T}\mathcal{Q}_\beta^\dag\big(I_n- F(\overline{X})\big)\overline{P}_2$$ has a solution $\widehat{\Lambda} \in \mathbb{S}_+^{n-r}$. Moreover, in this case, $$\label{eqndelvaldelpos}
\widehat{\Delta} = \mathcal{Q}_\beta^\dag\big(\overline{P}_2\widehat{\Lambda}\,\overline{P}_2^\mathbb{T}-I_n + F(\overline{X})\big).$$
Note that Lemma \[lemlocrk\] still holds for the positive semidefinite case if $\overline{U}_2^\mathbb{T} \Delta \overline{V}_2$ is replaced by $\overline{P}_2^\mathbb{T} \Delta \overline{P}_2$. Therefore, in line with the rectangular case, from Lemma \[lemdelcondpos\], we have the following necessary condition for rank consistency.
\[thmnespos\] If $\rho_m \rightarrow 0, \sqrt{m}\rho_m\rightarrow \infty$ and $\gamma_m = O_p(1)$, then a necessary condition for the rank consistency of $\widehat{X}_m$ is that the linear system (\[eqndelcondpos\]) has a solution $\widehat{\Lambda} \in \mathbb{S}^{n-r}_+$.
Analogies to Theorem \[thmsuf\], we have the following sufficient condition for rank consistency for the positive semidefinite case.
\[thmsufpos\] If $\rho_m \rightarrow 0, \sqrt{m}\rho_m\rightarrow \infty$ and $\gamma_m = O_p(1)$, then a sufficient condition for the rank consistency of $\widehat{X}_m$ is that the linear system (\[eqndelcondpos\]) has a unique solution $\widehat{\Lambda} \in \mathbb{S}^{n-r}_{++}$.
Constraint nondegeneracy and rank consistency
---------------------------------------------
In this subsection, with the help of constraint nondegeneracy, we provide conditions to guarantee that the linear systems (\[eqndelcond\]) and (\[eqndelcondpos\]) have a unique solution. The concept of constraint nondegeneracy was pioneered by Robinson [@Rob84] and later extensively developed by Bonnans and Shapiro [@BonS00]. Consider the following constrained optimization problem $$\label{conic-prob}
\min_{X\in\mathbb{V}^{n_1\times n_2}}\Big\{\Phi(X)+\Psi(X):\ \mathcal{A}(X)-b\in K\Big\},$$ where $\Phi:\mathbb{V}^{n_1\times n_2} \to\mathbb{R}$ is a continuously differentiable function, $\Psi:\mathbb{V}^{n_1\times n_2}\to\mathbb{R}$ is a convex function, $\mathcal{A}: \mathbb{V}^{n_1\times n_2}\to\mathbb{R}^l$ is a linear operator, $b\in\mathbb{R}^l$ is a given vector and $K \subseteq \mathbb{R}^l$ is a closed convex set. Let $\widehat{X}$ be a given feasible point of (\[conic-prob\]) and $\widehat{z}:=\mathcal{A}(\widehat{X})-b$. When $\Psi$ is differentiable at $\widehat{X}$, we say that the constraint nondegeneracy holds at $\widehat{X}$ if $$\label{nondegeneracy1}
\mathcal{A}\,\mathbb{V}^{n_1\times n_2} + {\rm lin}\big(\mathcal{T}_{K}(\widehat{z})\big) = \mathbb{R}^l,$$ where $\mathcal{T}_{K}(\widehat{z})$ denotes the tangent cone of $K$ at $\widehat{z}$ and ${\rm lin}(\mathcal{T}_{K}(\widehat{z}))$ denotes the largest linearity space contained in $\mathcal{T}_{K}(\widehat{z})$, i.e., ${\rm lin}(\mathcal{T}_{K}(\widehat{z}))=\mathcal{T}_{K}(\widehat{z})\cap(-\mathcal{T}_{K}(\widehat{z}))$. When the function $\Psi$ is nondifferentiable, we can rewrite the optimization problem (\[conic-prob\]) equivalently as $$\min_{X\in\mathbb{V}^{n_1\times n_2},t\in\mathbb{R}}\Big\{\Phi(X)+t:\ \widetilde{\mathcal{A}}(X,t)\in K \times {\rm epi}\Psi \Big\},$$ where ${\rm epi}\Psi:=\left\{(X,t)\in\mathbb{V}^{n_1\times n_2}\times\mathbb{R}\ |\ \Psi(X)\le t\right\}$ denotes the epigraph of $\Psi$ and $\widetilde{\mathcal{A}}:\mathbb{V}^{n_1\times n_2}\times\mathbb{R}\rightarrow \mathbb{R}^l\times \mathbb{V}^{n_1\times n_2}\times\mathbb{R}$ is a linear operator defined by $$\widetilde{\mathcal{A}}(X,t) := \begin{pmatrix}
\mathcal{A}(X) - b\\ X\\ t
\end{pmatrix}, \quad\ (X,t)\in \mathbb{V}^{n_1\times n_2} \times \mathbb{R}.$$ From (\[nondegeneracy1\]) and [@RocW98 Theorem 6.41], the constraint nondegeneracy holds at $(\widehat{X}, \widehat{t})$ with $\widehat{t}=\Psi(\widehat{X})$ if $$\widetilde{\mathcal{A}}\begin{pmatrix}
\mathbb{V}^{n_1\times n_2} \\ \mathbb{R}
\end{pmatrix}
+\begin{pmatrix}
{\rm lin}\big(\mathcal{T}_{K}(\widehat{X})\big) \\ {\rm lin}\big(\mathcal{T}_{{\rm epi}\Psi}(\widehat{X},\widehat{t})\big)
\end{pmatrix}
= \begin{pmatrix}
\mathbb{R}^l\\ \mathbb{V}^{n_1\times n_2}\\ \mathbb{R}
\end{pmatrix}.$$ By the definition of $\widetilde{\mathcal{A}}$, it is not difficult to verify that this condition is equivalent to $$\label{nondegeneracy2}
[\mathcal{A}\ \ 0]\big({\rm lin}(\mathcal{T}_{{\rm epi}\Psi}(\widehat{X},\widehat{t}))\big)
+ {\rm lin}\big(\mathcal{T}_{K}(\widehat{X})\big) = \mathbb{R}^l.$$ By letting $\Psi=\|\cdot\|_*, \mathcal{A}=\mathcal{R}_{\alpha}$ and $K\!=\{0\}$, one can see that the problem (\[eqnrcs\]) takes the form of (\[conic-prob\]). By the expression of $\mathcal{T}_{{\rm epi}\Psi}(\overline{X},\overline{t})$ with $\overline{t}=\|\overline{X}\|_*$ (e.g., see [@JiaST12]), we see that for the problem (\[eqnrcs\]), the condition (\[nondegeneracy2\]) reduces to $$\label{eqncndc}
\mathcal{R}_\alpha\big(\mathcal{T}(\overline{X})\big) = \mathbb{R}^{d_1},$$ where $$\label{TX}
\mathcal{T}(\overline{X}) = \big\{H \in \mathbb{V}^{n_1\times n_2} \mid \overline{U}_2^\mathbb{T} H \overline{V}_2 = 0 \big\}.$$ Hence, we say that the constraint nondegeneracy holds at $\overline{X}$ to the problem (\[eqnrcs\]) if the condition (\[eqncndc\]) holds. By letting $\Psi =\delta_{\mathbb{S}_{+}^n}, \mathcal{A}=\mathcal{R}_{\alpha}$ and $K=\{0\}$, we can see that the problem (\[eqnrcspos\]) takes the form of (\[conic-prob\]) , and now that the condition (\[nondegeneracy2\]) reduces to $$\label{eqncndcpos}
\mathcal{R}_\alpha\big(\text{lin}(\mathcal{T}_{\mathbb{S}^n_+}(\overline{X}))\big) = \mathbb{R}^{d_1}.$$ Thus, we say that the constraint nondegeneracy holds at $\overline{X}$ to the problem (\[eqnrcspos\]) if the condition (\[eqncndcpos\]) holds. From Arnold’s characterization of the tangent cone $\mathcal{T}_{\mathbb{S}^n_+}(\overline{X})=\big\{H \in\mathbb{S}^n \mid \overline{P}_2^\mathbb{T} H \overline{P}_2 \in \mathbb{S}_+^{n-r} \big\}$ in [@Arn71], we can write the linearity space $\text{lin}(\mathcal{T}_{\mathbb{S}^n_+}(\overline{X}))$ explicitly as $$\text{lin}(\mathcal{T}_{\mathbb{S}^n_+}(\overline{X}))=\big\{H \in \mathbb{S}^{n} \mid \overline{P}_2^\mathbb{T} H \overline{P}_2 = 0 \big\}.$$
Interestingly, for some special matrix completion problems, the constraint nondegeneracy automatically hold at $\overline{X}$, as stated in the following proposition.
\[propcorcn\] For the following matrix completion problems:
(i)
: the covariance matrix completion with partial positive diagonal entries being fixed , in particular, the correlation matrix completion with all diagonal entries being fixed as ones;
(ii)
: the density matrix completion with its trace being fixed as one,
the constraint nondegeneracy (\[eqncndcpos\]) holds at $\overline{X}$.
Next, we take a closer look at the solutions to the linear systems (\[eqndelcond\]) and (\[eqndelcondpos\]). Define linear operators $\mathcal{B}_1:\mathbb{V}^{r\times r} \to \mathbb{V}^{(n_1-r)\times (n_2-r)}$ and $\mathcal{B}_2:\mathbb{V}^{(n_1-r)\times(n_2-r)} \to \mathbb{V}^{(n_1-r)\times(n_2-r)}$ associated with $\overline{X}$, respectively, by $$\label{linear-operator}
\mathcal{B}_1(Y): = \overline{U}_2^\mathbb{T}\mathcal{Q}_{\beta}^{\dag}(\overline{U}_1 Y \overline{V}_1^T)\overline{V}_2
\ \ {\rm and}\ \
\mathcal{B}_2(Z): = \overline{U}_2^\mathbb{T}\mathcal{Q}_{\beta}^{\dag}(\overline{U}_2 Z \overline{V}_2^T)\overline{V}_2,$$ where $Y \in \mathbb{V}^{r\times r}$ and $Z \in \mathbb{V}^{(n_1-r)\times (n_2-r)}$. Note that the operator $\mathcal{B}_2$ is self-adjoint and positive semidefinite according to the definition of $\mathcal{Q}_{\beta}^{\dag}$. Let $\widehat{g}(\overline{X})$ be the vector in $\mathbb{R}^r$ defined by $$\label{eqnvecgr}
\widehat{g}(\overline{X}):=\big(1-f_1(\sigma(\overline{X})),\ldots,1-f_r(\sigma(\overline{X}))\big)^{\mathbb{T}}.$$ Then, by the definition of the spectral operator $F$, we can rewrite (\[eqndelcond\]) in the following concise form $$\label{eqndelcond1eqiv}
\mathcal{B}_2(\Gamma) = \mathcal{B}_1\big({\rm Diag}(\widehat{g}(\overline{X}))\big), \quad \Gamma \in \mathbb{V}^{(n_1-r)\times (n_1-r)}.$$ For the positive semidefinite case $\mathbb{V}^{n_1\times n_2} = \mathbb{S}^n$ and $\overline{X} \in \mathbb{S}_+^n$, both $\overline{U}_i$ and $\overline{V}_i$ reduce to $\overline{P}_i$ for $i=1,2$. In this case, the linear system (\[eqndelcondpos\]) can be concisely written as $$\label{eqndelcond2eqiv}
\mathcal{B}_2(\Lambda) =\mathcal{B}_2(I_{n-r})+\mathcal{B}_1\big({\rm Diag}(\widehat{g}(\overline{X}))\big), \quad \Lambda \in \mathbb{S}^{n-r}.$$
\[propsoluni\] For the rectangular case, if the constraint nondegeneracy (\[eqncndc\]) holds at $\overline{X}$ to the problem (\[eqnrcs\]), then the linear operators $\mathcal{B}_2$ defined by (\[linear-operator\]) is self-adjoint and positive definite. For the positive semidefinite case, if the constraint nondegeneracy (\[eqncndcpos\]) holds at $\overline{X}$ to the problem (\[eqnrcspos\]), then the linear operators $\mathcal{B}_2$ is also self-adjoint and positive definite.
According to Proposition \[propsoluni\], the constraint nondegeneracy at $\overline{X}$ to the problem (\[eqnrcs\]) and (\[eqnrcspos\]), respectively, implies that the linear system (\[eqndelcond\]) has a unique solution $\widehat{\Gamma} = \mathcal{B}_2^{-1}\mathcal{B}_1\big({\rm Diag}({\widehat g}(\overline{X}))\big)$ and the linear system (\[eqndelcondpos\]) has a unique solution $\widehat{\Lambda} = I_{n-r} + \mathcal{B}_2^{-1}\mathcal{B}_1\big({\rm Diag}({\widehat g}(\overline{X}))\big)$. Then, from Theorems \[thmsuf\] and \[thmsufpos\], we can obtain the following main result for rank consistency.
\[thmgenconsis\] Suppose that $\rho_m \rightarrow 0,\,\sqrt{m}\rho_m\rightarrow \infty$ and $\gamma_m = O_p(1)$. For the rectangular case, if the constraint nondegeneracy (\[eqncndc\]) holds at $\overline{X}$ to the problem (\[eqnrcs\]) and $$\label{eqnsufso}
\big\|\mathcal{B}_2^{-1}\mathcal{B}_1 \big({\rm Diag}(\widehat{g}(\overline{X}))\big)\big\| <1,$$ then the estimator $\widehat{X}_m$ generated from the rank-correction step (\[eqnrcs\]) is rank consistent. For the positive semidefinite case, if the constraint nondegeneracy (\[eqncndcpos\]) holds at $\overline{X}$ to the problem (\[eqnrcspos\]) and $$\label{eqnsufsopos}
I_{n-r} + \mathcal{B}_2^{-1}\mathcal{B}_1 \big({\rm Diag}(g_r(\overline{X}))\big) \in \mathbb{S}_{++}^{n-r},$$ then the estimator $\widehat{X}_m$ generated from the rank-correction step (\[eqnrcspos\]) is rank consistent.
From Theorem \[thmgenconsis\], it is not difficult to see that there exists some threshold $\overline{\varepsilon} >0$ (depending on $\overline{X}$) such that the condition (\[eqnsufso\]) holds if $|1-f_i(\sigma(\overline{X}))| \leq \overline{\varepsilon} \ \forall\, 1\leq i \leq r$. In other words, when $F(\overline{X})$ is sufficiently close to $\overline{U}_1\overline{V}_1^\mathbb{T}$, the condition (\[eqnsufso\]) holds automatically and so does the rank consistency. Thus, Theorem \[thmgenconsis\] provides us a guideline to construct a suitable rank-correction function for rank consistency. This is another important aspect of what we can benefit from the rank-correction step, besides the reduction of recovery error discussed in Section \[section3\].
The next theorem shows that for the covariance (correlation) and density matrix completion problems with fixed basis coefficients described in Proposition \[propcorcn\], if observations are sampled uniformly at random, the rank consistency can be guaranteed for a broad class of rank-correction functions $F$.
\[thmrccordencons\] For the covariance (correlation) and density matrix completion problems defined in Proposition \[propcorcn\] under uniform sampling, if $\rho_m \rightarrow 0,\,\sqrt{m}\rho_m\rightarrow \infty,\,\gamma_m = O_p(1)$ and $F$ is a spectral operator associated with a symmetric function $f:\mathbb{R}^n \rightarrow \mathbb{R}^n$ such that for $i=1,\ldots,n$, $$\label{eqnfcorden}
f_i(x) \geq 0 \ \ \forall\, x\in \mathbb{R}_+^n \quad \text{and} \quad f_i(x)=0 \ \ \text{if and only if} \ \ x_i =0,$$ then the estimator $\widehat{X}_m$ generated from the rank-correction step is rank consistent.
Construction of the rank-correction function {#section5}
============================================
In this section, we focus on the construction of a suitable rank-correction function $F$ based on the results obtained in Sections \[section3\] and \[section4\]. As can be seen from Theorem \[thmstobd\], a smaller value of $a_m/b_m$ potentially leads to a smaller recovery error. Thus, we desire a construction of the rank-correction function such that $F(\widetilde{X}_m)$ is close to $\overline{U}_1\overline{V}_1^\mathbb{T}$. Meanwhile, according to Theorem \[thmgenconsis\], we also desire that $F(\overline{X})$ is close to $\overline{U}_1\overline{V}_1^\mathbb{T}$ for rank consistency. Notice that a reasonable initial estimator $\widetilde{X}_m$ should not deviate too much from the true matrix $\overline{X}$. Therefore, the above two criteria consistently suggest a natural idea to construct a rank-correction function $F$, if possible, such that $$\label{eqnrcfchoopt}
F(X)\rightarrow \overline{U}_1\overline{V}_1^\mathbb{T}
\quad \text{as} \quad X\rightarrow \overline{X}.$$ Next, we proceed the construction of the rank-correction function $F$ for the rectangular case. For the positive semidefinite case, one may just replace the singular value decomposition with the eigenvalue decomposition and conduct exactly the same analysis.
The rank is known {#subsecrankknown}
-----------------
If the rank of the true matrix $\overline{X}$ is known in advance, we construct the rank-correction function $F$ by $$\label{eqnchofun1}
F(X) := U_1 V_1^\mathbb{T},$$ where $(U,V)\in \mathbb{O}^{n_1,n_2}(X)$ and $X \in \mathbb{V}^{n_1\times n_2}$. Note that $F$ defined by (\[eqnchofun1\]) is not a spectral operator over the whole space of $\mathbb{V}^{n_1\times n_2}$, but in a neighborhood of $\overline{X}$ it is indeed a spectral operator and is actually twice continuously differentiable (see, e.g., [@DinST10 Proposition 8]). Hence, it satisfies the criterion (\[eqnrcfchoopt\]). With this rank-correction function, the rank-correction step is essentially the same as one step of the majorized penalty method developed in [@GaoS10]. By Theorem \[thmsuf\] and Proposition \[propsoluni\], we immediately obtain the following result.
Suppose that the rank of the true matrix $\overline{X}$ is known and the constraint nondegeneracy holds at $\overline{X}$. If $\rho_m\rightarrow 0,\,\sqrt{m}\rho_m\rightarrow \infty,\,\gamma_m = O_p(1)$ and $F$ is chosen by (\[eqnchofun1\]), then the estimator $\widehat{X}_m$ generated from the rank-correction step is rank consistent.
The rank is unknown
-------------------
If the rank of the true matrix $\overline{X}$ is unknown, then the rank-correction function $F$ cannot be defined by (\[eqnchofun1\]). What we will do is to construct a spectral operator $F$ to imitate the case when the rank is known. Here, we propose $F$ to be a spectral operator $$\label{eqnchofun}
F(X) := U \text{Diag}\big(f(\sigma(X))\big)V^\mathbb{T}$$ associated with the symmetric function $f:\mathbb{R}^n \rightarrow \mathbb{R}^n$ defined by $$\label{eqnchofun2}
f_i(x) = \begin{cases} {\displaystyle \phi\left(\frac{x_i}{\|x\|_\infty}\right)} \quad & \text{if}\ x \in \mathbb{R}^n \backslash \{0\},\\
0 \quad & \text{if}\ x =0,\end{cases}$$ where $(U,V)\in \mathbb{O}^{n_1,n_2}(X)$, $X \in \mathbb{V}^{n_1\times n_2}$, and the scalar function $\phi:\mathbb{R} \rightarrow \mathbb{R}$ takes the form $$\label{eqnchofun3}
\phi(t): = \text{sgn}(t) (1+\varepsilon^\tau)\frac{|t|^\tau}{|t|^\tau+\varepsilon^\tau}, \quad t \in \mathbb{R},$$ for some $\tau >0$ and $\varepsilon >0$. By noting that for each $t$, $\phi(t) \rightarrow \text{sgn}(t)$ as $\varepsilon \downarrow 0$, we directly obtain the following result.
\[cororankcons\] Suppose that the constraint nondegeneracy holds at $\overline{X}$. If $\rho_m \rightarrow 0,\,\sqrt{m}\rho_m\rightarrow \infty,\,\gamma_m = O_p(1)$, then for any given $\tau >0$, there exists some $\overline{\varepsilon}>0$ such that for any $F$ defined by (\[eqnchofun\]), (\[eqnchofun2\]) and (\[eqnchofun3\]) with $0< \varepsilon \leq \overline{\varepsilon}$, the estimator $\widehat{X}_m$ generated from the rank-correction step is rank consistent.
Corollary \[cororankcons\] indicates that one needs to choose a small $\varepsilon>0$ in pursuit of rank consistency. Meanwhile, we also need to take care of the influence of a small $\varepsilon>0$ on the recovery error bound which depends on the value of $a_m/b_m$. Certainly, we desire $a_m \approx 0$ and $b_m \approx 1$. This motivates us to choose a function $\phi$, if possible, such that $$\label{eqnfunhope}
\phi\bigg(\frac{\sigma_i(\widetilde{X}_m)}{\sigma_1(\widetilde{X}_m)}\bigg) \approx \begin{cases} 1 & \quad \text{if}\ 1\leq i\leq \text{rank}(\overline{X}), \\ 0 & \quad \text{if} \ \text{rank}(\overline{X})+1\leq i \leq n.\end{cases}$$ This is also why we normalize the function $\phi$ defined by (\[eqnchofun3\]) in the interval $t\in [0,1]$ such that $\phi(0)=0$ and $\phi(1)=1$. However, as indicated by Corollary \[cororkrhs\], the initial estimator $\widetilde{X}_m$ is very possible to have a higher rank than $\overline{X}$ when it approaches to $\overline{X}$. It turns out that when $\varepsilon>0$ is tiny, $\phi\big(\sigma_i(\widetilde{X}_m)/\sigma_1(\widetilde{X}_m)\big) \approx 1$ for $\text{rank}(\overline{X})+1 \leq i \leq \text{rank}(\widetilde{X}_m)$, which violates our desired property (\[eqnfunhope\]). As a result, $\varepsilon>0$ should be chosen to be small but balanced. Notice that $\phi(\varepsilon)= (1+\varepsilon^\tau)/2 \approx 1/2$ if $\varepsilon>0$ is small and $\tau >0$ is not too small. Thus, the value of $\varepsilon$ can be regarded as a divide of confidence on whether $\sigma_i(\widetilde{X}_m)$ is believed to come from a nonzero singular values of $\overline{X}$ with perturbation — positive confidence if $\sigma_i(\widetilde{X}_m) > \varepsilon \sigma_1(\widetilde{X}_m)$ and negative confidence if $\sigma_i(\widetilde{X}_m) < \varepsilon \sigma_1(\widetilde{X}_m)$. On the other hand, the parameter $\tau>0$ mainly controls the shape of the function $\phi$ over $t\in [0,1]$. The function $\phi$ is concave if $0<\tau\leq 1$ and $S$-shaped with a single inflection point at $\big(\frac{\tau-1}{\tau+1}\big)^{1/\tau}\varepsilon$ if $\tau>1$. Moreover, the steepness of the function $\phi$ increases when $\tau$ increases. In particular, if $0<\varepsilon<1$ and $\tau$ is very large, $\phi$ is very close to the step function taking the value $0$ if $0\leq t<\varepsilon$ and the value $1$ if $ \varepsilon < t \leq 1$. In this case, there exists some $\varepsilon$ such that the desired property (\[eqnfunhope\]) can be achieved and that the corresponding rank-correction function $F$ is very close to the one defined by (\[eqnchofun1\]). Thus, it seems to be a good idea to choose an $S$-shaped function $\phi$ with a large $\tau$. However, in practice, the parameter $\varepsilon$ should be pre-determined. Since $\text{rank}(\overline{X})$ is unknown and the singular values of $\widetilde{X}_m$ are unpredictable, it is hard to choose a suitable $\varepsilon$ in advance, and hence, it will be too risky to choose a large $\tau$ for recovery. As a result, one has to be somewhat conservative to choose $\tau$, sacrificing some optimality of recovery in exchange for robustness strategically. If the initial estimator is generated from the nuclear norm penalized least squares problem, we recommend the choices $\tau = 1$ or $2$ and $\varepsilon = 0.01 \sim 0.1$ as these choices show stable performance for plenty of problems, as validated in Section 6.
We also remark that for the positive semidefinite case, the rank-correction function defined by (\[eqnchofun\]), (\[eqnchofun2\]) and (\[eqnchofun3\]) is related to the reweighted trace norm for the matrix rank minimization proposed by Fazel et al. [@FazHB03; @MohF10]. The reweighted trace norm in [@FazHB03; @MohF10] for the positive semidefinite case is $\langle (X^k+\varepsilon I_n)^{-1}, X\rangle$, which arises from the derivative of the surrogate function $\log\det(X+\varepsilon I_n)$ of the rank function at an iterate $X^k$, where $\varepsilon$ is a small positive constant. Meanwhile, in our proposed rank-correction step, if we choose $\tau = 1$, then $I_n-\frac{1}{1+\varepsilon}F(\widetilde{X}_m) = \varepsilon'(\widetilde{X}_m+\varepsilon' I_n)^{-1}$ with $\varepsilon' =\varepsilon \|\widetilde{X}_m\|$. Superficially, similarity occurs; however, it is notable that $\varepsilon'$ depends on $\widetilde{X}_m$, which is different from the constant $\varepsilon$ in [@FazHB03; @MohF10]. More broadly speaking, the rank-correction function $F$ defined by (\[eqnchofun\]), (\[eqnchofun2\]) and (\[eqnchofun3\]) is not a gradient of any real-valued function. This distinguishes our proposed rank-correction step from the reweighted trace norm minimization in [@FazHB03; @MohF10] even for the positive semidefinite case.
Numerical experiments {#section6}
=====================
In this section, we validate the power of our proposed rank-corrected procedure on the recovery by applying it to the positive semidefinite matrix completion problems. In solving the optimization problem in the rank-correction step (\[eqnrcspos\]), we adopted the code developed by Jiang et al. [@JiaST12] for large scale linearly constrained convex semidefinite programming problems. The implemented code is based on an inexact version of the accelerated proximal gradient method [@Nes83; @BecT09]. All tests were run in MATLAB under Windows 7.0 operating system on an Intel Core(TM) i7-2720 QM 2.20GHz CPU with 8.00GB memory.
For convenience, in the sequel, the [NNPLS]{} estimator and the [RCS]{} estimator, respectively, stand for the estimators from the nuclear norm penalized least squares problem (i.e., the problem (\[eqnrcspos\]) with $F\equiv 0 $ and $\gamma_m=0$) and the rank-correction step. Let $X_m$ be an estimator. The [**relative error**]{} ([**relerr**]{} for short) of $X_m$ is defined by $${\rm relerr}=\frac{\|X_m - \overline{X}\|_{F}}{\max(10^{-8},\|\overline{X}\|_{F})}.$$
Influence of fixed basis coefficients on the recovery {#subsec6.1}
-----------------------------------------------------
In this subsection, we take the correlation matrix completion for example to test the performance of the [NNPLS]{} estimator and the [RCS]{} estimator with different patterns of fixed basis coefficients. We randomly generated the true matrix $\overline{X}$ by the following command:
M = randn(n,r); ML = weight*M(:,1:k); M(:,1:k) = ML; Xtemp = M*M';
D = diag(1./sqrt(diag(Xtemp))); X_bar = D*Gtemp*D
where the parameter [weight]{} is used to control the relative magnitude difference between the first [k]{} largest eigenvalues and the other nonzero eigenvalues. In our experiment, we set [weight]{} $=5$ and [k]{} $=1$, and took $\overline{X}=$ [X\_bar]{} with dimension [n]{} $=1000$ and rank [r]{} $=5$. We randomly fixed partial diagonal and off-diagonal entries of $\overline{X}$ and sampled the rest entries uniformly at random with i.i.d. Gaussian noise at the noise level $10\%$.
In Figure \[figure2\], we plot the curves of the relative error and the rank of the NNPLS estimator and the RCS estimator with different patterns of fixed entries. In the captions of the subfigures, [**diag**]{} means the number of fixed diagonal entries and [**non-diag**]{} means the number of fixed off-diagonal entries. The subfigures on the left-hand side and the right-hand side show the performance of the NNPLS estimator and the RCS estimator, respectively. For the RCS estimator, the rank-correction function $F$ is defined by (\[eqnchofun\]), (\[eqnchofun2\]) and (\[eqnchofun3\]) with $\tau = 2$ and $\varepsilon = 0.02$, and the initial $\widetilde{X}_m$ is chosen from those points of the corresponding subfigures on the left-hand side such that $\big|\|y-\mathcal{R}_\Omega(\widetilde{X}_m)\|_2/\|y\|_2-0.1\big|$ attains the smallest value.
\
\
\
\
From the subfigures on the left-hand side, we observe that as the number of fixed diagonal entries increases, the parameter $\rho_m$ for the smallest recovery error deviates more and more from the one for attaining the true rank. In particular, when [**diag**]{} $=n$, the NNPLS estimator reduces to the (constrained) least squares estimator so that one cannot benefit from the NNPLS estimator for encouraging a low-rank solution. This implies that the NNPLS estimator does not possess the rank consistency when some entries are fixed. However, the subfigures on the right-hand side indicate that the RCS estimator can yield a solution with the correct rank as well as a desired small recovery error simultaneously, with the parameter $\rho_m$ in a large interval. This exactly validates the theoretical result of Theorem \[thmrccordencons\] for rank consistency.
Performance of different rank-correction functions for recovery
---------------------------------------------------------------
In this subsection, we test the performance of different rank-correction functions for recovering a correlation matrix. We randomly generated the true matrix $\overline{X}$ by the command in Subsection 6.1 with [n]{} $=1000$, [r]{} $=10$, [weight]{} $=2$ and [k]{} $ = 5$. We fixed all the diagonal entries of $\overline{X}$ and sampled partial off-diagonal entries uniformly at random with i.i.d. Gaussian noise at the noise level $10\%$. We chose the (nuclear norm penalized) least squares estimator to be the initial estimator $\widetilde{X}_m$. In Figure \[figure3\], we plot four curves corresponding to the rank-correction functions $F$ defined by (\[eqnchofun\]), (\[eqnchofun2\]) and (\[eqnchofun3\]) with $\tau = 2$ and different $\varepsilon$, and another two curves corresponding to the rank-correction functions $F$ defined by (\[eqnchofun1\]) at $\widetilde{X}_m$ (i.e., $\widetilde{U}_1\widetilde{V}_1^\mathbb{T}$) and $\overline{X}$ (i.e., $\overline{U}_1\overline{V}_1^\mathbb{T}$), respectively. The values of $a_m$, $b_m$ and the optimal recovery error with different $\rho_m$ are listed in Table \[tab1\].
As can be seen from Figure \[figure3\], when $\rho_m$ increases, the recovery error decreases with the rank and then increases after the correct rank is attained, except for the case $\overline{U}_1\overline{V}_1^\mathbb{T}$. This validates our discussion about the recovery error at the end of Section \[section3\]. Moreover, for a smaller $\varepsilon$, the curve of recovery error changes more gently, though a certain optimality in the sense of recovery error is sacrificed. This means that the choice of a relatively small $\varepsilon$, say $0.01$ or $0.02$, is more robust for those ill-conditioned problems. From Table \[tab1\], we see that a smaller $a_m/b_m$ corresponds to a better optimal recovery error. It is worthwhile to point out that, even if $a_m/b_m$ is larger than $1$, the performance of the RCS estimator for recovery is still much better than that of the NNPLS estimator.
rank-correction function $a_m$ $b_m$ $a_m/b_m$ optimal relerr
------------------------------------------------------- ---------- ---------- ----------- ----------------
zero function $1$ $1$ $1$ $10.85\%$
$\varepsilon= 0.01, \tau = 2$ $0.1420$ $0.2351$ $0.6038$ $5.96\%$
$\varepsilon = 0.02, \tau = 2$ $0.1459$ $0.5514$ $0.2646$ $5.80\%$
$\varepsilon = 0.05, \tau = 2$ $0.1648$ $0.8846$ $0.1863$ $5.75\%$
$\varepsilon = 0.1, \tau = 2$ $0.2399$ $0.9681$ $0.2478$ $5.77\%$
$\widetilde{U}_1\widetilde{V}_1^\mathbb{T}$ (initial) $0.1445$ $0.9815$ $0.1472$ $5.75\%$
$\overline{U}_1\overline{V}_1^\mathbb{T}$ (true) $0$ $1$ $0$ $2.25\%$
: Influence of the rank-correction term on the recovery error \[tab1\]
\
Performance for different matrix completion problems
----------------------------------------------------
In this subsection, we test the performance of the RCS estimator for the covariance and density matrix completion problems. As can be seen from Figure \[figure2\], a good choice of the parameter $\rho_m$ for the RCS estimator could be the smallest one such that the rank becomes stable. Such a parameter $\rho_m$ can be found by the bisection search method. This is actually what we benefit from rank consistency. In the following numerical experiments, we apply the above strategy to find a suitable $\rho_m$ for the RCS estimator, and choose the rank-correction function $F$ defined by (\[eqnchofun\]), (\[eqnchofun2\]) and (\[eqnchofun3\]) with $\tau=2$ and $\varepsilon=0.02$.
We first take the covariance matrix completion for example to test the performance of the RCS estimator with different initial estimators $\widetilde{X}_m$. The true matrix $\overline{X}$ is generated by the command in Subsection 6.1 with [n]{} $=500$, [r]{} $=5$, [weight]{} $=3$ and [k]{} $=1$ except that [D = eye(n)]{}. We depict the numerical results in Figure 4, where the dash curves represent the relative recovery error and the rank of the NNPLS estimator with different $\rho_m$, and the solid curves represent the relative recovery error and the rank of the RCS estimator with $\widetilde{X}_m$ chosen to be the corresponding NNPLS estimator. As can be seen from Figure 4, the RCS estimator substantially improves the quality of the NNPLS estimator in terms of both the recovery error and the rank. We also observe that when the initial $\widetilde{X}_m$ has a large deviation from the true matrix, the quality of the RCS estimator may still not be satisfied. Thus, it is natural to ask whether further rank-correction steps could improve the quality. The answer can be found from Table \[tab2\] below, where the numerical results of the covariance matrix completion are reported. We also report the numerical results of the density matrix completion in Table \[tab3\].
{width="\textwidth"}
---------------- -------------- ---------------------------------------------------------------------------------------------------- ------------------- ------------------ ------------------- -------------------
[diag/]{} [NNPLS]{} [1st RCS]{} [2st RCS]{} [3rd RCS]{}
\[0pt\][$r$]{} [off-diag]{} \[0pt\][$\renewcommand{{1.2}}{0.85} \begin{array}{c} {\rm sample} \\ {\rm ratio} \end{array}$]{} [relerr (rank)]{} [relerr(rank)]{} [relerr (rank)]{} [relerr (rank)]{}
1000/0 2.40% 1.95e-1 (47) 1.27e-1 (5) 1.18e-1 (5) 1.12e-1 (5)
1000/0 7.99% 6.10e-2 (51) 3.41e-2 (5) 3.37e-2 (5) 3.36e-2 (5)
\[0pt\][5]{} 500/50 2.39% 2.01e-1 (45) 1.10e-1 (5) 9.47e-2 (5) 8.97e-2 (5)
500/50 7.98% 7.19e-2 (32) 3.77e-2 (5) 3.59e-2 (5) 3.58e-2 (5)
1000/0 5.38% 1.32e-1 (74) 7.68e-2 (10) 7.39e-2 (10) 7.36e-2 (10)
1000/0 8.96% 9.18e-2 (78) 5.15e-2 (10) 5.08e-2 (10) 5.08e-2 (10)
\[0pt\][10]{} 500/100 5.37% 1.58e-1 (57) 8.66e-2 (10) 7.74e-2 (10) 7.60e-2 (10)
500/100 8.96% 1.02e-1 (49) 5.36e-2 (10) 5.24e-2 (10) 5.25e-2 (10)
---------------- -------------- ---------------------------------------------------------------------------------------------------- ------------------- ------------------ ------------------- -------------------
: \[tab2\] Performance for covariance matrix completion problems with $n=1000$
-- -------------- --------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------- ------------------------------------- ------------------------------------- -------------------------------------
[NNPLS1]{} [NNPLS2]{} [RCS]{}
\[0pt\][$\renewcommand{{1.2}}{0.85} \begin{array}{c} {\rm noise} \\ {\rm level} \end{array}$]{} \[0pt\][$\renewcommand{{1.2}}{0.85} \begin{array}{c} {\rm sample} \\ {\rm ratio} \end{array}$]{} [fidelity]{} [relerr]{} [rank]{} [fidelity]{} [relerr]{} [rank]{} [fidelity]{} [relerr]{} [rank]{}
10.0% 1.5% 0.697 2.59e-1 3 0.955 2.50e-1 3 0.987 1.02e-1 3
\[0pt\][3]{} 10.0% 4.0% 0.915 8.04e-2 3 0.997 6.84e-2 3 0.998 4.13e-2 3
10.0% 2.0% 0.550 3.71e-1 5 0.908 4.23e-1 5 0.972 1.61e-1 5
\[0pt\][5]{} 10.0% 5.0% 0.889 1.03e-1 5 0.995 9.18e-2 5 0.997 4.91e-2 5
12.4% 1.5% 0.654 2.93e-1 3 0.957 2.43e-1 3 0.988 1.06e-1 3
\[0pt\][3]{} 12.4% 4.0% 0.832 1.49e-1 3 0.995 8.14e-2 3 0.997 6.41e-2 3
12.4% 2.0% 0.521 3.95e-1 5 0.912 4.09e-1 5 0.977 1.51e-1 5
\[0pt\][5]{} 12.5% 5.0% 0.817 1.61e-1 5 0.987 1.01e-1 5 0.996 7.09e-2 5
-- -------------- --------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------- ------------------------------------- ------------------------------------- -------------------------------------
: \[tab3\] Performance for density matrix completion problems with $n=1024$
For the covariance matrix completion problems, we generated the true matrix $\overline{X}$ by the command in Subsection \[subsec6.1\] with [n]{} $=1000$, [weight]{} $=3$ and [k]{} $=1$ except that [D = eye(n)]{}. The rank of $\overline{X}$ and the number of fixed diagonal and non-diagonal entries of $\overline{X}$ are reported in the first and the second columns of Table \[tab2\], respectively. We sampled partial off-diagonal entries uniformly at random with i.i.d. Gaussian noise at the noise level $10\%$. The first RCS estimator is using the NNPLS estimator as the initial estimator $\widetilde{X}_m$, and the second (third) RCS estimator is using the first (second) RCS estimator as the initial estimator $\widetilde{X}_m$. From Table \[tab2\], we see that when the sample ratio is reasonable, one rank-correction step is enough to yield a desired result. Meanwhile, when the sample ratio is very low, especially if some off-diagonal entries are further fixed, one or two more rank-correction steps can still improve the quality of estimation.
For the density matrix completion problems, we generated the true density matrix $\overline{X}$ by the following command:
M = randn(n,r)+i*randn(n,r); ML = weight*M(:,1:k); M(:,1:k) = ML;
Xtemp = M*M'; X_bar = Xtemp/sum(diag((Xtemp))).
During the testing, we set [n]{} $=1024$, [weight]{} $=2$ and [k]{} $=1$, and sampled partial Pauli measurements except the trace of $\overline{X}$ uniformly at random with i.i.d. Gaussian noise at the noise level $10\%$. Besides the above statistical noise, we further added the depolarizing noise, which frequently appears in quantum systems, with strength $0.01$. This case is labeled as the mixed noise in the last four rows of Table \[tab3\]. We remark here that the depolarizing noise differs from our assumption on noise since it does not have randomness. One may refer to [@GroLFBE10; @FlaGLE12] for details of the quantum depolarizing channel. In Table \[tab3\], the (squared) [**fidelity**]{} is a measure of the closeness of two quantum states, defined by $\big\|\widehat{X}_m^{1/2} \overline{X}^{1/2}\big\|_{*}^2$, the NNPLS1 estimator means the NNPLS estimator by dropping the trace one constraint, and the NNPLS2 estimator means the one obtained by normalizing the NNPLS1 estimator to be of trace one. Note that the NNPLS2 estimator was ever used by Flammia et al. [@FlaGLE12]. Table \[tab3\] shows that the RCS estimator is superior to the NNPLS2 estimator in terms of both the fidelity and the relative error.
Conclusions {#section7}
===========
In this paper, we proposed a rank-corrected procedure for low-rank matrix completion problems with fixed basis coefficients. This approach can substantially overcome the limitation of the nuclear norm penalization for recovering a low-rank matrix. We studied the impact of adding the rank-correction term on both the reduction of the recovery error bounds and the rank consistency (in the sense of Bach [@Bac08]). Due to the presence of fixed basis coefficients, constraint nondegeneracy plays an important role in our analysis. Extensive numerical experiments show that our approach can significantly improve the recovery performance in the sense of both the recovery error and the rank, compared with the nuclear norm penalized least square estimator. As a byproduct, our results also provide a theoretical foundation for the majorized penalty method of Gao and Sun [@GaoS10] and Gao [@Gao10] for structured low-rank matrix optimization problems.
Our proposed rank-correction step also allows additional constraints according to other possible prior information. In particular, for additional linear constraints, all the theoretical results in this paper hold with slight modifications. In order to better fit the under-sampling setting of matrix completion, in the future work, it would be of great interest to extend the asymptotic rank consistency results to the case that the matrix size is allowed to grow. It would also be interesting to extend this approach to deal with other low-rank matrix problems.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors would like to thank Dr. Kaifeng Jiang for helpful discussions on using the accelerated proximal gradient method to solve the density matrix completion problem.
Appendix {#appendix .unnumbered}
========
[**Proof of Theorem \[thmopbd\]**]{}
Let $\Delta_m:=\widehat{X}_m - \overline{X}$. Since $\widehat{X}_m$ is optimal to (\[eqnrcs\]) and $\overline{X}$ is feasible to (\[eqnrcs\]), it follows that $$\begin{aligned}
\frac{1}{2m} \|\mathcal{R}_\Omega(\Delta_m)\|_2^2 \leq & \ \Big\langle \frac{\nu}{m}\mathcal{R}_\Omega^*(\xi), \Delta_m\Big\rangle - \rho_m\big(\|\widehat{X}_m\|_*-\|\overline{X}\|_* - \langle F(\widetilde{X}_m)+\gamma_m\widetilde{X}_m, \Delta_m\rangle\big) \nonumber \\
& \ +\frac{\rho_m\gamma_m}{2}
\big( \|\overline{X}\|_F^2-\|\widehat{X}_m\|_F^2\big). \label{eqnopbd1}
\end{aligned}$$ Then, it follows from (\[eqndefrho\]) that $$\begin{aligned}
\Big\langle \frac{\nu}{m}\mathcal{R}_\Omega^*(\xi), \Delta_m \Big\rangle & \ \leq \nu \Big\|\frac{1}{m}\mathcal{R}_\Omega^*(\xi)\Big\|\big(\|\mathcal{P}_T(\Delta_m)\|_* +\|\mathcal{P}_{T^\perp}(\Delta_m)\|_*\big) \nonumber \\
& \ \leq \frac{\rho_mb_m}{\kappa} \big(\|\mathcal{P}_T(\Delta_m)\|_* +\|\mathcal{P}_{T^\perp}(\Delta_m)\|_*\big). \label{eqnopbd3}
\end{aligned}$$ From the directional derivative of the nuclear norm at $\overline{X}$ (see [@Wat92 Theorem 1]), we have $$\begin{aligned}
\|\widehat{X}_m \|_* - \|\overline{X}\|_* \ge
\langle \overline{U}_1\overline{V}_1^\mathbb{T}, \Delta_m\rangle +\|\overline{U}_2^\mathbb{T}\Delta_m\overline{V}_2\|_*.
\end{aligned}$$ This, together with equations (\[Operator-PT\]) and (\[eqndelalpbet\]), implies that $$\begin{aligned}
&\ \|\widehat{X}_m \|_* - \|\overline{X}\|_* -\langle F(\widetilde{X}_m)+\gamma_m \widetilde{X}_m, \Delta_m\rangle \nonumber\\
\geq & \, \langle \overline{U}_1\overline{V}_1^\mathbb{T}, \Delta_m\rangle +\|\overline{U}_2^\mathbb{T}\Delta_m\overline{V}_2\|_* -\langle F(\widetilde{X}_m)+\gamma_m \widetilde{X}_m, \Delta_m\rangle \nonumber\\
= & \, \langle \overline{U}_1\overline{V}_1^\mathbb{T}\! -\! \mathcal{P}_T(F(\widetilde{X}_m)\! + \!\gamma_m \widetilde{X}_m), \Delta_m\rangle \!+\! \|\mathcal{P}_{T^\perp}(\Delta_m)\|_*\! -\!\langle \mathcal{P}_{T^\perp}(F(\widetilde{X}_m)\!+\!\gamma_m \widetilde{X}_m), \Delta_m\rangle \nonumber \\
= & \, \langle \overline{U}_1\overline{V}_1^\mathbb{T}\! -\! \mathcal{P}_T(F(\widetilde{X}_m)\!+\! \gamma_m \widetilde{X}_m),\! \mathcal{P}_T(\Delta_m)\rangle \!+ \!\|\mathcal{P}_{T^\perp}(\Delta_m)\|_*\! -\!\langle \mathcal{P}_{T^\perp}(F(\widetilde{X}_m)\!+\!\gamma_m \widetilde{X}_m),\!\mathcal{P}_{T^\perp}(\Delta_m)\rangle \nonumber \\
\geq & \, -\!\|\overline{U}_1\overline{V}_1^\mathbb{T}\!-\!\mathcal{P}_T(F(\widetilde{X}_m)\! +\!\gamma_m\widetilde{X}_m)\|\|\mathcal{P}_T(\Delta_m)\|_*\!+ \!\big(1\!-\!\|\mathcal{P}_{T^\perp}(F(\widetilde{X}_m\!+\!\gamma_m \widetilde{X}_m)\|\big) \|\mathcal{P}_{T^\perp}(\Delta_m)\|_* \nonumber \\
= & \, -a_m \|\mathcal{P}_T(\Delta_m)\|_* +b_m \|\mathcal{P}_{T^\perp}(\Delta_m)\|_*. \label{eqnopbd2}
\end{aligned}$$ By substituting (\[eqnopbd2\]) and (\[eqnopbd3\]) into (\[eqnopbd1\]), we obtain that $$\label{eqnopbdtig}
\begin{aligned}
\frac{1}{2m} \|\mathcal{R}_\Omega(\Delta_m)\|_2^2 \leq & \ \rho_m\left(\Big(a_m+\frac{b_m}{\kappa}\Big)\|\mathcal{P}_T(\Delta_m)\|_*- \frac{\kappa-1}{\kappa}b_m\|\mathcal{P}_{T^\perp}(\Delta_m)\|_*\right) \\
& +\frac{\rho_m\gamma_m}{2} (\|\overline{X}\|_F^2 -\|\widehat{X}_m\|_F^2).
\end{aligned}$$ Note that $\text{rank}(\mathcal{P}_T(\Delta_m))\leq 2r$. Hence, $\|\mathcal{P}_T(\Delta_m)\|_* \leq \sqrt{2r} \|\mathcal{P}_T(\Delta_m)\|_F \leq\sqrt{2r}\|\Delta_m\|_F,$ and the desired result follows from (\[eqnopbdtig\]). Thus, we complete the proof.
[**Proof of Lemma \[lemiso\].**]{}
The proof is similar to that of [@Klo12 Lemma 12]. We need to show that the event $$\mathcal{E}=\!\left\{\exists \, \Delta \in \mathcal{C}(r)\ \text{such that} \ \left|\frac{1}{m}\|\mathcal{R}_\Omega(\Delta)\|_2^2-\!\langle \mathcal{Q}_\beta(\Delta),\Delta\rangle \right|\!\geq \frac{1}{2} \langle \mathcal{Q}_\beta(\Delta),\Delta\rangle + 128\mu_1 d_2 r \vartheta_m^2 \right\}$$ occurs with probability less than $2/(n_1+ n_2)$. For any given $\varepsilon\!>\!0$, we decompose $\mathcal{C}(r)$ as $$\mathcal{C}(r) = \bigcup_{k=1}^\infty \left\{\Delta \in \mathcal{C}(r) \mid 2^{k-1} \varepsilon \leq \langle \mathcal{Q}_\beta(\Delta),\Delta\rangle \leq 2^k \varepsilon\right\}.$$ For any $a>0$, let $\mathcal{C}(r,a):=\{\Delta \in \mathcal{C}(r) \mid \langle \mathcal{Q}_\beta(\Delta),\Delta\rangle \leq a\}.$ Then we get $\mathcal{E} \subseteq \cup_{k=1}^\infty \mathcal{E}_k$ with $$\mathcal{E}_k= \Big\{\exists \, \Delta \in \mathcal{C}(r,2^k \varepsilon)\ \text{such that} \ \Big|\frac{1}{m}\|\mathcal{R}_\Omega(\Delta)\|_2^2-\langle \mathcal{Q}_\beta(\Delta),\Delta\rangle \Big| \geq 2^{k-2}\varepsilon + 128\mu_1 d_2 r \vartheta_m^2 \Big\}.$$ Then, we need to estimate the probability of each event $\mathcal{E}_k$. Define $$Z_a:= \sup_{\Delta \in \mathcal{C}(r,a)} \Big| \frac{1}{m} \|\mathcal{R}_\Omega(\Delta)\|_2^2-\langle \mathcal{Q}_\beta(\Delta),\Delta \rangle \Big|.$$ Notice that for any $\Delta \in \mathbb{V}^{n_1\times n_2}$, $$\frac{1}{m} \|\mathcal{R}_\Omega(\Delta)\|_2^2 = \frac{1}{m}\sum_{i=1}^m \langle \Theta_{\omega_i}, \Delta\rangle^2 \stackrel{a.s.}{\rightarrow} \mathbb{E}(\langle \Theta_{\omega_i}, \Delta\rangle^2) = \langle \mathcal{Q}_\beta(\Delta),\Delta\rangle.$$ Since $\|\mathcal{R}_\beta(\Delta)\|_\infty \leq 1$ for all $\Delta \in \mathcal{C}(r)$, from Massart’s Hoeffding type concentration inequality [@Mas00 Theorem 9] for suprema of empirical processes, we have $$\label{eqnmasineq}
{\rm Pr}\left( Z_a \geq \mathbb{E}(Z_a) + t\right) \leq \exp(-mt^2/2) \quad \forall \, t>0.$$ Next, we use the standard Rademacher symmetrization in the theory of empirical processes to further derive an upper bound of $\mathbb{E}(Z_a)$. Let $\{\epsilon_1,\ldots, \epsilon_m\}$ be a Rademacher sequence. Then, we have $$\begin{aligned}
\mathbb{E}(Z_a) = & \ \mathbb{E} \bigg(\sup_{\Delta \in \mathcal{C}(r,a)} \Big| \frac{1}{m} \sum_{i=1}^m \langle \Theta_{\omega_i}, \Delta \rangle^2 -\mathbb{E}\big(\langle \Theta_{\omega_i}, \Delta \rangle^2\big) \Big|\bigg) \nonumber\\
\leq & \ 2 \mathbb{E} \bigg(\sup_{\Delta \in \mathcal{C}(r,a)} \Big|\frac{1}{m}\sum_{i=1}^m \epsilon_i \langle \Theta_{\omega_i}, \Delta\rangle^2 \Big|\bigg) \leq 8 \mathbb{E} \bigg(\sup_{\Delta \in \mathcal{C}(r,a)} \Big|\frac{1}{m}\sum_{i=1}^m \epsilon_i \langle \Theta_{\omega_i}, \Delta\rangle \Big|\bigg) \nonumber \\
= & \ 8 \mathbb{E} \bigg(\sup_{\Delta \in \mathcal{C}(r,a)} \Big|\frac{1}{m}\sum_{i=1}^m \langle \mathcal{R}_\Omega^*(\epsilon), \Delta\rangle \Big|\bigg) \leq 8 \mathbb{E} \Big\|\frac{1}{m}\mathcal{R}_\Omega^*(\epsilon)\Big\|\bigg( \sup_{\Delta \in \mathcal{C}(r,a)} \|\Delta\|_*\bigg), \label{eqnmasineq1}\end{aligned}$$ where the first inequality follows from the symmetrization theorem (e.g., see [@VanW96 Lemma 2.3.1] and [@BuhV11 Theorem 14.3]) and the second inequality follows from the contraction theorem (e.g., see [@LedT91 Theorem 4.12] and [@BuhV11 Theorem 14.4]). Moreover, from (\[eqndefiot\]), we have $$\label{eqnmasineq2}
\|\Delta\|_*\leq \sqrt{r} \|\Delta\|_F \leq \sqrt{\mu_1 r d_2 \langle \mathcal{Q}_\beta(\Delta), \Delta\rangle} \leq \sqrt{\mu_1 r d_2 a} \quad \forall \, \Delta \in \mathcal{C}(r,a).$$ Combining (\[eqnmasineq1\]) and (\[eqnmasineq2\]) with the definition of $\vartheta_m$ in (\[defvaritheta\]), we obtain that $$\begin{aligned}
\mathbb{E}(Z_a) +\frac{a}{8} \leq 8\vartheta_m\sqrt{\mu_1 r d_2 a} +\frac{a}{8} \leq 128 \mu_1 r d_2 \vartheta_m^2+\frac{a}{4}.
\end{aligned}$$ Then, by choosing $t = a/8$ in (\[eqnmasineq\]), it follows that $${\rm Pr}\left( Z_a \geq \frac{a}{4} + 128 \mu_1 r d_2 \vartheta_m^2 \right)\leq {\rm Pr}\left(Z_a \geq \mathbb{E}(Z_a)+\frac{a}{8}\right)\leq \exp\left(-\frac{ma^2}{128}\right).$$ This implies that ${\rm Pr}(\mathcal{E}_k)\!\leq \exp(-4^{k}\varepsilon^2m/128)$. Then, by choosing $\varepsilon =\!\sqrt{\frac{64\log(n_1+n_2)}{\log(2) m}}$ and using $e^x \geq 1+x >x$, we have $$\begin{aligned}
{\rm Pr}(\mathcal{E}) & \ \leq \sum_{k=1}^\infty {\rm Pr}(\mathcal{E}_k) \leq \sum_{k=1}^\infty \exp\left(-\frac{4^{k} \varepsilon^2 m}{128}\right) < \sum_{k=1}^\infty \exp\left(-\frac{\log(4)k \varepsilon^2 m}{128} \right) \\
& \ \leq \frac{\exp(-\log(2)m\varepsilon^2/64)}{1-\exp(-\log(2)\varepsilon^2 m/64)} = \frac{1}{n_1+n_2-1}.\end{aligned}$$ Thus, we complete the proof.
[**Proof of Theorem \[thmbdmid\]**]{}
The proof is similar to that of [@Klo12 Theorem 3]. Let $\Delta_m^c :=\widehat{X}_m^c-\overline{X}$. By noting that $\gamma_m = 0$ in this case, from (\[eqnopbdtig\]), we have $$\Big(a_m+\frac{b_m}{\kappa}\Big)\|\mathcal{P}_T(\Delta_m^c)\|_*- \frac{\kappa-1}{\kappa}b_m\|\mathcal{P}_{T^\perp}(\Delta_m^c)\|_*\geq 0.$$ Then, by setting $t_m\!:= \frac{\kappa}{\kappa-1}(1+\frac{a_m}{b_m})$, together with the above inequality, we obtain that $$\label{eqnbdmid1}
\|\Delta_m^{c}\|_* \le \|\mathcal{P}_T(\Delta_m^c)\|_* + \|\mathcal{P}_{T^\perp}(\Delta_m^c)\|_* \leq t_m\|\mathcal{P}_T(\Delta_m^{c})\|_* \leq \sqrt{2r}t_m \|\Delta_m^{c}\|_F.$$ Let $c_m := \|\mathcal{R}_\beta(\Delta_m^{c})\|_\infty$. Clearly, $c_m \leq 2c$. We proceed the discussions by two cases:
[**Case 1.**]{} Suppose that $\langle \mathcal{Q}_\beta(\Delta_m^{c}), \Delta_m^{c} \rangle \leq c_m^2\sqrt{\frac{64\log(n_1+n_2)}{\log(2)m}}$. From (\[eqndefiot\]), we obtain that $$\frac{\|\Delta_m^{c}\|_F^2}{d_2}\leq 4 b^2 \mu_1 \sqrt{\frac{64\log(n_1+n_2)}{\log(2)m}}.$$
[**Case 2.**]{} Suppose that $\langle \mathcal{Q}_\beta(\Delta_m^{c}), \Delta_m^{c} \rangle> c_m^2\sqrt{\frac{64\log(n_1+n_2)}{\log(2)m}}$. Then, from (\[eqnbdmid1\]), we have $\Delta_m^{c}/c_m \in \mathcal{C}(2t_m^2r)$. Together with Lemma \[lemiso\], it follows that $$\frac{1}{2}\langle \mathcal{Q}_\beta(\Delta_m^{c}), \Delta_m^{c}\rangle \leq \frac{1}{m}\|\mathcal{R}_\Omega(\Delta_m^{c})\|_2^2 + 128 c_m^2 t_m^2 \mu_1 d_2 r \vartheta_m^2.$$ Combining the last inequality with Theorem \[thmopbd\] and equation (\[eqndefiot\]), we obtain that $$\begin{aligned}
\frac{\|\Delta_m^{c}\|_F^2}{2d_2} \leq & \ \frac{\mu_1}{2}\langle \mathcal{Q}_\beta(\Delta_m^{c}), \Delta_m^{c}\rangle \leq \frac{\mu_1}{m}\|\mathcal{R}_\Omega(\Delta_m^{c})\|_2^2+ 128c_m^2 t_m^2 \mu_1^2 d_2 r\vartheta_m^2 \\
\leq & \ 2\sqrt{2r}\Big( a_m + \frac{b_m}{\kappa}\Big)\mu_1\rho_m\|\Delta_m^{c}\|_F +128 c_m^2 t_m^2 \mu_1^2 d_2 r\vartheta_m^2 \\
\leq & \ \frac{\|\Delta_m^{c}\|_F^2}{4d_2}+8\Big( a_m + \frac{b_m}{\kappa}\Big)^2\mu_1^2\rho_m^2rd_2 + 128 c_m^2 t_m^2 \mu_1^2 d_2 r \vartheta_m^2.
\end{aligned}$$ By plugging in $t_m$, we have that there exists some constant $C_1$ such that $$\frac{\|\Delta_m^{c}\|_F^2}{d_2} \leq C_1 \mu_1^2 d_2 r \left(\Big( a_m + \frac{b_m}{\kappa}\Big)^2\rho_m^2 + \frac{\kappa^2(a_m+b_m)^2}{(\kappa-1)^2b_m^2}c^2\vartheta_m^2\right).$$ This, together with Case 1, completes the proof.
[**Proof of Lemma \[lemben\].**]{}
Recall that $\frac{1}{m} \mathcal{R}^*_\Omega(\xi) = \frac{1}{m}\sum_{i=1}^m \xi_i \Theta_{\omega_i}$. Let $Z_i := \xi_i \Theta_{\omega_i}$. Since $\mathbb{E}(\xi_i)=0$, the independence of $\xi_i$ and $\Theta_{\omega_i}$ implies that $\mathbb{E}(Z_i)=0$. Since $\|\Theta_{\omega_i}\|_F=1$, we have that $$\|Z_i\|\leq \|Z_i\|_F = |\xi_i|\|\Theta_{\omega_i}\|_F = |\xi_i|.$$ It follows that $\big\|\|Z_i\| \big\|_{\psi_1} \leq \|\xi_i\|_{\psi_1}$. Thus, $\big\|\|Z_i\| \big\|_{\psi_1}$ is finite since $\xi_i$ is sub-exponential. Meanwhile, $\mathbb{E}^\frac{1}{2}(\|Z_i\|^2)\leq \mathbb{E}^\frac{1}{2}(\|Z_i\|_F^2) = \mathbb{E}^\frac{1}{2}(\xi_i^2)=1$. We also have $$\mathbb{E}\big(Z_iZ_i^\mathbb{T}\big)=\mathbb{E}\big(\xi^2_i \Theta_{\omega_i}\Theta_{\omega_i}^\mathbb{T}\big)=\mathbb{E}\big( \Theta_{\omega_i}\Theta_{\omega_i}^\mathbb{T}\big) =\sum_{k\in\beta}p_k \Theta_k \Theta_k^\mathbb{T}.$$ The calculation of $\mathbb{E}\big(Z_i^\mathbb{T} Z_i\big)$ is similar. From (\[eqndefL\]), we obtain that $\sqrt{1/n}\leq \sigma_Z \leq \sqrt{\mu_2/n}$. Then, applying the noncommutative Bernstein inequality yields (\[eqnnoibd\]). The proof of (\[eqnnoiexpd\]) is exactly the same as the proof of Lemma 6 in [@Klo12]. For simplicity, we omit the proof.
[**Proof of Lemma \[lemoper\]**]{}
\(i) From the definition of the sampling operator $\mathcal{R}_\Omega$ and its adjoint $\mathcal{R}_\Omega^*$, we have $$\frac{1}{m} \mathcal{R}_\Omega^* \mathcal{R}_\Omega(X) = \frac{1}{m} \sum_{i=1}^m \langle \Theta_{\omega_i}, X \rangle \, \Theta_{\omega_i}.$$ This is an average value of $m$ i.i.d. random matrices $\langle \Theta_{\omega_i}, X \rangle \Theta_{\omega_i}$. It is easy to see that $\mathbb{E}\big(\langle \Theta_{\omega_i}, X \rangle \Theta_{\omega_i}\big) = \mathcal{Q}_\beta(X).$ The result then follows directly from the strong law of large numbers.
\(ii) From the definition of $\mathcal{R}_\Omega^*$ and $\mathcal{R}_{\alpha\cup\beta}$, it is immediate to obtain that $$\frac{1}{\sqrt{m}} \mathcal{R}_{\mathcal{\alpha\cup\beta}} \mathcal{R}_\Omega^*(\xi) = \frac{1}{\sqrt{m}} \mathcal{R}_{\alpha\cup\beta} \bigg(\sum_{i=1}^m \xi_i \Theta_{\omega_i}\bigg) = \frac{1}{\sqrt{m}}\sum_{i=1}^m \xi_i \mathcal{R}_{\alpha\cup\beta}(\Theta_{\omega_i}).$$ Since $\mathbb{E}(\xi_i) = 0$ and $\mathbb{E}(\xi_i^2) =1$, from the independence of $\xi_i$ and $\mathcal{R}_{\alpha\cup\beta}(\Theta_{\omega_i})$, we have $\mathbb{E}\big(\xi_i \mathcal{R}_{\alpha\cup\beta}(\Theta_{\omega_i})\big) = 0$ and ${\rm cov}\big(\xi_i,\mathcal{R}_{\alpha\cup\beta}(\Theta_{\omega_i})\big)=p_i.$ Applying the central limit theorem then yields the desired result.
[**Proof of Theorem \[thmcons\]**]{}
Let $\Phi_m$ denote the objective function of (\[eqnrcs\]) and $\mathcal{F}$ denote the feasible set. Then, the problem (\[eqnrcs\]) can be concisely written as $$\min_{X \in \mathbb{V}^{n_1\times n_2}}\Phi_m(X) + \delta_\mathcal{F}(X).$$ By Assumptions \[asmpfun\] and \[asmpini\] and Lemma \[lemoper\], we have that $\Phi_m$ converges pointwise in probability to $\Phi$, where $\Phi(X):=\frac{1}{2}\|\mathcal{Q}_\beta(X-\overline{X})\|_2^2$ for any $X\in\!\mathbb{V}^{n_1\times n_2}$. As a direct extension of Rockafellar [@Roc70 Theorem 10.8], Andersen and Gill [@AndG82 Theorem II.1] proved that the pointwise convergence in probability of a sequence of random convex function implies the uniform convergence in probability on any compact subset. Then, from Proposition \[propepisum\] we obtain that $\Phi_m+\delta_\mathcal{F}$ epi-converges in distribution to $\Phi+\delta_\mathcal{F}$. Note that $\overline{X}$ is the unique minimizer of $\Phi(X) + \delta_\mathcal{F}(X)$ since $\Phi(X)$ is strongly convex over the feasible set $\mathcal{F}$. Using the convexity of $\Phi_m$ and $\Phi$, we complete the proof by Proposition \[propepicon\].
[**Proof of Lemma \[lemlocrk\]**]{}
By replacing $X$ and $H$ in (\[eqnlocsin\]) with $\overline{X}$ and $\rho\Delta$, respectively, and noting that $\sigma_{r+1}(\overline{X})=0$, we have $\sigma_{r+1}(\overline{X}+\rho\Delta)- \|\overline{U}_2^\mathbb{T}(\rho\Delta) \overline{V}_2\| = O(\|\rho\Delta\|_F^2).$ Since $\overline{U}_2^\mathbb{T}\overline{\Delta} \, \overline{V}_2 \neq 0$, for any $\rho\neq 0$ sufficiently small and $\Delta$ sufficiently close to $\overline{\Delta}$, $$\begin{aligned}
\frac{\sigma_{r+1}(\overline{X}+\rho\Delta)}{|\rho|} &=&\|\overline{U}_2^\mathbb{T} \Delta \overline{V}_2\| + O(|\rho|\|\Delta\|_F^2)\nonumber\\
&\ge& \|\overline{U}_2^\mathbb{T}\overline{\Delta} \, \overline{V}_2\| - \|\overline{U}_2^\mathbb{T}(\Delta-\overline{\Delta}) \overline{V}_2\|+ O(|\rho|\|\Delta\|_F^2) \\
&\ge& \frac{1}{2} \|\overline{U}_2^\mathbb{T}\overline{\Delta} \, \overline{V}_2\| >0.
\end{aligned}$$ This implies that ${\rm rank}(\overline{X}+\rho \Delta) > r$.
[**Proof of Proposition \[propdellim\]**]{}
By letting $\Delta:= \rho_m^{-1}(X-\overline{X})$ in the optimization problem (\[eqndellim\]), one can easily see that $\widehat{\Delta}_m$ is the optimal solution to $$\label{eqndellimappr}
\begin{aligned}
\min_{\Delta \in \mathbb{V}^{n_1\times n_2}}&\ {\displaystyle \frac{1}{2m}}\|\mathcal{R}_\Omega(\Delta)\|_2^2 - \frac{\nu}{m \rho_m}\langle \mathcal{R}_\Omega^*(\xi), \Delta\rangle + \frac{1}{\rho_m}\big(\|\overline{X}+\rho_m\Delta \|_* - \|\overline{X}\|_*\big) \\
&\ \ \hspace{2.4cm} - \langle F(\widetilde{X}_m), \Delta\rangle + \frac{\rho_m\gamma_m}{2}\|\Delta\|_F^2 + \gamma_m \langle \overline{X}-\widetilde{X}_m, \Delta \rangle \\
{\rm s.t.}\quad \ & \mathcal{R}_\alpha(\Delta) = 0.
\end{aligned}$$ Let $\Phi_m$ and $\Phi$ denote the objective functions of (\[eqndellimappr\]) and (\[eqndellim\]), respectively. Let $\mathcal{F}$ denote the feasible set of (\[eqndellim\]). By the definition of directional derivative and [@Wat92 Theorem 1], $$\lim_{\rho_m \rightarrow 0} \frac{1}{\rho_m}\big(\|\overline{X}+\rho_m\Delta \|_* - \|\overline{X}\|_*\big) = \langle \overline{U}_1\overline{V}_1^\mathbb{T}, \Delta \rangle + \|\overline{U}_2^\mathbb{T} \Delta \overline{V}_2\|_*.$$ Then, by combining Assumptions \[asmpfun\] and \[asmpini\] with Lemma \[lemoper\], we obtain that $\Phi_m$ converges pointwise in probability to $\Phi$. By using the same argument as in the proof of Theorem \[thmcons\], we obtain that $\Phi_m + \delta_\mathcal{F}$ epi-converges in distribution to $\Phi+\delta_\mathcal{F}$. Moreover, the optimal solution to (\[eqndellim\]) is unique due to the strong convexity of $\Phi$ over the feasible set $\mathcal{F}$. Therefore, we complete the proof by applying Proposition \[propepicon\] on the epi-convergence.
[**Proof of Lemma \[lemdelcond\]**]{}
Assume that $\overline{U}_2^\mathbb{T}\widehat{\Delta}\overline{V}_2=0$. Since $\widehat{\Delta}$ is the optimal solution to (\[eqndellim\]), from the optimality condition, the subdifferential of $\|X\|_*$ at $0$, and [@Roc70 Theorem 23.7], we obtain that there exist some $\widehat{\Gamma} \in \mathbb{V}^{(n_1-r)\times (n_2-r)}$ with $\|\widehat{\Gamma}\|\le 1$ and $\widehat{\eta} \in \mathbb{R}^{d_1}$ such that $$\label{eqndellimkkt}
\left\{
\begin{aligned}
& \mathcal{Q}_\beta(\widehat{\Delta}) + \overline{U}_1\overline{V}_1^\mathbb{T} - F(\overline{X}) - \mathcal{R}_\alpha^*(\widehat{\eta}) - \overline{U}_2 \widehat{\Gamma} \,\overline{V}_2^\mathbb{T}=0,\\
& \mathcal{R}_\alpha(\widehat{\Delta}) = 0.
\end{aligned}
\right.$$ Then, according to (\[relation-operator\]), we can easily obtain (\[eqndelvaldel\]) by applying the operator $\mathcal{Q}_\beta^\dag$ to the first equation of (\[eqndellimkkt\]) and using the second equation. By further combining (\[eqndelvaldel\]) and $\overline{U}_2^\mathbb{T}\widehat{\Delta}\overline{V}_2=0$, we obtain that $\widehat{\Gamma}$ is a solution to the linear system (\[eqndelcond\]). Conversely, if the linear system (\[eqndelcond\]) has a solution $\widehat{\Gamma}$ with $\|\widehat{\Gamma}\|\leq 1$, then it is easy to check that (\[eqndellimkkt\]) is satisfied with $\widehat{\Delta}$ being given by (\[eqndelvaldel\]) and $ \widehat{\eta} = \mathcal{R}_\alpha\big(\overline{U}_1\overline{V}_1^\mathbb{T}\!- F(\overline{X})-\overline{U}_2\widehat{\Gamma}\,\overline{V}_2^\mathbb{T}\big)$. Consequently, $\overline{U}_2^\mathbb{T} \widehat{\Delta} \overline{V}_2 = 0$ follows directly from the equations (\[eqndelcond\]) and (\[eqndelvaldel\]).
[**Proof of Theorem \[thmsuf\]**]{}
The estimator $\widehat{X}_m$ is the optimal solution to (\[eqnrcs\]) if and only if there exist a subgradient $\widehat{G}_m$ of the nuclear norm at $\widehat{X}_m$ and a vector $\widehat{\eta}_m \in \mathbb{R}^{d_1}$ such that $(\widehat{X}_m, \widehat{\eta}_m)$ satisfies the KKT conditions: $$\label{eqnsufkkt}
\left\{
\begin{aligned}
&\frac{1}{m}\mathcal{R}_\Omega^*\big(\mathcal{R}_\Omega(\widehat{X}_m)-y\big)+\rho_m \big(\widehat{G}_m\!-\!F(\widetilde{X}_m)+\gamma_m(\widehat{X}_m \!-\!\widetilde{X}_m)\big)- \mathcal{R}_\alpha^*(\widehat{\eta}_m)=0, \\
& \mathcal{R}_\alpha(\widehat{X}_m) = \mathcal{R}_\alpha(\overline{X}).
\end{aligned}
\right.$$ Let $ (\widehat{U}_m, \widehat{V}_m)\in \mathbb{O}^{n_1,n_2}(\widehat{X}_m)$ with $\widehat{U}_{m,1} \in \mathbb{O}^{n_1\times r}$, $\widehat{U}_{m,2} \in \mathbb{O}^{n_1\times (n_1-r)}$, $\widehat{V}_{m,1} \in \mathbb{O}^{n_2\times r}$ and $\widehat{V}_{m,2} \in \mathbb{O}^{n_2\times (n_2-r)}$. From Theorem \[thmcons\] and Corollary \[cororkrhs\], we know that $\text{rank}(\widehat{X}_m) \geq r$ with probability one. When $\text{rank}(\widehat{X}_m) \geq r$ holds, then from the characterization of the subdifferential of the nuclear norm [@Wat92; @Wat93], we have that $\widehat{G}_m = \widehat{U}_{m,1} \widehat{V}_{m,1}^\mathbb{T} + \widehat{U}_{m,2} \widehat{\Gamma}_m\widehat{V}_{m,2}^\mathbb{T}$ for some $\widehat{\Gamma}_m \in \mathbb{V}^{(n_1-r)\times (n_2-r)}$ satisfying $\|\widehat{\Gamma}_m\| \leq 1$. Moreover, if $\|\widehat{\Gamma}_m\| <1$, then $\text{rank}(\widehat{X}_m)=r$. Since $\widehat{X}_m \stackrel{p}{\rightarrow} \overline{X}$, by [@DinST10 Proposition 8] we have $\widehat{U}_{m,1}\widehat{V}_{m,1}^\mathbb{T} \stackrel{p}{\rightarrow} \overline{U}_1 \overline{V}_1^\mathbb{T}$. Together with Lemma \[lemoper\], the equation (\[eqnobs\]) and Lemma \[lemdelcond\], it is not hard to obtain that $$\begin{gathered}
\frac{1}{m\rho_m}\mathcal{R}_\Omega^*\big(\mathcal{R}_\Omega(\widehat{X}_m)-y\big)+\widehat{U}_{m,1}\widehat{V}_{m,1}^\mathbb{T} - F(\widetilde{X}_m)+\gamma_m(\widehat{X}_m-\widetilde{X}_m) \hspace{4cm} \nonumber\\
\hspace{6cm}\stackrel{p}{\rightarrow} \mathcal{Q}_\beta(\widehat{\Delta})+\overline{U}_1 \overline{V}_1^\mathbb{T} - F(\overline{X}) = \overline{U}_2\widehat{\Gamma}\overline{V}_2^\mathbb{T}, \label{eqnsufls}\end{gathered}$$ where the equality follows from (\[eqndelvaldel\]) and $\widehat{\Gamma}$ is the unique optimal solution to (\[eqndelcond\]). Then, by applying the operator $\mathcal{Q}_\beta^\dag$ to (\[eqnsufkkt\]), we obtain from (\[eqnsufls\]) that $$\label{eqnsuflsconv}
\overline{U}_2^\mathbb{T}\mathcal{Q}_\beta^\dag(\widehat{U}_{m,2}\widehat{\Gamma}_m\widehat{V}_{m,2}^\mathbb{T})\overline{V}_2 \stackrel{p}{\rightarrow} \overline{U}_2^\mathbb{T}\mathcal{Q}_\beta^\dag(\overline{U}_2\widehat{\Gamma}\overline{V}_2^\mathbb{T})\overline{V}_2.$$ Since $\widehat{X}_m \stackrel{p}{\rightarrow} \overline{X}$, according to [@DinST10 Proposition 7], there exist two sequences of matrices $Q_{m,U} \in \mathbb{O}^{n_1-r}$ and $Q_{m,V} \in \mathbb{O}^{n_2-r}$ such that $$\label{eqnsufuvprop}
\widehat{U}_{m,2}Q_{m,U} \stackrel{p}{\rightarrow} \overline{U}_2 \quad \text{and} \quad \widehat{V}_{m,2} Q_{m,V} \stackrel{p}{\rightarrow} \overline{V}_2.$$ Moreover, the uniqueness of the solution to the linear system (\[eqndelcond\]) is equivalent to the non-singularity of its linear operator. By combining (\[eqnsuflsconv\]) and (\[eqnsufuvprop\]), we obtain that $Q_{m,U}^\mathbb{T} \widehat{\Gamma}_m Q_{m,V} \stackrel{p}{\rightarrow} \widehat{\Gamma}.$ Hence, we obtain that $\|\widehat{\Gamma}_m\|<1$ with probability one since $\|\widehat{\Gamma}\|<1$. As discussed above, it follows that $\text{rank}(\widehat{X}_m) = r$ with probability one.
[**Proof of Proposition \[propdellimpos\]**]{}
It is easy to verify that $\widehat{\Delta}_m$ is the optimal solution to $$\label{eqndellimapprpos}
\begin{aligned}
\min_{\triangle\in\mathbb{S}^n} & \ \ {\displaystyle \frac{1}{2m}}\|\mathcal{R}_\Omega(\Delta)\|_2^2 - \frac{\nu}{m \rho_m}\langle \mathcal{R}_\Omega^*(\xi), \Delta\rangle + \langle I_n - F(\widetilde{X}_m), \Delta\rangle + \frac{\rho_m\gamma_m}{2}\|\Delta\|_F^2 \\
& \ \ \hspace{8.1cm} + \gamma_m \langle \overline{X}-\widetilde{X}_m, \Delta \rangle\\
{\rm s.t.} & \ \ \Delta \in \mathcal{F}_m := \rho_m^{-1}(\mathcal{C} \cap \mathbb{S}_+^n-\overline{X}),
\end{aligned}$$ where $\mathcal{C} := \big\{X \in \mathbb{S}^n \mid \mathcal{R}_\alpha(X) = \mathcal{R}_\alpha(\overline{X})\big\}$. Let $\Phi_m$ and $\Phi$ denote the objective functions of (\[eqndellimapprpos\]) and (\[eqndellimpos\]), respectively. Then $\Phi_m$ converges pointwise in probability to $\Phi$. Moreover, by considering the upper limit and lower limit of the family of feasible sets $\mathcal{F}_m$, we know that $\mathcal{F}_m$ converges in the sense of Painlev[é]{}-Kuratowski to the tangent cone $\mathcal{T}_{\mathcal{C}\cap \mathbb{S}_+^n}(\overline{X})$ (see [@RocW98; @BonS00]). Note that the Slater condition implies that $\mathcal{C}$ and $\mathbb{S}_+^n$ cannot be separated. Then, from [@RocW98 Theorem 6.42], we have $
\mathcal{T}_{\mathcal{C}\cap \mathbb{S}_+^n}(\overline{X}) = \mathcal{T}_\mathcal{C}(\overline{X}) \cap \mathcal{T}_{\mathbb{S}_+^n}(\overline{X}).
$ Clearly, $\mathcal{T}_\mathcal{C}(\overline{X}) = \{\Delta \in \mathbb{S}^n \mid \mathcal{R}_\alpha(\Delta)=0\}$. Moreover, by Arnold [@Arn71], $$\mathcal{T}_{\mathbb{S}^n_+}(\overline{X}) = \big\{\Delta \in\mathbb{S}^n \mid \overline{P}_2^\mathbb{T} \Delta \overline{P}_2 \in \mathbb{S}_+^{n-r} \big\}.$$ Since epi-convergence of functions corresponds to set convergence of their epigraphs [@RocW98], we obtain that $\delta_{\mathcal{F}_m}$ epi-converges to $\delta_{\mathcal{T}_{\mathcal{C} \cap \mathbb{S}_+^n}} = \delta_{\mathcal{T}_\mathcal{C}} + \delta_{\mathcal{T}_{\mathbb{S}_+^n}}$. Then, from Proposition \[propepisum\], $\Phi_m + \delta_{\mathcal{F}_m}$ epi-converges in distribution to $\Phi+\delta_{\mathcal{T}_\mathcal{C}} + \delta_{\mathcal{T}_{\mathbb{S}_+^n}}$. In addition, the optimal solution to (\[eqndellimpos\]) is unique due to the strong convexity of $\Phi$ over the feasible set $\mathcal{C}\cap \mathbb{S}_+^n$. Therefore, we complete the proof by applying Proposition \[propepicon\] on the epi-convergence.
[**Proof of Lemma \[lemdelcondpos\]**]{}
Note that the Slater condition also holds for the problem (\[eqndellimpos\]). (One may check the point $X^0-\overline{X}$.) Hence, $\widehat{\Delta}$ is the optimal solution to (\[eqndellimpos\]) if and only if there exists $(\widehat{\zeta}, \widehat{\Lambda}) \in \mathbb{R}^{d_1}\times \mathbb{S}^{n-r}$ such that $$\label{eqndellimkktpos}
\left\{
\begin{aligned}
& \mathcal{Q}_\beta(\widehat{\Delta}) + I_n - F(\overline{X}) - \mathcal{R}_\alpha^*(\widehat{\zeta}) - \overline{P}_2 \widehat{\Lambda} \overline{P}_2^\mathbb{T} =0,\\
& \mathcal{R}_\alpha(\widehat{\Delta}) = 0, \\
& \overline{P}_2^\mathbb{T}\widehat{\Delta} \overline{P}_2 \in \mathbb{S}_+^{n-r},\ \widehat{\Lambda} \in \mathbb{S}_+^{n-r},\ \langle \overline{P}_2^\mathbb{T} \widehat{\Delta} \overline{P}_2, \widehat{\Lambda}\rangle =0.
\end{aligned}\right.$$ Applying the operator $\mathcal{Q}_\beta^{\dag}$ to the first equation of (\[eqndellimkktpos\]) yields the equality (\[eqndelvaldelpos\]). Assume that $\overline{P}_2^\mathbb{T}\widehat{\Delta}\overline{P}_2=0$. Then, it is immediate to obtain from (\[eqndelvaldel\]) that $\widehat{\Lambda}$ is a solution to the linear system (\[eqndelcondpos\]).
Conversely, if the linear system (\[eqndelcondpos\]) has a solution $\widehat{\Lambda} \in \mathbb{S}_+^{n-r}$, then it is easy to check that (\[eqndellimkktpos\]) is satisfied with $\widehat{\Delta}$ given by (\[eqndelvaldelpos\]) and $\widehat{\zeta}=\mathcal{R}_\alpha\big(I_n - F(\overline{X})-\overline{P}_2\widehat{\Lambda}\,\overline{P}_2^\mathbb{T}\big)$. Then, $\overline{P}_2^\mathbb{T} \widehat{\Delta} \overline{P}_2 = 0$ directly follows from (\[eqndelvaldelpos\]) and the first equation of (\[eqndellimkktpos\]).
[**Proof of Theorem \[thmsufpos\]**]{}
The Slater condition implies that $\widehat{X}_m$ is the optimal solution to (\[eqnrcspos\]) if and only if there exists multipliers $(\widehat{\zeta}_m, \widehat{S}_m) \in \mathbb{R}^{d_1} \times \mathbb{S}^{n}$ such that $(\widehat{X}_m, \widehat{\zeta}_m, \widehat{S}_m)$ satisfy the KKT conditions: $$\label{eqnsufkktpos}
\left\{
\begin{aligned}
&\frac{1}{m}\mathcal{R}_\Omega^*\big(\mathcal{R}_\Omega(\widehat{X}_m)\!-\!y\big)+\rho_m \big(I_n\!-\! F(\widetilde{X}_m)+\gamma_m(\widehat{X}_m \!- \!\widetilde{X}_m)\big)\!- \!\mathcal{R}_\alpha^*(\widehat{\zeta}_m) - \widehat{S}_m=0, \\
&\mathcal{R}_\alpha(\widehat{X}_m) = \mathcal{R}_\alpha(\overline{X}), \\
&\widehat{X}_m \in \mathbb{S}_+^n,\ \widehat{S}_m \in \mathbb{S}_+^n,\ \langle \widehat{X}_m, \widehat{S}_m \rangle =0.
\end{aligned}\right.$$ The third equation of (\[eqnsufkktpos\]) implies that $\widehat{X}_m$ and $\widehat{S}_m$ can have a simultaneous eigenvalue decomposition. Let $\widehat{P}_m \in \mathbb{O}^{n}(\widehat{X}_m)$ with $\widehat{P}_{m,1} \in \mathbb{O}^{n\times r}$ and $\widehat{P}_{m,2} \in \mathbb{O}^{n\times (n-r)}$. From Theorem \[thmcons\] and Corollary \[cororkrhs\], we know that $\text{rank}(\widehat{X}_m) \geq r$ with probability one. When $\text{rank}(\widehat{X}_m) \geq r$ holds, we can write $\widehat{S}_m = \widehat{P}_{m,2} \widehat{\Lambda}_m\widehat{P}_{m,2}^\mathbb{T}$ for some diagonal matrix $\widehat{\Lambda}_m \in \mathbb{S}^{n-r}_+$. In addition, if $\widehat{\Lambda}_m \in \mathbb{S}^{n-r}_{++}$, then $\text{rank}(\widehat{X}_m)=r$. Since $\widehat{X}_m \stackrel{p}{\rightarrow} \overline{X}$, according to [@DinST10 Proposition 1], there exists a sequence of matrices $Q_m \in \mathbb{O}^{n-r}$ such that $\widehat{P}_{m,2}Q_m \stackrel{p}{\rightarrow} \overline{P}_2$. Then, using the similar arguments to the proof of Theorem \[thmsuf\], we obtain that $Q_m^\mathbb{T} \widehat{\Lambda}_m Q_m \stackrel{p}{\rightarrow} \widehat{\Lambda}$. Since $\widehat{\Lambda}\in \mathbb{S}_{++}^n$, we have $\widehat{\Lambda}_m \in \mathbb{S}_{++}^n$ with probability one. Thus, we complete the proof.
[**Proof of Proposition \[propcorcn\]**]{}
For the real covariance matrix case, the proof is given in [@QiS06 Lemma 3.3] and [@QiS11 Proposition 2.1]. For the complex covariance matrix case, one can use the similar arguments to prove the result.
We next consider the density matrix case. Suppose that $\overline{X}$ satisfies the density constraint, i.e., $\mathcal{R}_{\alpha}(\overline{X})=\text{Tr}(\overline{X})=1$. Note that for any $t\in\mathbb{R}$, we have $t \overline{X} \in \text{lin}(\mathcal{T}_{\mathcal{H}_+^n}(\overline{X}))$. This, along with $\text{Tr}(\overline{X})=1$, implies that $
\text{Tr}\big(\text{lin}(\mathcal{T}_{\mathcal{H}_+^n}(\overline{X}))\big)
= \mathcal{R}_{\alpha}\big(\text{lin}(\mathcal{T}_{\mathcal{H}_+^n}(\overline{X}))\big) =\mathbb{R}.
$ This means that the constraint nondegeneracy condition (\[eqncndcpos\]) holds.
[**Proof of Proposition \[propsoluni\]**]{}
We prove for the rectangular case by contradiction. Assume that there exists some $\mathbb{V}^{(n_1-r)\times (n_2-r)} \ni \overline{\Gamma} \neq 0$ such that $\mathcal{B}_2(\overline{\Gamma}) = \overline{U}_2^\mathbb{T}\mathcal{Q}_\beta^\dag(\overline{U}_2 \overline{\Gamma} \, \overline{V}_2^\mathbb{T})\overline{V}_2 = 0$. By noting that $\mathcal{Q}_\beta^\dag$ is a self-adjoint and positive semidefinite operator, we obtain $(\mathcal{Q}_\beta^\dag)^{1/2}(\overline{U}_2 \overline{\Gamma} \, \overline{V}_2^\mathbb{T})= 0$. It follows that $\mathcal{P}_\beta(\overline{U}_2 \overline{\Gamma} \, \overline{V}_2^\mathbb{T})= 0$. This, together with $\overline{\Gamma}\neq 0$, implies that $\overline{U}_2 \overline{\Gamma} \, \overline{V}_2^\mathbb{T} = \mathcal{P}_\alpha(\overline{U}_2 \overline{\Gamma} \, \overline{V}_2^\mathbb{T}) \neq 0$ and moreover $\mathcal{R}_\alpha(\overline{U}_2 \overline{\Gamma} \, \overline{V}_2^\mathbb{T}) \neq 0$. However, for any $H \in \mathcal{T}(\overline{X})$, we have $$\langle \mathcal{R}_\alpha(\overline{U}_2 \overline{\Gamma}\, \overline{V}_2^\mathbb{T}), \mathcal{R}_\alpha(H)\rangle = \langle\mathcal{P}_\alpha(\overline{U}_2 \overline{\Gamma} \, \overline{V}_2^\mathbb{T}), H\rangle = \langle \overline{U}_2 \overline{\Gamma} \, \overline{V}_2^\mathbb{T}, H\rangle = \langle \overline{\Gamma}, \overline{U}_2^{\mathbb{T}} H \overline{V}_2 \rangle =0.$$ Thus, the constraint nondegeneracy condition (\[eqncndc\]) implies that $\mathcal{R}_\alpha(\overline{U}_2 \overline{\Gamma} \, \overline{V}_2^\mathbb{T})=0$. This leads to a contradiction. Therefore, the linear operator $\mathcal{B}_2$ is positive definite. The proof for the positive semidefinite case is similar.
[**Proof of Theorem \[thmrccordencons\]**]{}
From Propositions \[propcorcn\] and \[propsoluni\], for both cases, the linear system (\[eqndelcondpos\]) has a unique solution $\widehat{\Lambda}$. Moreover, uniform sampling results in $\mathcal{Q}_\beta^\dag = \mathcal{P}_\beta /d_2$. Thus, from (\[eqndelcondpos\]), we get $$\label{eqnrccordenconseq}
\widehat{\Lambda} - \overline{P}_2^\mathbb{T}\mathcal{P}_\alpha(\overline{P}_2\widehat{\Lambda}\overline{P}_2^\mathbb{T})\overline{P}_2=\overline{P}_2^\mathbb{T}\mathcal{P}_\beta(\overline{P}_2\widehat{\Lambda}\overline{P}_2^\mathbb{T})\overline{P}_2 = \overline{P}_2^\mathbb{T}\mathcal{P}_\beta(I_n-F(\overline{X}))\overline{P}_2.$$
We first prove the covariance matrix completion by contradiction. Without loss of generality, we assume that the first $l$ diagonal entries are fixed and positive. Then, for any $X\in\mathbb{S}_+^n$, $\mathcal{P}_\alpha(X)$ is the diagonal matrix whose first $l$ diagonal entries are $X_{ii}, 1\leq i\leq l$ respectively and the other entries are $0$. Assume that $\widehat{\Lambda} \notin \mathbb{S}^{n-r}_{++}$, i.e., $\lambda_{\rm min}(\widehat{\Lambda})\leq 0$, where $\lambda_{\rm min}(\cdot)$ denotes the smallest eigenvalue. Then, we have $$\lambda_{\rm min}(\widehat{\Lambda}) = \lambda_{\rm min}(\overline{P}_2\widehat{\Lambda}\overline{P}_2^\mathbb{T}) \leq \lambda_{\rm min}\big(\mathcal{P}_\alpha(\overline{P}_2\widehat{\Lambda}\overline{P}_2^\mathbb{T})\big) \leq \lambda_{\rm min}\big(\overline{P}_2^\mathbb{T}\mathcal{P}_\alpha (\overline{P}_2\widehat{\Lambda}\overline{P}_2^\mathbb{T})\overline{P}_2\big),$$ where the equality follows from the fact that $\widehat{\Lambda}$ and $\overline{P}_2 \widehat{\Lambda}\overline{P}_2^\mathbb{T}$ have the same nonzero eigenvalues, the first inequality follows from the fact that the vector of eigenvalues is majorized by the vector of diagonal entries, (e.g., see [@MarOA10 Theorem 9.B.1]), and the second inequality follows from the Courant-Fischer minmax theorem, (e.g., see [@MarOA10 Theorem 20.A.1]). As a result, the left-hand side of (\[eqnrccordenconseq\]) is not positive definite. Notice that $\overline{P}_2^\mathbb{T} F(\overline{X})\overline{P}_2=0$. Thus, the right-hand side of (\[eqnrccordenconseq\]) can be written as $$\overline{P}_2^\mathbb{T}\mathcal{P}_\beta(I_n-F(\overline{X}))\overline{P}_2 = \overline{P}_2^\mathbb{T}\mathcal{P}_\beta(I_n)\overline{P}_2 +\overline{P}_2^\mathbb{T} \mathcal{P}_\alpha(F(\overline{X}))\overline{P}_2 = \overline{P}_2^\mathbb{T} \big(\mathcal{P}_\beta(I_n)+\mathcal{P}_\alpha(F(\overline{X}))\big)\overline{P}_2.$$ Since $\text{rank}(\overline{X})=r$, with the choice (\[eqnfcorden\]) of $F$, we have that for any $1\leq i\leq l$, $$\overline{X}_{ii} = \sum_{j=1}^r\lambda_j(\overline{X}) |\overline{P}_{ij}|^2 >0 \quad \text{implies} \quad
\big(F(\overline{X})\big)_{ii} = \sum_{j=1}^r f_i\big(\lambda_j(\overline{X})\big) |\overline{P}_{ij}|^2 >0.$$ Moreover, $\mathcal{P}_\beta(I_n)$ is the diagonal matrix with the last $n-r$ diagonal entries being ones and the other entries being zeros. Thus, $\mathcal{P}_\beta(I_n)+\mathcal{P}_\alpha(F(\overline{X}))$ is a diagonal matrix with all positive diagonal entries. It follows that the right-hand side of (\[eqnrccordenconseq\]) is positive definite. Thus, we obtain a contradiction. Hence, $\widehat{\Lambda} \in \mathbb{S}^{n-r}_{++}$. Then, from Theorem \[thmsufpos\], we obtain the rank consistency.
For the density matrix completion, $\mathcal{P}_\alpha(\cdot) = \frac{1}{n}\text{Tr}(\cdot)I_n$. By further using $\overline{P}_2^\mathbb{T} F(\overline{X})\overline{P}_2= 0$ and $\mathcal{P}_\beta(I_n)=0$, we can rewrite (\[eqnrccordenconseq\]) as $$\widehat{\Lambda}-\frac{1}{n}\text{Tr}(\widehat{\Lambda})I_{n-r} = \frac{1}{n}\text{Tr}(F(\overline{X}))I_{n-r}.$$ By taking the trace on both sides, we obtain that $\widehat{\Lambda} = \frac{1}{r}\text{Tr}(F(\overline{X}))I_{n-r}.$ Since $\overline{X}$ is a density matrix of rank $r$, with the choice (\[eqnfcorden\]) of $F$, we have that $$\text{Tr}(\overline{X}) = \sum_{i=1}^n \sum_{j=1}^r\lambda_j(\overline{X}) |\overline{P}_{ij}|^2 = 1 \quad \text{implies} \quad
\text{Tr}\big(F(\overline{X})\big) = \sum_{i=1}^n\sum_{j=1}^r f_i\big(\lambda_j(\overline{X})\big) |\overline{P}_{ij}|^2 >0.$$ It follows that $\widehat{\Lambda} \in \mathbb{S}_{++}^{n-r}$ and thus we obtain the rank consistency.
[^1]: Department of Mathematics, National University of Singapore, 10 Lower Kent Ridge Road, Singapore 119076 ([email protected]).
[^2]: Department of Mathematics, South China University of Technology, Tianhe District of Guangzhou City, China ([email protected]).
[^3]: Department of Mathematics and Risk Management Institute, National University of Singapore, 10 Lower Kent Ridge Road, Singapore 119076 ([email protected]). This author’s research is supported in part by Academic Research Fund under grant R-146-000-149-112.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A single physical interaction might not be universal for quantum computation in general. It has been shown, however, that in some cases it can achieve universal quantum computation over a subspace. For example, by encoding logical qubits into arrays of multiple physical qubits, a single isotropic or anisotropic exchange interaction can generate a universal logical gate-set. Recently, encoded universality for the exchange interaction was explicitly demonstrated on three-qubit arrays, the smallest nontrivial encoding. We now present the exact specification of a discrete universal logical gate-set on four-qubit arrays. We show how to implement the single qubit operations exactly with at most $3$ nearest neighbor exchange operations and how to generate the encoded controlled-NOT with $27$ parallel nearest neighbor exchange interactions or $50$ serial gates, obtained from extensive numerical optimization using genetic algorithms and Nelder-Mead searches. We also give gate-switching times for the three-qubit encoding to much higher accuracy than previously and provide the full specification for exact $CNOT$ for this encoding. Our gate-sequences are immediately applicable to implementations of quantum circuits with the exchange interaction.'
address: |
$^1$ Department of Chemistry, $^2$ Computer Science Division, University of California, Berkeley\
$^3$ CNRS-LRI, UMR 8623, Université de Paris-Sud, 91405 Orsay, France
author:
- 'M. Hsieh$^{1}$, J. Kempe$^{1,2,3}$, S. Myrgren$^1$ and K. B. Whaley$^1$'
title: 'An Explicit Universal Gate-set for Exchange-Only Quantum Computation'
---
Introduction
============
To implement universal computation in the quantum regime, one must be able to generate any unitary transformation on the logical qubit states. By now it has become part of the quantum computation folklore that the group $SU(2)$ of single-qubit operations and an entangling two-qubit operation such as the controlled-NOT ($CNOT$) can generate any unitary transformation exactly [@DiVincenzo:95a; @Barenco:95a]. Furthermore it has been shown that there are [*discrete*]{} universal elementary gate-sets which approximate any unitary transformation with arbitrary precision efficiently[^1] (see [@Kitaev:book; @Nielsen:book] for details). One such set is comprised of $\{H,\frac{\pi}{8},CNOT\}$ [@Boykin:99a], where $H$ is the Hadamard transform and $\frac{\pi}{8}$ is a phase gate, both acting on a single qubit. In this sense, $H$, $\frac{\pi}{8}$ and $CNOT$ comprise a quantum analogue to a classical *universal logical gate-set*.
The traditional paradigm of quantum computation of “one physical qubit = one logical qubit” is often hard to implement because in the presently known menu of physical implementation schemes, it is usually difficult to control at least one of either the single-body or the two-body operations [@Kempe:02a].
A prime example is the Heisenberg interaction (with Hamiltonian $H_E^{i,j}=J_{ij} \vec{S}_i \otimes \vec{S}_j$ between spin particles $i$ and $j$, where $\vec{S}_i=\frac{1}{2} \vec{\sigma}^i$ and $\sigma^i_{x,y,z}$ are the usual Pauli matrices acting on qubit $i$). It has many attractive features [@Loss:98a; @Burkard:99b] that have led to its being chosen as the fundamental two-qubit interaction in a large number of recent proposals: Its functional form is very accurate — deviations from the isotropic form of the interaction, arising only from relativistic corrections, can be very small in suitably chosen systems. It is a strong interaction, so that it should permit very fast gate operation, well into the GHz range for several of the proposals. At the same time, it is very short ranged, arising from the spatial overlap of electronic wavefunctions, so that it should be possible to have an on-off ratio of many orders of magnitude. We will assume that the interaction can be switched on and off between coupled qubits [@Burkard:99b]. Unfortunately, the Heisenberg interaction by itself is not a universal gate, in the sense that it cannot generate any arbitrary unitary transformation on a collection of spin-1/2 qubits. So, every proposal has supplemented the Heisenberg interaction with some other means of applying independent one-qubit gates (which can be thought of as time-dependent local magnetic fields). But the need to add this capability to the device adds considerably to the complexity of the structures, by putting unprecedented demands on “g-factor” engineering of heterostructure materials [@DiVincenzo:99a; @Vrijen:00a], requiring that strong, inhomogeneous magnetic fields be applied, or involving microwave manipulations of the spins that may be slow and may cause heating of the device. These added complexities may well exact a high cost, perhaps degrading the quantum coherence and clock rate of these devices by several orders of magnitude.
*Encoded universality* [@Kempe:01b] provides a way around this problem in some crucial cases, for example when the “easy” interaction is the exchange interaction, by entirely eliminating the need for single-body physical operations. By encoding each logical qubit in an array of multiple physical qubits, sequences of two-body nearest-neighbor exchange interactions are sufficient to generate the logical $SU(2)$ and $CNOT$ operations[^2] on the encoded qubits [@Kempe:01a; @Kempe:01b; @Bacon:99b] and single-spin operations and all their attendant difficulties can be avoided.
One drawback of the theory of encoded universality [@Kempe:01a] is that it establishes the sufficiency of certain two-body interactions for universality in a non-constructive way, not offering explicit methods with which to specify the sequences of physical implementable Hamiltonians corresponding to the encoded logical gates. In particular it is not clear at the outset how many physical interactions are required to implement each of the logical gates in some layout of the qubits. Encoded computation schemes are only viable if the number of physical interactions to be applied to the qubits is not too large, where the threshold is determined by currently achievable decoherence and switching times. In most cases, numerical methods are the only way to find explicit sequences of Hamiltonians for a set of universal gates for some realistic arrangement of the physical qubits. Recently, more or less explicit universal logical gate-sets have been given for a three-qubit encoding using only the exchange interaction [@DiVincenzo:00a], for the $XY$-interaction [@Kempe:02a] and for the generalized anisotropic exchange Hamiltonian [@Vala:02a]. In [@DiVincenzo:00a] an initial encoding of three physical qubits per logical qubit is used and a sequence of 19 Hamiltonians is presented that implements the encoded $CNOT$. However this $CNOT$ is given up to local unitary operations only, and the encoded single-qubit operations are given in terms of Euler-angle rotations for the group $SU(2)$. Some further processing is needed to obtain a universal discrete gate-set, needed to implement quantum circuits [*in terms of the computational basis*]{} [@Myrgren:03a].
We present here a complete scheme for universal quantum computation on four-qubit encodings in a one- (or two-)dimensional layout with nearest neighbor interactions only. We specify the encoding and layout and give all the gate switching times to obtain the encoded $H$, $\frac{\pi}{8}$ and $CNOT$ in the computational basis without further post-processing. This scheme provides an immediately applicable building block for exchange-only quantum circuits. We also provide new gate sequences for the $CNOT$ in the three-qubit encoding to higher precision and with different symmetries than in [@DiVincenzo:00a] and provide the complete set of gates for the exact encoded $CNOT$ in this smaller encoding. Although the four-qubit encoding has a slightly larger overhead in spatial resources than the three-qubit encoding it offers several advantages. A quantum computation begins by setting all encoded qubits to the (logical) zero state. In our scheme this state is a tensor product of singlet states. This state is easily obtained using the exchange interaction: if a strong $H_{12}$ is turned on in each coded block and the temperature made lower than the strength $J$ of the interaction, these two spins will equilibrate to their ground state, which is the singlet state. Unlike the smaller three-qubit encoding we do not require here any additional weak magnetic fields for initialization. This aspect renders the four-qubit encoding particularly attractive. Another advantage is that the four-qubit scheme is conceptually simpler for use in quantum logic when the properties of robustness to noise are also taken into account. Whereas the four-qubit logical states constitute a decoherence free subspace (DFS) under collective decoherence [@Kempe:01a], the three-qubit logical states constitute a decoherence free subsystem in which the logical state evolution is defined by only one component of the tensor product space. A third advantage of the four-qubit encoding is that additional protection against single qubit errors can be achieved in this case by control of extra exchange interactions to form a supercoherent qubit [@Bacon:01b].
The structure of this paper is as follows. First, we describe the four-qubit encodings which define our logical space and give the physical layout of the qubits. We then specify the Hamiltonians required to generate the single-qubit operations and the $CNOT$ gate on the encoded qubits. The numerical methods used to obtain the exchange gate sequence for the encoded $CNOT$ and further details are described in Appendices \[App:1\] and \[App:2\]. The same numerical methods are used to obtain new high accuracy gate sequences for the exact $CNOT$ in the three-qubit encoding. We conclude with a brief discussion of a solid-state implementation scheme in which these results can be readily applied, and some open problems meriting further consideration.
Four-qubit encoding
===================
We define the logical zero-state and one-state on one encoded qubit as $$\begin{aligned}
|0_{L}\rangle = \frac{1}{2}\left(|01\rangle - |10\rangle
\right)\otimes\left(|01\rangle - |10\rangle \right)
\label{E:encoding1} \\
|1_{L}\rangle =
\frac{1}{\sqrt{3}}\left(|t_{+}\rangle\otimes|t_{-}\rangle -
|t_{0}\rangle\otimes|t_{0}\rangle
+|t_{-}\rangle\otimes|t_{+}\rangle \right), \label{E:encoding2}\end{aligned}$$ where $|t_{0}\rangle = \frac{1}{\sqrt{2}}\left(|01\rangle +
|10\rangle\right)$, $|t_{-}\rangle = |00\rangle$, and $|t_{+}\rangle =
|11\rangle$. For a more detailed discussion on how to obtain these encodings, refer to [@Kempe:01b; @Bacon:99b]. To initialize a computation all logical qubits have to prepared in the $|0_L\ra$ state. Note that here the $|0_L\ra$ state is a tensor product of singlets $\frac{1}{\sqrt{2}} (|01\ra - |10\ra)$. As detailed in the previous section this will be advantageous in many experimental settings since it will permit easy initializion of the logical qubits at the beginning of a computation.
The arrays are spatially configured to permit serial nearest-neighbor exchanges between the physical qubits in a one-dimensional layout:
\[QubitLayout\]
(0,80)
We could also imagine these qubit arrays in a two-dimensional layout, where several of the one-dimensional layers are stacked on top of each other. Our construction of gate sequences will also hold for the two-dimensional setting. It suffices to note that the only difference is that along the layers of arrays the fourth qubit of each array is coupled to the first qubit of the next, whereas between layers the first qubit of one array in one layer is interacting with the first qubit of an array in the other. But note that both encoded states $|0_L\ra,|1_L\ra$ are symmetric with respect to swapping qubit $1$ with qubit $4$ and qubit $2$ with qubit $3$, so we obtain exactly the same gate-sequence for $CNOT$ for a coupling of the two first qubits of an array - all we need to do is to relabel the qubits on the bottom array as $4,3,2,1$.
The basis states of the logical space defined by two encoded qubits are $|0_{L} 0_{L}\rangle$, $|0_{L} 1_{L}\rangle$, $|1_{L}
0_{L}\rangle$ and $|1_{L} 1_{L}\rangle$.
Single-Qubit Operations
=======================
Our goal is to construct the single-qubit Hadamard $H$ and $\frac{\pi}{8}$ gates, defined as $$\frac{\pi}{8} = e^{i\pi/8} \left( \begin{array}{cc} e^{-i\pi/8} & 0 \\
0 & e^{i\pi/8} \end{array} \right),~H = \frac{1}{\sqrt{2}} \left( \begin{array}{cc} 1 & 1 \\
1 & -1 \end{array} \right). \label{E:HadPi8}$$ on the encoded qubits, using a sequence of exchange interactions $H_E^{i,i+1}=\frac{J_{i,i+1}}{4} (\sigma_{x}^i\sigma_{x}^{i+1} + \sigma_{y}^i\sigma_{y}^{i+1} + \sigma_{z}^i\sigma_{z}^{i+1})$ on adjacent pairs of physical qubits $i$ and $i+1$. The matrices $\sigma_{x,y,z}^i$ are the usual Pauli matrices acting on the $i$th qubit and $J_{i,i+1}$ is the coupling constant. When we write $H_E^{i,i+1}$ we assume that the Hamiltonian acts on the $i$th and $i+1$st qubit as specified and as the identity on all the other qubits. For convenience we are going to add a multiple of the identity to $H_E$ (which just gives an unobservable global phase), and work with the rescaled interaction $$E^{i,i+1}=\frac{1}{2}(\sigma_x^i \otimes \sigma_x^{i+1} +\sigma_y^i \otimes \sigma_y^{i+1} +\sigma_z^i \otimes \sigma_z^{i+1} +I^i \otimes I^{i+1})=
\left(
\begin{array}{llll} 1 & 0 & 0 & 0 \\ 0 & 0 & 1 & 0\\0 & 1 & 0 & 0\\ 0 & 0 & 0 & 1
\end{array}
\right). \label{E:Hexchange}$$ so that $exp \left( - (J_{ij}/\hbar)t {\bar S}_i \cdot {\bar S}_j \right) \equiv exp \left( -i(J_{ij}/2\hbar)t E^{i,j} \right)$ (up to a global phase). We will give all exchange times in units of $2\hbar/J_{ij}$ [@Loss:98a].
Consider the effect of two particular exchanges on the logical states of a single encoded qubit. First, we note that in the code-subspace the action of the exchange $E^{1,2}$ is equal to that of $E^{3,4}$, with both of these generating the transformation $|0_{L}\rangle \rightarrow -|0_{L}\rangle$ and $|1_{L}\rangle
\rightarrow |1_{L}\rangle$. So -$E^{1,2}$ is equivalent to a $\sigma_{z}$ operation on a single logical qubit. Therefore the Hamiltonian for the encoded $\frac{\pi}{8}$ operation, up to an unobservable global phase, is exactly $e^{i\frac{\pi}{8}E^{1,2}}$.
Next, consider the exchange $E^{2,3}$. The action of this in the code-space is equivalent to $E^{1,4}$. The effect of the $E_{23}$ operation on a single logical qubit is: $$E^{2,3} = \left(
\begin{array}{cc} \frac{1}{2} & \frac{\sqrt{3}}{2}\\ \frac{\sqrt{3}}{2} &
-\frac{1}{2}\\ \end{array} \right). \label{E:SigXcomp}$$
By examining the effect of the $E^{1,2}$ and $E^{2,3}$ operations on the code-space, we can obtain an exact encoded Hadamard operation as $H
= e^{it_{1}E^{1,2}}e^{it_{2}E^{2,3}}e^{it_{1}E^{1,2}}$, where $t_{1} = \frac{1}{2}\arcsin\sqrt{\frac{2}{3}}=0.4777$ and $t_{2}
= \arccos{\sqrt{\frac{1}{3}}}=0.9553$.
The exact specifications of exchange operations for these single qubit gates are depicted in Fig. 2.
\[OneQubitOperations\]
(0,130)
Encoded $CNOT$ operation {#sec:CNOT}
========================
To obtain the encoded $CNOT$ we used numerical methods and proceeded in two stages. In the first step we attempted to obtain a gate $U_{cnot}^{exchange}$ that is [*equivalent*]{} to the encoded $CNOT$ [*up to local unitary transformations*]{} on the encoded qubits. $U_{cnot}^{exchange}$ is [*locally equivalent*]{} to $CNOT$ if there are single-qubit unitary operations $U_1,U_2,V_1,V_2$ (acting on the first and second encoded qubit, respectively) such that $$\label{locequivalent}
CNOT = (U_1 \otimes U_2) U_{cnot}^{exchange} (V_1 \otimes V_2)$$ In the second stage we numerically obtained the local unitary operations $U_1,U_2,V_1,V_2$ to get from $U_{cnot}^{exchange}$ to the real $CNOT$ in the computational basis and to obtain the gate-sequences of exchange interactions corresponding to each of these local operations $U_i,V_i$, $i=1,2$. The reason for splitting the task into these two stages is the following. A result by Makhlin [@Makhlin:00a] shows that all locally equivalent gates are characterized by only three real parameters, $M_1$ (a complex number) and $M_2$ (real), which we will refer to as the [*Makhlin-invariants*]{} in what follows. Appendix \[App:1\] gives a brief summary of how $M_1$ and $M_2$ are calculated. For the $CNOT$, $M_1=0$ and $M_2=1$. The reduction to three parameters greatly simplifies the numerical search and allowed us to obtain the gate sequences by a combination of genetic algorithms [@Buckles:book] and Nelder-Mead simplex searches [@Lagarias:95a; @Nelder:65a]. The Makhlin invariants give rise to a simple fitness function $f = \|M_{1}(CNOT) - M_{1}(Candidate)\| + \|M_{2}(CNOT) -
M_{2}(Candidate)\|$, where $\| .\|$ is the complex norm. A gate sequence that generates a value $f=0$ is therefore equivalent to a $CNOT$.
It is imperative for succesful quantum computation over a subspace that any permissible operation over the encoded qubits must act unitarily on linear combinations of these basis states and not “leak” any amplitude out of the subspace into its complement and vice versa. We will capture this requirement by defining a *leakage parameter* $\Lambda$. Then any permissible two-qubit physical operation $W$ must keep the code-subspace invariant, i.e. obey the following equation[^3]: $$\begin{aligned}
\Lambda = 4-\sum_{s=0}^{1}\sum_{t=0}^{1}\sum_{u=0}^{1}\sum_{v=0}^{1} |\langle
s_{L} t_{L}| W |u_{L} v_{L} \rangle|^{2} = 0. \label{E:Leakage}\end{aligned}$$ If leakage occurs $\Lambda >0$. We note that any $W$ that is locally equivalent to $CNOT$ or any other unitary logical operation over two encoded qubits must *by definition* generate a leakage parameter of $\Lambda=0$. However, we have found through numerical inspection that the search space in this problem is heavily pocked with local minima in which the Makhlin invariants are close to the desired values for $CNOT$, but which “leak” out of the logical space by generating a $\Lambda$ value larger than $0$. In our numerical searches, we overcame this problem by defining our fitness function as $\mathcal{F}$ = f + $\Lambda$, optimizing explicitly for not only a Makhlin invariant match, but also for non-leakage.
A detailed description of our search algorithms and of the accuracy of our gate sequence can be found in Appendix \[App:2\].
For a gate $U_{cnot}^{exchange}$, locally equivalent to the $CNOT$, we found a gate sequence of $34$ nearest neighbor exchange interactions. Figure 3 and Table \[tab:4qbcnot\] show the layout and time parameters.
\[CNOTcircuit\]
(0,170)
---------- ------- ------- --------- ---------- ------- ------- ---------
Exchange Qubit Qubit Exchange Qubit Qubit
Time 1 2 Times Time 1 2 Times
$t_{1}$ 4 5 1.90680 $t_{18}$ 7 8 0.95629
$t_{2}$ 3 4 1.59536 $t_{19}$ 2 3 1.06260
$t_{3}$ 5 6 1.26290 $t_{20}$ 6 7 0.68131
$t_{4}$ 2 3 1.59745 $t_{21}$ 3 4 0.59800
$t_{5}$ 6 7 2.06920 $t_{22}$ 5 6 1.19942
$t_{6}$ 1 2 0.05331 $t_{23}$ 4 5 1.04719
$t_{7}$ 7 8 0.76951 $t_{24}$ 3 4 3.14138
$t_{8}$ 2 3 1.59747 $t_{25}$ 5 6 0.95529
$t_{9}$ 6 7 0.71337 $t_{26}$ 2 3 1.63957
$t_{10}$ 3 4 1.59958 $t_{27}$ 6 7 1.91303
$t_{11}$ 5 6 1.26287 $t_{28}$ 1 2 2.47920
$t_{12}$ 4 5 1.90667 $t_{29}$ 7 8 2.18627
$t_{13}$ 3 4 0.59810 $t_{30}$ 2 3 1.05736
$t_{14}$ 5 6 1.71467 $t_{31}$ 6 7 0.94814
$t_{15}$ 2 3 1.06264 $t_{32}$ 3 4 3.14170
$t_{16}$ 6 7 0.91559 $t_{33}$ 5 6 4.09690
$t_{17}$ 1 2 2.30240 $t_{34}$ 4 5 2.09434
---------- ------- ------- --------- ---------- ------- ------- ---------
: Gate switching times for the sequence of $34$ exchange interactions of Figure 3, given in units of $2 \hbar/ J$.[]{data-label="tab:4qbcnot"}
.1cm
Note that those exchanges that are on a vertical line in Figure 3 involve disjoint sets of qubits and can be applied [*in parallel*]{}, where each cycle of gates lasts as long as the longest switching time in the set of parallel gates. If we count the number of parallel operations, we obtain $19$ gate cycles.
In the second stage we searched for the encoded local unitary gates $U_1,U_2,V_1,V_2$ (see Eq. (\[locequivalent\])) that transform the $34$-gate sequence into an exact $CNOT$ on the computational basis states, Eqs. \[E:encoding1\] and \[E:encoding2\]. It has been shown previously [@Bacon:99b; @DiVincenzo:00a] that each encoded local unitary can be obtained by a sequence of four exchange gates. We employ here the nearest neighbor layout as shown in Figure \[fig:4qbUV4\]. The constructions in [@Bacon:99b; @DiVincenzo:00a] involve non-nearest neighbor interactions, with $E_{13}$ instead of $E_{23}$. It is easy to see, however, that replacing $E_{13}$ with $E_{23}$ in the arguments of [@Bacon:99b; @DiVincenzo:00a] also leads to a sequence of four exchanges.
![Local unitaries can be generated using a sequence of $4$ exchange gates as shown.[]{data-label="fig:4qbUV4"}](4qbuv4.eps)
These $16$ remaining gates ($4$ for each local unitary) can be obtained either as the result of numerical optimization of a suitable cost function, or from solving the system of non-linear equations derived from the element-wise equivalency condition between the objective matrix and the product of the four exchange matrices. We employed the optimization approach using a Nelder-Mead simplex search because of its high efficiency and generality. A similar approach has been used in [@Myrgren:03a], where the cost function (a combination of a matrix distance between the actual gate and the desired gate and a non-leakage requirement) and the details of our numerical search calculations can be found. To reduce the probability of sampling only local minima, we sampled a large number (5 million) of randomly selected initial points in parameter space. We were able to determine the four local unitaries and their corresponding $4$-gate exchange sequences to a precision of $10^{-4}$ in the cost function, with the corresponding maximum matrix element distance of the order of $10^{-5}$.
An alternative approach to finding gate sequences for local unitaries using the standard mapping from $SU(2)$ to $SO(3)$ and a quaternion representation of $SO(3)$ can be found in Ref. [@Myrgren:03a].
Table \[tab:localunitaries\] shows the gate times for each of the 4 exchange interactions required to implement the encoded local unitary operations.
---------- ------- ------- ---------- ---------- ---------- ----------
Exchange Qubit Qubit U1 U2 V1 V2
Time 1 2 Times Times Times Times
$t_1$ 1 2 2.218823 1.391831 4.865658 0.933012
$t_2$ 2 3 4.386508 1.977325 3.141319 2.025429
$t_3$ 1 2 3.442139 2.974488 1.493938 1.315318
$t_4$ 2 3 1.808165 2.105277 3.141314 0.042865
---------- ------- ------- ---------- ---------- ---------- ----------
: The exchange gates and corresponding gate times (in units of $2 \hbar/ J$) required to transform $W$ into the actual $CNOT$ gate.[]{data-label="tab:localunitaries"}
.1cm
Note that the exchange interactions implementing $U_1$ and $U_2$ can be applied in parallel, as can those for $V_1$ and $V_2$. Thus, transforming the 34-gate sequence of Fig. 3 into the exact $CNOT$ gate on the computational basis requires $16$ additional nearest-neighbor interactions, that can be realized as $8$ additional parallel gate cycles.
The total number of nearest neighbor interactions for the exact $CNOT$ amounts to $27$ if applied in parallel and $50$ if applied serially.
Gate sequences for the three-qubit encoding
===========================================
A $19$ exchange gate sequence for a gate locally equivalent to $CNOT$ for the three qubit encoding has been given in [@DiVincenzo:00a]. The layout in [@DiVincenzo:00a] obeyed certain symmetry constraints. We have recalculated this sequence to a higher accuracy with and without these symmetry constraints and have also computed the exchange-only implementation of the local unitaries $U_i,V_i$, $i=1,2$, needed to transform the $19$-gate sequence to the exact $CNOT$ in the computational basis. The resulting sequences for exact $CNOT$ in the three-qubit encoding are given in Appendix \[App:3\].
Conclusion
==========
We have presented an exact construction of a discrete universal logical gate-set using only the exchange operation with a four-qubit encoding. These results are readily applicable to physical implementation schemes in which exchange interactions are favored. These include the classical solid-state nuclear-spin qubit model proposed by Kane [@Kane:98a; @Kane:00a] and the electron-spin qubit proposal of Loss and DiVincenzo [@Loss:98a; @Burkard:99b]. For a four-fold increase in the number of system qubits and a twenty-nine-fold increase in the number of computational cycles for the two-qubit operation, we are able to simplify the implementation of spin coupled solid state systems by entirely eliminating the need for single-spin *A*-gates. In contrast to the rapid and relatively easily-tunable two-spin *J*-gates, the *A*-gates demand considerably greater device complexity and $g$-factor engineering on solid-state heterostructures [@Vrijen:00a; @DiVincenzo:99a].
Thus far, explicit constructions of universal logical gate-sets for exchange-only quantum circuits have been given on three-qubit encodings [@DiVincenzo:00a] and for the four-qubit scheme presented here. It should be noted that in principle the overhead in spatial resources can be made arbitrarily small: asymptotically the rate of encoding into subsystems converges to unity [@Kempe:01a]. However we have to carefully evaluate the trade-offs in space and time for each encoding. So far no constructive analytical methods to lower bound the number of nearest neighbor interactions for encoded gates exist. Using numerical methods yields an increase from $19$ to $34$ gates for a gate equivalent to the encoded $CNOT$ going from a three-qubit [@DiVincenzo:00a] to the present four-qubit encoding. This seems to indicate that the rate of growth of the number of nearest neighbor gates needed is rather large and that it is probably wise to stick to small encodings if the error correcting properties are not also to be incorporated. However, we note that these are all numerical solutions and are not guaranteed to be optimal. It would therefore be useful to obtain analytic bounds on the minimum number of exchange gates required for encoded operations.
Another open problem is the application of encoded universality to other interactions encountered in nature and in the laboratory, to facilitate the search towards optimal physical schemes for implementation of universal quantum computation. We believe that the scheme presented here provides a step in this direction and alleviates the task of the quantum engineer working towards spin-coupled solid state quantum computation.
Acknowledgements {#acknowledgements .unnumbered}
================
The effort of the authors is sponsored by the Defense Advanced Research Projects Agency (DARPA) and Air Force Laboratory, Air Force Materiel Command, USAF, under agreement number F30602-01-2-0524, and by DARPA and the Office of Naval Research under grant number FDN-00014-01-1-0826.
[10]{}
D. P. DiVincenzo, Phys. Rev. A [**51**]{}, 1015 (1995).
D. Barenco [*et al.*]{}, Phys. Rev. A [**52**]{}, 3457 (1995), ANL preprint quant-ph/9503016.
A. Kitaev, A. Shen, and M. Vyalyi, [*Classical and Quantum Computation*]{}, No. 47 in [*Graduate Series in Mathematics*]{} (AMS, Providence, RI, 2002).
M. Nielsen and I. Chuang, [*Quantum Computation and Quantum Information*]{} (Cambridge University Press, Cambridge, UK, 2000).
P. O. Boykin [*et al.*]{}, in [*Proceedings of 40th Annual Symposium on the Foundations of Computer Science (FOCS)*]{} (IEEE Press, Los Alamitos, CA, 1999), pp. 486–494, ANL preprint quant-ph/9906054.
J. Kempe and K. Whaley, Phys. Rev. A [**65 (5)**]{}, 052330 (2002).
D. Loss and D. DiVincenzo, Phys. Rev. A [**57**]{}, 120 (1998).
G. Burkard, D. Loss, and D. DiVincenzo, Phys. Rev. B [**59**]{}, 2070 (1999).
D. P. DiVincenzo, G. Burkard, D. Loss, and E. V. Sukhorukov, in [*Quantum Mesoscopic Phenomena and Mesoscopic Devices in Microelectronics.*]{}, edited by I. O. Kulk and R. Ellialtioglu (NATO Advanced Study Institute, Turkey, 1999), ANL preprint cond-mat/9911245.
R. Vrijen [*et al.*]{}, Phys. Rev. A [**62**]{}, 012306 (2000).
J. Kempe, D. Bacon, D. DiVincenzo, and K. Whaley, in [*Quantum Computation and Information*]{}, edited by R. Clark [*et al.*]{} (Rinton Press, New Jersey, 2001), Vol. 1, pp. 33–55.
D. Deutsch, A. Barenco, and A. Ekert, Proc. Roy. Soc. London Ser. A [**449**]{}, 669 (1995).
S. Lloyd, Phys. Rev. Lett. [**75**]{}, 346 (1995).
J. Kempe, D. Bacon, D. Lidar, and K. Whaley, Phys. Rev. A [**63**]{}, 042307 (2001).
D. Bacon, J. Kempe, D. Lidar, and K. Whaley, Phys. Rev. Lett. [**85**]{}, 1758 (2000), ANL preprint quant-ph/9909058.
D. P. DiVincenzo [*et al.*]{}, Nature [**408**]{}, 339 (2000).
J. Vala and K. B. Whaley, Phys. Rev. A [**66**]{}, 022304 (2002).
S. Myrgren and K. Whaley, Quantum Information Processing, in press, lanl-report quant-ph/0309051.
D. Bacon, K. Brown, and K. Whaley, Supercoherent quantum bits, 2001, lANL preprint quant-ph/0012018.
Y. Makhlin, Nonlocal properties of two-qubit gates and mixed states and optimization of quantum computations, 2000, ANL preprint quant-ph/0002045.
B. Buckles and F. Petry, [*Genetic Algorithms*]{} (IEEE Computer Society Press, Los Alamitos, CA, 1992).
J. C. Lagarias, J. A. Reeds, M. H. Wright, and P. E. Wright, SIAM J. Optim. [**9**]{}, 113 (1995).
J. Nelder and R. Mead, Computer Journal [**7**]{}, 308 (1965).
B. Kane, Nature [**393**]{}, 133 (1998).
B. Kane, Silicon-based Quantum Computation, 2000, ANL preprint quant-ph/0003031.
The Makhlin Invariants {#App:1}
======================
We give a brief description of how to calculate the [*Makhlin invariants*]{} [@Makhlin:00a] for an encoded two-qubit operation $W$. These invariants characterize a two-qubit operation up to equivalence by local unitaries (see Eq. (\[locequivalent\])).
In a first step project the physical operator $W$ onto the logical subspace: $$M = P^{\dag} U_{cnot}^{exchange} P.,\label{E:WtoM}$$ P is a 256-by-4 matrix whose column vectors are the basis states {$|0_{L}0_{L}\rangle$, $|0_{L}1_{L}\rangle$, $|1_{L}0_{L}\rangle$, $|1_{L}1_{L}\rangle$}, and $M$ is a matrix in $SU(4)$. We next transform $M$ into the “Bell-basis” as $M_{B} =Q^\dagger M Q$, where $$Q = \frac{1}{\sqrt{2}}\left( \begin{array}{cccc} 1 & 0 & 0 & i\\
0 & i & 1 & 0\\0 & i & -1 & 0\\ 1 & 0 & 0 & -i\\
\end{array} \right). \label{E:BellTrans}$$ Finally, we define $m = M_{B}^{T} M_{B}$, to obtain the invariants $M_{1} = tr^{2}(m)/16 det M$ and $M_{2}= (tr^{2}(m) -
tr (m^{2}))/4 det M$. For gates that are locally equivalent to the $CNOT$, $M_{1} = 0$ and $M_{2} = 1$.
Numerical search for a gate locally equivalent to $CNOT$ {#App:2}
========================================================
To obtain a gate $U_{cnot}^{exchange}$ which is locally equivalent to the encoded $CNOT$, we applied a combination of genetic algorithms and Nelder-Mead simplex searches. At the beginning of every search, we fixed a sequence of qubit pairs to be coupled with an exchange interaction, and optimized the fitness function $\mathcal{F}$ with respect to time parameters only. No restrictions on symmetry were imposed, unlike [@DiVincenzo:00a] (see Appendix \[App:3\]). We started with a small number of couplings and incremented the number of exchange interactions after each unsuccesful attempt to find a gate equivalent to the $CNOT$. The final layout of the exchanges is indicated in Figure 3. We found that space generated by $\mathcal{F}$ was sufficiently complex such that allowing the sequence of qubit-pairs to vary during the optimization only introduced unnecessary complications into the search.
Even with the incorporation of the leakage parameter $\Lambda$ into fitness function $\mathcal{F} = f + \Lambda$, the large space of parameters is still marked with many local minima. Therefore, the first stage of our search was a *genetic algorithm*, whose heuristic is well-equipped to score large spaces aggressively in order to identify basins in which a global minimum may occur. Whereas algorithms based on the hill-descent heuristic often trap themselves into basins of local minima, genetic algorithms are able to traverse rapidly through regions of the space between the basins, enabling them to descend from one basin to another. The pseudocode for the genetic algorithm is as follows:
**** Let the initial population consisting of 60 *candidates* be defined as the set $\textbf{P}_{t}$. Each member of $\textbf{P}_{t}$ is a 34-dimensional real-valued vector whose elements lie in the interval \[0, 2$\pi$\]. Each vector represents the *genome* of a candidate, and the $j^{th}$ element in each vector ( a [*gene*]{}) is the time parameter for the Hamiltonian in the $j^{th}$ exchange in a fixed sequence of qubit-pair exchanges.
**** Generate the Makhlin invariants $M_{1}, M_{2}$ and leakage parameter $\Lambda$ for each candidate and rank the candidates according to their fitness scores $\mathcal{F}$. Sort the top $20$ performing candidates into the *parental pool*.
**** Randomly pair the 20 members of the parental pool. Each parental pair generates two offspring. The first offspring is a random, pairwise convex combination of the genomes of the parental pair. Let $\alpha$ and $\beta$ be random real variables in the interval $[0,1]$. For parents *u* and *v*, we have: $$\begin{aligned}
\label{E:ConvexCO}
PARENT_{u} &=& (\gamma_{u,1},\dots, \gamma_{u,34}) \nonumber\\
PARENT_{v} &=& (\gamma_{v,1},\dots, \gamma_{v,34}) \nonumber\\
OFFSPRING1_{u,v} &=& (\alpha\gamma_{u,1} + (1-\alpha)\gamma_{v,1},
\dots, \alpha\gamma_{u,34} + (1-\alpha)\gamma_{v,34})\end{aligned}$$ The second offspring is a random geometric average of the genomes of the parental pair. $$\begin{aligned}
\label{E:GeoCO}
PARENT_{u} &=& (\gamma_{u,1},\dots, \gamma_{u,34}) \nonumber\\
PARENT_{v} &=& (\gamma_{v,1},\dots, \gamma_{v,34}) \nonumber\\
OFFSPRING2_{u,v} &=& ({\gamma_{u,1}})^{\beta} ({\gamma_{v,1}})^{1 -
\beta}, \dots, ({\gamma_{u,34}})^{\beta} ({\gamma_{v,34}})^{1 -
\beta}\end{aligned}$$ Intuitively, each candidate in the population represents a point on a simplex within the search space. By taking convex combinations and geometric averages between the points, we search the face planes of the simplex.
**** We now construct the population of the next generation $\textbf{P}_{t+1}$. The new population consists of:
1. The top (20 + *M*) candidates from $\textbf{P}_{t}$, where *M* is a randomly generated integer between $0$ and $20$
2. The 20 offspring generated from the crossover step
3. ($20$ - *M*) new, randomly-generated candidates
The purpose in inserting new candidates during each generation is to enable the search to extract itself from local minima. If the search simplex has converged to a local minimum, the new candidates will serve as vertex points that can pull the search into more promising regions within the space.
**** We now subject the population to a random mutation process, where each gene (component of $\textbf{P}_{t}$) in each genome is perturbed to a new value within $[0, 2 \pi]$ with probability $.03$. It is necessary to introduce these mutations, corresponding to small steps in the search simplex, because the cross-over operations tend to pass over global minima too rapidly. However, even small perturbations in the genome cause increasingly violent movements in $\mathcal{F}$ as global minima are approached. So to balance these considerations, the top ten performers in each generation are exempted from mutations to stabilize the performance of the algorithm.
**** Check if the top-ranked candidate satisfies the condition $\mathcal{F} < \epsilon$ for a sufficiently small $\epsilon$. If not, return to Step 2.
We ran four simultaneous genetic algorithms with four distinct, randomly-generated populations of size $60$, and coordinated the search by inserting a clone of the top candidate from the population with the best top-performer into the other three populations. After 2394 generations, we obtained a candidate with error magnitudes of $O(10^{-2})$ with respect to the Makhlin invariants and $O(10^{-1})$ with respect to the leakage parameter. At this point, the pace of progress in the genetic algorithm slowed down dramatically, so we used the top performer as the starting point for the second stage of our algorithm, a Nelder-Mead simplex direct search. At this point, the simplex search was considerably more robust, because the simplex heuristic enables the simplex to flex and squeeze itself through narrow valleys of the space more sensitively. After 5296 iterations of the simplex search, we obtained a candidate with error magnitudes of $O(10^{-6})$ with respect to the Makhlin invariants and $O(10^{-2})$ with respect to the leakage parameter. Since this was the first time we had advanced to such a low point in the space, we ran the genetic algorithm again to see if any further improvement could be obtained in this manner, and to acquire some further intuition about the structure of the space. After 471 generations, only a trivial improvement was obtained, so we returned to the simplex method once more. After 22081 iterations of the simplex search, we obtained error magnitudes of $O(10^{-10})$ with respect to the Makhlin invariants and $O(10^{-8})$ with respect to the leakage parameter.
Gate sequences for $CNOT$ with the three-qubit encoding {#App:3}
=======================================================
In Ref. [@DiVincenzo:00a], a numerical search was utilized to generate a $CNOT$ gate from a sequence of nearest-neighbor exchange interactions on a system of two computational qubits, each encoded by three physical spin-1/2 qubits. The three-qubit encoding is $$\begin{aligned}
\label{E:NatureEncoding}
|0_{L} \rangle &= \frac{1}{2}|S \rangle \otimes | 1 \rangle \nonumber\\
|1_{L} \rangle &= \sqrt{\frac{2}{3}}|T_{+} \rangle \otimes |0
\rangle - \frac{1}{\sqrt{3}}|T_{0} \rangle \otimes | 1 \rangle,\end{aligned}$$ where $|S \rangle = \frac{1}{\sqrt{2}}\left( |10 \rangle - | 01
\rangle \right)$, $|T_{+} \rangle = |11 \rangle$ and $|T_{0}
\rangle = \frac{1}{\sqrt{2}}\left( |10 \rangle + | 01 \rangle
\right)$. The numerical search was made by minimization of a similar fitness function to that employed here, [*i.e.*]{}, including both matrix distance from $CNOT$ and leakage penalty functions. A candidate sequence of $19$ exchanges equivalent to $CNOT$ up to local transformations on the encoded basis ($U_{cnot}^{exchange}$) was obtained for a layout containing two simplifying symmetries in the exchange gate times. These symmetries are illustrated in Figure 5, with the gate switching times given in Table 3. Here we give also the exchange-only implementation of the local gates $U_i,V_i,$ $i=1,2$ (cf. Eq. (\[locequivalent\])), needed to convert the $19$-gate sequence into the exact $CNOT$ in the computational basis. These were obtained using the same procedures as described in Section \[sec:CNOT\] and Ref. [@Myrgren:03a].
\[fig:appendix1\]
(0,220)
\[Table:another\]
------------- ------- ------- ----------- ---------------- ------- ------- -----------
Exchange Qubit Qubit Exchange Qubit Qubit
Time 1 2 Times Time 1 2 Times
$t_1$ 5 6 0.863060 $t_8$ 2 3 1.302881
$t_2$ 4 5 0.303496 $t_9$ 3 4 0.463869
$t_3$ 5 6 0.863060 $t_{10}$ 2 3 2.554511
$t_4$ 3 4 1.290877 $t_6$ 4 5 0.871873
$t_5$ 2 3 0.650655 $\bar{t_{10}}$ 1 2 1.249644
$t_6$ 4 5 0.871873 $\bar{t_6}$ 5 6 -1.034121
$\bar{t_5}$ 1 2 -1.207108 $t_{10}$ 2 3 2.554511
$\bar{t_6}$ 5 6 -1.034121 $t_6$ 4 5 0.871873
$t_5$ 2 3 0.650655 $t_4$ 3 4 1.290877
$t_6$ 4 5 0.871873 $t_{11}$ 1 2 0.612497
$t_7$ 3 4 2.012205 $t_{12}$ 5 6 2.826113
$t_8$ 2 3 1.302881 $t_{13}$ 4 5 2.838096
$\bar{t_8}$ 1 2 -0.502098 $t_{14}$ 5 6 2.278532
------------- ------- ------- ----------- ---------------- ------- ------- -----------
: Gate switching times for the $26$ gate sequence of Figure 5, given in units of $2 \hbar/J$. This corresponds to $\pi$ times the time units employed in Ref. [@DiVincenzo:00a].
.1cm
The first symmetry is that certain gate times are repeated in a spatially symmetric configuration, [*e.g.*]{}, $t_4$ occurs at the beginning and end of the sequence between physical qubits $3$ and $4$ in each case. We denote this the *repetition symmetry*. A second symmetry apparent in Figure 5 is that certain sequential pairs of interaction times are related by analytic functions. In particular, for $k=5,6,8,10$, the functions $$\label{E:TbarRelation}
c_{k} = \textup{tan}(t_{k})\textup{tan}(\bar{t}_{k}) + 2$$ are exactly equal to zero. We denote this symmetry a *correlation symmetry*. With these symmetry restrictions the optimal solution for a $19$ exchange gate sequence obtained in Ref [@DiVincenzo:00a] yields a Makhlin invariant fitness value of $f$ on the order of $10^{-10}$ and a leakage parameter $\Lambda$ on the order of $10^{-8}$. The overall precision of the sequence, given by the maximum matrix element distance to the $CNOT$ in the computational basis is of order $10^{-6}$.
We have performed a new set of numerical searches for this same $19$-exchange gate layout in the three-qubit encoding, using the techniques described in this paper and without imposing any symmetry constraints on the exchange times. We find that not only can higher quality numerical solutions be obtained, but also that these are very significantly improved when the simplifying symmetry constraints are lifted. This results in a larger set of independent variables ($19$ instead of the $7$ given in Ref. [@DiVincenzo:00a]) but with the advantage of a considerably smaller value of the cost function.
We performed a search with a Nelder-Mead simplex algorithm whose optimization criterion was the minimization of the Makhlin fitness function $f$ and the leakage parameter $\Lambda$, starting with a random initial set of exchange gate times. Figure 6 and Table 4 shows the corresponding exchange sequence, together with the exchange-only implementation of the local gates $U_i, V_i$, $i=1,2$ needed to convert the $19$-gate sequence into the exact $CNOT$ in the computational basis. This search yielded zero values of Makhlin fitness function $f$ and leakage parameter $\Lambda$ to within machine precision ($10^{-16}$), and a maximum matrix element distance from the exact $CNOT$ of $10^{-9}$. This provides a significant improvement over the corresponding overall precision of $10^{-6}$ for the original $19$-gate sequence of Ref. [@DiVincenzo:00a].
\[fig:appendix2\]
(0,240)
\[Table:2\]
---------- ------- ------- ---------- ---------- ------- ------- ----------
Exchange Qubit Qubit Exchange Qubit Qubit
Time 1 2 Times Time 1 2 Time
$t_1$ 2 3 3.141592 $t_{17}$ 2 3 4.444461
$t_2$ 1 2 0.989737 $t_{18}$ 3 4 0.463873
$t_3$ 2 3 3.141593 $t_{19}$ 2 3 1.249608
$t_4$ 5 6 2.477807 $t_{20}$ 4 5 5.249065
$t_5$ 4 5 0.303496 $t_{21}$ 1 2 2.554454
$t_6$ 5 6 0.863060 $t_{22}$ 5 6 4.013466
$t_7$ 3 4 4.432470 $t_{23}$ 2 3 4.391200
$t_8$ 2 3 3.792238 $t_{24}$ 4 5 2.107472
$t_9$ 4 5 2.107472 $t_{25}$ 3 4 1.290877
$t_{10}$ 1 2 5.076069 $t_{26}$ 2 3 3.141592
$t_{11}$ 5 6 0.871873 $t_{27}$ 1 2 0.927636
$t_{12}$ 2 3 3.792237 $t_{28}$ 2 3 3.141592
$t_{13}$ 4 5 5.249065 $t_{29}$ 5 6 0.863060
$t_{14}$ 3 4 5.153789 $t_{30}$ 4 5 0.303496
$t_{15}$ 2 3 1.302870 $t_{31}$ 5 6 0.466283
$t_{16}$ 1 2 5.781068
---------- ------- ------- ---------- ---------- ------- ------- ----------
: Gate switching times for the $31$ gate sequence of Figure 6, given in units of $2 \hbar/ J$.
.1cm
In the solution described in [@DiVincenzo:00a], the correlation symmetries were satisfied to machine precision, while the repetition symmetries satisfied exactly. We note that these correlation and repetition symmetries are not essential to the task of implementing the encoded $CNOT$ operation. From an optimization perspective, they might even be interpreted as a hindrance that constrains the trajectory of numerical searches to lie in sub-optimal subspaces of the control parameter space. It is not clear whether the symmetries in the solution obtained in Ref. [@DiVincenzo:00a] suggest the existence of analytical solutions to the optimization problem of the cost function, and if so whether these correspond to local or global minima. The above example shows that, without any symmetry restrictions and allowing the number of independent time parameters to increase, improvement to optimization of the cost function to within machine prevision can be obtained.
[^1]: We use [*efficient*]{} in the computational sense, meaning that we can implement the transformation with a number of elementary gates polynomial in the number of qubits. Note that not all general unitary transformations can be implemented efficiently; in fact the generic unitary transformation on $n$ qubits requires an exponential amount of elementary gates. Our usage of [*efficient*]{} here means that given there is a sequence of one- and two-qubit gates that generates $U$ then we can approximate this $U$ to arbitrary accuracy with a sequence of gates drawn from our elementary set and such that we only have polynomial overhead in the number of gates used. Further, to double the precision we only need a constant amount of additional gates. This is the notion we need to define efficient computation.
[^2]: Note that it has been shown that a [*generic*]{} two-qubit interaction alone generates universal computation [@Deutsch:95a; @Lloyd:95a]. However, by an irony of nature most implementable interactions in current quantum computation schemes happen to fall in the set of exceptions to this. These exceptions include the ubiquitous exchange interaction (both isotropic and anisotropic) and several other interactions that exhibit a certain amount of symmetry, which makes them [*non-generic*]{} in the above sense. Even for interactions that fall into the category of being universal by themselves, explicit gate-constructions have to be found in a case by case basis.
[^3]: Note that exchange operations [*within*]{} an array of $4$ qubits keep the one-qubit code space invariant and do not leak. Leakage can only occur when we couple two arrays (see [@Kempe:01a] for details).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A class of discrete-time nonlinear positive time-delay switched systems with sector-type nonlinearities is studied. Sufficient conditions for the existence of common and switched diagonal Lyapunov–Krasovskii functionals for this system class are derived; these are expressed as feasibility conditions for systems of linear algebraic inequalities. Corresponding spectral conditions for the existence of common L–K functionals are also described. Furthermore, it is shown that the proposed approaches can be applied to discrete-time models of digital filters and neural networks. Finally, a numerical example is given to illustrate the effectiveness of theoretical results.'
author:
- 'Alexander Aleksandrov [^1], Oliver Mason[^2]'
title: 'Diagonal stability of a class of discrete-time positive switched systems with delay'
---
Submitted to IET Control Theory and Applications
Introduction {#sec1}
============
Positive systems, for which nonnegative initial conditions give rise to nonnegative trajectories [@FarRin; @ValCol], are of significant practical importance due to their role in various applications including: population dynamics; consensus problems; and congestion control in transmission control protocol networks [@ShoTCP]. While the fundamental properties of positive linear time-invariant (LTI) systems are now well studied and understood [@FarRin], there is a practical need to extend this theory to more complex, realistic models. In particular, issues such as nonlinearities, time-delay and switches in system dynamics [@ValCol] give rise to novel problems, many of which are not yet resolved. The issue of time-delay is of particular importance for applications to networked control systems [@ZhaoLiuRees09]. In the current paper, we are concerned with the fundamental question of stability for a class of nonlinear, switched positive systems subject to time-delay. The model class considered is in discrete time.
The stability theory of positive systems has several aspects that distinguish it from the corresponding theory for general systems. Two of the most significant of these are arguably the use of copositive Lyapunov functions [@MasSho07] and diagonal Lyapunov functions [@Kaz; @Lam; @AlexMas] in the stability analysis of positive systems. It is well known, and follows from classical Perron-Frobenius theory [@HJ], that the stability of a positive LTI system is equivalent to: i) the existence of a linear copositive Lyapunov function; and ii) the existence of a diagonal Lyapunov function. These simple facts have motivated researchers to investigate the existence of corresponding types of Lyapunov functions and Lyapunov–Krasovskii (L–K) functionals for systems subject to time-delay [@Lam; @AlexMas], switching [@Short; @Past; @Huang] and nonlinearities [@AiT; @Shaker]. Our work here continues in this vein, focusing on the existence of common diagonal and switched diagonal L–K functionals for a class of switched positive systems with time-delay. Specifically, we shall first present conditions for the existence of common diagonal L–K functionals for the system class that are less conservative than those given in [@AlexMas]. We shall also describe results on the existence of switched diagonal L–K functionals that give conditions in terms of simple algebraic inequalities in contrast to the LMI conditions given in [@Shaker] where undelayed systems subject to a related but narrower class of nonlinearities were considered.
There are several reasons for considering the question of diagonal stability for this system class. First of all, diagonal functionals are attractive because of their simple structure in which individual states are decoupled. In many cases, the existence of a diagonal Lyapunov function implies a more robust form of stability, ensuring that the system is stable even when subject to parameter uncertainty [@Kaz]. From a practical point of view, common and switched diagonal Lyapunov functions can be used to stabilize switched linear systems [@ZhenFen12]. Moreover, such functions have proven to be powerful tools in the analysis of neural networks, asynchronous computation and in the so-called large scale systems approach [@Kaz]. Finally, on a theoretical level, it is of course interesting to understand the degree to which properties of basic LTI systems can be extended to more complex system classes and results characterising the existence of diagonal L–K functionals allow us to understand the, somewhat intricate, relationship between the various types of Lyapunov function available. In this context, the interesting work of [@Past] on the links between linear, max-type and diagonal Lyapunov functions for switched positive linear systems is noteworthy; in fact, the work of this latter paper has inspired some of the results to be presented here.
Contributions {#sec:con}
-------------
The main contributions of this paper are described below.
- We describe less conservative conditions for the existence of common diagonal L–K functionals for the considered classes of subsystems than those described in [@AlexMas]. Equivalent spectral conditions are also described. An explicit numerical example to highlight the reduction in conservatism is given.
- We provide constructive, readily verifiable conditions for the existence of common and switched L–K functionals. Furthermore, while sufficient conditions in terms of LMIs have been given previously [@Shaker], our conditions are formulated in terms of linear algebraic inequalities, and can be solved using linear programming. Links to the spectral radii of matrices associated with the systems are also described.
- We extend the result of [@Past], which applied to linear switched positive delay-free systems, to a class of nonlinear switched positive systems with delay. In addition, compared with [@Past], we derive conditions for the existence of not only common diagonal L–K functionals, but also switched diagonal L–K functionals.
Notation and background
=======================
Throughout the paper $\mathbb{R}^n$ and $\mathbb{R}^{n \times n}$ denote the vector spaces of $n$-tuples of real numbers and of $n \times n$ matrices respectively. The notation $\|\cdot\|$ refers to the Euclidean vector norm. For vectors $v \in \mathbb{R}^n$, $v \geq 0$ ($v \leq 0$) means $v_i \geq 0$ ($v_i \leq 0$) for $1 \leq i\leq n$, and $v \gg 0 $ ($v \ll 0 $) means $v_i > 0$ ($v_i < 0$) for $1 \leq i \leq n$. We use the notation $A^T$ for the transpose of a matrix $A$ and $P \succ 0$ ($P \prec 0$) to denote that the matrix $P$ is positive (negative) definite. Let $\textrm{diag}\left(c_1,\ldots,c_n\right)$ indicate a diagonal matrix with the elements $c_1,\ldots,c_n$ along the main diagonal. The identity matrix is denoted by $I$; the dimension will be clear in context. We say that a matrix $A \in \mathbb{R}^{n \times n}$ is nonnegative if all of its entries are nonnegative. The *spectral radius* of $A \in \mathbb{R}^{n \times n}$ is denoted by $\rho(A)$. A matrix $A$ is Schur if all of its eigenvalues have modulus strictly less than 1 ($\rho(A) < 1$); $A$ is Hurwitz if all of its eigenvalues have negative real parts. A matrix $A$ is Metzler if its off-diagonal entries are all nonnegative.
The following known facts about nonnegative and Metzler matrices are useful for our later results, see [@HJ; @Kaz].
\[prop:Schur\] Let $A \in \mathbb{R}^{n\times n}$ be nonnegative. Then
- $A$ is Schur if and only if there exists some $v \gg 0$ with $A v \ll v$;
- if $A\theta \ll \theta$ and $A^T d \ll d$, where $\theta\gg 0$, $d\gg 0$, then defining $P = \textrm{diag}\left(d_1/\theta_1,\ldots,d_n/\theta_n\right)$, we obtain $A^TPA - P \prec 0$.
\[prop:Metzl\] Let $A \in \mathbb{R}^{n\times n}$ be Metzler. Then
- $A$ is Hurwitz if and only if there exists some $v \gg 0$ with $A v \ll 0$;
- if $A\theta \ll 0$ and $A^T d \ll 0$, where $\theta\gg 0$, $d\gg 0$, then defining $P = \textrm{diag}\left(d_1/\theta_1,\ldots,d_n/\theta_n\right)$, we obtain $A^TP+PA \prec 0$.
Our main results in this paper are expressed in terms of systems of algebraic linear inequalities, which are related to the existence of linear copositive Lyapunov functions or the so-called S-property for matrix sets; for background on these and related topics, see [@Knorn; @ForVal10; @Doan; @Gowda]. We here recall some notation and terminology related to these questions.
Given a finite set of matrices, $M$, the *row selection set* $\mathcal{R}(M)$ of $M$ consists of all matrices formed from the elements of $M$ in the following manner. The $i$th row of each matrix in $\mathcal{R}(M)$ is the $i$th row of some matrix in $M$ for each $i$, $1 \leq i \leq n$. Given a set of Metzler matrices $M$, there exists a $v \gg 0$ with $A v \ll 0$ for all $A \in M$ if and only if every matrix in $\mathcal{R}(M)$ is Hurwitz. The following result is a simple consequence of this fact and has been proven in a distinct manner using *Collatz-Wielandt* sets in [@Doan].
\[thmcoplin\] Let $M$ be a finite set of nonnegative matrices in $\mathbb{R}^{n \times n}_+$. There exists a vector $v \gg 0$ such that $Av \ll v$ for all $A \in M$ if and only if every matrix in $\mathcal{R}(M)$ is Schur.
Statement of the problem
========================
Consider the switched system $$\label{eq:sys1}
x(k+1)=A_{\sigma(k)} f(x(k))+B_{\sigma(k)} f(x(k-l))$$ and the corresponding family of subsystems
$$\label{eq:sys2}
x(k+1)=A_{s} f(x(k))+B_{s} f(x(k-l)), \quad s=1,\ldots,N.$$
Here $x(k)\in \mathbb{R}^n$; $A_1, \ldots, A_N$, $B_{1}, \ldots, B_{N}$ are constant matrices; the nonlinearity $f: \mathbb{R}^n \rightarrow \mathbb{R}^n$ is continuous and diagonal, meaning: $
f(x) = (f_1(x_1),\ldots, f_n(x_n))^T.
$ Furthermore, we assume that each $f_i$ satisfies the following conditions: $$\label{eq:f1}
x_if_i(x_i)>0 \ \ \ \text{for}\ \ x_i\neq 0,$$
$$\label{eq:f2}
|f_i(x_i)|\leq |x_i|.$$
Finally, $l$ is a positive integer delay. The switching signal $\sigma$ maps the nonnegative integers into $\{1, \ldots, N\}$ and selects which constituent subsystem is active at each time $k$.
Systems of this class are closely related to continuous time systems of Persidskii type [@Kaz] and are motivated by numerous applications, such as control systems, digital filters, neural networks and iterative numerical methods; see, for example, [@Kaz; @Phi; @Sontag; @Liao; @EM].
In what follows we will assume that the matrices $A_1, \ldots, A_N$, $B_{1}, \ldots, B_{N}$ are nonnegative. Under this assumption, defines a switched positive time-delay system in discrete time, meaning that if initial conditions $x(-l), \ldots, x(0)$ are nonnegative, then $x(k) \geq 0$ for all $k \geq 0$. For notational convenience, we write $x^{(k)}$ for the augmented state vector $x^{(k)} = \left
( x^T(k), x^T(k-1), \ldots , x^T(k-l)
\right
)^T$ in $\mathbb{R}^{(l+1)n}$.
We will derive conditions for the existence of a common diagonal L–K functional of the form $$\begin{aligned}
\nonumber
V(x^{(k)})&=&x^T(k)Px(k)+f^T(x(k-1))Q_1f(x(k-1)) \\
\label{eq:V1}
&+&\ldots+f^T(x(k-l))
Q_l f(x(k-l)) \end{aligned}$$ for the family . Here $P,Q_1,\ldots,Q_l$ are positive definite diagonal matrices. If the family admits such a functional whose differences, $V(x^{(k+1)}) - V(x^{(k)})$ along trajectories of all constituent subsystems are negative, then we say that the system is *diagonally stable*.
If the system is diagonally stable, then its zero solution is asymptotically stable for any nonlinearities $f_1(x_1),\ldots,f_n(x_n)$ and for any switching law.
The problem of diagonal stability of is equivalent to the feasibility of the associated system of LMIs. This problem has been well investigated, see, for instance, [@Kam; @Boid], and various numerical schemes are available to determine LMI feasibility. However, it should be noted that for systems with parametric uncertainty and for situations where understanding the relationship between the existence of certain types of Lyapunov function and the dynamical properties of the underlying system, the determination of simple analytic criteria is crucial.
In [@AlexMas], sufficient conditions for the diagonal stability of the system were derived; these conditions were formulated in terms of the feasibility of auxiliary systems of linear algebraic inequalities. Formally, the following theorem was proved in [@AlexMas].
\[thm1\] Let there exist vectors $d \gg 0$, $\theta \gg 0$ such that $$\label{eq:d1}
(A_{s}+B_{r})^T d \ll d, \qquad s,r=1,\ldots ,N,$$
$$\label{eq:theta1}
(A_{s}+B_{s})\theta \ll \theta, \qquad s=1,\ldots ,N.$$
Then the system is diagonally stable.
\[rem1a\] Using Theorem \[thmcoplin\], it is possible to state the conditions given in Theorem \[thm1\] in terms of the spectral radii of sets of matrices associated with the system. First of all, denote by $M_1$ the set of all matrices of the form $(A_s + B_r)^T$ for $1 \leq s, r \leq N$; similarly, denote by $M_2$ the set of all matrices $A_s+B_s$ for $1 \leq s \leq N$. Now letting $\mathcal{R}(M_1)$ and $\mathcal{R}(M_2)$ be the sets of all [row selections]{} from $M_1$, $M_2$ respectively, the theorem can be reformulated as follows.
\[thm1a\] Let $\rho_1 = \max\{\rho(A) \mid A \in \mathcal{R}(M_1)\}$ and $\rho_2 = \max\{\rho(A) \mid A \in \mathcal{R}(M_2)\}$. If $\rho_1 < 1$, $\rho_2 < 1$ then the system is diagonally stable.
Our first goal here is to relax the conditions for diagonal stability formulated in Theorem \[thm1\]. Furthermore, along with conditions of the existence of a common diagonal L–K functional, conditions of the existence of a switched diagonal L–K functional for system will be obtained.
Construction of a common diagonal L–K functional
================================================
In [@Past], an approach was described that allows us to relax conditions for the existence of a common quadratic diagonal Lyapunov function for a family of linear positive delay-free systems. In the present section, we apply the core idea of this approach to derive new conditions for diagonal stability of the nonlinear switched positive time-delay system .
\[thm:Swit2\] Let there exist numbers $\mu>0$, $\lambda>0$ and vectors $d \gg 0$, $\theta \gg 0$ satisfying the inequalities $$\label{eq:d2}
(A_{s}+B_{r})^T d \leq \mu d, \qquad s,r=1,\ldots ,N,$$
$$\label{eq:theta2}
(A_{s}+B_{s}) \theta \leq \lambda \theta, \qquad s=1,\ldots ,N,$$
$$\label{eq:mul}
\lambda \mu< 1.$$
Then the system is diagonally stable.
**Proof** We will show how to construct a common L–K functional for the family , given by formula . Set $P=\textrm{diag}\left(d_1/\theta_1,\ldots,d_n/\theta_n\right)$, $d_i$ and $\theta_i$ are the components of the vectors $d$ and $\theta$ respectively. We will show how to choose a positive definite diagonal $Q$ and a real number $\varepsilon>0$ such that defining $Q_m=Q+(l-m+1)\varepsilon I$, $m=1,\ldots,l$, yields a diagonal L–K functional of the form for the family of systems .
Consider the difference $\Delta V = V(x^{(k+1)}) - V(x^{(k)})$ of the functional along trajectories of the $s$-th subsystem from family for some $s$ in $\{1,\ldots,N\}$. We obtain $$\Delta V \leq W_{1s}(x^{(k)})+W_{2s}(x^{(k)}).$$ Here $$W_{1s}(x^{(k)})=
f^T(x(k))\left(A_s^T P A_s-P+ Q \right)f(x(k))$$ $$+2f^T(x(k)) A_s^T P B_sf(x(k-l))$$ $$+
f^T(x(k-l))\left(B_s^T P B_s- Q \right)f(x(k-l)),$$ $$W_{2s}(x^{(k)})=\varepsilon\left(l\|f(x(k))\|^2-
\sum_{m=1}^l\|f(x(k-m))\|^2\right).$$
The function $W_{1s}(x^{(k)})$ is a quadratic form in $f(x(k))$ and $f(x(k-l))$ defined by the matrix $$C_s = \left(\begin{array}{c c}
A_s^T P A_s-P+ Q & A_s^T P B_s\\
B_s^T P A_s & B_s^T P B_s- Q \end{array}\right).$$
It is straightforward to verify that $$C_s \left(\begin{array}{c}
\theta \\
\theta \end{array}\right)=
\left(\begin{array}{c}
A_s^T P (A_s+B_s)\theta-d+Q\theta \\
B_s^T P (A_s+B_s)\theta-Q\theta \end{array}\right)$$ $$\leq \left(\begin{array}{c}
\lambda A_s^T d -d+Q\theta \\
\lambda B_s^Td-Q\theta \end{array}\right).$$
Define $\eta=\max_{r=1,\ldots,N}\left\{B_r^Td\right\}$, where the maximum is taken componentwise. Next, choose $Q$ to be a diagonal positive definite matrix such that $Q\theta=\lambda \eta+\delta e$. Here $\delta$ is a positive parameter, and $e=(1,\ldots,1)^T$. Then the inequalities $$C_s \left(\begin{array}{c}
\theta \\
\theta \end{array}\right)\leq
\left(\begin{array}{c}
(\lambda\mu-1)d+\delta e \\
-\delta e \end{array}\right)\ll 0$$ are valid for sufficiently small values of $\delta$.
As $\theta\gg 0$ and $C_s$ is Metzler, it follows from Proposition \[prop:Metzl\] that $C_s$ is a Hurwitz matrix. Moreover, $C_s$ is a symmetric matrix. Therefore, it is negative definite. Hence, one can choose a value of the parameter $\varepsilon$ and a number $\beta>0$ such that the estimate $$\Delta V \leq -\beta\sum_{j=0}^l \|f(x(k-j))\|^2$$ holds.
\[rem2a\] Using Theorem \[thmcoplin\] and the notation defined in Remark \[rem1a\], we can reformulate the previous result and provide equivalent spectral conditions for diagonal stability. The following fact is key.
\[prop:equiv\] Let $\rho_1 = \max\{\rho(A) \mid A \in \mathcal{R}(M_1)\}$ and $\rho_2 = \max\{\rho(A) \mid A \in \mathcal{R}(M_2)\}$. There exist real numbers $\mu > 0$, $\lambda > 0$ and vectors $d \gg 0$, $\theta \gg 0$ in $\mathbb{R}^n$ satisfying – if and only if $\rho_1\rho_2 < 1$.
**Proof** First assume that $\rho_1 \rho_2 < 1$. It follows that there is some $\epsilon >0$ such that $(\rho_1 + \epsilon)(\rho_2 + \epsilon) < 1$. From the definitions of $\rho_1$, $\rho_2$, all matrices in the sets $\frac{1}{\rho_1 + \epsilon}\mathcal{R}(M_1)$, $\frac{1}{\rho_2 + \epsilon}\mathcal{R}(M_2)$ are Schur. Hence by Theorem \[thmcoplin\] there exist vectors $d \gg 0$ and $\theta \gg 0$ such that $(A_s + B_r)^Td \ll (\rho_1 + \epsilon) d$, $(A_s + B_s) \theta \ll (\rho_2 + \epsilon) \theta$ for $1 \leq s, r \leq N$, and $(\rho_1 + \epsilon)(\rho_2 + \epsilon) < 1$.
Conversely, assume that there exist real numbers $\mu > 0$, $\lambda > 0$ and vectors $d \gg 0$, $\theta \gg 0$ in $\mathbb{R}^n$ satisfying –. It follows that for any $\epsilon > 0$, $(A_s + B_r)^T d \ll (\mu+\epsilon)d$, $(A_s + B_s)\theta \ll (\lambda + \epsilon) \theta$ for $1 \leq s, r \leq N$. Theorem \[thmcoplin\] then implies that $\rho_1 < \mu+\epsilon$, $\rho_2 < \lambda + \epsilon$. As $\epsilon > 0$ is arbitrary, it follows that $\rho_1 \leq \mu$, $\rho_2 \leq \lambda$ and hence that $\rho_1 \rho_2 < 1$. This completes the proof.
The following reformulation of Theorem \[thm:Swit2\] is now immediate.
\[thm2a\] Let $\rho_1 = \max\{\rho(A) \mid A \in \mathcal{R}(M_1)\}$ and $\rho_2 = \max\{\rho(A) \mid A \in \mathcal{R}(M_2)\}$. If $\rho_1 \rho_2 < 1$ then the system is diagonally stable.
This alternative formulation makes clear how the conditions of Theorem \[thm:Swit2\] relax those given in Theorem \[thm1\]. Moreover, it gives a spectral (albeit non-constructive) method of verifying the conditions of the theorem. It also shows how Theorem \[thm:Swit2\] is a direct extension of Theorem 3 in [@Past] to the time-delayed, nonlinear system class considered here.
To verify the existence of numbers $\mu>0$, $\lambda>0$ and vectors $d \gg 0$, $\theta \gg 0$ satisfying inequalities –, one can also adapt the procedure proposed in [@Past].
We note that Theorem \[thm:Swit2\] can be extended to systems with multiple delays.
Formally, consider the family of subsystems $$\begin{aligned}
\nonumber x(k+1)=A_{s}f(x(k))+B_{1s}
f(x(k-1))\\
\label{eq:11}
+\ldots+
B_{ls}f(x(k-l)), \quad s=1,\ldots,N, \end{aligned}$$ where $A_{s},B_{1s}\ldots,B_{l s}$ are nonnegative matrices for $s=1,\ldots,N$, and the remaining notation is the same as used for .
\[thm:Swit3\] Assume that there exist numbers $\mu>0$, $\lambda>0$ and vectors $d \gg 0$, $\theta \gg 0$ such that $$(A_{s}+B_{1 r_1}+\ldots+B_{l r_l})^T d \ll \mu d, \quad s,r_1,\ldots,r_l
=1,\ldots,N,$$ $$(A_{s}+B_{1 s}+\ldots+B_{l s}) \theta \ll \lambda \theta , \quad s=1,\ldots,N,$$ and inequality holds. Then there exists a common diagonal L–K functional of the form for the family .
The proof of the theorem is essentially the same as that of Theorem \[thm:Swit2\] although the notation is a little more involved.
Construction of a switched diagonal L–K functional
==================================================
One approach to obtain less conservative stability conditions for switched systems than those obtained with the aid of common Lyapunov functions is based on the construction of switched Lyapunov functions [@Daaf; @Shaker; @Liu; @zapp]. In particular, in [@Daaf] and [@Shaker] switched quadratic Lyapunov functions were used for the stability analysis of discrete linear delay-free switched systems and discrete nonlinear delay-free switched $\Phi$-systems respectively. We will adapt this approach to construct switched diagonal L–K functionals for the system .
Specifically, we will construct a switched diagonal L–K functional of the form $$\begin{aligned}
\nonumber \widetilde V(k,x^{(k)})=x^T(k)P^{(\sigma(k))}x(k)\\
\nonumber
+f^T(x(k-1))Q_1f(x(k-1))\\
\nonumber
+f^T(x(k-2))Q_2f(x(k-2))\\
\label{eq:12}
+\ldots+f^T(x(k-l))
Q_lf(x(k-l)) \end{aligned}$$ for the time-delay system . Here $P^{(1)},\ldots,P^{(N)},Q_1,\ldots,Q_l$ are positive definite diagonal matrices. In contrast to [@Daaf; @Shaker], we will give conditions for the existence of such functionals that are based on systems of linear algebraic inequalities.
\[thm:Swit4\] Assume that there exist numbers $\mu>0$, $\lambda>0$ and vectors $d^{(1)} \gg 0,\ldots,d^{(N)}\gg 0, \theta \gg 0$ such that $$\label{eq:rsj1}
A^T_{s}d^{(r)}+B^T_{m}d^{(j)} \leq \mu d^{(s)}, \qquad s,r,m,j=1,\ldots ,N,$$ and inequalities and hold. Then there exists a switched diagonal L–K functional of the form for system .
**Proof** We write $P^{(s)}=\textrm{diag}\left(d^{(s)}_1/\theta_1,\ldots,d^{(s)}_n/\theta_n\right)$, $Q_m=\tilde Q+(l-m+1)\varepsilon I$, $m=1,\ldots,l$, $s=1,\ldots,N$, for the matrices appearing in . Here $d^{(s)}_i$ and $\theta_i$ are components of the vectors $d^{(s)}$ and $\theta$ respectively, $\tilde Q$ is a positive definite diagonal matrix, and $\varepsilon>0$ is a parameter; as in the proof of Theorem \[thm:Swit2\], we shall show how to determine the values of $\tilde Q$, $\varepsilon$.
A relatively straightforward computation shows that the difference of the functional along trajectories of satisfies the estimate $$\Delta \widetilde V \leq
f^T(x(k))\left(A_{\sigma(k)}^T P^{(\sigma(k+1))} A_{\sigma(k)} \right.$$ $$\left. -P^{(\sigma(k))}+ \tilde Q \right)f(x(k))$$ $$+2f^T(x(k)) A_{\sigma(k)}^T P^{(\sigma(k+1))} B_{\sigma(k)}f(x(k-l))$$ $$+
f^T(x(k-l))\left(B_{\sigma(k)}^T P^{(\sigma(k+1))} B_{\sigma(k)}- \tilde Q \right)f(x(k-l))$$ $$+\varepsilon\left(l\|f(x(k))\|^2-
\sum_{m=1}^l\|f(x(k-m))\|^2\right).$$
Consider the matrices $$C_{sr}=
\left(\begin{array}{c c}
A_s^T P^{(r)} A_s-P^{(s)}+ \tilde Q & A_s^T P^{(r)} B_s\\
B_s^T P^{(r)} A_s & B_s^T P^{(r)} B_s- \tilde Q \end{array}\right)$$ for $s,r=1,\ldots,N$. We obtain $$C_{sr} \left(\begin{array}{c}
\theta \\
\theta \end{array}\right)
\leq \left(\begin{array}{c}
\lambda A_s^T d^{(r)} -d^{(s)}+\tilde Q\theta \\
\lambda B_s^Td^{(r)}-\tilde Q\theta \end{array}\right).$$
Choose $\tilde Q$ to be the diagonal positive definite matrix such that $$\tilde Q \theta=\lambda \max_{s, m=1,\ldots,N}\left\{B_s^Td^{(m)}\right\}+\delta e,
%\eqno(14)$$ where the maximum is taken componentwise, and $\delta$ is a positive parameter. If the value of $\delta$ is sufficiently small, then $$C_{sr} \left(\begin{array}{c}
\theta \\
\theta \end{array}\right)\leq
\left(\begin{array}{c}
(\lambda\mu-1)d^{(s)}+\delta e \\
-\delta e \end{array}\right)\ll 0$$ for $ s,r=1,\ldots,N$.
The subsequent proof is similar to that of Theorem \[thm:Swit2\]. In the next result, we consider the case where $l=1$. Using a virtually identical argument, we establish conditions for the existence of a closely related class of diagonal L–K functional.
\[thm:Cor\] Consider system with $l=1$. Assume that there exist numbers $\mu>0$, $\lambda>0$ and vectors $d^{(1)} \gg 0,\ldots,d^{(N)}\gg 0, \theta \gg 0$ such that $$\label{eq:14}
A^T_{s}d^{(r)}+B^T_{r}d^{(j)} \leq \mu d^{(s)}, \qquad s,r,j=1,\ldots ,N,$$ and that inequalities and hold. Then there exists a switched diagonal L–K functional of the form $$\begin{aligned}
\nonumber
\widetilde V(k,x^{(k)})=x^T(k)P^{(\sigma(k))}x(k)\\
\label{eq:15}
+f^T(x(k-1))Q^{(\sigma(k))}f(x(k-1))\end{aligned}$$ for system , where $P^{(s)}$ and $Q^{(s)}$ are positive definite diagonal matrices, $s=1,\ldots,N$.
**Proof** As in the proof of Theorem \[thm:Swit4\], we compute the difference $ \Delta \widetilde V$ along the trajectories of the system. In this case the difference can be bounded by quadratic forms in $f(x(k))$, $f(x(k-1))$ described by the matrices: $$C_{sr}=
\left(\begin{array}{c c}
A_s^T P^{(r)} A_s-P^{(s)}+ Q^{(r)} & A_s^T P^{(r)} B_s\\
B_s^T P^{(r)} A_s & B_s^T P^{(r)} B_s- Q^{(s)} \end{array}\right)$$ for $s,r=1,\ldots,N$. Here, the matrices $P^{(s)}$ are defined in the same way as in Theorem \[thm:Swit4\] while the positive definite diagonal matrices $Q^{(s)}$ are given by $$Q^{(s)} \theta=\lambda \max_{m=1,\ldots,N}\left\{B_s^Td^{(m)}\right\}+\delta e,$$ for $s = 1, \ldots , N$.
In a similar way as in the proof of Theorem \[thm:Swit4\], we can see that if $\delta>0$ is sufficiently small, then $$C_{sr} \left(\begin{array}{c}
\theta \\
\theta \end{array}\right)\leq
\left(\begin{array}{c}
\lambda A_s^T d^{(r)} -d^{(s)}+ Q^{(r)}\theta \\
\lambda B_s^Td^{(r)}- Q^{(s)}\theta \end{array}\right)$$ $$\leq
\left(\begin{array}{c}
(\lambda\mu-1)d^{(s)}+\delta e \\
-\delta e \end{array}\right)\ll 0$$ for $ s,r=1,\ldots,N$.
Theorem \[thm:Swit4\] can be extended to systems with multiple delays.
Some applications of the proposed approaches
============================================
In this section, we briefly describe some potential applications of the results and approaches outlined in the previous sections.
Models of digital filters
-------------------------
Let the family of subsystems $$\label{eq:16}
x(k+1)=f\left(A_{s} x(k)+B_{s} x(k-l)\right), \ \ s=1,\ldots,N,$$ be given. As for the system , we assume that $x(k)\in \mathbb{R}^n$; $A_1, \ldots, A_N$, $B_{1}, \ldots, B_{N}$ are constant nonnegative matrices; the nonlinearity $f: \mathbb{R}^n \rightarrow \mathbb{R}^n$ is continuous and diagonal, meaning: $
f(x) = (f_1(x_1),\ldots, f_n(x_n))^T,
$ where each $f_i$ satisfies conditions and ; $l$ is a positive integer delay.
Systems of the form are used as mathematical models of digital filters, see [@Kaz; @EM].
The switched system associated with is $$\label{eq:17}
x(k+1)=f\left(A_{\sigma(k)} x(k)+B_{\sigma(k)} x(k-l)\right).$$
We will first describe how to construct a common diagonal L–K functional for family . In this case, instead of , we will choose such a functional in the form $$\begin{aligned}
\nonumber
V(x^{(k)})=x^T(k)Px(k)+x^T(k-1)Q_1x(k-1) \\
%\nonumber
\label{eq:18}
+x^T(k-2)Q_2x(k-2)
+\ldots+x^T(k-l)
Q_lx(k-l),\end{aligned}$$ where $P,Q_1,\ldots,Q_l$ are positive definite diagonal matrices.
\[thm:Digital-1\] If there exist real numbers $\mu>0$, $\lambda>0$ and vectors $d \gg 0$, $\theta \gg 0$ such that the inequalities – are valid, then there exists a common L–K functional of the form for the family .
**Proof** Consider the difference $\Delta V = V(x^{(k+1)}) - V(x^{(k)})$ of the functional along trajectories of the $s$-th subsystem from the family for some $s$ in $\{1,\ldots,N\}$. We obtain $$\Delta V =f^T\left(A_{s} x(k)+B_{s} x(k-l)\right)Pf\left(A_{s} x(k)+B_{s} x(k-l)\right)$$ $$+
x^T(k)(Q_1-P)x(k)
+x^T(k-1)(Q_2-Q_1)x(k-1)$$ $$+\ldots+x^T(k-l+1)(Q_l-Q_{l-1})x(k-l+1)$$ $$-x^T(k-l)
Q_lx(k-l).$$ From , it follows that $$\Delta V \leq \left(A_{s} x(k)+B_{s} x(k-l)\right)^T P \left(A_{s} x(k)+B_{s} x(k-l)\right)$$ $$+
x^T(k)(Q_1-P)x(k)
+x^T(k-1)(Q_2-Q_1)x(k-1)$$ $$+\ldots+x^T(k-l+1)(Q_l-Q_{l-1})x(k-l+1)$$ $$-x^T(k-l)
Q_lx(k-l).$$ Taking into account this estimate, the rest of the proof follows from arguments identical to those used in the proof of Theorem \[thm:Swit2\].
Next, we shall present conditions of the existence of a switched diagonal L–K functional of the form $$\begin{aligned}
\nonumber \widetilde V(k,x^{(k)})=x^T(k)P^{(\sigma(k))}x(k)
%\\ \nonumber
+x^T(k-1)Q_1x(k-1)
\\
%\nonumber
+x^T(k-2)Q_2 x(k-2)
\label{eq:19}
+\ldots+x^T(k-l)
Q_l x(k-l) \end{aligned}$$ for the system . Here $P^{(1)},\ldots,P^{(N)},Q_1,\ldots,Q_l$ are positive definite diagonal matrices.
It is not too difficult to adapt the arguments of Theorems 7 and 8 to obtain the following result.
\[thm:Digital-2\] Assume that there exist numbers $\mu>0$, $\lambda>0$ and vectors $d^{(1)} \gg 0,\ldots,d^{(N)}\gg 0, \theta \gg 0$ such that the inequalities , and are satisfied. Then there exists a switched diagonal L–K functional of the form for system .
Discrete-time neural networks
-----------------------------
Next, consider the switched system $$\label{eq:20}
x(k+1)=A_{\sigma(k)} f(x(k))+B_{\sigma(k)} f(x(k-l))+u(k)$$ and the associated family of subsystems $$\label{eq:21}
x(k+1)=A_{s} f(x(k))+B_{s} f(x(k-l))+u(k),$$ $$s=1,\ldots,N.$$ Here $u(k)$ is a bounded input, and the rest of the notation is the same as for the original system given in . Systems of the form arise in models of neural networks, see [@Kaz; @TG] for background on their use in this context.
In general, the family will not admit a common equilibrium position. However, the approaches outlined in this paper permit us to derive conditions under which solutions of will be ultimately bounded.
Let $x\left(k,x^{(k_0)},k_0\right)$ denote the solution of with initial conditions $k_0\geq 0$, $x^{(k_0)}\in \mathbb{R}^{(l+1)n}$.
The system is said to be uniformly ultimately bounded with the ultimate bound $R$ if, for any $D>0$, there exists a positive integer $\tilde k$ such that $\left\|x\left(k,x^{(k_0)},k_0\right)\right\| \leq R$ for all $k_0\geq 0$, $\left\|x^{(k_0)}\right\| \leq D$, $k\geq \tilde k+ k_0$ and for arbitrary switching law $\sigma(k)$.
\[thm:Networks-1\] Let $f_i(x_i)\to -\infty$ as $x_i\to -\infty$, and $f_i(x_i)\to +\infty$ as $x_i\to +\infty$, $i=1,\ldots,n$. Assume that there exist numbers $\mu>0$, $\lambda>0$ and vectors $d \gg 0$, $\theta \gg 0$ such that inequalities – are valid. Then there exists a common L–K functional of the form for the family guaranteeing uniform ultimate boundedness of system .
**Proof** Define matrices $P,Q_1,\ldots,Q_l$ in the same way as in the proof of Theorem 4. Consider the difference of the functional along trajectories of the $s$-th subsystem from the family for some $s$ in $\{1,\ldots,N\}$. For sufficiently small values of the parameters $\varepsilon$ and $\delta$, we obtain $$\label{eq:22}
\Delta V \leq -\beta_1\sum_{j=0}^l \|f(x(k-j))\|^2+\beta_2,$$ where $\beta_1$ and $\beta_2$ are positive constants.
Using the estimate , the remainder of the proof is very similar to the proof of Theorem 5.2 in [@Al-Diff].
\[thm:Networks-2\] Let $f_i(x_i)\to -\infty$ as $x_i\to -\infty$, and $f_i(x_i)\to +\infty$ as $x_i\to +\infty$, $i=1,\ldots,n$. Assume that there exist numbers $\mu>0$, $\lambda>0$ and vectors $d^{(1)} \gg 0,\ldots,d^{(N)}\gg 0, \theta \gg 0$ such that inequalities , and are valid. Then there exists a switched diagonal L–K functional of the form guaranteeing uniform ultimate boundedness of system .
The proof of the theorem is similar to that of Theorem \[thm:Swit4\].
Numerical example {#sec4}
=================
In this section, we present a simple numerical example in order to illustrate how our results relate to each other and to earlier work.
Consider a family consisting of two subsystems of dimension 2 ($N=2$, $n=2$) with the system matrices $$A_1 =\frac14 \left(\begin{array}{c c}
0 & 0\\
1 & 1
\end{array}\right), \qquad
B_1 = \frac14\left(\begin{array}{c c}
a & 1\\
2 & 0
\end{array}\right),$$ $$A_2 =\frac14 \left(\begin{array}{c c}
0 & 1\\
0 & 2
\end{array}\right), \qquad
B_2 = \frac14\left(\begin{array}{c c}
1 & 1\\
0 & 0
\end{array}\right),$$ where $a$ is a positive parameter. In addition, assume that $l=1$.
In this case, the system of inequalities admits a positive solution if and only if $a<1$, whereas admits a positive solution if and only if $a<3$. Thus, we can apply Theorem 2 only for $a<1$.
Next, let $a=2$. For this value of the parameter $a$, admits a positive solution if and only if $\lambda\geq (1+\sqrt{6})/4$, whereas system admits a positive solution if and only if $\mu\geq (2+\sqrt{6})/4$. Verifying condition , we obtain $$\frac{2+\sqrt{6}}4 \cdot \frac{1+\sqrt{6}}4 <1.$$ Thus, Theorem 4 guarantees the existence of a common L–K functional of the form .
Finally, choose $a=9/4$. Then system admits a positive solution if and only if $\lambda\geq (13+\sqrt{217})/32$, and system admits a positive solution if and only if $\mu\geq (17+\sqrt{433})/32$.
Since $$\frac{13+\sqrt{217}}{32} \cdot \frac{17+\sqrt{433}}{32} >1,$$ condition is not fulfilled. Hence, for this value of parameter $a$,we can not apply Theorem 4.
On the other hand, if we consider system for $a=9/4$, $\mu=1.152$, then the system admits the solution $d^{(1)}=(1.179, 0.5)^T$, $d^{(2)}=(1.3, 1)^T$. We obtain $$\frac{13+\sqrt{217}}{32} \cdot 1.152<1.$$ Hence, all the conditions of Proposition 4 are fulfilled, and we can guarantee the existence of a switched diagonal L–K functional of the form for the associated switched system.
Conclusion {#sec5}
==========
In this short paper, we have described sufficient conditions for the existence of common and switched diagonal Lyapunov functionals for switched positive nonlinear systems subject to time-delay. In particular, the result of Theorem \[thm:Swit2\] relaxes the condition required in [@AlexMas] for the existence of a common diagonal Lyapunov functional, thereby giving a less conservative stability criterion. It also provides an extension of Theorem 3 of [@Past] to nonlinear time-delayed systems. The result of Theorem \[thm:Swit4\] describes conditions for a switched diagonal Lyapunov functional to exist for the same class of nonlinear switched systems, and gives an alternative type of condition to the LMIs described in [@Shaker].
Acknowledgments
===============
This work was partially supported by the Government of the Russian Federation (Grant no. 074-U01), the Russian Foundation for Basic Research (Grant no. 16-01-00587), the Science Foundation Ireland (Grant no. 13/RC/2094), and the European Regional Development Fund through the Southern & Eastern Regional Operational Programme to Lero – the Irish Software Research Centre (www.lero.ie).
[99]{}
Blanchini F., Colaneri P., Valcher M.E.: ‘Switched positive linear systems’, *Foundations and Trends in Systems and Control*, 2015, **2**, no. 2, pp. 101–273
Farina L., Rinaldi S.: ‘Positive linear systems: theory and applications’ (Wiley, New York, 2000)
Shorten R.N., Wirth F., Leith D.: ‘A positive systems model of TCP-like congestion control’, *IEEE Trans. Networking*, 2006, **14**, no. 3, pp. 616–629
Zhao Y.B., Liu G.P., Rees D.: ‘Stability and stabilisation of discrete-time networked control systems: a new time delay system approach’, *IET Control Theory Appl.*, 2009, **4**, no. 9, pp. 1859–1866
Mason O., Shorten R.N.: ‘On linear copositive Lyapunov functions and the stability of switched positive linear systems’, *IEEE Trans. Aut. Cont.*, 2007, **52**, no. 7, pp. 1346–1349
Kaszkurewicz E., Bhaya A.: ‘Matrix diagonal stability in systems and computation’ (Birkhäuser, Boston, Basel, Berlin, 2000)
Wu L., Lam J., Shu Z., et al.: ‘On stability and stabilizability of positive delay systems’, *Asian Journal of Control*, 2009, **11**, pp. 226–234.
Aleksandrov A., Mason O.: ‘Diagonal Lyapunov–Krasovskii functionals for discrete-time positive systems with delay’, *Syst. Control Lett.*, 2014, **63**, pp. 63–67
Horn R.A., Johnson C.R.: ‘Matrix analysis’ (Cambridge University Press, New York, 1985)
Shorten R., Wirth F., Mason O., et al.: ‘Stability criteria for switched and hybrid systems’, *SIAM Rev.*, 2007, **49**, no. 4, pp. 545–592
Pastravanu O.C., Matcovschi M.-H.: ‘Max-type copositive Lyapunov functions for switching positive linear systems’, *Automatica*, 2014, **50**, pp. 3323–3327
Zhang J., Huang J., Zhao X.: ‘Further results on stability and stabilisation of switched positive systems’, *IET Control Theory Appl.*, 2015, **9**, no. 14, pp. 2132–2139
Aleksandrov A., Platonov A.: ‘On absolute stability of one class of nonlinear switched systems’, *Autom. Remote Control*, 2008, **69**, no. 7, pp. 1101–1116
Shaker H.R., How J.P.: ’Stability analysis for class of switched nonlinear systems’. Proc. American Control Conf., Marriott Waterfront, Baltimore, MD, USA, June 30-July 02 2010, pp. 2517–2520
Zheng Y., Feng G.: ‘Diagonal stabilisation of a class of single-input discrete-time switched systems’, *IET Control Theory Appl.*, 2012, **7**, no. 4, pp. 515–522
Knorn F., Mason O., Shorten R.N.: ‘On linear co-positive Lyapunov functions for sets of linear positive systems’, *Automatica*, 2009, **45**, pp. 1943–1947
Fornasini E., Valcher M.E..: ‘Linear copositive functions for continuous time positive switched systems’, *IEEE Trans. Aut. Cont.*, 2010, **55**, no. 8, pp. 1933–1937
Doan T.S., Kalauch A., Siegmund S.: ‘A constructive approach to linear Lyapunov functions for positive switched systems using Collatz-Wielandt sets’, *IEEE Trans. Aut. Cont.*, 2013, **58**, no. 3, pp. 748–751
Song Y., Seetharama-Gowda M., Ravindran G.: ‘On some properties of P-matrix sets’, *Lin. Alg. and Appl.*, 1999, **290**, pp. 237–246
Sun X., Li J., Zhao J.: ‘Stabilization for a class of discrete-time switched $\Phi$-systems’, *Circuits, Systems, and Signal Processing*, 2017, **36**, pp. 834–844
Sontag E.D., Karny M., Warwick K., et al.: ‘Recurrent neural networks: some systems-theoretic aspects in dealing with complexity: a neural network approach’ (Springer, London, 1997)
Liao X., Yu P.: ‘Absolute stability of nonlinear control systems’ (Springer, New York, Heidelberg, 2008)
Erickson K.T., Michel A.N.: ‘Stability analysis of fixed-point digital filters using computer generated Lyapunov functions – Part I: Direct form and coupled form filters’, *IEEE Trans. on Circuits and Systems*, 1985, **32**, pp. 113–132
Kamenetskiy V.A., Pyatnitskiy Ye.S.: ‘An iterative method of Lyapunov function construction for differential inclusions’, *Syst. Control Lett.*, 1987, **8**, pp. 445–451
Boyd S., Ghaoui E., Feron E., et al.: ‘Linear matrix inequalities in system and control theory’ (SIAM, Philadelphia, 1994)
Liu X.: ‘Stability analysis of switched positive systems: a switched linear copositive Lyapunov function method’, *IEEE Trans. Circuits and Systems* (II), 2009, **56**, pp. 414–418
Daafous J., Riedinger P., Iung C.: ‘Stability analysis and control synthesis for switched systems: a switched Lyapunov function approach’, *IEEE Trans. Autom. Control*, 2002, **47**, pp. 1883–1887
Zappavigna A., Colaneri P., Geromel J.C., et al.: ’Dwell time analysis for continuous-time switched linear positive systems’. Proc. American Control Conf., Marriott Waterfront, Baltimore, MD, USA, June 30-July 02 2010, pp. 6256–6261
Takeda M., Goodman J.W.: ‘Neural networks for computation: number representations and programming complexity’, *Applied Optics*, 1986, **25**, pp. 3033–3046
Aleksandrov A., Chen Y., Platonov A., et al.: ‘Stability analysis and uniform ultimate boundedness control synthesis for a class of nonlinear switched difference systems’, *J. Difference Equ. Appl.*, 2012, **18**, pp. 1545–1561
[^1]: Faculty of Applied Mathematics and Control Processes, Saint Petersburg State University, Saint Petersburg, 199034, Russia, and Department of Control of Complex Systems, ITMO University, Saint Petersburg, 197101, Russia. email: [email protected]
[^2]: Dept. of Mathematics and Statistics/Hamilton Institute, Maynooth University, Maynooth, Co. Kildare, Ireland and Lero, the Irish Software Research Centre. email: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'I show how to construct Monte Carlo algorithms (programs), prove that they are correct and document them. Complicated algorithms are build using a handful of elementary methods. This construction process is transparently illustrated using graphical representation in which complicated graphs consist of only several elementary building blocks. In particular I discuss the equivalent algorithms, that is different MC algorithms, with different arrangements of the elementary building blocks, which generate [*the same*]{} final probability distribution. I also show how to transform a given MC algorithm into another equivalent one and discuss advantages of the various “architectures”.'
---
[Version of\
]{}
[**Practical Guide to Monte Carlo** ]{}
[**S. Jadach**]{}\
[*Institute of Nuclear Physics, ul. Kawiory 26a, Kraków, Poland*]{}\
[ and]{}\
[*DESY, Theory Group, Notkestrasse 85, D-22603 Hamburg, Germany*]{}\
[*To be submitted somewhere, sometime (or may be not)*]{}
Introduction
============
The aim of this report is to provide:
- Elementary description of elementary Monte Carlo (MC) methods for a graduate student in physics, who is supposed to learn them in a couple of days,
- Methods of transparent documenting of the MC programs,
- Reference in publications where there is no space for description of the elementary MC methodology.
In my opinion there is certain gap in the published literature on the MC methods. The elementary MC methods like rejection according to weight, branching (multichannel method) or mapping of variables are so simple and intuitive that it seems to be not worth to write anything on them. On the other hand in the practical MC applications these methods are often combined in such a complicated and baroque way that sometimes one may wonder if the author is really controlling what he is doing, especially if the documentation is incomplete/sparse and we lack commonly accepted terminology graphical notation for describing MC algorithms. There are also many mathematically oriented articles and textbooks on the MC methods which in my opinion seem to have very little connection with the practical every day work of someone constructing MC program. The aim of this report is to fill at least partly this gap. This report is extension of a section in ref. [@jadach:1985]. I would like also to recommend the classical report [@James:1980] of James on the elementary MC methods.
Section 1 describes elementary MC methods of the single and multiple level rejection (reweighting), including detailed description of the weight book-keeping and recipes for keeping correct normalisation for the total integrand and the differential distributions (histograms). Section 2 introduces branching (multi-channel) method and section 3 demonstrates the simplest combinations of the rejection and branching. Section 4 and 5 reviews more advanced aspects of combining rejection and branching, in particular I show examples of “equivalent” algorithms, i.e. different algorithm which provide the same distributions, pointing out advantages of certain arrangements of the the rejection and branching. Another common method of the variable mapping is discussed in section 5, again in the context of various arrangement of the rejection and branching methods.
Rejection, compensating weights
===============================
We intend to generate randomly [*events*]{}, that is points $x=(x_1,x_2,x_3,...,x_n)$, according to a distribution $$\rho(x_i) = { d^n\sigma \over d x^n }(x_i)$$ within certain domain $\Omega$ and, simultaneously, we want to calculate (estimate) the integral $$\sigma = \int\limits_\Omega \rho(x_i)\; dx^n.$$ as precisely as possible. In our notation, change of integration variables induces in $\rho$-density a [*Jacobian*]{} factor[^1] $$\rho(y_i) = \left| { \partial x \over \partial y } \right|\; \rho(x_i).$$ The normalised to unity probability density is simply $$d^n p = {1\over \sigma } d^n \sigma,\qquad
\int\limits_\Omega d^n p =
\int\limits_\Omega {d^n p \over dx^n}\; dx^n =1.$$
(1400,850) ( 400,800)[(0,0)\[t\]]{} ( 400,850)
(0,0)\[t\]
( 600,850) (300,700)[(0,-1)[250]{}]{} (300,450)[(0,0)\[t\][(150,100)[$\rho^{(1)}(x_i)$]{}]{}]{} (300, 350)( 300, 300) (300,300)
(0,0)\[t\]
( 600,200) (300, 200)[(0,-1) [75]{}]{} (300, 75) (300, 75)[(0,0)\[t\]]{} (300, 25)( 300, 0) (350, 75)[( 1, 0)[175]{}]{} (525, 75)[( 0, 1)[375]{}]{} (525,450)[(-1, 0)[225]{}]{}
(300,100)[(0,-1)[100]{}]{} (100,100)[(450,475)[ ]{}]{} (300,300)[(0,0)\[l\]]{} (300, 70)[(0,0)\[l\]]{}
(1000,800)[(0,0)\[t\]]{} (1000,850)
(0,0)\[t\]
( 600,850) (300,700)[(0,-1)[250]{}]{} (300,450)[(0,0)\[t\][(150,100)[$\rho^{(1)}(x_i)$]{}]{}]{} (300, 350)( 300, 300) (300,300)
(0,0)\[t\]
( 600,200) (300, 200)[(0,-1) [75]{}]{} (300, 75) (300, 75)[(0,0)\[t\]]{} (300, 25)( 300, 0)
(300,100)[(0,-1)[100]{}]{} (300,300)[(0,0)\[l\]]{} (300, 70)[(0,0)\[l\]]{}
In fig. \[fig:simple1\](a) I show the simple single-line algorithm with rejection according to a [*weight*]{}[^2] defined as a ratio of an [*exact*]{} distribution $\rho$ to [*approximate*]{} $\rho^{(1)}$ $$w(x) = { \rho(x_i) \over \rho^{(1)}(x_i)}
= { d^n \sigma \over d^n \sigma^{(1)} }.$$ We assume that we are able to generate randomly events according to $\rho^{(1)}$ and we know the numerical value of the integral $$\sigma^{(1)} = \int\limits_\Omega \rho(x_i)^{(1)}\; dx^n.$$ The box represents part of the algorithm (part of computer code) which provides us events according to $\rho^{(1)}$ and the value of $\sigma^{(1)}$. The content of the box can be a complicated algorithm (computer code) and we treat it as a “black box”, i.e. we may know nothing about its content. In particular it can be taken from ready-to-use library of the programs generating standard distributions, or even a physical process providing “random events”. The circle with the return line depicts [*rejection method*]{}. For each event leaving we calculate weight $w$ and we accept event (downward arrow) if $$r W < w,$$ where W is a [*maximum weight*]{} and $r$ is a [*uniform random number*]{} $ 0<r<1.$ Otherwise event is rejected (return arrow in the graph). It is easy to see that events exiting our algorithm are generated according to density $\rho$. Probability density of accepted events at the point $x_i$ is equal to product of probability $d^n p^{(1)}$ of events produced in the box $\rho^{(1)}$ times the probability $p_{accept}=w(x)/W$ of accepting an event $$d^n p(x) = {\cal N} d^n p^{(1)}\; p_{accept} =
{\cal N} { d^n \sigma^{(1)}(x) \over \sigma^{(1)} }\;
{ w(x) \over W},$$ where ${\cal N}$ is normalisation factor. Substituting the definition of the weight $w(x)=d^n\sigma/d^n\sigma^{(1)}$ and imposing normalisation condition $\int d^n p(x) =1$ we find $$d^n p(x) = { {\cal N} \over W}\; { d^n \sigma(x) \over \sigma^{(1)} },
\qquad
{\cal N} = { W \sigma^{(1)} \over \sigma}$$ and as a result we obtain $$d^n p(x) = { d^n \sigma(x) \over \sigma }$$ as desired. The dashed box $\rho(x_i)$ can be used as part in a bigger algorithm (box in a bigger graph) because it provides events generated according to density $\rho(x_i)$. The question is whether within the dashed box we are able to estimate the integral $\sigma$. In fact we can, and there are even two ways to do it. In the first method we use the ratio on accepted events $N$ to the total number $N^{(1)}$ of the events generated in the box . The number of accepted events is, on the average, proportional to probability of generating an event event $d^n \sigma^{(1)}/\sigma^{(1)}$ times probability of accepting an event $$\bar{w}={ w(x) \over W},\qquad$$ averaged all over the points $x_i$ in the entire integration (generation) domain $\Omega$ $$\label{naccept}
N = N^{(1)} \int\limits_\Omega
{d^n \sigma^{(1)} \over \sigma^{(1)}}\; \bar{w}
= N^{(1)} { \sigma \over \bar{\sigma}^{(1)} },\quad
\bar{\sigma}^{(1)}= W\;\sigma^{(1)}.$$ The above relation can be used to calculate the unknown integral $\sigma$ as follows $$\label{norm-nacc}
\sigma = \bar{\sigma}^{(1)}\; { N \over N^{(1)} }.$$ using known $\sigma^{(1)}$ and counting accepted events $N$. Of course, the error of the above estimator of $\sigma$ is given by the usual statistical error from the binomial distribution. In the second method we calculate the average weight where the averaging is done over all accepted and rejected events $$<w> = \int\limits_\Omega
{d^n \sigma^{(1)} \over \sigma^{(1)}}\; w(x)
= { \sigma \over \sigma^{(1)} }.$$ The above gives us second equivalent estimator of the unknown integral $\sigma$ in terms of the known $\sigma^{(1)}$ and the measured average weight $$\label{norm-avewt}
\sigma = \sigma^{(1)}\; <w> = \bar{\sigma}^{(1)} \; <\bar{w}>.$$ Another often asked question is: how to calculate the integral $\Delta\sigma$ over a [*subdomain*]{} $\Delta\Omega$, which is for instance a single bin in a histogram? The following formula can be easily derived $$\Delta\sigma = \sigma\; {\Delta N \over N}
= \bar{\sigma}^{(1)} {\Delta N \over N^{(1)}}$$ where $\Delta N$ is number of events falling into subdomain $\Delta\Omega$. A particular case is the proper normalisation of the histogram. Let us take the one dimensional distribution $${d\sigma \over d z} = \int\limits_\Omega d^n\sigma\; \delta(z - z(x_i))$$ which we estimate/calculate by means of collecting generated events in a histogram with $n_b$ equal bins within a $(z_{\min},z_{\max})$ range. The relevant formula reads $${d\sigma \over d z}
\simeq {\sigma \Delta N \over \Delta z N }
= { n_b \sigma \Delta N \over (z_{\max}-z_{\min}) N}
= { n_b \bar{\sigma}^{(1)} \Delta N \over (z_{\max}-z_{\min}) N^{(1)} }$$ In fig. \[fig:simple1\](b) I show the same algorithm for variable-weight events. In this case we do not reject events but we associate the weight $w$ with each event. For the total integral over entire $\Omega$ I may use the same formula of eq. (\[norm-avewt\]) as for the constant-weight algorithm. In the case of the histogram we accumulate in each bin a sum of weights $\sum\limits_{z\in bin} w$. The properly normalised distribution is obtained as follows[^3] $${d\sigma \over d z}
= { n_b \sigma^{(1)} \over (z_{\max}-z_{\min}) N^{(1)} }
\sum\limits_{z\in bin} w
= { n_b \bar{\sigma}^{(1)} \over (z_{\max}-z_{\min}) N^{(1)} }
\sum\limits_{z\in bin} \bar{w}$$
(1400,1200) ( 400,1200)[(0,0)\[t\]]{} ( 400,1100)
(0,0)\[t\]
( 600,1100) (175,550)[(325,375)[ ]{}]{} (400,550)[(0,0)\[t\]]{} (100,300)[(450,675)[ ]{}]{} (450,300)[(0,0)\[t\]]{} ( 25,100)[(575,950)[ ]{}]{} (500,100)[(0,0)\[t\]]{} (300,1100)[(0,-1)[250]{}]{} (300,850)[(0,0)\[t\][(130,100)[$\rho^{(n)}(x_i)$]{}]{}]{} (300,750)
(0,0)\[t\]
( 600,200) (300, 200)[(0,-1) [75]{}]{} (300, 75) (300, 95)[(0,0)\[t\]]{} (300, 25)( 300, 0) (350, 75)[( 1, 0)[125]{}]{} (475, 75)[( 0, 1)[275]{}]{} (475,350)[(-1, 0)[175]{}]{}
(300,562)[(0,0)\[t\]]{} (300,500)
(0,0)\[t\]
( 600,200) (300, 200)[(0,-1) [75]{}]{} (300, 75) (300, 95)[(0,0)\[t\]]{} (300, 25)( 300, 0) (350, 75)[( 1, 0)[175]{}]{} (525, 75)[( 0, 1)[575]{}]{} (525,650)[(-1, 0)[225]{}]{}
(300,300)
(0,0)\[t\]
( 600,200) (300, 200)[(0,-1) [75]{}]{} (300, 75) (300, 95)[(0,0)\[t\]]{} (300, 25)( 300, 0) (350, 75)[( 1, 0)[225]{}]{} (575, 75)[( 0, 1)[825]{}]{} (575,900)[(-1, 0)[275]{}]{}
(300,100)[(0,-1)[100]{}]{}
(1000,1200)[(0,0)\[t\]]{} (1000,1100)
(0,0)\[t\]
( 600,1100) (300,1100)[(0,-1)[250]{}]{} (300,850)[(0,0)\[t\][(130,100)[$\rho^{(n)}(x_i)$]{}]{}]{} (300,750)
(0,0)\[t\]
( 600,200) (300, 200)[(0,-1) [75]{}]{} (300, 75) (300, 95)[(0,0)\[t\]]{} (300, 25)( 300, 0)
(300,562)[(0,0)\[t\]]{} (300,500)
(0,0)\[t\]
( 600,200) (300, 200)[(0,-1) [75]{}]{} (300, 75) (300, 95)[(0,0)\[t\]]{} (300, 25)( 300, 0)
(300,300)
(0,0)\[t\]
( 600,200) (300, 200)[(0,-1) [75]{}]{} (300, 75) (300, 95)[(0,0)\[t\]]{} (300, 25)( 300, 0)
(300,100)[(0,-1)[100]{}]{}
In fig. \[fig:single-line\](a) I show the simple single-line algorithm with several nested rejection loops. The meaning of the graph is rather obvious. The original distribution $\rho_0$ goes through $n$-step simplification procedure $$\rho^{(0)} \to \rho^{(1)} \to \rho^{(2)} \dots \to \rho^{(n)}$$ and the compensation weights $$w^{(k)} = { \rho^{(k)} \over \rho^{(k-1)}}$$ are used for rejections “locally” in a standard way: each weight $w^{(k)}$ is compared with $r W^{(k)}$ where $0<r<1$ is uniform random number, and if $w^{(k)}< r W^{(k)}$ the event is accepted (down-ward arrow), otherwise rejected (return loop). The average weights $<w^{(k)}>$ are calculated for each rejection loop. The most inward box represents generation algorithm of the points $x_i$ according to maximally simplified (crude) distribution $\rho^{(n)}$ for which we know the integral $\sigma^{(n)}=\int \rho^{(n)}$ analytically. The integral of the original distribution $\rho$ is obtained from the crude integral and the average weights $$\sigma^{(0)} = \int \rho(x_i)\; dx^n
= \sigma^{(n)}\; \prod\limits_{i=1}^n <w^{(i-1)}>
= \bar{\sigma}^{(n)}\; \prod\limits_{i=1}^n <\bar{w}^{(i-1)}>$$ The above is completely standard and can be found in ref. [@jadach:1985]. Note also that all $n$ rejection loops may be combined into single rejection loop with the weight being product of all weights along the line $$w = \prod\limits_{i=1}^n w^{(i-1)}.$$ Usually, the version with nested loops is more efficient and the corresponding program is more modular. The weights for the internal loops are related to more technical aspects of the MC algorithm (Jacobians) and do not evolve quickly (during the development of the program) while external weights correspond to physics model and may change more frequently. It is therefore profitable to keep in practice several levels of the weights.
Finally, we may decide to perform calculation for weighted events. In fig. \[fig:single-line\](b) I show the version of the simple single-line MC algorithm with variable-weight events. The event at the exit of the graph gets associated weight $w$ which is the product of all weights along the line.
Branching {#sect:branching}
=========
(1200,800) (600,800)
(0,0)\[t\]
( 600,800) ( 0,150)[(625,600)[ ]{}]{} (500,150)[(0,0)\[t\]]{} (300,800)
(0,0)\[t\]
( 600,200) (300,200)[(0,-1)[175]{}]{} (300, 00) (280, 15)[(0,0)\[r\]]{} (330, 15)[(0,0)\[l\]]{}
(300,600)
(0,0)\[t\]
( 600,100) (300,100)(100,0) (300,100)(300,0) (300,100)(500,0)
(300,500)
(0,0)\[t\]
( 600,100) (100,100)[(0,0)\[t\][(130,100)[$\rho_1(x_i)$]{}]{}]{} (300,100)[(0,0)\[t\][(130,100)[$\rho_2(x_i)$]{}]{}]{} (500,100)[(0,0)\[t\][(130,100)[$\rho_n(x_i)$]{}]{}]{} (400, 50)[(0,0)\[t\]]{}
(300,400)
(0,0)\[t\]
( 600,50) (100,50)[(0,-1)[50]{}]{} (300,50)[(0,-1)[50]{}]{} (500,50)[(0,-1)[50]{}]{}
(300,350)
(0,0)\[t\]
( 600,100) (100,100)(300,0) (300,100)(300,0) (500,100)(300,0)
(300,250)[(0,-1)[250]{}]{}
In fig. \[fig:branching\] I show the general MC algorithm with branching into $n$ branches. This kind of algorithm is used when the distribution to be generated $\rho$ can be split into sum of several distinct (positive) subdistributions $$\rho(x_i) = \sum_{i=1}^n \rho_i$$ and we are able, one way or another, to generate each $\rho_i$ separately. Usually each $\rho_i$ contains single peak or one class of peaks in the distribution $\rho$. In the beginning of the algorithm (black disc) we pick randomly one branch (subdistribution) according to probability $$P_k = { \int \rho_k \over \sum\limits_i \int \rho_i }
= {\sigma_k \over \sigma},$$ i.e. we have to know [*in advance*]{} the integrals $\sigma_i= \int \rho_i$ analytically or numerically. Typically, in each branch one uses different integration variables $x_i$ to parameterise the integral. The particular choice of variables will be adjusted to leading “singularities” in the branch distribution.
Let us give a formal proof of the correctness of the branching method. $$d^n p(x) = \sum\limits_k P_k d^n p_k(x)
= \sum\limits_k {\sigma_k \over \sigma}\;
{d^n \sigma_k(x) \over \sigma_k}
= {1\over \sigma} \sum\limits_k d^n \sigma_k(x)
= {d^n \sigma(x) \over \sigma}$$ Finally, note that at the exit of the branched MC algorithm (exit of the graph in fig. \[fig:branching\]), we may be forced for various reason (saving computer memory), trash all information on the origin of an event, consequently, we may be not able to use any information specific to the branch from which the event has came. This is rather important practical aspect to be kept in mind.
Branching and internal rejection
================================
(1400,1100) (300,1100)[(0,0)\[t\]]{} (300,1000)
(0,0)\[t\]
( 600,1100) ( 15,200)[(600,850)[ ]{}]{} (500,200)[(0,0)\[t\]]{} (300,1100)
(0,0)\[t\]
( 600,200) (300,200)[(0,-1)[175]{}]{} (300, 00) (280, 15)[(0,0)\[r\]]{} (330, 15)[(0,0)\[l\]]{}
(300,900)
(0,0)\[t\]
( 600,100) (300,100)(100,0) (300,100)(300,0) (300,100)(500,0)
(300,800)
(0,0)\[t\]
( 600,150) (100,150)[(0,-1)[150]{}]{} (300,150)[(0,-1)[150]{}]{} (500,150)[(0,-1)[150]{}]{}
(300,650)
(0,0)\[t\]
( 600,100) (100,100)[(0,0)\[t\][(130,100)[$\rho^{(1)}_1(x_i)$]{}]{}]{} (300,100)[(0,0)\[t\][(130,100)[$\rho^{(1)}_2(x_i)$]{}]{}]{} (500,100)[(0,0)\[t\][(130,100)[$\rho^{(1)}_n(x_i)$]{}]{}]{} (400, 50)[(0,0)\[t\]]{}
(300,550)
(0,0)\[t\]
( 600,150) (100, 50)[(0,0)\[t\]]{} (100, 50) (150, 50)[( 1, 0)[ 50]{}]{} (200, 50)[( 0, 1)[275]{}]{} (200,325)[(-1, 0)[100]{}]{} (300, 50)[(0,0)\[t\]]{} (300, 50) (350, 50)[( 1, 0)[ 50]{}]{} (400, 50)[( 0, 1)[275]{}]{} (400,325)[(-1, 0)[100]{}]{} (500, 50)[(0,0)\[t\]]{} (500, 50) (550, 50)[( 1, 0)[ 50]{}]{} (600, 50)[( 0, 1)[275]{}]{} (600,325)[(-1, 0)[100]{}]{} (100,150)[( 0,-1)[50]{}]{} (300,150)[( 0,-1)[50]{}]{} (500,150)[( 0,-1)[50]{}]{}
(300,400)
(0,0)\[t\]
( 600,50) (100,50)[(0,-1)[50]{}]{} (300,50)[(0,-1)[50]{}]{} (500,50)[(0,-1)[50]{}]{}
(300,350)
(0,0)\[t\]
( 600,100) (100,100)(300,0) (300,100)(300,0) (500,100)(300,0)
(300,250)[(0,-1)[150]{}]{}
(1000,1100)[(0,0)\[t\]]{} (1000,1000)
(0,0)\[t\]
( 600,1100) ( 15,200)[(600,850)[ ]{}]{} (500,200)[(0,0)\[t\]]{} (300,1100)
(0,0)\[t\]
( 600,200) (300,200)[(0,-1)[175]{}]{} (300, 00) (280, 15)[(0,0)\[r\]]{} (330, 15)[(0,0)\[l\]]{}
(300,900)
(0,0)\[t\]
( 600,100) (300,100)(100,0) (300,100)(300,0) (300,100)(500,0)
(300,800)
(0,0)\[t\]
( 600,100) (100,100)[(0,0)\[t\][(130,100)[$\rho^{(1)}_1(x_i)$]{}]{}]{} (300,100)[(0,0)\[t\][(130,100)[$\rho^{(1)}_2(x_i)$]{}]{}]{} (500,100)[(0,0)\[t\][(130,100)[$\rho^{(1)}_n(x_i)$]{}]{}]{} (400, 50)[(0,0)\[t\]]{}
(300,700)
(0,0)\[t\]
( 600,300) (100, 50)[(0,0)\[t\]]{} (100, 50) (150, 50)[(1,0)[450]{}]{} (300,125)[(0,0)\[t\]]{} (300,125) (350,120)[(1,0)[250]{}]{} (500,200)[(0,0)\[t\]]{} (500,200) (550,200)[(1,0)[ 50]{}]{} (600, 50)[( 0, 1)[550]{}]{} (600,600)[(-1, 0)[300]{}]{} (100,300)[( 0,-1)[200]{}]{} (300,300)[( 0,-1)[125]{}]{} (500,300)[( 0,-1)[ 50]{}]{} (300, 75)[( 0,-1)[ 75]{}]{} (500,150)[( 0,-1)[150]{}]{}
(300,400)
(0,0)\[t\]
( 600,50) (100,50)[(0,-1)[50]{}]{} (300,50)[(0,-1)[50]{}]{} (500,50)[(0,-1)[50]{}]{}
(300,350)
(0,0)\[t\]
( 600,100) (100,100)(300,0) (300,100)(300,0) (500,100)(300,0)
(300,250)[(0,-1)[150]{}]{}
In fig. \[fig:bran-wt\](a) I show the simplest combination of the branching algorithm with the standard rejection method. This type of the algorithm is potentially very efficient but is used not so often in the MC event generators because it requires that we know in advance branching probabilities $P_k$. In most of situations we do not know them analytically. In principle, they can be calculated numerically using $\sigma_k=\bar{\sigma}^{(1)}_k <\bar{w}_k>$ but this is not very handy because it requires two MC runs – first run which determines average weights $<\bar{w}_k>$ an second run with $P_k$ calculated from $<\bar{w}_k>$ from the first run.
The solution to the above problem is the variant of the algorithm presented in fig. \[fig:bran-wt\](b) where $$\bar{P}_k = {\bar{\sigma}^{(1)}_k \over \bar{\sigma}^{(1)} }
= { \int \rho^{(1)}_k \over \sum\limits_l \int \rho^{(1)}_l }$$ and all rejection returns are done to a point before the branching point. Let us check correctness of the algorithm in fig. \[fig:bran-wt\](b). The probability density $d^n p(x_i)$ at the point $x_i$ at the exit of the algorithm (graph) is proportional to product of probability of getting event in the box equal $d^n p^{(1)}_k(x) = d^n \sigma^{(1)}_k(x) /\sigma_k^{(1)}$ times probability of accepting event being $\bar{w}_k(x)= d^n \sigma_k(x)/d^n \bar{\sigma}^{(1)}_k(x)$, all that averaged over branches with probabilities $\bar{P}_k$. The same statement is expressed mathematically as follows: $$\begin{split}
d^n p(x_i)
&= {\cal N} \sum\limits_k
\bar{P}_k\; d^n p^{(1)}_k(x)\; \bar{w}_k(x)
\\&
= {\cal N} \sum\limits_k
{\bar{\sigma}_k^{(1)} \over \bar{\sigma}^{(1)} }\;
{d^n \sigma^{(1)}_k(x) \over \sigma_k^{(1)} }\;
{d^n \sigma_k(x) \over d^n \bar{\sigma}^{(1)}_k(x) }
= {\cal N} { d^n\sigma(x) \over \bar{\sigma}^{(1)} }.
\end{split}$$ Normalisation ${\cal N}= \bar{\sigma}^{(1)}/\sigma$ is determined from the condition $\int_\Omega d^n p(x_i) =1$. Finally we obtain $d^n p(x_i) = d^n\sigma(x_i) / \sigma$, as expected.
The total integral is a sum over integrals from all branches $\sigma=\sum_k \sigma_k$ where for each branch we may use the formula $\sigma_k = \sigma^{(1)}_k <w_k>$ (see eq. (\[norm-avewt\])). Slightly rearranged formula $$\sigma = \sum_k \sigma^{(1)}_k <w_k>
= \bar{\sigma}^{(1)} \sum_k \bar{P}_k <\bar{w}_k>
= \bar{\sigma}^{(1)} <\bar{w}>,$$ where in $<\bar{w}>$ we average also over branches, is a straightforward generalisation of eq. (\[norm-avewt\]). We can also generalised formula of eq. (\[norm-nacc\]) for the total integral based on the number of accepted events $$\label{norm-bran-nacc}
\sigma = \bar{\sigma}^{(1)} { N \over N^{(1)} }
= \bar{\sigma}^{(1)} { \sum_k N_k \over \sum_k N^{(1)}_k }$$ [*Proof*]{}: number $N$ of events accepted in all branches is $$N = \sum_k N_k = \sum_k N^{(1)}_k { \sigma_k \over \bar{\sigma}^{(1)}_k }$$ where $N^{(1)}_k$ is total number of events in a given branch (that is before rejection), see also eq. (\[naccept\]). Inserting $N^{(1)}_k= N^{(1)} \bar{P}_k$ we get $N = \sigma / \bar{\sigma}^{(1)}$ and therefore eq. (\[norm-bran-nacc\]).
Summarising, we see that the two algorithms in fig. \[fig:bran-wt\] are equivalent, ie. they provide the same distribution of events and the same total integral. The algorithm (a) is probably slightly more efficient but also more difficult to realize because it requires precise knowledge of the branching probabilities $P_k$. The algorithm (b) is usually less efficient but the branching probabilities $\bar{P}^{(1)}_k$ are easier to evaluate because they correspond to [*simplified*]{} distributions $\rho^{(1)}_k$.
Let us now consider the case of variable-weight events for which all return loops in fig. \[fig:bran-wt\] are removed and the two cases (a) and (b) are identical. The event at the exit of the algorithm carries the weight from one of the branches! For the calculation of the total integral we may use the same formulas of eqs. (\[norm-bran-nacc\]) and (\[norm-avewt\]) as for the constant-weight method. Let us check whether we may proceed as usual for the calculation of the integrals in the subdomain $\Delta\Omega$ being single bin in any kind of the differential (or multi-differential) distribution. Let us generate long series of $N$ weighted events and accumulate sum of the weights which fall into $\Delta\Omega$. Of course, in the sum of accumulated weights we have contributions from all branches $$\sum\limits_{x_i \in \Delta\Omega} \bar{w}(x_i)
= \sum\limits_k \int\limits_{\Delta\Omega} d^n N_k\; \bar{w}_k
= \sum\limits_k N \bar{P}^{(1)}_k
\int\limits_{\Delta\Omega} d^n p^{(1)}_k\; \bar{w}_k$$ Substituting definitions for $\bar{P}^{(1)}_k$ and of $d^n p^{(1)}_k$ we get $$\begin{split}
\sum\limits_{x_i \in \Delta\Omega} \bar{w}(x_i)
&= N \sum\limits_k
{ \bar{\sigma}^{(1)}_k \over \bar{\sigma}^{(1)} }\;
\int\limits_{\Delta\Omega}
{ d^n \bar{\sigma}^{(1)}_k(x) \over \bar{\sigma}^{(1)}_k }\;
{ d^n \sigma_k(x) \over d^n \bar{\sigma}^{(1)}_k(x) }
\\
&= {N \over \bar{\sigma}^{(1)} }\;
\sum\limits_k \int\limits_{\Delta\Omega} d^n \sigma_k(x)
= {N \over \bar{\sigma}^{(1)} }\; \Delta\sigma
\end{split}$$ Reverting the above formula we get an estimate of the integrated or (multi-) differential distribution in terms of sum of the weights $$\Delta\sigma \equiv \int\limits_{\Delta\Omega} d^n \sigma_k
= { \bar{\sigma}^{(1)} \over N }
\sum\limits_{x_i \in \Delta\Omega} \bar{w}(x_i)
= { \sigma^{(1)} \over N }
\sum\limits_{x_i \in \Delta\Omega} w(x_i)$$ Let us finally discuss the role of the maximum weight $W_k$ and the apparently unnecessary complication of keeping two kinds of crude distributions $\sigma^{(1)}$ and $\bar{\sigma}^{(1)}$. For variable-weight events without branching, $W$ is merely a scale factor which cancels out completely among $<\bar{w}>$ and $\bar{\sigma}^{(1)}$ in the overall normalisation. Its only role is to keep weights in certain preferred range, for example it is often preferred to have weights of order 1. In the case of the variable-weights with branching the relative values of $W_k$ start to play certain role. Although, for infinite number of events, final results (distributions and integrals) do not depend on $W_k$ the efficiency (convergence) of the calculation depends on the relative ratios of $W_k$ [@Kleiss:1994qy]. The maximum weights $W_k$ are more important/useful for constant-weight algorithm. They are chosen in such a way that $\bar{w}<1$. The rejection method does not work if this condition is not fulfilled. In most cases we do not know analytically the maximum weight for a given approximate $\rho^{(1)}$ and the maximum weight $W$ is adjusted empirically. Of course, the same adjustments can be done by scaling (multiplying by a constant) the entire $\rho^{(1)}$ but the long-standing tradition tells us to keep $\rho^{(1)}$ unchanged and rather introduce an explicit adjustment factor $W$. In the case of the constant-weight algorithm the values of $W_k$ determine the efficiency (rejection rate) in each branch. Let us stress again that it is always possible to enforce $W_k=1$ and the presence of $W_k$ is in fact pure conventional.
Branching and external rejection
================================
(1400,1100) (300,1100)[(0,0)\[t\]]{} (300,1000)
(0,0)\[t\]
( 600,1100) ( 15,150)[(600,900)[ ]{}]{} (500,150)[(0,0)\[t\]]{} (300,1100)
(0,0)\[t\]
( 600,200) (300,200)[(0,-1)[175]{}]{} (300, 00) (280, 15)[(0,0)\[r\]]{} (330, 15)[(0,0)\[l\]]{}
(300,900)
(0,0)\[t\]
( 600,100) (300,100)(100,0) (300,100)(300,0) (300,100)(500,0)
(300,800)
(0,0)\[t\]
( 600,100) (100,100)[(0,0)\[t\][(130,100)[$\rho^{(1)}_1(x_i)$]{}]{}]{} (300,100)[(0,0)\[t\][(130,100)[$\rho^{(1)}_2(x_i)$]{}]{}]{} (500,100)[(0,0)\[t\][(130,100)[$\rho^{(1)}_n(x_i)$]{}]{}]{} (400, 50)[(0,0)\[t\]]{}
(300,700)
(0,0)\[t\]
( 600,300) (100, 50)[(0,0)\[t\]]{} (100, 50) (150, 50)[(1,0)[450]{}]{} (300,125)[(0,0)\[t\]]{} (300,125) (350,120)[(1,0)[250]{}]{} (500,200)[(0,0)\[t\]]{} (500,200) (550,200)[(1,0)[ 50]{}]{} (600, 50)[( 0, 1)[550]{}]{} (600,600)[(-1, 0)[300]{}]{} (100,300)[( 0,-1)[200]{}]{} (300,300)[( 0,-1)[125]{}]{} (500,300)[( 0,-1)[ 50]{}]{} (300, 75)(300,0) (500,150)(500,0)
(300,400)
(0,0)\[t\]
( 600,100) (100,100)[(0,-1)[100]{}]{} (300,100)[(0,-1)[100]{}]{} (500,100)[(0,-1)[100]{}]{} (110, 60)[(0,0)\[l\]]{} (310, 60)[(0,0)\[l\]]{} (510, 60)[(0,0)\[l\]]{}
(300,300)
(0,0)\[t\]
( 600,100) (100,100)(300,0) (300,100)(300,0) (500,100)(300,0)
(300,250)[(0,-1)[150]{}]{}
(1000,900)[(0,0)\[t\]]{} (1000,800)
(0,0)\[t\]
( 600,800) ( 0,50)[(625,700)[ ]{}]{} (500,50)[(0,0)\[t\]]{} ( 15,230)[(570,410)[ ]{}]{} (500,230)[(0,0)\[t\]]{} (300,800)
(0,0)\[t\]
( 600,200) (300,200)[(0,-1)[175]{}]{} (300, 00) (280, 15)[(0,0)\[r\]]{} (330, 15)[(0,0)\[l\]]{}
(300,600)
(0,0)\[t\]
( 600,100) (300,100)(100,0) (300,100)(300,0) (300,100)(500,0)
(300,500)
(0,0)\[t\]
( 600,100) (100,100)[(0,0)\[t\][(130,100)[$\rho^{(1)}_1(x_i)$]{}]{}]{} (300,100)[(0,0)\[t\][(130,100)[$\rho^{(1)}_2(x_i)$]{}]{}]{} (500,100)[(0,0)\[t\][(130,100)[$\rho^{(1)}_n(x_i)$]{}]{}]{} (400, 50)[(0,0)\[t\]]{}
(300,400)
(0,0)\[t\]
( 600,50) (100,50)[(0,-1)[50]{}]{} (300,50)[(0,-1)[50]{}]{} (500,50)[(0,-1)[50]{}]{}
(300,350)
(0,0)\[t\]
( 600,100) (100,100)(300,0) (300,100)(300,0) (500,100)(300,0)
(300,250)
(0,0)\[t\]
( 600,250) (300,250)[(0,-1)[50]{}]{} (300,150) (300,150)[(0,0)\[t\]]{} (300,100)[( 0,-1)[100]{}]{} (350,150)[( 1, 0)[250]{}]{} (600,150)[( 0, 1)[550]{}]{} (600,700)[(-1, 0)[300]{}]{}
In fig. \[fig:bran-wt2\] we transform our algorithm one step further. In fig. \[fig:bran-wt2\](a) we repeat essentially the algorithm of fig. \[fig:bran-wt\](b) while in fig. \[fig:bran-wt2\](b) we have single rejection outside branched part. The weights $w_k$ and $w$ are related such that two algorithms are equivalent, that is both algorithms provide the same distributions and calculate the same integral. The relations is very simple $$w
= \sum\limits_k \bar{p}^{(1)}_k(x)\; w_k(x)
= \sum\limits_k { \bar{\rho}^{(1)}_k(x) \over \bar{\rho}^{(1)}(x) }\; w_k(x)
= { \rho(x) \over \rho^{(1)}(x) }
= { \sum\limits_k \rho_k(x) \over \sum\limits_k \rho^{(1)}_k(x) }.$$ The algorithm of fig. \[fig:bran-wt2\](b) can be also obtained independently by combining in a straightforward way the rejection and branching methods. We proceed as follows: first we simplify $$\rho(x) \to \rho^{(1)}(x)$$ and this simplification is compensated by the weight $$w = { \rho(x) \over \rho^{(1)}(x) }$$ and for the internal part we apply the branching method as described in section \[sect:branching\]. Consequently, we may apply all standard formulas for the calculation of the total integral, for instance $\sigma = \sigma^{(1)} <w>$, and we do not need to worry about additional proofs of the correctness of the algorithm of fig. \[fig:bran-wt2\](b); we already know that it generates properly the distribution $\rho(x_i)$.
Note that the algorithm of fig. \[fig:bran-wt2\](b) looks more general than that of fig. \[fig:bran-wt2\](a) in the following sense: the simplified distribution $\rho^{(1)}$ can be written a sum from contributions from all branches $\rho^{(1)} = \sum\limits_k \rho^{(1)}_k$ and the same is true for $\rho$ in the case (a) while, in general, it needs not be true in the case (b). In other words algorithm (a) can be transformed into (b) but the transformation in the opposite direction is less obvious. There is always a trivial transformation of (b) into (a) in which we set $w_k\equiv w$. In other words, if in the graph (a) all weights $w_k$ are the same then we are allowed to contract all rejection loop into a single one as in graph (b), and vice versa. This sounds trivial but may be useful in the case of the several levels of the compensation/rejection weights.
In the case of the variable-weight we simply omit the rejection return-loops and sum up weights of the events. Again, since fig. \[fig:bran-wt2\](b) is a direct superposition of the standard weighting and branching methods all standard rules apply. It is amusing to observe that in spite of the fact that the the weights in the two algorithm of figs. \[fig:bran-wt2\] are different, the two algorithm provide exactly the same distributions and integrals – only efficiency may differ. Which one is more convenient or efficient depends on the details of a particular problem. In the next section we shall elaborate on advantages and disadvantages of the two.
Branching, compensating weights and mapping
===========================================
Branching is a very powerful tool in the case of the distribution with many peaks. Usually, we are able to split $$\rho(x) = \sum\limits_k \rho_k(x)$$ in such a way that each $\rho_k(x)$ contains one kind of a spike in the distribution. In each branch we generate different spike with help of the dedicated change of the variables $x_i \to y_i^{(k)}$ such that in $$\rho_k(x_i) = \rho_k(y_i^{(k)})\;
\left| { \partial x \over \partial y^{(k)} } \right|$$ new distribution $\rho_k(y_i^{(k)})$ is completely flat and the whole spike is located in the Jacobian function $| \partial y^{(k)} / \partial x | $. In the following approximation $$\rho_k(x_i) \to \rho^{(1)}_k(x_i) = r^{(1)}_k\;
\left| { \partial x \over \partial y^{(k)} } \right|$$ the $\rho_k(y_i^{(k)})$ is simply replaced by the constant residue $r^{(1)}_k$. The relevant compensating weight reads as follows $$w_k = { \rho_k(x_i) \over \rho^{(1)}_k(x_i) }
= { \rho_k(x_i) \over r^{(1)}_k }\;
\left| { \partial y^{(k)} \over \partial x } \right|$$ The approximate cross sections needed for branching probabilities are $$\sigma^{(1)}_k
= \int\limits_{\Omega} d^n x \rho^{(1)}_k
= r^{(1)}_k \int\limits_{\Omega_k} d^n y^{(k)}
= r^{(1)}_k\; V(\Omega_k)$$ where $\Omega_k$ is the integration domain expressed in the variables $y^{(k)}$ and $V(\Omega_k)$ is simply the Cartesian volume of the domain $\Omega_k$ in the $y$-space.
Now comes the interesting question: Which of the two algorithms of fig. \[fig:bran-wt2\] is more convenient and/or efficient. Finally, the answer will always depend on the individual properties of a given distribution $\rho$. Nevertheless, let us point out some general advantages of the case (a). In the algorithm of fig. \[fig:bran-wt2\](a) we need a single weight $w_k$, for the $k$-th branch from which an event originates. The distribution $\rho^{(1)}_k$ might be a simple function directly expressed in terms of $x_i$ but it may happen that the Jacobian $|\partial y^{(k)}/ \partial x|$ is a more complicated function which requires the knowledge of $y^{(k)}_i$ and the whole transformation $y^{(k)}_i \to x_i$. Of course, it is not the problem for the single branch, as in the case (a), since in the process of generating an event we calculate primarily (generate randomly) the point $y^{(k)}_i$ and we transform it into $x_i$; the calculation of this Jacobian is usually a byproduct of the generation process. The situation in the algorithm of fig. \[fig:bran-wt2\](b) might be worse because in this case the global weight $$w = { \rho(x) \over \rho^{(1)}(x) }
= { \sum\limits_k \rho_k(x) \over \sum\limits_k \rho_k^{(1)}(x) }
= { \sum\limits_k \rho_k(x)
\over
\sum\limits_k r^{(1)}_k \left|{ \partial y^{(k)} \over \partial x }\right|
}
\label{wt5b}$$ contains in the denominator $\rho_k^{(1)}(x)$ (or Jacobians) for all branches. Consequently, in some cases we may be forced to perform [*for each event*]{}, (often quite complicated) transformations $x_i \to y_i^{(k)}$ and calculate Jacobians [*for all branches*]{}. This is cumbersome, especially if we have large number of branches. It may also consume a lot of computer time. Just imagine that due to permutation symmetry we have $N!$ branches – even if $N$ is some moderately high number the summation over all branches might consume almost infinite amount of computer time. The procedure of summation over branches might be also numerically instable in the case of very strong spikes in $\rho_k^{(1)}(x)$ because computer arithmetic is usually optimised for a given single spike in a given branch and it might break down for other branch, unless special labour-hungry methods are employed.
We conclude that the algorithm in fig. \[fig:bran-wt2\](a) seems to have certain advantages over the algorithm in fig. \[fig:bran-wt2\](b) although in many cases the difference might be unimportant and one might find algorithm (b) more simple (it is perhaps easier to explain and document).
(1400,1200) (700,1200)
(0,0)\[t\]
( 600,1200) (300,1200)
(0,0)\[t\]
( 600,200) (300,200)[(0,-1)[175]{}]{} (300, 00) (280, 15)[(0,0)\[r\]]{} (330, 15)[(0,0)\[l\]]{}
(300,1000)
(0,0)\[t\]
( 600,100) (300,100)(100,0) (300,100)(300,0) (300,100)(500,0)
(300,900)
(0,0)\[t\]
( 600,100) (100,100)[(0,0)\[t\][(130,100)[$\rho^{(2)}_1(x_i)$]{}]{}]{} (300,100)[(0,0)\[t\][(130,100)[$\rho^{(2)}_2(x_i)$]{}]{}]{} (500,100)[(0,0)\[t\][(130,100)[$\rho^{(2)}_n(x_i)$]{}]{}]{} (400, 50)[(0,0)\[t\]]{}
(300,800)
(0,0)\[t\]
( 600,300) (100, 65)[(0,0)\[t\]]{} (100, 50) (150, 50)[(1,0)[450]{}]{} (300,140)[(0,0)\[t\]]{} (300,125) (350,120)[(1,0)[250]{}]{} (500,215)[(0,0)\[t\]]{} (500,200) (550,200)[(1,0)[ 50]{}]{} (600, 50)[( 0, 1)[550]{}]{} (600,600)[(-1, 0)[300]{}]{} (100,300)[( 0,-1)[200]{}]{} (300,300)[( 0,-1)[125]{}]{} (500,300)[( 0,-1)[ 50]{}]{} (300, 75)(300,0) (500,150)(500,0)
(300,500)
(0,0)\[t\]
( 600,100) (100,100)[(0,-1)[100]{}]{} (300,100)[(0,-1)[100]{}]{} (500,100)[(0,-1)[100]{}]{} (110,60)[(0,0)\[l\]]{} (310,60)[(0,0)\[l\]]{} (510,60)[(0,0)\[l\]]{}
(300,400)
(0,0)\[t\]
( 600,100) (100,100)(300,0) (300,100)(300,0) (500,100)(300,0)
(300,300)
(0,0)\[t\]
( 600,300) (300,300)[(0,-1)[100]{}]{} (300,150) (300,165)[(0,0)\[t\]]{} (300,100)[( 0,-1)[100]{}]{} (350, 150)[( 1, 0)[ 300]{}]{} (650, 150)[( 0, 1)[1000]{}]{} (650,1150)[(-1, 0)[ 350]{}]{} (310, 60)[(0,0)\[l\]]{}
( 15,250)[(600,875)[ ]{}]{} (500,245)[(0,0)\[t\]]{}
In the two examples of fig. \[fig:bran-wt2\] we have required that $\rho(x)$ can be split [*immediately*]{} into a sum of singular terms, each of them generated in a separate branch. In the real life it is often not true and fig. \[fig:bran-wt3\] illustrates more realistic scenario. Here, before branching can be applied, we make simplification $$\rho(x_i) \to \rho^{(1)}(x_i)$$ compensated by the weight $$w^{(0)} = {\rho(x_i) \over \rho^{(1)}(x_i) }
= {\rho(x_i) \over \sum\limits_k \rho^{(1)}_k(x_i) }.
\label{w0}$$ This simplification removes fine details from $\rho(x)$ (for example quantum mechanical interferences) which are numerically unimportant and prevent us from writing $\rho(x)$ as a sum of several positive and relatively simple singular terms. (Note that in $w^{(0)}$ we still do not have any Jacobians and we know nothing about transformation to variables $y_i^{(k)}$!) The branching is done in the next step for $\rho^{(1)}(x_i)$ and the weights $$w^{(1)}_k = {\rho^{(1)}_k(x_i) \over \rho^{(2)}_k(x_i) }.$$ compensate for the fact that the Jacobian $|\partial y^{(k)}/ \partial x|$ do not match exactly the $\rho^{(1)}_k(x_i)$, see discussion above. As in the example of fig. \[fig:bran-wt2\], $w^{(1)}_k$ involves elements of the calculation for a single branch only (Jacobian!). The branching probabilities $\bar{P}_k^{(2)}$ are easily calculated using known integrals $\bar{\sigma}_k^{(2)}$.
The total weight $$w = w^{(0)} w^{(1)}_k =
{\rho(x_i) \over \sum\limits_k \rho^{(1)}_k(x_i) }\;
{\rho^{(1)}_k(x_i) \over \rho^{(2)}_k(x_i) }
=
{\rho(x_i)
\over
\sum\limits_k \rho^{(1)}_k(x_i) }\;
{\rho^{(1)}_k(x_i)
\over
r^{(1)}_k \left|{ \partial y^{(k)} \over \partial x }\right| }
\label{local}$$ consists of global component $w^{(0)}$ which knows nothing about technicalities of generation in the individual branch $k$ and the local component $w^{(1)}_k$ which bears responsibility for technicalities of generation in the individual branch $k$ (in particular it may encapsulate cumbersome Jacobian functions). The lack of sum over $k$ in eq. (\[local\]) is not a mistake – the local part of the weight is calculated only for a SINGLE $k$-th branch!!! This is a great practical advantage and such an arrangement of the weights is generally much more convenient contrary to the algorithm being the straightforward extension of the algorithm in fig \[fig:bran-wt2\](b) for which the weight $$w = w^{(0)} w^{(1)} =
{\rho(x_i) \over \sum\limits_k \rho^{(1)}_k(x_i) }\;
{\sum\limits_k \rho^{(1)}_k(x_i) \over \sum\limits_k \rho^{(2)}_k(x_i) }
=
{\rho(x_i)
\over
\sum\limits_k \rho^{(1)}_k(x_i) }\;
{\sum\limits_k \rho^{(1)}_k(x_i)
\over
\sum\limits_k r^{(1)}_k \left|{ \partial y^{(k)} \over \partial x }\right| }
\label{global}$$ [*does*]{} include sum all over branches for the local part $w^{(1)}$ of the weight. The efficiency of the two algorithms depends on the details of the distribution and in order to see which of the above of two algorithms is more efficient has to be checked case by case.
Another important advantage of the algorithm of eq. (\[local\]) is that the part of the algorithm generating $\rho^{(1)}_k(x_i)$ can be encapsulated into single subprogram which generated $x_i$ according to $k$-th crude distribution $\rho^{(1)}_k(x_i)$ and provides to the outside world the weight $w^{(1)}_k$. The main program does not to need to know more about any details of the algorithm encapsulated in the subprogram. The rather annoying feature of the algorithm of eq. (\[global\]) is that for the construction of the total weight in the main program we need to know all nuts and bolts of the sub-generator for $\rho^{(1)}_k(x_i)$, thus encapsulation cannot be realized, leading to cumbersome non-modular program.
Finally let us note that the total integral is calculated with the usual formula $$\sigma = \bar{\sigma}^{(2)}\; <w^{(1)} \; w^{(0)}>, \quad
\bar{\sigma}^{(2)} = \sum\limits_k \bar{\sigma}^{(2)}_k$$ where we understand that for $<w^{(1)}>$ in eq. (\[local\]) the average is taken over all branches. For variable-weight events the the weight is $w=w^{(1)} w^{(0)}$ where $w^{(1)}=w^{(1)}_k$ for the [*actual*]{} $k$-th branch.
Conclusions
===========
I have described how to combine three elementary Monte Carlo methods rejection, branching and change of variables in the difficult task of generating multi-dimensional distributions. I have spend some time giving formal mathematical proofs of these methods, thus providing useful reference for papers describing MC event generators, where usually authors lack space/time to discuss such proofs. I have also discussed in quite some detail advantages and disadvantages various combinations of branching and rejection methods. Again, although these aspects may be known to authors of various MC programs they are practically never discussed. The most important for practical applications is probably the discussion on the advantages and disadvantages of the two arrangements of rejection and branching in fig. \[fig:bran-wt2\].
[1]{}
S. Jadach, Acta. Phys. Polon. [**B16**]{}, 1007 (1985).
F. James, Rep. Prog. Phys. [**43**]{}, 1145 (1980).
S. Jadach, E. Richter-Was, B. F. L. Ward, and Z. Was, Comput. Phys. Commun. [**70**]{}, 305 (1992).
R. Kleiss and R. Pittau, Comput. Phys. Commun. [**83**]{}, 141 (1994).
[^1]: I assume that the reader is familiar only with the elementary calculus methods, and purposely avoid to call $\rho$ a measure.
[^2]: I assume that the distribution $\rho$ may be quite singular, in particular it may include Dirac $\delta$ functions. The weight I expect to be analytical almost everywhere. It may, however, include discontinuities.
[^3]: Operationally this formula is identical for the case of variable-weight and constant-weight events and is therefore handy in practical calculations [@bhlumi2:1992]. The values of $\bar{\sigma}^{(1)}$ and $N^{(1)}$ can be accumulated using a dedicated, one-bin-histogram. This arrangement facilitates dumping all MC results on the disk and restarting MC run at the later time.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This paper addresses the problem of video summarization. Given an input video, the goal is to select a subset of the frames to create a summary video that optimally captures the important information of the input video. With the large amount of videos available online, video summarization provides a useful tool that assists video search, retrieval, browsing, etc. In this paper, we formulate video summarization as a sequence labeling problem. Unlike existing approaches that use recurrent models, we propose fully convolutional sequence models to solve video summarization. We firstly establish a novel connection between semantic segmentation and video summarization, and then adapt popular semantic segmentation networks for video summarization. Extensive experiments and analysis on two benchmark datasets demonstrate the effectiveness of our models.'
author:
- |
Mrigank Rochan, Linwei Ye, and Yang Wang\
`{mrochan,yel3,ywang}@cs.umanitoba.ca`
bibliography:
- 'wang.bib'
title: Video Summarization Using Fully Convolutional Sequence Networks
---
intro.tex related.tex approach.tex experiment.tex analysis.tex conclude.tex
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
This paper presents a fully automated procedure for controller synthesis for a general class of multi-agent systems under coupling constraints. Each agent is modeled with dynamics consisting of two terms: the first one models the coupling constraints and the other one is an additional bounded control input. We aim to design these inputs so that each agent meets an individual high-level specification given as a Metric Interval Temporal Logic (MITL). Furthermore, the connectivity of the initially connected agents, is required to be maintained. First, assuming a polyhedral partition of the workspace, a novel decentralized abstraction that provides controllers for each agent that guarantee the transition between different regions is designed. The controllers are the solution of a Robust Optimal Control Problem (ROCP) for each agent. Second, by utilizing techniques from formal verification, an algorithm that computes the individual runs which provably satisfy the high-level tasks is provided. Finally, simulation results conducted in MATLAB environment verify the performance of the proposed framework.\
author:
- 'Alexandros Nikou, Shahab Heshmati-alamdari, Christos Verginis and Dimos V. Dimarogonas [^1]'
bibliography:
- 'references.bib'
title: '**Decentralized Abstractions and Timed Constrained Planning of a General Class of Coupled Multi-Agent Systems**'
---
multi-agent systems, cooperative control, hybrid systems.
Introduction
============
Cooperative control of multi-agent systems has traditionally focused on designing distributed control laws in order to achieve global tasks such as consensus and formation control, and at the same time fulfill properties such as network connectivity and collision avoidance. Over the last few years, the field of control of multi-agent systems under temporal logic specifications has been gaining attention. In this work, we aim to additionally introduce specific time bounds into these tasks, in order to include specifications such as: “Robot 1 and robot 2 should visit region $A$ and $B$ within 4 time units respectively or “Both robots 1 and 2 should periodically survey regions $A_1$, $A_2$, $A_3$, avoid region $X$ and always keep the longest time between two consecutive visits to $A_1$ below 8 time units".
The qualitative specification language that has primarily been used to express the high-level tasks is Linear Temporal Logic (LTL) (see, e.g., [@loizou_2004; @muray_2010_receding]). There is a rich body of literature containing algorithms for verification and synthesis of multi-agent systems under temporal logic specifications ([@guo_2015_reconfiguration; @frazzoli_vehicle_routing; @zavlanos_2016_multi-agent_LTL]). A three-step hierarchical procedure to address the problem of multi-agent systems under LTL specifications is described as follows ([@fainekos_girard_2009_temporal; @belta_2010_product_system; @belta_cdc_reduced_communication]): first the dynamics of each agent is abstracted into a Transition System (TS). Second, by invoking ideas from formal verification, a discrete plan that meets the high-level tasks is synthesized for each agent. Third, the discrete plan is translated into a sequence of continuous time controllers for the original continuous dynamical system of each agent.
Controller synthesis under timed specifications has been considered in [@liu_MTL; @murray_2015_stl; @topcu_2015; @fainekos_mtl_2015_robot; @baras_MTL_2016_new]. However, all these works are restricted to single agent planning and are not extendable to multi-agent systems in a straightforward way. The multi-agent case has been considered in [@frazzoli_MTL], where the vehicle routing problem was addressed, under Metric Temporal Logic (MTL) specifications. The corresponding approach does not rely on automata-based verification, as it is based on a construction of linear inequalities and the solution of a Mixed-Integer Linear Programming (MILP) problem.
An automata-based solution was proposed in our previous work [@alex_2016_acc], where MITL formulas were introduced in order to synthesize controllers such that every agent fulfills an individual specification and the team of agents fulfills a global specification. Specifically, the abstraction of each agent’s dynamics was considered to be given and an upper bound of the time that each agent needs to perform a transition from one region to another was assumed. Furthermore, potential coupled constraints between the agents were not taken into consideration. Motivated by this, in this work, we aim to address the aforementioned issues. We assume that the dynamics of each agent consists of two parts: the first part is a nonlinear function representing the coupling between the agent and its neighbors, and the second one is an additional control input which will be exploited for high-level planning. Hereafter, we call it a free input. A decentralized abstraction procedure is provided, which leads to an individual Weighted Transition System (WTS) for each agent and provides a basis for high-level planning.
Abstractions for both single and multi-agent systems have been provided e.g. in [@alur_2000_discrete_abstractions; @belta_2004_abstraction; @helwa2014block; @habets_2006_reachability; @zamani_2012_symbolic; @liu_2016_abstraction; @abate_2014_finite_abstractions; @pola_2016_symbolic; @boskos_cdc_2015]. In this paper, we deal with the complete framework of both abstractions and controller synthesis of multi-agent systems. We start from the dynamics of each agent and we provide controllers that guarantee the transition between the regions of the workspace, while the initially connected agents remain connected for all times. The decentralized controllers are the solution of an ROCP. Then, each agent is assigned an individual task given as an MITL formulas. We aim to synthesize controllers, in discrete level, so that each agent performs the desired individual task within specific time bounds as imposed by the MITL formulas. In particular, we provide an automatic controller synthesis method of a general class of coupled multi-agent systems under high-level tasks with timed constraints. Compared to existing works on multi-agent planning under temporal logic specifications, the proposed approach considers dynamically coupled multi-agent systems under timed temporal specifications in a distributed way.
In our previous work [@alex_acc_8_pages], we treated a similar problem, but the under consideration dynamics were linear couplings and connectivity maintenance was not guaranteed by the proposed control scheme. Furthermore, the procedure was partially decentralized, due to the fact that a product Wighted Transition System (WTS) was required, which rendered the framework computationally intractable. To the best of the authors’ knowledge, this is the first time that a fully automated framework for a general class of multi-agent systems consisting of both constructing purely decentralized abstractions and conducting timed temporal logic planning is considered.
This paper is organized as follows. In Section \[sec: preliminaries\] a description of the necessary mathematical tools, the notations and the definitions are given. Section \[sec: prob\_formulation\] provides the dynamics of the system and the formal problem statement. Section \[sec: solution\] discusses the technical details of the solution. Section \[sec: simulation\_results\] is devoted to a simulation example. Finally, conclusions and future work are discussed in Section \[sec: conclusions\].
Notation and Preliminaries {#sec: preliminaries}
==========================
Notation
--------
We denote by $\mathbb{R}, \mathbb{Q}_+, \mathbb{N}$ the set of real, nonnegative rational and natural numbers including 0, respectively. $\mathbb{R}^{n}_{\ge 0}$ and $\mathbb{R}^{n}_{> 0}$ are the sets of real $n$-vectors with all elements nonnegative and positive, respectively. Define also $\mathbb{T}_{\infty} = \mathbb{T} \cup \{\infty\}$ for a set $\mathbb{T} \subseteq \mathbb{R}$. Given a set $S$, denote by $|S|$ its cardinality, by $S^N = S \times \dots \times S$ its $N$-fold Cartesian product, and by $2^S$ the set of all its subsets. Given the sets $S_1, S_2$, their *Minkowski addition* is defined by $S_1 \oplus S_2 = \{s_1 + s_2 : s_1 \in S_1, s_2 \in S_2\}$. $I_n \in \mathbb{R}^{n \times n}$ stands for the identity matrix. The notation $\|x\|$ is used for the Euclidean norm of a vector $x \in \mathbb{R}^n$. $\|A\| = \text{max} \{\|Ax\| : \|x\| = 1\}$ stands for the induced norm of a matrix $A \in \mathbb{R}^{n \times n}$. $\mathcal{B}(c, \underline{r}) = \{x \in \mathbb{R}^2: \|x-c\|\le \underline{r} \}$ is the disk of center $c \in \mathbb{R}^2$ and $\underline{r} \in \mathbb{R}_{> 0}$. The absolute value of the maximum singular value and the absolute value of the minimum eigenvalue of a matrix $A \in \mathbb{R}^{n \times n}$ are denoted by $\sigma_{\max}(A), \lambda_{\min}(A)$, respectively. The indexes $i$ and $j$ stand for agent $i$ and its neighbors (see Sec. \[sec: prob\_formulation\] for the definition of neighbors), respectively; $\mu, z \in \mathbb{N}$ are indexes used for sequences and sampling times, respectively.
Consider two sets $A, B \subseteq \mathbb{R}^n$. Then, the *Pontryagin difference* is defined by: $$A \sim B = \{x \in \mathbb{R}^n: x+y \in A, \forall \ y \in B\}.$$
\[def:class\_K\] ([@khalil_nonlinear_systems]) A continuous function $\alpha: [0, a) \to \mathbb{R}_{\ge 0}$ is said to belong to *class* $\mathcal{K}$, if it is strictly increasing and $\alpha(0) = 0$. It is said to belong to class $\mathcal{K}_{\infty}$ if $a = \infty$ and $\alpha(r) \to \infty$, as $r \to \infty$.
\[def:class\_KL\] ([@khalil_nonlinear_systems]) A continuous function $\beta: [0, a) \times \mathbb{R}_{\ge 0} \to \mathbb{R}_{\ge 0}$ is said to belong to *class* $\mathcal{KL}$, if:
- For each fixed $s$, $\beta(r, s) \in \mathcal{K}$ with respect to $r$.
- For each fixed $r$, $\beta(r, s)$ is decreasing with respect to $s$ and $\beta(r, s) \to 0$, at $s \to \infty$.
\[def:ISS\] ([@sontag_2008_ISS]) A nonlinear system $\dot{x} = f(x,u)$ with initial condition $x(t_0)$ is said to be *Input to State Stable (ISS)* if there exist functions $\beta \in \mathcal{KL}$ and $\sigma \in \mathcal{K}_{\infty}$ such that: $$\|x(t)\| \le \beta(\|x(t_0)\|, t)+\sigma(\|u\|).$$
\[def:ISS\_lyapunov\] ([@sontag_2008_ISS]) A Lyapunov function $V(x,u)$ for the nonlinear system $\dot{x} = f(x,u)$ with initial condition $x(t_0)$ is said to be *ISS-Lyapunov function* if there exist functions $\alpha, \sigma \in \mathcal{K}_\infty$ such that: $$\label{eq:lyapunov_iss}
\dot{V}(x,u) \le - \alpha(\|x\|)+\sigma(\|u\|), \forall x, u.$$
A nonlinear system $\dot{x} = f(x,u)$ with initial condition $x(t_0)$ is said to be ISS if and only if it admits a ISS-Lyapunov function.
The proof can be found in [@sontag_1995_ISS_proofs].
Partitions
----------
In the subsequent analysis a discrete partition of the workspace will be considered which is formalized through the following definition.
\[def:partition\] Given a set $S$, we say that a family of sets $\{S_\ell\}_{\ell \in \mathbb{I}}$ forms a partition of $S$ if $S \neq \emptyset$, $\displaystyle \bigcup_{\ell \in \mathbb{I}} S_\ell = S$ and for every $S,S' \in S$ with $S \neq S'$ it holds $S \cap S' = \emptyset$.
Hereafter, every region $S_\ell$ of a partition $S$ will be called *region*.
Time Sequence, Timed Run and Weighted Transition System
-------------------------------------------------------
In this section we include some definitions that are required to analyze our framework.
An infinite sequence of elements of a set $X$ is called an *infinite word* over this set and it is denoted by $\chi = \chi(0)\chi(1) \ldots $ The $z$-th element of a sequence is denoted by $\chi(z)$. For certain technical reasons that will be clarified in the sequel, we will assume hereafter that $\mathbb{T} = \mathbb{Q}_{+}$.
([@alur1994]) A *time sequence* $\tau = \tau(0) \tau(1) \ldots$ is an infinite sequence of time values $\tau(\mu) \in \mathbb{T} = \mathbb{Q}_{+}$, satisfying the following properties: 1) Monotonicity: $\tau(\mu) < \tau(\mu+1)$ for all $j \geq 0$; 2) Progress: For every $t \in \mathbb{T}$, there exists $\mu \ge 1$, such that $\tau(\mu) > t$.
An *atomic proposition* $\sigma$ is a statement that is either True $(\top)$ or False $(\bot)$.
([@alur1994]) Let $\Sigma$ be a finite set of atomic propositions. A *timed word* $w$ over the set $\Sigma$ is an infinite sequence $w^t = (w(0), \tau(0)) (w(1), \tau(1)) \ldots$ where $w(0) w(1) \ldots$ is an infinite word over the set $2^{\Sigma}$ and $\tau(0) \tau(1) \ldots$ is a time sequence with $\tau(\mu) \in \mathbb{T}, \mu \geq 0$.
\[def: WTS\] A Weighted Transition System (*WTS*) is a tuple $(S, S_0, Act, \longrightarrow, d, \Sigma, L)$ where $S$ is a finite set of states; $S_0 \subseteq S$ is a set of initial states; $Act$ is a set of actions; $\longrightarrow \subseteq S \times Act \times S$ is a transition relation; $d: \longrightarrow \rightarrow \mathbb{T}$ is a map that assigns a positive weight to each transition; $\Sigma$ is a finite set of atomic propositions; and $L: S \rightarrow 2^{\Sigma}$ is a labeling function.
\[run\_of\_WTS\] A *timed run* of a WTS is an infinite sequence $r^t = (r(0), \tau(0))(r(1), \tau(1)) \ldots$, such that $r(0) \in S_0$, and for all $\mu \geq 1$, it holds that $r(\mu) \in S$ and $(r(\mu), \alpha(\mu), r(\mu+1)) \in \longrightarrow$ for a sequence of actions $\alpha(1) \alpha(2) \ldots$ with $\alpha(\mu) \in Act, \forall \ \mu \geq 1$. The *time stamps* $\tau(\mu), \mu \geq 0$ are inductively defined as: 1) $\tau(0) = 0$; 2) $\displaystyle \tau(\mu+1) = \tau(\mu) + d(r(\mu), \alpha(\mu), r(\mu+1)), \forall \ \mu \geq 1$. Every timed run $r^t$ generates a *timed word* $w(r^t) = (w(0), \tau(0)) \ (w(1), \tau(1)) \ldots$ over the set $2^{\Sigma}$ where $w(\mu) = L(r(\mu))$, $\forall \ \mu \geq 0$ is the subset of atomic propositions that are true at state $r(\mu)$.
Metric Interval Temporal Logic (MITL)
-------------------------------------
The syntax of *Metric Interval Temporal Logic (MITL)* over a set of atomic propositions $\Sigma$ is defined by the grammar: $$\varphi := p \ | \ \neg \varphi \ | \ \varphi_1 \wedge \varphi_2 \ | \ \bigcirc_I \varphi \ | \ \Diamond_I \varphi \mid \square_I \varphi \mid \varphi_1 \ \mathcal{U}_I \ \varphi_2,$$ where $\sigma \in \Sigma$, and $\bigcirc$, $\Diamond$, $\square$ and $\mathcal U$ are the next, eventually, always and until temporal operators, respectively; $I = [a, b] \subseteq \mathbb{T}$ where $a, b \in [0, \infty]$ with $a < b$ is a non-empty timed interval. MITL can be interpreted either in continuous or point-wise semantics [@pavithra_expressiveness]. In this paper, the latter approach is utilized, since the consideration of point-wise (event-based) semantics is more suitable for the automata-based specifications considered in a discretized state-space. The MITL formulas are interpreted over timed words like the ones produced by a WTS it is given in Def. \[run\_of\_WTS\].
\[def:mitl\_semantics\] ([@pavithra_expressiveness], [@quaknine_decidability]) Given a timed word $w^t = (w(0),\tau(0))(w(1),\tau(1)) \dots$, an MITL formula $\varphi$ and a position $i$ in the timed word, the satisfaction relation $(w^t, i) \models \varphi$, for $\ i \geq 0$ (read $w^t$ satisfies $\varphi$ at position $\mu$) is inductively defined as follows: $$\begin{aligned}
\label{eq: for1}
&(w^t, \mu) \models p \Leftrightarrow p \in w(\mu), \\
&(w^t, \mu) \models \neg \varphi \Leftrightarrow (w^t, i) \not \models \varphi, \\
&(w^t, \mu) \models \varphi_1 \wedge \varphi_2 \Leftrightarrow (w^t, \mu) \models \varphi_1 \ \text{and} \ (w^t, \mu) \models \varphi_2, \\
&(w^t, \mu) \models \bigcirc_I \ \varphi \Leftrightarrow (w^t, \mu+1) \models \varphi \notag \\
&\hspace{35mm} \text{and} \ \tau(\mu+1) - \tau(i) \in I, \\
&(w^t, \mu) \models \Diamond_I \varphi \Leftrightarrow \exists \mu' \ge \mu, \ \text{such that} \notag \\
&\hspace{35mm} (w^t, j) \models \varphi, \tau(\mu')-\tau(\mu) \in {I},
\end{aligned}$$ $$\begin{aligned}
&(w^t, \mu) \models \square_I \varphi \Leftrightarrow \forall \mu' \ge \mu, \notag \\
&\hspace{25mm} \tau(\mu')-\tau(\mu) \in {I} \Rightarrow (w^t, \mu') \models \varphi, \\
&(w^t, \mu) \models \varphi_1 \ \mathcal{U}_I \ \varphi_2 \Leftrightarrow \exists \mu' \ge \mu, \ \text{s.t. } (w^t, \mu') \models \varphi_2, \\
&\qquad \tau(\mu')-\tau(\mu) \in I \ \text{and} \ (w^t, \mu'') \models \varphi_1, \forall \ \mu \leq \mu'' < \mu'.
\end{aligned}$$ We say that a timed run $r^t = (r(0),\tau(0))(r(1),\tau(1)) \dots$ satisfies the MITL formula $\varphi$ (we write $r^t \models \varphi$) if and only if the corresponding timed word $w(r^t) = (w(0),\tau(0))(w(1),\tau(1)) \dots$ with $w(\mu) = L(r(\mu)), \forall \mu \ge 0$, satisfies the MITL formula ($w(r^t) \models \varphi$).
It has been proved that MITL is decidable in infinite words and point-wise semantics, which is the case considered here (see [@alur_mitl; @reynold] for details). The model checking and satisfiability problems are *EXPSPACE*-complete. It should be noted that in the context of timed systems, EXSPACE complexity is fairly low [@bouyer_phd]. An example with a WTS and two runs $r_1^t, r_2^t$ that satisfy two MITL formulas can be found in [@alex_acc_2017 Section II, page 4].
Timed Büchi Automata {#sec: timed_automata}
--------------------
*Timed Büchi Automata (TBA)* were introduced in [@alur1994] and in this work, we also partially adopt the notation from [@bouyer_phd; @tripakis_tba]. Let $C = \{c_1, \ldots, c_{|C|}\}$ be a finite set of *clocks*. The set of *clock constraints* $\Phi(C)$ is defined by the grammar $$\phi := \top \mid \ \neg \phi \ | \ \phi_1 \wedge \phi_2 \ | \ c \bowtie \psi,$$ where $c \in C$ is a clock, $\psi \in \mathbb{T}$ is a clock constant and $\bowtie \ \in \{ <, >, \geq, \leq, = \}$. A clock *valuation* is a function $\nu: C \rightarrow\mathbb{T}$ that assigns a value to each clock. A clock $c_i$ has valuation $\nu_i$ for $i \in \{1, \ldots, |C|\}$, and $\nu = (\nu_1, \ldots, \nu_{|C|})$. We denote by $\nu \models \phi$ the fact that the valuation $\nu$ satisfies the clock constraint $\phi$.
A *Timed Büchi Automaton* is a tuple $\mathcal{A} = (Q, Q^{\text{init}}, C, Inv, E, F, \Sigma, \mathcal{L})$ where $Q$ is a finite set of locations; $Q^{\text{init}} \subseteq Q$ is the set of initial locations; $C$ is a finite set of clocks; $Inv: Q \rightarrow \Phi(C)$ is the invariant; $E \subseteq Q \times \Phi(C) \times 2^C \times Q$ gives the set of edges; $F \subseteq Q$ is a set of accepting locations; $\Sigma$ is a finite set of atomic propositions; and $\mathcal{L}: Q \rightarrow 2^{\Sigma}$ labels every state with a subset of atomic propositions.
For the semantics of TBA we refer the reader to [@alex_acc_2017 Section II, page 4]. The problem of deciding the emptiness of the language of a given TBA $\mathcal{A}$ is PSPACE-complete [@alur1994]. Any MITL formula $\varphi$ over $\Sigma$ can be algorithmically translated to a TBA with the alphabet $2^{\Sigma}$, such that the language of timed words that satisfy $\varphi$ is the language of timed words produced by the TBA ([@alur_mitl; @maler_MITL_TA; @nickovic_timed_aut]). An example of a TBA and accepting runs of it can be found in [@alex_acc_2017 Section II, page 4].
Traditionally, the clock constraints and the TBAs are defined with $\mathbb{T} = \mathbb{N}$. However, they can be extended to accommodate $\mathbb{T} = \mathbb{Q}_+$, by multiplying all the rational numbers that are appearing in the state invariants and the edge constraints with their least common multiple.
Problem Formulation {#sec: prob_formulation}
===================
System Model
------------
Consider a system of $N$ agents, with $\mathcal{V} = \{1,\dots, N\}, N \ge 2$, operating in a workspace $W \subseteq \mathbb{R}^2$. The workspace is assumed to be closed, bounded and connected. Let $x_i: \mathbb{R}_{\ge 0} \to D$ denotes the position of each agent in the workspace at time $t \in \mathbb{R}_{\ge 0}$. Each agent is equipped with a sensor device that can sense omnidirectionally. Let the disk $\mathcal{B}(x_i(t), \underline{r})$ model the sensing zone of agent $i$ at time $t \in \mathbb{R}_{\ge 0}$, where $\underline{r} \in \mathbb{R}_{\ge 0}$ is the sensing radius. The sensing radius is the same for all the agents. Let also $h > 0$ denote the constant sampling period of the system. We make the following assumption:
\[ass:measurement\_assumption\] (Measurements Assumption) It is assumed that each agent $i$, is able to measure its own position and all agents’ positions that are located within agent’s $i$ sensing zone without any delays.
According to Assumption \[ass:measurement\_assumption\], the agent’s $i$ neighboring set at time $t_0$ is defined by $\mathcal{N}_i = \{j \in \mathcal{V}: x_j(t_0) \in \mathcal{B}(x_i(t_0), \underline{r})\}$. For the neighboring set $\mathcal{N}_i$ define also $N_i = |\mathcal{N}_i|$. Note that $i \in \mathcal{N}_j \Leftrightarrow j \in \mathcal{N}_i, \forall \ i,j \in \mathcal{V}, i \neq j$. The control design for every agent $i$ should guarantee that it remains connected with all its neighbors $j \in \mathcal{N}_i$, for all times.
Consider the neighboring set $\mathcal{N}_i$. The coupled dynamics of each agent are given in the form: $$\label{eq:system}
\dot{x}_i = f(x_i, \bar{x}_i)+u_{i}, \ x_i \in W, \ i \in \mathcal{V},$$ where $f:W$ $\times$ $W^{N_i}$ $\to W$, is a nonlinear function representing the coupling between agent $i$ and its neighbors $i_1, \dots, i_{N_i}$. The notation $\bar{x}_i = [ x_{i_1}^\top, \ldots, x_{i_{N_i}}^\top]^\top \in W^{N_i}$ is used for the vector of the neighbors of agent $i$, and $u_i: \mathbb{R}_{\ge 0} \to \mathbb{R}^2, \ i \in \mathcal{V}$ is the control input of each agent. For the dynamics of each agent the following assumption are taken.
\[ass:dynamic\_control\_bounds\] There exist constants $u_{\max}, \bar{M}$ with $0 < u_{\max} < \bar{M} < \infty$ such that the following holds $\forall i \in \mathcal{V}, (x_i, \bar{x}_i) \in W \times W^{N_i}$:
$$\begin{aligned}
&\|f_i(x_i, \bar{x}_i) \| \leq \bar{M}, \ \\
& u_i \in \mathcal{U}_i \triangleq \{u_i \in \mathbb{R}^2: \|u_i\| \le u_{\max}\}. \label{eq:control_constraints}
\end{aligned}$$
\[ass:lipsitch\_f\_x\_bar\_x\] The functions $f_i(x_i, \bar{x}_i), i \in \mathcal{V}$ are *Lipschitz continuous* in $W \times W^{N_i}$. Thus, there exists constants $L_i, \bar{L}_i > 0$ such that the following inequalities hold:
$$\begin{aligned}
\|f_i(x_i, \bar{x}_i) - f_i(y_i, \bar{x}_i) \| &\le L_i \|x_i - y_i\|, \label{eq:lip_1} \\
\|f_i(x_i, \bar{x}_i) - f_i(x_i, \bar{y}_i) \| &\le \bar{L}_i \|\bar{x}_i - \bar{y}_i\|, \label{eq:lip_2}\end{aligned}$$
for all $x_i, y_i \in W, \bar{x}_i, \bar{y}_i \in W^{N_i}, i \in \mathcal{V}$.
The coupling terms $f_i(x_i, \bar{x}_i), i \in \mathcal{V}$ are encountered in a large set of multi-agent protocols [@mesbahi_2010_graph_theory], including consensus, connectivity maintenance, collision avoidance and formation control. In addition, may represent internal dynamics of the system as for instance in the case of smart buildings (see e.g., [@andreasson_2014_smart_building]) where the temperature $T_i, i \in \mathcal{V}$ of each room evolves according to the law $\dot{T}_i = \sum_{j \in \mathcal{N}_i} \alpha_{ij}(T_j-T_i)+u_i$, with $\alpha_{ij}$ representing the heat conductivity between rooms $i$ and $j$ and $u_i$ the heating/cooling capabilities of the room.
Specification
-------------
Our goal is to control the multi-agent system so that each agent obeys a given individual specification. In particular, it is required to drive each agent to a sequence of desired subsets of the *workspace* $W$ within certain time limits and provide certain atomic tasks there. Atomic tasks are captured through a finite set of atomic propositions $\Sigma_i, i \in \mathcal{V}$, with $\Sigma_i \cap \Sigma_j = \emptyset$, for all $i,j \in \mathcal{V}, i \neq j,$ which means that the agents do not share any atomic propositions. Each position $x_i$ of each agent $i \in \mathcal{V}$ is labeled with atomic propositions that hold there. Initially, a labeling function $$\label{eq:label_lambda}
\Lambda_i: W \to 2^{\Sigma_i},$$ is introduced for each agent $i \in \mathcal{V}$ which maps each state $x_i \in \mathbb{R}^2$ with the atomic propositions $\Lambda_i(x_i)$ which hold true at $x_i$ i.e., the subset of atomic propositions that hold for agent $i$ in position $x_i$. Define also by $\displaystyle \Lambda(x) = \bigcup_{i \in \mathcal V} \Lambda_i(x)$ the union of all the labeling functions. Let us now introduce the following assumption which is important for defining the problem properly.
\[assumption: AP\_cell\_decomposition\] There exists a partition $D = \{D_\ell\}_{\ell \in \mathbb I}$ of the workspace $W$ which respects the labeling function $\Lambda$ i.e., for all $D_\ell \in D$ it holds that $\Lambda(x) = \Lambda(x'), \forall \ x, x' \in D_\ell$. This assumption, intuitively, and without loss of generality, means that the same atomic propositions hold at all the points that belong to the same region of the partition.
Although the regions $D_\ell, \ell \in \mathbb{I}$ of the partition $D$ may have different geometric shape, without loss of generality, we assume that they are hexagons with side length $R$. Define also for each agent $i$ a labeling function: $$\label{eq:label_mathcal_lambda}
{L}_i: D \to 2^{\Sigma_i},$$ which maps every region of the partition $D$ to the subset of the atomic propositions which hold true there. Furthermore, we assume that a time step $T > h > 0$ is given. This time step models the required time in which each agent should transit from a region to a neighboring region and is the same for all the agents.
The trajectory of each agent $i$ is denoted by $x_i(t), t \geq 0, i \in \mathcal{V}$. The trajectory $x_i(t)$ is associated with a unique sequence: $$r_{x_i}^t = (r_i(0), \tau_i(0))(r_i(1), \tau_i(1))(r_i(2), \tau_i(2))\ldots,$$ of regions that the agent $i$ crosses, where for all $\mu \ge 0$ it holds that: $x_i(\tau_i(\mu) \in r_i(\mu)$ and $\Lambda_i(x_i(t)) = {L}_i(r_i(\mu)), \forall \ t \in [\tau_i(\mu), \tau_i(\mu+1))$ for some $r_i(\mu) \in D$ and $r_i(\mu) \ne r_i(\mu+1)$. The timed word: $$\begin{aligned}
w_{x_i}^t &= (L_i(r_i(0)), \tau_i(0))({L}_i(r_i(1)), \tau_i(1)) \notag \\
&\hspace{38mm}(L_i(r_i(2)), \tau_i(2))\ldots,\end{aligned}$$ where $w_i(\mu) = {L}_i(r_i(\mu)), \mu \ge 0, i \in \mathcal{V}$, is associated uniquely with the trajectory $x_i(t)$.
For each agent $i \in \mathcal{V}$ we define the *relaxed timed word* as: $$\begin{aligned}
\label{eq:relaxed_timed_word}
\widetilde{w}_{i}^t &= (w_i(0), \widetilde{\tau}_i(0))(w_i(1), \widetilde{\tau}_i(1)) (w_i(2), \widetilde{\tau}_i(2))\ldots,\end{aligned}$$ where $w_i(\mu) = L_i(r_i(\mu)), \tilde{\tau}_i(\mu) \in [\tau_i(\mu), \tau_i(\mu+1)), \forall \ \mu \ge 0$.
The time stamp $\tau_i(0) = \widetilde{\tau}_i(0) = t_0, i \in \mathcal{V}$ models the initial starting time of the agents. The time stamps $\tau_i(\mu), \mu \ge 1$ models the exact time in which the agent $i$ crosses the boundary of the regions $r_i(\mu-1)$ and $r_i(\mu)$. The time stamps $\widetilde{\tau}_i(\mu)$ model a time instant in which the agent $i$ is in the region $r_i(\mu)$ of the workspace (see Example \[ex: example\_01\] below). The specification task $\varphi_i$ given as an MITL formula over the set of atomic propositions $\Sigma_i$, represents desired tasks that are imposed to each agent $i \in \mathcal{I}$. We say that a trajectory $x_i(t)$ *satisfies* a formula $\varphi_i$ given in MITL over the set $\Sigma_i$, and we formally write: $$\label{eq:traj_satisf_formula}
x_i(t) \models \varphi_i, \forall t \ge 0,$$ if and only if there exists a *relaxed timed word* $\widetilde{w}_{i}^t$ that complies with $x_i(t)$ and satisfies $\varphi_i$ according to the semantics of MITL in \[def:mitl\_semantics\].
(-2.5, -5.0) grid (0,0); (-7.5,0.0) – (-2.5,0.0); (-7.5,-2.5) – (-2.5,-2.5); (-7.5,-5.0) – (-2.5,-5.0); (-10.0, -5.0) grid (-7.5,0);
(-10, -2.5) rectangle +(2.5, 2.5); (-10, -5.0) rectangle +(2.5, 2.5); (-7.5, -2.5) rectangle +(5.0, 2.5); (-2.5, -2.5) rectangle +(2.5, 2.5); (-2.5, -5.0) rectangle +(2.5, 2.5); (-7.5, -5.0) rectangle +(5.0, 2.5);
(-8.8, -3.7) .. controls (-7.50, -0.0) .. (-5.8, -3.2);
(-5.1, -3.7) .. controls (-0.2, -2.5) .. (-3.0, -1.1);
(-8.8, -3.7) node\[circle, inner sep=0.8pt, fill=black, label=[below:[$x_1(0)$]{}]{}\] (A1) ; (-8.35, -2.5) node\[circle, inner sep=0.8pt, fill=black\] (B1) ; (-7.50, -0.85) node\[circle, inner sep=0.8pt, fill=black, label=[left:[$x_1(t_2)$]{}]{}\] (C1) ; (-6.2, -2.5) node\[circle, inner sep=0.8pt, fill=black\] (D1) ;
(-1.15, -2.5) node\[circle, inner sep=0.8pt, fill=black\] (A2) ; (-2.5, -3.05) node\[circle, inner sep=0.8pt, fill=black\] (B2) ; (-5.1, -3.7) node\[circle, inner sep=0.8pt, fill=black, label=[below:[$x_2(0)$]{}]{}\] (C2) ; (-2.5, -1.35) node\[circle, inner sep=0.8pt, fill=black\] (D2) ;
at (-8.7, 0.5) [$D_1$]{}; at (-5.0, 0.5) [$D_2$]{}; at (-1.3, 0.5) [$D_3$]{}; at (-1.3, -5.5) [$D_4$]{}; at (-5.0, -5.5) [$D_5$]{}; at (-8.7, -5.5) [$D_6$]{}; at (-9,-2.2) [$x_1(t_1)$]{}; at (-5.5,-2.2) [$x_1(t_3)$]{}; at (-1.8,-3.3) [$x_2(t_1')$]{}; at (-0.57,-2.1) [$x_2(t_2')$]{}; at (-1.9,-1.1) [$x_2(t_3')$]{};
\[ex: example\_01\] Consider $N = 2$ agents performing in the partitioned environment of Fig. \[fig: example\_01\]. Both agents have the ability to pick up, deliver and throw two different balls. Their sets of atomic propositions are $\Sigma_1 = \{\rm pickUp1, deliver1, throw1\}$ and $\Sigma_2 = \{\rm pickUp2, deliver2, throw2\}$, respectively, and satisfy $\Sigma_1 \cap \Sigma_2 = \emptyset$. Three points of the agents’ trajectories that belong to different regions with different atomic propositions are captured. Assume that $t_1 < t_1' < t_2 < t_2 < t_2' < t_3 < t_3'.$ The trajectories $x_1(t), x_2(t), t \ge 0$ are depicted with the red lines. According to Assumption \[assumption: AP\_cell\_decomposition\], the partition $D = \{D_\ell\}_{\ell \in \mathbb I} = \{D_1, \ldots, D_6\}$ is given where $\mathbb I = \{1, \ldots, 6\}$ respects the labeling functions $\Lambda_i, {L}_i, i \in \{1,2\}$. In particular, it holds that: $$\begin{aligned}
&\Lambda_1(x_1(t)) = {L}_1(r_1(0)) = \{\rm pickUp1\}, t \in [0, t_1), \notag \\
&\Lambda_1(x_1(t)) = {L}_1(r_1(1)) = \{\rm throw1\}, t \in [t_1, t_2), \notag \\
&\Lambda_1(x_1(t)) = {L}_1(r_1(2)) = \{\rm deliver1\}, t \in [t_2, t_3), \notag \\
&\Lambda_1(x_1(t)) = {L}_1(r_1(3)) = \emptyset, t \ge t_3. \notag \\
&\Lambda_2(x_2(t)) = {L}_2(r_2(0)) = \{\rm pickUp2\}, t \in [0, t_1'), \notag \\
&\Lambda_2(x_2(t)) = {L}_2(r_2(1)) = \{\rm deliver2\}, t \in [t_1', t_2'), \notag \\
&\Lambda_2(x_2(t)) = {L}_2(r_2(2)) = \{\rm throw2\}, t \in [t_2', t_3'), \notag \\
&\Lambda_2(x_2(t)) = {L}_2(r_2(3)) = \emptyset, t \ge t_3'. \notag
\end{aligned}$$ By the fact that $w_i(\mu) = {L}(r_i(\mu)), \forall \ i \in \{1,2\}, \mu \in \{1,2,3\}$, the corresponding individual timed words are given as: $$\begin{aligned}
w^t_{x_1} &= (\{\rm pickUp1\}, 0)(\{\rm throw1\}, t_1)(\{\rm deliver1\}, t_2)(\emptyset, t_3), \notag \\
w^t_{x_2} &= (\{\rm pickUp2\}, 0)(\{\rm deliver2\}, t_1')(\{\rm throw2\}, t_2')(\emptyset, t_3'). \notag
\end{aligned}$$ According to , two two relaxed timed words (depicted with red in Fig. \[fig: example\_01\]) are given as: $$\begin{aligned}
&w^t_{1} = (\{\rm pickUp1\}, \widetilde{\tau}_1(0))(\{\rm throw1\}, \widetilde{\tau}_1(1)) \notag \\
&\hspace{40mm} (\{\rm deliver1\}, \widetilde{\tau}_1(2)) (\emptyset, \widetilde{\tau}_1(3)), \notag \\
&w^t_{2} = (\{\rm pickUp2\}, \widetilde{\tau}_2(0))(\{\rm deliver2\}, \widetilde{\tau}_2(1)) \notag \\
&\hspace{40mm}(\{\rm throw2\}, \widetilde{\tau}_2(2))(\emptyset, \widetilde{\tau}_2(3)). \notag
\end{aligned}$$ The time stamps $\widetilde{\tau}_1(\mu), \widetilde{\tau}_2(\mu), \mu \in \{1, 2, 3\}$ should satisfy the following conditions: $$\begin{aligned}
\widetilde{\tau}_1(0) &\in [\tau_1(0), \tau_1(1)) = [0, t_1), \\
\widetilde{\tau}_1(1) &\in [\tau_1(1), \tau_1(2)) = [t_1, t_2), \\
\widetilde{\tau}_1(2) &\in [\tau_1(2), \tau_1(3)) = [t_2, t_3), \\
\widetilde{\tau}_1(3) &\in [\tau_1(3), \cdot) = [t_3, \cdot), \\
\widetilde{\tau}_2(0) &\in [\tau_2(0), \tau_2(1)) = [0, t_1), \\
\widetilde{\tau}_2(1) &\in [\tau_2(1), \tau_2(2)) = [t_1, t_2), \\
\widetilde{\tau}_2(2) &\in [\tau_2(2), \tau_2(3)) = [t_2, t_3), \\
\widetilde{\tau}_2(3) &\in [\tau_2(3), \cdot) = [t_3, \cdot). \\
\end{aligned}$$
Problem Statement
-----------------
We can now formulate the problem treated in this paper as follows:
\[problem: basic\_prob\] Given $N$ agents operating in the bounded workspace $W \subseteq \mathbb{R}^2$, their initial positions $x_1(t_0), \dots, x_N(t_0)$, their dynamics as in , a time step $T > h > 0$, $N$ task specification formulas $\varphi_1, \ldots, \varphi_N$ expressed in MITL over the sets of services $\Sigma_1, \ldots, \Sigma_{{N}}$, respectively, a partition of the workspace $W$ into hexagonal regions $\{D_\ell\}_{\ell \in \mathbb I}$ with side length $R$ as in Assumption \[assumption: AP\_cell\_decomposition\] and the labeling functions $\Lambda_1, \ldots, \Lambda_N, {L}_1, \ldots, {L}_N$, as in , , assign control laws $u_1, \ldots, u_N$ to each agent $1, \dots, N$, respectively, such that the connectivity between the agents that belong to the neighboring sets $\mathcal{N}_1, \dots, \mathcal{N}_N$ is maintained, as well as each agent fulfills its individual MITL specification $\varphi_1, \dots, \varphi_N$, respectively, i.e., $x_1(t) \models \varphi_1, \dots, x_N(t) \models \varphi_N, \forall \ t \in \mathbb{R}_{\ge 0}$.
The initial positions $x_1(t_0), \dots, x_N(t_0)$ should be such that the agents which are required to remain connected for all times need to satisfy the inequality $\|x_i(t_0)-x_{i'}(t_0)\| < 2\underline{r}, i, i' \in \mathcal{V}, i \neq i'$.
It should be noted that, in this work, the dependencies between the agents are induced through the coupled dynamics and not in the discrete level, by allowing for couplings between the services (i.e., $\Sigma_i \cap \Sigma_j \ne \emptyset$, for some $i, j \in \mathcal{V}$). Hence, even though the agents do not share atomic propositions, the constraints on their motion due to the dynamic couplings and the connectivity maintenance specifications may restrict them to fulfill the desired high-level tasks. Treating additional couplings through individual atomic propositions in the discrete level is a topic of current work.
In our previous work on the multi-agent controller synthesis framework under MITL specifications [@alex_2016_acc], the multi-agent system was considered to have fully-actuated dynamics. The only constraints on the system were due to the presence of time constrained MITL formulas. In the current framework, we have two types of constraints: the constraints due to the coupling dynamics of the system , which constrain the motion of each agent, and, the timed constraints that are inherently imposed from the time bounds of the MITL formulas. Thus, there exist formulas that cannot be satisfied either due to the coupling constraints or the time constraints of the MITL formulas. These constraints, make the procedure of the controller synthesis in the discrete level substantially different and more elaborate than the corresponding multi-agent LTL frameworks in the literature ([@guo_2015_reconfiguration; @frazzoli_vehicle_routing; @belta_2010_product_system; @belta_cdc_reduced_communication]).
Proposed Solution {#sec: solution}
=================
In this section, a systematic solution to Problem \[problem: basic\_prob\] is introduced. Our overall approach builds on abstracting the system in through a WTS for each agent and exploiting the fact that the timed runs in the $i$-th WTS project onto the trajectories of agent $i$ while preserving the satisfaction of the individual MITL formulas $\varphi_i, i \in \mathcal{V}$. In particular, the following analysis is performed:
1. We propose a novel decentralized abstraction technique for the multi-agent system, i.e., discretization of the time into time steps $T$ for the given partition $D = \{D_\ell\}_{\ell \in \mathbb{I}}$, such that the motion of each agent is modeled by a WTS $\mathcal{T}_i, \ i \in \mathcal{I}$ (Section \[sec: abstration\]). We adopt here the technique of designing Nonlinear Model Predictive Controllers (NMPC), for driving the agents between neighboring regions.
2. A three-step automated procedure for controller synthesis which serves as a solution to Problem \[problem: basic\_prob\] is provided in Section \[sec: synthesis\].
3. Finally, the computational complexity of the proposed approach is discussed in Section \[sec:complexity\].
The next sections provide the proposed solution in detail.
Discrete System Abstraction {#sec: abstration}
---------------------------
In this section we provide the abstraction technique that is designed in order to capture the dynamics of each agent into WTSs. Thereafter, we work completely at discrete level, which is necessary in order to solve Problem \[problem: basic\_prob\].
(reg1) at (0,0); at (0.0, 0.0) [$\bullet$]{}; at (0.0, -0.32) [$x_i(t_k)$]{}; at (0.0, +0.33) [$P(i, k)$]{};
at (1.52, 0.87) [$\widetilde{P}(i,k,2)$]{}; (reg2) at (1.52,0.87);
at (1.52,-0.87) [$\widetilde{P}(i,k,3)$]{}; (reg3) at (1.52,-0.87);
at (0,-1.74) [$\widetilde{P}(i,k,4)$]{}; (reg4) at (0,-1.74);
at (-1.52,-0.87) [$\widetilde{P}(i,k,5)$]{}; (reg5) at (-1.52,-0.87);
at (-1.52,0.87) [$\widetilde{P}(i,k,6)$]{}; (reg5) at (-1.52,0.87);
at (0,1.74) [$\widetilde{P}(i,k,1)$]{}; (reg5) at (0,1.74);
### Workspace Geometry
Consider an enumeration $\mathbb{I}$ of the regions of the workspace, the index variable $\ell \in \mathbb{I}$ and the given time step $T$. The time step $T$ models the time duration that each agent needs to transit between two neighboring regions of the workspace. Consider also a timed sequence: $$\label{eq:time_sequence}
\mathcal{S} = \{t_0, t_1 = t_0+T, \dots, t_k = t_0+k T, \dots \}, k \in \mathbb{N}.$$ $S$ models the time stamps in which the agents are required to occupy different neighboring regions. For example, if at time $t_k$ agent $i$ occupies region $D_{\ell}$, at the next time stamp $t_k +T$ is required to occupy a neighboring region of $D_{\ell}$. The agents are always forced to change region for every different time stamp. Let us define the mapping: $$P:\mathcal{V} \times \mathbb{N}\to D,$$ which denotes the fact that the agent $i \in \mathcal{V}$, at time instant $$t_k = t_0+k T, k \in \mathbb{N},$$ occupies the region $D_{\scriptscriptstyle \ell_i} \in D$ for an index $\ell_i \in \mathbb{I}$. Define the mapping: $$\widetilde{P}: \mathcal{V} \times \mathbb{N} \times \mathbb{L} \to D.$$ where $\mathbb{L} = \{1,\dots,6\}$. By $\widetilde{P}(i, k, \widetilde{\ell}), \widetilde{\ell} \in \mathbb{L}$ we denote one and only one out of the six neighboring regions of region $P(i, k)$ that agent $i$ occupies at time $t_k$. Define also by $\bar{P}(i,k)$ the union of all the six neighboring regions of region $P(i, k)$, i.e., $$\bar{P}(i,k) = \bigcup_{\widetilde{\ell} \in \mathbb{L}} \widetilde{P}(i,k, \widetilde{\ell}),$$ with $|\bar{P}(i,k)| = 6$. An example of agent $i$ being at the region $P(i, k)$ along with its neighboring regions is depicted in Fig. \[fig:agent\_i\_widetilde\_P\].
We start by giving a graphical example for the abstraction technique that will be adopted in this work. Consider agent $i$ occupying the green region $P(i, k) = D_{\scriptscriptstyle \ell_i}$ at time $t_k = t_0+kT$ and let its neighbors $j_1, j_2$ occupying the red and blue regions $P(j_1, k) = D_{\scriptscriptstyle \ell_{j_1}}, P(j_2, k) = D_{\scriptscriptstyle \ell_{j_2}}$, respectively, as is depicted in Fig. \[fig:agent\_i\_j1\_j2\_example\]. The neighboring regions $\bar{P}(i,k), \bar{P}(j_1,k)$ and $\widetilde{P}(j_2,k, \widetilde{\ell}), \widetilde{\ell} \in \{4,5,6\}$ for agent $i, j_1, j_2$, respectively, are also depicted. All the agents start their motion at time $t_k$ simultaneously. The goal is to design a decentralized feedback control law $u_i(x_i, x_{j_1}, x_{j_2})$, that drives agent $i$ in the neighboring region $D_{\scriptscriptstyle \ell_{\text{des}}}$ exactly at time $T$, *regardless of the transitions of its neighbors to their neighboring regions*. If such controller exists, it is stored in the memory a new search for the next region is performed. This procedure is repeated for all possible neighboring regions i.e., six times, and for all the agents. For the example of Fig. \[fig:agent\_i\_j1\_j2\_example\], the procedure is performed $6^3$ times (six times for each agent). With this procedure, we are able to: 1) synchronize the agents so that each of them knows at every time step $T$ its position in the workspace as well as the region that occupies; 2) know which controller brings each agent in its desired region for any possible choice of controllers of its corresponding neighbors. We will hereafter present a formal approach of this procedure. We will hereafter present a formal approach of this procedure.
(reg01) at (0,0); at (0.0, -0.32) [$x_i(t_k)$]{}; (reg11) at (0,1.74);
(reg21) at (1.52,0.87);
(reg31) at (1.52,-0.87);
(reg41) at (0,-1.74);
(reg51) at (-1.52,-0.87);
(reg61) at (-1.52,0.87);
(reg12) at (3,1.74);
(reg12) at (4.52,0.87); at (4.52,0.45) [$x_{\scriptscriptstyle j_1}(t_k)$]{};
(reg31) at (4.52,-0.87);
(reg41) at (3,-1.74);
(reg03) at (1.52,2.60); at (1.52, 2.98) [$x_{\scriptscriptstyle j_2}(t_k)$]{};
(reg13) at (1.52,4.33);
(reg12) at (3,3.48);
(reg11) at (0,3.48);
(0.0, 0.0) – (1.52,2.60); (0.0, 0.0) – (4.52,0.87); (1.52,2.60) – (4.52,0.87);
at (0.0, 0.0) [$\bullet$]{}; at (1.52,2.60) [$\bullet$]{}; at (4.52,0.87) [$\bullet$]{}; at (-1.52,0.87) [$\bullet$]{}; at (-1.52,1.05) [$x_{i}(t_k+T)$]{}; at (-1.52, 0.40) [$D_{\scriptscriptstyle \ell_{\text{des}}}$]{}; (0.0, 0.0) .. controls (-0.2, -0.2) .. (-1.52,0.87); (0.0, -0.5) to (0.0, -1.5); at (0.0, -1.7) [$P(i,k)$]{}; at (1.80,1.90) [$P(j_2,k)$]{}; at (4.52, 1.40) [$P(j_1,k)$]{}; at (-0.5, 0.4) [$u_i$]{};
### Decentralized Controller Specification
Consider a time interval $[t_k, t_k+T]$. We state here the specifications that a decentralized feedback controller $u_i(x_i, \bar{x}_i)$ needs to guarantee so as agent $i$ to have a *well-defined transition* between two neighboring regions within the time interval $[t_k, t_k+T]$.
**(S1)** The controller needs to take into consideration the dynamics and the constraints that are imposed by the bounds of Assumption 1.
**(S2)** Agent $i$ should move from one region $P(i,k) \in D$ to a neighboring region $\widetilde{P}(i, k, \widetilde{\ell})$, without intersecting other regions, irrespectively of which region its neighbors are moving to. Thus, since the duration of the transition is $T$, it is required that $x_i(t_k) \in P(i, k)$, $x_i(t_k+T) \in \widetilde{P}(i, k, \widetilde{\ell})$ and $x_i(t) \in P(i,k) \cup \widetilde{P}(i, k, \widetilde{\ell}), t \in (t_k, t_k+T)$. The neighbors of agent $i$ will move also to exactly one of their corresponding neighboring regions.
The reason for imposing the aforementioned constraints is due to the need of imposing safety specifications to the agents. Thus, it is required to be guaranteed that the agents will not cross more than one neighboring region within the duration of a transition $T$.
### Error Dynamics
Let us define by $x_{i,k,\widetilde{\ell},\text{des}} \in \widetilde{P}(i, k, \widetilde{\ell})$ a reference point of the desired region $\widetilde{P}(i, k, \ell)$ which agent $i$ needs to occupy at time $t_k+T$. Define also by: $$e_i(t) = x_i(t)-x_{i, k, \widetilde{\ell}, \text{des}}, t \in [t_k, t_k+T],$$ the error which the controller $u_i$ needs to guarantee to become zero in the time interval $t \in [t_k, t_k+T]$. Then, the *nominal error dynamics* are given by: $$\label{eq:error_dynamics}
\dot{e}_i(t) = g_i(e_i(t), \bar{x}_i(t), u_i(t)), t \in [t_k, t_k+T],$$ with initial condition $e_i(t_k) = x_i(t_k)-x_{i, k, \widetilde{\ell}, \text{des}}$, where: $$g_i(e_i(t), \bar{x}_i(t), u_i(t)) = f_i(e_i(t)+x_{i, k, \widetilde{\ell}, \text{des}}, \bar{x}_i(t))+u_i(t).$$
\[property:error\_bound\] According to Assumption , at every time $s \in [t_k, t_k+T]$, with $t_k = t_0+ k T$, the error $e_i(s)$ of the state of agent $i$ is upper bounded by: $$\label{eq:error_bound}
\|e_i(s)\| \le \|e_i(t_k)\|+ (s-t_k)(M+u_{\max}), i \in \mathcal{V}.$$
The proof can be found in Appendix \[app:proof\_of\_property\_1\].
### State Constraints
Before defining the ROCP we state here the state constraints that are imposed to the state of each agent. Define the set: $$\begin{aligned}
& X_i = \{x_i \in W, \bar{x}_i \in W^{N_i}: \notag \\
&\hspace{3mm} \|f_i(x_i, \bar{x}_i)\| \le M, \|x_i-x_j\| < \underline{r}, \forall j \in \mathcal{N}_i(0), \notag \\
&\hspace{3mm} x_i \in P(i,k) \cup \widetilde{P}(i,k,\widetilde{\ell}), \widetilde{\ell} \in \mathbb{L} \}, \notag\end{aligned}$$ as the set that captures the state constraints of agent $i$. The first constraint in the set $X$ stands for the bound of Assumption \[ass:dynamic\_control\_bounds\]; the second one stands for the connectivity requirement of agent $i$ with all its neighbors; the last one stands for the requirement each agent to transit from one region to exactly one desired neighboring region. In order to translate the constraints that are dictated for the state $x_i(t)$ into constraints regarding the error state $e_i(t)$ from , define the set $E_i = X_i \oplus (-x_{i,k,\widetilde{\ell},\text{des}}).$ Then, the following implication holds: $x_i \in X_i \Rightarrow e_i \in E_i.$
### Control Design
This subsection concerns the control design regarding the transition of agent $i$ to one neighboring region $\widetilde{P}(i, k, \widetilde{\ell})$, for some $\widetilde{\ell}\in\mathbb{L}$. The abstraction design, however, concerns all the neighboring regions $\bar{P}(i,k)$, for which we will discuss in the next subsection.
The timed sequence $\mathcal{S}$ consists of intervals of duration $T$. Within every time interval $[t_k, t_k+T]$, each agent needs to be at time $t_k$ in region $P(i,k)$ and at time $t_k+T$ in a neighboring region $\widetilde{P}(i,k,\widetilde{\ell}), \widetilde{\ell} \in \mathbb{L}$. We assume that $T$ is related to the sampling time $h$ according to: $T = m h, m \in \mathbb{N}$. Therefore, within the time interval $[t_k, t_k+T]$, there exists $m+1$ sampling times. By introducing the notation $t_{\scriptscriptstyle k_{z}} \triangleq t_k+z h, \forall z \in \mathbb{M} \triangleq \{0, \dots, m\}$, we denote by $\{t_{\scriptscriptstyle k_{z}}\}_{z \in \mathbb{M}}$ the sampling sequence within the interval $[t_k, t_k+T]$. Note that $t_{\scriptscriptstyle k_{0}} = t_k$ and $t_{\scriptscriptstyle k_{m}} = t_k+T$. The indexes $k, z$ stands for the interval and for the sampling times within this interval, respectively. As it will be presented hereafter, at every sampling time $t_{\scriptscriptstyle k_{z}}, z \in \mathbb{M}$, each agents solves a ROCP.
Our control design approach is based on Nonlinear Model Predictive Control (NMPC). NMPC has been proven to be efficient for systems with nonlinearities and state/input constraints. For details about NMPC we refer the reader to [@morrari_npmpc; @cannon_2001_nmpc; @frank_2003_nmpc_bible; @frank_1998_quasi_infinite; @frank_2003_towards_sampled-data-nmpc; @grune_2011_nonlinear_mpc; @camacho_2007_nmpc; @borrelli_2013_nmpc; @fontes_2001_nmpc_stability; @camacho_2002_input_to_state]. We propose here a sampled-data NMPC with decreasing horizon in order to design a controller that respects the desired specifications and guarantees the transition between regions at time $T$. In the proposed sampled-data NMPC, an open-loop Robust Optimal Control Problem (ROCP) is solved at every discrete sampling time instant $t_{\scriptscriptstyle k_z}, z \in \mathbb{M}$ based on the current error state information $e_i(t_{\scriptscriptstyle k_z})$. The solution is an optimal control signal $\hat{u}_{i}(t)$, for $t \in [t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z}]$, where $T_{\scriptscriptstyle z}$ is defined as follows.
A *decreasing horizon policy* is defined by: $$\label{eq:decr_horizon}
T_{z} = T - z h, z \in \mathbb{M}.$$
This means that at every time sample $t_{\scriptscriptstyle k_z}$ in which the ROCP is solved, the horizon is decreased by a sampling time $h$. The specific policy is adopted in order to enforce the controllers $u_i$ to guarantee that agent $i$ will reach the desired neighboring region at time $T$. implies also that $t_{\scriptscriptstyle k_{z}} +T_z = t_k+T, \forall z \in \mathbb{M}$ A graphical illustration of the presented time sequences is given in Fig. \[fig:time\_sequence\].
The *open-loop input signal* is applied in between the sampling instants and is given by the solution of the following Robust Optimal Control Problem (ROCP): $\mathcal{O}(k, x_i(t), \bar{x}_i(t), P(i,k), \widetilde{\ell}, x_{i, k, \widetilde{\ell}, \text{des}}), t \in [t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z}],$ which is defined as:
$$\begin{aligned}
&\hspace{-1mm}\min\limits_{\hat{u}_i(\cdot)} J_i(e_i(t_{\scriptscriptstyle k_z})),\hat{u}_i(\cdot)) = \notag \\
&\hspace{-1mm} \min\limits_{\hat{u}_i(\cdot)} \left\{ V_i(\hat{e}_i(t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z})) + \int_{t_{\scriptscriptstyle k_z}}^{t_{\scriptscriptstyle k_z}+T_{z}} \Big[ F_i(\hat{e}_i(s), \hat{u}_i(s)) \Big] ds \right\} \label{eq:mpc_minimazation} \\
&\hspace{-1mm}\text{subject to:} \notag \\
&\hspace{1mm} \dot{\hat{e}}_i(s) = g_i(\hat{e}_i(s), \hat{\bar{x}}_i(s), \hat{u}_i(s)), \hat{e}_i(t_{\scriptscriptstyle k_z}) = e_i(t_{\scriptscriptstyle k_z}), \label{eq:diff_mpc}
\end{aligned}$$
$$\begin{aligned}
&\hspace{1mm} \hat{e}_i(s) \in E_{s-t_{\scriptscriptstyle k_z}}^{i}, \hat{u}_i(s) \in \mathcal{U}_i, s \in [t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z}], \label{eq:mpc_constrained_set} \\
&\hspace{1mm} \hat{e}_i(t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z})\in\mathcal{E}_i. \label{eq:mpc_terminal_set}
\end{aligned}$$
The ROCP has as inputs the terms $k, x_i(t)$, $\bar{x}_i(t), P(i,k)$, $\widetilde{\ell}$, $x_{i, k, \widetilde{\ell}, \text{des}}$, for time $t \in [t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z}]$. We will explain hereafter all the terms appearing in the ROCP problem -. By hat $\hat{(\cdot)}$ we denote the predicted variables (internal to the controller), corresponding to the system i.e., $\hat{e}_i(\cdot)$ is the solution of driven by the control input $\hat{u}_{i}(\cdot): [t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z}] \to \mathcal{U}_i$ with initial condition $\hat{e}_i(t_{\scriptscriptstyle k_z}) = e_i(t_{\scriptscriptstyle k_z})$. The set $E_{s-t_{\scriptscriptstyle k_z}}^{i}$ is a subset of $E_i$ and will be explicitly defined later.
(-0.5, 0 ) – (7.5, 0);
(-0.5,0) circle (2.5 pt); (1.6,0) circle (2.5 pt); (5.5,0) circle (2.5 pt); (7.5,0) circle (2.5 pt);
at (-0.3, -0.4) [$t_{\scriptscriptstyle k_z}$]{}; at (1.5, -0.4) [$t_{\scriptscriptstyle k_{z+1}}$]{}; at (5.5, -0.4) [$t_{\scriptscriptstyle k_{z}}+ T_{\scriptscriptstyle z+1}$]{}; at (7.3, -0.4) [$t_{\scriptscriptstyle k_{z}}+ T_{\scriptscriptstyle z}$]{}; (-0.5, 0.4) to (1.6, 0.4); (1.6, 0.4) to (-0.5, 0.4); at (0.6, 0.8) [$h$]{}; (5.5, 0.4) to (7.5, 0.4); (7.5, 0.4) to (5.5, 0.4); at (6.5, 0.8) [$h$]{}; (-0.5, -1.0) to (7.5, -1.0); (7.5, -1.0) to (-0.5, -1.0); at (3.8, -1.3) [$T_{\scriptscriptstyle z} = T$]{};
In sampled-data NMPC bibliography an ROCP is defined over the time interval $s \in \{t_i, t_{i+1} = t_i + h, \dots , t_i+T\}$, where $T$ is the prediction horizon. Due to the fact that we have denoted by $i$ the agents, and the fact that the ROCP is solved for every time interval, we use the notation $s \in \{t_{\scriptscriptstyle k_z} = t_k, t_{\scriptscriptstyle k_{z+1}} = t_k + h, \dots, t_{\scriptscriptstyle k_z} + T_{\scriptscriptstyle z} = t_{\scriptscriptstyle k_z} +T\}$, instead. The indexes $k, z$ stands for the interval and for the sampling time, respectively. A graphical illustration of the presented time sequence is given in Fig. \[fig:time\_sequence\].
Note that the predicted values are not the same with the actual closed-loop values due to the fact that agent $i$, can not know the estimation of the trajectories of its neighbors $\hat{\bar{x}}$, within a predicted horizon. Thus, the term $\hat{\bar{x}}$ is treated as a disturbance to the *nominal system* .
The term $F_i: E_i \times \mathcal{U}_i \to \mathbb{R}_{\ge 0}$, stands for the *running cost*, and is chosen as: $$F_i(e_i, u_i) = e_i^\top Q_i e_i + u_i^\top R_i u_i,$$ where $Q_i=\text{diag}\{q_{i_1},q_{i_2}\}, R_i=\text{diag}\{\xi_{i_1},\xi_{i_2}\}$, with $q_{i_\zeta}\in\mathbb{R}_{\ge 0}, \xi_{i_\zeta}\in\mathbb{R}_{>0}, \zeta\in\{1,2\} $. For the running cost, it holds that $F_i(0,0) = 0$, as well as: $$\begin{aligned}
\label{eq:F_lower_bound}
\underline{m}_i \|e_i\|^2 \leq F_i(e_i, u_i) \leq \bar{m}_i \|e_i\|^2,\end{aligned}$$ where $\underline{m}_i, \bar{m}_i$ will be defined later. Note that $\underline{m}_i\|e_i\|^2$ is $\mathcal{K}$ function, according to Definition \[def:class\_K\].
The running cost function $F_i(e_i, u_i)$ is Lipschitz continuous in $E_i \times \mathcal{U}_i$, with Lipschitz constant: $$L_{\scriptscriptstyle F_i} = 2\bar{\varepsilon}_i \sigma_{\max}(Q_i),$$ where: $$\bar{\varepsilon}_i = \sup_{e_i \in E_i} \{\|e_i\|\},$$ for all $e_i \in E_i, u_i \in \mathcal{U}_i$.
The proof can be found in Appendix \[app:lemma\_1\].
Note that, according to , the terms $\|e_i\|$ are bounded, for all $i \in \mathcal{V}$. The terms $V_i: E_i \to \mathbb{R}_{ > 0}$ and $\mathcal{E}_i \subseteq E_i$ are the *terminal penalty cost* and *terminal set*, respectively, and are used to enforce the stability of the system. The terminal cost is given by: $$V_i(e_i(t)) = e_i(t)^\top P_i e_i(t).$$ where $P_i=\text{diag}\{p_{i_1},p_{i_2}\}$, with $p_{i_\zeta}\in\mathbb{R}_{> 0}, \zeta\in\{1,2\}$. We choose $\underline{m}_i = \{q_{i_1},q_{i_2}, \xi_{i_1}, \xi_{i_2}\}$ and $\bar{m}_i = \{q_{i_1},q_{i_2}, \xi_{i_1}, \xi_{i_2}\}$.
The solution of the nominal model at time $s \in [t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z}]$, starting at time $t_{\scriptscriptstyle k_z}$ from an initial condition $e_i(t_{\scriptscriptstyle k_z})$, applying a control input $u_{i}: [t_{\scriptscriptstyle k_z}, s] \to \mathcal{U}_i$ is denoted by: $$e_i(s; u_i(\cdot), e_i(t_{\scriptscriptstyle k_z})), s \in [t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z}].$$ The predicted state of the system at time $s \in [t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z}]$ is denoted by: $$\hat{e}_i(s; u_i(\cdot), e_i(t_{\scriptscriptstyle k_z})), s \in [t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z}],$$ and it is based on the measurement of the state $e_i(t_{\scriptscriptstyle k_z})$ at time $t_{\scriptscriptstyle k_z}$, when a control input $u_{i}(\cdot; e_i(t_{\scriptscriptstyle k_z}))$ is applied to the system for the time period $[t_{\scriptscriptstyle k_z}, s]$. Thus, it holds that: $$\label{eq:predicted_state_relation}
e_i(s) = \hat{e}_i(s; u_i(\cdot), e_i(s)), s \in [t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z}].$$
The state measurement enters the system via the initial condition of at the sampling instant, i.e. the system model used to predict the future system behavior is initialized by the actual system state. The solution of the ROCP - at time $t_{\scriptscriptstyle k_z}$ provides an optimal control input denoted by $\hat{u}^\star_{i}(t; e(t_{\scriptscriptstyle k_z}))$, for $t \in [t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z}]$. It defines the open-loop input that is applied to the system until the next sampling instant $t_{\scriptscriptstyle k_{z+1}}$: $$\label{eq:control_input_star}
u_{i}(t; e_i(t_i)) = \hat{u}^\star_{i}(t_{\scriptscriptstyle k_z}; e_i(t_{\scriptscriptstyle k_z})), t \in [t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_{z+1}}).$$ The corresponding *optimal value function* is given by: $$\label{eq:J_star}
J_{i}^\star(e_i(t_{\scriptscriptstyle k_z})) \triangleq J_{i}(e_i(t_{\scriptscriptstyle k_z}), \hat{u}_{i}^\star(\cdot; e_i(t_{\scriptscriptstyle k_z}))).$$ with $J_i(\cdot)$ as is given in . The control input $u_i(t; e_i(t_{\scriptscriptstyle k_z}))$ is of the feedback form, since it is recalculated at each sampling instant using the new state information. Define an admissible control input as:
\[def:admissible\_control\_input\] A control input $u_i: [t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z}] \to \mathbb{R}^{2}$ for a state $e(t_{\scriptscriptstyle k_z})$ is called *admissible*, if all the following hold:
1. $u_i(\cdot)$ is piecewise continuous;
2. $u_i(s) \in \mathcal{U}_i, \forall \ s \in [t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z}]$;
3. $e_i(s; u_i(\cdot), e(t_{\scriptscriptstyle k_z})) \in E_i, \forall \ s \in [t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z}]$;
4. $e_i(T_{\scriptscriptstyle z}; u_i(\cdot), e(t_{\scriptscriptstyle k_z})) \in \mathcal{E}_i$;
(reg1) at (0,0); at (0.0, 0.0) [$\bullet$]{}; (reg2) at (1.52,0.87);
(reg3) at (1.52,-0.87);
(reg4) at (0,-1.74);
(reg5) at (-1.52,-0.87);
(reg5) at (-1.52,0.87);
(reg5) at (0,1.74);
(0,0) circle (0.87cm); (1.52,0.87) circle (0.87cm); (-0.0, -0.0) – (-0.70, -0.55); (0.0, 0.0) .. controls (0.3, -0.5) .. (1.70,1.2);
at (0.0, 0.0) [$\bullet$]{}; at (0.2, +0.32) [$x_j(t_k)$]{}; at (2.4, 0.8) [$x_j(t_k+T)$]{}; at (-0.55, -0.20) [$\underline{r}_{\scriptscriptstyle h}$]{};
For the given hexagonal regions with side length $R$, the radius of the inscribed circle is given by $\underline{r}_{\scriptscriptstyle h} = \frac{\sqrt{3}}{2}R$ (two inscribed circles for the given regions are depicted with orange in Fig. \[fig:largest\_distance\]). Thus, according to Fig. \[fig:largest\_distance\], an upper bound of the norm of differences between the actual position $x_j$ and the estimated position $\hat{x}_j$ of the agent’s $i$ neighbors states, is given by: $$\label{eq:bound_geitonon}
\|x_j-\hat{x}_j\| \le 4 \underline{r}_{\scriptscriptstyle h} = 2 \sqrt{3} R, j \in \mathcal{N}_i,$$ due to the fact that each agent can transit at most to a neighboring region, according to the constraint set $X_i$.
\[lemma:e-hat\_e\] In view of Assumptions \[ass:dynamic\_control\_bounds\], \[ass:lipsitch\_f\_x\_bar\_x\], the difference between the actual measurement $e_i(s) = e_i(s; u_i(s; e_i(t_{\scriptscriptstyle k_z})), e_i(t_{\scriptscriptstyle k_z}))$ at time $s \in [t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z}]$ and the predicted state $\hat{e}_i(s; u_i(s; e_i(t_{\scriptscriptstyle k_z})), e_i(t_{\scriptscriptstyle k_z}))$ at the same time under the same control law $u_i(s; e_i(t_{\scriptscriptstyle k_z}))$, starting at the same initial state $e_i(t_{\scriptscriptstyle k_z})$, is upper bounded by: $$\begin{aligned}
&\|e_i(s) - \hat{e}_i(s; u_i(s; e_i(t_{\scriptscriptstyle k_z})), e_i(t_{\scriptscriptstyle k_z}))\| \le \rho_i(s-t_{\scriptscriptstyle k_z}),
\end{aligned}$$ where $\rho_i: \mathbb{R}_{\ge 0} \to \mathbb{R}$, with: $$\begin{aligned}
&\rho_i(y) = \min \Big\{ \widetilde{\rho}_i \left[ e^{L_i y}- 1 \right], \notag \\
&\hspace{15mm} 2 \|e_i(t_{\scriptscriptstyle k_z})\|+ 2 y(M+u_{\max}) \Big\},
\end{aligned}$$ and $$\widetilde{\rho}_i = \frac{2 \sqrt{3} R \bar{L}_i N_i}{L_i}.$$
The proof can be found in Appendix \[app:lemma\_e\_ehat\].
The satisfaction of the constraint on the state along the prediction horizon depends on the future evolution of the neighboring agents trajectories. Under Assumptions , of Lipschitz continuity and bounds of the nominal model, respectively, it is possible to compute a bound on the future effect of the disturbance on the system as is given by Lemma \[lemma:e-hat\_e\]. Then, by considering this effect on the state constraint on the nominal prediction, it is possible to guarantee that the evolution of the real state of the system will be admissible all the time. In view of latter, the state constraint set $E$ of the standard NMPC formulation, is being replaced by a restricted constrained set $E_{s-t_{\scriptscriptstyle k_z}}^{i} \subseteq E_i$ in . This state constraint’s tightening for the nominal system is a key ingredient of the robust NMPC controller and guarantees that the evolution of the real system will be admissible. Authors in [@camacho_2002_input_to_state; @alina_ecc_2011] has considered such a Robust NMPC formulation. The restricted constrained set is then defined as $E_{s-t_{\scriptscriptstyle k_z}}^{i}= E_i \sim B^i_{s-t_{\scriptscriptstyle k_{z}}}$, where: $$\begin{aligned}
&B^i_{s-t_{\scriptscriptstyle k_{z}}} = \notag \\
&\left\{e_i \in \mathbb{R}^{2}: \|e_i(s)\| \le \rho_i(s-t_i)\right\}, s \in [t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z}]. \notag\end{aligned}$$
For every $s \in [t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z}]$, we have that if $$\hat{e}_i(s; u_i(s; e(t_{\scriptscriptstyle k_z})), e_i(t_{\scriptscriptstyle k_z})) \in E^i_{s-t_{\scriptscriptstyle k_{z}}} = E_i \sim B^i_{s-t_{\scriptscriptstyle k_{z}}} \subseteq E_i,$$ then the real state satisfies the constraint $E_i$, i.e., $e_i(s) \in E_i.$
The proof can be found in Appendix \[app:property\_tighthened\_sets\].
For the feasibility and convergence proofs of the ROCP the following assumptions are required.
\[ass:admissible\_u\_f\] Assume that there exists a *local stabilizing controller* $u_{f,i} = \kappa_i(e_i) \in \mathcal{U}_i$ satisfying: $$\label{eq:admiss_controller}
\displaystyle \frac{\partial V_i}{\partial{e_i}} \left[g_i(e_i, \bar{x}_i, \kappa_i(e_i))\right] + F_i(e_i, \kappa_i(e_i)) \le 0, \forall \ e_i \in \Phi_i,$$ where $\Phi_i$ is a set given by: $$\Phi_i \triangleq \{e_i \in \mathbb{R}^{2}: V_i(e_i) \le \alpha_{1,i}\}, \alpha_{1,i} > 0,$$ such that: $$\Phi_i \subseteq \mathbb{E}_i \triangleq \{e_i \in E^i_{\scriptscriptstyle T_z}: \kappa_i(e_i) \in \mathcal{U}_i\},$$ where $E^i_{\scriptscriptstyle T_z} = E_i \sim B^i_{\scriptscriptstyle T_z}.$
\[eq:lemma\_L\_v\] The terminal penalty function $V_i(\cdot)$ is Lipschitz in $\Phi_i$, with Lipschitz constant $L_{\scriptscriptstyle V, i} = 2 \sigma_{\max}(P_i) \sqrt{ \frac{\alpha_{1,i}}{\lambda_{\min}(P_i)}}$, for all $e_i(t) \in \Phi_i$.
The proof can be found in Appendix \[app:lemma\_lip\_L\_v\].
Once the set $\Phi_i$ is computed, the terminal constraint set $\mathcal{E}_i$ is given by the following. Supposing that Assumption \[ass:admissible\_u\_f\] holds. Then, by choosing: $\mathcal{E}_i = \{e_i \in \mathbb{R}^2: \|e_i\| \le \sqrt{\frac{\alpha_{2,i}}{\lambda_{\min}(P_i)}} < r_{\scriptscriptstyle h}\}$, with $\alpha_{2,i} \in (0, \alpha_{1,i})$, we guarantee the following: $1)$ $\mathcal{E}_i \subseteq \widetilde{P}(i, k, \widetilde{\ell})$, i.e. the terminal set is a subset of the desired neighboring region; $2)$ for all $e_i \in \Phi_i$ it holds that $g_i(e_i, \kappa_i(e_i)) \in \mathcal{E}_i$.
The following two lemmas are required in order to prove the basic Theorem or this paper.
\[lemma:x\_y\_proof\] Let $s \ge t_{\scriptscriptstyle k_{z+1}}$, $x \in E^i_{s-t_{\scriptscriptstyle k_{z}}}$ and $y \in \mathbb{R}^{2}$ such that: $\|x-y\| \le \rho_i(t_{\scriptscriptstyle k_{z+1}}-t_{\scriptscriptstyle k_{z}}) = \rho_i(h)$, as $\rho_i$ is given in Lemma \[lemma:e-hat\_e\]. Then, it holds that $y \in E^i_{s-t_{\scriptscriptstyle k_{z+1}}}$.
The proof can be found in Appendix \[app:y\_E\_t\_i\_plus\_1\].
\[lemma:bounded\_trajectories\] Let $s \ge t_{\scriptscriptstyle k_{z}}$. The difference between two estimated trajectories $\hat{e}_i(s; u_i(\cdot), e_i(t_{\scriptscriptstyle k_{z+1}})), \hat{e}_i(s; u_i(\cdot), e_i(t_{\scriptscriptstyle k_{z}}))$ at time $s$, starting from from initial points $t_{\scriptscriptstyle k_{z+1}}$, $t_{\scriptscriptstyle k_{z}}$, respectively, under the same control input $u_i(\cdot)$, is upper bounded by: $$\begin{aligned}
&\|\hat{e}_i(s; u_i(\cdot), e_i(t_{\scriptscriptstyle k_{z+1}}))- \hat{e}_i(s; u_i(\cdot), e_i(t_{\scriptscriptstyle k_{z}}))\| \le \notag \\
&\hspace{40mm}\rho_i(t_{\scriptscriptstyle k_{z+1}}-t_{\scriptscriptstyle k_{z}}) = \rho_i(h).
\end{aligned}$$
The proof can be found in Appendix \[app:bounded\_trajectories\].
Suppose that Assumptions \[ass:measurement\_assumption\]-\[ass:admissible\_u\_f\] hold. If the ROCP is feasible at time $t_k$, then, the closed loop system of agent $i$, under the control input , starting its motion at time $t_k = t_0 + k T$ from region $P(i,k)$, is Input to State Stable (ISS) (for ISS see [@sontag_2008_ISS]) and its trajectory converges to the admissible positively invariant terminal set $\mathcal{E}_i$ exactly at time $t_k+T$, if it holds that $\rho_i(T_{\scriptscriptstyle z}) \le \bar{\rho}_i \triangleq \frac{\alpha_{1,1}-\alpha_{2,i}}{L_{\scriptscriptstyle V_i}}$.
The proof consists of two parts: in the first part it is established that initial feasibility implies feasibility afterwards. Based on this result it is then shown that the error $e_i(t)$ converges to the terminal set $\mathcal{E}_i$. The *feasibility analysis* as well as the *convergence analysis* can be found in Appendix \[app:feasibility\_convergence\].
Assumption \[ass:admissible\_u\_f\] is common in the NMPC literature. Many methodologies on how to compute $\Phi_i$ and controllers $u_{f,i} = \kappa_i(e_i)$, if they exist, have been proposed. We refer the reader to [@frank_1998_quasi_infinite; @mayne_2000_nmpc]. Regarding the initial feasibility, numerical tools (e.g. [@grune_2011_nonlinear_mpc]) can be utilized in order to solve the ROCP and check if the problem is feasible at time $t_k = t_{\scriptscriptstyle k_{z}}$.
The term $\bar{\rho}_i, i \in \mathcal{V}$ gives an upper bound on the deviation of the trajectories of the neighboring agents of agent $i$ from their real values. If this bound is satisfied, agent $i$ can transit between the corresponding two neighboring regions, provided the ROCP is feasible at $t_{\scriptscriptstyle k_z}$.
It should be noted that, due to the nonlinear coupling terms $f_i(x_i, \bar{x}_i)$, the desired connectivity specifications and the bounds of Assumption \[ass:dynamic\_control\_bounds\], some of the ROCPs for $k \in N$ might not have a feasible solution. Let $i' \in \mathcal{V}, k' \in \mathbb{N}, \widetilde{\ell}' \in \mathbb{L}$ represent an agent $i'$ that at time step $t_{k'} = t_0 + k'T$ is desired to transit from region $P(i', k')$ to region $\widetilde{P}(i', k', \widetilde{\ell}')$. If the ROCP $\mathcal{O}(k', x_{i'}(t), \bar{x}_{i'}(t), P(i',k'), \widetilde{\ell}', x_{i', k', \widetilde{\ell}', \text{des}}), t \in [t_{\scriptscriptstyle k'_{z}}, t_{\scriptscriptstyle k'_{z}}+T_{z}],$ has no solution, then there does not exist admissible controller that can drive agent $i'$ from $P(i', k')$ to region $\widetilde{P}(i', k', \widetilde{\ell}')$. Our goal, through the proposed approach, is to seek all the possible solutions of the ROCP, which implies to seek for all possible transitions that will form later the individual WTS $\mathcal{T}_i$ of each agent. In this way, the resulting WTS $\mathcal{T}_i$ will capture the coupling dynamics and the transition possibilities of agent $i$ in the best possible way.
### Generating the WTSs
Each agent $i \in \mathcal{V}$ solves the ROCP \[eq:mpc\_minimazation\]-\[eq:mpc\_terminal\_set\] for every time interval $[t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z}], k \in \mathbb{N}$, for all the desired neighboring regions $\widetilde{P}(i,k,\widetilde{\ell}), \widetilde{\ell} \in \mathbb{L}$. This procedure is performed by off-line simulation, i.e., at each sampling time $t_{\scriptscriptstyle k_z}, z \in \mathbb{M}$, each agent exchanges information about its new state with its neighbors and simulates the dynamics . Between the sampling times the estimation $\hat{\bar{x}}_i$ is considered to be a disturbance, as discussed earlier.
Algorithm 1 provides the off-line procedure in order to generate the transition relation for each agent. At time $t_0$ each agent $i$ calls the algorithm in order to compute all possible admissible controllers to all possible neighboring regions of the workspace. The term $\text{Transit}$, which is the output of the algorithm, is a matrix of control input sequences for all pairs of neighboring regions in the workspace, initialized at sequences of zeros. The function $\text{Point}2\text{Region}(\cdot)$ maps the point $x_i(t_k)$ to the corresponding region of the workspace. The function $\text{Sampling}(\cdot)$ takes as input the interval $[t_k, t_k+T]$ and returns the $m+1$ samples of this interval. The notation $(u_i^\star)_{\scriptstyle k_z}$ stands for the $z$-th element of the vector $(u_i^\star)$. The function $\text{OptSolve}(k,x_i(t),\bar{x}_i(t),p,\widetilde{\ell})$ (i) solves the ROCP and the function $\text{UpdateStates}(x_i,\bar{x}_i)$ updates the states of agent $i$ and its neighbors after every sampling time. If the OptSolve function does not return a solution, then there does not exist an admissible control input that can drive agent $i$ to the desired neighboring region. After utilizing Algorithm 1, the WTS of each agent is defined as follows:
**Input:** $i, x_i(t_0), \bar{x}_i(t_0);$ **Output:** $\text{Transit}$; Matrix with regions$\backslash$control inputs;\
$\text{Transit} \leftarrow \text{zeros}(|\mathbb{I}|, 6)$; k = 0; Flag = False; $\text{List} \leftarrow \{\text{Point}2\text{Region}(x_i(t_0))\}$; Initialize p is a region of the List; $t \leftarrow \text{Sampling} (t_k, t_k+T)$; $(u^\star_i)_{\scriptscriptstyle k_z} \leftarrow \text{OptSolve}(k, x_i(t), \bar{x}_i(t), p, \widetilde{\ell});$ UpdateStates($x_i, \bar{x}_{i}$); $\nexists$ controller; Flag = True; search next region; break; $u^\star_i \leftarrow \{(u^\star_i)_{\scriptscriptstyle k_z} \}_{z \in \mathbb{M}}$ $u_i$ found $\text{Transit}(p, \widetilde{\ell})\leftarrow u^\star_i$; $\text{List} \leftarrow \text{List} \cup \widetilde{P}(i,k,\widetilde{\ell})$ Flag = False; $\text{List} \leftarrow \text{List} \backslash p$; $k = k +1;$
\[alg:basic\_algorithm\]
\[def: indiv\_WTS\] The motion of each agent $i \in \mathcal{V}$ in the workspace is modeled by the WTS $\ \mathcal{T}_i$ $= (S_i$, $S_i^{\text{init}}$, $Act_i$, $\longrightarrow_i$, $d_i$, $\Sigma_i$, $L_i)$ where: $S_i = \{D_\ell\}_{\ell \in \mathbb{I}}$ is the set of states of each agent; $S_i^{\text{init}} = P(i,0) \subseteq S_i$ is a set of initial states defined by the agents’ initial positions $x_i(t_0) \in P(i,0)$ in the workspace; $Act_i$ is the set of actions containing the union of all the admissible control inputs $u_i \in \mathcal{U}_i$ that are a feasible solution to the ROCP and can drive agent $i$ between neighboring regions; $\longrightarrow_i \subseteq S_i \times Act_i \times S_i$ is the transition relation. We say that $(P(i,k), u_i, \widetilde{P}(i,k,\widetilde{\ell})) \in \longrightarrow_i$, $k \in \mathbb{N}, \widetilde{\ell} \in \mathbb{L}$ if there exist an admissible controller $u_i \in Act_i$ which at step $k$ drives the agent $i$ from the region $P(i,k)$ to the desired region $\widetilde{P}(i,k,\widetilde{\ell})$. Algorithm 1 gives the steps how the transition relation can be constructed. $d_i: \longrightarrow_i \rightarrow \mathbb{R}_{\ge 0}$, is a map that assigns a positive weight (duration) to each transition. The duration of each transition is exactly equal to $T$; $\Sigma_i$, is the set of atomic propositions; $L_i: S_i \rightarrow 2^{\Sigma_i}$, is the labeling function.
The individual WTSs of the agents will allow us to work directly in the discrete level and design sequences of controllers that solve Problem \[problem: basic\_prob\]. Every WTS $\mathcal{T}_i, i \in \mathcal{V}$ generates timed runs and timed words of the form $r_i^t = (r_i(0), \tau_i(0))$ $(r_i(1), \tau_i(1))\ldots$, $w_i^t$ $= (w_i(0), \tau_i(0))$ $(w_i(1), \tau_i(1))\ldots$, respectively, over the set $2^{\Sigma_i}$ with $w_i(\mu) = L_i(r_i(\mu)), \tau_i(\mu) = \mu T, \forall \ \mu \ge 0$. The transition relation $\longrightarrow_i$ along with the output of the Algorithm 1, i.e, $\text{Transit}(\cdot)$, allows each agent to have all the necessary information in order to be able to make a decentralized plan in the discrete level that is presented hereafter. The relation between the timed words that are generated by the WTSs $\mathcal{T}_i, i \in \mathcal{V}$ with the timed service words produced by the trajectories $x_i(t), i \in \mathcal{V}, t \ge 0$ is provided through the following remark:
\[lemma:compliant\_WTS\_runs\_with\_trajectories\] By construction, each timed word produced by the WTS $\mathcal{T}_i$ is a *relaxed timed word* associated with the trajectory $x_i(t)$ of the system . Hence, if we find a timed word of $\mathcal{T}_i$ satisfying a formula $\varphi_i$ given in MITL, we also find for each agent $i$ a desired timed word of the original system, and hence trajectories $x_i(t)$ that are a solution to the Problem \[problem: basic\_prob\]. Therefore, the produced timed words of $\mathcal{T}_i$ are compliant with the relaxed timed words of the trajectories $x_i(t)$.
Controller Synthesis {#sec: synthesis}
--------------------
(-15.6, 6.0) rectangle +(2.3, 0.9); at (-14.45, 6.45) [$\text{MITL2TBA}$]{}; (-15.6, 0.80) rectangle +(2.3, 0.9); at (-14.45, 1.25) [$\text{MITL2TBA}$]{};
(-14.5,6.0) – (-14.5,4.6); (-14.5,7.6) – (-14.5,6.9); (-14.5,0.1) – (-14.5,0.8); (-14.5,1.7) – (-14.5,3.1);
at (-14.5, 7.80) [$\varphi_1$]{}; at (-14.5, -0.1) [$\varphi_N$]{};
at (-14.0, 5.35) [$\mathcal{A}_1$]{}; at (-14.0, 2.40) [$\mathcal{A}_N$]{};
at (-14.5, 3.25) [$\otimes$]{}; at (-14.5, 4.45) [$\otimes$]{};
at (-14.5, 3.95) [$\vdots$]{};
(-14.30,4.45) – (-12.6,4.45); (-14.30,3.25) – (-12.6,3.25);
at (-12.3, 4.47) [$\widetilde{\mathcal{T}}_1$]{}; at (-12.3, 3.95) [$\vdots$]{}; at (-12.3, 3.26) [$\widetilde{\mathcal{T}}_N$]{};
(-16.35,4.45) – (-14.7,4.45); (-16.35,3.25) – (-14.7,3.25);
at (-16.7, 4.47) [$\mathcal{T}_1$]{}; at (-16.7, 3.95) [$\vdots$]{}; at (-16.7, 3.26) [$\mathcal{T}_N$]{};
(-12.00,4.45) – (-11.3,4.45); (-12.00,3.25) – (-11.3,3.25);
(-11.3, 2.85) rectangle +(1.85, 0.70); (-11.3, 4.15) rectangle +(1.85, 0.70); at (-10.32, 4.50) [$\text{synthesis}$]{}; at (-10.32, 3.20) [$\text{synthesis}$]{};
(-9.35,4.45) – (-8.65,4.45); (-9.35,3.25) – (-8.65,3.25);
at (-8.35, 4.47) [$\widetilde{r}_1^t$]{}; at (-8.35, 3.95) [$\vdots$]{}; at (-8.35, 3.26) [$\widetilde{r}_N^t$]{};
(-17.70,4.47) – (-17.00,4.47); (-17.70,3.26) – (-17.00,3.26);
(-19.6, 2.90) rectangle +(1.8, 0.7); (-19.6, 4.20) rectangle +(1.8, 0.7); at (-18.70, 4.60) [$\text{abstraction}$]{}; at (-18.70, 3.30) [$\text{abstraction}$]{};
(-20.35,4.47) – (-19.65,4.47); (-20.35,3.26) – (-19.65,3.26);
at (-22.70, 4.47) [$\displaystyle \dot{x}_1 = f_1(x_1, \bar{x}_1)+u_{1}$]{}; at (-22.70, 3.95) [$\vdots$]{}; at (-22.70, 3.26) [$\displaystyle \dot{x}_N = f_N(x_N, \bar{x}_N)+u_{N}$]{};
(-8.35, 4.80) – (-8.35, 8.50); (-8.35, 8.50) – (-18.00, 8.50); (-19.00, 8.50) – (-23.00, 8.50); (-23.00, 8.50) – (-23.00, 5.1);
(-8.35, 2.90) – (-8.35, -0.80); (-8.35, -0.80) – (-18.00, -0.80); (-19.00, -0.8) – (-23.00, -0.8); (-23.00, -0.8) – (-23.00, 2.6);
(-19.0, 7.95) rectangle +(1.0, 1.0); at (-18.5, 8.40) [$u_1(t)$]{}; (-19.0, -1.25) rectangle +(1.0, 1.0); at (-18.5, -0.80) [$u_N(t)$]{};
The proposed controller synthesis procedure is described with the following steps:
1. $N$ TBAs $\mathcal{A}_i, \ i \in \mathcal{V}$ that accept all the timed runs satisfying the corresponding specification formulas $\varphi_i, i \in \mathcal{V}$ are constructed.
2. A Büchi WTS $\widetilde{\mathcal T}_i = \mathcal{T}_i \otimes \mathcal{A}_i$ (see Def. \[def: buchi\_WTS\] below) is constructed for every $i \in \mathcal{V}$. The accepting runs of $\widetilde{\mathcal T}_i$ are the individual runs of $\mathcal{T}_i$ that satisfy the corresponding MITL formula $\varphi_i, \ i \in \mathcal{V}$.
3. The abstraction procedure allows to find an explicit feedback law for each transition in $\mathcal T_i$. Therefore, an accepting run $\widetilde{r}^t_i$ in $\mathcal T_i$ that takes the form of a sequence of transitions is realized in the system in via the corresponding sequence of feedback laws.
\[def: buchi\_WTS\] Given a WTS $\mathcal{T}_i =(S_i, S_{i}^{\text{init}}, Act_i, \longrightarrow_i, d_i, \Sigma_i, L_i)$, and a TBA $\mathcal{A}_i = (Q_i, Q^\text{init}_i, C_i, Inv_i, E_i, F_i, \\ \Sigma_i, \mathcal{L}_i)$ with $|C_i|$ clocks and let $C^{\max}_i$ be the largest constant appearing in $\mathcal{A}_i$. Then, we define their *Büchi WTS* $\widetilde{\mathcal{T}}_i = \mathcal{T}_i \otimes \mathcal{A}_i = (\widetilde{S}_i, \widetilde{S}_{i}^{\text{init}}, \widetilde{Act}_i, {\rightsquigarrow}_{i}, \widetilde{d}_i, \widetilde{F}_i, \Sigma_i, \widetilde{L}_i)$ as follows:
- $\widetilde{S}_i \subseteq \{(s_i, q_i) \in S_i \times Q_i : L_i(s_i) = \mathcal{L}_i(q_i)\} \times \mathbb{T}_\infty^{|C_i|}$.
- $\widetilde{S}_{i}^{\text{init}} = S_i^{\text{init}} \times Q_i^{\text{init}} \times \{0\}^{|C_i|}$.
- $\widetilde{Act}_i = Act_i$.
- $(\widetilde{q}, act_i, \widetilde{q} ') \in {\rightsquigarrow}_i$ iff
- $\widetilde{q} = (s, q, \nu_1, \ldots, \nu_{|C_i|}) \in \widetilde{S}_i$,\
$\widetilde{q} ' = (s', q', \nu_1', \ldots, \nu_{|C_i|}') \in \widetilde{S}_i$,
- $act_i \in Act_i$,
- $(s, act_i, s') \in \longrightarrow_i$, and
- there exists $\gamma, R$, such that $(q, \gamma, R, q') \in E_i$, $\nu_1,\ldots,\nu_{|C_i|} \models \gamma$, $\nu_1',\ldots,\nu_{|C_i|}' \models Inv_i(q')$, and for all $i \in \{1,\ldots, |C_i|\}$ $$\nu_i' =
\begin{cases}
0, & \text{if } c_i \in R \\
\nu_i + d_i(s, s'), & \text{if } c_i \not \in R \text{ and } \\ & \nu_i + d_i(s, s') \leq C^{\mathit{max}}_i \\ \infty, & \text{otherwise}.
\end{cases}$$
Then, $\widetilde{d}_i(\widetilde{q}, \widetilde{q}') = d_i(s, s')$.
- $\widetilde{F}_i = \{(s_i, q_i,\nu_1,\ldots,\nu_{|C_i|}) \in Q_i : q_i \in F_i\}$.
- $\widetilde{L}_i(s_i, q_i, \nu_1, \ldots, \nu_{|C_i|}) = L_i(s_i)$.
Each Büchi WTS $\widetilde{\mathcal{T}}_i, i \in \mathcal{V}$ is in fact a WTS with a Büchi acceptance condition $\widetilde{F}_i$. A timed run of $\widetilde T_i$ can be written as $\widetilde{r}_i^t = (q_i(0), \tau_i(0))(q_i(1), \tau_i(1)) \ldots$ using the terminology of Def. \[run\_of\_WTS\]. It is *accepting* if $q_i(\mu) \in \widetilde F_i$ for infinitely many $j \geq 0$. An accepting timed run of $\widetilde{\mathcal{T}}_i$ projects onto a timed run of $\mathcal{T}_i$ that satisfies the local specification formula $\varphi_i$ by construction. Formally, the following lemma, whose proof follows directly from the construction and and the principles of automata-based LTL model checking (see, e.g., [@katoen]), holds:
\[eq: lemma\_1\] Consider an accepting timed run $\widetilde{r}_i^t = (q_i(0), \tau_i(0))(q_i(1), \tau_i(1)) \ldots$ of the Büchi WTS $\widetilde T_i$ defined above, where $q_i(\mu) = (r_i(\mu), s_i(\mu), \nu_{i, 1}, \ldots, \nu_{i, |C_i|})$ denotes a state of $\mathcal{\widetilde T}_i$, for all $\mu \geq 0$. The timed run $\widetilde{r}_i^t$ projects onto the timed run $r_i^t = (r_i(0), \tau_i(0))(r_i(1), \tau_i(1)) \ldots$ of the WTS $\mathcal{T}_i$ that produces the timed word $w(r_i^t) = ({L}_i(r_i(0)), \tau_i(0))({L}_i(r_i(1)), \tau_i(1)) \ldots$ accepted by the TBA $\mathcal{A}_i$ via its run $\chi_i = s_i(0)s_i(1) \ldots$. Vice versa, if there exists a timed run $r_i^t = (r_i(0),\tau_i(0))(r_i(1),\tau_i(1))\ldots$ of the WTS $T_i$ that produces a timed word $w(r_i^t) = (L_i(r_i(0)), \tau_i(0))(L_i(r_i(1)), \tau_i(1)) \ldots$ accepted by the TBA $A_i$ via its run $\chi_i = s_i(0)s_i(1)\ldots$ then there exist the accepting timed run $\widetilde{r}_i^t = (q_i(0), \tau_i(0))(q_i(1), \tau_i(1)) \ldots$ of $\widetilde{T}_i$, where $q_i(z)$ denotes $(r_i(z), s_i(z), \nu_{i,1}, \ldots, \nu_{i,|C_i|})$ in $\widetilde{T}_i$.
The proposed framework is depicted in Fig. \[fig:solution\_scheme\]. The dynamics of each agent $i$ is abstracted into a WTS $\mathcal{T}_i$ (orange rectangles). Then the product between each WTS $\mathcal{T}_i$ and the $TBA$ $\mathcal{A}_i$ is computed according to Def. \[def: buchi\_WTS\]. The TBA $\mathcal{A}_i$ accepts all the words that satisfy the formula $\varphi_i$ (blue rectangles). For every Büchi WTS $\widetilde{\mathcal{T}}_i$ the controller synthesis procedure that was described in this Section (red rectangles) is performed and a sequence of accepted runs $\{\widetilde{r}_1^t, \dots, \widetilde{r}_N^t\}$ is designed. Every accepted run $\widetilde{r}_i^t$ maps into a decentralized controller $u_i(t)$ which is a solution to Problem \[problem: basic\_prob\].
The solution that we obtain from Steps 1-5, if one found, gives a sequence of controllers $u_1, \ldots, u_N$ that guarantees the satisfaction of the formulas formulas $\varphi_1, \ldots, \varphi_N$ of the agents $1, \ldots, N$ respectively, governed by dynamics as in . Thus, we solved Problem \[problem: basic\_prob\].
Complexity {#sec:complexity}
----------
In the proposed abstraction technique $6^N$ MPC optimization problems are solved for every time interval $t \in [t_k, t_k + T]$. Assume that the desired horizon for the system to run is $M$ steps i.e. the timed sequence $\mathcal{S}$ is written as: $\mathcal{S} = \{t_0, t_1 = t_0+T, \dots, t_M = t_0+M T \}$. Then the complexity of the abstraction is $M 6^N.$ As for the controller synthesis framework now we have the following. Denote by $|\varphi|$ the length of an MITL formula $\varphi$. A TBA $\mathcal{A}_i, i \in \mathcal{V}$ can be constructed in space and time $2^{\mathbb{O}(|\varphi_i)|}, i \in \mathcal{V}$ ($\mathbb{O}$ stands for the “big O" from complexity theory). Let $\varphi_{\text{max}} = \text{max} \{ |\varphi_i\}, i \in \mathcal{V}$ be the MITL formula with the longest length. Then, the complexity of Step 1 is $2^{\mathbb{O}(|\varphi_{\text{max}})|}$. The model checking of Step 2 costs $\mathbb{O}(|\mathcal{T}_i| 2^{|\varphi_i|}), i \in \mathcal{V}$ where $|\mathcal{T}_i|$ is the length of the WTS $\mathcal{T}_i$ i.e., the number of its states. Thus, $\mathbb{O}(|\mathcal{T}_i| 2^{|\varphi_i|}) = \mathbb{O}(|S_i| 2^{|\varphi_i|}) = \mathbb{O}(|{\mathbb{I}}| 2^{|\varphi_i|})$, where $|\mathbb{I}|$ is the number of cells of the cell decomposition $D$. The worst case of Step 2 costs $\mathbb{O}(|{\mathbb{I}}| 2^{|\varphi_{\text{max}}|})$ due to the fact that all WTSs $\mathcal{T}_i, i \in \mathcal{I}$ have the same number of states. Therefore, the complexity of the total framework is $\mathbb{O}(M |{\mathbb{I}}| 6^N 2^{|\varphi_{\text{max}}|})$.
Simulation Results {#sec: simulation_results}
==================
For a simulation example, a system of three agents with $x_i \in \mathbb{R}^2, \ i \in \mathcal{V} = \{1, 2, 3\}$, $\mathcal{N}_1 = \{2,3\}$ $\mathcal{N}_2 = \{1,3\}$, $\mathcal{N}_3$ $= \{1, 1\}$ is considered. The workspace $W = [-10, 10] \times [-10, 10] \subseteq \mathbb{R}^2$ is decomposed into hexagonal regions with $R = 1, r_{\scriptscriptstyle h} = \frac{\sqrt{3}}{2}$, which are depicted in Fig. \[fig:simulation\]. The initial agents’ positions are set to $x_1(0) = (0,10r_{\scriptscriptstyle h}), x_2(0) = (-6,-8r_{\scriptscriptstyle h})$ and $x_3(0)=(7.5,-7r_{\scriptscriptstyle h})$. The sensing radius is $\underline{r} = 18$. The dynamics are set to: $\dot{x}_1 = -2x_1+x_2+x_3-\sin^2(x_1-x_2)+u_1$, $\dot{x}_2 = -2x_2+x_1+x_3-\sin^2(x_2-x_1)+u_2$ and $\dot{x}_3 = -2x_3+x_1+x_2+u_3$. The time step is $T = 3$. The specification formulas are set to $\varphi_1 = \Diamond_{[15, 27]} \{\rm red\}, \varphi_2 = \Diamond_{[7.5, 22]} \{\rm green\}, \varphi_3 = \Diamond_{[0, 19]} \{grey\}$ respectively. We set: $Q_i, P_i, R_i = I_2, \forall i \in \mathcal{V}$. Fig. \[fig:simulation\] shows a sequence of transitions for agents $1,2$ and $3$ which form the accepting timed words $\widetilde{r}_1^t$, $\widetilde{r}_2^t$ and $\widetilde{r}_3^t$, respectively. Every timed word maps to a sequence of admissible control inputs for each agent, which is the outcome of solving the ROCPs. The agents remain connected for all $t \in [0, 6T]$. The simulations were carried out in MATLAB Environment by using the NMPC toolbox [@grune_2011_nonlinear_mpc], on a desktop with 8 cores, 3.60GHz CPU and 16GB of RAM.
Conclusions and Future Work {#sec: conclusions}
===========================
A systematic method of both decentralized abstractions and controller synthesis of a general class of coupled multi-agent systems has been proposed in which timed temporal specifications are imposed to the system. The solution involves a repetitive solving of an ROCP for every agent and for every desired region in order to build decentralized Transition Systems that are then used in the derivation of the controllers that satisfy the timed temporal formulas. Future work includes further computational improvement of the proposed decentralized abstraction method.
Proof of Property 1 {#app:proof_of_property_1}
===================
By integrating in the time interval $s \in [t_k, t_k+T]$ and taking the norms, we get: $$\begin{aligned}
\|e_i(t)\| &= \left\|e_i(t_k) + \int_{t_k}^{t} \left[ g_i(x_i(s), \bar{x}_i(s), u_i(s))\right]ds \right\| \notag \\
&\le \left\|e_i(t_k) \right\| + \left\| \int_{t_k}^{t} \left[ f_i(e_i(s), \bar{x}_i(s)) + u_i(s)\right]ds \right\| \notag \\
&\le \left\|e_i(t_k) \right\| + \int_{t_k}^{t} \left\| f_i(e_i(s), \bar{x}_i(s) + u_i(s) \right\|ds \notag \\
&\le \left\|e_i(t_k) \right\| + \int_{t_k}^{t} \Big\{ \left\| f_i(e_i(s), \bar{x}_i(s) \right\| + \left\| u_i(s) \right\| \Big\} ds \notag \\
&\le \left\|e_i(t_k) \right\| +\int_{t_k}^{t} (M+u_{\max}) ds \notag \\
&= \left\|e_i(t_k) \right\| +(t-t_k) (M+u_{\max}), \notag\end{aligned}$$ which concludes the proof.
Proof of Lemma 1 {#app:lemma_1}
================
For every $e_1,e_2 \in E_i, u \in \mathcal{U}_i, i \in \mathcal{V}$, the following holds: $$\begin{aligned}
|F_i(e_1, u)-F_i(e_2, u) | &= \notag \\
&\hspace{-12mm}\|e_1^\top Q_i e_1+u^\top R_i u - e_2^\top Q_i e_2 - u^\top R_i u | \notag
\end{aligned}$$ $$\begin{aligned}
&\hspace{-12mm}= |e_1^\top Q_i e_1 - e_2^\top Q_i e_2 | \notag \\
&\hspace{-12mm}= |e_1^\top Q_i e_1 +e_1^\top Q_i e_2 -e_1^\top Q_i e_2 - e_2^\top Q_i e_2 | \notag \\
&\hspace{-12mm}= |e_1^\top Q_i (e_1-e_2) - e_2^\top Q_i ( e_1 - e_2) | \notag \\
&\hspace{-12mm}\le |e_1^\top Q_i (e_1-e_2)| + |e_2^\top Q_i ( e_1 - e_2) |. \label{eq:lemma_1_proof_step_11}
\end{aligned}$$ By employing the property that: $$\label{eq:sigma_max}
|x^\top A y| \le \sigma_{\max}(A) \|x\| \|y\|, \forall \ x,y \in \mathbb{R}^n, A \in \mathbb{R}^{n \times n},$$ is written as: $$\begin{aligned}
&|F_i(e_1, u)-F_i(e_2, u) | \le \sigma_{\max}(Q_i) \|e_1\| \|e_1-e_2\| \notag \\
&\hspace{35mm} +\sigma_{\max}(Q_i) \|e_2\| \|e_1-e_2\| \notag \\
&\hspace{4mm}= \sigma_{\max}(Q_i) (\|e_1\| + \|e_2\|) \|e_1-e_2\| \notag \\
&\hspace{4mm}= \sigma_{\max}(Q_i) \left[\sup_{e_1, e_2 \in E_i} \{\|e_1\| + \|e_2\|\}\right] \|e_1-e_2\| \notag \\
&\hspace{4mm}= 2 \sigma_{\max}(Q_i) \left[\sup_{e_i \in E_i} \{\|e\|\}\right] \|e_1-e_2\| \notag \\
&\hspace{4mm}= \left[ 2 \bar{\varepsilon}_i \sigma_{\max}(Q_i) \right] \|e_1-e_2\|.
\end{aligned}$$ which completes the proof.
Proof of Lemma 2 {#app:lemma_e_ehat}
================
Let us denote by: $$\begin{aligned}
u_i(\cdot) \triangleq u_i(s; e(t_{\scriptscriptstyle k_z})), \notag \\
e_i(s) \triangleq e_i(s; u_i(\cdot), e_i(t_{\scriptscriptstyle k_z})). \notag\end{aligned}$$ the control input and real trajectory of the system for $s \in [t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_z} + T_{\scriptscriptstyle z}]$. Also, denote for sake of simplicity: $$\hat{e}_i(s) \triangleq \hat{e}_i(s; u_i(\cdot), e_i(t_{\scriptscriptstyle k_z})).$$ the corresponding estimated trajectory. By integrating , for the time interval $[t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_z}+s]$ we have the following: $$\begin{aligned}
e_i(s) &= e_i(t_{\scriptscriptstyle k_z}) + \int_{t_{\scriptscriptstyle k_z}}^{s} \left[ g_i(e_i(s'), \bar{x}_i(s'), u_i(\cdot)) \right] ds', \notag \\
\hat{e}_i(s) &= e_i(t_{\scriptscriptstyle k_z}) + \int_{t_{\scriptscriptstyle k_z}}^{s} \left[ g_i(\hat{e}_i(s'), \hat{\bar{x}}_i(s'), u_i(\cdot)) \right] ds', \notag\end{aligned}$$ respectively. Then, we have that: $$\begin{aligned}
&\|e_i(s)-\hat{e}_i(s)\| \notag \\
&= \Bigg\|\int_{t_{\scriptscriptstyle k_z}}^{s} \left[ g(e_i(s'), \bar{x}_i(s'), u_i(\cdot)) \right] ds'- \notag \\
&\hspace{35mm} \int_{t_{\scriptscriptstyle k_z}}^{s} \left[ g(\hat{e}_i(s'), \hat{\bar{x}}_i(s'), u_i(\cdot)) \right] ds' \Bigg\| \notag \\
&= \Bigg\|\int_{t_{\scriptscriptstyle k_z}}^{s} \big[ f(e_i(s'), \bar{x}_i(s'))+u_{i}(s') \notag \\
&\hspace{35mm}-f(\hat{e}_i(s'), \hat{\bar{x}}_i(s'))-u_{i}(s')\big] ds' \Bigg\| \notag \\
&= \Bigg\|\int_{t_{\scriptscriptstyle k_z}}^{s} \big[ f(e_i(s'), \bar{x}_i(s'))- f(\hat{e}_i(s'), \hat{\bar{x}}_i(s'))\big] ds' \Bigg\| \notag \\
&\le \int_{t_{\scriptscriptstyle k_z}}^{s} \Big\| f(e_i(s'), \bar{x}_i(s'))- f(\hat{e}_i(s'), \hat{\bar{x}}_i(s')) \Big\| ds' \notag\end{aligned}$$ $$\begin{aligned}
&= \int_{t_{\scriptscriptstyle k_z}}^{s} \Big\| f(e_i(s'), \bar{x}_i(s')) -f(\hat{e}_i(s'), \bar{x}_i(s')) \notag \\
&\hspace{20mm}+ f(\hat{e}_i(s'), \bar{x}_i(s'))- f(\hat{e}_i(s'), \hat{\bar{x}}_i(s'))\Big\| ds' \notag \\
&\le \int_{t_{\scriptscriptstyle k_z}}^{s} \Big\| f(e_i(s'), \bar{x}_i(s')) -f(\hat{e}_i(s'), \bar{x}_i(s')) \big\|ds' \notag \\
&\hspace{15mm}+ \int_{t_{\scriptscriptstyle k_z}}^{s} \Big\| f(\hat{e}_i(s'), \bar{x}_i(s'))- f(\hat{e}_i(s'), \hat{\bar{x}}_i(s'))\Big\| ds'. \notag\end{aligned}$$ By using the bounds of - we obtain: $$\begin{aligned}
&\|e_i(s)-\hat{e}_i(s)\| \notag \\
&\le \int_{t_{\scriptscriptstyle k_z}}^{s} L_i \big\| e_i(s') -\hat{e}_i(s') \big\|ds' \notag \\
&\hspace{30mm}+ \int_{t_{\scriptscriptstyle k_z}}^{s} \bar{L}_i \big\| \bar{x}_i(s')- \hat{\bar{x}}_i(s') \big\| ds'. \label{eq:before_bellman}\end{aligned}$$ The following property holds: $$\|\bar{x}_i-\hat{\bar{x}}_i\| \le \sum_{j \in \mathcal{N}_i} \|x_j-\hat{x}_j\|, \forall i \in \mathcal{V}, j \in \mathcal{N}_i.$$ Then, by combining the last inequality with from Property 2, we have that: $$\|\bar{x}_i-\hat{\bar{x}}_i\| \le \sum_{j \in \mathcal{N}_i} 2 \sqrt{3} R = 2 \sqrt{3} R N_i, \forall i \in \mathcal{V}, j \in \mathcal{N}_i.$$ By combining the last result with we get: $$\begin{aligned}
&\|e_i(s)-\hat{e}_i(s)\| \le \notag \\
&\int_{t_{\scriptscriptstyle k_z}}^{s} L_i \big\| e_i(s') -\hat{e}_i(s') \big\|ds' + \int_{t_{\scriptscriptstyle k_z}}^{s} \bar{L}_i 2 \sqrt{3} R N_i ds' \notag \\
&= \int_{t_{\scriptscriptstyle k_z}}^{s} L_i \big\| e_i(s') -\hat{e}_i(s') \big\|ds' + 2 \sqrt{3} R \bar{L}_i N_i (s-t_{\scriptscriptstyle k_z}). \label{eq:bbefore_bellman}\end{aligned}$$ By employing the Gronwall-Bellman inequality from [@khalil_nonlinear_systems], becomes: $$\begin{aligned}
&\|e_i(s)-\hat{e}_i(s)\| \notag \\
&\le 2 \sqrt{3} R L_i \bar{L}_i N_i \int_{t_{\scriptscriptstyle k_z}}^{s} (s'-t_{\scriptscriptstyle k_z}) \exp\left[\int_{s'}^{s} L_i ds'' \right] ds' \notag \\
&\hspace{20mm} + 2 \sqrt{3} R \bar{L}_i N_i (s-t_{\scriptscriptstyle k_z}) \notag \\
&= 2 \sqrt{3} R L_i \bar{L}_i N_i \int_{t_{\scriptscriptstyle k_z}}^{s} (s'-t_{\scriptscriptstyle k_z}) \exp\left[ L_i(-s'+s) \right] ds' \notag \\
&\hspace{20mm} + 2 \sqrt{3} R \bar{L}_i N_i (s-t_{\scriptscriptstyle k_z}) \notag \\
&= -2 \sqrt{3} R \bar{L}_i N_i(s-t_{\scriptscriptstyle k_z}) +2 \sqrt{3} R \bar{L}_i N_i (s-t_{\scriptscriptstyle k_z}) \notag \\
&\hspace{10mm} + 2 \sqrt{3} R \bar{L}_i N_i\int_{t_{\scriptscriptstyle k_z}}^{s} \exp\left[ L_i(-s'+s) \right] ds' \notag \\
&= 2 \sqrt{3} R \bar{L}_i N_i\int_{t_{\scriptscriptstyle k_z}}^{s} \exp\left[ L_i(-s'+s) \right] ds' \notag \\
&= -\frac{2 \sqrt{3} R \bar{L}_i N_i}{L_i} \left[ 1- e^{L_i(s-t_{\scriptscriptstyle k_z})} \right] \notag \\
&= \frac{2 \sqrt{3} R \bar{L}_i N_i}{L_i} \left[ e^{L_i(s-t_{\scriptscriptstyle k_z})}- 1 \right]. \label{eq:e_e_hat_bound_1}\end{aligned}$$ By employing of Property \[property:error\_bound\] for the terms $e(s), \hat{e}(s)$ we have that: $$\begin{aligned}
&\|e_i(s)-\hat{e}_i(s)\| \le \|e_i(s)\| + \|\hat{e}_i(s)\| \notag \\
&\le \|e_i(t_{\scriptscriptstyle k_z}; u_i(\cdot), e_i(t_{\scriptscriptstyle k_z}))\|+ (s-t_{\scriptscriptstyle k_z})(M+u_{\max}) + \notag \\
&\hspace{15mm} \|\hat{e}_i(t_{\scriptscriptstyle k_z}; u_i(\cdot), e_i(t_{\scriptscriptstyle k_z}))\|+ (s-t_{\scriptscriptstyle k_z})(M+u_{\max}) \notag \\
&\le 2 \|e_i(t_{\scriptscriptstyle k_z}; u_i(\cdot), e_i(t_{\scriptscriptstyle k_z}))\|+ 2 (s-t_{\scriptscriptstyle k_z})(M+u_{\max})\notag \\
&= 2 \|e_i(t_{\scriptscriptstyle k_z})\|+ 2 (s-t_{\scriptscriptstyle k_z})(M+u_{\max}). \label{eq:e_e_hat_bound_2}\end{aligned}$$ By combining , we get: $$\begin{aligned}
&\|e_i(s)-\hat{e}_i(s)\| \le \min \big\{ \frac{2 \sqrt{3} R \bar{L}_i N_i}{L_i} \left[ e^{L_i(s-t_{\scriptscriptstyle k_z})}- 1 \right], \notag \\
&\hspace{10mm} 2 \|e_i(t_{\scriptscriptstyle k_z})\|+ 2 (s-t_{\scriptscriptstyle k_z})(M+u_{\max}) \big\}.\end{aligned}$$ which leads to the conclusion of the proof.
Proof of Property 3 {#app:property_tighthened_sets}
===================
Let $s \in [t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z}]$. Let us also define: $$z_i(s) \triangleq e_i(s) -\hat{e}_i(s; u_i(s; e(t_{\scriptscriptstyle k_z})), e_i(t_{\scriptscriptstyle k_z})).$$ Then, according to Lemma \[lemma:e-hat\_e\], for $s \in [t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z}]$, we get: $$\begin{aligned}
\|z_i(s)\| &= \| e_i(s) -\hat{e}_i(s; u_i(s; e_i(t_{\scriptscriptstyle k_z})), e_i(t_{\scriptscriptstyle k_z})) \| \notag \\
&\le \rho_i(s-t_{\scriptscriptstyle k_z}).
\end{aligned}$$ Hence, $z_i \in B^i_{s-t_{\scriptscriptstyle k_z}}$, which implies that: $-z_i \in B^i_{s-t_{\scriptscriptstyle k_z}}$. The following implications hold: $$\begin{aligned}
\hat{e}_i(s; u_i(s; e_i(t_{\scriptscriptstyle k_z})), e_i(t_{\scriptscriptstyle k_z})) &\in E_i \sim B^i_{s-t_{\scriptscriptstyle k_z}} \notag \\
\Rightarrow e_i(s) - z_i &\in E \sim B^i_{s-t_{\scriptscriptstyle k_z}} \notag \\
\Rightarrow e_i(s) + (-z_i) &\in E \sim B^i_{s-t_{\scriptscriptstyle k_z}} \notag \\
\Rightarrow e_i(s) &\in E_i, \forall \ s \in [t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z}],
\end{aligned}$$ which concludes the proof.
Proof of Lemma 3 {#app:lemma_lip_L_v}
================
For every $e_i \in \Phi_i$ we have that: $$\begin{aligned}
V_i(e_i) \le \alpha_{1,i} &\Rightarrow e_i^\top P_i e_i \le \alpha_{1,i} \notag \\
&\Rightarrow \lambda_{\min}(P_i) \|e_i\|^2 \le e_i^\top P_i e_i \le \alpha_{1,i} \notag \\
&\Rightarrow \|e_i\| \le \sqrt{ \frac{\alpha_{1,i}}{\lambda_{\min}(P_i)}}. \label{eq:lamda_min_P}
\end{aligned}$$
For every $e_1,e_2 \in \Phi_i$, it also holds: $$\begin{aligned}
|V_i(e_{1})-V_i(e_{2}) | &= |e_{1}^\top P_i e_{1} - e_{2}^\top P_i e_{2} | \notag \\
&\hspace{-2mm}= |e_1^\top P_i e_1 +e_1^\top P_i e_2 -e_1^\top P_i e_2 - e_2^\top P_i e_2 | \notag \\
&\hspace{-2mm}= |e_1^\top P_i (e_1-e_2) - e_2^\top P_i ( e_1 - e_2) | \notag \\
&\hspace{-2mm}\le |e_1^\top P_i (e_1-e_2)| + |e_2^\top P_i ( e_1 - e_2) |, \notag
\end{aligned}$$ which by using leads to: $$\begin{aligned}
|V_i(e_1)-V_i(e_2) | &\le \sigma_{\max}(P_i) \|e_1\| \|e_1-e_2\| \notag \\
&\hspace{5mm} +\sigma_{\max}(P_i) \|e_2\| \|e_1-e_2\| \notag \\
&= \sigma_{\max}(P_i) (\|e_1\| + \|e_2\|) \|e_1-e_2\|, \notag
\end{aligned}$$ which by employing , becomes: $$\begin{aligned}
|V_i(e_1)-V_i(e_2) | &\le \sigma_{\max}(P_i) \|e_1\| \|e_1-e_2\| \notag \\
&\hspace{5mm} +\sigma_{\max}(P_i) \|e_2\| \|e_1-e_2\| \notag \\
&= \sigma_{\max}(P_i) (\|e_1\| + \|e_2\|) \|e_1-e_2\| \notag \\
&\le 2 \sigma_{\max}(P_i) \sqrt{ \frac{\alpha_1}{\lambda_{\min}(P_i)}} \|e_1-e_2\|, \notag
\end{aligned}$$ which completes the proof.
Proof of Lemma 4 {#app:y_E_t_i_plus_1}
================
For every $s \ge t_{\scriptscriptstyle k_{z+1}}, L_g > 0$ the following inequality holds: $$\left[e^{L_g(t_{\scriptscriptstyle k_{z+1}}-t_{\scriptscriptstyle k_z})}-1\right]+\left[e^{L_g(s-t_{\scriptscriptstyle k_{z+1}})}-1\right] \le \left[e^{L_g(s-t_{\scriptscriptstyle k_z})}-1\right],$$ which implies that: $$\begin{aligned}
&\widetilde{\rho}_i \left[e^{L_g(t_{\scriptscriptstyle k_{z+1}}-t_{\scriptscriptstyle k_z})}-1\right]+\widetilde{\rho}_i \left[e^{L_g(s-t_{\scriptscriptstyle k_{z+1}})}-1\right] \notag \\
&\hspace{40mm}\le \widetilde{\rho}_i \left[e^{L_g(s-t_{\scriptscriptstyle k_z})}-1\right]. \label{eq:inequaliti_min_rwho_1}
\end{aligned}$$ It holds also that: $$\begin{aligned}
& t_{\scriptscriptstyle k_{z+1}}-t_{\scriptscriptstyle k_z}+s-t_{\scriptscriptstyle k_{z+1}} \le s-t_{\scriptscriptstyle k_z} \notag \\
\Leftrightarrow \ & 2 \|e_i(t_{\scriptscriptstyle k_z})\|+ 2 (t_{\scriptscriptstyle k_{z+1}}-t_{\scriptscriptstyle k_z})(M+u_{\max}) \notag \\
&+ 2 \|e_i(t_{\scriptscriptstyle k_z})\|+ 2 (s-t_{\scriptscriptstyle k_{z+1}})(M+u_{\max}) \le \notag \\
& 2 \|e_i(t_{\scriptscriptstyle k_z})\|+ 2 (s-t_{\scriptscriptstyle k_z})(M+u_{\max}). \label{eq:inequaliti_min_rwho_2}
\end{aligned}$$ By setting: $$\begin{aligned}
A_1 &= \widetilde{\rho}_i \left[e^{L_g(t_{\scriptscriptstyle k_{z+1}}-t_{\scriptscriptstyle k_z})}-1\right], \notag \\
A_2 &= \widetilde{\rho}_i \left[e^{L_g(s-t_{\scriptscriptstyle k_{z+1}})}-1\right], \notag \\
A_3 &= \widetilde{\rho}_i \left[e^{L_g(s-t_{\scriptscriptstyle k_z})}-1\right], \notag \\
B_1 &= 2 \|e_i(t_{\scriptscriptstyle k_z})\|+ 2 (t_{\scriptscriptstyle k_{z+1}}-t_{\scriptscriptstyle k_z})(M+u_{\max}), \notag \\
B_2 &= 2 \|e_i(t_{\scriptscriptstyle k_z})\|+ 2 (s-t_{\scriptscriptstyle k_{z+1}})(M+u_{\max}), \notag \\
B_3 &= 2 \|e_i(t_{\scriptscriptstyle k_z})\|+ 2 (s-t_{\scriptscriptstyle k_z})(M+u_{\max}), \notag
\end{aligned}$$ and taking account , we get: $$\begin{aligned}
&\rho_i(t_{\scriptscriptstyle k_{z+1}}-t_{\scriptscriptstyle k_z})+\rho_i(s-t_{\scriptscriptstyle k_{z+1}}) \notag \\
&\hspace{15mm} \le \min\{A_1, B_1\} + \min\{B_1, B_2\} \notag \\
&\hspace{15mm} \le \min\{A_1+A_2, B_1+B_2\} \notag \\
&\hspace{15mm} \le \min\{A_3, B_3\} \notag \\
&\hspace{15mm} = \rho_i(s-t_{\scriptscriptstyle k_z}), \notag
\end{aligned}$$ or: $$\label{eq:gamma_inequality}
\rho_i(t_{\scriptscriptstyle k_{z+1}}-t_{\scriptscriptstyle k_z})+\rho_i(s-t_{\scriptscriptstyle k_{z+1}}) \le \rho_i(s-t_{\scriptscriptstyle k_z}).$$ Let us consider $\phi \in B^i_{s-t_{\scriptscriptstyle k_{z+1}}}$. Then, it holds $\|\phi\| \le \rho_i(s-t_{\scriptscriptstyle k_{z+1}})$. Let us denote $z = x-y+\phi$. It is clear that: $$\begin{aligned}
\|z\| &\le \|x-y+\phi \| \notag \\
&\le \|x-y\| + \|\phi\| \notag \\
&\le \rho_i(t_{\scriptscriptstyle k_{z+1}}-t_{\scriptscriptstyle k_z})+\rho_i(s-t_{\scriptscriptstyle k_{z+1}}). \label{eq:z_inequlity}
\end{aligned}$$ By employing , becomes: $$\|z\| \le \rho_i(s-t_{\scriptscriptstyle k_z}),$$ which implies that $z \in B^i_{s-t_{\scriptscriptstyle k_z}}$. We have that: $$\begin{aligned}
x + (-z) &= y + (-\phi), \notag \\
x &\in E_{s-t_{\scriptscriptstyle k_z}} = E \sim B_{s-t_{\scriptscriptstyle k_z}}, \notag \\
-z &\in B^i_{s-t_{\scriptscriptstyle k_z}}, \notag \\
-\rho &\in B^i_{s-t_{\scriptscriptstyle k_{z+1}}}, \notag
\end{aligned}$$ which implies that $y \in E_{s-t_{\scriptscriptstyle k_{z+1}}} = E \sim B_{s-t_{\scriptscriptstyle k_{z+1}}}$.
Proof of Lemma 5 {#app:bounded_trajectories}
================
Let $s \ge t_{\scriptscriptstyle k_z}$. The following equalities hold: $$\begin{aligned}
&\| \hat{e}_i(s; u_i(\cdot), e_i(t_{\scriptscriptstyle k_{z+1}})) - \hat{e}_i(s; u_i(\cdot), e_i(t_{\scriptscriptstyle k_z})) \| \notag \\
&= \Bigg\| \hat{e}_i(s; u_i(\cdot), e_i(t_{\scriptscriptstyle k_{z+1}})) + \int_{t_{\scriptscriptstyle k_{z+1}}}^{s} g_i(\hat{e}_i(s'), \hat{\bar{x}}_i(s'), u_i(\cdot))ds' \notag \\
&-\hat{e}_i(t_{\scriptscriptstyle k_{z}}; u_i(\cdot), e_i(t_{\scriptscriptstyle k_z})) - \int_{t_{\scriptscriptstyle k_z}}^{s} g_i(\hat{e}_i(s'), \hat{\bar{x}}_i(s'), u_i(\cdot)))ds \Bigg\| \notag \\
&=\Bigg\| e_i(t_{\scriptscriptstyle k_{z+1}}) -e_i(t_{\scriptscriptstyle k_z}) - \int_{t_{\scriptscriptstyle k_z}}^{s} g_i(\hat{e}_i(s'), \hat{\bar{x}}_i(s'), u_i(\cdot)) ds'\notag
\end{aligned}$$ $$\begin{aligned}
&\hspace{30mm}- \int_{s}^{t_{\scriptscriptstyle k_{z+1}}} g_i(\hat{e}_i(s'), \hat{\bar{x}}_i(s'), u_i(\cdot)) ds'\Bigg\| \notag \\
&= \left\| e_i(t_{\scriptscriptstyle k_{z+1}}) -e_i(t_{\scriptscriptstyle k_z}) - \int_{t_{\scriptscriptstyle k_z}}^{t_{\scriptscriptstyle k_{z+1}}} g_i(\hat{e}_i(s'), \hat{\bar{x}}_i(s'), u_i(\cdot)) ds'\right\| \notag \\
&= \left\| e_i(t_{\scriptscriptstyle k_{z+1}}) -e_i(t_{\scriptscriptstyle k_z}) - \int_{t_{\scriptscriptstyle k_z}}^{t_{\scriptscriptstyle k_{z+1}}} \frac{d}{dt} \left[ \hat{e}_i(s'; u_i(\cdot), e_i(t_{\scriptscriptstyle k_z})\right] ds'\right\| \notag \\
&= \big\| e_i(t_{\scriptscriptstyle k_{z+1}}) -e_i(t_{\scriptscriptstyle k_z}) - \hat{e}(t_{\scriptscriptstyle k_{z+1}}; u(\cdot), e_i(t_{\scriptscriptstyle k_z})) \notag \\ &\hspace{45mm} +\hat{e}_i(t_{\scriptscriptstyle k_z}; u_i(\cdot), e_i(t_{\scriptscriptstyle k_z}))\big\| \notag \\
&= \left\| e_i(t_{\scriptscriptstyle k_{z+1}}) -e_i(t_{\scriptscriptstyle k_z}) - \hat{e}(t_{\scriptscriptstyle k_{z+1}}; u(\cdot), e_i(t_{\scriptscriptstyle k_z}))+e_i(t_{\scriptscriptstyle k_z})\right\| \notag \\
&= \left\| e_i(t_{\scriptscriptstyle k_{z+1}}) - \hat{e}_i(t_{\scriptscriptstyle k_{z+1}}; u_i(\cdot), e_i(t_{\scriptscriptstyle k_z}))\right\|, \notag
\end{aligned}$$ which, by employing Lemma \[lemma:e-hat\_e\] for $s = t_{\scriptscriptstyle k_{z+1}}$, becomes: $$\begin{aligned}
&\| \hat{e}_i(s; u(\cdot), e_i(t_{\scriptscriptstyle k_{z+1}})) - \hat{e}_i(s; u_i(\cdot), e_i(t_{\scriptscriptstyle k_z})) \| \le \notag \\
&\hspace{30mm} \rho_i(t_{\scriptscriptstyle k_{z+1}}-t_{\scriptscriptstyle k_z}) = \rho_i(h), \notag
\end{aligned}$$ since $t_{\scriptscriptstyle k_{z+1}}-t_{\scriptscriptstyle k_z} = h$, which concludes the proof.
Feasibility and Convergence {#app:feasibility_convergence}
===========================
The proof consists of two parts: in the first part it is established that initial feasibility implies feasibility afterwards. Based on this result it is then shown that the error $e_i(t)$ converges to the terminal set $\mathcal{E}_i$.
*Feasibility Analysis*: Consider any sampling time instant for which a solution exists, say $t_{\scriptscriptstyle k_z}$. In between $t_{\scriptscriptstyle k_z}$ and $t_{\scriptscriptstyle k_{z+1}}$, the optimal control input $\hat{u}_i^\star (s; e_i(t_{\scriptscriptstyle k_z})), s \in [t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_{z+1}})$ is implemented. The remaining part of the optimal control input $\hat{u}_i^\star (s; e_i(t_{\scriptscriptstyle k_z})), s \in [t_{\scriptscriptstyle k_{z+1}}, t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z}],$ satisfies the state and input constraints $E_i, \mathcal{U}_i$, respectively. Furthermore, since the problem is feasible at time $t_{\scriptscriptstyle k_z}$, it holds that:
$$\begin{aligned}
\hat{e}_i(s; \hat{u}^\star (s; e_i(t_{\scriptscriptstyle k_z})), e_i(t_{\scriptscriptstyle k_z})) &\in E^i_{s-t_{\scriptscriptstyle k_z}}, \label{eq:e_s_hat_in_E_s_t_i}\\
\hat{e}_i(t_{\scriptscriptstyle k_z}+T; \hat{u}_i^\star (s; e_i(t_{\scriptscriptstyle k_z})), e_i(t_{\scriptscriptstyle k_z})) &\in \mathcal{E}_i, \label{eg:t_k_z_plus_T_in_E_i}
\end{aligned}$$
for $s \in [t_{\scriptscriptstyle k_z}, t_{\scriptscriptstyle k_z}+T_{\scriptscriptstyle z}]$. By using Property 1, implies also that $e_i(s; \hat{u}_i^\star (s; e_i(t_{\scriptscriptstyle k_z})), e_i(t_{\scriptscriptstyle k_z})) \in E_i$. We know also from Assumption \[ass:admissible\_u\_f\] that for all $e_i \in \mathcal{E}_i$, there exists at least one control input $u_{f,i}(\cdot)$ that renders the set $\mathcal{E}_i$ invariant over $h$. Picking any such input, a feasible control input $\bar{u}_i(\cdot; e_i(t_{\scriptscriptstyle k_{z+1}}))$, at time instant $t_{\scriptscriptstyle k_{z+1}}$, may be the following: $$\begin{aligned}
\label{eq:u_bar_feas}
&\bar{u}_i(s; e(t_{\scriptscriptstyle k_{z+1}})) = \notag \\
&\hspace{-2mm} \begin{cases}
\hat{u}_i^\star (s; e_i(t_{\scriptscriptstyle k_z})), \ & s \in [t_{\scriptscriptstyle k_{z+1}}, t_{\scriptscriptstyle k_{z}}+ T_{\scriptscriptstyle z+1}], \\
u_{f,i} (t_{\scriptscriptstyle k_{z}}+ T_{\scriptscriptstyle z+1}; \hat{u}^\star(\cdot), e(t_i))), & s \in [t_{\scriptscriptstyle k_{z}}+ T_{\scriptscriptstyle z+1}, t_{\scriptscriptstyle k_{z}}+ T_{\scriptscriptstyle z}].
\end{cases}
\end{aligned}$$ For the time intervals it holds that (see Fig. \[fig:time\_sequence\]): $$\begin{aligned}
&t_{\scriptscriptstyle k_{z}}+ T_{\scriptscriptstyle z+1} = t_{\scriptscriptstyle k_{z}} + T_{\scriptscriptstyle z} - h = t_{\scriptscriptstyle k_{z}} + T - h.\notag
\end{aligned}$$ For the feasibility of the ROCP, we have to prove the following three statements for every $s \in [t_{\scriptscriptstyle k_{z+1}}, t_{\scriptscriptstyle k_{z}}+T_{\scriptscriptstyle z}]$:
1. $\bar{u}_i(s; e(t_{\scriptscriptstyle k_{z+1}})) \in \mathcal{U}_i$.
2. $\hat{e}_i(t_{\scriptscriptstyle k_{z}}+T_{\scriptscriptstyle z}; \bar{u}(s; e(t_{\scriptscriptstyle k_{z+1}})), e(t_{\scriptscriptstyle k_{z+1}})) \in \mathcal{E}_i$.
3. $\hat{e}_i(s; \bar{u}_i(s; e(t_{\scriptscriptstyle k_{z+1}})), e(t_{\scriptscriptstyle k_{z+1}})) \in E^i_{s-t_{\scriptscriptstyle k_{z+1}}}$.
Statement 1: From the feasibility of $\hat{u}_i^\star(s, e(t_{\scriptscriptstyle k_{z}}))$ and the fact that $u_{f,i}(e_i(\cdot)) \in \mathcal{U}_i$, for all $e_i(\cdot) \in \Phi_i$, it follows that: $$\bar{u}_i(s; e_i(t_{\scriptscriptstyle k_{z+1}})) \in \mathcal{U}_i, \forall \ s \in [t_{\scriptscriptstyle k_{z+1}}, t_{\scriptscriptstyle k_{z}}+T_{\scriptscriptstyle z}].$$ Statement 2: We need to prove in this step that for every $s \in [t_{\scriptscriptstyle k_{z+1}}, t_{\scriptscriptstyle k_{z}}+T_{\scriptscriptstyle z}]$ it holds that $\hat{e}_i(t_{\scriptscriptstyle k_{z}}+T_{\scriptscriptstyle z}; \bar{u}_i(s; e_i(t_{\scriptscriptstyle k_{z+1}}))), e_i(t_{\scriptscriptstyle k_{z+1}})) \in \mathcal{E}_i$. Since $V_i(\cdot)$ is Lipschitz continuous, we get: $$\begin{aligned}
&\hspace{-2mm} V_i(\hat{e}_i(t_{\scriptscriptstyle k_{z}}+T_{\scriptscriptstyle z+1}; \bar{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_{z+1}}))) - \notag \\
&\hspace{20mm} V_i(\hat{e}_i(t_{\scriptscriptstyle k_{z}}+T_{\scriptscriptstyle z+1}; \bar{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_{z}}))) \le \notag \\
&\hspace{-2mm} L_{\scriptscriptstyle V_i} \| \hat{e}_i(t_{\scriptscriptstyle k_{z}}+T_{\scriptscriptstyle z+1}; \bar{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_{z+1}})) \notag \\
&\hspace{20mm}- \hat{e}(t_{\scriptscriptstyle k_{z}}+T_{\scriptscriptstyle z+1}; \bar{u}_i(\cdot), e(t_{\scriptscriptstyle k_{z}})) \|. \label{eq:feasib_lipsitch}
\end{aligned}$$ for the same control input $\bar{u}_i(\cdot) = u_i^\star(s; e_i(t_{\scriptscriptstyle k_{z}}))$. By employing Lemma \[lemma:bounded\_trajectories\] for $\alpha = t_{\scriptscriptstyle k_{z}}+T_{\scriptscriptstyle z+1}$ and $u(\cdot) = \bar{u}_i(\cdot) = u_i^\star(s; e_i(t_{\scriptscriptstyle k_{z}}))$, we have that: $$\begin{aligned}
&\hspace{-2mm}\| \hat{e}_i(t_{\scriptscriptstyle k_{z}}+T_{\scriptscriptstyle z}; \bar{u}_i(\cdot),
e_i(t_{\scriptscriptstyle k_{z+1}})) \notag \\
&\hspace{-1mm}- \hat{e}_i(t_{\scriptscriptstyle k_{z}}+T_{\scriptscriptstyle z+1};
\bar{u}_i(\cdot), e(t_{\scriptscriptstyle k_{z}})) \| \le \rho_i(t_{\scriptscriptstyle k_{z+1}}-t_{\scriptscriptstyle k_{z}}) = \rho_i(h). \label{eq:bound_gamma_h}
\end{aligned}$$ Note also that for the function $\rho_i(\cdot)$ the following implication holds: $$h \le T_{\scriptscriptstyle z} \Rightarrow \rho_i(h) \le \rho_i(T_{\scriptscriptstyle z}).$$ By employing the latter result, becomes: $$\begin{aligned}
&\hspace{-2mm}\| \hat{e}_i(t_{\scriptscriptstyle k_{z}}+T_{\scriptscriptstyle z}; \bar{u}_i(\cdot),
e_i(t_{i+1})) \notag \\
&\hspace{-1mm}- \hat{e}_i(t_{\scriptscriptstyle k_{z}}+T_{\scriptscriptstyle z+1};
\bar{u}_i(\cdot), e(t_{\scriptscriptstyle k_{z}})) \| \le \rho_i(h) \le \rho_i(T_{\scriptscriptstyle z}). \label{eq:bound_gamma_h2}
\end{aligned}$$ By combining and we get: $$\begin{aligned}
&\hspace{-2mm} V_i(\hat{e}_i(t_{\scriptscriptstyle k_{z}}+T_{\scriptscriptstyle z+1}; \bar{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_{z+1}}))) - \notag \\
&\hspace{10mm} V_i(\hat{e}_i(t_{\scriptscriptstyle k_{z}}+T_{\scriptscriptstyle z+1}; \bar{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_{z}}))) \le L_{\scriptscriptstyle V_i} \rho_i(T_{\scriptscriptstyle z}), \notag
\end{aligned}$$ or equivalently: $$\begin{aligned}
&\hspace{-2mm} V_i(\hat{e}_i(t_{\scriptscriptstyle k_{z}}+T_{\scriptscriptstyle z+1}; \bar{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_{z+1}}))) \le \notag \\
&\hspace{10mm} V_i(\hat{e}_i(t_{\scriptscriptstyle k_{z}}+T_{\scriptscriptstyle z+1}; \bar{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_{z}}))) + L_{\scriptscriptstyle V_i} \rho_i(T_{\scriptscriptstyle z}). \label{eq:feasib_lipsitch_1}
\end{aligned}$$ By using , we have that $\hat{e}_i(t_{\scriptscriptstyle k_{z}}+T_{\scriptscriptstyle z+1}; \bar{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_{z}})) \in \mathcal{E}_i$. Then, gives: $$\begin{aligned}
V_i(\hat{e}_i(t_{\scriptscriptstyle k_{z}}+T_{\scriptscriptstyle z+1}; \bar{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_{z+1}}))) \le \alpha_{2,i} + L_{\scriptscriptstyle V_i} \rho_i(T_{\scriptscriptstyle z}) \label{eq:feasib_lipsitch_2}
\end{aligned}$$ From of the Theorem 1, we get equivalently: $$\begin{aligned}
& \rho_i(T_{\scriptscriptstyle z}) \le \frac{\alpha_{1,1}-\alpha_{2,i}}{L_{\scriptscriptstyle V_i}} \notag \\
\Leftrightarrow & \alpha_{2,i} + L_{\scriptscriptstyle V_i} \rho_i(T_{\scriptscriptstyle z}) \le \alpha_{1,i}. \label{eq:alpha_1_L_V}
\end{aligned}$$ By combining and , we get: $$\begin{aligned}
V_i(\hat{e}_i(t_{\scriptscriptstyle k_{z}}+T_{\scriptscriptstyle z+1}; \bar{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_{z+1}}))) \le \alpha_{1,i},
\end{aligned}$$ which, from Assumption \[ass:admissible\_u\_f\], implies that: $$\label{eq:e_hat_in_phi}
\hat{e}_i(t_{\scriptscriptstyle k_{z}}+T_{\scriptscriptstyle z+1}; \bar{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_{z+1}})) \in \Phi_i.$$ But since $\bar{u}_i(\cdot)$ is chosen to be local admissible controller from Assumption \[ass:admissible\_u\_f\], according to our choice of terminal set $\mathcal{E}_i$, leads to: $$\hat{e}_i(t_{\scriptscriptstyle k_{z}}+T_{\scriptscriptstyle z}; \bar{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_{z+1}})) \in \mathcal{E}_i.$$ Thus, statement 2 holds.
Statement 3: By employing Lemma \[lemma:bounded\_trajectories\] for: $$\begin{aligned}
x &= \hat{e}_i(s; \bar{u}_i(s; e(t_{\scriptscriptstyle k_{z}})), e(t_{\scriptscriptstyle k_{z}})) \in E^i_{s-t_i}, \notag \\
y &= \hat{e}_i(s; \bar{u}_i(s; e(t_{\scriptscriptstyle k_{z+1}})), e(t_{\scriptscriptstyle k_{z+1}})), \notag
\end{aligned}$$ we get that: $$\begin{aligned}
&\|y-x\| = \|\hat{e}(s; \bar{u}(s; e(t_{i+1})), e(t_{i+1})) \notag \\
&\hspace{10mm} - \hat{e}(s; \bar{u}(s; e(t_{i})), e(t_{i})) \in E_{s-t_i} \| \le \rho_i(h).
\end{aligned}$$ Furthermore, by employing Lemma \[lemma:x\_y\_proof\] for $s \in [t_{\scriptscriptstyle k_{z+1}}, t_{\scriptscriptstyle k_{z}}+T_{\scriptscriptstyle z}]$ and the same $x, y$ as previously we get that $ y = \hat{e}_i(s; \bar{u}(s; e_i(t_{\scriptscriptstyle k_{z+1}})), e(t_{\scriptscriptstyle k_{z+1}})) \in E^i_{s-t_{\scriptscriptstyle k_{z+1}}}$, which according to Property 1, implies that $e_i(s; \bar{u}_i(s; e_i(t_{\scriptscriptstyle k_{z+1}})), e_i(t_{\scriptscriptstyle k_{z+1}})) \in E_i$. Thus, Statement 3 holds. Hence, the feasibility at time $t_{\scriptscriptstyle k_{z}}$ implies feasibility at time $t_{\scriptscriptstyle k_{z+1}}$. Therefore, if the ROCP - is feasible at time $t_{\scriptscriptstyle k_{z}}$, i.e., it remains feasible for every $t \in [t_k, t_k + T]$.
*Convergence Analysis*: The second part involves proving convergence of the state $e_i$ to the terminal set $\mathcal{E}_i$. In order to prove this, it must be shown that a proper value function is decreasing along the solution trajectories starting at a sampling time $t_i$. Consider the optimal value function $J_i^\star(e_i(t_{\scriptscriptstyle k_z}))$, as is given in , to be a Lyapunov-like function. Consider also the cost of the feasible control input, indicated by: $$\label{eq:J_bar}
\bar{J}_i(e_i(t_{\scriptscriptstyle k_{z+1}})) \triangleq \bar{J}_i(e_i(t_{\scriptscriptstyle k_{z+1}}), \bar{u}_i(\cdot; e_i(t_{\scriptscriptstyle k_{z+1}}))),$$ where $t_{\scriptscriptstyle k_{z+1}} = t_{\scriptscriptstyle k_z} + h$. Define:
$$\begin{aligned}
&\hspace{-4mm} \bar{u}_1(s) \triangleq \bar{u}_i(s; e_i(t_{\scriptscriptstyle k_{z+1}})), \label{eq:u_1} \\
&\hspace{-3mm} \bar{e}_1(s) \triangleq \bar{e}_i(s; u_1(s), e_i(t_{\scriptscriptstyle k_{z+1}})), s \in [t_{\scriptscriptstyle k_{z+1}}, t_{\scriptscriptstyle k_{z}}+T], \label{eq:e_1}
\end{aligned}$$
where $\bar{e}_1(s)$ stands for the predicted state $e_i$ at time $s$, based on the measurement of the state $e_i$ at time $t_{\scriptscriptstyle k_{z+1}}$, while using the feasible control input $\bar{u}_i(s; e(t_{\scriptscriptstyle k_{z+1}}))$ from . Let us also define the following terms: $$\begin{aligned}
\hat{u}_2(s) &\triangleq \hat{u}_i^\star(s; e_i(t_{\scriptscriptstyle k_z})), \label{eq:u_2}\\
\hat{e}_2(s) &\triangleq \hat{e}_i(s; \hat{u}_2(s), e_i(t_{\scriptscriptstyle k_z})), s \in [t_{\scriptscriptstyle k_{z}}, t_{\scriptscriptstyle k_{z}}+T-h]. \notag
\end{aligned}$$ where $\hat{e}_1(s)$ stands for the predicted state $e_i$ at time $s$, based on the measurement of the state $e_i$ at time $t_{\scriptscriptstyle k_{z}}$, while using the control input $\hat{u}_i(s; e(t_{\scriptscriptstyle k_{z}})), s \in [t_{\scriptscriptstyle k_{z}}, t_{\scriptscriptstyle k_{z}}+T-h]$ from . By employing , and , the difference between the optimal and feasible cost is given by: $$\begin{aligned}
&\bar{J}(e_i(t_{\scriptscriptstyle k_{z+1}})) - J^\star(e_i(t_{\scriptscriptstyle k_z})) = \notag \\
& V_i(\bar{e}_1(t_{\scriptscriptstyle k_{z}}+T) + \int_{t_{\scriptscriptstyle k_{z+1}}}^{t_{\scriptscriptstyle k_{z}}+T} \Big[ F_i(\bar{e}_1(s), \bar{u}_1(s)) \Big] ds \notag \\
&\hspace{-1mm}-V_i(\hat{e}_2(t_{\scriptscriptstyle k_z}+T-h) - \int_{t_{\scriptscriptstyle k_z}}^{t_{\scriptscriptstyle k_z}+T-h} \Big[ F_i(\hat{e}_2(s), \hat{u}_2(s)) \Big] ds \notag \\
&=V_i(\bar{e}_1(t_{\scriptscriptstyle k_{z}}+T)) + \int_{t_{\scriptscriptstyle k_{z+1}}}^{t_{\scriptscriptstyle k_{z+1}}+T-h} \Big[ F_i(\bar{e}_1(s), \bar{u}_1(s)) \Big] ds \notag \\
&+\int_{t_{\scriptscriptstyle k_z}+T-h}^{t_{\scriptscriptstyle k_z}+T} \Big[ F_i(\bar{e}_1(s), \bar{u}_1(s)) \Big] ds -V_i(\hat{e}_2(t_{\scriptscriptstyle k_z}+T - h)) \notag \\
&\hspace{13mm}-\int_{t_{\scriptscriptstyle k_z}}^{t_{\scriptscriptstyle k_{z+1}}} \Big[ F_i(\hat{e}_2(s), \hat{u}_2(s)) \Big] ds \notag \\
&\hspace{23mm}-\int_{t_{\scriptscriptstyle k_{z+1}}}^{t_{\scriptscriptstyle k_z}+T - h} \Big[ F_i(\hat{e}_2(s), \hat{u}_2(s)) \Big] ds. \label{eq:lyap1}
\end{aligned}$$ Note that, from , the following holds: $$\label{eq:verify_u_bar}
\bar{u}_i(s; e_i(t_{\scriptscriptstyle k_{z+1}})) = \hat{u}_i^\star(s; e_i(t_{\scriptscriptstyle k_z})), \forall \ s \in [t_{\scriptscriptstyle k_{z+1}}, t_{\scriptscriptstyle k_z}+T - h].$$ By combining , and , we have that: $$\begin{aligned}
&\bar{u}_1(s) = \hat{u}_2(s) = \bar{u}_i(s; e_i(t_{\scriptscriptstyle k_{z+1}})) = \hat{u}_i^\star(s; e_i(t_{\scriptscriptstyle k_z})), \notag \\
&\hspace{25mm}\forall \ s \in [t_{\scriptscriptstyle k_{z+1}}, t_{\scriptscriptstyle k_z}+T - h], \label{eq:controllers_equality_convergence}
\end{aligned}$$ By applying the last result and the fact that $F_i(e,u)$ is Lipschitz, the following holds: $$\begin{aligned}
&\int_{t_{\scriptscriptstyle k_{z+1}}}^{t_{\scriptscriptstyle k_z}+T - h} \Big[ F_i(\bar{e}_1(s), \bar{u}_1(s)) \Big] ds \notag \\
&\hspace{30mm}-\int_{t_{\scriptscriptstyle k_{z+1}}}^{t_{\scriptscriptstyle k_z}+T - h} \Big[ F_i(\hat{e}_2(s), \hat{u}_2(s)) \Big] ds \notag \\
&=\int_{t_{\scriptscriptstyle k_{z+1}}}^{t_{\scriptscriptstyle k_z}+T - h} \Big[ F_i(\bar{e}_1(s), \bar{u}_1(s)) - F_i(\hat{e}_2(s), \hat{u}_2(s)) \Big] ds \notag
\end{aligned}$$ $$\begin{aligned}
&= \notag \int_{t_{\scriptscriptstyle k_{z+1}}}^{t_{\scriptscriptstyle k_z}+T - h} \Big[ F_i(\bar{e}_1(s), \bar{u}_i(s; e_i(t_{\scriptscriptstyle k_{z+1}}))) \notag \\
&\hspace{30mm}-F_i(\hat{e}_2(s), \bar{u}_i(s; e_i(t_{\scriptscriptstyle k_{z+1}}))) \Big] ds \notag \\
&\le \Bigg\| \int_{t_{\scriptscriptstyle k_{z+1}}}^{t_{\scriptscriptstyle k_z}+T - h} \Big[ F_i(\bar{e}_1(s), \bar{u}_i(\cdot)) - F_i(\hat{e}_2(s), \bar{u}_i(\cdot)) \Big] ds \Bigg\| \notag \\
&\le \int_{t_{\scriptscriptstyle k_{z+1}}}^{t_{\scriptscriptstyle k_z}+T - h} \Big\| F_i(\bar{e}_1(s), \bar{u}_i(\cdot)) - F_i(\hat{e}_2(s), \bar{u}_i(\cdot)) \Big\| ds \notag \\
&\le L_{\scriptscriptstyle F_i} \int_{t_{\scriptscriptstyle k_{z+1}}}^{t_{\scriptscriptstyle k_z}+T - h} \left\| \bar{e}_1(s) - \hat{e}_2(s) \right\| ds. \label{eq:F_1_F_2}
\end{aligned}$$ By employing the fact that $\forall s \in [t_{\scriptscriptstyle k_{z+1}}, t_{\scriptscriptstyle k_z}+T - h]$ the following holds: $$\label{eq:bar_e_equals_hat_e}
\bar{e}_i(s; \bar{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_{z+1}})) = \hat{e}_i(s; \bar{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_z})),$$ the term $\left\| \bar{e}_1(s) - \hat{e}_2(s) \right\|$ can be written as: $$\begin{aligned}
&\left\| \bar{e}_1(s) - \hat{e}_2(s) \right\| \notag \\
&= \|\bar{e}_i(s; \bar{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_{z+1}})) - \hat{e}_i(s; \hat{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_z})) \| \notag \\
&= \Bigg\| \bar{e}_i(t_{\scriptscriptstyle k_{z+1}}; \bar{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_{z+1}})) \notag \\
&+ \int_{t_{\scriptscriptstyle k_{z+1}}}^{s} g_i(\bar{e}_i(s'), \hat{\bar{x}}_i(s') , \bar{u}_i(\cdot))ds' \notag \\
&- \hat{e}_i(t_{\scriptscriptstyle k_z}; \hat{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_z})) - \int_{t_{\scriptscriptstyle k_z}}^{t_{\scriptscriptstyle k_{z+1}}} g_i(\hat{e}_i(s), \hat{\bar{x}}_i(s'), \hat{u}_i(\cdot))ds \notag \\
&\hspace{25mm}-\int_{t_{\scriptscriptstyle k_{z+1}}}^{s} g_i(\hat{e}_i(s'), \hat{\bar{x}}_i(s'), \bar{u}_i(\cdot))ds' \Bigg\| \notag \\
&\le \Bigg\| e_i(t_{\scriptscriptstyle k_{z+1}}) -e_i(t_{\scriptscriptstyle k_z}) - \int_{t_{\scriptscriptstyle k_z}}^{t_{\scriptscriptstyle k_{z+1}}} g_i(\hat{e}_i(s'), \hat{\bar{x}}_i(s'), \hat{u}_i(\cdot))ds' \Bigg\| \notag \\
&+\Bigg\| \int_{t_{\scriptscriptstyle k_{z+1}}}^{s} g_i(\bar{e}_i(s'), \hat{\bar{x}}_i(\cdot), \bar{u}_i(\cdot))ds' \notag \\
&\hspace{20mm} -\int_{t_{\scriptscriptstyle k_{z+1}}}^{s} g_i(\hat{e}_i(s'), \hat{\bar{x}}_i(\cdot), \bar{u}_i(\cdot))ds' \Bigg\| \notag \\
&= \Bigg\| e_i(t_{\scriptscriptstyle k_{z+1}}) -e_i(t_{\scriptscriptstyle k_z}) - \int_{t_{\scriptscriptstyle k_z}}^{t_{\scriptscriptstyle k_{z+1}}} \frac{d}{dt} \left[ \hat{e}_i(s; \hat{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_z}))\right] ds \Bigg\| \notag \\
&+\Bigg\| \int_{t_{\scriptscriptstyle k_{z+1}}}^{s} \frac{d}{dt} \left[ \bar{e}_i(s'; \bar{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_{z+1}}))\right]ds' \notag \\
&\hspace{25mm} -\int_{t_{\scriptscriptstyle k_{z+1}}}^{s} \frac{d}{dt} \left[ \hat{e}_i(s'; \bar{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_z}))\right]ds' \Bigg\| \notag \\
&= \big\| e_i(t_{\scriptscriptstyle k_{z+1}}) -e_i(t_{\scriptscriptstyle k_z}) - \hat{e}_i(t_{\scriptscriptstyle k_{z+1}}; \hat{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_z})) \notag \\
&\hspace{40mm}+\hat{e}_i(t_{\scriptscriptstyle k_z}; \hat{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_z})) \big\| \notag \\
&+\big\| \bar{e}_i(s; \bar{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_{z+1}}))-\bar{e}_i(t_{i+1}; \bar{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_{z+1}})) \notag \\
&- \hat{e}_i(s; \bar{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_z}))+\hat{e}_i(t_{\scriptscriptstyle k_{z+1}}; \bar{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_z})) \big\| \notag \\
&= \left\| e_i(t_{\scriptscriptstyle k_{z+1}}) -e_i(t_{\scriptscriptstyle k_z}) - \hat{e}_i(t_{\scriptscriptstyle k_{z+1}}; \hat{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_z})) +e_i(t_{\scriptscriptstyle k_z}) \right\| \notag \\
&= \left\| e_i(t_{\scriptscriptstyle k_{z+1}})- \hat{e}_i(t_{\scriptscriptstyle k_{z+1}}; \hat{u}_i(\cdot), e_i(t_{\scriptscriptstyle k_z}))\right\|, \notag
\end{aligned}$$ which, by employing Lemma \[lemma:e-hat\_e\], leads to: $$\left\| \bar{e}_1(s) - \hat{e}_2(s) \right\| \le \rho_i(t_{\scriptscriptstyle k_{z+1}}-t_{\scriptscriptstyle k_z}) = \rho_i(h).$$ By combining the last result with we get: $$\begin{aligned}
&\int_{t_{\scriptscriptstyle k_{z+1}}}^{t_{\scriptscriptstyle k_z}+T - h} \Big[ F_i(\bar{e}_1(s), \bar{u}_1(s)) \Big] ds \notag \\
&\hspace{27mm}-\int_{t_{\scriptscriptstyle k_{z+1}}}^{t_{\scriptscriptstyle k_z}+T - h} \Big[ F_i(\hat{e}_2(s), \hat{u}_2(s)) \Big] ds \notag \\
&\le L_{\scriptscriptstyle F_i} \int_{t_{\scriptscriptstyle k_{z+1}}}^{t_{\scriptscriptstyle k_z}+T - h} \rho_i(h) ds = (T - 2h) \rho_i(h) L_{\scriptscriptstyle F_i} . \label{eq:F_1_F_23}
\end{aligned}$$ By combining the last result with , becomes: $$\begin{aligned}
&\bar{J}(e_i(t_{\scriptscriptstyle k_{z+1}})) - J^\star(e_i(t_{\scriptscriptstyle k_z})) \le (T -2 h) \rho_i(h) L_{\scriptscriptstyle F_i} \notag \\
&\hspace{13mm}+V_i(\bar{e}_1(t_{\scriptscriptstyle k_{z}}+T)) - V_i(\hat{e}_2(t_{\scriptscriptstyle k_z}+T-h))\notag \\
&\hspace{20mm}+\int_{t_{\scriptscriptstyle k_z}+T-h}^{t_{\scriptscriptstyle k_{z}}+T} \Big[ F_i(\bar{e}_1(s), \bar{u}_1(s)) \Big] ds \notag \\
&\hspace{30mm}-\int_{t_{\scriptscriptstyle k_z}}^{t_{\scriptscriptstyle k_{z+1}}} \Big[ F_i(\hat{e}_2(s), \hat{u}_2(s)) \Big] ds \label{eq:lyap_2}
\end{aligned}$$ By integrating inequality from $t_{\scriptscriptstyle k_z}+T - h$ to $t_{\scriptscriptstyle k_{z}}+T$ and we get the following: $$\begin{aligned}
&\hspace{-3mm} \int_{t_{\scriptscriptstyle k_z}+T - h}^{t_{\scriptscriptstyle k_{z}}+ T} \Big[ \frac{\partial V}{\partial{e}} \cdot g_i(\bar{e}_1(s), \hat{\bar{x}}_i(s), \bar{u}_1(s)) \notag \\
&\hspace{30mm}+ F_i(\bar{e}_1(s), \bar{u}_1(s)) \Big] ds \le 0 \notag \\
\Leftrightarrow & V_i(\bar{e}_1(t_{\scriptscriptstyle k_{z}}+T) - V_i(\bar{e}_1(t_{\scriptscriptstyle k_z}+T-h)) \notag \\
&\hspace{22mm} + \int_{t_{\scriptscriptstyle k_z}+ T - h}^{t_{\scriptscriptstyle k_{z}}+ T} \Big[F_i(\bar{e}_1(s), \bar{u}_1(s)) \Big] ds \le 0, \notag
\end{aligned}$$ which by adding and subtracting the term $V_i(\hat{e}_2(t_{\scriptscriptstyle k_z}+T-h))$ becomes: $$\begin{aligned}
& V_i(\bar{e}_1(t_{\scriptscriptstyle k_{z}}+T) - V_i(\hat{e}_2(t_{\scriptscriptstyle k_z}+T-h)) \notag \\
&\hspace{10mm} + \int_{t_{\scriptscriptstyle k_z}+ T - h}^{t_{\scriptscriptstyle k_{z}}+T} \Big[F_i(\bar{e}_1(s), \bar{u}_1(s)) \Big] ds \le \notag \\
&\hspace{15mm} V_i(\bar{e}_1(t_{\scriptscriptstyle k_z}+T-h))-V_i(\hat{e}_2(t_{\scriptscriptstyle k_z}+T-h)). \notag
\end{aligned}$$ By employing the property $y \le |y|, \forall y \in \mathbb{R}$, we get: $$\begin{aligned}
& V_i(\bar{e}_1(t_{\scriptscriptstyle k_{z}}+T) - V_i(\hat{e}_2(t_{\scriptscriptstyle k_z}+T-h)) \notag \\
&\hspace{4mm} + \int_{t_{\scriptscriptstyle k_z}+ T - h}^{t_{\scriptscriptstyle k_{z}}+T} \Big[F_i(\bar{e}_1(s), \bar{u}_1(s)) \Big] ds \le \notag \\
&\hspace{4mm} \left|V_i(\bar{e}_1(t_{\scriptscriptstyle k_z}+T-h)) - V_i(\hat{e}_2(t_{\scriptscriptstyle k_z}+T-h)) \right|. \label{eq:V_e_1-V_e_2}
\end{aligned}$$ By employing Lemma \[eq:lemma\_L\_v\], we have that: $$\begin{aligned}
&\left|V_i(\bar{e}_1(t_{\scriptscriptstyle k_z}+T-h)) - V_i(\hat{e}_2(t_{\scriptscriptstyle k_z}+T-h)) \right| \le \notag \\
&\hspace{18mm} L_{\scriptscriptstyle V_i} \|\bar{e}_1(t_{\scriptscriptstyle k_z}+T-h) - \hat{e}_2(t_{\scriptscriptstyle k_z}+T-h) \|,
\end{aligned}$$ which by employing Lemma \[lemma:bounded\_trajectories\] and , becomes: $$\begin{aligned}
&\left|V_i(\bar{e}_1(t_{\scriptscriptstyle k_z}+T-h)) - V_i(\hat{e}_2(t_{\scriptscriptstyle k_z}+T-h)) \right| \le \notag \\
&\hspace{18mm} L_{\scriptscriptstyle V_i} \rho_i(t_{\scriptscriptstyle k_{z+1}}-t_{\scriptscriptstyle k_z}) = \rho_i(h) L_{\scriptscriptstyle V_i}.
\end{aligned}$$ By combining the last result with , we get: $$\begin{aligned}
& V_i(\bar{e}_1(t_{\scriptscriptstyle k_{z}}+T) - V_i(\hat{e}_2(t_{\scriptscriptstyle k_z}+T-h)) \notag \\
&\hspace{4mm} + \int_{t_{\scriptscriptstyle k_z}+ T - h}^{t_{\scriptscriptstyle k_{z}}+T} \Big[F_i(\bar{e}_1(s), \bar{u}_1(s)) \Big] ds \le \rho_i(h) L_{\scriptscriptstyle V_i}. \notag
\end{aligned}$$ The last inequality along with leads to: $$\begin{aligned}
&\bar{J}(e(t_{\scriptscriptstyle k_{z+1}})) - J^\star(e(t_{\scriptscriptstyle k_z})) \le (T -2 h) \rho_i(h) L_{\scriptscriptstyle F_i} +\rho_i(h) L_{\scriptscriptstyle V_i} \notag \\
&\hspace{25mm}-\int_{t_{\scriptscriptstyle k_z}}^{t_{\scriptscriptstyle k_{z+1}}} \Big[ F_i(\hat{e}_2(s), \hat{u}_2(s)) \Big] ds. \label{eq:lyap_3}
\end{aligned}$$ By substituting $e_i = \hat{e}_2(s), u_i = \hat{u}_2(s)$ in we get $F_i(\hat{e}_2(s), \hat{u}_2(s)) \ge m_i \|\hat{e}_2(s)\|^2$, or equivalently: $$\begin{aligned}
&\int_{t_{\scriptscriptstyle k_z}}^{t_{\scriptscriptstyle k_{z+1}}} \Big[ F_i(\hat{e}_2(s), \hat{u}_2(s)) \Big] ds \ge \underline{m}_i \int_{t_{\scriptscriptstyle k_z}}^{t_{\scriptscriptstyle k_{z+1}}} \|\hat{e}_2(s)\|^2 ds \notag \\
\Leftrightarrow &-\int_{t_{\scriptscriptstyle k_z}}^{t_{\scriptscriptstyle k_{z+1}}} \Big[ F(\hat{e}_2(s), \hat{u}_2(s)) \Big] ds \le -\underline{m}_i \int_{t_{\scriptscriptstyle k_z}}^{t_{\scriptscriptstyle k_{z+1}}} \|\hat{e}_2(s)\|^2 ds.
\end{aligned}$$ By combining the last result with , we get: $$\begin{aligned}
&\bar{J}_i(e(t_{\scriptscriptstyle k_{z+1}})) - J^\star_i(e(t_{\scriptscriptstyle k_z})) \le (T -2 h) \rho_i(h) L_{\scriptscriptstyle F_i} +\rho_i(h) L_{\scriptscriptstyle V_i} \notag \\
&\hspace{35mm}-\underline{m}_i \int_{t_{\scriptscriptstyle k_z}}^{t_{\scriptscriptstyle k_{z+1}}} \|\hat{e}_2(s)\|^2 ds \label{eq:lyap_4}
\end{aligned}$$ It is clear that the optimal solution at time $t_{\scriptscriptstyle k_{z+1}}$ i.e., $J^\star(e_i(t_{\scriptscriptstyle k_{z+1}}))$ will not be worse than the feasible one at the same time i.e. $\bar{J}(e_i(t_{\scriptscriptstyle k_{z+1}}))$. Therefore, implies: $$\begin{aligned}
&J^\star_i(e_i(t_{\scriptscriptstyle k_{z+1}})) - J^\star_i(e_i(t_{\scriptscriptstyle k_z})) \le (T -2 h) \rho_i(h) L_{\scriptscriptstyle F_i} +\rho_i(h) L_{\scriptscriptstyle V_i} \notag \\
&\hspace{35mm}-\underline{m}_i \int_{t_{\scriptscriptstyle k_z}}^{t_{\scriptscriptstyle k_{z+1}}} \|\hat{e}_2(s)\|^2 ds,
\end{aligned}$$ which is equivalent to: $$\begin{aligned}
&J^\star_i(e_i(t_{\scriptscriptstyle k_{z+1}})) - J^\star_i(e_i(t_{\scriptscriptstyle k_z})) \le \notag \\
&\hspace{10mm}-m_i \int_{t_{\scriptscriptstyle k_z}}^{t_{\scriptscriptstyle k_{z+1}}} \|\hat{e}_i(s; \hat{u}_i^\star(s; e_i(t_{\scriptscriptstyle k_z})), e_i(t_{\scriptscriptstyle k_z}))\|^2 ds \notag \\
&\hspace{20mm}+(T -2 h) \rho_i(h) L_{\scriptscriptstyle F_i} +\rho_i(h) L_{\scriptscriptstyle V_i}.
\end{aligned}$$ which, according to , is in the form: $$\begin{aligned}
&J^\star_i(e_i(t_{\scriptscriptstyle k_{z+1}})) - J^\star_i(e_i(t_{\scriptscriptstyle k_z})) \le -\alpha(\|e_i\|)+\sigma(\|\bar{x}_i\|) \label{eq:lyap_7}
\end{aligned}$$ Thus, the optimal cost $J$ has been proven to be decreasing, and according to Definition \[def:ISS\] and Theorem \[def:ISS\_lyapunov\], the closed loop system is ISS stable. Therefore, the closed loop trajectories converges to the closed set $\mathcal{E}_{i}$.
[^1]: Alexandros Nikou, Christos Verginis and Dimos V. Dimarogonas are with the ACCESS Linnaeus Center, School of Electrical Engineering, KTH Royal Institute of Technology, SE-100 44, Stockholm, Sweden and with the KTH Center for Autonomous Systems. Email: [{anikou, cverginis, dimos}@kth.se]{}. Shahab Heshmati-alamdari is with the Control Systems Lab, Department of Mechanical Engineering, National Technical University of Athens, 9 Heroon Polytechniou Street, Zografou 15780, Athens, Greece. Email: [{shahab}@mail.ntua.gr]{}. This work was supported by the H2020 ERC Starting Grant BUCOPHSYS, the Swedish Research Council (VR), the Swedish Foundation for Strategic Research, the Knut och Alice Wallenberg Foundation, the European Union’s Horizon 2020 Research and Innovation Programme under the Grant Agreement No. 644128 (AEROWORKS) and the EU H2020 Research and Innovation Programme under GA No. 731869 (Co4Robots).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The Fermi edge singularity and the Anderson orthogonality catastrophe describe the universal physics which occurs when a Fermi sea is locally quenched by the sudden switching of a scattering potential, leading to a brutal disturbance of its ground state. We demonstrate that the effect can be seen in the controllable domain of ultracold trapped gases by providing an analytic description of the out-of-equilibrium response to an atomic impurity, both at zero and at finite temperature. Furthermore, we link the transient behavior of the gas to the decoherence of the impurity, and, in particular to the amount of non-markovianity of its dynamics.'
author:
- 'A. Sindona$^{1,2}$, J. Goold$^{3,4}$, N. Lo Gullo$^{4,5}$, S. Lorenzo$^{1,2}$, F. Plastina$^{1,2}$'
title: 'Orthogonality catastrophe and decoherence in a trapped-Fermion environment'
---
A Fermi gas may be shaken-up by the switching of even a single, weakly interacting impurity, producing a complete rearrangement of the many body wave-function that, as a consequence, loses essentially any overlap with the initial, unperturbed one. This is the essence of Anderson’s orthogonality catastrophe , witnessed by the singular (edge-like) behavior of the energy distribution of the impurity induced excitations. An example of how such a many-body effect comes into play is provided by X-ray photoemission spectra from most simple metals, where the expected sharp symmetric peak at the binding energy of a core level is converted into a power law singularity, as predicted by the Mahan-Nozi[è]{}res-Dominics (MND) theory [@Mahan1; @Nozieres]. Similar patterns have been observed in electron emission via X-ray absorption and Auger neutralization from carbon based nanomaterials [@AntoEdge], and for quantum dots [@heyl]. Fermi edge resonance and orthogonality catastrophe have been also revealed by non-equilibrium current fluctuations (shot noise) in nanoscale conductors [@novotny], and enter prominently the physics of phenomena as diverse as the Kondo effect and the scattering or sticking of a low-energy atom or ion on a metal surface [@scatt; @stick].
Recently, it has been proposed to observe this universal physics in controllable ultracold atomic setups where the singular behavior may be probed either in the time domain, by Ramsey interference type experiments performed on the impurity atom [@gooldOC], or in the frequency domain, by radio-frequency spectroscopy [@knap]. However, an analytic framework for the case of a *trapped* Fermi gas is lacking. In this letter we provide such an analytic description, and discuss the transient response of a harmonically trapped Fermi gas following the ‘*sudden*’ switching of an embedded two-level atom excited by a fast pulse. The interaction with the *local* impurity produces a local quench of the gas, giving rise to the Anderson catastrophe. We study the Fermi-edge physics at zero and finite temperature and both in the frequency domain, by looking at the excitation spectrum of the gas, and in the time domain, by analyzing the dynamics of the impurity. Thus, we link the Fermi edge behavior of the excitation energy distribution to the decoherence of the impurity. In particular, we investigate the Loschmidt echo [@diciotto; @diciassette] and the non-Markovianity, using recently developed tools [@breuermeasure; @rivas; @fisherNM; @geo; @pinja], employed so far to study open systems in different environments, ranging from spins [@noi] to Bose-Einstein condensates [@haikka], and experimentally tested in optical set-ups [@experimentNM; @chiuri]. We find that the non-Markovianity of the decoherent dynamics of the impurity provides a novel interpretation of the essential physics of the shake-up process.
We consider a gas of non-interacting cold fermions confined by a one-dimensional trapping harmonic potential of frequency $\omega
$, and described by the Hamiltonian $\hat{H}_{0} = \sum_{n,\xi}
{\varepsilon_{n}}\hat{c}_{n\xi}^{{\dagger}} \hat{c}_{n\xi}$, with $\hat{c}_{n\xi}$ being the annihilation operator for the $n$-th single particle state of energy ${\varepsilon_{n}}= {\hbar\omega}(n+1/2)$ and spin $\xi$. We add a two-level impurity (an atom of a different species from the trapped component), with internal states ${\left|{g}\right\rangle}$ and ${\left|{e}\right\rangle}$ and Hamiltonian $\Hat{H}_{{\text{\textsc{i}}}}=\sum_{i=e,g} \epsilon_i \, {\left|{i}\right\rangle} {\left\langle{i}\right|}$, which is trapped in an auxiliary potential and brought in contact with the Fermi gas. Such a situation can be achieved using a species selective dipole potential that has a frequency much greater than the trap which contains the gas, so that the impurity motion is essentially frozen. We assume that when the impurity is in the ${\left|{g}\right\rangle}$ state, it has a negligible scattering interaction with the gas, hence the Hamiltonian of the composite system is given by $\hat{H}=\hat{H}_{0} +
\Hat{H}_{{\text{\textsc{i}}}}+\hat{V} \otimes {\left|{e}\right\rangle}{\left\langle{e}\right|}$. With the fermions in their equilibrium configuration, set by $\hat{H}_{0}$, we suppose the impurity to be quickly excited so that the gas feels a sudden perturbation $\hat{V}(t)=\hat{V} \theta(t)$ due to the interaction, assumed to have an $s$-wave like character.
As it is standard in ultracold atoms, at sufficiently low temperatures, the pseudo-potential approximation for the interaction is invoked, which amounts to replacing the complicated atomic interaction potential with an effective short range potential of strength $V_{0}$, localized at the minimum of the harmonic well, which we scale with the trap length $x_{0}$ such that $V(x)=\pi V_{0}x_{0}\delta (x)$. Due to the parity of the single particle wave-functions, only the fermions lying in even-parity states ($n=2r$, with $r=0,1,2,\cdots$) feel the impurity and are involved in the shake-up process. Explicitly, the fermion-impurity interaction is given by $\hat{V}=\sum_{r,{r^{{\prime}}},\xi}
V_{r{r^{{\prime}}}} \hat{c}^{{\dagger}}_{2r\xi} \hat{c}_{2{r^{{\prime}}}\xi}$, where $V_{r{r^{{\prime}}}}=V_{0}{\,}(-1)^{r+{r^{{\prime}}}}{\,}\gamma_r^{1/2}{\,}\gamma_{{r^{{\prime}}}}^{1/2}$, and $\gamma_r=2^{-2r} \pi^{1/2} (2r)!/r!^{2}$ (See appendix A). We label the highest occupied level by $n_F= 2 {r_{{\text{\textsc{f}}}}}$, with ${r_{{\text{\textsc{f}}}}}$ a positive integer, so that the Fermi energy reads ${\varepsilon_{{\text{\textsc{f}}}}}= {\hbar\omega}(2 {r_{{\text{\textsc{f}}}}}+1/2)$.
A key quantity for the following is the *vacuum persistence amplitude*$$\nu _{\beta }(t>0) = {\big\langle{e^{\frac{i}{\hbar}\hat{H}_{0}t}{\,}e^{-\frac{i}{\hbar} (\hat{H}_{0}+\hat{V}) t}}\big\rangle}\text{,}
\label{Eq:VacAmp}$$ with $\langle{\cdots}\rangle$ denoting the grand canonical average over the unperturbed fermion state. $\nu _{\beta }(t)$ is the probability amplitude that the gas will retrieve its equilibrium state at time $t$, after the switching on of the perturbation and, as discussed below, its modulus gives the decoherence factor for the impurity.
The Fourier transform $\tilde{\nu}_{\beta }(E)$, subject to the constraint $\nu _{\beta }(t<0)=\nu _{\beta
}^{\ast}(-t)$, gives the excitation spectrum of the gas. Turning to the interaction picture, we get $$\nu_{\beta }(t)={\big\langle{T e^{\frac{1}{i \hbar
}\int_{0}^{t}d{t^{{\prime}}}\tilde{V}({t^{{\prime}}})}}\big\rangle} \text{,} \quad
\tilde{V}(t)=e^{\frac{i}{\hbar
}\hat{H}_{0}t}\hat{V}e^{-\frac{i}{\hbar }\hat{H}_{0}t}\text{,}
\label{IP}$$ which, by virtue of the linked cluster theorem, reduces to an exponential sum of connected Feynman diagrams, $\nu _{\beta }(t)=e^{\Lambda _{\beta}(t)}$, with:\
\
The closed graphs in $\Lambda _{\beta}(t)$ contain products of vertices ($V_{r{r^{{\prime}}}}$) connected by lines ($G_{r}^{\beta }$) that represent the unperturbed propagators $$i \hbar G_{r}^{\beta }(t)=e^{-i {\varepsilon_{2r}}t/\hbar} [ \theta(t)\,f^{-}_{r}
- \theta(-t)\,f^{+}_{r} ], \label{prop}$$ where $f^{\pm}_{r}=[1+e^{\pm \beta ({\varepsilon_{2r}}-\mu)}]^{-1}$ are the particle-hole distributions, and $\mu$ denotes the chemical potential (see Appendix C).
We focus on the lowest order loops, namely $$\label{Lambda1} \hbar\Lambda _{1}^{\beta }(t)=-i t \, \chi_s
V_{0}\lambda_{+}^{\beta}(0),$$ and $$\label{Lambda2} \hbar^{2}\Lambda _{2}^{\beta }(t)=- \chi_s
V_{0}^{2} {}
\int_{0}^{t}d{t^{{\prime}}}\int_{0}^{{t^{{\prime}}}}d{t^{{\prime}{\hskip-0.5pt}{\prime}}}\,\lambda_{+}^{\beta
}({t^{{\prime}{\hskip-0.5pt}{\prime}}}){\,}\lambda_{-}^{\beta }({t^{{\prime}{\hskip-0.5pt}{\prime}}}),$$ with $\chi_s={(2s+1)}$ accounting for the spin degeneracy and $
\lambda_{\pm }^{\beta
}(t)=\sum_{r=0}^{\infty}{\,}\gamma_r{\,}e^{\pm 2 i r \omega t}
f^{\pm}_{r}$.
This approximation will prove to accurately describe the singular response of the gas (contained in the two-vertex term) and to give the dominant contribution to the shake-up process if the interaction strength is small in the energy scale of the problem. The latter is set by both the level separation ${\hbar\omega}$ and Fermi energy ${\varepsilon_{{\text{\textsc{f}}}}}$, which allow us to introduce $\alpha =\frac{\chi_s
V_{0}^{2}}{2{\hbar\omega}{\varepsilon_{{\text{\textsc{f}}}}}}$ as a sensible interaction strength parameter.
-12pt
-12pt
The contribution (\[Lambda1\]) may be written as $\hbar \Lambda _{1}^{\beta }(t)=-i t E_{1}^{\beta} $. Here, $$E_{1}^{\beta}= \sqrt{2 \chi_s
{\hbar\omega}{\varepsilon_{{\text{\textsc{f}}}}}\alpha}{\,}\sum_{r=0}^{\infty}{\,}\gamma_r{\,}f^{+}_{r}
\label{eq:E1}$$ is the first-order shift to the gas energy, as provided by the Rayleigh-Schrödinger perturbation theory. The behavior of the unperturbed energy $E_{0}^{\beta}=\chi_s
\sum_{n}{\varepsilon_{n}}\,f^{+}_{n/2}$, and of its first and second order corrections (the latter obtained from the two vertex term, see below) vs ${\varepsilon_{{\text{\textsc{f}}}}}$ is shown in Fig. \[eradue\] for various temperatures. We notice that $E_{0}^{\beta}$ is $1$ to $3$ orders of magnitude larger than $E_{1}^{\beta }$ for $\alpha \lesssim 1$, while temperature plays an appreciable role in both $E_{0}^{\beta}$ and $E_{1}^{\beta}$ for $\beta{\hbar\omega}$ less than $\sim
0.05$.
While $\Lambda _{1}^{\beta}(t)$ only brings a phase factor to $\nu_{\beta }(t)$, which corresponds to shifting the spectrum $\tilde{\nu}_{\beta }(E)$ by $E_{1}^{\beta}$, the two-vertex connected graph gives the crucial contribution to the persistence amplitude. As detailed in Appendix B, it can be split into three parts with well defined trends and physical meaning, i.e., $\Lambda _{2}^{\beta}(t)= \Lambda _{2{\text{\textsc{s}}}}^{\beta }(t) +
\Lambda_{2{\text{\textsc{g}}}}^{\beta }(t) + \Lambda _{2{\text{\textsc{p}}}}^{\beta }(t)$. These represent a (further) energy shift, a gaussian envelope due to finite temperature effects, and a periodic terms originating from the equal spacing of the unperturbed single-particle states, respectively, and are separately analyzed in Figs. \[eradue\]**C**, \[eratre\]**A**, and \[eratre\]**B**.
The first one, $\hbar\Lambda_{2{\text{\textsc{s}}}}^{\beta}(t)=-itE_{2}^{\beta
}$, provides the second-order correction to the energy of the gas (the $n>2$-vertex graphs would complete the perturbation series): $$E_{2}^{\beta } =\alpha {\varepsilon_{{\text{\textsc{f}}}}}\sum\limits_{r{\neq}{r^{{\prime}}}=0}^{\infty}
\frac{f^{+}_{r}\gamma_r\,\gamma_{{r^{{\prime}}}}f^{-}_{{r^{{\prime}}}}}{r-{r^{{\prime}}}}.
\label{Delta2B}$$ Comparing Fig. \[eradue\]**B** and **C**, we notice that the chosen value of $\alpha$ let $E_{2}^{\beta}$ take absolute values smaller than $E_{1}^{\beta }$. However, $E_{2}^{\beta}$ is more sensitive to temperature than $E_{1}^{\beta }$ for $\beta {\hbar\omega}<0.05$.
The second contribution, $
\Lambda_{2{\text{\textsc{g}}}}^{\beta}(t)=-\delta_{\beta}^{2} \omega^2
t^{2}/2,\label{Lambda2G}$ produces a Gaussian damping in $\nu_{\beta }(t)$ and, therefore, a Gaussian broadening in $\tilde{\nu}_{\beta}(E)$ with standard deviation $$\delta_{\beta} = \sqrt{2 \alpha g_{\beta }},\quad g_{\beta }=\frac{{\varepsilon_{{\text{\textsc{f}}}}}}{{\hbar\omega}} \sum_{r=0}^{\infty}{\,}\gamma_r^{2}\,f^{+}_{r}\,f^{-}_{r}.
\label{eq:gbeta}$$ Here, the coefficient $g_{\beta}$ is weakly influenced by the Fermi energy, but strongly affected by temperature, changing by various orders of magnitude for $\beta {\hbar\omega}\lesssim 0.5$. No damping/broadening effects are present at the absolute zero, since $\delta_{\beta} \to 0$ for $\beta {\hbar\omega}\to
\infty$ (Fig.\[eratre\]**A**).
The most important content of the second diagram, giving a non trivial structure to $\nu_{\beta }(t)$, arises from the third contribution (see the Appendix): $$\Lambda _{2{\text{\textsc{p}}}}^{\beta }(t)=-\frac{\alpha {\varepsilon_{{\text{\textsc{f}}}}}}{2{\hbar\omega}}\sum\limits_{r{\neq }{r^{{\prime}}}}^{\infty}\,\gamma_r f_{r}^{+}\,\frac{1-e^{2i(r-{r^{{\prime}}})t\omega }}{(r-{r^{{\prime}}})^2}\,\gamma_{{r^{{\prime}}}}f_{{r^{{\prime}}}}^{-}.
\label{Lambda2E}$$ Due to the harmonic form of trapping potential, this is a periodic function of time with frequency $2\omega $, see Fig. \[eratre\]**B**. The zeroes of this sub-graph (at $\omega t = m \pi$ with $m=0,\pm1,\pm2$), when combined with the Gaussian damping (\[eq:gbeta\]), yield modulations in the vacuum persistence amplitude which, as discussed below, are a signature of non-markovian dynamics of the impurity.
Leaving aside the shifts, the persistence amplitude is then:$$\nu_{\beta }^{{\prime}}(t)=e^{-\delta_{\beta } \omega^2
t^{2}/2}e^{\Lambda _{2{\text{\textsc{p}}}}^{\beta }(t)}\text{.} \label{Vpbet}$$ Of particular interest for the discussion below is the behavior of $|\nu _{\beta}(t)|$ exhibiting spikes at $\omega t \sim \pi ,2\pi
,\cdots $, which become more and more pronounced with increasing $\beta {\hbar\omega}$, see left panels in Fig. \[cinque\]. The periodicity in the time domain is reflected in the excitation spectrum $\tilde{\nu}_{\beta }(E)$ that offers an asymmetric, broadened, signature of the singular behavior of the Fermi gas. The monotonic structure turns into a sequence of sub-peaks, separated by $2 {\hbar\omega}$ and related to even-level transitions in the gas as $\beta {\hbar\omega}$ gets above $\sim 0.5$ (see Fig. \[cinque\]**B**). These features are observed for any ${r_{{\text{\textsc{f}}}}}$ in the range of $5$ to $100$ (see Appendix C).
The coefficient (\[eq:gbeta\]) of the Gaussian power law and the periodic contribution $\Lambda _{2{\text{\textsc{p}}}}^{\beta }(t)$ can be approximated as $$g_{\beta } \approx 2 \sum_{m=1}^{\infty } (-1)^{m} m
\frac{e^{\beta {\hbar\omega}m/2}}{e^{\beta {\hbar\omega}m}-1}, \label{Lambda2GA}$$ and $$\Lambda _{2{\text{\textsc{p}}}}^{\beta }(t) \approx \alpha
\sum_{m=-\infty}^{\infty } \ln \frac{e^{2\tau _{m}\omega
}-1}{e^{2(it+\tau _{m})\omega }-1}. \label{LambXX}$$ At low temperatures, the leading behavior of the Gaussian standard deviation is $\delta _{\beta }\approx 2\alpha^{1/2} e^{-\beta
{\hbar\omega}/4}$ for thermal energies $\beta {\hbar\omega}\gtrsim 6$ (see Fig. \[eratre\]**A**). On the other hand, Eq. (\[LambXX\]) contains a singularity at the absolute zero, that we regularized by introducing a cut-off parameter $\tau_{0}$. This regularization is only needed to remove a zero temperature indefiniteness of the analytic approximation, whereas the numerical evaluation of the vacuum persistence amplitude does not suffer from divergence problems. As shown below and as detailed in the Appendices below, a similar parameter enters the original MND theory, and we can interpret it as the typical time-scale over which transitions occur in the gas. On the other hand, thermal fluctuations introduce other characteristic times $\tau_{m}=m\beta\hbar$.
-10pt
-8pt -8pt
-8pt
Taking $g_{\beta}$ and $\Lambda _{2{\text{\textsc{p}}}}^{\beta }(t)$ as in Eqs. (\[Lambda2GA\]) and (\[LambXX\]), and using them in Eq. (\[Vpbet\]) gives an accurate approximation to the numerical results for $\beta {\hbar\omega}\gtrsim 0.1$, number of particles larger than $10$ and for suitable values of the cut-off parameter, say, $\omega\tau_0 < 0.02$ (see Fig. \[nonmark\]**A**). In particular, at $T=0$, the vacuum persistence amplitude takes the form: $$\nu_{\beta \to \infty }^{{\prime}}(t) \approx
\left[\frac{e^{2\tau_{0}\omega }-1}{e^{2\omega
(\tau_{0}+it)}-1}\right]^{\alpha} \text{.} \label{Vbet0}$$
To compare our findings to the one-dimensional free-fermion theory, one needs to fix $\alpha $ and let the harmonic frequency go to zero by keeping the number of particles in the gas ($2{r_{{\text{\textsc{f}}}}}\approx {\varepsilon_{{\text{\textsc{f}}}}}/{\hbar\omega}$) finite. No Gaussian damping occurs in this case, and the two vertex graph tends to $$\Lambda_{{\text{\textsc{mnd}}}}(t)=-\alpha \ln (it/\tau _{0}+1),
\label{MNDLamb}$$ yielding the Nozieres-De Dominicis propagator $\nu_{{\text{\textsc{mnd}}}}(t)=\left( it/\tau _{0}+1\right)^{-\alpha}$, originally calculated for a suddenly switched on core-hole in a free electron gas [@Nozieres]. Eq. was obtained by writing down a generalized Dyson equation for the electron Green’s function in a constant window potential of width $\hbar/\tau _{0}$, and solving it for all connected graphs in the long-time limit. For this reason, the MND spectrum lacks formal justification away from the threshold. In the present derivation, we have taken into account the full perturbation at an arbitrary time $t>0$, retaining only the first non adiabatic contribution in the linked cluster expansion [@schotte]. We expect the effect of higher order diagrams to be mainly concerned with the adiabatic correction to the equilibrium energy and some additional broadening of the excitation peaks. The latter should provide a renormalization to the critical parameter. Nevertheless, in the investigated ranges of temperatures and particle numbers, the definition of $\alpha$ given here produces a markedly singular response with the same range of criticality as the MND edge response parameter ($\alpha =0-1$).
From this comparison with the free-gas case, we learn that the trapping frequency $\omega$ enters crucially the physics of the shake up process. Indeed, it modifies the long time response of the gas as all single particle excitations involve energy exchanges which are now even multiples of $\hbar \omega$. This gives rise to the periodic part of the fermion response and to the corresponding spectral peaks which are then broadened at finite temperatures due to the gaussian envelope, the latter being a typical effect of suddenly switched perturbations .
-8pt
-8pt
Up to now, we treated the response of the Fermi gas without any reference to the dynamics of the impurity, that has just been assumed in the excited states for $t\geq 0$. If, instead, the two-level atom is subject to (say) a fast $\pi /2$ pulse and quickly prepared in the superposition $({\left|{g}\right\rangle}+{\left|{e}\right\rangle})/\sqrt{2}$, it experiences a purely dephasing dynamics due to the coupling with the gas, such that its state at later times is $\rho_{{\text{\textsc{imp}}}}(t)=( {\left|{g}\right\rangle} {\left\langle{g}\right|} + {\left|{e}\right\rangle}
{\left\langle{e}\right|} + \nu _{\beta }(t) {\left|{g}\right\rangle} {\left\langle{e}\right|} + \mbox{h.c.})/2$. The decoherence factor that enters the off-diagonal elements is just the persistence amplitude that we obtained before, going to zero at long times due to the orthogonality catastrophe. In the theory of open systems, one typically uses a related function, the so called Loschmidt echo $L(t)=|\nu_{\beta }(t)|^{2}$, which gives a measure of the environmental response to the perturbation induced by the system [@diciotto; @diciassette; @zana] and which, as shown in Ref. [@pinja], is linked with non-Markovianity of the open system dynamics. The amount of non-Markovianity of a dynamical map can be evaluated in different manners [@breuermeasure; @rivas; @fisherNM; @geo], which are, however, essentially equivalent for a purely dephasing quantum channel [@altricin; @pinja]. By adopting the definition in terms of information flow given in [@breuermeasure], one finds $$\mathcal{N}=\sum_{n} \left \{\left |\nu_{\beta}(t_{{\rm
max},n})\right |-\left | \nu_{\beta}(t_{{\rm min},n}) \right |
\right \}, \label{Ndef}$$ where the sum is performed over all maxima and minima of the $|\nu_{\beta}(t)|$, occurring at $t_{{\rm max},n}$ and $t_{{\rm
min},n}$, respectively. Using our previous results for the amplitude, we can then obtain the non-Markovianity of the dynamics of a two-level system in a trapped Fermi environment. The results are shown in Fig. \[nonmark\]**B**, where we see that $\mathcal{N}$ depends on the temperature and on the critical parameter $\alpha$. In particular, it has a maximum at small $\alpha$, increasing with low temperatures, and goes to zero both for large temperatures (due to the fact that thermal fluctuations suppress oscillations in the persistence amplitude) and for $\alpha >1$. In the latter case, excitations are generated at every energy scale in the fermion gas, as witnessed by the fact that the spectrum becomes structure-less. This implies that the gas becomes more and more stiff (in the sense that it is not able to react on the impurity any more) and explains why $\mathcal{N}$ is zero: the open system does not receive information back, its Loschmidt echo decays monotonously and thus the dynamics is Markovian. As a result, we conclude that a non-Markovian dynamics can be characterized, in our case, by the appearance of specific spectral features in the excitation energy distribution [@nori12].
We conclude with two remarks. First, the spectral distribution of energy excitations obtained in the present work coincides with the so called work distribution function, which is a central quantity in non-equilibrium processes [@campisi]. In the set-up that we have described above, it is simple enough to conceive a ‘reverse’ protocol, with the fermi gas brought to thermal equilibrium in the presence of the impurity (i.e. with the two-level atom in the excited state) which is then switched off. The comparison of the work distribution functions in the direct and reverse protocols would lead to a direct experimental test of the Crooks relation in the quantum regime [@crook]. The second remark is on the experimental realization of the model that we have described. Many experiments have recently dealt with the effects of impurities in trapped Fermi gas [@fermionexp], and state-dependent scattering lengths have been discussed [@theo]. This would lead to a direct test of our theory. Another viable candidate could be a gas of hard-core bosons in one-dimension, where the Loschmidt echo is equivalent to that of the corresponding Fermi gas [@johnpap] and in which impurities have recently been experimentally generated [@exp1d].
[99]{} P. W. Anderson, Phys. Rev. Lett. **18**, [1049](http://link.aps.org/doi/10.1103/PhysRevLett.18.1049) (1967).
P. W. Anderson and G. Yuval, Phys. Rev. Lett. **23**, [89](http://link.aps.org/doi/10.1103/PhysRevLett.23.89) (1969); P. W. Anderson, G. Yuval and R. D. Hamann, Phys. Rev. B **1**, [4464](http://link.aps.org/doi/10.1103/PhysRevB.1.4464) (1970).
G.D. Mahan, Phys. Rev. **163**, 612 (1967).
P. Nozières and C.T. De Dominicis, Phys.Rev. **178**, 1097 (1969); R. Roulet, J. Gavoret and P. Nozières, Phys.Rev. **178**, 1072 (1969); P. Nozieres, J. Gavoret and R. Roulet, Phys.Rev. **178**, 1084 (1969).
A. Sindona [*et al*]{}., Surf. Sci., [**601**]{}, 2805 (2007); A. Sindona, M. Pisarra, P. Riccardi, G. Falcone, Nanosci. Nanotechnol. Lett. [**4**]{}, 1050 (2012). A. Sindona [*et al*]{}., J. Phys.: Condens. Matter [**25**]{} 115301 (2013)
K. A. Matveev and A. I. Larkin, Phys. Rev. B [**46**]{}, 15337 (1992); A. K. Geim [*et al*]{}., Phys. Rev. Lett. [**72**]{}, 2061 (1994); H. Frahm, C. von Zobeltitz, N. Maire, R. J. Haug, Phys. Rev. B [**74**]{}, 035329 (2006); M. Heyl and S. Kehrein, Phys. Rev. B **85**, 155413 (2012).
N. Ubbelohde *et al.*, Sci. Rep. **2**:374, 1 (2012).
M. Hentschel and F. Guinea, Phys. Rev. B **76**, 115407 (2007).
A. Sindona, R.A. Baragiola, G. Falcone, A. Oliva, P. Riccardi, Phys. Rev. A [**71**]{}, 052903 (2005)
D. P. Clougherty and Y. Zhang, Phys. Rev. Lett **109**, 120401 (2012).
J. Goold, T. Fogarty, N. Lo Gullo, M. Paternostro, and T. Busch, Phys. Rev. A **84**, 063632 (2011).
M. Knap *et al*., Phys. Rev. X [**2**]{}, 041020 (2012).
A. Peres, Phys. Rev. A **30**, 1610 (1984); T. Gorin *et al*., Phys. Rep. **435**, 33 (2006).
F. M. Cucchietti, *et al*., Phys. Rev. Lett. **91**, 210403 (2003).
H. P. Breuer, E.-M. Laine, and J. Piilo, Phys. Rev. Lett. **103**, 210401 (2009).
A. Rivas, S. F. Huelga, and M. B. Plenio, Phys. Rev. Lett. **105**, 050403 (2010).
X. M. Lu, X. Wang, and C. P. Sun, Phys. Rev. A **82**, 042103 (2010).
S. Lorenzo, F. Plastina, M. Paternostro, Phys. Rev. A (R) to be published (2013), arXiv:1302.6673 (2013).
P. Haikka et al., Phys. Rev. A **85**, 060101(R) (2012).
T. J. G. Apollaro *et al*., Phys. Rev. A **83**, 032103 (2011); S. Lorenzo, F. Plastina, and M. Paternostro, Phys. Rev. A [**87**]{}, 022317 (2013).
P. Haikka *et al*. Phys. Rev. A **84**, 031602(R) (2012).
B-H. Lieu et al., Nat. Phys. **7**, 931 (2011).
A. Chiuri *et al*., Scientific Report [**2**]{}, 968 (2012).
A complementary, bosonization based approach to the response of a free electron gas is provided by K. D. Schotte and U. Schotte, Phys. Rev. [**182**]{}, 479 (1969).
P. Zanardi and N. Paunkovíc, Phys. Rev. E. **74**, 031123 (2006).
H.-S. Zeng *et al*., Phys. Rev. A **84**, 032118 (2011).
Similar conclusions have been reported in W.-M. Zhang *et al*., Phys. Rev. Lett. **109**, 170402 (2012).
M. Campisi, P. Hänggi, and P. Talkner, Rev. Mod. Phys. **83**, 771 (2011).
M. Heyl and S. Kehrein, Phys. Rev. Lett. **108**, 190601 (2012).
A. Schirotzek *et al*., Phys. Rev. Lett. **102**, 230402 (2009); C. Kohstall *et al*., Nature **485**, 615 (2012); M. Koschorreck *et al*., Nature **485**, 619 (2012).
K. M. Daily, D. Rakshit, and D. Blume, Phys. Rev. Lett. **109**, 030401 (2012).
K. Lelas, T. Seva, and H. Buljan, Phys. Rev. A **84**, 063601 (2011); K. Lelas, T. Seva, H. Buljan and J. Goold, Phys. Rev. A **86**, 033620 (2012).
S. Palzer, C. Zipkes, C. Sias, and M. Köhl, Phys. Rev. Lett. **103**, 150601 (2009); J. Catani [*et al*]{}., Phys. Rev. A **85**, 023623 (2012).
Impurity potential {#AppA}
==================
As explained in the main text, the non-interacting fermions in the harmonic trap lie in their equilibrium configuration, set by $\hat{H}_{0}$, until the impurity is excited and the sudden perturbation $\hat{V}(t)=\hat{V}\theta (t)$ is felt by the gas. To mimic a very strong difference in scattering length depending on the internal state ${\left|{e}\right\rangle}$ of the impurity, we have modelled it by a spatially localized potential, activated by the population of the excited state, with the structure-less form $V(x)=\pi
V_{0}x_{0}\delta (x)$.
-12pt
-16pt
For mathematical simplicity we have placed the impurity at the minimum of the harmonic potential. Thus, the coupling matrix elements between two unperturbed one-fermion states, $$\int{dx} \psi_{n}^{\ast}(x) V(x) \psi_{{n^{{\prime}}}}(x) = \pi V_{0} x_{0}
\psi_{n}^{\ast }(0) \psi _{{n^{{\prime}}}}(0)\text{,}$$ involve the Harmonic oscillator wave-functions at $x=0$. These have the usual expression $$\psi_{n}(x) = \frac{x_{0}^{-1/2} \pi^{-1/4}}{2^{n/2}
n!^{1/2}}{\,}H_{n}\!\!\left(\frac{x}{x_{0}}\right){\,}e^{-\frac{x^2}{2x_{0}^2}}\text{,}$$ in terms of the Hermite polynomials $H_{n}$ with $x_{0}{}$ being the characteristic oscillator length. By the parity of the Hermite polynomials, $H_{n}(-x/x_{0})=(-1)^{n}H_{n}(x/x_{0})$, we have $H_{n}(0)=0$ for odd-$n$, i.e., $n=2r+1$, with $r=0,1,\cdots ,\infty $. Therefore, the impurity potential induces excitations which connect only unperturbed one-fermion states labelled by even numbers $n=2r$, with $r=0,1,\cdots ,\infty $. The representation of the Hermite polynomials with even numbers in power series, $$H_{2r}\!\!\left(\frac{x}{x_{0}}\right)=\sum_{k=0}^{r}\frac{4^{k}(2r)!{\,}(-1)^{r-k}}{(2k)!{\,}(r-k)!}\left(\frac{x}{x_{0}}\right)^{2k}\text{,}$$ allow us to write $H_{2r}(0)=(-1)^{r}\frac{(2r)!}{r!}$, which leads to the matrix elements $$\begin{aligned}
V_{r{r^{{\prime}}}}=&\int{dx} \psi_{2r}^{\ast}(x) V(x) \psi_{2{r^{{\prime}}}}(x) \\
=&\sqrt{\pi}V_{0}{\,}\frac{(-1)^{r+{r^{{\prime}}}}}{2^{r+{r^{{\prime}}}}}{\,}\frac{(2r)!^{1/2}}{r!}\frac{(2{r^{{\prime}}})!^{1/2}}{{r^{{\prime}}}!},\end{aligned}$$ appearing in the second quantized representation of the impurity potential in the harmonic oscillator basis. Using the identity $$\gamma_r = \frac{\Gamma (r+1/2)}{\Gamma (r+1)}{}=2^{-2r} \pi^{1/2} \frac{(2r)!}{r!^2}\text{,}$$ where $\Gamma$ is the Euler gamma function , we get $V_{r{r^{{\prime}}}}=V_{0}(-1)^{r+{r^{{\prime}}}}{\,}{\gamma_{r}^{1/2}}{\,}{\gamma_{{r^{{\prime}}}}^{1/2}}$. The diagonal matrix elements $V_{rr}=V_{0}\gamma_{r}$, shown in Fig. \[figuno\]**A**, entirely determine the one-vertex graph of Eq. (\[Lambda1\]), representing first order corrections to the single-particle energies ${\varepsilon_{2r}}={\hbar\omega}(2r+1/2)$. On the other hand, both diagonal and off-diagonal matrix elements of $V(x)$, shown in Fig. \[figuno\]**B**, appear as absolute squares, $|V_{r{r^{{\prime}}}}|^{2}=V_{0}^{2}\,\gamma _{r}\gamma _{{r^{{\prime}}}}$, in the two-vertex graph given in Eq. (\[Lambda2\]).
Linked Cluster expansion of the vacuum persistence amplitude
============================================================
The vacuum persistence amplitude, introduced in Eqs. (\[Eq:VacAmp\]) and , may be expanded by the Dyson-Wick series $$\nu_{\beta }(t)=1+\sum_{m=1}^{\infty }\nu _{\beta }^{(m)}(t),$$ whose coefficients account for processes where the gas retrieves its equilibrium unperturbed configuration after $m=1,2,\cdots$ ‘scatterings’ with the impurity potential: $$\nu _{\beta }^{(m)}(t)=\frac{(-i)^m}{\hbar^{m}m!}
\int_{0}^{t}dt_{1} \cdots \int_{0}^{t}dt_{m} \big\langle
T\tilde{V}(t_{1})\cdots \tilde{V}(t_{m}) \big\rangle.$$ In the interaction picture, with $$\begin{aligned}
\tilde{V}(t)=\sum_{r,{r^{{\prime}}},\xi}V_{r{r^{{\prime}}}}\hat{c}_{2r\xi}^{{\dagger}}(t)\hat{c}_{2{r^{{\prime}}}\xi}(t),\end{aligned}$$ the time-evolution of creation and annihilation operators is that of the undisturbed Harmonic oscillator, i.e., $$\hat{c}^{{\dagger}}_{2r\xi}(t)=e^{i\omega \left(2r+\frac{1}{2}\right) t
}\hat{c}^{{\dagger}}_{2r\xi}, \quad \hat{c}_{2r\xi}(t)=e^{-i\omega
\left(2r+\frac{1}{2}\right) t}\hat{c}_{2r\xi}.$$ Accordingly, $$\begin{aligned}
& \big\langle T\tilde{V}(t_{1})\cdots \tilde{V}(t_{m})\big\rangle =\sum_{r_{1},r_{1}^{{\prime}},\xi _{1}}V_{r_{1}r_{1}^{{\prime}}}\cdots \sum_{r_{m},r_{m}^{{\prime}},\xi _{m}}V_{r_{m}r_{m}^{{\prime}}} \\
& \qquad \times \big\langle T\hat{c}_{2r_{1}\xi _{1}}^{{\dagger}}(t_{1})\hat{c}_{2r_{1}^{{\prime}}\xi _{1}}(t_{1})\cdots
\hat{c}_{2r_{m}\xi _{m}}^{{\dagger}}(t_{m})\hat{c}_{2r_{m}^{{\prime}}\xi
_{m}}(t_{m})\big\rangle.\end{aligned}$$ Here, the time-ordered average at the right-hand side may be decomposed using the Wick’s theorem into sums of products where each factor is a contracted pairs of creation/anihilation operators. The central approximation of the work is to retain terms that contain only equal time and two-time contractions, namely $$\big\langle {\cdots} { \contraction{} {\hat{c}} {_{2 r_{j}
\xi_{j}}^{{\dagger}}(t_{j})} {\hat{c}} } \hat{c}_{2 r_{j}
\xi_{j}}^{{\dagger}}(t_{j}) \hat{c}_{2r_{j}^{{\prime}}\xi_j}(t_{j})
{\cdots}\big\rangle={\cdots} f_{r_{j}}^{+} \delta
_{r_{j}\,r_{j}^{{\prime}}} {\cdots}$$ and $$\begin{aligned}
&\big\langle {\cdots} { \bcontraction{} {\hat{c}}
{_{2r_{j}\xi_j}^{{\dagger}}(t_{j})\hat{c}_{2r_{j}^{{\prime}}\xi_j}(t_{j}){\cdots}\hat{c}_{2r_{i}\xi}^{{\dagger}}(t_{i})}
{\hat{c}_{2}}
\acontraction{{\:\:}\hat{c}_{2r_{j}\xi}^{{\dagger}}(t_{j})} {\hat{c}}
{_{2r_{j}^{{\prime}}\xi_j}(t_{j}} {){\cdots}\hat{c}_{2r\xi}} }
\hat{c}_{2r_{j}\xi_j}^{{\dagger}}(t_{j})\hat{c}_{2r_{j}^{{\prime}}\xi_j}(t_{j})
{\cdots}
\hat{c}_{2r_{i}\xi_i}^{{\dagger}}(t_{i})\hat{c}_{2r_{i}^{{\prime}}\xi_i}(t_{i})
{\cdots}\big\rangle
\\
& \qquad =-\hbar ^{2} {\cdots} \delta_{\xi_i \xi_j}
G_{r_{j}}^{\beta}(t_{i}-t_{j}) \delta_{r_{j}r_{i}^{{\prime}}}
G_{r_{i}}^{\beta }(t_{j}-t_{i}) \delta_{r_{i}r_{j}^{{\prime}}} {\cdots},\end{aligned}$$ where the Fermion occupation numbers $f_{r}^{+}$ and the unperturbed Fermion propagators $G_{r}^{\beta }(t)$ have been introduced in the main text (see Eq. (\[prop\])). When these expressions are summed over even level numbers and integrated over time variables, we are left with products including either $$\Lambda _{1}^{\beta }(t)=-\frac{i{(2s+1)}t}{\hbar }
\sum_{r}V_{rr}f_{r}^{+}$$ or $$\begin{aligned}
& \Lambda _{2}^{\beta }(t)=-\frac{{(2s+1)}}{2}\sum_{{r^{{\prime}}},{r^{{\prime}{\hskip-0.5pt}{\prime}}}}|V_{{r^{{\prime}}}{r^{{\prime}{\hskip-0.5pt}{\prime}}}}|^{2} \notag \\
& \qquad \qquad \times \int_{0}^{t}d{t^{{\prime}}}\int_{0}^{t}d{t^{{\prime}{\hskip-0.5pt}{\prime}}}G_{{r^{{\prime}}}}^{\beta }({t^{{\prime}}}-{t^{{\prime}{\hskip-0.5pt}{\prime}}})G_{{r^{{\prime}{\hskip-0.5pt}{\prime}}}}^{\beta }({t^{{\prime}{\hskip-0.5pt}{\prime}}}-{t^{{\prime}}}),\end{aligned}$$ which are just the connected diagrams reported in Eqs. (\[Lambda1\]) and (\[Lambda2\]). Each product equals $\Lambda _{1}^{\beta }(t)^{j} \Lambda _{2}^{\beta }(t)^{m-j}$, obtained by $\binom{m}{j}$ distinct contractions for some $j$ between $0$ and $m$.
This means that $$\nu _{\beta }^{(m)}(t)\approx \frac{1}{m!} \sum_{j=0}^{m}
\binom{m}{j} \Lambda_{1}^{\beta }(t)^{j} \Lambda _{2}^{\beta
}(t)^{m-j}.$$ so that the Dyson-Wick series for the vacuum persistence amplitude takes the exponential form $$\nu _{\beta }(t)=\sum_{m=0}^{\infty }\frac{[\Lambda _{1}^{\beta
}(t)+\Lambda _{2}^{\beta }(t)]^{m}}{m!}=e^{\Lambda _{1}^{\beta
}(t)+\Lambda _{2}^{\beta }(t)}.$$ The single vertex graph (\[Lambda1\]) gives rise to the first order energy shift discussed in the main text, while the two-vertex connected graph has a more involved structure. Performing the time-ordered integrals in (\[Lambda2\]) we rewrite it as $$\Lambda_{2}^{\beta }(t)=-\frac{\alpha
{\varepsilon_{{\text{\textsc{f}}}}}}{\hbar}\sum_{r,{r^{{\prime}}}=0}^{\infty}
\left[it\varphi_{r{r^{{\prime}}}}+\frac{\psi _{r{r^{{\prime}}}}(t)}{2\omega }\right]
f^{+}_{r} f^{-}_{{r^{{\prime}}}}, \label{L2Start}$$ in which $$\varphi_{r{r^{{\prime}}}}=\frac{\gamma_r{\,}\gamma_{{r^{{\prime}}}}}{r-{r^{{\prime}}}}, \qquad
\psi_{r{r^{{\prime}}}}(t)=\varphi_{r{r^{{\prime}}}}{\,}\frac{1-e^{2 i (r-{r^{{\prime}}}) t
\omega}}{r-{r^{{\prime}}}}.
$$ Here, we may separate the sums over even-state labels ($r$,${r^{{\prime}}}$), so that the two-vertex connected graph can then be split into three contributions: $$\Lambda _{2}^{\beta}(t) = \Lambda_{2{\text{\textsc{s}}}}^{\beta }(t) +
\Lambda_{2{\text{\textsc{g}}}}^{\beta }(t) + \Lambda_{2{\text{\textsc{p}}}}^{\beta }(t),$$ where the subscripts stand for Shift, Gaussian and Periodic, respectively. In particular:
- the off-diagonal summands in (\[L2Start\]) that multiply $\varphi_{r{r^{{\prime}}}}$ give rise to $\Lambda_{2{\text{\textsc{s}}}}^{\beta}(t)=-itE_{2}^{\beta }/\hbar$, where $$E_{2}^{\beta } =\alpha {\varepsilon_{{\text{\textsc{f}}}}}\sum\limits_{r{\neq}{r^{{\prime}}}=0}^{\infty}
f^{+}_{r}\,\varphi_{r{r^{{\prime}}}}\,f^{-}_{{r^{{\prime}}}}.
$$ is the energy correction reported in Eq.
- the diagonal elements of Eq. (\[L2Start\]) yield the quadratic power law in $\Lambda_{2G}^{\beta}(t)$ \[see Eq. \];
- the remaining terms of the series in Eq. (\[L2Start\]) give the time periodic sub-diagram $$\Lambda _{2{\text{\textsc{p}}}}^{\beta }(t)=-\frac{\alpha
{\varepsilon_{{\text{\textsc{f}}}}}}{2{\hbar\omega}}\sum\limits_{r{\neq
}{r^{{\prime}}}}^{\infty}\,f_{r}^{+}\,\psi_{r{r^{{\prime}}}}\,f_{{r^{{\prime}}}}^{-},
$$ also reported in Eq. (\[Lambda2E\]).
Numerical computations {#AppC}
======================
As shown in the main text, the real and imaginary parts of the connected graphs $\Lambda _{1}^{\beta }(t)$ \[Eqs. , Fig. \[eradue\]**B**\] and $\Lambda _{2}^{\beta
}(t)$ \[Eq. , Fig. \[eradue\]**C**, Fig. 2\] combine in the vacuum persistence amplitude to give: $$\nu _{\beta }(t)= e^{-\frac{i t}{\hbar}(E_{1}^{\beta
}+E_{2}^{\beta })}\,
e^{-\alpha\,g_{\beta}\,\omega^{2}t^{2}}\,e^{\Lambda
_{2{\text{\textsc{p}}}}^{\beta }(t)}.$$ The knowledge of $\nu _{\beta }(t)$, allows us to determine the decoherence factor $|\nu _{\beta
}(t)|$ (Fig. \[cinque\] **A**, **C**, **E**), the shake up spectrum $|\tilde{\nu} _{\beta
}(E)|$ (Fig. \[cinque\]**B**, **D**, **F**, Fig \[nonmark\]**A**), and the non-Markovianity measure $\mathcal{N}$ \[Eq. , Fig \[nonmark\]**B**\]. Then, the basic quantities in our calculations are:
- the first and second order corrections, $E_{1}^{\beta}$ \[Eq. (\[eq:E1\]), Fig. \[eradue\]**B**\] and $E_{2}^{\beta}$ \[Eqs. (\[Delta2B\]), Fig. \[eradue\]**C**\], to the equilibrium energy $E_{0}^{\beta }$ (Fig. \[eradue\]**A**);
- the coefficient $g_{\beta }$ determining the standard deviation $\delta_{\beta}$ \[Eq. (\[eq:gbeta\]), Fig. \[eratre\]**A**\] of the Gaussian sub-diagram $\Lambda
_{2{\text{\textsc{g}}}}^{\beta }$;
- the shake-up sub-diagram $\Lambda_{2{\text{\textsc{p}}}}^{\beta}$ \[Eqs. (\[Lambda2E\]), Fig. \[eratre\]**B**\].
These contributions contain summations running over all one-fermion eigenstates of the trap weighted by the fermi factors $f_{r}^{+}$, $f_{{r^{{\prime}}}}^{-}$. These are expressed as $f^{+}_{r}=[1+e^{2 \beta{\hbar\omega}(r-r_{\mu})}]^{-1}$ and $f^{-}_{{r^{{\prime}}}}=[1+e^{-2 \beta{\hbar\omega}(r-r_{\mu})}]^{-1}$, using the parametrization $\mu ={\hbar\omega}(2r_{\mu }+1/2)$ for the chemical potential. The index $r_{\mu }$ tends to ${r_{{\text{\textsc{f}}}}}+ 1/4$ for $\beta
\to \infty$, so that the $\mu$ lies in the middle between the highest occupied (${\varepsilon_{{\text{\textsc{f}}}}}=\varepsilon_{2{r_{{\text{\textsc{f}}}}}}$) and the lowest unoccupied one-fermion levels ($\varepsilon_{2{r_{{\text{\textsc{f}}}}}+1}$) of the gas. For finite $\beta $ we determined $r_{\mu}$ by numerically constraining the conservation of particle number: $$2{r_{{\text{\textsc{f}}}}}+1=\sum_{r}\left(f_{r}^{+}+f_{r+1/2}^{+}\right),$$ As shown in Fig. \[figMmu\], $r_{\mu}$ and hence $\mu$, reach their maximum values at the absolute zero ($\beta \to \infty $). They decrease with decreasing $\beta$ and take largely negative values for $\beta \to 0$ where the classical limit applies. Interestingly enough, both $r_{\mu}$ and $\mu$ are almost independent on temperature for $\beta{\hbar\omega}\gtrsim 0.4$ and ${r_{{\text{\textsc{f}}}}}=5-500$.
With the computed $f^{\pm}_{r}$-distributions, we run numerical computations of the basic quantities (i)-(iii) using a high energy cut-off $\varepsilon_{{\text{\textsc{cut}}}}\gg {\varepsilon_{{\text{\textsc{f}}}}}$. To fix $\varepsilon_{{\text{\textsc{cut}}}}= {\hbar\omega}(2r_{{\text{\textsc{cut}}}}+1/2)$, we performed convergency tests by changing $r_{{\text{\textsc{cut}}}}$ in order to have a maximum instability error below $0.1\%$.\
In our applications to a spin $1/2$ gas, we observed that accurate estimations of $E_{2}^{\beta }$ and $g_{\beta }$ require values of $r_{{\text{\textsc{cut}}}}$ of the order of $10^{4}$ for ${r_{{\text{\textsc{f}}}}}$ below $\sim
200$ and $\beta{\hbar\omega}$ larger than $\sim 10^{-5}$. In particular, the plots of Fig. 1**C** and Fig. 2**A** of the main text were generated with a energy cut-off of $10-10^3~{\varepsilon_{{\text{\textsc{f}}}}}$. As for $E_{0}^{\beta }$, $E_{1}^{\beta }$, and $\Lambda_{2{\text{\textsc{p}}}}^{\beta
}$, we found out a cut-off of $\sim 10~{\varepsilon_{{\text{\textsc{f}}}}}$ to be sufficient in the investigated ranges of fermion numbers and temperatures. Accordingly, we included up to $10^{3}$ one-fermion states in Fig. 1**A**-**B** and Fig. 2**C** of the main text.\
To provide a more complete picture of the basic quantities involved in the numerical calculations, in Fig. \[FigE\]**A**-**C** we report the behavior of the equilibrium energy and the two energy shifts *vs* $\beta
{\hbar\omega}$ for gases made of different fermion numbers. We remark that both $E_{1}^{\beta }$ and $E_{2}^{\beta }$ are indeed small corrections to the unperturbed value $E_{0}^{\beta }$ of a spin-$1/2$ gas, for values of the critical exponent $\alpha
\lesssim 1$. In addition, $E_{1}^{\beta }$ is generally larger that $E_{2}^{\beta }$. These energies are strongly affected by the number of particles in the gas and weakly dependent on temperature for $\beta{\hbar\omega}\gtrsim 0.4$.\
On the other hand the Gaussian damping/broadening brought by $\Lambda_{2{\text{\textsc{g}}}}^{\beta }$, with standard deviation $\delta_{\beta}$, is almost entirely dependent on the thermal energy $\beta{\hbar\omega}$ and the critical exponent $\alpha$ (Fig. \[FigE\]**D**). This contribution leads to a smearing of the shake up response of the system in way that resembles the Anderson-Yuval approach to the Kondo problem . Indeed, $\delta_{\beta}$ decreases exponentially to zero with increasing $\beta{\hbar\omega}$, following the limiting trend $$\delta _{\beta }{\approx }2^{3/2}\alpha ^{1/2} e^{-\beta\omega
\hbar/4}$$ for $\beta{\hbar\omega}\gtrsim 7-8$ (Figs. 2**A** and \[FigE\]**B**).
-8pt
In addition, as evident by comparing Fig. 2**B** of the main text with the plots of Fig. \[FigGP\], the sub-diagram $\Lambda_{2{\text{\textsc{p}}}}^{\beta }$ has a time period of $\pi/\omega$. Its modulus $|\Lambda_{2{\text{\textsc{p}}}}^{\beta }|$ presents zeroes at $\omega t = m \pi$ and maxima at $\omega t = m \pi/2$, with $m=0,\pm 1,\pm 2,\cdots$. The intensities of such maxima (Fig. \[FigGP\]**A**-**C**) increase with increasing the Fermi number ($2{r_{{\text{\textsc{f}}}}}$), the critical exponent ($\alpha$), and the thermal energy ($\beta^{-1}$). The phase of $\Lambda_{2{\text{\textsc{p}}}}^{\beta }$ is discontinuous at the extremes of $|\Lambda_{2{\text{\textsc{p}}}}^{\beta }|$ and less dependent on these parameters (Fig. \[FigGP\]**D**-**F**).
-4pt
-4pt
Low thermal energy approximation
================================
We work at sufficiently low temperatures such that the chemical potential is well approximated by its maximum value $$\label{Mmu} \mu \underset{\beta{\hbar\omega}\rightarrow \infty }{\rightarrow
}{\hbar\omega}(2{r_{{\text{\textsc{f}}}}}+1),$$ corresponding to $$\label{Rmu} r_{\mu }\underset{\beta{\hbar\omega}\rightarrow \infty
}{\rightarrow }{r_{{\text{\textsc{f}}}}}+1/4,$$ as shown in Fig. \[figMmu\]. Then, we consider systems with a relatively large number of particles (${r_{{\text{\textsc{f}}}}}\gtrsim 10$) and focus on the sub-diagrams $\Lambda _{2{\text{\textsc{g}}}}^{\beta }(t)$ and $\Lambda _{2{\text{\textsc{p}}}}^{\beta }(t)$. In this, way we provide the formal justifications for the analytical approximations introduced in Eqs. (\[Lambda2GA\]) and (\[LambXX\]), which determine the ‘unshifted’ amplitude $\nu
_{\beta }^{{\prime}}(t)$ given in Eq. (\[Vpbet\]) and the excitation spectrum $\tilde{\nu}_{\beta }^{{\prime}}(E)$, shown in Fig \[nonmark\]**A**.
Using the power series $$f_{r}^{+}f_{r}^{-}\underset{{\varepsilon_{2r}}\lessgtr \mu
}{=}\sum\limits_{m=1}^{\infty }(-1)^{m+1}\,m\,e^{\pm \beta
m({\varepsilon_{2r}}-\mu )},$$ we rewrite the Gaussian damping parameter, reported in Eq. (\[eq:gbeta\]), as $$g_{\beta }=\sum_{m=1}^{\infty }(-1)^{m+1}\,m\,g_{m}^{\beta }.$$The coefficients of this expansion read$$\begin{aligned}
g_{m}^{\beta }& =\frac{{\varepsilon_{{\text{\textsc{f}}}}}}{{\hbar\omega}}\sum_{r<r_{\mu}}\gamma _{r}^{2}e^{\beta m({\varepsilon_{2r}}-\mu )}+\frac{{\varepsilon_{{\text{\textsc{f}}}}}}{{\hbar\omega}}\sum_{r>r_{\mu}}\gamma _{r}^{2}e^{-\beta m({\varepsilon_{2r}}-\mu )} \\
& =\frac{{\varepsilon_{{\text{\textsc{f}}}}}}{{\hbar\omega}}\sum_{r<r_{\mu}}\gamma _{r}^{2}e^{-2\omega \tau
_{m}(r_{\mu }-r)}+\frac{{\varepsilon_{{\text{\textsc{f}}}}}}{{\hbar\omega}}\sum_{r>r_{\mu}}\gamma
_{r}^{2}e^{-2\omega \tau _{m}(r-r_{\mu })},\end{aligned}$$ with $\tau _{m}$ being the characteristic times $\tau
_{m}=m\beta\hbar $, induced by thermal fluctuations. Then, we use Eq. (\[Rmu\]) and perform a change of summation indices to write $$\begin{aligned}
g_{m}^{\beta } &=&\frac{{\varepsilon_{{\text{\textsc{f}}}}}}{{\hbar\omega}}e^{-2\omega \tau _{m}(r_{\mu }-{r_{{\text{\textsc{f}}}}})}\sum_{r=0}^{{r_{{\text{\textsc{f}}}}}}\gamma _{{r_{{\text{\textsc{f}}}}}-r}^{2}e^{-2\omega \tau _{m}r} \\
&&+\frac{{\varepsilon_{{\text{\textsc{f}}}}}}{{\hbar\omega}}e^{2\omega \tau _{m}(r_{\mu
}-{r_{{\text{\textsc{f}}}}})}\sum_{r=1}^{\infty }\gamma _{{r_{{\text{\textsc{f}}}}}+r}^{2}e^{-2\omega \tau
_{m}r}.\end{aligned}$$The transformed summations in this last line are dominated by low $r$ terms. In a many fermion environment, we may use the asymptotic relation $\gamma _{{r_{{\text{\textsc{f}}}}}\pm r}^{2}\approx \gamma _{{r_{{\text{\textsc{f}}}}}}^{2}\approx {r_{{\text{\textsc{f}}}}}^{-1}$ and obtain $$g_{m}^{\beta }\underset{{r_{{\text{\textsc{f}}}}}\gg1}{\approx }2\frac{e^{2\omega \tau
_{m}(r_{\mu
}-{r_{{\text{\textsc{f}}}}})}+e^{2\omega \tau _{m}({r_{{\text{\textsc{f}}}}}-r_{\mu }+1)}-e^{-2\omega \tau _{m}r_{\mu }}}{e^{2\omega \tau _{m}}-1}.$$Then, using the asymptotic form and neglecting $e^{-2\omega \tau _{m}r_{\mu }}$, we find $$g_{m}^{\beta }\approx 2\frac{e^{\omega \tau _{m}/2}}{e^{\omega
\tau _{m}}-1},$$which leads to Eq. (\[Lambda2GA\]), i.e.,$$g_{\beta }\approx 2\sum_{m=1}^{\infty }(-1)^{m}m\frac{e^{\omega \tau _{m}/2}}{e^{\omega \tau _{m}}-1},$$and let us approximate the standard deviation with $$\delta _{\beta }\approx 2\alpha ^{1/2} \left[ \sum_{m=1}^{m_{{\text{\textsc{cut}}}}}(-1)^{m}m\frac{e^{\omega \tau _{m}/2}}{e^{\omega \tau
_{m}}-1}\right] ^{1/2}. \label{deltabAPP}$$Eq. (\[deltabAPP\]), plotted in Fig. \[LambdaPASX\]**A** for $m_{{\text{\textsc{cut}}}}=1,100$, is independent of the number of particles in the gas. We have verified that the truncated series for $m_{{\text{\textsc{cut}}}}=100$ works extremely well for ${r_{{\text{\textsc{f}}}}}=5-500$ and $\beta{\hbar\omega}\gtrsim 0.4$. The asymptotic form of $\delta _{\beta }$ leads to the result reported in Fig2**A** of the main text.
As for the Fermi-edge component $\Lambda
_{2{\text{\textsc{p}}}}^{\beta }(t)$, we consider the auxiliary functions $\lambda _{\pm }^{\beta }(t)$ introduced in the main text, which enter the connected graph $\Lambda _{2}^{\beta }(t)$ (see Eq. ). We replace the particle-hole distributions $f_{r}^{\pm }$ with the power series expansion$$\begin{aligned}
f_{r}^{\pm }=& \sum\limits_{m=0}^{\infty }(-1)^{m}e^{\pm \beta
m({\varepsilon_{2r}}-\mu )}
\qquad \quad {\varepsilon_{2r}}\lessgtr \mu \\
=& -\sum\limits_{m=1}^{\infty }(-1)^{m}e^{\mp \beta m({\varepsilon_{2r}}-\mu )}
\qquad {\varepsilon_{2r}}\gtrless \mu\end{aligned}$$ to write $$\lambda _{\pm }^{\beta }(t)=\sum_{m=0}^{\infty }(-1)^{m}\,\lambda
_{m\pm }^{\beta }(t). \label{lambdasexp}$$Here, the coefficients $\lambda _{m\pm }^{\beta }(t)$ may be computed exactly by the finite summations $$\sum_{r=r_{1}}^{r_{2}}\gamma
_{r}\,z^{r}=z^{r_{1}}\,{{
\setbox1=\hbox{\ensuremath{^{ }}}
\setbox2=\hbox{\ensuremath{_{2}}}
\setbox5=\hbox{\ensuremath{\tilde{F}}}
\hspace{\ifnum\wd1>\wd2\wd1\else\wd2\fi}
\ensuremath{\copy5^{\hspace{-\wd1}\hspace{-\wd5} \hspace{\wd5} }_{\hspace{-\wd2}\hspace{-\wd5}2\hspace{\wd5}1}}
}}(r_{1},z)-z^{r_{2}+1}\,{{
\setbox1=\hbox{\ensuremath{^{ }}}
\setbox2=\hbox{\ensuremath{_{2}}}
\setbox5=\hbox{\ensuremath{\tilde{F}}}
\hspace{\ifnum\wd1>\wd2\wd1\else\wd2\fi}
\ensuremath{\copy5^{\hspace{-\wd1}\hspace{-\wd5} \hspace{\wd5} }_{\hspace{-\wd2}\hspace{-\wd5}2\hspace{\wd5}1}}
}}(r_{2}+1,z),$$ holding for any $z\neq 1$, in which ${{
\setbox1=\hbox{\ensuremath{^{ }}}
\setbox2=\hbox{\ensuremath{_{2}}}
\setbox5=\hbox{\ensuremath{\tilde{F}}}
\hspace{\ifnum\wd1>\wd2\wd1\else\wd2\fi}
\ensuremath{\copy5^{\hspace{-\wd1}\hspace{-\wd5} \hspace{\wd5} }_{\hspace{-\wd2}\hspace{-\wd5}2\hspace{\wd5}1}}
}}$ is the regularized Hypergeometric function $${{
\setbox1=\hbox{\ensuremath{^{ }}}
\setbox2=\hbox{\ensuremath{_{2}}}
\setbox5=\hbox{\ensuremath{\tilde{F}}}
\hspace{\ifnum\wd1>\wd2\wd1\else\wd2\fi}
\ensuremath{\copy5^{\hspace{-\wd1}\hspace{-\wd5} \hspace{\wd5} }_{\hspace{-\wd2}\hspace{-\wd5}2\hspace{\wd5}1}}
}}({r_{{\text{\textsc{f}}}}},z)=\gamma _{{r_{{\text{\textsc{f}}}}}}\;{{
\setbox1=\hbox{\ensuremath{^{ }}}
\setbox2=\hbox{\ensuremath{_{2}}}
\setbox5=\hbox{\ensuremath{F}}
\hspace{\ifnum\wd1>\wd2\wd1\else\wd2\fi}
\ensuremath{\copy5^{\hspace{-\wd1}\hspace{-\wd5} \hspace{\wd5} }_{\hspace{-\wd2}\hspace{-\wd5}2\hspace{\wd5}1}}
}}(1,1/2+{r_{{\text{\textsc{f}}}}},1+{r_{{\text{\textsc{f}}}}};z)$$ Working in the temperature range where Eq. holds, the $m=0$ coefficients of the series (\[lambdasexp\]) turn out to be independent on $\beta $: $$\begin{aligned}
\lambda _{0+}^{\beta }(t) &=&\frac{\sqrt{\pi
}}{\sqrt{1-e^{2it\omega }}}
\label{lambdasexpM} \\
&&-e^{\frac{it{\varepsilon_{{\text{\textsc{f}}}}}}{\hbar }+\frac{3\omega
it}{2}}\,{{
\setbox1=\hbox{\ensuremath{^{ }}}
\setbox2=\hbox{\ensuremath{_{2}}}
\setbox5=\hbox{\ensuremath{\tilde{F}}}
\hspace{\ifnum\wd1>\wd2\wd1\else\wd2\fi}
\ensuremath{\copy5^{\hspace{-\wd1}\hspace{-\wd5} \hspace{\wd5} }_{\hspace{-\wd2}\hspace{-\wd5}2\hspace{\wd5}1}}
}}({r_{{\text{\textsc{f}}}}}+1,e^{2it\omega
}),\qquad \qquad \notag \\
\lambda _{0-}^{\beta }(t) &=&e^{-\frac{it{\varepsilon_{{\text{\textsc{f}}}}}}{\hbar }-\frac{3\omega it}{2}}\,{{
\setbox1=\hbox{\ensuremath{^{ }}}
\setbox2=\hbox{\ensuremath{_{2}}}
\setbox5=\hbox{\ensuremath{\tilde{F}}}
\hspace{\ifnum\wd1>\wd2\wd1\else\wd2\fi}
\ensuremath{\copy5^{\hspace{-\wd1}\hspace{-\wd5} \hspace{\wd5} }_{\hspace{-\wd2}\hspace{-\wd5}2\hspace{\wd5}1}}
}}({r_{{\text{\textsc{f}}}}}+1,e^{-2it\omega }). \label{lambdasexpP}\end{aligned}$$ The other coefficients, accounting for low temperature effects, have the form: $$\begin{aligned}
\lambda _{m\pm }^{\beta }(t) &=&\pm \frac{\sqrt{\pi }e^{-\beta m{\varepsilon_{{\text{\textsc{f}}}}}}}{\sqrt{1-e^{2\omega \left( \tau _{m}\pm it\right) }}} \\
&&\mp e^{\pm \frac{it{\varepsilon_{{\text{\textsc{f}}}}}}{\hbar }\pm \frac{3}{2}\omega \left(
it+\tau _{m}\right) }{{
\setbox1=\hbox{\ensuremath{^{ }}}
\setbox2=\hbox{\ensuremath{_{2}}}
\setbox5=\hbox{\ensuremath{\tilde{F}}}
\hspace{\ifnum\wd1>\wd2\wd1\else\wd2\fi}
\ensuremath{\copy5^{\hspace{-\wd1}\hspace{-\wd5} \hspace{\wd5} }_{\hspace{-\wd2}\hspace{-\wd5}2\hspace{\wd5}1}}
}}\lbrack {r_{{\text{\textsc{f}}}}}+1,e^{2\omega \left( \tau
_{m}\pm it\right) }]
\nonumber \\
&&\mp e^{\pm \frac{it{\varepsilon_{{\text{\textsc{f}}}}}}{\hbar }\pm \frac{3}{2}\omega \left(
it-\tau _{m}\right) }{{
\setbox1=\hbox{\ensuremath{^{ }}}
\setbox2=\hbox{\ensuremath{_{2}}}
\setbox5=\hbox{\ensuremath{\tilde{F}}}
\hspace{\ifnum\wd1>\wd2\wd1\else\wd2\fi}
\ensuremath{\copy5^{\hspace{-\wd1}\hspace{-\wd5} \hspace{\wd5} }_{\hspace{-\wd2}\hspace{-\wd5}2\hspace{\wd5}1}}
}}\lbrack {r_{{\text{\textsc{f}}}}}+1,e^{2\omega \left( \pm
it-\tau _{m}\right) }],
$$ With the zero temperature parts (\[lambdasexpP\]) and (\[lambdasexpM\]), we need to add an imaginary time regularization to the ${t^{{\prime}{\hskip-0.5pt}{\prime}}}$-integral in the two-loop term (\[Lambda2\]), i.e., we have to shift the ${t^{{\prime}{\hskip-0.5pt}{\prime}}}$ integration domain by $i\tau _{0}$ to prevent ${{
\setbox1=\hbox{\ensuremath{^{ }}}
\setbox2=\hbox{\ensuremath{_{2}}}
\setbox5=\hbox{\ensuremath{\tilde{F}}}
\hspace{\ifnum\wd1>\wd2\wd1\else\wd2\fi}
\ensuremath{\copy5^{\hspace{-\wd1}\hspace{-\wd5} \hspace{\wd5} }_{\hspace{-\wd2}\hspace{-\wd5}2\hspace{\wd5}1}}
}}({r_{{\text{\textsc{f}}}}}+1,e^{\pm 2\omega
ri{t^{{\prime}{\hskip-0.5pt}{\prime}}}})$ from being singular. We may, then, insert Eq. (\[lambdasexp\]) in the expression for $\Lambda _{2}^{\beta
}(t) $, compute the $r$-summations, and use the large ${r_{{\text{\textsc{f}}}}}$ expansion $${{
\setbox1=\hbox{\ensuremath{^{ }}}
\setbox2=\hbox{\ensuremath{_{2}}}
\setbox5=\hbox{\ensuremath{\tilde{F}}}
\hspace{\ifnum\wd1>\wd2\wd1\else\wd2\fi}
\ensuremath{\copy5^{\hspace{-\wd1}\hspace{-\wd5} \hspace{\wd5} }_{\hspace{-\wd2}\hspace{-\wd5}2\hspace{\wd5}1}}
}}({r_{{\text{\textsc{f}}}}}\gg 1,z)=\frac{{r_{{\text{\textsc{f}}}}}^{-1/2}}{1-z}+\mathrm{o}\left(
{r_{{\text{\textsc{f}}}}}^{-3/2}\right)\text{.}
$$ Finally, we may perform the ${t^{{\prime}}}$ and ${t^{{\prime}{\hskip-0.5pt}{\prime}}}$-integrals, excluding terms proportional to $t$ and $t^{2}$. What is left is a combination of logarithmic and polylogarithmic functions, dominated by the Fermi-edge term reported in Eq. (\[LambXX\]). Indeed, as shown in Fig. \[LambdaPASX\], the zero temperature form $$\Lambda _{2{\text{\textsc{p}}}}^{\infty }(t)\approx \ln \left[ \frac{e^{2\tau
_{0}\omega }-1}{e^{2\omega (\tau _{0}+it)}-1}\right] ^{\alpha }
\label{LambFES}$$ obtained from this procedure is in excellent agreement with the numerical calculations reported in Fig. \[eratre\]**B** for ${r_{{\text{\textsc{f}}}}}=100$ and $\beta{\hbar\omega}$ larger than $\sim 1$. The reliability of such an approximation is also attested by the comparison in Fig \[nonmark\]**A**. We will see in the following appendix that Eq. provides an accurate description of the singular response of a Fermi gas with low numbers of particles.
In the main text, we observed that $\Lambda _{2{\text{\textsc{p}}}}^{\infty
}(t)$ correctly tends to the MND form given by Eq. (\[MNDLamb\]) when the harmonic trap frequency is lowered to zero, keeping the number of particles in the gas finite. The singularity index is proportional to the height of the impurity potential barrier. Nozières and De Dominicis [@Nozieres] calculated the propagator for a free electron in the transient potential activated by a structure-less core-hole, by solving the associated Dyson equation in the long-time limit with Muskhelishvili techniques for singular integral equations. They assumed a constant potential of arbitrary height and width $\hbar/\tau_0$ around the Fermi level of the gas and found the singularity index to depend of the phase-shifts of this potential. Therefore they provided an asymptotic expression for all closed loops $\Lambda^{\infty
}(t)=\sum_n \Lambda_n^{\infty }(t)$, which may be written as Eq. . In our derivation, we have given a model to the impurity potential, being described by non constant matrix elements $V_{r{r^{{\prime}}}}$ (see appendix \[AppA\]). In addition we have found an analytical form for two vertex graph, Eq. (\[LambXX\]), which accurately describes $\Lambda _{2{\text{\textsc{p}}}}^{\beta}(t)$ at any time $t$ for a sufficiently wide range of temperatures and particle number. We expect the non trivial part of the higher order contributions $\Lambda _{n>2}^{\beta}(t)$ to change the value of the $\alpha$-parameter in such a way that it will depend on the phase shifts of the impurity potential.
-12pt
Decoherence factor, excitation spectrum and Non markovianity measure
====================================================================
With the arguments given in the previous section, the decoherence factor takes the analytical approximation $$\begin{aligned}
| \nu _{\beta }(t)| =&e^{-2 \alpha \omega^{2}t^{2}
\sum_{m=1}^{\infty }(-1)^{m} m e^{-\beta{\hbar\omega}m/2} }\\
&\times\prod_{m=-\infty }^{\infty}\left| \frac{e^{2\tau _{m}\omega
}-1}{e^{2(it+\tau _{m})\omega }-1}\right| ^{\alpha },\end{aligned}$$ and the excitation spectrum relative to the *perturbed* equilibrium energy of the gas may be written $$\tilde{\nu}^{\prime}_{\beta }(E)=\int_{0}^{\infty }\frac{dt}{2\pi
\hbar }e^{\frac{it}{\hbar }(E+E_1^{\beta}+E_2^{\beta})}\nu
^{\prime}_{\beta }(t).$$ We performed numerical calculations of both $| \nu_{\beta }(t) |$ and $\tilde{\nu}^{\prime}_{\beta }(E)$ by selecting different fermion numbers (${r_{{\text{\textsc{f}}}}}=5-100$), coupling parameters ($\alpha
=0.1-1.05$), and thermal energies ($\beta{\hbar\omega}=0.001-10$). Then, we evaluated the maxima/minima of $|\nu _{\beta }(t)|$ in a finite time-window with $t=0-100 \delta_{\beta}^{-1}$ and plugged the differences $|\nu _{\beta }(t_{{\rm max},n})|-|\nu _{\beta
}(t_{{\rm min},n})|$ into Eq. to estimate the non-Markovianity $\mathcal{N}$ of the two-level impurity in the gas.
-12pt
-8pt
In Fig. 3 and 4**A**, we have presented an application to a Fermi gas of $402$ particles (${r_{{\text{\textsc{f}}}}}=100$) where the Fermi-edge behavior, superimposed on a Gaussian damping trend, appears as a sequence of spikes in $|\nu _{\beta }(t)|=|\nu^{{\prime}}_{\beta }(t)|$ and an asymmetric peak structure in $\tilde{\nu}^{{\prime}}_{\beta
}(E)$. Such features becoming more and more marked with decreasing temperature, which reduces the effect of the Gaussian damping $\delta_{\beta }$ (see Fig. 2**A**), correspond to a sharp peak of $\mathcal{N}$ at $\alpha<0.2$ (Fig. 4**B**). Similar considerations hold for environments containing low fermion numbers (i.e, for ${r_{{\text{\textsc{f}}}}}=5$ in Fig. \[vtvbetrf5\] and Fig. \[nonmarkB\]**A**) where shake-up effects are even more visible because of the decreasing of $\delta_{\beta }$ with decreasing ${\varepsilon_{{\text{\textsc{f}}}}}$, leading to a sharp peaks in $\mathcal{N}$ at $\alpha<0.5$ (Fig. \[nonmarkB\]**B**). We finally notice that the analytical the approximation given by Eqs. (\[Lambda2GA\]) and (\[LambXX\]) works extremely well with environments containing both low (Fig. \[nonmark\]**A**) and large (Fig. \[nonmarkB\]**A**) fermion numbers.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Marilyn A. Walker [^1]'
title: |
Limited Attention and Discourse Structure\
[cmp-lg/9512003]{}
---
Hierarchical versus Linear Recency {#intro-sec}
==================================
In computational theories of discourse, there are at least three processes presumed to operate under a [limited attention constraint]{} of some type: (1) ellipsis interpretation; (2) pronominal anaphora interpretation; and (3) inference of discourse relations between representations A and B of utterances in a discourse, e.g. B [motivates]{} A. In each case, the interpretation of the current element B of a discourse depends on the accessibility of another earlier element A. According to the [limited attention constraint]{} only a limited number of candidates need to be considered in the processing of B, e.g. only a limited number of entities in the discourse model are potential cospecifiers for a pronoun.
The limited attention constraint has been defined by some researchers by [linear recency]{}: a representation of an utterance A is [linearly recent]{} for a representation of an utterance B if A is linearly adjacent to B. Using linear recency as a model of the limited attention constraint would mean that an antecedent for an anaphor is determined by a linear backward search of the text, or of a discourse model representation of the text [@CS79] [*inter alia*]{}.
In contrast, other work has formulated the limited attention constraint in terms of [hierarchical recency]{} [@GS86; @Hobbs85a; @MannThompson87] [*inter alia*]{}. A representation of an utterance A is [hierarchically recent]{} for a representation of an utterance B if A is adjacent to B in the tree structure of the discourse. Of all theories based on hierarchical recency, only Grosz and Sidner’s theory of discourse structure provides an operationalization of hierarchical recency in terms of their [stack]{} model of attentional state [@Sidner79; @Grosz77; @GS86]. Thus, below, the relationship between limited attention and hierarchical recency will be discussed in terms of their stack model, but the discussion should also apply to claims about the role of hierarchical recency in other work.
In the remainder of this squib, I will argue that the limited attention constraint must account for three types of evidence: (1) the occurrence of [informationally redundant utterances]{} in naturally occurring dialogues [@Walker93c]; (2) the infelicity of discourses that depend on accessing discourse entities that are not linearly recent; and (3) experiments that show that humans have limited attentional capacity [@Miller56; @Baddeley86].
Evidence for Limited Attention from Anaphoric Processing
========================================================
------------------------------------------ ------------------------------------------
Dialogue A Dialogue B
\(4) C: Ok Harry, I’m have a problem \(4) C: Ok Harry, I’m have a problem
that uh my - with today’s economy that uh my - with today’s economy
[*my daughter is working*]{}, [*my daughter is working*]{}
\(5) H: I missed your name. \(5) H: I missed your name.
\(6) C: Hank. \(6) C: Hank
(6.2) H: Is that H A N K?
(6.3) C: Yes.
\(7) H: Go ahead Hank \(7) H: Go ahead Hank
(8a) C: [*as well as her uh husband*]{}. (8a) C: [*as well as her uh husband*]{}.
(8b) They have a child. (8b) They have a child.
(8c) and they bring the child to us (8c) and they bring the child to us
every day for babysitting. every day for babysitting.
------------------------------------------ ------------------------------------------
In figure \[examp-ab-fig\], Dialogue A, hierarchical recency supports the interpretation of the proforms in utterance (8a) from a radio talk show for financial advice [@PHW82]. In utterance A-5, H interrupts C’s narrative to ask for his name, but in A-8, C continues as though A-4 had just been said. Utterance A-8a realizes the proposition [*My daughter’s husband is working as well*]{}, but this realization depends on both an anaphoric referent and an anaphoric property.
According to the stack model, since utterances A-5 $\ldots$ A-7 are part of an embedded segment, A-4 is [hierarchically recent]{} when A-8 is interpreted. A new focus space is pushed on the stack during the processing of dialogue A when the intention of utterance 5 is recognized. Since utterance 7 clearly indicates completion of the interrupting segment, the focus space for the interruption in 5 to 7 is popped from the stack after utterance 7, leaving the focus space for utterances 1 to 4 on the top of the stack. This focus space supports the interpretation of the proforms in A-8a.
However, consider the variation of dialogue A in dialogue B in figure \[examp-ab-fig\]. Here, the segment between B-5 $\ldots$ B-7 is also an embedded segment. Utterance B-7 indicates completion of the embedded segment and signals a pop. So, by the stack model, this segment is handled by the same focus stack popping mechanism as we saw for dialogue A.
However, in dialogue B, utterance 8a is more difficult, if not impossible, to interpret. This is surprising because utterance B-4 is hierarchically recent for B-8a, just as it is in dialogue A. The interruption in dialogue B is but a slightly longer version of that in dialogue A. Inasmuch as the stack model is a precise formulation of hierarchical recency, it does not predict the infelicity of dialogue B. The problem arises partly because the stack model includes no constraints related to the length, depth, or the amount of processing required for an embedded segment. Thus, these types of extended embedded segments suggest that the limited attention constraint must be sensitive to some aspect of linear recency.
Evidence for Limited Attention from Informational Redundancy
============================================================
Additional evidence for the influence of linear recency arises from the analysis of [informationally redundant]{} utterances (IRUs) in naturally-occurring discourse [@Walker93c].[^2] IRUs realize propositions already established as mutually believed in the discourse. IRUs have [antecedents]{} in the discourse, which are those utterances that originally established the propositions realized by the IRU as mutually believed. Consider excerpt C from the financial advice corpus. Here E has been telling H about how her money is invested, and then poses a question in C-3. IRUs in the examples below are capitalized and their antecedents are italicized.
The utterances in 22b and 22c realize propositions previously established as mutually believed, so they are IRUs.[^3] The cue word [*but*]{} in utterance 22a indicates a push, a new intention [@GS86]. The phrase [*as far as the certificates are concerned*]{} indicates that this new intention is subordinate to the previous discussion of the certificates. Thus, utterance 22a, [*but as far as the certificates are concerned*]{}, has the effect that the focus space related to the discussion of retirement investments, corresponding to utterances 8 to 21, is popped from the stack.
This means that the focus space representation of the intention for utterances 4 to 7 are on the top of the stack after C-22a, when 22b and 22c are processed. Therefore there are two reasons why it is surprising that H restates the content of utterances 4, 5 and 6 in 22b and 22c: (1) The propositions realized by 22b and 22c are already mutually believed; and (2) These mutual beliefs should be salient by virtue of being on top of the stack. If they are salient by virtue of being on top of the stack, they should be accessible for processes such as content-based inferences or the inference of discourse relations. If they must be accessible for these inferences to take place, as I will argue below, their reintroduction suggests that in fact they are not accessible. Many similar examples of IRUs are found in the corpus [@Walker93c]. Thus, these types of IRUs show that hierarchical recency, as realized by the stack model, does not predict when information is accessible.
The Cache Model of Attentional State {#cache-sec}
====================================
The evidence above suggests the need for a model of attentional state in discourse that reflects the limited attentional capacity of human processing. Here, I propose an alternate model to the stack model, which I will call the [cache model]{}, and discuss the evidence for this model. In section \[discuss-sec\], I compare a number of dimensions of the cache and stack models.
The notion of a cache in combination with main memory, as is standard in computational architectures, is a good basis for a computational model of human attentional capacity in processing discourse. All conversants in a dialogue have their own cache and some conversational processes are devoted to keeping these caches synchronized.
The cache model consists of: (1) basic mechanisms and architectural properties; (2) assumptions about processing; (3) specification of which mechanism is applied at which point. The [cache]{} represents working memory and [main memory]{} represents long-term memory. The cache is a limited capacity, almost instantaneously accessible, memory store. The exact specification of this capacity must be determined by future work, but previous research suggests a limit of 2 or 3 sentences, or approximately 7 propositions [@Kintsch88; @Miller56]. Main memory is larger than the cache, but is slower to access [@Baddeley86; @Kintsch88].
There are three operations involving the cache and main memory. Items in the cache can be preferentially [retained]{} and items in main memory can be [retrieved]{} to the cache. Items in the cache can also be [stored]{} to main memory.
When new items are retrieved from main memory to the cache, or enter the cache directly due to events in the world, other items may be [displaced]{}, because the cache has limited capacity. Displaced items are stored in main memory. The determination of which items to displace is handled by a cache replacement policy. The specification of the cache replacement policy is left open, however, replacing items that haven’t been recently used, with the exception of those items that are preferentially retained, is a good working assumption, as shown by previous work on linear recency.[^4]
The cache model includes specific assumptions about processing. Discourse processes execute on elements that are in the cache. All of the premises for an inference must be simultaneously in the cache for the inference to be made [@McKoonRatcliff92; @Walker93c]. If a discourse relation is to be inferred between two separate segments, a representation of both segments must be simultaneously in the cache [@FHM90; @Walker93c]. The cospecifier of an anaphor must be in the cache for automatic interpretation, or be strategically retrieved to the cache in order to interpret the anaphor [@GMR92]. Thus what is contained in the cache at any one time is a [working set]{} consisting of discourse entities such as entities, properties and relations that are currently being used for some process.
Two factors determine when cache operations are applied: (1) the speaker’s intentions and the hearer’s recognition of intention; (2) expectations about what will be discussed.
The cache model maintains the distinction between intentional structure and attentional state first proposed by Grosz and Sidner (1986). This distinction is critical. Just as a cache can be used for processing the references and operations of a hierarchically structured program, so can a cache be used to model attentional state when discourse intentions are hierarchically structured. The intentions of a conversant and the recognition of the other’s intentions determine what is [retrieved]{} from main memory and what is preferentially [retained]{} in the cache.
When conversants start working towards the achievement of a new intention, that intention may utilize information that was already in the cache. If so, that information will be preferentially retained in the cache because it is being used. Whenever the new intention requires information that is not currently in the cache, that information must be retrieved from main memory. Thus the process of initiating the achievement of the new intention has the result that some, and perhaps all, of the items currently in the cache are replaced with items having to do with the new intention.
When conversants return to a prior intention, information relevant to that intention must be retrieved from main memory if it has not been retained in the cache.
When an intention is completed, it is not necessary to strategically retain information relevant to the completed segment in the cache. It does not necessarily mean that there is an automatic retrieval of information related to other intentions. However, automatic retrieval processes can be triggered by associations between information being currently discussed and information stored in main memory [@GMR92]. These processes make items salient that have not been explicitly mentioned.
[expectations]{} about what will be discussed also determine operations on the cache. Expectations can arise from shared knowledge about the task, and from the prior discourse [@Grosz77; @Malt84]. Expectations can arise from interruptions when the nature of the interruption makes it obvious that there will be a return to the interrupted segment. When the pursuit of an intention is momentarily interrupted, as in dialogue A, the conversants attempt to retain the relevant material in the cache during the interruption.
Evaluating Critical Evidence: comparing the cache with the stack {#discuss-sec}
================================================================
In this section, I wish to examine evidence for the cache model, look at further predictions of the model,and then discuss evidence relevant to both stack and cache models in order to draw direct comparisons between them. First, I contrast the mechanisms of the models with respect to certain discourse processes.
- New intention subordinate to current intention: (1) Stack pushes new focus space; (2) Cache retrieves entities related to new intention
- Intention completed: (1) Stack pops focus space for intention from stack, entities in focus space are no longer accessible; (2) Cache doesn’t retain entities for completed intention, but they remain accessible until displaced
- New intention subordinate to prior intention: (1) Stack pops focus spaces for intervening segments, focus space for prior intention accessible after pop; (2) Cache retrieves entities related to prior intention from main memory to cache, unless retained in the cache
- Informationally redundant utterances: (1) Stack predicts no role for IRUs when they are represented in focus space on top of stack, because information should be immediately available; (2) Cache predicts that IRUs reinstantiate or refresh known information in the cache
- Returning from interruption: (1) In the stack model, the length and depth of the interruption and the processing required is irrelevant; (2) In the cache model, the length of the interruption or the processing required predicts retrievals from main memory
First, consider the differences in the treatment of interruptions. The state of the stack when returning from an interruption is identical for interruptions of various lengths and depths of embedding. In the cache model, an interruption may give rise to an expectation of a return to a prior intention, and each participant may attempt to retain information relevant to pursuing that intention in their cache. However, it may not be possible to retain the relevant material in the cache. In dialogue B, the interruption is too long and the working set for the interruption uses all of the cache. When this happens, the relevant material is displaced to main memory. On returning after an interruption, the conversants must initiate a cued retrieval of beliefs and intentions. This will require some processing effort, yielding the prediction that there will be a short period of time in which the cache does not have the necessary information. This would mean that the processing of incoming information would be slower until all of the required information is in the cache.[^5] The ease with which the conversants can return to a previous discussion will then rely on the retrievability of the required information from main memory, and this in turn depends on what is stored in main memory and the type of cue provided by the speaker as to what to retrieve. For example, if processing involves the surface form of the utterance, as it might dialogue B, we can explain the clear-cut infelicity by the fact that surface forms are not normally stored in main memory [@Sachs67].
Next, consider the differences between the models with respect to the function of IRUs. In dialogue C, a version of the dialogue without the IRUs is possible but is harder to interpret. Consider dialogue C without 22b, 22c and 23, i.e. replace 22a to 24 with [*But as far as the certificates are concerned, I don’t like all my eggs in one basket.*]{} Interpreting this alternate version requires the same inference, namely that having all your investments in six month certificates constitutes the negatively evaluated condition of having all your eggs in one basket. However the inference requires more effort to process.
The stack model doesn’t predict a function for the IRUs. However, according to the cache model, IRUs make information accessible that is not accessible by virtue of hierarchical recency, so that processes of content-based inferences, inference of discourse relations, and interpretation of anaphors can take place with less effort. Thus, one prediction of the cache model is that a natural way to make the anaphoric forms in dialogue B more easily interpretable is to re-realize the relevant proposition with an IRU, as in 8a’:[*My problem is that my daughter is working, as well as her uh husband.* ]{}
The IRU may function this way since: (1) the IRU reinstantiates the necessary information in the cache; or (2) the IRU is a retrieval cue for retrieval of information to the cache. Here reinstantiation is certainly sufficient, but in general these cases cannot be distinguished from corpus analysis. It should be possible to test psychologically using reaction time methods, whether and under what conditions IRUs function to simply reinstantiate an entity in the cache, and when they serve as retrieval cues.
Next, consider the differences in status of the entities in completed discourse segments. In the stack model, focus spaces for segments that have been closed are popped from the stack and entities in those focus spaces are not accessible. In the cache model, “popping” only occurs via displacement. Thus even when a segment is clearly closed, if a new topic has not been initiated, the popped entities should still be available. Some support for the cache model predictions about popped entities is that (1) rules proposed for deaccenting noun phrases treat popped entities as accessible [@DavisHirschberg88]; and (2) Rules for referring expressions in argumentative texts treat the conclusions of popped sisters as salient [@Huang94]. Stronger evidence would be the reaction times to the mention of entities in a closed segment, after it is clear that a new segment has been initiated, but before the topic of that new segment has initiated a retrieval to, and hence displacement from, the cache.
It should also be possible to test whether entities that are in the focus spaces on the stack, according to the stack model, are more accessible than entities that have been popped off the stack. In the cache model, the entities in these focus spaces would not have a privileged attentional status, unless of course they had been refreshed in the cache by being realized implicitly or explicitly in the intervening discussion.
Finally, consider one of the most studied predictions of the stack model: cases where a pronoun has an antecedent in a prior focus space. These cases have been called [return pops]{} or [focus pops]{} [@Grosz77; @Sidner79; @Reichman85; @Fox87; @PassonneauLitman94]. In the stack model, any of the focus spaces on the stack can be returned to, and the antecedent for a pronoun can be in any of these focus spaces. As a potential alternative to the stack model, the cache model appears to be unable to handle return pops since a previous state of the cache can’t be popped to. Since return pops are a primary motivation for the stack model, I will re-examine all of the naturally-occurring return pops that I was able to find in the literature. There are 21 of them. While it would be premature to draw final conclusions from such a small sample size, I will argue that the data supports the conclusion that return pops are [**cued retrieval from main memory**]{} and that the cues reflect the context of the pop [@RatcliffMcKoon88]. Thus, return pops are not problematic for the cache model.
In the cache model, there are at least three possibilities for how the context is created so that pronouns in [return pops]{} can be interpreted: (1) The pronoun alone functions as a retrieval cue [@GMR92]; or (2) The content of the first utterance in a return indicates what information to retrieve from main memory to the cache, which implies that the interpretation of the pronoun is delayed; (3) The shared knowledge of the conversants creates expectations that determines what is in the cache, e.g. shared knowledge of the task structure.
Let us consider the first possibility. The view that pronouns must be able to function as retrieval cues is contrary to the view that pronouns indicate entities that are currently salient [@Prince81]. However, there are certain cases where a pronoun alone is a good retrieval cue, such as when only one referent of a particular gender or number has been discussed in the conversation. If competing antecedents are those that match the gender and number of the pronoun [@Fox87], then only 11 of the 21 return pops found in the literature have competing antecedents.
Thus, the numbers suggest that in about half the cases we could expect the pronoun to function as an adequate retrieval cue based on gender and number alone. In fact, Sidner proposed that return pops might always have this property in her [stacked focus constraint]{}: “Since anaphors may co-specify the focus or a potential focus, an anaphor which is intended to co-specify a stacked focus must not be acceptable as co-specifying either the focus or potential focus. If, for example, the focus is a noun phrase which can be mentioned with an [*it*]{} anaphor, then [*it*]{} cannot be used to co-specify with a stacked focus.” [@Sidner79], p. 88,89.
In addition, the representation of the anaphor should include selectional restrictions from the verb’s subcategorization frame as retrieval cues [@Dieugenio90]. Of the 11 tokens with competing antecedents, 5 tokens have no competing antecedents if selectional restrictions are also applied. For example, in the dialogues about the construction of a pump from [@Grosz77], only some entities can be bolted, loosened, or made to work. Only 4 pronouns of the 21 return pops have competing referents if a selectional constraint can arise from the dialogue, e.g. if only one of the male discourse entities under discussion has been riding a bike, then the verb [*rode*]{} serves as a cue for retrieving that entity [@PassonneauLitman94]. Thus in 17 cases, an adequate retrieval cue is constructed from processing the pronoun and the matrix verb [@Dieugenio90].
The second hypothesis is that the content of the return utterance indicates what information to retrieve from main memory to the cache. The occurrence of IRUs as in dialogue C is one way of doing this. IRUs at the locus of a return can: (1) reinstantiate required information in the cache so that no retrieval is necessary; (2) function as excellent retrieval cues for information from main memory. An examination of the data shows that IRUs occur in 6 of the 21 return pops. IRUs in combination with selectional restrictions leave only 2 cases of pronouns in return pops with competing antecedents.
In the remaining 2 cases, the competing antecedent is not and was never prominent in the discourse, i.e. it was never the discourse center, suggesting that it may never compete with the other cospecifier.
It should be possible to test how long it takes to resolve anaphors in return pops and under what conditions it can be done, considering the data presented here on competing referents, IRUs, explicit closing, and selectional restrictions. A probe just after a pronoun in a return pop and before the verb could determine whether the pronoun alone is an adequate retrieval cue, or whether selectional information from the verb is required or simply speeds processing.
Finally, it should be possible to test whether pronouns in return pops are accented, which signals to the hearer that the most recent antecedent is not the correct one [@Cahn91].
To conclude, the analysis presented here suggests many hypotheses that could be empirically tested, which the currently available evidence does not enable us to resolve.
Discussion and Conclusion
=========================
This squib has discussed the role of limited attention in a computational model of discourse processing. The cache model was proposed as a computational implemention of human working memory and operations on attentional state are formulated as operations on a cache. Just as a cache can be used for processing the references and operations of a hierarchically structured program, so can a cache be used to model attentional state when discourse intentions are hierarchically structured.
The store and retrieve operations of the cache model casts discourse processing as a [**gradient**]{} phenomenon, predicting that the contents of the cache will change gradually, and that change requires processing effort. The notion of processing effort for retrieval operations on main memory makes predictions that can be experimentally tested. In the meantime, the notion of increased processing effort in the cache model explains the occurrence of a class of [informationally redundant]{} utterances in discourse, as well as cases of infelicitous discourses constructed as variations on naturally occurring ones, while remaining consistent with evidence on human limited attentional capacity. Finally, the cache model appears to handle the class of “return pops” which prima facie should be problematic for the model.
I’d like to thank Aravind Joshi, Ellen Prince, Mark Liberman, Karen Sparck Jones, Bonnie Webber, Scott Weinstein, Susan Brennan, Janet Cahn, Mitch Marcus, Cindie McLemore, Owen Rambow, Candy Sidner, Ellen Germain, Megan Moser, Becky Passonneau, Pam Jordan, Jennifer Arnold, and Steve Whittaker for extended discussion of the issues in this paper. Thanks also to the two anonymous reviewers.
Baddeley, Alan. 1986. . Oxford University Press.
Cahn, Janet. 1991. The effect of intonation on pronoun referent resolution. Technical report, MIT Media Lab.
Clark, Herbert H. and C.J. Sengul. 1979. In search of referents for nouns and pronouns. , 7:35–41.
Davis, James R. and Julia Hirschberg. 1988. Assigning intonational features in synthesized spoken directions. In [*ACL88*]{}.
D[i Eugenio]{}, Barbara. 1990. Centering theory and the italian pronominal system. In [*COLING 90*]{}.
Fletcher, Charles R., John E. Hummel, and Chad J. Marsolek. 1990. Causality and the allocation of attention during comprehension. .
Fox, Barbara A. 1987. . Cambridge University Press.
Greene, S.B., Gail McKoon, and R. Ratcliff. 1992. Pronoun resolution and discourse models. .
Grosz, Barbara J. 1977. The representation and use of focus in dialogue understanding. Technical Report 151, SRI International, 333 Ravenswood Ave, Menlo Park, Ca. 94025.
Grosz, Barbara J. and Candace L. Sidner. 1986. Attentions, intentions and the structure of discourse. , 12:175–204.
Hobbs, Jerry R. 1985. On the coherence and structure of discourse. Technical Report CSLI-85-37, Center for the Study of Language and Information, Ventura Hall, Stanford University, Stanford, CA 94305.
Huang, Xiorong. 1994. Planning references choices for argumentative texts. In [*The 7th International Conference on Natural Language Generation*]{}.
Kintsch, W. 1988. The role of knowledge in discourse comprehension: A construction-integration model. , 95:163–182.
Malt, Barbara. 1984. The role of discourse structure in understanding anaphora. .
Mann, W.C. and S.A. Thompson. 1987. Rhetorical structure theory: [D]{}escription and construction of text structures. In Gerard Kempen, editor, [*Natural Language Generation*]{}. Martinus Nijhoff, pages 83–96.
McKoon, Gail and Roger Ratcliff. 1992. Inference during reading. , 99(3):440–466.
Miller, G. A. 1956. The magical number seven, plus or minus two: Some limits on our capacity for processing information. , pages 81–97.
Passonneau, Rebecca J. and Diane Litman. 1994. Empirical analysis of three dimension of spoken discourse: Segmentation, coherence and linguistic devices. In Donia Scott and Eduard Hovy, editors, [*Burning Issues in Discourse*]{}.
Pollack, Martha, Julia Hirschberg, and Bonnie Webber. 1982. User participation in the reasoning process of expert systems. In [*AAAI82*]{}.
Prince, Ellen F. 1981. Toward a taxonomy of given-new information. In [*Radical Pragmatics*]{}. Academic Press, pages 223–255.
Ratcliff, Roger and Gail McKoon. 1988. A retrieval theory of priming in memory. , 95(3):385–408.
Reichman, Rachel. 1985. . Press, Cambridge, MA.
Sachs, Jacqueline D. 1967. . thesis, University of California Berkeley.
Sidner, Candace L. 1979. Toward a computational theory of definite anaphora comprehension in [English]{}. Technical Report AI-TR-537, MIT.
Walker, Marilyn A. 1992. Redundancy in collaborative dialogue. In [*Fourteenth International Conference on Computational Linguistics*]{}, pages 345–351.
Walker, Marilyn A. 1993. . thesis, University of Pennsylvania.
[^1]: ATT Laboratories, 600 Mountain Ave., Murray Hill, N.J. 07974, USA, [[email protected]]{}
[^2]: A subclass of Attention IRUs, Open-Segment IRUs, is discussed here.
[^3]: The antecedents are in utterances 4, 5 and 6: H asserted the content of 22b to E in 6. E indicated understanding and implicated acceptance of this assertion in 7 [@Walker92a], and E confirmed the truth of the content of 22c for H in 5.
[^4]: Obviously, linear recency is simply an approximation to what is in the cache. If something has been recently discussed, it was recently in the cache, and thus is is more likely to still be in the cache than other items. However, linear recency ignores the effects of retention and retrieval.
[^5]: This could predict the observed occurrence of disfluencies at segment boundaries [@PassonneauLitman94].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The properties of the ferromagnetic frustrated spin-S one-dimensional Heisenberg model in the vicinity of the transition point from the ferromagnetic to the singlet ground state is studied using the perturbation theory (PT) in small parameter characterizing the deviation from the transition point. The critical exponents defining the behavior of the ground state energy and spin correlation functions are determined using scaling estimates of infrared divergencies of the PT. It is shown that the quantum fluctuations for $s=1/2$ are sufficiently strong to change the classical critical exponents, while for spin systems with $s\geq 1$ the critical exponents remain classical. The dimerization in the singlet phase near the transition point is discussed.'
author:
- 'D.V.Dmitriev'
- 'V.Ya.Krivnov'
- 'J.Richter'
title: 'Zigzag spin-S chain near ferromagnet-antiferromagnet transition point'
---
Introduction
============
The quantum spin chains with nearest-neighbor (NN) $J_1$ and next-nearest-neighbor (NNN) $J_2$ interactions have been subject of numerous studies [@review]. The model with both antiferromagnetic interactions $J_1,J_2>0$ (AF-AF model) is well studied [@Haldane; @Tonegawa87; @Okamoto; @Bursill; @Majumdar; @White]. The case of F-AF interactions ($J_1<0,J_2>0$) is less studied. Though the latter model has been subject of many studies [@Tonegawa89; @Chubukov; @KO; @Aligia; @Vekua; @Lu], the complete picture of the phases of this model as a function of the frustration parameter $J_2/J_1$ is unclear up to now. An additional motivation to study this model is related to the fact that a class of recently synthesized compounds containing $CuO$ chains with edge-sharing $CuO_4$ units are described by the F-AF zigzag model [@Mizuno; @Drechsler1; @Drechsler2; @Hase]. The $Cu-O-Cu$ angle in these compounds is close to $90^{\circ}$ and usual antiferromagnetic NN exchange between $Cu$ ions is suppressed. This means that the sign of $J_1$ can be negative, while the NNN exchange is antiferromagnetic.
It is well known that there is a critical value $J_2/J_1=-1/4$, where the transition from the ferromagnetic ground state to the incommensurate singlet state occurs [@Schilling; @Hamada]. The study of the character of this quantum transition is one of the interesting problem related to the F-AF model. In this paper we focus on the behavior of the model in the vicinity of the transition point. We hope that this analysis will be useful for the study of the properties of the edge-shared copper oxides where the frustration ratio is close to the critical point. In particular, for edge-shared cuprate $Li_2ZrCuO_4$ the ratio is $J_2/J_1\sim -0.28$ and for $Rb_2Cu_2Mo_3O_{12}$ it is $J_2/J_1\sim -0.37$ [@helimagnetism1; @helimagnetism2].
The Hamiltonian of the F-AF model is $$H=-\sum_{n=1}^{N}(\mathbf{S}_{n}\cdot \mathbf{S}_{n+1}-s^{2})+J
\sum_{n=1}^{N}(\mathbf{S}_{n}\cdot \mathbf{S}_{n+2}-s^{2})
\label{H}$$ where we put $J_1=-1$ and $J_2=J$, $s$ is a spin value and periodic boundary conditions are imposed. The constant shifts in Eq.(\[H\]) secure the energy of the ferromagnetic state to be zero.
Unfortunately, this model is not solved exactly. As was noted above, the ground state of the model (\[H\]) is ferromagnetic at $0<J<1/4$ and it becomes a singlet incommensurate state for $J>1/4$. Though the transition point between these phases is $J=1/4$ for any $s$ [@Schilling], the spectra of the model with $s=1/2$ and $s\geq 1$ in this point are different. For $s=1/2$ the singlet ground state wave function in the transition point is known exactly [@Hamada; @DKO97]. It is degenerate with the ferromagnetic state for any even $N$. For $s\geq 1$ the singlet ground state wave function is unknown. Finite-size calculations shows that at $J=1/4$ the singlet state lies slightly higher than the ferromagnetic level and the energies of the singlet and the ferromagnetic states are equal in the limit $N\to
\infty $ only.
In the vicinity of the transition point at $0<\gamma \ll 1$ $(\gamma =J-1/4) $ the singlet ground state energy $E_0$ behaves as $E_0\sim \gamma ^{\beta }$, where $\beta $ is a critical exponent. The classical approximation gives $\beta =2$. The spin-wave theory as well as some other approximations [@Bursill; @KO] do not change this critical exponent. Unfortunately, the exact diagonalization of finite chains shows a complicated irregular size dependence of the ground state energy, which makes the numerical estimation of the critical exponent $\beta $ impossible. In the paper [@DK06] we conjectured that for $s=1/2$ strong quantum fluctuations changes the critical exponent and $\beta =5/3$.
We note that the model (\[H\]) with $s\geq 1$ has not been studied before and the critical exponent for these cases is unknown. In this paper we confirm our conjecture for $s=1/2$ using scaling estimates of the perturbation theory (PT) in small parameter $\gamma $. We show also that $\beta =2$ for $s\geq 1$, though the corresponding factor at $\gamma^2$ is different from the classical value and it depends on $s$.
One of the most important and open questions in the zigzag model (\[H\]) is the possibility of the spontaneous dimerization of the system in the singlet phase accompanying by a gap in the spectrum. This problem has been mostly studied in the limit of two weakly coupled AF $s=1/2$ chains ($J\gg 1 $). The one-loop renormalization group analysis indicates [@Nersesyan; @Cabra] that the gap is open. However, the existence of the gap has not been verified numerically [@Cabra]. On the basis of a field theory consideration it has been proposed [@Itoi] that a finite gap exists, but it is so tiny that it can not be observed numerically. On the opposite side of the singlet phase, $J\to
1/4$, there are no any reliable results about the dimerization and the gap. Strong nonmonotonic finite-size effects do not allow to study the dimerization numerically.
In order to study the problem of the spontaneous dimerization in the singlet phase of the model (\[H\]) close to the transition point $J=1/4$, we consider the generalization of the model (\[H\]) by adding to the Hamiltonian $H$ the perturbation in a form of dimerization term. Unfortunately, the used special version of the PT did not give us a rigorous answer about the spontaneous dimerization in the model (\[H\]). However, it allowed us to estimate the critical exponent of the dimer order, in case if the spontaneous dimerization in the model (\[H\]) exists. Besides, it allowed us to obtain the critical exponents of the ground state energy and the dimer order for the dimerized version of the model (\[H\]).
The paper is organized as follows. In Sec.II we present the scaling estimate of the critical exponent $\beta $ using the PT in $\gamma $ for the Hamiltonian (\[H\]) starting from the singlet ground state at $J_{c}=1/4$. In Sec.III we analyze the PT for the Hamiltonian which is transformed to new local axes forming a spiral structure. We establish the scaling behavior associated with $\gamma $ and with the pitch angle of the spiral. It is shown that the critical exponents for the ground state energy are different for the spins $s=1/2$ and $s\geq 1$. In Sec.IV we study the problem of the spontaneous dimerization in the model (\[H\]). In Sec.V we summarize our results.
Scaling estimate of the critical exponent near the transition point $J=1/4$
===========================================================================
We are interested in the behavior of the model (\[H\]) in the vicinity of the transition point $J_{c}=1/4$. For this aim it is natural to develop the perturbation theory $$\begin{aligned}
H &=&H_{0}+V_{\gamma } \nonumber \\
H_{0} &=&-\sum (\mathbf{S}_{n}\cdot
\mathbf{S}_{n+1}-s^{2})+\frac{1}{4}\sum (
\mathbf{S}_{n}\cdot \mathbf{S}_{n+2}-s^{2}) \nonumber \\
V_{\gamma } &=&\gamma \sum (\mathbf{S}_{n}\cdot \mathbf{S}_{n+2}-s^{2})
\label{H0}\end{aligned}$$ with a small parameter $\gamma =J-1/4\ll 1$ ($\gamma >0$).
At $\gamma >0$ the ground state of the Hamiltonian $H$ is a singlet. Since the perturbation $V_{\gamma }$ conserves the total spin $S^{2}$, the PT to the lowest singlet state $\left| \Psi
_{0}\right\rangle $ of the Hamiltonian $H_{0}$ involves only singlet excited states. The low-lying singlet excitations at the transition point have very small energies as shown in Figs.(1) and (2), where we present finite-size calculations of the energy gap between the two lowest singlets. These calculations show that the low-lying singlet excitations have different powers in $N$ for the cases $s=1/2$ and $s\geq 1$. As it will be shown below this fact leads to different critical exponents for spin systems with $s=1/2$ and $s\geq 1$.
The perturbation series for the singlet ground state energy can be written in a form: $$E_{0}(\gamma )=\left\langle \Psi _{0}\right| V_{\gamma }+V_{\gamma
}\frac{1}{ E_{0}-H_{0}}V_{\gamma }+\ldots \left| \Psi
_{0}\right\rangle \label{Eseries}$$
Suppose that the low-lying excitations acting in the PT behave as $$E_{k}-E_{0}\sim N^{-\delta } \label{dEexp}$$
The higher orders of perturbation series contain more dangerous denominators, and, therefore, have higher powers of the infrared divergency. Therefore, we use scaling arguments to estimate the critical exponent for the ground-state energy. Below we will follow only powers of divergencies and omit all numerical factors.
Suppose that the matrix elements of the perturbation operator $V_{\gamma }$ between low-lying states involved into the PT at $N\to \infty $ behave as $$\left\langle \Psi _{i}\right| V_{\gamma }\left| \Psi _{j}\right\rangle \sim
\gamma N^{1-d} \label{Vgexp}$$ with some exponent $d$.
Looking after powers of infrared divergencies in all orders of the perturbation series the correction to the ground state energy takes a form: $$E_{0}(\gamma )\sim \left\langle \Psi _{0}\right| V_{\gamma }\left| \Psi
_{0}\right\rangle \sum_{m=0}^{\infty }c_{m}x^{m}\sim \gamma N^{1-d}\cdot f(x)
\label{Eseries2}$$ where $c_{m}$ are unknown constants and $$x\sim \frac{\left\langle \Psi _{i}\right| V_{\gamma }\left| \Psi
_{k}\right\rangle }{E_{k}-E_{0}}\sim \gamma N^{\delta +1-d} \label{x}$$ is a scaling parameter, which absorbs the infrared divergencies.
The scaling function $f(x)$ at $x\to 0$ is given by the first order correction. In the thermodynamic limit ($x\to \infty $) the behavior of $f(x)$ is generally unknown, but the natural condition $E_{0}(\gamma )\sim N$ at $ N\to \infty $ requires $$f(x)\sim x^{\frac{d}{\delta +1-d}} \label{f(x)}$$ and, finally $$E_{0}(\gamma )\sim -N\gamma ^{\frac{\delta +1}{\delta +1-d}} \label{Eexp}$$
The perturbation series for the lowest excited state $E_{1}(\gamma
)$ has the same form as Eq.(\[Eseries\]). But a requirement of a finite mass gap (if any) $m=E_{1}(\gamma )-E_{0}(\gamma )\sim
O(1)$ leads to another critical exponent $$m\sim \gamma ^{\frac{\delta }{\delta +1-d}} \label{mexp}$$
We note, that for the models in fixed points with a linear spectrum ($\delta =1$), Eqs.(\[x\])-(\[mexp\]) reduce to the well-known formulae [@Cardy] $$\begin{aligned}
x &=&\gamma N^{2-d} \nonumber \\
E(\gamma ) &\sim &-N\gamma ^{\frac{2}{2-d}} \nonumber \\
m &\sim &\gamma ^{\frac{1}{2-d}} \label{CFT}\end{aligned}$$ where the exponent $d$ represents the scaling dimension of the perturbation operator.
However, the transition point $J=1/4$ is not a fixed point. Finite size calculations for the gap between lowest singlet states give exponents $\delta =4$ for $s=1/2$ chain and $\delta =3$ for $s\geq
1$ chain (see Figs.(1) and (2)).
In order to determine the value of the exponent $d$, we notice that the singlet ground state of $H_{0}$ has a spiral ordering at $N\to \infty $ with a period of the spiral equal to $N$ $$\left\langle \Psi _{0}\right| \mathbf{S}_{n}\cdot \mathbf{S}_{n+l}\left|
\Psi _{0}\right\rangle =s^{2}\cos \frac{2\pi l}{N} \label{spiral}$$
For the case $s=1/2$ this expression is an exact one [@Hamada],[@DKO97], while for $s\geq 1$ we have observed the spin spiral structure in exact diagonalization of finite-size systems (the spin correlation function for spin $s=1$ chain of size $N=20$ is shown in Fig.3). This means that the first order correction to the ground state energy in $\gamma $ is $$\left\langle \Psi _{0}\right| V_{\gamma }\left| \Psi _{0}\right\rangle
=-\gamma \frac{\left( 4\pi s\right) ^{2}}{N} \label{E1singlet}$$
We assume that all matrix elements between any low-lying singlet states $\left| \Psi _{i}\right\rangle $ and $\left| \Psi
_{j}\right\rangle $ have the same $N$-dependence $$\left\langle \Psi _{i}\right| V_{\gamma }\left| \Psi _{j}\right\rangle \sim
\frac{\gamma }{N} \label{Vgsinglet}$$ and, therefore, the exponent $d=2$.
Thus, as follows from Eq.(\[Eexp\]) the critical exponents for the ground state energy are different for $s=1/2$ and $s\geq 1$ $$\begin{aligned}
E_{0}(\gamma ) &\sim &-N\gamma ^{5/3},\qquad s=1/2 \nonumber \\
E_{0}(\gamma ) &\sim &-N\gamma ^{2},\qquad s\geq 1 \label{Egexp}\end{aligned}$$
For the case $s\geq 1$ the above scaling estimates reproduce the classical value for the critical exponent of the ground state energy. But in the special case $s=1/2$ the quantum fluctuations are strong enough to change the critical exponent. In order to understand the nature of the difference between $s=1/2$ and $s\geq
1$ systems and determine factors in Eqs.(\[Egexp\]) in the next Section we develop a special version of the PT.
Perturbation theory for the transformed Hamiltonian
===================================================
Let us start from the classical picture of the ground state of the model (\[H\]). In the classical approximation the spins are vectors which form the spiral structure with a pitch angle $\varphi $ between neighboring spins. The classical energy per site $$E_{\mathrm{cl}}(\varphi )=Ns^{2}\left[ 1-\cos \varphi -J(1-\cos
(2\varphi ) \right] \label{Ecl}$$ is minimized by the angle $$\varphi _{\mathrm{cl}}=\cos ^{-1}\frac{1}{4J} \label{phicl}$$
The classical ground state energy is $$E_{\mathrm{cl}}(\varphi _{\mathrm{cl}})=-N\frac{2s^{2}}{J}\gamma ^{2}
\label{Eclphicl}$$
Following this picture we transform local axes on $n$-th site by a rotation about the $Y$ axis by $\varphi n$. This rotation transforms the original spin wave functions $\left| \psi
_{n}\right\rangle $ to a new basis depending on the angle $\varphi
$ $$\left| \psi _{n,\varphi }\right\rangle =U_{\varphi }\left| \psi
_{n}\right\rangle \label{rotate}$$ where $$U_{\varphi }=\exp \left( i\varphi \sum_{n=1}^{N}nS_{n}^{y}\right)
\label{Uphi}$$ is the rotation operator and $U_{\varphi }^{\dagger
}=U_{-\varphi}$. For finite cyclic systems the pitch angle $\varphi $ takes quantized values $\varphi _{m}=\frac{2\pi m}{N}$.
Under the unitary transformation $U_{\varphi }$ the Hamiltonian $H$ takes a form $$\begin{aligned}
H_{\varphi } &=&U_{\varphi }HU_{-\varphi }=H+V_{\varphi } \nonumber \\
V_{\varphi } &=&(1-\cos \varphi )\sum \left[
S_{n}^{x}S_{n+1}^{x}+S_{n}^{z}S_{n+1}^{z}\right] \nonumber \\
&&-J(1-\cos 2\varphi )\sum \left[
S_{n}^{x}S_{n+2}^{x}+S_{n}^{z}S_{n+2}^{z}
\right] \nonumber \\
&&-\sum [\sin \varphi (S_{n}^{x}S_{n+1}^{z}-S_{n}^{z}S_{n+1}^{x})-J\sin
2\varphi (S_{n}^{x}S_{n+2}^{z}-S_{n}^{z}S_{n+2}^{x})] \label{Vphi}\end{aligned}$$
Now let us choose some eigen state $\left| \psi _{n}\right\rangle$ of the Hamiltonian $H$ $$H\left| \psi _{n}\right\rangle =E_{n}\left| \psi _{n}\right\rangle
\label{Heigen}$$
The state $\left| \psi _{n}\right\rangle $ is the eigen state of the Hamiltonian $H$, but not of $H_{\varphi }$. Therefore, if we develop and exactly calculate the perturbation theory in $V_{\varphi }$ to this state we arrive to some eigen state $\left|
\psi _{m,\varphi }\right\rangle $ of the Hamiltonian $H_{\varphi
}$ $$H_{\varphi }\left| \psi _{m,\varphi }\right\rangle =E_{m}(\varphi )\left|
\psi _{m,\varphi }\right\rangle \label{Hphieigen}$$ corresponding to, generally speaking, another energy level $E_{m}(\varphi )\neq E_{n}$. Obviously, the unitary transformation $U_{\varphi }$ does not change the spectrum. Therefore, the found energy level $E_{m}(\varphi )$ is also one of the eigen values of the original Hamiltonian $H$. Thus, taking different values of the pitch angle $\varphi _{m}=\frac{2\pi m}{N}$ ($m=1\ldots N$) and developing the PT in $V_{\varphi }$ to some definite eigen state $\left| \psi _{n}\right\rangle $ of the Hamiltonian $H$ we obtain a set of $N$, generally different, levels $E_{m}(\varphi )$ of $H$. So, we do not need to fix the value of $\varphi $ to its classical value in contrast to the spin-wave approximation. Instead, we are free to pick out the minimal energy from the set of the found $N$ levels $E_{m}(\varphi )$. In the thermodynamic limit, when $\varphi $ becomes continuous variable, this procedure means the minimization of the found energy $E(\varphi )$ over $\varphi $.
As a ‘source’ function $\left| \psi _{n}\right\rangle $ of $H$ it is natural to choose the ferromagnetic state with all spins pointing up $$\left| F\right\rangle =\left| \uparrow \uparrow \ldots \uparrow
\right\rangle \label{Fstate}$$
This choice is equivalent to taking the function $\left|
F_{\varphi }\right\rangle =U_{\varphi }\left| F\right\rangle $ as a probe ground state for the model (\[H\]). The function $\left|
F_{\varphi }\right\rangle $ has a spiral structure arising in the classical approximation. The expectation value of the total $S^{2}$ in this state is [@KO] $$\left\langle F_{\varphi }\right| S^{2}\left| F_{\varphi
}\right\rangle = \frac{N}{2}$$
This means that $\left| F_{\varphi }\right\rangle $ is not a pure singlet state, but contains an admixture of states with $S\neq 0$. However, it is clear that the weights of states with $S\neq 0$ are negligible at $N\to \infty $ and we can treat the state $\left|
F_{\varphi }\right\rangle $ as a singlet one.
Since we are interested in the behavior of the model near the transition point $J=1/4$, it is convenient to represent the Hamiltonian $H_{\varphi }$ in the form $$H_{\varphi }=H_{0}+V_{\gamma }+V_{\varphi } \label{Hphi}$$ with $H_{0}$ and $V_{\gamma }$ defined above in Eq.(\[H0\]) and to develop the perturbation theory to the ferromagnetic state in $V=V_{\varphi }+V_{\gamma }$. So, there are two channels $V_{\varphi }$ and $V_{\gamma }$ in the perturbation theory characterized by two small parameters $\varphi $ and $\gamma $. The ferromagnetic state $\left| F\right\rangle $ is the eigen state of the Hamiltonian $H_{0}$ with the energy $E_{0}=0$ and also of the perturbation $V_{\gamma }$, but not of $V_{\varphi }$. The obvious relation $V_{\gamma }\left| F\right\rangle =0$ means that the perturbation series for the energy contains terms $\sim
\varphi ^{m}\gamma ^{n}$, but does not contain terms $\sim \gamma
^{n}$ without $\varphi $.
At first sight it seems that as a result of the rotation Eq.(\[Uphi\]) we obtain more complex Hamiltonian $H_{\varphi }$ and more complicated perturbation theory with two channels. But the advantage of this method is to construct the perturbation theory in $V=V_{\varphi }+V_{\gamma }$ to the simple ferromagnetic state instead of the perturbation theory in $V_{\gamma } $ to very complicated (and even unknown for $s\geq 1$) lowest singlet state of $H_{0}$, which was analyzed using scaling arguments and numerical calculations in the previous section. The fact that we separate the term $V_{\gamma }$ from $H$ and treat it as the perturbation does not change our arguments about minimization of the found expression for energy $E(\varphi ,\gamma )$ over $\varphi $.
The ground state of the Hamiltonian $H_{0}$ is manifold degenerate: all the ferromagnetic states $\left|
F_{S_{z}}\right\rangle $ with different total $S_{z}=\sum
S_{n}^{z}$ have zero energy. Therefore, at first we have to split this degeneration of the ground state with use of secular equation. It turns out that diagonal elements are proportional to $\left\langle F_{S_{z}}\right| V\left| F_{S_{z}}\right\rangle \sim
N$, while non-diagonal matrix elements are $\left\langle
F_{S_{z}}\right| V\left| F_{S_{z}^{\prime }}\right\rangle \sim
O(1)$. Therefore, in the thermodynamic limit we can neglect non-diagonal matrix elements and develop regular perturbation theory directly to the ferromagnetic state $\left| F\right\rangle
$ with all spins pointing up.
The first-order correction to the energy reproduces the leading terms of the classical result (\[Eclphicl\]): $$E^{(1)}=\left\langle F\right| V_{\varphi }\left| F\right\rangle
=-2Ns^{2}\gamma \varphi ^{2}+Ns^{2}\frac{\varphi ^{4}}{8} \label{E1}$$
The second-order correction to the energy $$E^{(2)}=\sum_{k}^{\prime }\frac{\left\langle \Psi _{k}\right| V\left|
F\right\rangle ^{2}}{E_{0}-E_{k}} \label{E2formula}$$ relates to a two-magnon states, because operator $V$ (actually $V_{\varphi }$) have non-zero matrix elements in (\[E2formula\]) only with the states $\left| \Psi _{k}\right\rangle $ containing two magnons with total quasi-momentum $Q=0$ and relative quasi-momentum $k$. Exact calculation of the two-magnon problem gives for the sum the following result $$E^{(2)}=-N\frac{3s^{2}\varphi ^{4}}{16(s+1)} \label{E2}$$
This sum converges, because a dangerous denominator $$\varepsilon _{2}(k)=E_{k}-E_{0}=\frac{s}{2}k^{4}$$ for small $k$ is compensated by the matrix elements in a numerator $$\left\langle \Psi _{k}\right| V_{\varphi }\left| F\right\rangle
=\frac{ 3s^{2}\varphi ^{2}}{4(s+1)}k^{2} \label{Vphiexp}$$
As one can see, the two-magnon spectrum of $H_{0}$ at $Q=0$ and $k\ll 1$ is simply twice an energy of one magnon $\varepsilon
_{2}(k)=2\varepsilon _{1}(k)$ where $$\varepsilon _{1}(k)=2s(1-\cos k)-\frac{s}{2}(1-\cos (2k)) \label{Eonemagnon}$$ and $\varepsilon _{1}(k)=sk^{4}/4$ at small $k$. So, the low-lying states of $H_{0}$ with small number of magnons have energies $\varepsilon _{m}=m\varepsilon _{1}(k)\sim sN^{-4}$, which leads to infrared divergencies in the next-order corrections to the energy. Similar to Eq.(\[Eseries\]) we sum them up using the scaling arguments.
The PT for Eq.(\[Hphi\]) contains two channels $V_{\gamma }$ and $V_{\varphi }$, which are described by two independent scaling parameters. In order to determine these scaling parameters one should estimate large-$N$ behavior of the matrix elements of the operators $V_{\gamma }$ and $V_{\varphi }$ between low-lying states $\left| \Psi _{i}\right\rangle $ and $\left| \Psi
_{j}\right\rangle $, acting in the PT. Since the operators $V_{\gamma }$ and $V_{\varphi }$ create (annihilate) not more than two magnons, we look after only low-lying states with small number of magnons and energies $$\varepsilon _{m}\sim sN^{-4} \label{Emexp}$$
We note that these states are very different from singlet states (with $N/2$ magnons) presented in Eq.(\[Eseries\]) and this fact is crucial.
The diagonal matrix elements for one-magnon states with small quasi-momentum $k$ behave as $$\begin{aligned}
\left\langle k\right| V_{\gamma }\left| k\right\rangle &=&-4s\gamma k^{2}
\nonumber \\
\left\langle k\right| V_{\varphi }\left| k\right\rangle
&=&\frac{3}{4} s\varphi ^{2}k^{2}-s\frac{\varphi ^{4}}{4}+4s\gamma
\varphi ^{2} \label{VgVphi1}\end{aligned}$$ (non-diagonal elements in the one-magnon sector are zero).
For a small number of magnons $m\ll N$ we can treat them as almost independent, because the interactions between magnons gives only corrections of the order of magnon density $\rho =m/N$ to the excitation energies and to the matrix elements. Therefore, large-$N$ behavior of the matrix elements ($k\sim 1/N$) are $$\begin{aligned}
\left\langle \Psi _{i}\right| V_{\gamma }\left| \Psi _{j}\right\rangle &\sim
&\gamma sN^{-2} \nonumber \\
\left\langle \Psi _{i}\right| V_{\varphi }\left| \Psi _{j}\right\rangle
&\sim &\varphi ^{2}sN^{-2} \label{VgVphi2}\end{aligned}$$
These formulae are validated by the exact solution of two-magnon problem.
Now we are ready to identify the scaling parameters of the perturbations $V_{\gamma }$ and $V_{\varphi }$. Similar to Eq.(\[x\]), they are $$\begin{aligned}
\frac{\left\langle \Psi _{i}\right| V_{\gamma }\left| \Psi _{j}\right\rangle
}{\varepsilon _{m}} &\sim &\gamma N^{2} \nonumber \\
\frac{\left\langle \Psi _{i}\right| V_{\varphi }\left| \Psi
_{i}\right\rangle }{\varepsilon _{m}} &\sim &\varphi ^{2}N^{2}
\label{VgVphiexp}\end{aligned}$$
The scaling parameter $\varphi N$ looks natural, because for the finite cyclic system the pitch angle $\varphi $ is quantized as $\varphi _{m}=\frac{2\pi m}{N}$.
The infrared divergencies are absorbed by these scaling parameters so that the divergent part of the perturbation series in both channels has a form $$E^{(div)}=\left\langle \Psi _{i}\right| V_{\varphi }\left| \Psi
_{j}\right\rangle \sum_{m,n=0}^{\infty }c_{mn}\left( \varphi N\right)
^{2m}\left( \gamma N^{2}\right) ^{n} \label{Ediv1}$$ with unknown constants $c_{mn}$.
In order to satisfy the thermodynamic relation $E^{(div)}\sim N$, we rewrite Eq.(\[Ediv1\]) as $$E^{(div)}=Ns\varphi ^{5}\sum_{n=0}^{\infty }g_{n}\left( \varphi N\right)
\left( \frac{\gamma }{\varphi ^{2}}\right) ^{n} \label{Ediv2}$$ where $$g_{n}\left( \varphi N\right) =\sum_{m=0}^{\infty }c_{mn}\left( \varphi
N\right) ^{2n+2m-3} \label{gn}$$ are a set of (generally unknown) scaling functions. They should converge in the thermodynamic limit $N\to \infty $ to some constants $$a_{n}=\lim_{N\to \infty }g_{n}\left( \varphi N\right) \label{an}$$ which are Taylor coefficients of an unknown scaling function $f\left( \frac{\gamma }{\varphi ^{2}}\right) $: $$E^{(div)}=Ns\varphi ^{5}\sum_{n=0}^{\infty }a_{n}\left(
\frac{\gamma }{ \varphi ^{2}}\right) ^{n}=Ns\varphi ^{5}f\left(
\frac{\gamma }{\varphi ^{2}} \right) \label{Ediv3}$$
Thus, collecting the converged first-order $E^{(1)}$ and second-order $E^{(2)}$ corrections with the divergent part $E^{(div)}$ the energy takes the form $$E=-2Ns^{2}\gamma \varphi ^{2}+Ns^{2}\frac{\varphi
^{4}}{8}\frac{s-1/2}{s+1} +Ns\varphi ^{5}f\left( \frac{\gamma
}{\varphi ^{2}}\right) \label{E}$$
At $\gamma =0$ the estimate of the energy relates to the spectrum of the Hamiltonian $H_{0}$. One can see that the energy for $s\geq
1$ is $E\sim N\varphi ^{4}$, which for small $\varphi \sim
\frac{1}{N}$ agrees with the numerical estimate $E\sim N^{-3}$(see Fig.2). However, in the special case $s=1/2$ the second term in Eq.(\[E\]) vanishes and the energy becomes $E\sim N\varphi
^{5}f\left( 0\right) $, which for $\varphi \sim \frac{1}{N}$ agrees again with the numerical estimate $E\sim N^{-4}$ (see Fig.1). From the positivity of the spectrum of the Hamiltonian $H_{0}$ we conclude that $f\left( 0\right) >0$.
Now we need to minimize the ground state energy over $\varphi $. As follows from Eq.(\[E\]) this procedure is different for $s=1/2$ and $s\geq 1$. For the case $s=1/2$ $$E=-N\frac{\gamma \varphi ^{2}}{2}+N\frac{\varphi ^{5}}{2}f\left( \frac{%
\gamma }{\varphi ^{2}}\right) \label{Es12}$$
The comparison of powers in $\varphi $ and $\gamma $ of two terms in Eq.(\[Es12\]) shows that the minimum of $E$ is reached at $\varphi _{\min }\sim \gamma ^{1/3}$. Therefore, $f\left(
\frac{\gamma }{\varphi _{\min }^{2}} \right) \to f\left( 0\right)
$ at $\gamma \to 0$ and the expression for $\varphi _{\min }$ takes a form $$\varphi _{\min }(\gamma )=\left( \frac{2\gamma }{5f(0)}\right) ^{1/3}
\label{phimin12}$$
As was shown above $f\left( 0\right) >0$, which justifies Eq.(\[phimin12\]). The corresponding minimal energy is $$E_{\min }=-\frac{0.3}{\left( 2.5f(0)\right) ^{2/3}}N\gamma ^{5/3}
\label{Emin12}$$
For the case $s\geq 1$ the minimum is defined by the first two terms in Eq.(\[E\]) and $$\begin{aligned}
\varphi _{\min } &=&\sqrt{8\frac{s+1}{s-1/2}}\sqrt{\gamma } \nonumber \\
E_{\min } &=&-8Ns^{2}\frac{s+1}{s-1/2}\gamma ^{2} \label{phimin}\end{aligned}$$
The last term in Eq.(\[E\]) gives the correction to the energy proportional to $\sim Ns\gamma ^{5/2}$.
Thus, we reproduce the critical exponents obtained in Sec.II. However, this special type of the PT allowed us to determine also the factor at $\gamma ^{2}$ for the case $s\geq 1$, which at $s\to
\infty $ tends to the classical result Eq.(\[Eclphicl\]).
According to Eqs.(\[phimin12\]) and (\[phimin\]) the pitch angle $\varphi _{\min }$ has different behavior at $\gamma \to 0$ for $s=1/2 $ and $s\geq 1$. It does not coincide with its classical value (\[phicl\]) for any $s$, but it naturally tends to $\varphi _{\mathrm{cl}}$ at $s\to \infty $. The found non-zero pitch angle $\varphi _{\min }$ indicates the helical (spiral) structure of the ground state. Of course, this does not imply the helical long range order, which is destroyed by strong quantum fluctuations. Instead, this means an incommensurate behavior of the spin correlation function and the pitch angle $\varphi _{\min
}$ can be identified with the quasi-momentum $q_{\max }$ at which the static structure factor takes its maximal value.
Dimerized zigzag model
======================
In order to study the problem of the spontaneous dimerization in the zigzag model (\[H\]) close to transition point $J=1/4$ we add to the Hamiltonian $H$ the dimerization perturbation $V_{\alpha }$ $$\begin{aligned}
H_{d} &=&H+V_{\alpha } \nonumber \\
V_{\alpha } &=&\alpha \sum (-1)^{n}\mathbf{S}_{n}\cdot \mathbf{S}_{n+1}
\label{Hd}\end{aligned}$$
Then, the behavior of the ground state energy $E_{0}(\alpha
,\gamma )$ of the model (\[Hd\]) gives us the dimerization order parameter: $$p(\alpha ,\gamma )=\frac{1}{N}\left\langle \sum (-1)^{n}\mathbf{S}_{n}\cdot
\mathbf{S}_{n+1}\right\rangle =-\frac{1}{N}\frac{\partial E_{0}(\alpha
,\gamma )}{\partial \alpha } \label{pdef}$$
If $E_{0}(\alpha ,\gamma )\sim -N\alpha p(\gamma )$ at $\alpha \to
0$, then the singlet phase of the model (\[H\]) is spontaneously dimerized and $p(\gamma )$ is the corresponding order parameter.
The classical approximation for the model (\[Hd\]) shows that the spins form a double-spiral structure defined by two pitch angles $\varphi $ and $\theta $ so that the rotation angle about the $Y$ axis on $n$-th site is $$\varphi _{n}=n\varphi +\frac{(-1)^{n}}{2}\theta \label{phin}$$
The expansion of the classical energy at $(\alpha ,\gamma ,\varphi
,\theta )\ll 1$ $$E_{\mathrm{cl}}(\alpha ,\gamma ,\varphi ,\theta )=Ns^{2}\left( \frac{\varphi
^{4}}{8}+\frac{\theta ^{2}}{2}-2\gamma \varphi ^{2}-\alpha \varphi \theta
\right) \label{Eclfull}$$ is minimized by the angles $$\begin{aligned}
\varphi _{\mathrm{cl}} &=&\sqrt{8\gamma +2\alpha ^{2}} \nonumber \\
\theta _{\mathrm{cl}} &=&\alpha \varphi _{\mathrm{cl}} \label{thetacl}\end{aligned}$$ which gives the ground state energy at $\alpha ,\gamma \ll 1$ $$E_{\mathrm{cl}}(\alpha ,\gamma )=-\frac{1}{2}Ns^{2}\left( 4\gamma +\alpha
^{2}\right) ^{2} \label{Eclag}$$
As follows from Eq.(\[Eclag\]) $E_{\mathrm{cl}}(\alpha ,\gamma
)$ vanishes on the line $$4\gamma +\alpha ^{2}=0 \label{line}$$ which determines the transition line between the ferromagnetic and the singlet phases for the model (\[Hd\]).
For the case $s=1/2$ the exact singlet ground state on this line is known [@DKO97],[@DKOEPJ]. It has double-spiral long-range order $$\left\langle \mathbf{S}_{i}\cdot \mathbf{S}_{i+n}\right\rangle =\frac{1}{4}%
\cos \varphi _{n}$$ where the angles $\varphi _{n}$ are defined by Eq.(\[phin\]) with pitch angle $\varphi =\frac{2\pi }{N}$ and small shift angle between spirals $\theta =\frac{2\pi }{N}\alpha $. It is interesting that the classical relation $\theta =\alpha \varphi $ (see Eq.(\[thetacl\])) remains for the strong quantum $s=1/2$ case on the transition line. The dimerization parameter on the transition line behaves as $$p_{\mathrm{tr}}=\frac{\varphi \theta }{4}=\frac{\pi ^{2}}{N^{2}}\alpha
\label{pline}$$
Though $p_{\mathrm{tr}}\neq 0$ the spontaneous dimerization is absent on the transition line in the thermodynamic limit.
As follows from Eq.(\[pdef\]) the classical approximation yields the dimerization order for the model (\[Hd\]): $$p_{\mathrm{cl}}(\alpha ,\gamma )=s^{2}\varphi _{\mathrm{cl}}\theta _{\mathrm{%
cl}}=2s^{2}\alpha \left( 4\gamma +\alpha ^{2}\right) \label{pcl}$$
Eq.(\[pcl\]) shows that the dimerization vanishes on the transition line Eq.(\[line\]), which accords with Eq.(\[pline\]), and it vanishes also at $\alpha =0$, which implies the absence of the spontaneous dimerization for the model (\[H\]). Since the classical approximation describes the limit $s\to \infty $, one can expect that at least in the limit $s\to
\infty $ the spontaneous dimerization in the model (\[H\]) is absent.
Following the classical picture we transform the local axes on $n$-th site by a rotation about the $Y$ axis by angle $\varphi
_{n}$ as written in Eq.(\[phin\]), but not fixing $\varphi $ and $\theta $ to their classical values. Under this unitary transformation the Hamiltonian $H_{d}$ (\[Hd\]) takes a form $$H_{\varphi ,\theta }=H_{0}+V(\alpha ,\gamma ,\varphi ,\theta )
\label{Hdimer}$$ where the perturbation $V(\alpha ,\gamma ,\varphi ,\theta )$ has a very cumbersome form and we do not present it here.
Similar to the analysis done in Sec.III we develop PT in $V(\alpha
,\gamma ,\varphi ,\theta )$ to the fully polarized state (\[Fstate\]). The first order in $V$ exactly reproduces the classical energy (\[Eclfull\]) $$E^{(1)}=E_{\mathrm{cl}}(\alpha ,\gamma ,\varphi ,\theta ) \label{dimerE1}$$
The second order correction to the ground state energy gives $$E^{(2)}=-Ns^{2}\left( \frac{3\varphi ^{4}}{16(s+1)}+\frac{(\theta -\alpha
\varphi )^{2}}{2}\right) \label{dimerE2}$$
As one can see the terms containing the angle $\theta $ in the first order $E^{(1)}$ are exactly compensated by the contributions of the second order $E^{(2)}$. This result is rather unexpected. The classical approximation corresponds to the limit $s\to \infty
$ and it would seem that the quantum effects will give relative corrections $\sim s^{-1}$ to the energy. However, in this case the quantum corrections have the same order in $s$ as the classical energy.
The next-order corrections contain infrared divergencies and we treat them using the scaling arguments similar to that done in Sec.III. The analysis shows that the most divergent parts of the PT are accumulated in the following scaling parameters: $$\begin{aligned}
x_{\alpha } &\sim &\alpha N \nonumber \\
x_{\varphi } &\sim &\varphi N \nonumber \\
x_{\gamma } &\sim &\gamma N^{2} \label{scaling}\end{aligned}$$
It turns out that the angle $\theta $ is not accompanied by the infrared divergencies and, therefore, it does not form a scaling parameter. After the algebraic manipulations with the divergent series of the PT similar to Eq.(\[Ediv1\])-(\[Ediv3\]), the main contribution of the next-order corrections to the ground state energy at $N\to \infty $ takes a form $$E^{(div)}=Ns\varphi ^{5}f\left( \frac{\gamma }{\varphi
^{2}},\frac{\alpha }{\sqrt{\gamma }}\right) \label{dimerEdiv}$$ where $f\left( \frac{\gamma }{\varphi ^{2}},\frac{\alpha
}{\sqrt{\gamma }} \right) $ is unknown scaling function of two scaling variables. The angle $\theta $ does not contribute to the most divergent parts of the PT Eq.(\[dimerEdiv\]), because it is not accompanied by the infrared divergencies.
Collecting the corrections $E^{(1)}$ and $E^{(2)}$ with the scaling part $E^{(div)}$ we obtain the leading terms for the ground state energy: $$E=-\frac{1}{2}Ns^{2}(4\gamma +\alpha ^{2})\varphi ^{2}+Ns^{2}\frac{\varphi
^{4}}{8}\frac{s-1/2}{s+1}+Ns\varphi ^{5}f_{s}\left( \frac{\gamma }{\varphi
^{2}},\frac{\alpha }{\sqrt{\gamma }}\right) \label{dimerE}$$
As follows from Eq.(\[dimerE\]) the leading terms do not contain the angle $\theta $. In fact, we have checked that the energy does not contain terms up to $\sim \theta ^{4}$. This result is not surprising. In general, the PT in $V(\alpha ,\gamma ,\varphi
,\theta )$ for the ferromagnetic state results in the energy $E(\alpha ,\gamma ,\varphi ,\theta )$ depending on $\theta $. On the other hand, the spectra of the Hamiltonians $H_{d}$ (\[Hd\]) and $H_{\varphi ,\theta }$ (\[Hdimer\]) coincide and the eigenvalues $E_{n}(\alpha ,\gamma )$ of both Hamiltonians do not depend on $\theta $ and $\varphi $. Therefore, for any values of $\theta $ and $\varphi $ the PT leads to one of the determinate levels $E_{n}(\alpha ,\gamma )$ of the Hamiltonian $H$. The pitch angle $\varphi $ is quantized as $\varphi _{n}= \frac{2\pi n}{N}$ for finite $N$, and the PT with different $\varphi _{n}$ leads to generally different levels $E_{n}(\alpha ,\gamma )$. At the same time, in contrast to the pitch angle $\varphi $ the angle $\theta
$ is a continuous variable even for finite $N$. Therefore, the continuity condition of the dependence $E(\alpha ,\gamma ,\varphi
,\theta )$ on $\theta $ implies that the PT in $V(\alpha ,\gamma
,\varphi ,\theta )$ with any value of $\theta $ leads to the same energy level as at $\theta =0$. In other words, the obtained in the PT ground state energy does not depend on $\theta $.
This fact is an argument for the absence of the spontaneous dimerization in the zigzag model (\[H\]). Really, the PT with any value of $\theta $ brings to the same state as it does at $\theta =0$. But at $\theta =0$ and $\alpha =0$ the PT in $V(\alpha ,\gamma ,\varphi ,\theta )$ reduces to the PT (\[Vphi\]) in $V_{\varphi }$ considered in Sec.III. There are no terms in the perturbation $V_{\varphi }$ which break translational symmetry and can potentially lead to the dimer order.
However, the above arguments do not prove the absence of the spontaneous dimerization in the model (\[H\]). The rigorous method is to calculate the dimer order parameter directly from Eq.(\[pdef\]), which we follow below.
The minimization of the ground state energy over $\varphi $ is performed in the same manner as was done in Sec.III. For the case $s=1/2$ the second term in Eq.(\[dimerE\]) disappears. The comparison of powers in $\varphi $ of two terms in Eq.(\[dimerE\]) shows that the minimum of $E$ is reached at $\varphi _{\min }\sim (4\gamma +\alpha ^{2})^{1/3}$. Therefore, we substitute $f\left( \frac{\gamma }{\varphi _{\min
}^{2}},\frac{\alpha }{\sqrt{\gamma }} \right) \to f\left(
0,\frac{\alpha }{\sqrt{\gamma }}\right) $ at $\gamma \to 0$ and the expression for $\varphi _{\min }$ becomes $$\varphi _{\min }(\alpha ,\gamma )=\left( 4\gamma +\alpha ^{2}\right)
^{1/3}g(\eta ) \label{dimerphi12}$$ where $\eta =\frac{\alpha }{\sqrt{\gamma }}$ and $g(\eta )=\left[ 10f\left(
0,\eta \right) \right] ^{-1/3}$.
The corresponding minimal energy is $$E_{\min }=-\frac{3N}{40}\left( 4\gamma +\alpha ^{2}\right) ^{5/3}g^{2}(\eta )
\label{dimerEmin12}$$
For the case $s\geq 1$ the energy minimum is defined by the first two terms in Eq.(\[dimerE\]) and $$\begin{aligned}
\varphi _{\min } &=&\sqrt{2\frac{s+1}{s-1/2}}\sqrt{4\gamma +\alpha ^{2}}
\nonumber \\
E_{\min } &=&-Ns^{2}\frac{s+1}{2s-1}\left( 4\gamma +\alpha ^{2}\right)
^{2}+Ns\left( 4\gamma +\alpha ^{2}\right) ^{5/2}g_{s}\left( \eta \right)
\label{dimerEmin}\end{aligned}$$ where $g_{s}\left( \eta \right) =\left[ \frac{2\left( s+1\right)
}{s-1/2} \right] ^{5/2}$ $f_{s}\left( \frac{\gamma }{\varphi
_{\min }^{2}},\eta \right) $. So, the leading term in the ground state energy for $s\geq 1$ is determined by the regular parts of the PT, while the scaling part (the last term in Eq.(\[dimerEmin\])) gives only small correction to the energy.
We see that the difference in the critical exponents for the cases $s=1/2$ and $s\geq 1$ remains for more general dimerized model (\[Hdimer\]) as well. The pitch angle $\varphi _{\min }$ and the ground state energy $E_{\min }$ naturally vanish on the transition line Eq.(\[line\]) for both cases $s=1/2$ and $s\geq 1$.
The dimerization of the model (\[Hdimer\]) is defined as a derivative of the energy with respect to $\alpha $ (\[pdef\]). As follows from Eqs.(\[dimerEmin12\]),(\[dimerEmin\]) the dimerization at $\gamma =0$ appears with critical exponents $$\begin{aligned}
\left. p\right| _{\gamma =0} &\sim &\alpha ^{7/3},\qquad s=1/2 \nonumber \\
\left. p\right| _{\gamma =0} &\sim &\alpha ^{3},\qquad s\geq 1 \label{pg0}\end{aligned}$$
As for the model (\[H\]) ($\alpha =0$), the dimerization depends on the behavior of the scaling functions $f_{s}\left( \eta \right)
$ at small $\eta $. There are two possible scenarios. First, the expansion of $f_{s}\left( \eta \right) $ at $\eta \to 0$ is $f_{s}\left( \eta \right) =a+b\eta ^{\mu }$ with some constants $a$ and $b$ and $\mu >1$, so that $f_{s}^{\prime }\left( 0\right)
=0$. In this case the dimer parameter is zero for the model (\[H\]). Second, $f_{s}\left( \eta \right) =a+b\eta $ at $\eta
\to 0$ and $f_{s}^{\prime }\left( 0\right) =b$. For this case the translation symmetry of the zigzag model (\[H\]) is spontaneously broken and dimer long-range order $p(\gamma )$ appears as $$\begin{aligned}
\left. p\right| _{\alpha =0} &\sim &\gamma ^{7/6},\qquad s=1/2 \nonumber \\
\left. p\right| _{\alpha =0} &\sim &\gamma ^{2},\qquad s\geq 1 \label{pa0}\end{aligned}$$
Here the critical exponent for the dimerization (\[pa0\]) for $s\geq 1$ comes not from the leading term in Eq.(\[dimerEmin\]), but from the scaling correction (the last term).
Unfortunately, we do not have any information about the behavior of the scaling functions $f_{s}\left( \eta \right) $. Therefore, we can only state that if the zigzag model (\[H\]) is in the dimerized singlet phase at $\gamma >0$, then the critical exponents for the dimer LRO are given by Eqs.(\[pa0\]).
Summary
=======
We have studied the frustrated Heisenberg chain with the nearest ferromagnetic and the next-nearest neighbor antiferromagnetic exchange interactions. It was shown that the behavior of the model in the vicinity of the transition point between the ferromagnetic and the singlet phases depends on the value of the spin. For $s=1/2$ the critical exponent characterizing the behavior of the energy is $\beta =5/3$ in contrast to the ‘classical’ exponent $\beta =2$ for $s\geq 1$. This difference is a result of different finite-size dependencies of the spectrum at the transition point $\gamma =0$ for the cases $s=1/2$ and $s\geq 1$. The pitch angles characterizing the incommensurate behavior of the spin correlation functions are different for $s=1/2$ and $s\geq 1$, too. In particular, the pitch angle $ \varphi $ of the spiral is proportional to $\gamma ^{1/3}$ for $s=1/2$ and to $\gamma ^{1/2}$ for $s\geq 1$. It means that the considered model with $s=1/2$ is special and the quantum effects for this value of $s$ are the most strong.
One more intriguing question is related to the existence of the spontaneous dimerization in the singlet phase. In order to study this problem, we added to the Hamiltonian $H$ the dimerization term and treated it as a perturbation. Unfortunately, the used special version of the PT did not give us a rigorous answer about the spontaneous dimerization in the singlet phase. Instead, under assumption of the existence of the spontaneous dimerization, the PT allowed us to estimate the critical exponent of the dimer order parameter. Besides, using the special version of the PT we obtain the critical exponents of the ground state energy and the dimer order for the dimerized version of the model (\[H\]).
We would like to thank S.-L. Drechsler and R. O. Kuzian for helpful discussions. D.D. was supported by INTAS YS Grant Nr. 05– 109–4916 and President of RF Grant MK-3987.2005.2. The numerical exact diagonalization for finite systems were performed using the J.Schulenburg’s [*spinpack*]{}.
[99]{}
H.-J. Mikeska, A.K. Kolezhuk, in [*Quantum Magnetism*]{}, Lecture Notes in Physics vol. [**645**]{}, pp. 1, U. Schollwöck, J. Richter, D.J.J. Farnell, R.F. Bishop, Eds. (Springer-Verlag, Berlin, 2004).
F.D.M. Haldane, Phys.Rev. B [**25**]{}, R4925 (1982).
T. Tonegawa and I. Harada, J.Phys.Soc.Jpn [**56**]{}, 2153 (1987).
K. Nomura and K. Okamoto, Phys.Lett. [**169A**]{}, 433 (1992).
R. Bursill, G.A. Gehring, D.J.J. Farnell, J.B. Parkinson, T. Xiang, and C. Zeng, J.Phys: Condens.Matter [**7**]{}, 8605 (1995).
C.K. Majumdar and D.K. Ghosh, J.Math.Phys. [**10**]{}, 1388 (1969).
S.R. White and I. Affleck, Phys.Rev. B [**54**]{}, 9862 (1996).
T. Tonegawa and I. Harada, J.Phys.Soc.Jpn [**58**]{}, 2902 (1989).
A.V. Chubukov, Phys.Rev. B [**44**]{}, R4693 (1991).
V.Ya. Krivnov and A.A. Ovchinnikov, Phys.Rev. B [**53**]{}, 6435 (1996).
A.A. Aligia, Phys.Rev. B [**63**]{}, 14402 (2001).
F. Heidrich-Meisner, A. Honecker, and T. Vekua, Phys.Rev. B [**74**]{}, R020403 (2006).
H.T. Lu, Y.J. Wang, S. Qin, and T. Xiang, cond-mat/0603519.
Y. Mizuno, T. Tohyama, S. Maekawa, T. Osafune, N. Motoyama, H. Eisaki, and S. Uchida, Phys.Rev. B [**57**]{}, 5326 (1998).
M. Enderle, C. Mukherjee, B. Fak, R.K. Kremer, J.-M. Broto, H. Rosner, S.-L. Drechsler, J. Richter, J. Malek, A. Prokofiev, W. Assmus, S. Pujol, J.-L. Raggazzoni, H. Rakoto, M. Rheinstaedter, and H.M. Ronnow, Europhys. Lett. [**70**]{}, 237 (2005).
S.-L. Drechsler, J. Richter, A.A. Gippius, A. Vasiliev, A.A. Bush, A.S. Moskvin, J. Malek, Y. Prots, W. Schnelle, and H. Rosner, Europhys. Lett. [**73**]{}, 83 (2006).
M. Hase, H. Kuroe, K. Ozawa, O. Suzuki, H. Kitazawa, G. Kido and T. Sekine, Phys.Rev. B [**70**]{}, 104426 (2004).
H.P. Bader and R. Schilling, Phys.Rev. B [**19**]{}, 3556 (1979).
T. Hamada, J. Kane, S. Nakagawa, and Y. Natsume, J.Phys.Soc.Jpn. [**57**]{}, 1891 (1988); [**58**]{}, 3869 (1989).
S.-L. Drechsler, O. Volkova, A.N. Vasiliev, N. Tristan, J. Richter, M. Schmitt, H. Rosner, J. Malek, R. Klingeler, A.A. Zvyagin, and B. Büchner (to be published)
S.-L. Drechsler, N. Tristan, R. Klingeler, B. Büchner, J. Richter, J. Malek, O. Volkova, A. Vasiliev, M. Schmitt, A. Ormeci, C. Loison, W. Schnelle, and H. Rosner, J.Phys.: Condens. Matter [**18**]{}, 1 (2006).
D.V. Dmitriev, V.Ya. Krivnov, and A.A. Ovchinnikov, Phys.Rev. B [**56**]{}, 5985 (1997).
D.V. Dmitriev and V.Ya. Krivnov, Phys.Rev. B [**73**]{}, 024402 (2006).
A.A. Nersesyan, A.O. Gogolin, and F.H.L. Essler, Phys.Rev.Lett. [**81**]{}, 910 (1998).
D.C. Cabra, A. Honecker and P. Pujol, Eur.Phys.J. B [**13**]{}, 55 (2000).
C. Itoi and S. Qin, Phys.Rev. B [**63**]{}, 224423 (2001).
J.L. Cardy in *Phase transition and critical phenomena*, eds. Domb and Lebowitz, Vol.XI, (Academic, New York, 1986).
D.V. Dmitriev, V.Ya. Krivnov, and A.A. Ovchinnikov, Eur.Phys.J. B [**14**]{}, 91 (2000).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We use new multi-wavelength radio observations, made with the VLA and Effelsberg telescopes, to study the magnetic field of the nearby galaxy M51 on scales from $200{\,\mathrm{pc}}$ to several ${\,\mathrm{kpc}}$. Interferometric and single dish data are combined to obtain new maps at [$\lambda\lambda3,6\,\mathrm{cm}$]{} in total and polarized emission, and earlier [$\lambda20\,\mathrm{cm}$]{} data are re-reduced. We compare the spatial distribution of the radio emission with observations of the neutral gas, derive radio spectral index and Faraday depolarization maps, and model the large-scale variation in Faraday rotation in order to deduce the structure of the regular magnetic field. We find that the [$\lambda20\,\mathrm{cm}$]{} emission from the disc is severely depolarized and that a dominating fraction of the observed polarized emission at [$\lambda6\,\mathrm{cm}$]{} must be due to anisotropic small-scale magnetic fields. Taking this into account, we derive two components for the regular magnetic field in this galaxy: the disc is dominated by a combination of azimuthal modes, $m=0+2$, but in the halo only an $m=1$ mode is required to fit the observations. We disuss how the observed arm-interarm contrast in radio intensities can be reconciled with evidence for strong gas compression in the spiral shocks. In the inner spiral arms, the strong arm–interarm contrasts in total and polarized radio emission are roughly consistent with expectations from shock compression of the regular and turbulent components of the magnetic field. However, the average arm–interam contrast, representative of the radii $r>2{\,\mathrm{kpc}}$ where the spiral arms are broader, is not compatible with straightforward compression: lower arm–interarm contrasts than expected may be due to resolution effects and *decompression* of the magnetic field as it leaves the arms. We suggest a simple method to estimate the turbulent scale in the magneto-ionic medium from the dependence of the standard deviation of the observed Faraday rotation measure on resolution. We thus obtain an estimate of $50{\,\mathrm{pc}}$ for the size of the turbulent eddies.'
author:
- |
A. Fletcher,$^{1}$[^1] R. Beck,$^2$ A. Shukurov,$^1$ E. M. Berkhuijsen$^2$ and C. Horellou$^3$\
$^1$School of Mathematics and Statistics, Newcastle University, Newcastle-upon-Tyne NE1 7RU, U.K.\
$^2$Max-Planck-Institut für Radioastronomie, Auf dem Hügel 69, 53121 Bonn, Germany\
$^3$Onsala Space Observatory, Chalmers University of Technology, 439 92 Onsala, Sweden
title: Magnetic fields and spiral arms in the galaxy M51
---
galaxies: spiral – galaxies: magnetic fields – galaxies: ISM – galaxies: individual: M51
Introduction {#sec:intro}
============
The Whirlpool galaxy, M51 or NGC5194, is one of the classical grand-design spiral galaxies. The two spiral arms of M51 can be traced through more than 360 in azimuthal angle in numerous wavebands. M51 is probably perturbed by a recent encounter with its companion galaxy NGC5195. Such interactions usually result in enhanced star formation, either localized or global, as tidal forces and density waves compress the interstellar medium. In M51 this may have resulted in two systems of density waves [@Elmegreen89].
M51 was the first external galaxy where polarized radio emission was detected [@Mathewson72] and one of the few external galaxies where optical polarization has been studied [@Scarrott:1987]. @Neininger:1992 and @Horellou:1992 found that the lines of the regular magnetic field in M51 have a spiral shape, but their resolution was too low to determine how well the field is aligned with the optical spiral arms. @Horellou:1992 realized that, at wavelengths of $\lambda\ge18$ cm, only polarized emission from a foreground layer reaches the observer because of Faraday depolarization, and the Faraday rotation measures obtained at the longer wavelengths are much smaller than those at shorter wavelengths. @Heald:2009 observed a fractional polarization at [$\lambda22\,\mathrm{cm}$]{} of around $5\%$ in the optical disc, increasing to $30\%$ at large radii, and Faraday rotation measures of a similar magnitude to those found by @Horellou:1992. @Berkhuijsen:1997 analyzed the global magnetic field of M51 using radio data at four frequencies and found that the average orientation of the fitted magnetic field is similar to the average pitch angle of the optical spiral arms measured by @Howard90. @Berkhuijsen:1997 represented the magnetic field in the disc of M51 as a superposition of periodic azimuthal modes, with about equal contribution from the axisymmetric $m=0$ and the bisymmetric $m=1$ ones. Their fit contains a magnetic field reversal at about 5 kpc radius which extends over a few kpc in azimuth. Furthermore, @Berkhuijsen:1997 found evidence for an axisymmetric magnetic field in the halo of M51 (visible at [$\lambda18\,\mathrm{cm}$]{} and [$\lambda20\,\mathrm{cm}$]{}) with a reversed direction (inwards) with respect to the axisymmetric mode of the disc field (outwards).
The results of @Berkhuijsen:1997 indicate that the magnetic field of M51 is strong and partly regular, with some interesting properties. However, some of their results may be affected by the low angular resolution which was limited by the Effelsberg single-dish data at [$\lambda2.8\,\mathrm{cm}$]{}. Furthermore, no correction for missing large-scale structure could be applied to the VLA [$\lambda6\,\mathrm{cm}$]{} polarization data because no single-dish map was available at that time.
Here we present a refined analysis based on new data with higher resolution and better sensitivity. Our new surveys of M51 in total and linearly polarized [$\lambda3\,\mathrm{cm}$]{} and [$\lambda6\,\mathrm{cm}$]{} radio continuum emission combines the resolution of the VLA with the sensitivity of the Effelsberg single-dish telescope. The new maps are of comparable resolution to the maps of the CO(1–0) emission [@Helfer03] and mid-infrared dust emission [@Sauvage96; @Regan06] and are of unprecedented sensitivity. The shape of the spiral radio arms and the comparison to the arms seen with different tracers were discussed in a separate paper [@Patrikeev06].
The spiral shocks in M51 are strong and regular [@Aalto99] and offer the possibility to compare arm–interarm contrasts of gas and the magnetic field. We analyse our new data to try to separate the contribution to the observed polarized emission from regular (or mean) magnetic fields and anisotropic random magnetic fields produced by compression and/or shear in the spiral arms.
The new high-resolution polarization maps allow us for the first time to investigate in detail the interaction between the magnetic fields and the shock fronts. Results from the barred galaxies NGC 1097 and NGC 1365 showed that the small-scale and the large-scale magnetic field components behave differently, i.e. the small-scale field is compressed significantly in the bar’s shock, while the large-scale field is hardly compressed [@Beck:2005]. This was interpreted as a strong indication that the large-scale field is coupled to the warm, diffuse gas which is only weakly compressed. We investigate whether a similar decoupling of the regular magnetic field from the dense gas clouds is suggested by the observed arm–interarm contrast in M51.
The basic parameters we adopt for M51 are: centre’s right ascension $\mathrm 13^\mathrm{h}29^\mathrm{m}52\fs709$ and declination (J2000) $+47\degr 11\arcmin 42\farcs 59$ [@Ford85]; distance 7.6 Mpc, [@Ciardullo:2002], thus $1"\approx 37{\,\mathrm{pc}}$; position angle of major axis $-10\degr$ ($0\degr$ is North) and inclination $-20\degr$ ($0\degr$ is face-on) [@Tully74]. The inclination is measured from the galaxy’s rotation axis to the line-of-sight, viewed from the northern end of the major axis, and its sign is important for the geometry of the model discussed in Sect. \[sec:B\].
Observations and data reduction {#sec:obs}
===============================
{width="48.00000%"} {width="48.00000%"}
VLA observations
----------------
M51 was observed in October 2001 at [$\lambda3.5\,\mathrm{cm}$]{} with the VLA[^2] in the compact D configuration and in August 2001 at [$\lambda6.2\,\mathrm{cm}$]{} in the C configuration. Two pointings of the array, at the northern and southern parts of the galaxy, were required to obtain complete coverage of M51 at [$\lambda3.5\,\mathrm{cm}$]{}. The data were edited, calibrated and imaged using standard AIPS procedures and VLA calibration sources. After initial examination of the data and the flagging of bad visibilities, self-calibration was used. After correction for the pattern of the primary beam the two pointings at [$\lambda3.5\,\mathrm{cm}$]{} were mosaiced using the AIPS task `LTESS`. The calibrated [$\lambda6.2\,\mathrm{cm}$]{} C-array uv-data were combined with existing [$\lambda6.2\,\mathrm{cm}$]{} D-array data [@Neininger:1996].
A new [$\lambda20\,\mathrm{cm}$]{} map in total power and polarization is presented in Fig. \[fig:20\], based on the C-array data of @Neininger:1996 combined with D-array data of @Horellou:1992. All of the observations were re-reduced for this work, the two data sets were combined and then smoothed to a resolution of $15\arcsec$ to improve the signal-to-noise ratio. The resulting maps at [$\lambda20\,\mathrm{cm}$]{} contain information on all scales down to the beamsize as the primary beam of the VLA at [$\lambda20\,\mathrm{cm}$]{} in the D-array is $\sim 30\arcmin$, about twice the size of M51 on the sky.
Different weighting schemes were used in the final imaging of the data to produce maps with either high resolution (natural weighting) or high signal-to-noise (uniform weighting) and for a compromise between the two extremes (robust weighting, obtained by setting the parameter `ROBUST=0` in the AIPS task `IMAGR`). Maps of the Stokes parameters $I$, $Q$ and $U$ were produced in each case. The optimum number of iterations of the CLEAN algorithm used to produce the images was determined individually for each map. Following slight smoothing, the $Q$ and $U$ maps were combined to give the polarized intensity $PI=\sqrt{Q^2 + U^2}$, using a first-order correction for the positive bias [@Wardle74].
----------- ------ ----------- ----------------------------- -----------------------------
$\lambda$ HPBW Weighting $\sigma_{I}$ $\sigma_{PI}$
(cm) (${\,\mu\mathrm{Jy/beam}}$) (${\,\mu\mathrm{Jy/beam}}$)
3 8 robust $12$ $10$
3 15 natural $20$ $ \phantom{1}8$
6 4 uniform $15$ $10$
6 8 robust $25$ $10$
6 15 natural $30$ $10$
20 15 natural $20$ $13$
----------- ------ ----------- ----------------------------- -----------------------------
: \[tab:obs\] The merged VLA and Effelsberg maps discussed in this paper, their half-power beam widths, imaging weighting schemes and their r.m.s. noises in total and polarized emission, $\sigma_{I}$ and $\sigma_{PI}$, respectively
{width="49.00000%"} {width="49.00000%"}
![\[fig:20\] Contours of [$\lambda20\,\mathrm{cm}$]{} total radio emission at $15\arcsec$ resolution, overlaid on the same optical image as in Fig. \[fig:i15\]. Total intensity contours are at 6, 12, 24, 36, 48, 96, 192 times the noise level of $20{\,\mu\mathrm{Jy/beam}}$. Also shown are the $B$-vectors of polarized emission: the plane of polarization of the observed electric field rotated by 90, not corrected for Faraday rotation, with a length proportional to the polarized intensity $PI$ and only plotted where $PI\ge 3\sigma_{PI}$.](fig3){width="47.00000%"}
Effelsberg observations and merging
-----------------------------------
In order to correct the VLA maps at [$\lambda3\,\mathrm{cm}$]{} for missing extended emission we made observations of M51 in total intensity and polarization with the 100-m Effelsberg telescope [^3] in December 2001 and April 2002 using the sensitive [$\lambda3.6\,\mathrm{cm}$]{} (1.1 GHz bandwidth) receiver. We obtained 44 maps of a 12$\arcmin$ by 12$\arcmin$ field around M51, scanned in orthogonal directions. Each map in $I$, $Q$ and $U$ was edited and baseline corrected individually, then all maps in each Stokes parameter were combined using a basket weaving method [@Emerson88]. The r.m.s. noise in the final maps, after slight smoothing to $90\arcsec$, is $200{\,\mu\mathrm{Jy/beam}}$ in total intensity and $20{\,\mu\mathrm{Jy/beam}}$ in polarized intensity.
At [$\lambda6.2\,\mathrm{cm}$]{}, ten maps of a 41$\arcmin$ by 34$\arcmin$ field were observed in November 2003 with the 4.85 GHz (500 MHz bandwidth) dual-horn receiver. The combined data resulted in a new $180\arcsec$ resolution image with r.m.s. noise of $250{\,\mu\mathrm{Jy/beam}}$ in total intensity and $25{\,\mu\mathrm{Jy/beam}}$ in polarization.
Maps from the VLA and Effelsberg were combined using the AIPS task `IMERG`. A useful description of the principles of merging single dish and interferometric data is given by @Stanimirovic02. The range of overlap in the uv plane between the two images (parameter `uvrange` in `IMERG`) was estimated as follows: we assumed an effective Effelsberg diameter of about 60 m to estimate the maximum extent of the single dish in the uv space ($1.7{\,\mathrm{k}\lambda}$ at [$\lambda3.6\,\mathrm{cm}$]{} and $1.0{\,\mathrm{k}\lambda}$ at [$\lambda6.2\,\mathrm{cm}$]{}) and used the minimum separation of the VLA antennas in the D-array configuration, $35$ m, to calculate the minimum coverage of the interferometer in the uv space ($1.0{\,\mathrm{k}\lambda}$ at [$\lambda3.6\,\mathrm{cm}$]{} and $0.6{\,\mathrm{k}\lambda}$ at [$\lambda6.2\,\mathrm{cm}$]{}). We then varied these parameters in order to find the optimum overlap in the uv space by comparing the integrated total flux in the merged maps with that of the single dish maps; the optimal uv-ranges for merging were found to be $1.0\rightarrow1.6{\,\mathrm{k}\lambda}$ at [$\lambda3.6\,\mathrm{cm}$]{} and $0.5\rightarrow0.7{\,\mathrm{k}\lambda}$ at [$\lambda6.2\,\mathrm{cm}$]{}. Merging of the $Q$ and $U$ maps was carried out using the same optimum uv range as found for $I$.
The fraction of total emission (Stokes $I$) present in the VLA maps – those produced using natural weighting, and hence with the highest signal-to-noise ratio – is about 30% at [$\lambda3\,\mathrm{cm}$]{} and close to 50% at [$\lambda6\,\mathrm{cm}$]{} compared to the merged maps. Small-scale fluctuations in $Q$ and $U$ due to variations in the magnetic field orientation and Faraday rotation in M51 mean that the polarized emission is less severely affected by missing large-scales (alternatively, the single dish detects the large-scale emission missed by an interferometer *but* simultaneously suffers from stronger wavelength-independent beam depolarization). At [$\lambda3\,\mathrm{cm}$]{} the VLA map contains about 75% of the polarized emission present in the merged map, with about 85% present at [$\lambda6\,\mathrm{cm}$]{}.
Following the merging, the maps in $I$, $Q$ and $U$ were convolved with a Gaussian beam to give a slightly coarser resolution and higher signal-to-noise ratio. The maps that are discussed in this paper are listed in Table \[tab:obs\] along with the r.m.s. noises in total and polarized intensity.
The M51 maps
============
The spiral arms of M51 as seen in radio emission {#sec:desc}
------------------------------------------------
Figure \[fig:i15\] shows the total radio continuum emission and the $B$-vectors of polarized emission (the observed plane of linear polarization rotated by 90) at [$\lambda3\,\mathrm{cm}$]{} and [$\lambda6\,\mathrm{cm}$]{}, overlaid on a Hubble Space Telescope optical image. At the assumed distance of 7.6Mpc, the $15\arcsec$ resolution corresponds to $560{\,\mathrm{pc}}$ and $590{\,\mathrm{pc}}$ along the major and minor axes, respectively. The distribution of polarized emission at [$\lambda\lambda3,6\,\mathrm{cm}$]{} is shown in Fig. \[fig:pi15\] at $15\arcsec$ resolution. The extensive [$\lambda20\,\mathrm{cm}$]{} total emission disc is shown in Fig. \[fig:20\] also at $15\arcsec$ resolution.
The total emission at [$\lambda3\,\mathrm{cm}$]{} and [$\lambda6\,\mathrm{cm}$]{} in Fig. \[fig:i15\] shows a close correspondence with the optical spiral arms whereas the [$\lambda20\,\mathrm{cm}$]{} total emission (Fig. \[fig:20\]) and [$\lambda3\,\mathrm{cm}$]{} and [$\lambda6\,\mathrm{cm}$]{} polarized emission (Fig. \[fig:pi15\]) are spread more evenly across the galactic disc. Compact, bright peaks of total emission coincide with complexes of H[ii]{} regions in the spiral arms, as expected if thermal bremsstrahlung is a significant component of the radio signal at these peaks. The flatter spectral index ($\alpha\lesssim
0.6$ where $I\propto \nu^{-\alpha}$, Fig. \[fig:spx\]) and the absence of the corresponding peaks in polarized radio emission (Fig. \[fig:pi15\]) suggest that a significant proportion of the centimeter wavelength radio emission in these peaks is thermal. The extended [$\lambda20\,\mathrm{cm}$]{} total emission and [$\lambda3\,\mathrm{cm}$]{} and [$\lambda6\,\mathrm{cm}$]{} polarized emission accurately trace the synchrotron component of the radio continuum.
Ridges of enhanced polarized emission are prominent in the inner galaxy; some are located on the optical spiral arms but others are located in-between the arms. A detailed analysis of the location and pitch angles of the spiral arms traced by different observations and a comparison of the pitch angles with the orientation of the regular magnetic field is given by @Patrikeev06. They find systematic shifts between the spiral ridges seen in polarized and total radio emission, integrated CO line emission and infrared emission, which are consistent with the following sequence in a density wave picture: firstly, shock compresses gas and magnetic fields (traced by polarized radio emission), then molecules are formed (traced by CO) and finally thermal emission is generated (traced by infrared). @Patrikeev06 also show that while the pitch angle of the regular magnetic field is fairly close to that of the gaseous spiral arms at the location of the arms, the magnetic field pitch angle changes from the above by around $\pm15\degr$ in interarm regions.
The connection between polarized radio emission and gaseous spiral arms {#sec:conn2}
-----------------------------------------------------------------------
![ Contours of the [$\lambda6\,\mathrm{cm}$]{} polarized radio emission (VLA and Effelsberg combined) in the central $\sim 3\arcmin \times 4\arcmin$ of M51 at $8\arcsec$ resolution, along with Faraday-rotation corrected B-vectors, overlaid on the map of integrated CO(1–0) line emission of @Helfer03. Contours are at $3,5,7,9$ times the noise level of $10{\,\mu\mathrm{Jy/beam}}$.[]{data-label="fig:inner:co"}](fig4){width="45.00000%"}
In Fig. \[fig:inner:co\], the [$\lambda6\,\mathrm{cm}$]{} polarized emission and the orientation of the regular magnetic field in the central $8{\,\mathrm{kpc}}$ of M51 are overlaid onto an image of the spiral arms as traced by the CO(1–0) integrated line emission. Part of the polarized emission appears to be concentrated in elongated arm-like structures that sometimes coincide with the gas spiral.
The correspondence between polarized and CO arms is good along most of the northern arm in Fig. \[fig:inner:co\] (called Arm 1 in the rest of the paper) and in the inner part of the southern arm (called Arm 2). Arm 1 continues towards the south where the polarized emission longer coincides with the CO and optical arms, but becomes broader further out (Fig. \[fig:pi15\]).
Moving along Arm 2, the excellent overlap of the radio polarization and CO in the inner arm ends abruptly further in the south (Fig. \[fig:inner:co\]), beyond which the distribution of polarized emission is very broad and located on the galactic centre side of the gaseous spiral arm: the peak of this interarm polarized emission corresponds very closely with the pronounced peaks in radial velocity, $v_r\simeq 100{\,\mathrm{km\,s^{-1}}}$, derived by @Shetty:07 (shown in their Fig. 14). This may be an indication that strong shear in the interarm gas flow is producing this magnetic feature. At about 6 kpc radius, the polarization Arm 2 *crosses* the CO and optical arms in the east, followed in the northeast by a bending away from the optical arm towards the companion galaxy NGC5195 (Fig. \[fig:pi15\]). The northern part of Arm 2 is weakly polarized and rather irregular in total and CO emission. The whole space between the northern Arm 2 and the companion galaxy is filled with highly polarized radio emission (typically 15% at [$\lambda6\,\mathrm{cm}$]{}). Arm 2 becomes well organized again at larger radii (located at the western edge of Fig. \[fig:pi15\]), where the total radio, polarized radio and CO emission perfectly coincide.
West of the central region, between Arms 1 and 2 in Fig. \[fig:inner:co\], another polarization feature emerges which appears similar to the magnetic arms observed e.g. in NGC6946 [@Beck96]. However, in contrast to NGC6946, Faraday rotation is not enhanced in the interarm feature of M51 (see Fig. \[fig:rm:15\]). Some peaks of polarized emission between Arms 1 and 2 in the south and southeast (see a low-resolution image of Fig. \[fig:pi15\]) and may indicate the outer extension of this magnetic arm. Inside of the inner corotation radius, located at 4.8 kpc [@Elmegreen89], this phenomenon can be explained by enhanced dynamo action in the interarm regions [@Moss98; @Shukurov98; @Rohde99].
Polarized radio emission from the inner arms and central region {#sec:inner}
---------------------------------------------------------------
{width="49.00000%"} {width="49.00000%"}
In the CO and H$\alpha$ line emissions (Fig. \[fig:inner:co\] and the red regions in Fig. \[fig:inner:8\]), the spiral arms continue towards the galaxy center. The high-resolution CO map by @Aalto99 shows that the arms are sharpest and brightest between about $25\arcsec$ and $50\arcsec$ distance from the center. The arms become significantly broader and less pronounced inside a radius of about 0.8 kpc: this is inside the inner Lindblad resonance of the inner density-wave system at $r\approx 1.3{\,\mathrm{kpc}}$ identified by @Elmegreen89.
The polarized emission at $4\arcsec$ resolution (see Fig \[fig:inner:4b\]) is also strongest along the inner arms 1–2 kpc distance from the center, with typically 20% polarization. The arm–interarm contrast is at least 4 in polarized intensity (this is a lower limit as the interarm polarized emission is below the noise level at this resolution and we take $\sigma_{PI}$ as an upper limit for the interarm value), larger than that of the outer arms, and is consistent with the expectations from compression of the magnetic field in the density-wave shock (Sect. \[sec:contrast\]). The contrast weakens significantly for $r<0.8 {\,\mathrm{kpc}}$. This may be an indication that the inner Lindblad resonance of the inner spiral density wave is at $r\simeq 0.8{\,\mathrm{kpc}}$ rather than $r\simeq 1.3{\,\mathrm{kpc}}$ (as located by @Elmegreen89): the shock is probably weak around the inner Lindblad resonance. In total intensity, the typical arm–interarm contrast for the region of the inner arms is about 5. The actual contrast in the M51 disc alone may be stronger than this if there is significant diffuse emission in the central region from a radio halo, but this effect is hard to estimate.
In the central region, two new features appear in polarized intensity which are the brightest in the entire galaxy (Fig. \[fig:inner:4b\]). The first is a region $11\arcsec$ north of the nucleus with a mean fractional polarization of 10% and an almost constant polarization angle. This feature coincides with the ring-like radio cloud observed in total intensity at [$\lambda6\,\mathrm{cm}$]{} and at $1\arcsec$ resolution by @Ford85 who also detected polarization in this region. The polarized emission indicates that the plasma cloud expands against an external medium and compresses the gas and magnetic field. The second feature of similar intensity in polarization is a ridge located along the eastern edge of the first region, extending east of the nuclear source, with 15% mean polarization and a magnetic field almost perfectly aligned along the ridge. Field compression is apparently also strong in this ridge. The nuclear source itself appears unpolarized at this resolution.
![\[fig:inner:4b\] [$\lambda6\,\mathrm{cm}$]{} total radio emission (VLA and Effelsberg combined) in the central $1.4\arcmin \times 1.4 \arcmin$ of M51 at $4\arcsec$ resolution. The $B$-vectors (not corrected for Faraday rotation) are shown where $PI\ge 3\sigma_{PI}$.](fig6){width="47.00000%"}
Spectral index and magnetic field strength {#sec:spx}
==========================================
{width="48.00000%"} {width="48.00000%"}
Spectral index maps of the total radio emission at $15\arcsec$ resolution are shown in Fig. \[fig:spx\]. The spectral indices are calculated using combined Effelsberg and VLA maps that contain signal on all scales down to the beamsize. There is therefore no missing flux in these maps due to the short-spacings problem of interferometers.
Although there is good general agreement between the two spectral index maps, the spectral index between [$\lambda\lambda20,3\,\mathrm{cm}$]{} is generally slightly flatter than that between [$\lambda\lambda20,6\,\mathrm{cm}$]{}; this is particularly noticeable in the spiral arms. This is due to the lower resolution of the Effelsberg [$\lambda6\,\mathrm{cm}$]{} map (180$\arcsec$ beam against $90\arcsec$ at [$\lambda3\,\mathrm{cm}$]{}): $180\arcsec$ is about the radius of the M51 disc. The coarse resolution made it more difficult to fine-tune the merging of the VLA and Effelsberg data in order to match the integrated fluxes in the single-dish and merged maps. This in turn led to a slight under-representation of the single-dish data at [$\lambda6\,\mathrm{cm}$]{} in the merged map and thus to a slightly steeper spectral index. We believe the [$\lambda\lambda20,3\,\mathrm{cm}$]{} spectral index, shown in Fig. \[fig:spx\]a, to be more reliable.
In both spectral index maps, one can clearly distinguish arms and interarm regions. The spectral index is typically in the range $-0.9\le\alpha\le-0.6$ in the spiral arms and $-1.2\le\alpha\le-0.9$ in the interarm zones. Since the spectral index in the arms is not characteristic of thermal emission ($\alpha\ne -0.1$) the arm emission must comprise a mixture of thermal and synchrotron radiation. The flatter arm spectral index can then be explained by two factors: stronger thermal emission in the arms due to recent star formation and the production of H[ii]{} regions and energy losses of cosmic ray electrons as they spread into the interarm from their acceleration sites in the arms.
We can use the observed steepening of the spectral index of the total radio emission $\alpha$ to estimate the diffusion coefficient of cosmic ray electrons in M51. We assume that the sources of the electrons are supernovae in the arms, that the initial spectral index is $\alpha_\mathrm{syn}\simeq -0.5$ and that the radio emission in the interarm region is predominantly synchrotron, so that $\alpha_\mathrm{syn}\simeq -1.1$ between the arms. A difference in the spectral indices of $\Delta\alpha= 0.5$ is expected if the main mechanisms of energy losses for the electrons are synchrotron emission and inverse Compton scattering [@Longair:1994].
In the next section we estimate magnetic field strengths of around $20{\,\mu\mathrm{G}}$ for the inter-arm regions. In these magnetic fields, an electron emitting at $5{\,\mathrm{GHz}}$ has an energy of about $4\,\mathrm{GeV}$. We can estimate the lifetime of cosmic ray electrons emitting the synchrotron radiation as $$\tau\simeq\frac{8.4\times 10^9}{\nu_\mathrm{16}^{1/2} {B_\mathrm{tot\perp}}^{3/2}}{\,\mathrm{yr}}\approx
8.2\times 10^6 {\,\mathrm{yr}},$$ [@Lang:1999 Sect. 1.25] where the frequency $\nu_\mathrm{16}$ is measured in units of $16{\, \mathrm{MHz}}$, the total magnetic field strength in the plane of the sky ${B_\mathrm{tot\perp}}$ is measured in ${\,\mu\mathrm{G}}$ and we have taken ${B_\mathrm{tot\perp}}\simeq 15{\,\mu\mathrm{G}}$ in the interarm regions. Taking $L=1{\,\mathrm{kpc}}$ as the typical distance a cosmic ray electron travels from its source in a supernova remnant to the inter-arm region, yields the diffusion coefficient $D$ of the electrons, $$D=\frac{L^2}{\tau}\simeq 4\times
10^{28}\,{\,\mathrm{cm}}^2{\,\mathrm{s}}^{-1},$$ which is compatible with the value of $D\simeq1$–$10\times10^{28}{\,\mathrm{cm}}^2{\,\mathrm{s}}^{-1}$ estimated by @Strong:1998 for the Milky Way. Note that cosmic ray electrons producing radio emission at cm wavelengths, propagating for a few ${\,\mathrm{kpc}}$ in ${\,\mu\mathrm{G}}$ strength magnetic fields, give diffusion coefficients in this range: our estimate for $D$ is not a unique property of the cosmic rays in M51.
Total magnetic field
--------------------
![\[fig:Bfield\] Total magnetic field strength, derived from the [$\lambda6\,\mathrm{cm}$]{} emission assuming equipartition between the energy densities of magnetic fields and cosmic rays (colour scale in ${\,\mu\mathrm{G}}$) along with contours of neutral gas density (a combination of CO [@Helfer03] and HI [@Rots:90] observations) plotted at $1,4,8,16,32,64\%$ of the maximum value.](fig8){width="45.00000%"}
In order to derive the strength and distribution of the total magnetic field, we have made a crude separation of the [$\lambda6\,\mathrm{cm}$]{} map at $15\arcsec$ into its nonthermal $I_\mathrm{syn}$ and thermal $I_\mathrm{th}$ components. We assumed that the thermal spectral index is everywhere $\alpha_\mathrm{th}=-0.1$ as expected at cm wavelengths [e.g. @Rohlfs:1999] and that the synchrotron spectral index is $\alpha_\mathrm{syn}=-1.1$ everywhere, as observed in the interarm regions (Fig. \[fig:spx\]a). We are constrained in our choice of $\alpha_\mathrm{syn}$ by two considerations: if $\alpha_\mathrm{syn}>-1.1$ then thermal emission is absent from the whole inter-arm region, whereas H-$\alpha$ emission is detected; if $\alpha_\mathrm{syn}<-1.1$ we find that the degree of polarization approaches is maximum theoretical value of $70\%$ in many regions, which is implausible for our resolution of $570{\,\mathrm{pc}}$. The average thermal emission fraction at [$\lambda6\,\mathrm{cm}$]{} is 25%.
Assuming equipartition between the energy densities of the magnetic field and cosmic rays, a proton-to-electron ratio of 100 and a pathlength through the synchrotron-emitting regions of 1 kpc, estimates for the total field strength are shown in Fig. \[fig:Bfield\], applying the revised formulae by @BeckKrause05.
The strongest total magnetic fields of about $30{\,\mu\mathrm{G}}$ are observed in the central region of M51. The main spiral arms host total fields of 20–$25{\,\mu\mathrm{G}}$, while the interarm regions still reveal total fields of 15–$20{\,\mu\mathrm{G}}$. This is significantly larger than in spiral galaxies like NGC6946 [@Beck07] and M33 [@Tabatabaei08]. These two galaxies have similar star-formation rates per unit area as M51, but weaker density waves, so that compression is probably higher in M51.
Ordered magnetic field {#subsec:Breg}
----------------------
The strength of the ordered magnetic field can be estimated from that of the total field using the degree of polarization. This method gives field strengths of $11$–$13{\,\mu\mathrm{G}}$ in the inner spiral arms, $8$–$10{\,\mu\mathrm{G}}$ in the outer spiral arms and $10$–$12{\,\mu\mathrm{G}}$ in the inter-arm regions. However, these values can only be attributed to a regular (or mean) magnetic field *if* the unresolved random component is purely isotropic [see @Sokoloff98 Sect. 5.1]. The observed maximum degree of polarization of around $40\%$ can equally be produced by an anisotropic random field whose degree of anisotropy is about 2: that is if the standard deviation of the fluctuations in one direction on the plane of the sky is twice as large as in the orthogonal direction. In Sect. \[sec:B\] we shall see that there is only a weak signature of a regular field in the observed multi-frequency polarization angles and that most of the polarized emission does indeed arise due to anisotropy in the random field.
The difference between the total and ordered magnetic field strengths gives an estimate for the isotropic random magnetic field strength of $18{\,\mu\mathrm{G}}$, or 1.5 times the ordered field, in the arms and $13{\,\mu\mathrm{G}}$, or 1.2 times the ordered field, in the inter-arms. If the main drivers of (isotropic) turbulence are supernova remnants, then the preferential clustering of Type II supernovae in the spiral arms is compatible with the higher fraction of isotropic random field in the arms. So a significant fraction of the magnetic field consists of a random component that is isotropic on scales less than $500{\,\mathrm{pc}}$.
Uncertainties
-------------
Our assumption of a synchrotron spectral index that is constant across the galaxy is crude and an oversimplification, even though the value used, $\alpha_\mathrm{syn}=-1.1$, can be somewhat constrained by other data as described above. We would expect that $\alpha_\mathrm{syn}$ should be closer to $-0.5$, the theoretical injection spectrum for electrons accelerated in supernova remnants, in parts of the spiral arms. This limitation results in an overestimate (underestimate) of the thermal (nonthermal) emission in the arms.
In principle we could combine the data at all three frequencies and simultaneously recover $I_\mathrm{syn}$, $I_\mathrm{th}$ and $\alpha_\mathrm{syn}$ at each pixel. However we defer a more robust calculation, interpretation and discussion to a later paper.
The equipartition estimate depends on the input parameters with a power of only $1/(3+\alpha_\mathrm{syn}) \approx0.24$, so that even large input errors hardly affect the results. Further errors are induced by the underestimate of the synchrotron emission in the spiral arms by the standard separation method (see above). In M33, @Tabatabaei:2007 found that the standard method undestimates the average nonthermal fraction by about 25%. In star-forming regions of the spiral arms, the nonthermal intensity can be a factor of two too small, which leads to an equipartition field strength which is 20% too low. In the same regions, the synchrotron spectral index is too steep by about 0.5 which overestimates the field strength by 15%. Interestingly, both effects almost cancel in M33.
The equipartition assumption itself is subject of debate. Equipartition between cosmic rays and magnetic fields likely does not hold on small spatial scales (e.g. smaller than the diffusion length of cosmic rays) and on small time scales (e.g. smaller than the diffusion time of cosmic rays).
Faraday rotation and depolarization
===================================
Faraday rotation {#sec:rm}
----------------
![\[fig:rm:15\] Rotation measures between [$\lambda\lambda3,6\,\mathrm{cm}$]{}, at $15\arcsec$ resolution, overlaid with contours of H$\alpha$ emission [@Greenawalt98] at the same resolution, plotted at $4, 8, 16, 32\%$ of the map maximum. Data were only used where the signal-to-noise ratio in polarized intensity exceeds three. ](fig9){width="45.00000%"}
![\[fig:rm:hist\] Distribution of rotation measures between [$\lambda\lambda3,6\,\mathrm{cm}$]{}, shown in Fig. \[fig:rm:15\]: note that the map is oversampled and so the histogram of pixel counts does not represent statistically independent data points. Data were only used where the signal-to-noise ratio in polarized intensity exceeds three. Solid line is the best-fitting Gaussian to the histogram.](fig10){width="45.00000%"}
The nonthermal radio emission from the arms has a relatively low degree of polarization (typically 25% at [$\lambda6\,\mathrm{cm}$]{} and $15\arcsec$ resolution) so that unresolved, tangled or turbulent, magnetic field dominates in the arms. In contrast, the interarm regions are up to $40\%$ polarized and host a significant fraction of magnetic fields with orientation ordered at large-scales. Whether these fields are coherent (regular) or incoherent (anisotropic turbulent) can be decided only with the help of Faraday rotation measures.
One might expect from the well-ordered, large-scale spiral patterns of the polarization vectors of Figs. \[fig:pi15\] and \[fig:inner:8\] that the regular magnetic field would produce an obvious pattern in Faraday rotation. (See the rotation measure map of M31 in Fig. 11 of @Berkhuijsen:2003 for an example of clear rotation measure signal arising from a well-ordered magnetic field.) However, the Faraday rotation measure map shown in Fig. \[fig:rm:15\] is dominated by strong fluctuations in rotation measure, with a magnitude of order $100{\,\mathrm{rad\,m^{-2}}}$. We apparently have a paradox: the orientation of the regular magnetic field follows a systematic spiral pattern on scales exceeding $1{\,\mathrm{kpc}}$ but it does not produce any obvious large scale pattern in ${\mathrm{RM}}$. Even the magnetic arms located between the CO arms do not immediately exhibit any large-scale non-zero ${\mathrm{RM}}$. If the ordered field seen in polarized intensity with an equipartition strength of about $10{\,\mu\mathrm{G}}$ (Sect. \[sec:spx\]) was fully regular, we would have $|{\mathrm{RM}}|\simeq 700{\,\mathrm{rad\,m^{-2}}}$ near the major axis of M51 with a systematic decrease moving away from this axis in azimuth, which is not observed (we adopt $\langle n_\mathrm{e}\rangle=0.1{\,\mathrm{cm^{-3}}}$, $h=400$ pc and an inclination of $20{^{\circ}}$ for this estimate, see Sect. \[sec:B:results\]). Note also that the overlaid H$\alpha$ contours in Fig. \[fig:rm:15\] are not generally coincident with regions of strong rotation measure.
Observational uncertainties in the measured Stokes parameters can be a source of the fluctuations in Fig. \[fig:rm:15\]. The uncertainty in ${\mathrm{RM}}$ between [$\lambda3\,\mathrm{cm}$]{} and [$\lambda6\,\mathrm{cm}$]{}, denoted here $\delta{\mathrm{RM}}$, depends on the signal-to-noise ratio in the polarized intensity, $\Sigma$, and the difference in the squares of the wavelengths $\delta\lambda^2$ as $$\delta{\mathrm{RM}}=\frac{1}{\sqrt{2}\,\Sigma\,\delta\lambda^2}\approx
\frac{280}{\Sigma} {\,\mathrm{rad\,m^{-2}}}.$$ Due to the steep spectral index of the synchrotron emission and the weak Faraday depolarization between [$\lambda3\,\mathrm{cm}$]{} and [$\lambda6\,\mathrm{cm}$]{}, $\Sigma$ is lower at [$\lambda3\,\mathrm{cm}$]{} and we use these values to estimate $\delta{\mathrm{RM}}$. The ${\mathrm{RM}}$ fluctuations from the noise in the observed polarization signal are $\pm 10{\,\mathrm{rad\,m^{-2}}}$ (for typical $\Sigma\gtrsim30$ in the central $r\lesssim 90\arcsec$ of Fig. \[fig:rm:15\]). ${\mathrm{RM}}$ maps at $8\arcsec$ resolution have lower $\Sigma$ and are dominated by noise fluctuations, although some strong rotation measures, above the noise level, are also present.
The structure function of the ${\mathrm{RM}}$ fluctuations is flat on scales up to $3\arcmin$, whereas ${\mathrm{RM}}$ fluctuations due to the Milky Way foreground in the direction of M51 in the model of @Sun:2009 have a slope of around $0.8$ (Reich, private communication). This is a strong indication that these fluctuations are mostly due to the magnetic field in M51.
In Fig. \[fig:rm:15\] the ${\mathrm{RM}}$ fluctuations with an amplitude exceeding $45{\,\mathrm{rad\,m^{-2}}}$, throughout the central $r\lesssim 90\arcsec$, and $55{\,\mathrm{rad\,m^{-2}}}$ along the outer spiral arms *cannot* be explained by the noise. There are therefore around ten patches where ${\mathrm{RM}}$ changes sign over a distance of $1$–$2{\,\mathrm{kpc}}$ due to intrinsic fluctuations of magnetic field.
The dispersion in ${\mathrm{RM}}$ is $15$–$20{\,\mathrm{rad\,m^{-2}}}$ measured in several regions in the inner spiral arms and $25$–$30{\,\mathrm{rad\,m^{-2}}}$ in the outer arms. After correction for the dispersion due to noise, the intrinsic dispersion is $11\pm3{\,\mathrm{rad\,m^{-2}}}$ in the inner arms and $19\pm5{\,\mathrm{rad\,m^{-2}}}$ in outer arms, hence constant within the errors. The distribution of rotation measures is shown in Fig. \[fig:rm:hist\], along with the best-fit Gaussian, which has a mean ${\mathrm{RM}}$ of $10\pm 1{\,\mathrm{rad\,m^{-2}}}$ and a dispersion of $28\pm 1{\,\mathrm{rad\,m^{-2}}}$.
Intrinsic fluctuations dominate the ${\mathrm{RM}}$ maps; even smoothed to linear scales of order $1{\,\mathrm{kpc}}$, no large-scale pattern in rotation measure is apparent, in contrast to the clear large-scale spiral structure in polarization angles. This result is quite surprising, as we would expect to see the components of the same field in the sky plane and along the line of sight in polarization angle and in Faraday rotation, respectively. As polarization angles are not sensitive to field reversals, the observation of ordered pattern in angles does not demonstrate the existence of a regular (coherent) field. The spiral field seen in polarization angle could be anisotropic with many small-scale reversals, e.g. produced by strong shearing gas motions and compression, and hence would not contribute to Faraday rotation. Alternatively, the field may have significant components perpendicular to the galaxy plane (due to loops, outflows etc.) which are mostly visible in Faraday rotation and hide the large-scale pattern. Such an underlying large-scale pattern indeed exists, as we discuss in the next section.
The close alignment of the observed field lines along the CO arms and the lack of enhanced Faraday rotation in the polarized ridges can be understood if the turbulent magnetic field is anisotropic. An anisotropic turbulent field can produce strong polarized emission, but not Faraday rotation. This picture is similar to that obtained for the effect of large-scale shocks on magnetic fields in the barred galaxies NGC 1097 and NGC 1365 [@Beck:2005] and will be investigated in detail in Sect. \[sec:contrast\].
### The size of turbulent cells
The ${\mathrm{RM}}$ dispersion $\sigma_\mathrm{RM,D}$ observed within a beam of a linear diameter $D$ is related to $\sigma_{\mathrm{RM}}$ (Eq. \[eq:sigRM\]) as $$\label{sRMD} \sigma_{{\mathrm{RM}},D}\simeq
N^{-1/2}\sigma_{\mathrm{RM}}=\sigma_{{\mathrm{RM}}}\frac{d}{D}\,,$$ where $N=(D/d)^2$ is the number of turbulent cells within the beam area, assumed to be large. We confirmed the approximate scaling of $\sigma_{{\mathrm{RM}},D}$ with $D^{-1}$ using ${\mathrm{RM}}$ maps smoothed from 8 to 12 where the noise fluctuations are not dominant. Combination with Eq. (\[eq:sigRM\]) allows us to estimate the least known quantity involved, the diameter of a turbulent cell (or twice the correlation scale of the turbulence): $$\begin{aligned}
d\!\!\!&\simeq&\!\!\!
\left[\frac{D\sigma_{{\mathrm{RM}},D}}{0.81\langle n_\mathrm{e}\rangle
B_\mathrm{r} (L)^{1/2}}\right]^{2/3}\\ \nonumber
&=&\!\!\! 50{\,\mathrm{pc}}\left(\frac{D}{600{\,\mathrm{pc}}}\right)^{2/3}
\left(\frac{\sigma_{{\mathrm{RM}},D}}{15{\,\mathrm{rad\,m^{-2}}}}\right)^{2/3}
\left(\frac{\langle n_\mathrm{e}\rangle}{0.1{\,\mathrm{cm}}^{-3}}\right)^{-2/3}\\ \nonumber
&&\mbox{}\times\left(\frac{B_\mathrm{r}}{20{\,\mu\mathrm{G}}}\right)^{-2/3}
\left(\frac{L}{1{\,\mathrm{kpc}}}\right)^{-1/3}.
\end{aligned}$$
Faraday depolarization {#sec:dp}
----------------------
Faraday depolarization ${\mathrm{DP}}$ gives important information about the density of ionized gas, the strength of the regular and turbulent field components, and the typical length scale (or integral scale) of turbulent magnetic fields. ${\mathrm{DP}}$ is usually defined as the ratio of the degrees of polarization of the synchrotron emission at two wavelengths. This requires subtraction of the thermal emission which is subject to major uncertainties (see Sect. \[sec:spx\]). Instead ${\mathrm{DP}}$ was computed, from the polarized intensities $P$, as ${\mathrm{DP}}=(PI_1/PI_2) \times (\nu_2/\nu_1)^{\alpha_\mathrm{syn}}$, where $\alpha_\mathrm{syn}=-1.1$ is the synchrotron spectral index, assumed to be constant across the galaxy. Variations in $\alpha_\mathrm{syn}$ affect ${\mathrm{DP}}$ less severely than errors in the estimate of the thermal fraction of the total radio emission. The ${\mathrm{DP}}$ maps derived for [$\lambda6\,\mathrm{cm}$]{} and [$\lambda3\,\mathrm{cm}$]{} and between [$\lambda20\,\mathrm{cm}$]{} and [$\lambda6\,\mathrm{cm}$]{} are shown in Fig. \[fig:dp\].
${\mathrm{DP}}(6{\,\mathrm{cm}}/3{\,\mathrm{cm}})$ (Fig. \[fig:dp\]a) is around unity (i.e. no Faraday depolarization) in most of the galaxy. Small patches with noticeable Faraday depolarization, where ${\mathrm{DP}}=0.6$–$0.7$ are generally found in the spiral arms. There is no systematic connection between the depolarization and the intensity of H$\alpha$ emission, indicating that variations in the thermal electron density are not the main source of the depolarization. The average value of ${\mathrm{DP}}(20{\,\mathrm{cm}}/6{\,\mathrm{cm}})$ (Fig. \[fig:dp\]b) is 0.28, smaller by a factor of about 3 than ${\mathrm{DP}}(6{\,\mathrm{cm}}/3{\,\mathrm{cm}})$. In the inner arms, ${\mathrm{DP}}(20{\,\mathrm{cm}}/6{\,\mathrm{cm}})$ is lower than 0.2. Only in the outer regions of the disc does ${\mathrm{DP}}(20{\,\mathrm{cm}}/6{\,\mathrm{cm}})$ increase to 0.5 or higher.
{width="48.00000%"} {width="46.50000%"}
Differential Faraday rotation within the emitting layer leads to depolarization which varies as a $\sin(x)/x$, with $x=2 \,|{\mathrm{RM}}_i| \,
\lambda^2$, where ${\mathrm{RM}}_i$ is the intrinsic Faraday rotation measure within the emitting layer [@Burn66; @Sokoloff98]. With typical values $|{\mathrm{RM}}_i| = 50{\,\mathrm{rad\,m^{-2}}}$ (Fig. \[fig:rm:15\]), we expect little depolarization (${\mathrm{DP}}\approx 0.98$) at [$\lambda6.2\,\mathrm{cm}$]{} and even less at [$\lambda3.6\,\mathrm{cm}$]{}. At [$\lambda20.5\,\mathrm{cm}$]{}, significant ${\mathrm{DP}}$ is expected for $|{\mathrm{RM}}_i| >
30{\,\mathrm{rad\,m^{-2}}}$. Furthermore, lines of zero polarization (‘canals’) are expected along level lines with $|{\mathrm{RM}}_i|= 37.5n{\,\mathrm{rad\,m^{-2}}}$ (with integer $n\neq0$) [@Shukurov:2003; @Fletcher:2008], but not a single ‘canal’ is observed in the [$\lambda20.5\,\mathrm{cm}$]{} polarized intensity map. This suggests that the average $|{\mathrm{RM}}_i|$ at [$\lambda20.5\,\mathrm{cm}$]{} is significantly smaller than $37.5{\,\mathrm{rad\,m^{-2}}}$. Since the depolarization *is* relatively strong, it must be due to a different mechanism, e.g., Faraday dispersion.
Internal Faraday dispersion by turbulence in the magneto-ionic interstellar medium is the probable source of strong depolarization at longer wavelengths, producing the degree of polarization given by [@Burn66; @Sokoloff98]: $$p = p_0 \frac{1-\exp(-2S)}{2S}\, ,$$ where $S=\sigma_{\mathrm{RM}}^2 \lambda^4$ and the maximum degree of polarization is $p_0\approx 0.7$. Here $\sigma_{\mathrm{RM}}$ is the dispersion of intrinsic rotation measure ${\mathrm{RM}}_i$ within the volume traced by the telescope beam, $$\sigma_{\mathrm{RM}}=0.81 \langle n_\mathrm{e}\rangle
B_\mathrm{r} (Ld)^{1/2}\,,
\label{eq:sigRM}$$ where $\langle n_\mathrm{e}\rangle$ is the average thermal electron density along the line of sight (in ${\,\mathrm{cm^{-3}}}$), $B_\mathrm{r}$ the strength of the component of the random field along the line of sight (in ${\,\mu\mathrm{G}}$), $L$ the total path-length through the ionized gas (in pc), $d$ the size (diameter) of a turbulent cell (in pc). Reasonable values of $\langle n_\mathrm{e}\rangle=0.1{\,\mathrm{cm^{-3}}}$, $L=800$ pc [estimated for $2.4<r<4.8{\,\mathrm{kpc}}$ by @Berkhuijsen:1997], $B_\mathrm{r}=20{\,\mu\mathrm{G}}$ (Sect. \[sec:spx\]), $d=50$ pc (see Sect. \[sec:rm\]) yield $\sigma_{\mathrm{RM}}\approx 300{\,\mathrm{rad\,m^{-2}}}$ resulting in $p/p_0\approx 0.002$, a very strong depolarizing effect. The resulting Faraday depolarization is ${\mathrm{DP}}=0.002$ between [$\lambda20.5\,\mathrm{cm}$]{} and [$\lambda6.2\,\mathrm{cm}$]{}, much smaller than the observed depolarization of ${\mathrm{DP}}=0.2$.
[$\lambda20\,\mathrm{cm}$]{} [$\lambda3\,\mathrm{cm}$]{} [$\lambda6\,\mathrm{cm}$]{} [$\lambda20\,\mathrm{cm}$]{} Fig. \[fig:20\] Two possibilities, that are not mutually exclusive, are that (i) the scale height of the [$\lambda20\,\mathrm{cm}$]{} synchrotron disc is greater than the scale height of the thermal electron disc, or (ii) the [$\lambda20\,\mathrm{cm}$]{} polarized emission is produced in a synchrotron halo. In either case, since Faraday rotation is observed at wavelengths around [$\lambda20\,\mathrm{cm}$]{} [@Horellou:1992; @Heald:2009], with about $1/5$ the amplitude as between [$\lambda3\,\mathrm{cm}$]{} and [$\lambda6\,\mathrm{cm}$]{} (see below), thermal electrons and a magnetic field are required in the halo.[^4]
In contrast, at [$\lambda3.6\,\mathrm{cm}$]{} $\sigma_{\mathrm{RM}}\approx 300{\,\mathrm{rad\,m^{-2}}}$ produces virtually no depolarization with $p/p_0\approx 0.85$, so we expect that the galaxy is transparent (or Faraday thin) at [$\lambda3.6\,\mathrm{cm}$]{}. At [$\lambda6.2\,\mathrm{cm}$]{} we have $p/p_0\approx 0.3$, moderate depolarization, with $DP=0.2$ between [$\lambda6.2\,\mathrm{cm}$]{} and [$\lambda3.6\,\mathrm{cm}$]{}. This rough estimate for $DP$ is the same order of magnitude as the $DP$ shown in Fig. \[fig:dp\](a), albeit about three times lower than the typical observed value of $DP\approx 0.9$, indicating that our value of $\sigma_{\mathrm{RM}}$ is a slight overestimate. (The uncertainty in the adopted values of $\langle n_\mathrm{e}\rangle$, $L$ and $B_\mathrm{r}$ can easily explain the discrepancy: for $\sigma_{\mathrm{RM}}\approx 200{\,\mathrm{rad\,m^{-2}}}$ we have $DP=0.6$, much closer to the observed values.) Since Faraday effects are so small at [$\lambda3.6\,\mathrm{cm}$]{}, Fig. \[fig:dp\](a) strongly indicates that the disc is also Faraday thin at [$\lambda6.2\,\mathrm{cm}$]{}.
Regular magnetic field structure {#sec:B}
================================
The method
----------
The map of Faraday rotation discussed in Sect. \[sec:rm\] clearly shows the effects of strong magnetic field fluctuations on scales of $300{\,\mathrm{pc}}$ to $\sim 1{\,\mathrm{kpc}}$; from this map one could expect the magnetic field to be rather chaotic and disordered. However, the observed polarization angles, shown in Figs. \[fig:i15\] and \[fig:pi15\], suggest an underlying spiral pattern to the magnetic field on scales $\gtrsim 1{\,\mathrm{kpc}}$ even when corrected for Faraday rotation (Fig. \[fig:inner:co\]). If the large-scale spiral pattern in the magnetic field is due to a regular field component, such as might be expected due to mean-field dynamo action, we can expect to find a signature of such a field in the Faraday rotation signal at relevant scales. If no such signal can be uncovered then we may conclude that the data do not support the presence of a mean field in M51 and that the spiral patterns seen in Figs. \[fig:i15\] and \[fig:pi15\] are being imprinted on a purely random magnetic field through e.g. large-scale compression in the gas flow.
![\[fig:rm:30\] Rotation measures between [$\lambda\lambda3,6\,\mathrm{cm}$]{}, at $30\arcsec$ resolution, overlaid with contours of mid-infrared 15$\,\mu$m emission [@Sauvage96] plotted at $1,2,4,8,16,32\%$ of the maximum value. Data were only used where the signal-to-noise ratio in polarized intensity exceeds five. ](fig12){width="45.00000%"}
\[fig:rm:30\] [$\lambda3\,\mathrm{cm}$]{} [$\lambda6\,\mathrm{cm}$]{} \[fig:rm:15\]). [$\lambda18\,\mathrm{cm}$]{} [$\lambda20\,\mathrm{cm}$]{} [@Horellou:1992 Fig. 9],
In order to look for the signature of a regular magnetic field in the multi-frequency polarization maps we applied a Fourier filter to the $15\arcsec$ resolution Stokes $Q$ and $U$ maps at [$\lambda\lambda3,6,20\,\mathrm{cm}$]{} to remove the signal on scales $\lesssim 30\arcsec \approx 1.1{\,\mathrm{kpc}}$. We also used the [$\lambda18\,\mathrm{cm}$]{} data of @Horellou:1992 which has a resolution of $43\arcsec$ resolution and was not filtered. Then maps of polarization angle were constructed at each wavelength. These were subsequently averaged in sectors with an opening angle of $20{^{\circ}}$ and radial ranges of $2.4-3.6{\,\mathrm{kpc}}$, $3.6-4.8{\,\mathrm{kpc}}$, $4.8-6.0{\,\mathrm{kpc}}$ and $6.0-7.2{\,\mathrm{kpc}}$ (see Fig. \[fig:Bmodel\]). We estimated that the minimum systematic errors in polarization angle arising from this method are about $4\degr$ at [$\lambda\lambda3,6\,\mathrm{cm}$]{} and $10\degr$ at [$\lambda\lambda18,20\,\mathrm{cm}$]{}, with the main source of error being uncertainties arising from Faraday rotation by the random magnetic field [see @Ruzmaikin:1990 Sect. 2]. We set these as minimum errors in the average polarization angles, otherwise using the standard deviation in the sector.
We applied a method that seeks to find statistically good fits to the polarization angles using a superposition of azimuthal magnetic field modes $\exp(im\phi)$ with integer $m$, where $\phi$ is the azimuthal angle in the galaxy’s plane measured anti-clockwise from the north end of the major axis. A three-dimensional model of the regular magnetic field is fitted to the observations of polarization angles at several wavelengths simultaneously. The polarization angle affected by Faraday rotation is given by $\psi=\psi_0+{\mathrm{RM}}\lambda^2+{\mathrm{RM_{fg}}}\lambda^2$, where the intrinsic angle of polarized emission is $\psi_0$, ${\mathrm{RM}}$ is the Faraday rotation caused by the magneto-ionic medium of M51 and ${\mathrm{RM_{fg}}}$ is foreground Faraday rotation arising in the Milky Way. The method of modelling is described in detail in @Berkhuijsen:1997 and @Fletcher:2004 and has been successfully applied to normal spiral galaxies [@Berkhuijsen:1997; @Fletcher:2004; @Tabatabaei08] and barred galaxies [@Moss:2001; @Beck:2005]. We have derived a new model of the regular magnetic field in M51 as our combined VLA + Effelsberg maps at [$\lambda\lambda3,6\,\mathrm{cm}$]{} are a significant improvement in both sensitivity and resolution over the maps used by @Berkhuijsen:1997.
The fitted parameters of the regular magnetic fields are given in Appendix \[app:Bmodel\]. These fit parameters can be used to reconstruct the global magnetic structure in M51. In order to obtain statistically good fits to the observed data we needed plane-parallel field components only (no vertical fields). No satisfactory fits could be found, using various combinations of the azimuthal modes m=0, 1, 2 for the horizontal field and m=0,1 for the vertical field, for the observations at all 4 wavelengths for a single layer. Therefore at least two separate regions of Faraday rotation are required. This is because the patterns of polarization angle and Faraday rotation at [$\lambda\lambda3,6\,\mathrm{cm}$]{} and [$\lambda\lambda18,20\,\mathrm{cm}$]{} are very different: at [$\lambda18\,\mathrm{cm}$]{} and [$\lambda20\,\mathrm{cm}$]{} the disc emission is heavily depolarized by Faraday dispersion (see Sect. \[sec:dp\]), so we only see polarized emission at these wavelengths from the top of the disc. A similar requirement for two Faraday rotating layers in M51 was also found by @Berkhuijsen:1997.
![Geometry of the disc and halo layers. The thermal disc and halo have thickness ${h_\mathrm{th}}$ and $Z$, and electron densities $n_{{_{\mathrm{e}}},\mathrm{d}}$ and $n_{{_{\mathrm{e}}},\mathrm{h}}$. The scale-height of the region emitting polarized synchrotron radiation at [$\lambda\lambda18,20\,\mathrm{cm}$]{} is $h_{20}$ and the *observed* [$\lambda20\,\mathrm{cm}$]{} polarized emission comes from the layer ${h_\mathrm{th}}<z<h_{20}$, due to Faraday depolarization. The regular magnetic field has two layers: the disc $B_\mathrm{d}$, extending to the same height as $h_{20}$, which is both an emitting and Faraday rotating layer, and the halo $B_\mathrm{h}$ where only Faraday rotation occurs.[]{data-label="fig:layers"}](fig13){width="49.00000%"}
To describe the two layers in the model, the Faraday rotation from M51 is split into two components, arising from a disc and halo, ${\mathrm{RM}}=\xi^{(D)}{\mathrm{RM}}^{(D)}+\xi^{(H)}{\mathrm{RM}}^{(H)}$, where $\xi^{(D)}$ and $
\xi^{(H)}$ are parameters that allow us to model how much of the disc and halo are visible in polarized emission at a given wavelength. We use this decomposition of the ${\mathrm{RM}}$ into disc and halo contributions to take into account the depolarization of the [$\lambda20\,\mathrm{cm}$]{} emission from the thermal disc discussed in Sect. \[sec:dp\] by setting $\xi^{(D)}=0$ and $\xi^{(H)}=1$ at [$\lambda\lambda18,20\,\mathrm{cm}$]{} and $\xi^{(D)}=1$ and $\xi^{(H)}=1$ at [$\lambda\lambda3,6\,\mathrm{cm}$]{}. In other words, at [$\lambda\lambda18,20\,\mathrm{cm}$]{} the polarized emission is produced in a thin layer that lies above the thermal disc (see Sect. \[sec:dp\]) and has the same regular magnetic field configuration that produces the [$\lambda\lambda3,6\,\mathrm{cm}$]{} polarized emission. The [$\lambda\lambda3,6\,\mathrm{cm}$]{} emission is Faraday rotated in the thermal disc and the halo whereas the [$\lambda\lambda18,20\,\mathrm{cm}$]{} emission is only Faraday rotated in the halo: see Fig. \[fig:layers\].
Results {#sec:B:results}
-------
{width="49.00000%"} {width="49.00000%"}
The resulting regular magnetic field structure is shown in Fig. \[fig:Bmodel\]. The regular field in the disc is best described by a superposition of m=0 and m=2 horizontal azimuthal modes and has a radial component directed outwards from the galaxy centre, whereas the halo field has a strong $m=1$ horizontal azimuthal mode and is directed inwards in the north, opposite to the direction of the disc field, and outwards in the south, same as the disc field. Oppositely directed components of the field in the disc and halo were also found by @Berkhuijsen:1997, however our new observations place the strong $m=1$ mode, and the resulting strong asymmetry in the magnetic field, in the halo rather than the disc. We believe that the difference in fitted fields arises from the higher quality of the new [$\lambda\lambda3,6\,\mathrm{cm}$]{} data, as discussed above. @Heald:2009 derived rotation measure maps from multi-channel WSRT data at [$\lambda22\,\mathrm{cm}$]{} using the RM-synthesis technique that qualitatively show the pattern expected from a $m=1$ magnetic field. Since depolarization due to Faraday dispersion cannot be removed by RM-synthesis, this gives a second, independent, indication that the halo of M51 hosts an $m=1$ regular magnetic field.
In the ring $2.4<r<3.6{\,\mathrm{kpc}}$ a weak $m=0$ mode is required to fit the data. This arises due to a sharp change in the observed polarization angles at [$\lambda\lambda18,20\,\mathrm{cm}$]{} between $\phi=120\degr$ and $\phi=140\degr$. Even with two halo modes we cannot capture the rapid change in angles and had to exclude the [$\lambda20\,\mathrm{cm}$]{} data at $\phi=140\degr$: the alternative would be to add an extra mode, with three new parameters, to model one data point.
The process by which two different regular magnetic field patterns in two layers of the same galaxy are produced is not clear and is beyond the scope of this paper. We only offer some speculative suggestions: (i) both fields might be generated by mean field dynamos but operating in different regimes, with the interaction of M51 with NGC 5195 driving a $m=1$ mode in the halo; (ii) the halo field could be a relic of the magnetic field present in the tenuous intergalactic medium from which the galaxy formed; (iii) as the disc field is advected into the halo it can be modified by the halo velocity field into the $m=1$ pattern. All of these possibilities will require careful modelling to determine their applicability to the problem.
In the disc of M51, in the four rings used in our model the $m=0$ azimuthal field component is $1$–$2$ times the strength of the $m=2$ mode (Table \[tab:fit\]). While the strength of the $m=2$ mode remains approximately constant between the rings, the $m=0$ mode is of similar strength in the inner ring, but is much stronger in the other three rings.
The r.m.s. regular field strength $\bar{B}$ in each ring can be determined by integrating the fitted modes over azimuth (Table \[tab:fit\]) via $$\bar{B} = \frac{\bar{R}}{88{\,\mathrm{rad\,m^{-2}}}}\left(\frac{{\langlen_\mathrm{e,d}\rangle}}{0.11{\,\mathrm{cm^{-3}}}}\right)^{-1}
\left(\frac{h_{\mathrm{th}}}{1{\,\mathrm{kpc}}}\right)^{-1}\, {\,\mu\mathrm{G}},$$ where $$\bar{R}=\frac{1}{2\pi}\int^{2\pi}_{0}\sqrt{R_r^2 + R_{\phi}^2}\,\mathrm{d}\phi$$ with $R_r$ and $R_{\phi}$ given by Eq. (\[eq:Bmod\]). @Berkhuijsen:1997 estimated that ${\langlen_\mathrm{e,d}\rangle}=0.11{\,\mathrm{cm^{-3}}}$ and ${h_\mathrm{th}}=400{\,\mathrm{pc}}$ in the radial range[^5] $2.4<r<4.8{\,\mathrm{kpc}}$ and for $4.8<r<7.2{\,\mathrm{kpc}}$ they estimated ${\langlen_\mathrm{e,d}\rangle}=0.06{\,\mathrm{cm^{-3}}}$ and ${h_\mathrm{th}}=600{\,\mathrm{pc}}$. The r.m.s. strengths of the large-scale magnetic field, using these parameters, are $\bar{B}=1.4\pm0.1{\,\mu\mathrm{G}}$, $\bar{B}=1.7\pm0.5{\,\mu\mathrm{G}}$, $\bar{B}=2.7\pm1.0{\,\mu\mathrm{G}}$ and $\bar{B}=2.8\pm0.1{\,\mu\mathrm{G}}$ in the rings $2.4<r<3.6{\,\mathrm{kpc}}$, $3.6<r<4.8{\,\mathrm{kpc}}$, $4.8<r<6.0{\,\mathrm{kpc}}$ and $6.0<r<7.2{\,\mathrm{kpc}}$ respectively. These are a factor of 4 lower than the strength of the ordered field derived from the equipartition assumption (Sect. \[subsec:Breg\]). The equipartition field strength is based on the observed polarized intensity and an anisotropic random magnetic field can contribute to the polarized signal (see Sect. \[subsec:Breg\]) but will not produce any systematic pattern in polarization angles at different frequencies. This is probably the reason for the discrepancy: most of the polarized radio emission in M51 does not trace a mean magnetic field, only the modelled large-scale pattern in Faraday rotation does.
The average pitch angle of the $m=0$ mode is $-20{^{\circ}}$ with little variation in radius between the rings. This means that the spiral structure of the regular field is coherent over the whole galaxy. The weaker $m=2$ modes produce an azimuthal variation in pitch angle of about $15\degr$ in the inner ring and $5\degr$ in the other rings. This variation of the pitch angle of the *regular* magnetic field with azimuth is much lower than the variation of the pitch angle of the *ordered* magnetic field observed in polarization: @Patrikeev06 showed that the orientation of Faraday rotation corrected polarization angles change by about $30\degr$. Thus the anisotropic random magnetic field, that we believe produces most of the polarized emission in M51, has a stronger azimuthal variation in its orientation: this is to be expected if the anisotropy arises from some periodic mechanism such as compression in spiral arms or localised enhanced shear.
The bisymmetric halo field has a much larger pitch angle of about $-50{^{\circ}}$ in the inner three rings. If this field is generated by a mean field dynamo the high pitch angles may be an indication that differential rotation in the halo is weak and an $\alpha^2$-dynamo action is significant. (In an $\alpha^2$-dynamo, shear due to differential rotation is negligible and so $|B_r|\approx|B_{\phi}|$.) The average thermal electron density and size of the halo in this galaxy are unknown and we have no specific constraints to apply. Taking reference values of ${\langlen_\mathrm{e,h}\rangle}=0.01{\,\mathrm{cm^{-3}}}$, $Z=5{\,\mathrm{kpc}}$ and ${\langlen_\mathrm{e,h}\rangle}=0.06{\,\mathrm{cm^{-3}}}$, $Z=3.3{\,\mathrm{kpc}}$ for the radial ranges $2.4\le r\le4.8{\,\mathrm{kpc}}$ and $4.8\le r\le7.2{\,\mathrm{kpc}}$ respectively, where the densities are one tenth of the disc density as in the Milky Way and $Z$ is as estimated by @Berkhuijsen:1997, the r.m.s. strength of the halo field (note that the $m=1$ mode means there are two values of the azimuth in each ring where the field is zero) is $$\bar{B}\approx \frac{\bar{R}_{h}}{8{\,\mathrm{rad\,m^{-2}}}}\left(\frac{{\langlen_\mathrm{e,h}\rangle}}{0.01{\,\mathrm{cm^{-3}}}}\right)^{-1}
\left(\frac{Z}{1{\,\mathrm{kpc}}}\right)^{-1}\, {\,\mu\mathrm{G}},$$ where $\bar{R_{h}}$ is the average amplitude of the halo mode. The fitted amplitudes of the halo field given in Table \[tab:fit\] give r.m.s. regular field strengths in the halo of $1.3\pm0.3$, $1.2\pm0.4$, $2.2\pm1.0$ and $1.6\pm0.6{\,\mu\mathrm{G}}$ in the four rings, with increasing radius.
We are confident that the azimuthal modes and pitch angles, fitted to our data, for the regular magnetic field in the disc and halo are robust as we have carried out extensive checks and searches of the parameter space. For example, if the [$\lambda\lambda18,20\,\mathrm{cm}$]{} data is ignored and a disc only model used the [$\lambda\lambda3,6\,\mathrm{cm}$]{} data produce a very similar fitted field to that given in Table \[tab:fit\]. However, we cannot consider the mode amplitudes to be reliable other than to reflect the relative strengths of the $m=0$ and $m=2$ modes in the disc.
A field reversal between the regular fields in the disc and the inner halo has also been suggested in the Milky Way [@Sun08], perhaps similar to the northern half of M51 (Fig. \[fig:Bmodel\]).
It is probable that our model for the vertical structure of M51 (Fig. \[fig:layers\]) is too simple and that this has lead to too much Faraday rotation being put into the halo field at the expense of the disc. In particular we do not allow for Faraday rotation from the emitting part of the [$\lambda20\,\mathrm{cm}$]{} disc nor for emission at any wavelength from the halo. Furthermore, each azimuthal mode is assumed to have an azimuth-independent intrinsic pitch angle. Adding more parameters to this model, given the limited number of sectors and frequencies that we can use, will not likely resolve this question. A more productive approach will be to develop a new model that also takes into account depolarizing effects directly and whose outputs are statistically compared directly to the individual maps, including the unpolarized emission (perhaps using the maps of Stokes parameters themselves).
Arm–interarm contrasts {#sec:contrast}
======================
---------------------------- ----------------------------------------- ------ ---------- -------------------- ------ ---------- --------------------
Quantity Units
Arm Interarm Arm/interarm ratio Arm Interarm Arm/interarm ratio
Neutral gas column density $10^{21}\mathrm{H}{\,\mathrm{cm}}^{-2}$ 18.0 4.0 4.5 200 40 5
Total radio intensity ${\,\mathrm{mJy/beam}}$ 1.1 0.5 2.2 0.6 0.12 5
Polarized radio intensity ${\,\mathrm{mJy/beam}}$ 0.1 0.1 1.0 0.07 $<0.01$ $\ge 7$
---------------------------- ----------------------------------------- ------ ---------- -------------------- ------ ---------- --------------------
\[tab:arminter\]
We wish to investigate the effect of magnetic field compression in the large-scale shocks associated with the spiral arms. In particular, how the regular and random magnetic field components are changed by the shocks and whether shock compression of isotropic random magnetic fields can produce enough anisotropic field to account for most of the polarized emission as inferred in Sections \[sec:dp\] and \[sec:B\]. We have carefully examined the azimuthal variations in the various maps at different radii: these emission profiles clearly demonstrate that there is not a simple azimuthal behaviour of any of the measured quantities. One cannot identify “typical” arm to interarm contrasts. So we have used a mask to separate arm (pixels in the mask set to $1$) and interarm (mask pixels set to $-1$) regions in each of the maps and hence calculate the average contrast over a wide radial range.
We combined the CO map of @Helfer03 with the HI map of @Rots:90 to produce a map of the total neutral gas density at $8\arcsec$ resolution, assuming a constant conversion factor $N_{\mathrm{H}_2}=1.9\times 10^{20} I_\mathrm{CO}\,{\,\mathrm{cm}}^{-2}$. The arm–interarm mask was determined by making a wavelet transform, using the Mexican hat wavelet with a linear scale of about $1{\,\mathrm{kpc}}$, of this map. This scale was selected by examining a range of transform maps: a $1{\,\mathrm{kpc}}$ scale wavelet produces transform coefficients that are continuously positive along spiral arms and continuously negative in interarm regions. Also, $1{\,\mathrm{kpc}}$ seems reasonable as a typical width of the spiral arms in M51. This mask was used to separate the arm and interarm components of each of the maps listed in Table \[tab:arminter\], over the radial range $1.6{\,\mathrm{kpc}}<r<4.8{\,\mathrm{kpc}}$, and then the average arm and interarm values were determined.
In addition to the data in Table \[tab:arminter\] we also separated the rotation measures shown in Fig. \[fig:rm:15\] into arm and interarm components. We calculated the average magnitude $\left<|{\mathrm{RM}}|\right>$ and standard deviation $\sigma_{\mathrm{RM}}$ of ${\mathrm{RM}}$ and found that $\left<|{\mathrm{RM}}|\right>\approx 22{\,\mathrm{rad\,m^{-2}}}$ and $\sigma_{\mathrm{RM}}\approx
39$ in the interarm and $\left<|{\mathrm{RM}}|\right>\approx 15{\,\mathrm{rad\,m^{-2}}}$ and $\sigma_{\mathrm{RM}}\approx 29$ in the arms. This may indicate that the regular magnetic field is stronger in the interarm regions than in the arms. However, the interpretation is difficult as the ${\mathrm{RM}}$ distribution depends on the signal-to-noise ratio, which tends to be higher in the arms.
The contrast in the neutral gas column density (2H$_2$ + HI) is compatible with what might be expected from compression by a strong adiabatic shock, $$\epsilon_n=n{^{\mathrm{(d)}}}/n{^{\mathrm{(u)}}}=4,$$ where the superscripts ${^{\mathrm{(d)}}}$ and ${^{\mathrm{(u)}}}$ refer to downstream and upstream of the shock front. We note that the scale height of the gas layer $h$ is not expected to be much affected by the spiral pattern [@Shukurov98]; H[i]{} observations in the Milky Way suggest $h{^{\mathrm{(d)}}}/h{^{\mathrm{(u)}}}\simeq1$–$1.5$ in the outer Milky Way [@K07].
General considerations
----------------------
Explaining the arm–interarm contrast in the observed radio intensity is a long-standing problem. @Mathewson72 argued that their [$\lambda20\,\mathrm{cm}$]{} observations of M51 are consistent with the density wave theory of @Roberts:1970. They assumed that both cosmic ray number density and the tangential magnetic field increase in proportion to the gas density at the spiral shock. The resulting arm–interam contrast in radio intensity, after taking into account of the telescope beamwidth, is expected to be of order $10$ or more.
@Tilanus:1988, using observations at a higher resolution, found that the shape of cross-sectional profiles across the radio intensity arms is not compatible with the density wave theory and concluded that the synchrotron emitting interstellar medium is *not* compressed by shocks and decouples from the molecular clouds as it traverses the arms. Thus there is clearly a discrepancy between the physically appealing theory of @Roberts:1970 and observations [see also @Condon:1992 p. 590].
@Mouschovias:1974 and @Mouschovias:2009 suggest that only a moderate increase in synchrotron emission in spiral arms is expected due to the Parker instability: rather than being strongly compressed the regular magnetic field rises out of the disc, in loops with a scale of 500–1000 pc. However, the substantial random component of the magnetic field in M51 may supress the instability or reduce it to a simple uniform buoyancy [@Kim:2001]. Furthermore we do not observe the periodic pattern of enhanced Faraday rotation along the spiral arms that would be expected from the vertical magnetic fields at the loop footpoints (Fig. \[fig:rm:15\]).
In this section we reconsider this question with additional emphasis on the polarized intensity. In Sect. \[subsec:cr\], we consider how shock compression affects the emission of cosmic-ray electrons at a *fixed* frequency. The effect of the modest observational resolution on the arm–interarm contrasts is discussed in Sect. \[subsec:beam\]. In Sect. \[subsec:comp\] we consider the case of compression of an isotropic random magnetic field and conclude that this may be the dominant origin of the arm–interam contrast in radio intensity only in the inner galaxy. Finally, in Sect. \[subsec:decomp\] we show how the decompression of an isotropic random magnetic field as it leaves the spiral arm affects the arm–interarm contrast.
Cosmic rays in compressed gas {#subsec:cr}
-----------------------------
Assuming that magnetic field is parallel to the shock front of the spiral density wave and is frozen into the gas, its strength increases in proportion to the gas density, $B\propto\rho$, as appropriate for one-dimensional compression. The ultrarelativistic gas of cosmic rays, whose speed of sound is $c/\sqrt3$, with $c$ the speed of light, is not compressed in the arms. However, compression of magnetic field will affect the cosmic rays (including the electron component) because $p_\perp^2/B\approx{\mathrm{const}}$ is an adiabatic invariant [@Rybicki:1979], where $p_\perp=\gamma m{_{\mathrm{e}}}c$ is the component of the particle momentum perpendicular to the magnetic field and $\gamma$ is the particle Lorentz factor. More precisely, only the part of the Lorentz factor related to the particle velocity perpendicular to the magnetic field should be included, but we ignore this detail for a rough estimate. In terms of the Larmor radius $r_B$, one can write $p_\perp=eBr_B/c$, with $e$ the electron charge, to obtain another form of the adiabatic invariant, $B r_B^2={\mathrm{const}}$, i.e.magnetic flux through the electron’s orbit remains constant. For $B\propto\rho$, we then obtain $r_B\propto\rho^{-1/2}$ and $p_\perp\propto\rho^{1/2}$, or $\gamma\propto\rho^{1/2}$.
Thus, compression of magnetic field leads to an increase in the Lorentz factor of the cosmic-ray electrons, $\gamma\propto\rho^{1/2}$. If the initial range of the Lorentz factors is $\gamma_\mathrm{min}{^{\mathrm{(u)}}}\leq\gamma\leq\gamma_\mathrm{max}{^{\mathrm{(u)}}}$, compression transforms it into $\gamma_\mathrm{min}{^{\mathrm{(d)}}}\leq\gamma\leq\gamma_\mathrm{max}{^{\mathrm{(d)}}}$ such that $\gamma_\mathrm{min}{^{\mathrm{(d)}}}/\gamma_\mathrm{min}{^{\mathrm{(u)}}}=(\rho{^{\mathrm{(d)}}}/\rho{^{\mathrm{(u)}}})^{
1/2}$. Of course, the total number density of cosmic-ray particles does not change because the cosmic-ray gas is not compressed. Adopting a power-law spectrum of cosmic-ray electrons, $$n_\gamma\, {\,\mathrm{d}}\gamma = K_\gamma \gamma^{-s}\, {\,\mathrm{d}}\gamma\;,$$ where $n_\gamma\,{\,\mathrm{d}}\gamma$ is the number of relativistic electrons per unit volume in the range $(\gamma,\gamma+{\,\mathrm{d}}\gamma)$, the total number density of the particles follows as $\int_{\gamma_\mathrm{min}}^{\gamma_\mathrm{max}}
K_\gamma\gamma^{-s}\,{\,\mathrm{d}}\gamma\simeq
\gamma_\mathrm{min}^{1-s}K_\gamma={\mathrm{const}}$, where we have assumed that the energy spectrum is broad enough to have $\gamma_\mathrm{max}\ll\gamma_\mathrm{min}$ and $s>1$. However, the energy of each cosmic-ray particle increases as $\rho^{1/2}$: the energy shifts along the energy (or $\gamma$) axis, and $$K_\gamma\propto\rho^{(s-1)/2}\;,$$ i.e., the number density of particles with given $\gamma$ increases with $\rho$.
Now we can estimate the effect of compression on the synchrotron emissivity observed at a fixed frequency, $\nu={\mathrm{const}}$ and fixed frequency interval ${\,\mathrm{d}}\nu={\mathrm{const}}$. Denoting $\epsilon_I(\nu)$ the arm–interarm contrast in $I(\nu)$ we have $$\epsilon_I(\nu)=\frac{I^{(d)}}{I^{(u)}}
\propto\rho^s\;,$$ since $I\propto K_\gamma B^{(s+1)/2}\propto \gamma_\mathrm{min}^{s-1} B^{(s+1)/2}$. Note that the Lorentz factor of the electrons which radiate at a fixed frequency $\nu\simeq\gamma^2 B={\mathrm{const}}$ reduces as $B$ increases, $\gamma(\nu)\propto B^{-1/2}$; this also leads to an increase in the number of cosmic-ray electrons radiating at a given frequency after compression since $n_{\gamma}\propto\gamma^{-s}$ and $s>0$.
With $s\simeq3$ and the arm-interarm density contrast of about four, the number of cosmic-ray electrons with a given $\gamma$ is proportional to $K_{\gamma}\propto\rho^{(s-1)/2}\propto\rho$, and the synchrotron intensity in the spiral arms would then be 50–100 times stronger than between the arms. There are other reasons to expect enhanced synchrotron emission in the arms as supernova remnants, sites of cosmic ray acceleration, are localized in the arms. However, Table \[con\] clearly shows that such an enormous contrast is not observed.
It is plausible that cosmic rays are rather uniformly distributed in galactic discs and only weakly perturbed by the spiral arms [Section 3.10 of @Ginzburg:1990]. During their lifetime within the galaxy, $\tau\simeq3\times10^7{\,\mathrm{yr}}$, the cosmic ray particles become well mixed over distances of order $(D\tau)^{1/2}\simeq2$–$3{\,\mathrm{kpc}}$, where $D\simeq4\times10^{28}{\,\mathrm{cm}}^2{\,\mathrm{s}}^{-1}$ is the cosmic ray diffusivity. This scale exceeds the width of spiral arms, so diffusion can significantly reduce the arm-interarm contrast, but it can hardly result in the almost uniform distribution of cosmic rays which would explain the low contrasts in Table \[tab:arminter\]. Anyway, even assuming that the cosmic ray intensity is the same within the ams and between them, the compression of magnetic field by a factor of four would result in an enhancement of the synchrotron emissivity by a factor 16 for $s=3$, and this is already larger than what is observed. Note that although the contrast in total radio emission given in Table \[tab:arminter\] includes the thermal radio emission, this is concentrated in the arms as can be seen in H$\alpha$ images [@Greenawalt98] and so the actual contrast in synchrotron emissivity will be lower than the contrast in total radio emissivity.
Telescope resolution and the width of the compressed region {#subsec:beam}
-----------------------------------------------------------
One factor that can help to explain the lower than expected arm-interam contrast in total radio emission (which for simplicity we assume to be all synchrotron emission) is the possibility that the compressed region is narrow compared to the width of the spiral arms. Given the complex processes that take place as gas passes through the arms, such as the formation of molecular clouds and star formation, it is plausible that shock compression is followed by a decompression before the gas leaves the arms. Then, if the compressed region is narrower than the telescope beam the observed contrast between the arm and interarm will be reduced. We can estimate the width of the compressed region that is compatible with the observations in Table \[con\] using $$\epsilon_I=\epsilon_I^0\frac{w}{D},$$ where $\epsilon_I\simeq 5$ is the observed arm-interam contrast in total radio emission in the inner spiral arms (which we assume to be all synchrotron emission to make a conservative estimate), $\epsilon_I^0\simeq 50$ is the expected contrast due to compressive effects in an adiabatic shock, $D\simeq
150{\,\mathrm{pc}}$ is the beamwidth at our highest resolution and $w$ is the width of the compressed region. Using the values of the parameters just quoted we obtain $w\simeq 15{\,\mathrm{pc}}$.
@Patrikeev06 showed that the ridge of strongest polarized emission (tracing the peak of the compressed magnetic field) is generally shifted upstream of the ridge of strongest CO emission (tracing the highest neutral gas density) in M51. We can assume that this shift is due to a spiral-shock triggering the formation of molecular clouds. The dense clouds will fill only a small fraction of the volume occupied by the ISM and may become decoupled from the magnetic field threading the diffuse ISM during their formation, as originally suggested in the case of M51 by @Tilanus:1988, by @Beck:2005 for barred galaxies and with a plausible mechanism for the separation outlined in @Fletcher:2009. We expect that the expansion of the compressed magnetic field will begin once the clouds have formed. Thus after a time $\tau_c\sim 10^6{\,\mathrm{yr}}$ the ridge in strong radio emission will begin to decay. If we estimate the magnetic field strength to be $B\simeq 20{\,\mu\mathrm{G}}$ (Sect. \[sec:spx\]) then an Alfvén wave will propagate over the compressed distance $w=15{\,\mathrm{pc}}$ if the density, of the compressed diffuse gas in which the clouds are embedded, is $n\simeq 8{\,\mathrm{cm^{-3}}}$. This density is not implausible if the upstream diffuse gas density is $n\simeq 2{\,\mathrm{cm^{-3}}}$.
So one possible explanation for the observed low arm-interam contrast in total radio emission, compared to that expected from a simple consideration of cosmic ray energies, is that the compressed region is narrower than the beam. In this case we estimate that the ridge in compressed magnetic field, along the upstream edge of the spiral arm, will be a few tens of parsecs wide. This possibility can be tested using higher resolution observations. The cosmic rays are not the only source of the arm-interam contrast in radio emission though: next we consider the effect of a large-scale shock on the regular and random components of the magnetic field.
Compression of a partially ordered magnetic field {#subsec:comp}
-------------------------------------------------
We consider compression of a partially ordered magnetic field by the spiral shock. We assume that the random magnetic field upstream of the shock is statistically isotropic but one-dimensional compression makes it anisotropic and so it contributes to the polarized radio emission but not to Faraday rotation. We assume that both the spiral arms and the field lines of the large-scale magnetic field are logarithmic spirals with the pitch angles $p{_{\mathrm{a}}}$ and $p_b$, respectively. An acceptable estimate is $p{_{\mathrm{a}}}\approx p_b\approx-20^\circ$, where negative values correspond to a trailing spiral.
The formulae required to calculate the compression of a magnetic field with both regular and isotropic random components and the associated total and polarized synchrotron emission and Faraday rotation are derived in Appendix \[CPOMF\]. Using these equations we now consider two cases that encompass the range of observed contrasts in polarized emission in M51: the inner spiral arms, where there is a strong arm–interarm contrast in polarization of at least $\epsilon_{PI}\sim 4$, and the average contrast of $\epsilon_{PI}\sim 1$ which is more representative of the greater radii. The upstream ratio of random to regular magnetic field strength $b/B$ and the increase in gas density $\epsilon_n=4$, in regular field $\epsilon_B$ and in random field $\epsilon_b$ fields due to the shock are fixed and then the consequential increases in nonthermal $\epsilon_I$ and polarized $\epsilon_{PI}$ emission across the shock are calculated.
Firstly the inner arms. Here we can obtain a reasonable match with the maximum observed arm–interarm contrasts $\epsilon_I\simeq 5$ (assumed to be mostly synchrotron emission) and $\epsilon_{PI}\simeq 7$ (a lower limit due to weak interarm intensity at $4\arcsec$ resolution). With the parameters $\epsilon_B=1$, i.e. the regular field is not increased by the shock — this can be justified, for example, if the regular magnetic field becomes detached from dense clouds as they form [@Beck:2005] — and $\epsilon_b=2.7$ we obtain expected arm–interarm contrasts in synchrotron emission of $3.3$ and polarized emission of $4.8$. If $\epsilon_B>1$ then the expected increase in polarization becomes much larger. For example, for $\epsilon_B=\epsilon_b=2.7$ we obtain $\epsilon_{PI}=11$: since we only have a lower limit on the observed $\epsilon_{PI}$ we cannot rule out this possibility. We conclude that in the innermost spiral arms ($r<1.6{\,\mathrm{kpc}}$) anisotropic random magnetic fields produced by compression of the interstellar gas in the spiral arms can account for the observed increases in total and polarized radio emission.
Now for the average arm–interarm contrasts. Here, since the average contrast in polarization is $\epsilon_{PI}\simeq 1$, this requires no increase in regular magnetic field in the arm and also no anisotropy of the random magnetic field, but simultaneously we require a small increase in synchrotron emission $\epsilon_I=2.2$. This is not possible in our model; it is also unlikely to happen when galaxies contain strong spiral density waves. The closest we can come is to set $\epsilon_B=1$ coupled with a weak upstream random field $b^2/B^2=0.2$ and strong compression of the random field $\epsilon_b=4.5$ to produce the required contrast in total emission. Then we can obtain arm–interarm contrasts of $2.3$ in total synchrotron emission and $1.7$ in polarization. However, $b^2/B^2=0.2$ is strongly contradicted by the equipartition estimates of the various field strengths in Sect. \[sec:spx\], which likely overestimate the strength of the regular field as discussed in Sect. \[sec:dp\] and \[sec:B\].
To summarise this subsection: enhancement of the regular and random magnetic field components parallel to a large-scale spiral shock can *partly* account for the observed arm-interam contrasts in radio emission but no single set of parameters is compatible with the full range of the observations.
Decompression of an isotropic magnetic field {#subsec:decomp}
--------------------------------------------
Finally, we consider the *decompression* of an isotropic field as the magnetized gas flows from a high density to a low density region. This decompression also leads to the generation of an anisotropic random magnetic field, as the field component parallel to lines of constant density increases while the perpendicular component is unaffected, similar to the compressive case. But decompression will also help to lower the arm-interam contrast in nonthermal emission, thus alleviating one of the difficulties encountered in the previous subsection, if the random field in the arms is (partly) isotropized: this can readily occur due to turbulence being driven by star formation activity and the expansion of supernova remnants.
First consider the case of straightforward compression. Let us define the $x$-direction as perpendicular to the shock and the $y$-direction as parallel: so only $y$-components of the magnetic field are affected by the shock. Thus $${\langleb_x^{2(d)}\rangle}={\langleb_x^{2(u)}\rangle}=\frac{1}{3}b^{2(u)},$$ and $${\langleb_y^{2(d)}\rangle}=\epsilon_{\rho}^2{\langleb_y^{2(u)}\rangle}=\frac{1}{3}\epsilon_{\rho}^2 b^{2(u)},$$ where $(u)$ and $(d)$ refer to upstream and downstream and $\vec{b}$ is the total random field. Now the plane-of-sky component of the random field is $b_{\perp}^2=b_x^2+b_y^2$ so we obtain $$\frac{b_{\perp}^2(d)}{b_{\perp}^2(u)}=\frac{1}{2}(1+\epsilon_{\rho}^2)\simeq 8,$$ for an adiabatic shock, where $\epsilon_{\rho}=4$. So we expect adiabatic compression in a simple, plane parallel shock to produce a contrast in nonthermal emission $\epsilon_I\simeq 8$, in the case that the cosmic rays are smoothly distributed. This is a simplified version of part of the calculation described above and in detail in Appendix \[CPOMF\].
Now consider the case of decompression. We follow a similar calculation but here $${\langleb_y^{2(d)}\rangle}=\frac{1}{\epsilon_{\rho}^2}{\langleb_y^{2(u)}\rangle}=\frac{1}{3\epsilon_{\rho}^2} b^{2(u)},$$ which leads to $$\frac{b_{\perp}^2(d)}{b_{\perp}^2(u)}=\frac{1}{2}(1+\frac{1}{\epsilon_{\rho}^2})\simeq \frac{1}{2}.$$ In this case the arm is the upstream region so the expected arm-interam contrast in nonthermal emission is $\epsilon_I\simeq 2$ which is close to the average contrast observed (Table \[con\]).
Summary: compression and arm-interarm contrasts
-----------------------------------------------
In Sect. \[sec:dp\] and \[sec:B\] we have shown that much of the polarized emission detected across the disc of M51 must come from anisotropic random fields. Combined with the problem described in Sect. \[subsec:comp\] above, of how isotropic random fields can be compressed in a spiral shock but not produce an increase in polarized emission, we are led to the view that anisotropic random fields are already present in interarm regions, perhaps as a result of enhanced localised shear or decompression. @Patrikeev06 showed that the orientation of the magnetic field in M51 varies in azimuth by $\pm
15\degr$ and in the interarm generally has a different pitch angle to the CO spiral arm, only becoming well aligned with the CO arm at its location. In this case compression of the already anisotropic field in the spiral shock will only weakly amplify the random field and hence lead to a weak change in polarized emission.
We conclude that the roughly constant average polarized emission across the arms and interam region cannot be easily explained with simple models of shock compression of magnetic field, if one simultaneously considers the weak contrasts that are observed in the total emission. The arm-interarm contrast in the radio emission probably results from a complex interplay of compression *and* decompression of the dominant random field, occurring as the ISM undergoes phase changes on its passage through the arms. In addition the thickness of the compressed regions compared to the width of the beam likely plays a role.
Summary and Conclusions
=======================
- Polarized emission ($PI$) is present throughout most of the disc of M51. In some regions the strongest $PI$ coincides with the location of the strong spiral arms as seen in CO emission. In other regions $PI$ is concentrated in the interarm region, forming structures up to $\sim 5{\,\mathrm{kpc}}$ in size, reminiscent of the $\sim10{\,\mathrm{kpc}}$-scale magnetic arms observed in NGC 6946 [@Beck96]. The origin of these magnetic arms is still unknown.
- The observed polarization angles trace spiral patterns with pitch angles similar to, but not always the same as, the gaseous spiral arms. The apparently ordered large-scale magnetic field responsible for the well aligned polarization angles does not produce a systematic pattern in Faraday rotation, leading us to conclude that a large fraction of the polarized emission is caused by anisotropic small-scale magnetic fields (where small-scale refers to sizes smaller than the beam, typically $300$–$600{\,\mathrm{pc}}$): anisotropic random fields, whose anisotropy is caused by a large scale process and so is aligned over large distances, can produce well ordered polarization angles but a random Faraday rotation distribution.
- Faraday depolarization, caused by Faraday dispersion due to turbulent magnetic fields, leads to strong depolarization of the [$\lambda20\,\mathrm{cm}$]{} polarized emission from the disc. From the observed fluctuations in Faraday depolarization we were able to estimate a typical of a turbulent cell as $\sim 50{\,\mathrm{pc}}$.
- Fourier filtering followed by averaging in sectors is necessary to reveal the contribution of the regular (or mean) magnetic field to the observed polarization signal. This allowed us to fit a model of the 3D regular magnetic field of M51 to the observations of polarization angles at [$\lambda\lambda3,6,18,20\,\mathrm{cm}$]{}. Due to the strong depolarization at [$\lambda\lambda18,20\,\mathrm{cm}$]{} we were able to identify two different regular magnetic field patterns. In the thermal disc the regular field can be described as a combination of $m=0+2$ azimuthal modes, with the $m=0$ mode being the strongest: this combination can be the result of the strong two-armed spiral pattern modifying a dynamo generated $m=0$ mode (the easiest to excite according to mean-field galactic dynamo theory). The pitch angle of the $m=0$ mode is similar at all radii. In the halo the observed polarization angles at [$\lambda\lambda18,20\,\mathrm{cm}$]{}, whose emission from the thermal disc is heavily depolarized, reveal a $m=1$ azimuthal geometry for the regular magnetic field. The origin of this halo field is unclear.
- The arm–interarm contrast in gas density and radio emission was compared to a model where a regular and (isotropic) random magnetic field is compressed by shocks along the spiral arms. We found that where a strong arm–interarm contrast in $PI$ is observed, in the inner arms $r\lesssim 1.6{\,\mathrm{kpc}}$, amplification of the random magnetic field by compression can successfully explain the data, provided that the regular magnetic field is not significantly increased. This constraint is similar to that obtained for two barred galaxies in @Beck:2005, where it was proposed that the regular magnetic field de-couples from molecular clouds as they rotate and collapse. We were unable to explain the average arm–interarm contrast in total and polarized radio emission, typical for much of the galaxy at $r>2{\,\mathrm{kpc}}$, by a model involving shock compression of magnetic fields. Even when the regular magnetic field is not compressed some increase in $PI$ in the arms is expected from compression of the random field, whereas the average arm–interarm contrast in $PI$ is about one: this problem could be resolved if the random field is isotropic in the arms and becomes anisotropic due to decompression as it enters the interam. Alternatively, the compressed region of magnetic field may be sufficiently narrow (with a width of about $20{\,\mathrm{pc}}$), as might occur if the molecular clouds are de-coupled from the synchrotron emitting gas, to reduce the arm–interarm contrast by the required degree.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank W. Reich for suggesting the comparison of RM structure functions of our maps and a Milky Way model and for kindly making the calculations. We thank K. Ferrière for helpful discussions and M. Krause for a careful reading of the manuscript. A.F. and A.S. gratefully acknowledge financial support under the Leverhulme Trust grant F/00 125/N and the STFC grant ST/F003080/1. This research made use of NASA’s Extragalactic database (NED) and Astrophysical Data System (ADS).
Parameters of the fitted regular magnetic fields {#app:Bmodel}
================================================
In Table \[tab:fit\] we give the parameters of the fitted regular magnetic field models discussed in Section \[sec:B\]. Although a component of the regular field perpendicular to the disc plane ($R_z$ in Eq. (\[eq:Bmod\])) is allowed in the model and we searched for fits using this component, a vertical field was not required to obtain a good fit in either ring. The greater of the standard deviation and the noise within a sector was taken as the error in polarization angle.
The regular magnetic field is modelled as $$\begin{aligned}
R_r & = & R_0\sin p_0 + R_1\sin p_1\cos(\phi-\beta_1)\nonumber \\
& & +\, R_2\sin p_2\cos(2\phi-\beta_2),\nonumber \\
R_{\phi} & = & R_0\cos p_0 + R_1\cos p_1\cos(\phi-\beta_1) \nonumber\\
& & +\, R_2\cos p_2\cos(2\phi-\beta_2),\label{eq:Bmod}\\
R_z & = & R_{z0} + R_{z1}\cos(\phi-\beta_{z1}) + R_{z2} \cos(2\phi-\beta_{z2}),\nonumber\\
R_{hr} & = & R_{h0}\sin p_{h0} + R_{h1}\sin p_{h1}\cos(\phi-\beta_{h1}),\nonumber \\
R_{h\phi} & = & R_{h0}\cos p_{h0} + R_{h1}\cos p_{h1}\cos(\phi-\beta_{h1}),\nonumber
\end{aligned}$$ where $R_i$ is the amplitude of the $i$’th mode in units of ${\,\mathrm{rad\,m^{-2}}}$, $p_i$ is its pitch angle, $\beta_i$ ($i\ge1$) determines the azimuth where the corresponding non-axisymmetric mode is maximum and the subscript $_{h}$ denotes the components of the halo field. The magnetic field in each mode of this model is approximated by a logarithmic spiral, $p_i=\mbox{const}$, within a given ring. However, the superposition of such modes with different pitch angles leads to deviations from a logarithmic spiral. For further details of the method, see @Berkhuijsen:1997 and @Fletcher:2004.
The foreground Faraday rotation due to the magnetic field of the Milky Way $R_\mathrm{fg}$ was also included in the fitting; we expect $R_\mathrm{fg}$ to be the same in all rings and this provides a useful consistency check to the results of independent fits to the four rings. A logarithmic spiral pattern has the same pitch angle in all rings. $S$ is the residual of the fit $S_\lambda$ The appropriate $\chi^2$ value, at the $95$% confidence level, is shown for the number of fit parameters and data points. A fit is statistically acceptable if $S\leq\chi^2$ and the Fisher criteria, that tests if $S$ is unduly influenced by a good fit to a single wavelength, is satisfied. The $\chi^2$ values vary from ring to ring (even when the same number of fit parameters are used) as some sectors are excluded from the model, either because the noise in the sector exceeds the standard deviation of the measurements or because the sector represents an outlier from the global pattern.
In Figs. \[fig:ring1\], \[fig:ring2\], \[fig:ring3\] and \[fig:ring4\] we show the observed sector-averaged polarization angles and the fitted model for each ring. The fit quality is excellent at [$\lambda3\,\mathrm{cm}$]{} and [$\lambda6\,\mathrm{cm}$]{} for all rings, but in the inner two rings sharp discontinuities in the [$\lambda20\,\mathrm{cm}$]{} polarization angles around $\theta\simeq 120\degr$ cannot be accommodated by the model. Since the parameters of the fitted halo field are largely determined by the data at the longer wavelengths we are therefore less satisfied with the parameters of the fitted halo field in these rings than with those of the disc field, as we discuss in Sect. \[sec:B\].
[llllll]{} & & $2.4<r<3.6{\,\mathrm{kpc}}$ & $3.6<r<4.8{\,\mathrm{kpc}}$ & $4.8<r<6.0{\,\mathrm{kpc}}$ & $6.0<r<7.2{\,\mathrm{kpc}}$\
$R_\mathrm{fg}$ & ${\,\mathrm{rad\,m^{-2}}}$ & $4$ & $5$ & $2$ & $4$\
\
$R_0$ & ${\,\mathrm{rad\,m^{-2}}}$ & $-46$ & $-57$ & $-76$ & $-76$\
$p_0$ & deg & $-20$ & $-24$ & $-22$ & $-18$\
$R_2$ & ${\,\mathrm{rad\,m^{-2}}}$ & $-33$ & $-25$ & $-40$ & $-44$\
$p_2$ & deg & $-12$ & $16$ & $8$ & $3$\
$\beta_2$ & deg & $-8$ & $-6$ & $-14$ & $-25$\
\
$R_{h0}$ & ${\,\mathrm{rad\,m^{-2}}}$ & $23$ & & &\
$p_{h0}$ & deg & $-43$ & & &\
$R_{h1}$ & ${\,\mathrm{rad\,m^{-2}}}$ & $76$ & $77$ & $57$ & $43$\
$p_{h1}$ & deg & $-45$ & $-49$ & $-50$ & $-90$\
$\beta_{h1}$ & deg & $44$ & $30$ & $-3$ & $-16$\
$S_{\lambda}$ & ([$\lambda\lambda3,6,18,20\,\mathrm{cm}$]{}) & $(14, 21, 21, 21)$ & $(9, 9, 20, 28)$ & $(10, 18, 11, 16)$ & $(10, 26, 21, 21)$\
$S$&& $77$ & $66$ & $55$ & $78$\
$\chi^2$ && $79$ & $83$ & $83$ & $79$\
One data point excluded in the ring $2.4<r<3.6{\,\mathrm{kpc}}$ at: [$\lambda20\,\mathrm{cm}$]{}, $\phi=140\degr$.\
Three data points excluded in the ring $6.0<r<7.2{\,\mathrm{kpc}}$ at: [$\lambda3\,\mathrm{cm}$]{}, $\phi=20\degr$; [$\lambda18\,\mathrm{cm}$]{}, $\phi=220\degr$; [$\lambda20\,\mathrm{cm}$]{}, $\phi=180\degr$.
![Polarization angles ($\psi$, measured from the local radial direction in the plane of M51) against azimuth in the galaxy plane ($\theta$) for the ring 2.4–3.6${\,\mathrm{kpc}}$. Fit (solid line) and observations (squares with error bars, horizontal lines with error bars show data points excluded from the fit) are shown for [$\lambda3\,\mathrm{cm}$]{}, [$\lambda6\,\mathrm{cm}$]{}, [$\lambda18\,\mathrm{cm}$]{} and [$\lambda20\,\mathrm{cm}$]{} from top to bottom. The error bars show the $1\sigma$ deviations.[]{data-label="fig:ring1"}](figA1){width="49.00000%"}
![As in Fig. \[fig:ring1\] but for the ring 3.6–4.8${\,\mathrm{kpc}}$.[]{data-label="fig:ring2"}](figA2){width="49.00000%"}
![As in Fig. \[fig:ring1\] but for the ring 4.8–6.0${\,\mathrm{kpc}}$.[]{data-label="fig:ring3"}](figA3){width="49.00000%"}
![As in Fig. \[fig:ring1\] but for the ring 6.0–7.2${\,\mathrm{kpc}}$.[]{data-label="fig:ring4"}](figA4){width="49.00000%"}
Compression of a partially ordered magnetic field in a spiral arm {#CPOMF}
=================================================================
We introduce a Cartesian frame in the sky plane $(\tilde x,\tilde y,\tilde z)$ centered at the galaxy center with the $\tilde x$-axis directed towards the western end of the major axis and the $\tilde y$-axis, in the northern direction; the $\tilde z$-axis is the directed towards the observer (and in the general direction of the galaxy’s north pole). We also introduce galaxy’s Cartesian frame $(x,y,z)$ where the $x$- and $\tilde
x$-axes coincide and the $z$-axis is also directed towards the galaxy’s north pole. Magnetic field components in the two frames are related by [@Berkhuijsen:1997]: $$\begin{aligned}
B_{\tilde x}&=&B_x,\nonumber\\
B_{\tilde y}&=&B_y\cos i+B_z\sin i,\label{trans}\\
B_{\tilde z}&=&-B_y\sin i+B_z\cos i,\nonumber\end{aligned}$$ where $i$ is the galaxy’s inclination angle, and we include $B_z$ for the future convenience (we recall that we can adopt $B_z=0$ in M51). The galaxy’s cylindrical frame $r,\phi,z$ then has the azimuthal angle measured counterclockwise (along the galaxy’s rotation) from the $x$-axis. And finally, we introduce a local Cartesian frame of the spiral shock $(x',y',z')$, with the $x'$-axis perpendicular to the shock and directed from the interarm region into the spiral arm, the $y'$-axis parallel to the shock, so that the $z'$-axis complements them to a right-handed triad i.e., is directed towards the north pole of the galaxy. Then the angle between the $x'$- and $x$-axes is $$\label{theta}
\theta=\phi-p{_{\mathrm{a}}}.$$
It is convenient to specify the upstream large-scale magnetic field in the galaxy’s cylindrical frame $\vec{{B}}=B(\sin p_B,\cos p_B,0)$, where we neglect the vertical field component (see Sect. \[sec:B\]). Then the unit normal to the shock, in the galaxy’s frame is given by $\widehat{\vec{n}}=\widehat{\vec{x}'}=(\cos\theta,\sin\theta,0)$, the tangent vector is given by $\widehat{\vec{t}}=\widehat{\vec{y}'}=(-\sin\theta,\cos\theta,0)$, and the regular magnetic field components normal and tangent to the shock then follow, respectively, as $$B_{x'}=B\sin(p_b-p{_{\mathrm{a}}}),\quad B_{y'}=B\cos(p_b-p{_{\mathrm{a}}}).$$
It is now easy to see that the compressed large-scale magnetic field in the spiral arm has the components $$B{^{\mathrm{(d)}}}_{x'}=B{^{\mathrm{(u)}}}_{x'},\quad B{^{\mathrm{(d)}}}_{y'}=\epsilon_n B{^{\mathrm{(u)}}}_{y'},$$ where $\epsilon_n=n{^{\mathrm{(d)}}}/n{^{\mathrm{(u)}}}$ is the gas density compression ratio. Now we can transform both the original and compressed fields to the galaxy’s frame by rotating it by the angle $-\theta$, $B_{x}=B_{x'}\cos\theta-B_{y'}\sin\theta,$ $B_{y}=B_{x'}\sin\theta+B_{y'}\cos\theta,$ $B_{z}=B_{z'},$ and then to the sky frame using Eq. (\[trans\]), where the $\tilde x$ and $\tilde y$ component contribute to the magnetic field in the plane of the sky and $B_{\tilde z}$ is directed along the line of sight: $$\vec{B}_\perp=(B_{\tilde x}, B_{\tilde y}),\quad B_\parallel =B_{\tilde z}.$$
Similar relations can be written for the random magnetic field $\vec{{b}}$, but now we cannot neglect the $z$-component of the random magnetic field in the galaxy frame. The compressed random field in the shock frame is given by $$\vec{{b}}{^{\mathrm{(d)}}}=(b{^{\mathrm{(u)}}}_{x'}, \epsilon_n b{^{\mathrm{(u)}}}_{y'},\epsilon_n b{^{\mathrm{(u)}}}_{z'})$$ which can be transformed to the galaxy’s frame and then to the sky frame to obtain $$\begin{aligned}
{b}_{\tilde x}&=&b_{x}=b_{x'}\cos\theta-\epsilon_nb_{y'}\sin\theta,\\
{b}_{\tilde y}&=&b_{x'}\sin\theta\cos i+\epsilon_n b_{y'}\cos\theta\cos i
+\epsilon_n b_{z'}\sin i,\\
{b}_{\tilde z}&=&-b_{x'}\sin\theta\sin i-\epsilon_n b_{y'}\cos\theta\sin i
+\epsilon_n b_{z'}\cos i.\end{aligned}$$ Averaging using these equations then yields $$\begin{aligned}
{\langle{{b}_{\tilde x}^2}\rangle}{^{\mathrm{(d)}}}&=&\frac13{\sigma_b^2}{^{\mathrm{(u)}}}\left[1+(\epsilon_n^2-1)\sin^2\theta\right],\\
{\langle{b}_{\tilde y}^2\rangle}{^{\mathrm{(d)}}}&=&\frac13{\sigma_b^2}{^{\mathrm{(u)}}}\left[1+(\epsilon_n^2-1)(-1\sin^2\theta\cos^2 i)\right],\\
{\langle{b}_{\tilde z}^2\rangle}{^{\mathrm{(d)}}}&=&\frac13{\sigma_b^2}{^{\mathrm{(u)}}}\left[1+(\epsilon_n^2-1)(-1\sin^2\theta\sin^2 i)\right],\end{aligned}$$ where ${\langle{{b}_{x'}^2}\rangle}{^{\mathrm{(u)}}}={\langle{{b}_{y'}^2}\rangle}{^{\mathrm{(u)}}}={\langle{{b}_{z'}^2}\rangle}{^{\mathrm{(u)}}}=\frac13{\sigma_b^2}{^{\mathrm{(u)}}}$ (isotropy of the upstream random magnetic field), and ${\langle{b}_{x'}{b}_{y'}\rangle}{^{\mathrm{(u)}}}={\langle{b}_{x'}{b}_{z'}\rangle}{^{\mathrm{(u)}}}={\langle{b}_{y'}{b}_{z'}\rangle}{^{\mathrm{(u)}}}=0$ (statistical independence of the upstream field components).
Now the arm–interarm contrasts in various observables can be estimated as follows. For the total synchrotron intensity, $$\epsilon_I\simeq
\frac{{B_\perp^2}{^{\mathrm{(d)}}}+{\langle{b}_\perp^2\rangle}{^{\mathrm{(d)}}}}{{B_\perp^2}{^{\mathrm{(u)}}}+{\langle{b}_\perp^2\rangle}{^{\mathrm{(u)}}}}
\,\frac{L_I{^{\mathrm{(d)}}}}{L_I{^{\mathrm{(u)}}}}\frac{n{^{\mathrm{(d)}}}_{\gamma}}{n{^{\mathrm{(u)}}}_{\gamma}},$$ where $L_I$ is the pathlength through the synchrotron layer and $n_{\gamma}$ the cosmic ray number density. The polarized emissivity in a partially ordered, anisotropic random magnetic field and uniform cosmic ray distribution can be calculated using Eq. (20) of @Sokoloff98 as $$PI\propto(B_{\tilde x}^2-B_{\tilde y}^2
+{\langleb_{\tilde x}^2\rangle}-{\langleb_{\tilde y}^2\rangle})^2
+4B_{\tilde x}^2B_{\tilde y}^2.$$ Applying this formula to the downstream magnetic field, we obtain the contrast in polarized intensity $\epsilon_{PI}$ by dividing it by $PI{^{\mathrm{(u)}}}\propto{B_\perp^2}{^{\mathrm{(u)}}}$, and also multiplying by the ratio of pathlengths $L_I{^{\mathrm{(d)}}}/L_I{^{\mathrm{(u)}}}$ and cosmic ray densities $n{^{\mathrm{(d)}}}_{\gamma}/n{^{\mathrm{(u)}}}_{\gamma}$.
The Faraday rotation measure can be calculated as ${\mathrm{RM}}\simeq 0.81{n_\mathrm{e}}B_{\tilde z} L_{{\mathrm{RM}}}$, where $L_{{\mathrm{RM}}}$ is an appropriate pathlength, and its standard deviation is obtained using Eq. (\[eq:sigRM\]), and then the arm–interarm contrast is calculated straightforwardly.
[99]{} Aalto, S., Hüttemeister, S., Scoville, N. Z., & Thaddeus, P. 1999, ApJ, 522, 165
Beck, R. 2007, A&A, 470, 539
Beck, R., & Hoernes, P. 1996, Nat, 379, 47
Beck, R., & Krause, M. 2005, AN, 326, 414
Beck, R., Fletcher, A., Shukurov, A., et al. 2005, A&A, 444, 739
Berezinskii, V. S., Bulanov, S. V., Dogiel, V. A., Ginzburg, V. L., & Ptuskin, V. S. 1990, Astrophysics of Comsic Rays, Amsterdam.
Berkhuijsen, E. M., Horellou, C., Krause, M., et al. 1997, A&A, 318, 700
Berkhuijsen, E. M., Beck, R., & Hoernes, P. 2003, A&A, 398, 937
Burn B. J., 1966, MNRAS, 133, 67
Ciardullo, R., Feldmeier, J. J., Jacoby, G. H., et al. 2002, ApJ, 577, 31
Condon, J. J. 1992, ARA&A, 30, 575
Elmegreen, B.G., Elmegreen, D. M., & Seiden, P. E. 1989, ApJ, 343, 602
Emerson, D. T., & Gräve, R. 1988, A&A, 190, 353
Fletcher, A., Berkhuijsen, E. M., Beck, R., & Shukurov, A. 2004, A&A, 414, 53
Fletcher, A. & Shukurov, A. 2008, MNRAS, 371, L21
Fletcher, A., Korpi, M. & Shukurov, A. 2009, in Cosmic Magnetic Fields: from Planets to Stars and Galaxies, ed. K. G. Strassmeier, A. G. Kosovichev & J. E. Beckman, Proc. IAU Symposium 259, 87
Ford, H. C., Crane, P. C., Jacoby, G. H., & van der Hulst, J. M. 1985, ApJ, 293, 132
Greenawalt, B., Walterbos, R. A. M., Thilker, D. & Hoopes, C. G. 1998, ApJ, 506, 135
Heald, G., Braun, R. & Edmonds, R. 2009, A&A, 503, 409.
Helfer, T. T., Thornley, M. D., Regan, M. W., Wong, T., Sheth, K., Vogel, S. N., Blitz, L., & Bock, D. C. J. 2003, ApJS, 145, 259
Horellou, C., Beck, R., Berkhuijsen, E. M., Krause, M., & Klein, U. 1992, A&A, 265, 417
Howard, S., & Byrd, G. 1990, AJ, 99, 1798
Kalberla,P. M. W. , Dedes,L., Kerp, J., & Haud, U. 2007, A&A, 469, 511
Kim, J. & Ryu, D. 2001, ApJ, 561, L135
Lang, K. R. 1999, Astrophysical Formulae (Berlin: Springer-Verlag)
Longair, M. 1994, High Energy Astrophysics: Volume 2 (Cambridge: CUP)
Mathewson, D. S., van der Kruit, P. C., & Brouw, W. N. 1972, A&A, 17, 468
Moss, D. 1998, MNRAS, 297, 860
Moss, D., Shukurov, A., Sokoloff, D., Beck, R., & Fletcher, A. 2001, A&A, 380, 55
Mouschovias, T. Ch., Shu, F., H. & Woodward, P. R. 1974, A&A, 33, 73
Mouschovias, T. Ch., Kunz, M. W. & Christie, D. A. 2009, MNRAS, 397, 14
Neininger, N. 1992, A&A, 263, 30
Neininger, N., & Horellou, C. 1996, in Polarimetry of the Interstellar Medium, ed. W. G. Roberge & D. C. B. Whittet, ASP Conf. Ser., 97, 592
Patrikeev, I., Fletcher, A., Stepanov, R., et al. 2006, A&A, 458, 441
Regan, M. W., Thornley, M. D., Vogel, S. N., et al. 2006, ApJ, 652, 1121
Roberts, W. W. & Yuan, C. 1970, ApJ, 161, 887
Rohde, R., Beck, R., & Elstner, D. 1999, A&A, 350, 423
Rohlfs, K., & Wilson, T. 1999, Tools of Radio Astronomy (Berlin: Springer-Verlag)
Rots, A. H., Bosma, A., van der Hulst, J. M., Athanassoula, E. & Crane, P. C. 1990, AJ, 1000, 387
Ruzmaikin, A., Sokolov, D., Shukurov, A. & Beck, R. 1990, A&A, 230, 284
Rybicki, G. B. & Lightman, A. P. 1979, Radiative Processes in Astrophysics (New York: Wiley)
Sauvage, M., Blommaert, J., Boulanger, F., et al. 1996, A&A, 315, L89
Scarrott, S. M., Ward-Thompson, D. & Warren-Smith, R. F. 1987, MNRAS, 224, 299
Shetty, R., Vogel, S. N., Ostriker, E. C. & Teuben, P. J. 2007, ApJ, 665, 1138
Shukurov, A. 1998, MNRAS, 299, L21
Shukurov, A. & Berkhuijsen, E. M. 2003, MNRAS, 342, 496
Sokoloff, D. D., Shukurov, A., & Krause, M. 1992, A&A, 264, 396
Sokoloff, D. D., Bykov, A. A., Shukurov, A., Berkhuijsen, E. M., Beck, R., & Poezd, A. D. 1998, MNRAS, 299, 189, and MNRAS, 303, 207, Erratum
Stanimirovic, S. 2002, in Single-Dish Radio Astronomy: Techniques and Applications, ed. S. Stanimirovic, D. Altschuler, P. Goldsmith, & C. Salter, ASP Conf. Proc., 278, 375
Strong, A. W. & Moskalenko, I. V. 1998, ApJ, 509, 212
Sun, X. H., Reich, W., Waelkens, A., & En[ß]{}lin, T. A. 2008, A&A, 477, 573
Sun, X. H. & Reich, W. 2009, A&A, 507, 1087
Tabatabaei, F. S., Beck, R., Krügel, E., Krause, M., Berkhuijsen, E. M., Gordon, K. D., & Menten, K. M. 2007, A&A, 475, 133
Tabatabaei, F. S., Krause, M., Fletcher, A., & Beck, R. 2008, A&A, 490, 1005
Tilanus, R. P. J., Allen, R. J., van der Hulst, J. M., Crane, P. C. & Kennicutt, R. C. 1988, ApJ, 330, 667
Tully, R. B. 1974 ApJS, 27, 437
Wardle, J. F. C., & Kronberg, P. P. 1974, ApJ, 194, 249
\[lastpage\]
[^1]: E-mail: [email protected]
[^2]: The VLA is operated by the NRAO. The NRAO is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc.
[^3]: The Effelsberg 100-m telescope is operated by the Max-Planck-Institut für Radioastronomie on behalf of the Max-Planck-Gesellschaft.
[^4]: The terminology used here is unavoidably imprecise. Instead of a disc and halo we could just as easily refer to a thin and thick disc. We are unable to say anything about the geometry (e.g. flat or spheroidal) of the two layers using our data, only that there must be two layers producing the Faraday rotation observed in M51. In Section \[sec:B\] we look at the magnetic field structure in the two layers in more detail.
[^5]: @Berkhuijsen:1997 adopted a distance to M51 of $8{\,\mathrm{Mpc}}$: we have rescaled their radii to our distance of $7.6{\,\mathrm{Mpc}}$.
|
{
"pile_set_name": "ArXiv"
}
|
September, 1998
M. Wakamatsu[^1] and T. Kubota[^2]
Department of Physics, Faculty of Science,\
Osaka University,\
Toyonaka, Osaka 560, JAPAN
PACS numbers : 13.60.Hb, 12.39.Fe, 12.39.Ki
We carry out a systematic investigation of twist-two spin dependent structure functions of the nucleon within the framework of the chiral quark soliton model (CQSM) by paying special attention to the role of chiral symmetry of QCD. The importance of chiral symmetry is illustrated through the good reproduction of the recent SLAC data for the neutron spin structure function $g_1^n (x,Q^2)$. We also observe substantial difference between the predictions of the longitudinally polarized distribution functions and those of the transversity distribution functions. That the chiral symmetry may be responsible for this difference is seen in the isospin dependence of the corresponding first moments, i.e. the axial and tensor charges. The CQSM predicts $g_A^{(0)} / g_A^{(3)} \simeq 0.25$ for the ratio of the isoscalar to isovector axial charges, while $g_T^{(0)} /
g_T^{(3)} \simeq 0.46$ for the ratio of the isoscalar to isovector tensor charges, which should be compared with the prediction $g_A^{(0)} / g_A^{(3)} = g_T^{(0)} / g_T^{(3)}
= 3 / 5$ of the constituent quark model or of the naive MIT bag model without proper account of chiral symmetry. Another prominent prediction of the CQSM is the opposite polarization of the $\bar{u}$ and $\bar{d}$ antiquarks, thereby indicating the SU(2) asymmetric sea quark (spin) polarization in the nucleon.
Introduction
============
Undoubtedly, the so-called “nucleon spin crisis” caused by the EMC measurement in 1988 is one of the most exciting topics in the field of hadron physics \[1\]. The recent renaissance of nucleon structure function physics is greatly owing to this epoch-making finding. Naturally, the physics of nucleon structure functions has two different aspects. One is a perturbative aspect, while the other is a nonperturbative aspect. Because of the asymptotic freedom of QCD, the $Q^2$-evolution of quark distribution functions can be controlled by the perturbative QCD at least for large enough $Q^2$ \[2\]. However, the perturbative QCD is entirely powerless for predicting distribution functions themselves. Here we need to solve nonperturbative QCD in some way. Unfortunately, we have no reliable analytical method for handling this aspect of QCD. For the present moment, we are then left with two tentative choices. One is to rely upon lattice QCD, while the other is to use effective models of QCD. If one takes the first choice, one must first evaluate infinite towers of moments of distribution functions, since the direct calculation of distribution functions does not match this numerical simulation method \[3\]. Here we take the second choice, which allows us a direct calculation of quark distribution functions. Still, there are quite a lot of effective model of baryons. We advocate that the chiral quark soliton model (CQSM) is a unique model of baryons which has several appealing features not possessed by other models of baryons, especially when applied to the physics of quark distribution functions. First of all, it is an effective model of baryons maximally incorporating spontaneous chiral symmetry breaking of QCD vacuum \[4-6\]. The nucleon in this model is a composite of three valence quarks and infinitely many Dirac sea quarks moving in a slowly rotating M.F. of hedgehog shape. As a natural consequence, it automatically simulates cloud of pions surrounding the core of three valence quarks. Nevertheless, since everything is described in terms of effective quark fields only, we need not worry about a double counting of quark and pion degrees of freedom. (We recall that this kind of double counting occurs, for instance, in models of hadrons based on the linear-sigma-quark-model type lagrangian \[7,8\].) This also means that we do not need to use such an ambiguous procedure as convoluting the pion structure functions with [*pion probability function*]{} (or more precisely a light-cone momentum distribution of the pion) inside the nucleon \[9-11\].
Several group have already attempted to calculate nucleon structure functions within the CQSM or the Nambu-Jona-Lasinio (NJL) soliton model. For instance, Weigel et al. investigated the polarized as well as unpolarized structure functions of the nucleon under the so-called “valence quark approximation” \[12\]. This is not an extremely bad approximation, but it is known to have several unpleasant features. Probably, most serious would be the violation of positivity condition for the unpolarized antiquark (or sea quark) distribution functions. Although such an apparent disaster does not happen for the spin dependent quark distribution functions, a lesson learned from the above observation is that a reliable prediction of [*antiquark*]{} distributions would not be obtainable unless incorporating effects of Dirac sea quarks or equivalently vacuum polarization effects.
More consistent calculation including vacuum polarization effects have been performed by Diakonov et al. \[13,14\] and also by Tanikawa and Saito \[15\] with different regularization schemes, but by confining to the isosinglet unpolarized as well as isovector longitudinally polarized distribution functions, which have values at the leading order of $1 / N_c$ expansion (or at the $0$th order of the expansion in the collective angular velocity $\Omega$ of the hedgehog soliton). Unfortunately, an abundance of interesting physics like the physics of “nucleon spin contents” is contained in the next order of $1 / N_c$ expansion \[5\]. This is easily understood because the inclusion of $O (\Omega^1)$ terms is the minimum condition for the collective quantization treatment of hedgehog solitons to hold. Otherwise, the nucleon cannot have correct quantum numbers \[4-6\].
We have recently reported the first calculation of the $O (\Omega^1)$ contributions to the isovector unpolarized quark distribution function related to the physics of Gottfried sum \[16\] with full inclusion of the vacuum polarization effects \[17\]. It was shown that the model can explain the excess of the $\bar{d}$ sea over the $\bar{u}$ sea in the proton very naturally \[17-19\]. However, some of the treatments there were criticized in a recent paper by Pobylitsa et al. \[20\]. In the process of obtaining theoretical quark distribution functions, we need to evaluate nucleon matrix elements of quark bilinear operators which are nonlocal in time. Their criticism is that the calculation in \[17\] does not treat this nonlocality in time to the full extent.
Now the purpose of the present paper is to carry out a systematic calculation of all the twist-2 spin dependent quark distribution functions of the nucleon as consistently as possible. We evaluate both of the $O (\Omega^0)$ and $O (\Omega^1)$ contributions with full inclusion of the vacuum polarization effects. The above-mentioned nonlocality effects are also carefully taken into account. We believe that these unique features of our theoretical analysis would give new and important information on the nonperturbative aspect of the spin dependent quark distribution functions including the [*antiquark*]{} distributions as well.
The plan of the paper is as follows. For completeness, we give in sect.2 a precise definition of twist-2 quark distribution functions which we shall investigate in the present paper. How to evaluate these quark distribution functions within the framework of the CQSM is explained in sect.3. Sect.4 is devoted to the discussion of the numerical results. We then summarize what we have found in sect.5.
Definition of quark distribution functions
==========================================
Most theoretical analyses of quark distribution functions of the nucleon are based on a field-theoretical formulation given by Collins and Soper \[21\]. As a natural extension, Jaffe and Ji recently carried out a systematic classification of quark distribution functions by including chiral-odd distribution functions which do not appear in the formulas of deep inelastic scattering cross sections \[22\]. According to them, there are nine independent distribution functions, from twist 2 to twist 4. Here we are interested in the twist-2 distribution functions, which are known to have simple parton model interpretation. There are three twist-2 distribution functions, the spin independent (or averaged) distribution $f_1 (x)$, the longitudinally polarized distribution $g_1 (x)$, and what is called the transversity distribution $h_1 (x)$. Following the notation of \[22\], they are represented as $$\begin{aligned}
f_1(x) &=& \frac{1}{\sqrt{2} p^+} \int
\frac{d \lambda}{2 \pi} \,e^{i \lambda x}
<P S | \psi_+^\dagger (0) \psi_+(\lambda n) | P S > ,\\
g_1 (x) &=& \frac{1}{\sqrt{2} p^+} \int
\frac{d \lambda}{2\pi} \,e^{i \lambda x}
<P S_z | \psi_+^\dagger (0) \gamma_5 \psi_+ (\lambda n) | P S_z > ,\\
h_1 (x) &=& \frac{1}{\sqrt{2} p^+} \int
\frac{d \lambda}{2 \pi} \,e^{i \lambda x}
<P S_{\bot} | \psi_+^\dagger (0) \gamma_{\bot} \gamma_5 \psi_+
(\lambda n) | P S_{\bot}> ,\end{aligned}$$ where $p^\mu$ and $n^\mu$ are two light-like (null) vectors, having the properties, $$p^{-} = 0, \ \ \ n^{+} = 0, \ \ \ p^2 = n^2 = 0, \ \ \ p \cdot n = 1 .$$ Without loss of generality, one can choose a frame in which the four-momentum $P^\mu$ of the initial nucleon and the four-momentum transfer $q^\mu$ from a lepton to a nucleon have the third and the time components only. In this frame, $p^\mu$ and $n^\mu$ take the form : $$p^\mu = \frac{{\cal P}}{\sqrt{2}} \,(1,0,0,1), \ \ \
n^\mu = \frac{1}{\sqrt{2} {\cal P}} \,(1,0,0,-1) ,$$ while $P^\mu$ and $q^\mu$ are represented as $$\begin{aligned}
P^\mu &=& p^\mu + \frac{M^2}{2} \,n^\mu , \\
q^\mu &=& \frac{1}{M_N^2} \left(
\nu - \sqrt{\nu^2 + M_N^2 Q^2} \,\right) p^\mu
+ \frac{1}{2} \left( \nu + \sqrt{\nu^2 + M_N^2 Q^2} \,\right)
n^\mu ,\end{aligned}$$ with $\nu = P \cdot q$ and $Q^2 = - q^2$. In the above definition of the twist-2 quark distribution functions, $\psi_{+}$ is a component of the quark field $\psi$ defined through the decomposition $$\psi = (P_{+} + P_{-}) \,\psi = \psi_{+} + \psi_{-} ,$$ by the projection operators $P_{\pm} = \frac{1}{2} \gamma^{\mp}
\gamma^{\pm}$ with $\gamma^{\pm} = \frac{1}{\sqrt{2}} (\gamma^0
\pm \gamma^3)$. According to the authors of \[22\], $\psi_{+}$ is called the “good” component of $\psi$, since it describes an independent propagating degrees of freedom in the light-cone quantization scheme \[23\]. On the other hand, $\psi_{-}$ is called the “bad” component, since it can be interpreted as quark-gluon composites. It is important to recognize that only the good component of $\psi$ appears in the definition of twist-2 quark distribution functions in conformity with the fact that they have simple parton model interpretation. In the actual model calculation of these distribution functions, it is more convenient to rewrite the above expressions with use of the identities : $$\begin{aligned}
P_{+}^2 \ \ \
&=& P_{+} \ = \ \frac{1}{2} \,(1 + \gamma^0 \gamma^3) ,\\
P_{+} \gamma_5 P_{+} &=& \frac{1}{2} \,(1 + \gamma^0 \gamma^3)
\gamma_5 ,\\
P_{+} \gamma_{\perp} \gamma_5 P_{+} &=& \frac{1}{2} \,
(1 + \gamma^0 \gamma^3) \,\gamma_{\perp} \gamma_5 .\end{aligned}$$ Since the distribution functions are in principle frame-independent, it is also convenient to go to the nucleon rest frame, in which one can set ${\cal P} = M_N / \sqrt{2}$. Now using the change of variable as $$\lambda n^\mu = \lambda \,\frac{1}{M_N} \,
(1,0,0,-1) \equiv z^\mu ,$$ we obtain $$\begin{aligned}
z_0 = \frac{\lambda}{M_N}, \ \ \
z_3 = - \frac{\lambda}{M_N} = - z_0, \ \ \
z_{\perp} = 0.\end{aligned}$$ Noting that $$\int_{- \infty}^{\infty} d \lambda \,e^{i \lambda x} \cdots =
M_N \int_{- \infty}^{\infty} d z_0 \,
e^{i \,x \,M_N \,z_0} ,$$ we are then led to the following expressions : $$\begin{aligned}
f_1 (x) &=& \frac{1}{4 \pi} \,\int \,d z^0 \,
e^{\,i \,x \,M_N \,z_0} \nonumber \\
&\times& < \mbox{\boldmath $P$} = 0,
S \,| \,\psi^\dagger (0) (1 + \gamma^0 \gamma^3)
\psi (z) \,| \,
\mbox{\boldmath $P$} = 0, S > |_{z_3 = -z_0, \ z_{\perp} = 0},
\\
g_1 (x) &=& \frac{1}{4 \pi} \,\int \,d z^0 \,
e^{\,i \,x \,M_N \,z_0} \nonumber \\
&\times& < \mbox{\boldmath $P$} = 0, S_z \,| \,
\psi^\dagger (0)
(1 + \gamma^0 \gamma^3) \gamma_5 \psi (z)
\,| \,\mbox{\boldmath $P$} = 0, S_z >
|_{z_3 = -z_0, \ z_{\perp} = 0}, \\
h_1 (x) &=& \frac{1}{4 \pi} \,\int \,d z^0 \,
e^{\,i \,x \,M_N \,z_0} \nonumber \\
&\times& < \mbox{\boldmath $P$} = 0, S_{\perp} \,| \,
\psi^\dagger (0)
(1 + \gamma^0 \gamma^3) \gamma_{\perp} \gamma_5
\psi(z) \,| \,\mbox{\boldmath $P$} = 0, S_{\perp} >|
_{z_3 = -z_0, \ z_{\perp} = 0} .\end{aligned}$$ What is left for us now is to evaluate nucleon matrix elements of quark bilinear operators containing two space-time coordinates with light-cone distance. How to evaluate these matrix elements of [*bilocal*]{} quark operators will be explained in the next section.
Theory of quark distribution functions
======================================
As shown in the previous section, the quark distribution functions of our present interest can generally be represented in the form : $$q (x) \ \ = \ \ \frac{1}{4 \pi} \,\,\int_{-\infty}^\infty
\,\,d z_0 \,\,e^{\,i \,x \,M_N \,z_0} \,\,
{\langle N (\mbox{\boldmath $P$} = 0) \,| \, \bar{\psi} (0) \,
\Gamma \,\psi(z) \,| \,
N (\mbox{\boldmath $P$} = 0) \rangle} \,\,
{|}_{z_3 = - z_0, \,z_\perp = 0} \,\, .$$ In the present study, we confine to spin-dependent distribution functions, so that we are to take $$O_a \ = \ (1 + \gamma^0 \gamma^3) \,\gamma_5, \ \ \
\tau_3 \,(1 + \gamma^0 \gamma^3) \gamma_5 ,$$ respectively for the isoscalar and isovector parts of the longitudinally polarized distribution functions, whereas $$O_a \ = \ (1 + \gamma^0 \gamma^3) \gamma_{\perp} \gamma_5, \ \ \
\tau_3 \,(1 + \gamma^0 \gamma^3) \,\gamma_{\perp} \gamma_5,$$ for the isoscalar and isovector parts of the transversity distributions. We recall here the fact that, extending the definition of distribution function $q(x)$ to interval $-1 \le x \le 1$, the relevant antiquark distributions are given as \[14\], $$\begin{aligned}
\Delta \bar{u} (x) + \Delta \bar{d} (x) &=&
\Delta u(-x) + \Delta d (-x) \hspace{10mm} (0 < x < 1) , \\
\Delta \bar{u} (x) - \Delta \bar{d} (x) &=&
\Delta u(-x) - \Delta d (-x) \hspace{10mm} (0 < x < 1) ,\end{aligned}$$ for the longitudinally polarized distributions, while $$\begin{aligned}
\ \delta \bar{u} (x) + \delta \bar{d} (x) &=&
- \,[ \delta u(-x) + \delta d (-x) ] \hspace{10mm} (0 < x < 1) , \\
\ \delta \bar{u} (x) - \delta \bar{d} (x) &=&
- \,[ \delta u(-x) - \delta d (-x) ] \hspace{10mm} (0 < x < 1) ,\end{aligned}$$ for the transversity distributions \[22\]. As explained in the previous paper \[17\], the basis of our analysis is the following path integral representation of a matrix element of an arbitrary (bilocal) quark bilinear operator between the nucleon states with definite momenta : $$\begin{aligned}
&\,& \langle N (\mbox{\boldmath $P$}) \,| \,\psi^\dagger (0) \,O \,
\psi(z) \,| \,N (\mbox{\boldmath $P$})\rangle
\ \ = \ \ \frac{1}{Z} \,\,\int \,\,d^3 x \,\,d^3 y \,\,
e^{\,- \,i \mbox{\boldmath $P$} \cdot \mbox{\boldmath $x$}} \,\,
e^{\,i \,\mbox{\boldmath $P$} \cdot \mbox{\boldmath $y$}} \,\,
\int {\cal D} U \nonumber \\
&\times& \!\!\!\! \int {\cal D}
\psi \,\,{\cal D} \psi^\dagger \,\,
J_N (\frac{T}{2}, \mbox{\boldmath $x$}) \,\,\psi^\dagger (0) \,
O \,\psi(z) \,\,J_N^\dagger (-\frac{T}{2}, \mbox{\boldmath $y$}) \,\,
\exp \,[\,\,i \int \,d^4 x \,\,\bar{\psi} \,\,
(\,i \! \not\!\partial \,- \,M
U^{\gamma_5}) \,\psi \, ] \, , \ \ \ \ \ \end{aligned}$$ where $$\begin{aligned}
{\cal L} \ \ = \ \ \bar{\psi} \,(\,i \not\!\partial \ - \
M U^{\gamma_5} (x) \,) \,\psi \,\, ,\end{aligned}$$ with $U^{\gamma_5} (x) = \exp [ \,i \gamma_5 \mbox{\boldmath $\tau$}
\cdot \mbox{\boldmath $\pi$} (x) / f_\pi \,]$ being the basic lagrangian of the CQSM, and $$J_N (x) \ \ = \ \ \frac{1}{N_c !} \,\,
\epsilon^{\alpha_1 \cdots \alpha_{N_c}} \,\,
\Gamma_{J J_3, T T_3}^{\{f_1 \cdots f_{N_c}\}} \,\,
\psi_{\alpha_1 f_1} (x) \cdots \psi_{\alpha_{N_c} f_{N_c}} (x) \,\, ,$$ is a composite operator carrying the quantum numbers $J J_3, T T_3$ (spin, isospin) of the nucleon, where $\alpha_i$ is the color index, while $\Gamma^{\{ f_1 \cdots f_{N_C} \}}_{J J_3,T T_3}$ is a symmetric matrix in spin-flavor indices $f_i$. By starting with a stationary pion field configuration of hedgehog shape $U^{\gamma_5}_0
(\mbox{\boldmath $x$}) = \exp \,[ \,i \gamma_5 \mbox{\boldmath $\tau$}
\cdot \hat{\mbox{\boldmath $r$}} F(r) \,]$, the path integral over the pion fields $U$ can be done in a saddle point approximation. Next, we consider two important fluctuations around the static configuration, i.e. the translational and rotational zero-modes. To treat the translational zero-modes, we use an approximate momentum projection procedure of the nucleon state, which amounts to integrating over all shift $\mbox{\boldmath $R$}$ of the soliton center-of-mass coordinates \[14\] : $$\langle N (\mbox{\boldmath $P$}) \,| \,
\psi^\dagger (0) \,O \,\psi (z) \,| \,N
(\mbox{\boldmath $P$}) \rangle
\ \longrightarrow \ \int \,d^3 R \,\,
\langle N (\mbox{\boldmath $P$}) \,| \,
\psi^\dagger (0,- \mbox{\boldmath $R$}) \,
O \,\psi (z_0,\mbox{\boldmath $z$} - \mbox{\boldmath $R$}) \,| \,N
(\mbox{\boldmath $P$}) \rangle \,\, .$$ The rotational zero-modes can be treated by introducing a rotating meson field of the form : $$\begin{aligned}
U^{\gamma_5} ( \mbox{\boldmath $x$}, t) \ \ = \ \ A(t) \,\,
U_0^{\gamma_5} (\mbox{\boldmath $x$}) \,\,A^\dagger (t) \,\, ,\end{aligned}$$ where $A(t)$ is a time-dependent $SU(2)$ matrix in the isospin space. Note first the identity $$\bar{\psi} \,( \,i \not\!\partial - M A (t) U^{\gamma_5}_0
(\mbox{\boldmath $x$}) A^\dagger (t) \,) \,\psi \ \ = \ \
\psi^\dagger_A \,( i \partial_t - H - \Omega \,) \,\psi_A \,\,$$ with $$\begin{aligned}
\psi_A \ = \ A^\dagger (t) \,\psi \,\, , \hspace{5mm}
H \ = \ \frac{\mbox{\boldmath $\alpha$} \cdot \nabla}{i}
\ + \ M \,\beta \,U^{\gamma_5}_0 (\mbox{\boldmath $x$})
\,\, , \hspace{5mm}
\Omega \ = \ - \,i \,A^\dagger (t) \,\dot{A} (t) \,\, .\end{aligned}$$ Here $H$ is a static Dirac Hamiltonian with the background pion fields $U_0^{\gamma_5} (\mbox{\boldmath $x$})$, playing the role of a mean field for quarks, while $\Omega = \frac{1}{2} \Omega_a \tau_a$ is the SU(2)-valued angular velocity matrix later to be quantized as $\Omega_a \rightarrow \hat{J}_a / I$ with $I$ the moment of inertia of the soliton and $\hat{J}_a$ the angular momentum operator \[4-6\]. We then introduce a change of quark field variables $\psi \rightarrow
\psi_A$, which amounts to getting on a body-fixed rotating frame. Denoting $\psi_A$ anew as $\psi$ for notational simplicity, the nucleon matrix element (25) can then be written as $$\begin{aligned}
&\,& \langle N (\mbox{\boldmath $P$}) \,| \,\psi^\dagger (0) \,O \,
\psi(z) \, | \,N (\mbox{\boldmath $P$}) \rangle \nonumber \\
&=& \frac{1}{Z} \,\,\Gamma^{\{ f \}} \,\,{\Gamma^{\{ g \}}}^* \,\,
\int \,\,d^3 x \,\,d^3 y \,\,
e^{-i \mbox{\boldmath $P$} \cdot \mbox{\boldmath $x$}} \,\,
e^{i \mbox{\boldmath $P$} \cdot \mbox{\boldmath $y$}} \,\,
\int \,\,d^3 R \nonumber \\
&\times& \int \,\,{\cal D} A \,\,{\cal D} \psi \,\,
{\cal D} \psi^\dagger
\,\,\exp \,[ \,\,i \,\int \,d^4 x \,\,\psi^\dagger
(\,i \partial_t - H - \Omega) \,\psi \,] \,\,
\prod^{N_c}_{i = 1} \,\,\,
[ \,A ( \frac{T}{2}) \,\,\psi_{f_i}
( \frac{T}{2}, \mbox{\boldmath $x$} ) \,] \nonumber \\
&\times& \psi^\dagger (0, - \mbox{\boldmath $R$})
\,\,A^\dagger (0) \,O \,A(z_0) \,\,
\psi(z_0,\mbox{\boldmath $z$} - \mbox{\boldmath $R$}) \,\,
\prod^{N_c}_{j = 1} \,\,\, [ \,\psi_{g_j}^\dagger
(- \frac{T}{2}, \mbox{\boldmath $y$}) \,\,
A^\dagger (-\frac{T}{2})] \,\, .\end{aligned}$$ Now performing the path integral over the quark fields, we obtain $$\begin{aligned}
&\,& \langle N (\mbox{\boldmath $P$}) \,
| \,\psi^\dagger (0) \,O \,\psi (z) \,
| \,N (\mbox{\boldmath $P$}) \rangle \nonumber \\
&=& \frac{1}{Z} \,\,\tilde{\Gamma}^{\{ f \}} \,\,
{\tilde{\Gamma}}^{{\{ g \}}^\dagger} \,\,N_c \,\,
\int \,\,d^3 x \,\,d^3 y \,\,e^{\,-i \,\mbox{\boldmath $P$}
\cdot \mbox{\boldmath $x$}}
\,\,e^{\,i \mbox{\boldmath $P$} \cdot \mbox{\boldmath $y$}} \,\,
\int d^3 R \,\,\int {\cal D} A \nonumber \\
&\times& \!\!\! \Bigl\{ \,
{}_{f_1} \langle\frac{T}{2}, \mbox{\boldmath $x$}
\,| \,\frac{i}
{\,i \partial_t - H - \Omega} \,| \,0,- \mbox{\boldmath $R$}
\rangle_{\gamma} \cdot
{( A^\dagger (0) O_a A(z_0) )}_{\gamma \delta} \cdot
{}_\delta \langle z_0, \mbox{\boldmath $z$} -
\mbox{\boldmath $R$} \,| \,\frac{i}
{\,i \partial_t - H - \Omega} \,
| - \frac{T}{2}, \mbox{\boldmath $y$} \rangle_{g_1} \nonumber \\
&-& \!\!\!
\mbox{Tr} \,\,{\bigl( \,\langle z_0, \mbox{\boldmath $z$} -
\mbox{\boldmath $R$} \,| \,
\frac{i}{i \partial_t - H - \Omega} \,| \,
0, - \mbox{\boldmath $R$} \rangle
A^\dagger (0) O_a A (z_0) \,\bigr)} \,\,\,
{}_{f_1} \langle \frac{T}{2},\mbox{\boldmath $x$} \,| \,
\frac{i}{\,i \partial_t - H - \Omega} \,| - \frac{T}{2},
\mbox{\boldmath $y$}\rangle_{g_1} \,\, \Bigr\} \nonumber \\
&\times& \prod^{N_c}_{j = 2} \,\,\,[ \,{}_{f_j}
\langle \frac{T}{2}, \mbox{\boldmath $x$}
\,| \,\frac{i}{\,i \partial_t - H - \Omega} \,
| -\frac{T}{2}, \mbox{\boldmath $y$}\rangle_{g_j} \,] \,\cdot \,
\exp \,[\,\mbox{Sp} \log \,(\,i \partial_t - H - \Omega) \, ] \,\, , \end{aligned}$$ with $\tilde{\Gamma}^{\{f\}} = \Gamma^{\{f\}} \,
{[A(\frac{T}{2})]}^{N_c}$ etc. Here $\mbox{Tr}$ is to be taken over spin-flavor indices. Assuming a slow rotation of the hedgehog soliton, we can make use of an expansion in $\Omega$. (Since $\Omega$ is known to be an $O (1 / N_c)$ quantity, this perturbative expansion in $\Omega$ can also be taken as a $1 / N_c$ expansion. For an effective action, this gives $$\mbox{Sp} \log \,
( \,i \partial_t - H - \Omega) \ \ = \ \
\mbox{Sp} \log \,(\,i \partial_t - H)
\ + \ \,i \,\,\,\frac{1}{2} \,\,I \,\,
\int \,\,\Omega_a^2 \,\,d t \,\, .$$ The second term here is essentially the action of a rigid rotor, which plays the role of the evolution operator in the space of collective coordinates. We also use the expansion of the single quark propagator as $$\begin{aligned}
&\,& \!\!\!\!\!
{}_{f_1} \langle \frac{T}{2}, \mbox{\boldmath $x$} \,| \,
\frac{i}{i \partial_t - H - \Omega} \,| \,0,
- \mbox{\boldmath $R$} \rangle {}_\gamma
\ \ = \ \
{}_{f_1} \langle \frac{T}{2}, \mbox{\boldmath $x$} \,| \,
\frac{i}{i \partial_t - H }
\,| \,0, - \mbox{\boldmath $R$} \rangle {}_\gamma \nonumber \\
&\,& - \ \ \int \,\,d z_0^\prime \,\,
d ^3 z^\prime \,\,{}_{f_1}
\langle \frac{T}{2}, \mbox{\boldmath $x$} \,| \,
\frac{i}{i \partial_t - H} \,
| \,z_0^\prime, \mbox{\boldmath $z$}^\prime \rangle {}_\alpha \,\,
\cdot i \,\Omega_{\alpha \beta} (z_0^\prime) \cdot
{}_\beta \langle z_0^\prime, \mbox{\boldmath $z$}^\prime \,| \,
\frac{i}{i \partial_t - H} \,| \,0, - \mbox{\boldmath $R$}
\rangle {}_\gamma \ \ \ \ \ \nonumber \\
&\,& + \hspace{65mm} \dots \ \ \ \ \ . \end{aligned}$$ An important suggestion made in a recent paper by Pobilytsa et al. \[20\] is that one must also take account of the nonlocality (in time) of the operator $A^\dagger (0) O_a A(z_0)$. Expanding this operator around $0$ or $z_0$, one respectively obtains $$\begin{aligned}
A^\dagger (0) O_a A(z_0) &=& A^\dagger (0) O_a A(0)
\ \,\,\,+ \ \,\,\,
z_0 A^\dagger (0) O_a \dot{A} (0) \ \,\,\,+ \ \,\,\,\cdots , \\
\mbox{or} \ \ \ \ \ \
A^\dagger (0) O_a A(z_0) &=& A^\dagger (z_0) O_a A(z_0) \ - \ z_0
\dot{A}^\dagger (z_0) O_a A (z_0) \ + \ \cdots .\end{aligned}$$ Since both choices are known to lead to the same answer \[20\], it is convenient to use a symmetrized form in the following manipulation. This amounts to performing the following replacement : $$\begin{aligned}
A^\dagger (0) O_a A(z_0) &\longrightarrow&
A^\dagger O_a A \ + \ \frac{1}{2} \,z_0 \,
(A^\dagger O_a A \,\, A^\dagger \dot{A} -
\dot{A}^\dagger A \,\, A^\dagger O_a A) , \nonumber \\
&=& \tilde{O}_a \ + \ i \,z_0 \,\frac{1}{2} \,
\{ \Omega , \tilde{O}_a \} ,\end{aligned}$$ in the process of collective quantization of the rotational motion. Here we have introduced the notation $$\tilde{O}_a \ \equiv \ A^\dagger O_a A ,$$ for saving space. Eq.(38) means that the nonlocality of the operator $A^\dagger (0) O_a A(z_0)$ causes a rotational correction [*proportional*]{} to the collective angular velocity $\Omega$. After taking all these into account, we are then led to a perturbative series in $\Omega$, which is also regarded as a $1 / N_c$ expansion : $$\begin{aligned}
&\,& \langle N (\mbox{\boldmath $P$}) |
\psi^\dagger (0) O_a \psi (z)
| N (\mbox{\boldmath $P$}) \rangle \nonumber \\
&=& {\langle N (\mbox{\boldmath $P$}) |
\psi^\dagger (0) O_a \psi (z)
| N (\mbox{\boldmath $P$}) \rangle}^{\Omega^0} \ + \
{\langle N (\mbox{\boldmath $P$}) |
\psi^\dagger (0) O_a \psi (z)
| N (\mbox{\boldmath $P$}) \rangle}^{\Omega^1} \ + \ \cdots ,\end{aligned}$$ where $$\begin{aligned}
&\,& \!\!\!\!\! {\langle N (\mbox{\boldmath $P$})
\,| \,\psi^\dagger (0) \,
O \,\psi (z) \,| \,
N (\mbox{\boldmath $P$}) \rangle}^{\Omega^0} \nonumber \\
&=& \frac{1}{Z} \,\,{\,\tilde{\Gamma}}^{\{ f \}} \,\,
{\tilde{\Gamma}}^{{\{ g \}}^\dagger} \,\,N_c \,\,
\int \,\,d^3 x \,\,d^3 y \,\,e^{\,-i \mbox{\boldmath $P$}
\cdot \mbox{\boldmath $x$}}
\,\,e^{\,i \mbox{\boldmath $P$} \cdot \mbox{\boldmath $y$}}
\,\,\int \,\,d^3 R \,\,\int \,\,{\cal D} A \,\,
{(\tilde{O}_a)}_{\gamma \delta} \nonumber \\
&\times& \!\!\! \Bigl[ \,{}_{f_1}
\langle\frac{T}{2}, \mbox{\boldmath $x$} \,| \,
\frac{i}{i \partial_t - H} \,| \,
0, -\mbox{\boldmath $R$} \rangle_{\gamma} \cdot
{}_ {\delta} \langle z_0,
\mbox{\boldmath $z$} - \mbox{\boldmath $R$}
\,| \,\frac{i}{i \partial_t - H} \,|\, -
\frac{T}{2}, \mbox{\boldmath $y$} \rangle {}_{g_1} \nonumber \\
&-& \!\!\! {}_{\delta} \langle z_0,
\mbox{\boldmath $z$} - \mbox{\boldmath $R$} \,| \,
\frac{i}{i \partial_t - H} \,| \,
0, - \mbox{\boldmath $R$} \rangle_{\gamma}
\cdot {}_{f_1} \langle \frac{T}{2}, \mbox{\boldmath $x$} \,| \,
\frac{i}{i \partial_t - H} \,| \,- \frac{T}{2},
\mbox{\boldmath $y$} \rangle {}_{g_1} \,\Bigr] \nonumber \\
&\times& \prod^{N_c}_{j = 2} \,\,\,\bigl[\, {}_{f_j}
\langle \frac{T}{2}, \mbox{\boldmath $x$}
\,| \,\frac{i}{i \partial_t - H} \,| \,
- \frac{T}{2}, \mbox{\boldmath $y$} \rangle_{g_j} \,\bigr]
\,\cdot \,
\,\,\,\exp \,[\,\,\mbox{Sp} \log (\,i \partial_t - H)
\,+ \,i \,\frac{I}{2}
\,\int \Omega_a^2 \,d t \,\,] \,\, ,\end{aligned}$$ and $$\begin{aligned}
&\,& \langle N (\mbox{\boldmath $P$}) \,| \,\psi^\dagger (0) \,
O \,\psi (z) \,| \,
N (\mbox{\boldmath $P$}) \rangle^{\Omega^1} \nonumber \\
&=& \frac{1}{Z} \,\,{\,\tilde{\Gamma}}^{\{ f \}} \,\,
{\tilde{\Gamma}}^{{\{ g \}}^\dagger} \,\,N_c \,\,
\int \,\,d^3 x \,\,d^3 y \,\,e^{\,-i \mbox{\boldmath $P$}
\cdot \mbox{\boldmath $x$}}
\,\,e^{\,i \mbox{\boldmath $P$} \cdot \mbox{\boldmath $y$}}
\,\,\int \,\,d^3 R \,\,\int \,\,{\cal D} {\cal A} \nonumber \\
&\times& \!\!\! \Biggl\{ \,
\int d^3 z^{\prime} \,\,d z_0^\prime \,\,\,\,
i \,\Omega_{\alpha \beta} (z_0^\prime) \,\,
{(A^\dagger (0) \Omega_a A(z_0))}_{\gamma \delta} \nonumber \\
&\times& \!\!\!\!
\Bigl[ \,{}_{f_1}
\langle\frac{T}{2}, \mbox{\boldmath $x$} \,| \,
\frac{i}{i \partial_t - H} \,| \,
z_0^\prime, \mbox{\boldmath $z$}^\prime \rangle_{\alpha} \cdot
{}_ {\beta} \langle z_0^\prime, \mbox{\boldmath $z$}^\prime
\,| \,\frac{i}{i \partial_t - H} \,|\, 0, - \mbox{\boldmath $R$}
\rangle_\gamma \cdot {}_\delta \langle z_0,
\mbox{\boldmath $z$} - \mbox{\boldmath $R$}
\,| \,\frac{i}{i \partial_t - H} \,| \,-
\frac{T}{2}, \mbox{\boldmath $y$} \rangle_{g_1} \ \ \ \ \nonumber \\
&+& \!\!\!\! {}_{f_1} \langle \frac{T}{2},
\mbox{\boldmath $x$} \,| \,
\frac{i}{i \partial_t - H} \,| \,
0, - \mbox{\boldmath $R$} \rangle_{\gamma}
\cdot {}_{\delta} \langle z_0, \mbox{\boldmath $z$} -
\mbox{\boldmath $R$} \,| \,
\frac{i}{i \partial_t - H} \,| \,z_0^\prime,
\mbox{\boldmath $z$}^\prime \rangle_{\alpha} \cdot {}_{\beta}
\langle z_0^\prime,
\mbox{\boldmath $z$}^\prime \,| \,\frac{i}{i \partial_t -H} |
- \frac{T}{2},
\mbox{\boldmath $y$}\rangle_{g_1} \ \ \ \ \nonumber \\
&-& \!\!\!\! {}_{f_1} \langle \frac{T}{2}, \mbox{\boldmath $x$}
\,|\, \frac{i}{i \partial_t - H} \,| \,
- \frac{T}{2}, \mbox{\boldmath $y$} \rangle_{g_1} \cdot
{}_{\delta} \langle z_0, \mbox{\boldmath $z$} -
\mbox{\boldmath $R$} \,|
\,\frac{i}{i \partial_t - H}
\, | \, z_0^\prime, \mbox{\boldmath $z$}^\prime \rangle_{\alpha}
\cdot {}_{\beta} \langle z_0^\prime,
\mbox{\boldmath $z$}^\prime \,| \,
\frac{i}{i \partial_t - H} \,| \, 0, - \mbox{\boldmath $R$}
\rangle_{\gamma} \,\Bigr] \ \ \ \ \nonumber \\
&+& i \,z_0 \,\frac{1}{2} \,
{\{ \Omega , \tilde{O}_a \}}_{\gamma \delta} \nonumber \\
&\times& \!\!\!\!
\Bigl[ \,{}_{f_1}
\langle\frac{T}{2}, \mbox{\boldmath $x$} \,| \,
\frac{i}{i \partial_t - H} \,| \,
0, -\mbox{\boldmath $R$} \rangle_{\gamma} \cdot
{}_ {\delta} \langle z_0,
\mbox{\boldmath $z$} - \mbox{\boldmath $R$}
\,| \,\frac{i}{i \partial_t - H} \,|\, -
\frac{T}{2}, \mbox{\boldmath $y$} \rangle {}_{g_1} \nonumber \\
&-& \!\!\!\! {}_{\delta} \langle z_0,
\mbox{\boldmath $z$} - \mbox{\boldmath $R$} \,| \,
\frac{i}{i \partial_t - H} \,| \,
0, - \mbox{\boldmath $R$} \rangle_{\gamma}
\cdot {}_{f_1} \langle \frac{T}{2}, \mbox{\boldmath $x$} \,| \,
\frac{i}{i \partial_t - H} \,| \,- \frac{T}{2},
\mbox{\boldmath $y$} \rangle {}_{g_1} \,\Bigr] \,\Biggr\} \nonumber \\
&\times& \prod^{N_c}_{j = 2} \,\,\,\bigl[\, {}_{f_j}
\langle \frac{T}{2}, \mbox{\boldmath $x$}
\,| \,\frac{i}{i \partial_t - H} \,| \,
- \frac{T}{2}, \mbox{\boldmath $y$} \rangle_{g_j} \,\bigr] \,\cdot \,
\,\,\,\exp \,[\,\,\mbox{Sp} \log (\,i \partial_t - H)
\,+ \,i \,\frac{I}{2}
\,\int \Omega_a^2 \,d t \,\,] \,\, .\end{aligned}$$ Let us first discuss the leading $O (\Omega^0)$ term. As usual \[4,5\], we introduce the eigenstates $| m \rangle$ and the associated eigenenergies $E_m$ of the static Dirac hamiltonian $H$, satisfying $$H \,| m \rangle \ = \ E_m \,| m \rangle .$$ This enables us to write down a spectral representation of the single quark Green function as follows : $$\begin{aligned}
{}_{\alpha} \langle\mbox{\boldmath $x$}, t \,|\,
\frac{i}{i \partial_t - H} \,| \,
\mbox{\boldmath $x$}^\prime, t^\prime\rangle_{\beta}
&=& \theta(t - t^\prime) \,\,\sum_{m>0} \,\,
e^{\,- i E_m (t - t^\prime)} \,\,
{}_{\alpha} \langle\mbox{\boldmath $x$}\,| \,m \rangle
\langle m \,| \,\mbox{\boldmath $x$}^\prime \rangle_{\beta}
\nonumber \\
&-& \theta(t^\prime - t) \,\,\sum_{m<0} \,\,
e^{\,-i E_m (t - t^\prime)} \,\,
{}_{\alpha} \langle\mbox{\boldmath $x$} \,| \,m \rangle
\langle m \,| \,\mbox{\boldmath $x$}^\prime\rangle_{\beta} \,\, . \ \\end{aligned}$$ Using this equation together with the relation $$\langle \mbox{\boldmath $z$} - \mbox{\boldmath $R$} \, | \,
\ \ = \ \ \langle -\mbox{\boldmath $R$} \,| \,\,
e^{\,i \mbox{\boldmath $p$} \cdot \mbox{\boldmath $z$}} \,\, ,$$ we can perform the integration over $\mbox{\boldmath $R$}$ in (41). The resultant expression is then put into (18) to carry out the integration over $z_0$. We then arrive at a formula, which provides us with a theoretical basis for evaluating the zeroth order contributions in $\Omega$ to quark distribution functions of the nucleon : $$q (x ; \Omega^0) \ \ = \ \ \int \,
\Psi_{J_3 T_3}^{{(J)}^*} [\xi_A] \,\,\,O^{(0)} [\xi_A] \,\,\,
\Psi_{J_3 T_3}^{(J)} [\xi_A] \,\,d \xi_A ,$$ where $$\Psi_{J_3 T_3}^{(J)} [\xi_A]
\ \ = \ \ \sqrt{\frac{2J+1}{8 \pi^2}} \,\,
{(-1)}^{T + T_3} \,\,D_{-T_3 J_3}^{(J)} (\xi_A) \,\, ,$$ are wave functions, describing the collective rotational motion of the hedgehog soliton, while $$O^{(0)} [\xi_A] \ \ = \ \ M_N \,\frac{N_c}{2} \,
\Bigl( \,\sum_{n \leq 0} - \sum_{n > 0} \,\Bigr) \,
\langle n | \,\tilde{O}_a \delta (x M_N - E_n - p^3) \,
| n \rangle .$$ Using the identity $$\Bigl( \,\sum_{n \leq 0} + \sum_{n > 0} \,\Bigr) \,
\langle n | \,\tilde{O}_a \delta (x M_N - E_n - p^3) \,
| n \rangle \ = \ 0 ,$$ eq.(48) can be expressed in either of the following two forms : $$\begin{aligned}
O^{(0)} [\xi_A] &=& \ \,\,\,\,M_N \,N_c \,
\sum_{n \leq 0} \, \langle n | \,\tilde{O}_a
\delta (x M_N - E_n - p^3) \, | n \rangle \nonumber \\
&=& - \,M_N \,N_c \,
\sum_{n > 0} \, \langle n | \,\tilde{O}_a
\delta (x M_N - E_n - p^3) \, | n \rangle ,\end{aligned}$$ i.e., as a sum over the occupied states or as a sum over the nonoccupied states. As was emphasized in \[14\], it is better to use the first form for $x > 0$, whereas the second form for $x < 0$, for the purpose of numerical calculation.
Next we turn to the $O (\Omega^1)$ contribution. In writing down (42), we have retained the time arguments $0, z_0$ and $z'_0$ in $A^\dagger$, $A$ and $\Omega$, since we have to pay attention to the time order of these collective space operators, which do not generally commute after collective quantization of the rotational zero-energy modes. In the previous paper \[17\], motivated by the physical picture that the time-scale of deep inelastic-scattering processes is much shorter than that of collective rotational motion of the soliton, we dropped special time-order diagrams in which the Coriolis coupling $\Omega$ between the collective rotational motion and the intrinsic quark motion operates in the time interval between $z_0$ and $0$. However, this procedure was criticized by Pobylitsa et al. in a recent paper \[20\]. According to the them, there is little reason to assume approximate degeneracy of $0$ and $z_0$ in $A^\dagger (0) O_a A(z_0)$, since the deep-inelastic scattering processes are not necessarily short distance phenomena. Taking this nonlocality in time arguments more seriously, one should retain all the possible time-order diagrams. In doing so, we must pay attention to the time order of collective space operators $A$ and $\Omega$. By ordering these operators according to their time orders, we are led to the replacement : $$\begin{aligned}
&\,& \!\!\!\!\! \Omega_{\alpha \beta} (z'_0) \,
{(A^\dagger (0) O_a A(z_0))}_{\gamma \delta} \nonumber \\
&\longrightarrow& [ \theta (z'_0, 0, z_0) + \theta (z'_0, z_0, 0) ]
\,\Omega_{\alpha \beta} \,\tilde{O}_{\gamma \delta} \,
\ + \ [ \theta (0, z_0, z'_0) + \theta (z_0, 0, z'_0) ] \,
\tilde{O}_{\gamma \delta} \Omega_{\alpha \beta} \nonumber \\
&+& \ \theta(0, z'_0, z_0) \,{(O_a)}_{\gamma^\prime \delta^\prime}
\,A^\dagger_{\gamma \gamma^\prime} \,
\Omega_{\alpha \beta} \,A_{\delta^\prime \delta}
\ \,\,+ \ \,\,\theta(z_0, z'_0, 0) \,
{(O_a)}_{\gamma^\prime \delta^\prime} \,
A_{\delta^\prime \delta} \,\Omega_{\alpha \beta} \,
A^\dagger_{\gamma \gamma^\prime} .\end{aligned}$$ Here the third and the fourth terms are new ones discarded in the treatment of \[17\]. In order to handle these somewhat peculiar terms, we first recall the rule of collective quantization : $$\Omega \ = \ \frac{1}{2} \,\Omega_a \tau_a
\longrightarrow \frac{1}{2 I} \,J_a \tau_a ,$$ where $J_a$ is the total angular momentum operator satisfying the commutation relations (CR) as follows : $$\begin{aligned}
\,[ J_a, J_b ] &=& i \,\epsilon_{a b c} \,J_c , \\
\,[ J_a, A ] &=& \,\,\,\frac{1}{2} \,A \,\tau_a , \\
\,[ J_a, A^\dagger ] &=& - \frac{1}{2} \,\tau_a \,A^\dagger .\end{aligned}$$ Using these CR, one can show that $$\begin{aligned}
{(O_a)}_{\gamma^\prime \delta^\prime} \,
A^\dagger_{\gamma \gamma^\prime} \,
\,\Omega_{\alpha \beta}
\,A_{\delta^\prime \delta}
&=& \frac{1}{2 I} \,{(\tau_c)}_{\alpha \beta} \,
{(O_a)}_{\gamma^\prime \delta^\prime} \,
A^\dagger_{\gamma \gamma^\prime} \,J_c A_{\delta^\prime \delta}
\nonumber \\
&=& \frac{1}{2 I} \,{(\tau_c)}_{\alpha \beta} \,
{(O_a)}_{\gamma^\prime \delta^\prime} \,
A^\dagger_{\gamma \gamma^\prime} \,
\,\bigl[\, \frac{1}{2} \,{(A \tau_c)}_{\delta^\prime \delta}
+ A_{\delta^\prime \delta} \,J_c \bigr] \nonumber \\
&=& \frac{1}{2 I} \,{(\tau_c)}_{\alpha \beta}
\,\bigl[\, \frac{1}{2} \,{(A^\dagger O_a A \tau_c)}_{\gamma \delta}
+ {(A^\dagger O_a A)}_{\gamma \delta} \,J_c \bigr] ,\end{aligned}$$ where we have used (54). Similarly, by using (55), one may obtain an alternative expression $$\begin{aligned}
&\,& {(O_a)}_{\gamma^\prime \delta^\prime} \,
A^\dagger_{\gamma \gamma^\prime} \,
\Omega_{\alpha \beta}
\,A_{\delta^\prime \delta} \ = \
\frac{1}{2 I} \,{(\tau_c)}_{\alpha \beta}
\,\bigl[\, \frac{1}{2} \,{(\tau_c A^\dagger O_a A)}_{\gamma \delta}
+ J_c \,{(A^\dagger O_a A)}_{\gamma \delta} \bigr] .\end{aligned}$$ In the following manipulation, we find it convenient to take an average of these two expressions as $$\begin{aligned}
{(O_a)}_{\gamma^\prime \delta^\prime} \,
A^\dagger_{\gamma \gamma^\prime} \,
\Omega_{\alpha \beta} \,A_{\delta^\prime \delta}
&=& \frac{1}{8 I} \,{(\tau_c)}_{\alpha \beta} \,
[ \,{(A^\dagger O_a A \tau_c)}_{\gamma \delta} +
{(\tau_c A^\dagger O_a A)}_{\gamma \delta} \,] \nonumber \\
&+& \frac{1}{4 I} \,{(\tau_c)}_{\alpha \beta} \,
[\, {(A^\dagger O_a A )}_{\gamma \delta} J_c +
J_c {(A^\dagger O_a A)}_{\gamma \delta} \,] .\end{aligned}$$ Now we must treat two cases separately. The first is the case in which the operator $O_a$ contains an isospin factor $\tau_a$ as $$O_a = \tau_a \bar{O} .$$ In this case, using the relation $A^\dagger O_a A = D_{ab} \tau_b
\bar{O}$, we can rewrite as $$\begin{aligned}
&\,& {(A^\dagger O_a A \tau_c)}_{\gamma \delta} +
{(\tau_c A^\dagger O_a A)}_{\gamma \delta}
\ = \ D_{ab} \,{\Bigl((\tau_b \tau_c + \tau_c \tau_b)
\bar{O}\Bigr)}_{\gamma
\delta} = 2 \,D_{ac} \,{(\bar{O})}_{\gamma \delta} . \ \ \ \end{aligned}$$ On the other hand, if $O_a$ contains no isospin factor as $$O_a = \bar{O} ,$$ we obtain $${(A^\dagger O_a A \tau_c)}_{\gamma \delta} +
{(\tau_c A^\dagger O_a A)}_{\gamma \delta} \ = \
2 \,{(\tau_c \bar{O})}_{\gamma \delta} .$$ Unifying the two cases, we can then write as $$\begin{aligned}
&\,& {(O_a)}_{\gamma^\prime \delta^\prime} \,
A^\dagger_{\gamma \gamma^\prime} \,
\Omega_{\alpha \beta} A_{\delta^\prime \delta}
\ = \ \frac{1}{4 I} \,{(\tau_c)}_{\alpha \beta}
\left\{ \begin{array}{c}
D_{ac} \bar{O}_{\gamma \delta} \\
{(\tau_c \bar{O})}_{\gamma \delta}
\end{array}
\right\} +
\frac{1}{2} \,\{ \Omega_{\alpha \beta},
{(A^\dagger O_a A)}_{\gamma \delta} \}_{+} .\end{aligned}$$ A similar manipulation for the fourth term in (51) leads to $$\begin{aligned}
&\,& {(O_a)}_{\gamma^\prime \delta^\prime}
A_{\delta^\prime \delta}
\Omega_{\alpha \beta} A^\dagger_{\gamma \gamma^\prime}
\ = \ - \,\frac{1}{4 I} \,{(\tau_c)}_{\alpha \beta}
\left\{ \begin{array}{c}
D_{ac} \bar{O}_{\gamma \delta} \\
{(\tau_c \bar{O})}_{\gamma \delta}
\end{array}
\right\} +
\frac{1}{2} \,\{ \Omega_{\alpha \beta},
{(A^\dagger O_a A)}_{\gamma \delta} \}_{+} .\end{aligned}$$ Retaining all these possible time order diagrams, the $O(\Omega^1)$ contribution to the distribution function now becomes $$\begin{aligned}
&\,& {\langle N (\mbox{\boldmath $P$}) \,| \,\psi^\dagger (z) \,
O \,\psi (0) \,| \,
N (\mbox{\boldmath $P$}) \rangle}^{\Omega^1} \nonumber \\
&=& \frac{1}{Z} \,\,{\,\tilde{\Gamma}}^{\{ f \}} \,\,
{\tilde{\Gamma}}^{{\{ g \}}^\dagger} \,\,N_c \,\,
\int \,\,d^3 x \,\,d^3 y \,\,e^{\,-i \mbox{\boldmath $P$}
\cdot \mbox{\boldmath $x$}}
\,\,e^{\,i \mbox{\boldmath $P$} \cdot \mbox{\boldmath $y$}}
\,\,\int \,\,d^3 R \,\,\int \,\,{\cal D} A \nonumber \\
&\times&
\Biggl\{ \,i \int d^3 z^\prime \,d z_0^\prime \nonumber \\
&\times&
\biggl( \,
[ \theta (z^\prime_0, 0, z_0) + \theta (z^\prime_0, z_0, 0) ]
\,\Omega_{\alpha \beta} \tilde{O}_{\gamma \delta} \ + \
[ \theta (0, z_0, z^\prime_0) + \theta (z_0, 0, z^\prime_0) ]
\,\tilde{O}_{\gamma \delta} \Omega_{\alpha \beta} \nonumber \\
&+& \theta(0, z^\prime_0, z_0) \,
\Bigl[ \,\frac{1}{2} {\{ \Omega_{\alpha \beta},
\,\tilde{O}_{\gamma \delta} \}}_{+} + \frac{1}{4 I}
{(\tau_c)}_{\alpha \beta}
\left\{ \begin{array}{c}
D_{ac} \bar{O}_{\gamma \delta} \\
{(\tau_c \bar{O})}_{\gamma \delta}
\end{array} \right\}
\Bigr] \nonumber \\
&+& \theta(z_0, z^\prime_0, 0) \,
\Bigl[ \,\frac{1}{2} {\{ \Omega_{\alpha \beta},
\tilde{O}_{\gamma \delta} \}}_{+} - \frac{1}{4 I}
{(\tau_c)}_{\alpha \beta}
\left\{ \begin{array}{c}
D_{ac} \bar{O}_{\gamma \delta} \\
{(\tau_c \bar{O})}_{\gamma \delta}
\end{array} \right\} \Bigr] \biggr)
\nonumber \\
&\times& \!\!\!\! \Bigl[ \,{}_{f_1}
\langle\frac{T}{2}, \mbox{\boldmath $x$} \,| \,
\frac{i}{i \partial_t - H} \,| \,
z_0^\prime, \mbox{\boldmath $z$}^\prime \rangle_{\alpha} \cdot
{}_ {\beta} \langle z_0^\prime, \mbox{\boldmath $z$}^\prime
\,| \,\frac{i}{i \partial_t - H} \,|\, 0, - \mbox{\boldmath $R$}
\rangle_\gamma \cdot {}_\delta \langle z_0, \mbox{\boldmath $z$} -
\mbox{\boldmath $R$}
\,| \,\frac{i}{i \partial_t - H} \,| \,-
\frac{T}{2}, \mbox{\boldmath $y$} \rangle_{g_1} \ \ \ \ \nonumber \\
&+& \!\!\!\! {}_{f_1} \langle \frac{T}{2},
\mbox{\boldmath $x$} \,| \,
\frac{i}{i \partial_t - H} \,| \,
0, - \mbox{\boldmath $R$} \rangle_{\gamma}
\cdot {}_{\delta} \langle z_0, \mbox{\boldmath $z$} -
\mbox{\boldmath $R$} \,| \,
\frac{i}{i \partial_t - H} \,| \,z_0^\prime,
\mbox{\boldmath $z$}^\prime \rangle_{\alpha} \cdot {}_{\beta}
\langle z_0^\prime,
\mbox{\boldmath $z$}^\prime \,| \,\frac{i}{i \partial_t -H} |
- \frac{T}{2},
\mbox{\boldmath $y$}\rangle_{g_1} \ \ \ \ \nonumber \\
&-& \!\!\!\! {}_{f_1} \langle \frac{T}{2}, \mbox{\boldmath $x$}
\,|\, \frac{i}{i \partial_t - H} \,| \,
- \frac{T}{2}, \mbox{\boldmath $y$} \rangle_{g_1} \cdot
{}_{\delta} \langle z_0, \mbox{\boldmath $z$} -
\mbox{\boldmath $R$} \,| \,\frac{i}{i \partial_t - H}
\, | \, z_0^\prime, \mbox{\boldmath $z$}^\prime\rangle_{\alpha}
\cdot {}_{\beta} \langle z_0^\prime,
\mbox{\boldmath $z$}^\prime \,| \,
\frac{i}{i \partial_t - H} \,| \, 0, - \mbox{\boldmath $R$}
\rangle_{\gamma} \Bigr] \ \ \ \ \nonumber \\
&+& i \,z_0 \,\frac{1}{2} \,
\{ \Omega, \tilde{O}_a \}_{\gamma \delta}
\nonumber \\
&\times& \!\!\!\! \Bigl[ \,{}_{f_1}
\langle\frac{T}{2}, \mbox{\boldmath $x$} \,| \,
\frac{i}{i \partial_t - H} \,| \,
0, -\mbox{\boldmath $R$} \rangle_{\gamma} \cdot
{}_ {\delta} \langle z_0,
\mbox{\boldmath $z$} - \mbox{\boldmath $R$}
\,| \,\frac{i}{i \partial_t - H} \,|\, -
\frac{T}{2}, \mbox{\boldmath $y$} \rangle {}_{g_1} \nonumber \\
&-& \!\!\!\! {}_{\delta} \langle z_0,
\mbox{\boldmath $z$} - \mbox{\boldmath $R$} \,| \,
\frac{i}{i \partial_t - H} \,| \,
0, - \mbox{\boldmath $R$} \rangle_{\gamma}
\cdot {}_{f_1} \langle \frac{T}{2}, \mbox{\boldmath $x$} \,| \,
\frac{i}{i \partial_t - H} \,| \,- \frac{T}{2},
\mbox{\boldmath $y$} \rangle {}_{g_1} \,\Bigr] \Biggr\}
\nonumber \\
&\times& \prod^{N_c}_{j = 2} \,\,\,[\, {}_{f_j}
\langle \frac{T}{2}, \mbox{\boldmath $x$}
\,| \,\frac{i}{i \partial_t - H} \,| \,
- \frac{T}{2}, \mbox{\boldmath $y$} \rangle_{g_j} \,] \,\cdot \,
\,\,\,\exp \,[\,\,\mbox{Sp} \log (\,i \partial_t - H)
\,+ \,i \,\frac{I}{2}
\,\int \Omega_a^2 \,d t \,\,] \,\, ,\end{aligned}$$ After stating all the delicacies inherent in the structure function problem, we can now proceed in the same way as \[17\] and \[24\]. Using the spectral representation of the single quark Green function (44) together with the relation (45), we can perform the integration over $\mbox{\boldmath $R$},
\mbox{\boldmath $z$}^\prime$, and $z'_0$. The resultant expression is then put into (18) to carry out the integration over $z_0$. We then arrive at a formula, which gives a theoretical basis for evaluating the $O (\Omega^1)$ contributions to quark distribution functions of the nucleon : $$q (x ; \Omega^1) \ \ = \ \ \int \,
\Psi_{J_3 T_3}^{{(J)}^*} [\xi_A] \,\,\,O^{(1)} [\xi_A] \,\,\,
\Psi_{J_3 T_3}^{(J)} [\xi_A] \,\,d \xi_A ,$$ where $$O^{(1)} [\xi_A] \ \ = \ \ O^{(1)}_A \ + \ O^{(1)}_B \ + \
O^{(1)}_{B^\prime} \ + \ O^{(1)}_C ,$$ with $$\begin{aligned}
O^{(1)}_A &=& M_N \,\frac{N_c}{4} \,
( \sum_{m > 0, n \leq 0} - \sum_{n > 0, m \leq 0} )
\,\frac{1}{E_m - E_n} \nonumber \\
&\times& \!\! [ \,\langle n \,| \,\tilde{O}_a
(\delta_n + \delta_m) \,| \,m \rangle \,\,
\langle m \,| \,\Omega \,| \,n \rangle +
\langle n \,| \,\Omega \,| \,m \rangle \,\,
\langle m \,| \, \tilde{O}_a \,(\delta_n + \delta_m) \,
| \,m \rangle \, ] , \\
O^{(1)}_B &=& M_N \,\frac{N_c}{4} \,
( \sum_{m \leq 0, n \leq 0} - \sum_{n > 0, m > 0} )
\,\frac{1}{E_m - E_n} \nonumber \\
&\times& \!\! [ \,\langle n \,| \,\tilde{O}_a
(\delta_n - \delta_m) \,| \,m \rangle \,\,
\langle m \,| \,\Omega \,| \,n \rangle
+ \,\,\langle n \,| \,\Omega \,| \,m \rangle \,\,
\langle m \,| \, \tilde{O}_a \,(\delta_n - \delta_m) \,
| \,m \rangle \, ] , \\
O^{(1)}_{B^\prime} &=& M_N \, \frac{N_c}{8 I} \,
( \sum_{m \leq 0, n \leq 0} - \sum_{n > 0, m > 0} )
\,\frac{1}{E_m - E_n} \nonumber \\
&\,& \hspace{35mm} \times \,
\langle n \,| \,\tau_c \,| \,m \rangle \,\,
\langle m \,| \,
\left\{ \begin{array}{c}
D_{ac} \bar{O} \\
\tau_c \bar{O}
\end{array} \right\} \, (\delta_n - \delta_m) \,
| \,n \rangle , \\
O^{(1)}_C &=& \frac{d}{d x} \,\frac{N_c}{4} \,
( \sum_{n \leq 0} - \sum_{n > 0} ) \,
\langle n \,| \, \{ \tilde{O}_a, \Omega \} \,\delta_n \,| \,n
\rangle .\end{aligned}$$ In the above equations, we have used the notation $$\delta_m \ \equiv \ \delta (x M_N - E_m - p^3), \ \ \mbox{and} \ \
\delta_n \ \equiv \ \delta (x M_N - E_n - p^3) .$$ for saving space. Here $O^{(1)}_A$ is the contribution from the diagram in which $z^\prime_0$ is later (or earlier) than both of $0$ or $z_0$. As was emphasized in \[17\], this term contains transitions between the occupied and nonoccupied single quark levels so that it is not in conflict with the Pauli principle. On the other hand, $O^{(1)}_B$ and $O^{(1)}_{B^\prime}$ are the contributions from diagrams in which $z^\prime_0$ lies between $0$ and $z_0$. Although these terms appear to contain Pauli-violating transitions between the occupied levels themselves or the nonoccupied ones, we take here the viewpoint advocated in \[20\] that there is no compulsory reason to drop them since we are here dealing with operators which are non-local in time. Finally, $O^{(1)}_C$ is the $O (\Omega^1)$ contribution resulting from the nonlocality of the operator $A^\dagger (0) O_a A(z_0)$, i.e. the second term of (38). In deriving $O^{(1)}_C$, use has been made of the identity, $$\frac{1}{2 \pi} \int_{- \infty}^{\infty} dz_0 \,\,i \,z_0 \,
e^{i \,(x M_N - E_n - p^3) \,z_0} \ = \
\frac{1}{M_N} \,\frac{\partial}{\partial x} \,
\delta (x M_N - E_n - p^3) .$$ As will become clear shortly, it is convenient to treat $O^{(1)}_A$ and $O^{(1)}_B$ in a combined way. To see it, first note that, after a simple change of summation indices, $O^{(1)}_A$ can be rewritten as $$\begin{aligned}
O^{(1)}_A &=& M_N \,\frac{N_c}{4} \,
\Bigl\{ \sum_{m > 0, n \leq 0} \frac{1}{E_m - E_n} \,
[ \langle n | \tilde{O}_a \delta_n | m \rangle
\langle m | \Omega | n \rangle + \langle n | \Omega | m \rangle
\langle m | \tilde{O}_a \delta_n | n \rangle ] \nonumber \\
&\,& \hspace{10mm} - \
\sum_{m \leq 0, n > 0} \frac{1}{E_m - E_n} \,
[ \langle m | \tilde{O}_a \delta_n | n \rangle
\langle n | \Omega | m \rangle + \langle m | \Omega | n \rangle
\langle n | \tilde{O}_a \delta_n | m \rangle ] \Bigr\} .\end{aligned}$$ From now on, we treat the two cases separately. First, assume that the relevant operator $O_a$ contains an isospin factor $\tau_a$ in such a form as $O_a = \tau_a \bar{O}$. In this case, in view of the relations $\tilde{O}_a = A^\dagger O_a A = D_{ab} \tau_b \bar{O}$ and $\Omega = \frac{1}{2 I} J_c \tau_c$, we must carefully treat the noncommutativity of the two collective space operators $D_{ab}$ and $J_c$. By keeping the order of $D_{ab}$ and $J_c$, $O^{(1)}_A$ can generally be divided into two pieces \[24\] as $$\begin{aligned}
O^{(1)}_A &=& M_N \,\frac{N_c}{4 I} \,
\frac{1}{2} \,\{ D_{ab}, J_c \}_{+} \nonumber \\
&\times& \Bigl\{ \sum_{m > 0, n \leq 0} \frac{1}{E_m - E_n} \,
[ \langle n | \tau_b \bar{O} \delta_n | m \rangle
\langle m | \tau_c | n \rangle + \langle n | \tau_c | m \rangle
\langle m | \tau_b \bar{O} \delta_n | n \rangle ] \nonumber \\
&-& \,\,\,\sum_{m \leq 0, n > 0} \frac{1}{E_m - E_n} \,
[ \langle n | \tau_b \bar{O} \delta_n | m \rangle
\langle m | \tau_c | n \rangle + \langle n | \tau_c | m \rangle
\langle m | \tau_b \bar{O} \delta_n | n \rangle ] \Bigr\}
\nonumber \\
&+& M_N \,\frac{N_c}{4 I} \,
\,\frac{1}{2} \,[ D_{ab}, J_c ] \nonumber \\
&\times& \Bigl\{ \sum_{m > 0, n \leq 0} \frac{1}{E_m - E_n} \,
[ \langle n | \tau_b \bar{O} \delta_n | m \rangle
\langle m | \tau_c | n \rangle - \langle n | \tau_c | m \rangle
\langle m | \tau_b \bar{O} \delta_n | n \rangle ] \nonumber \\
&-& \,\,\,\sum_{m \leq 0, n > 0} \frac{1}{E_m - E_n} \,
[ \langle n | \tau_b \bar{O} \delta_n | m \rangle
\langle m | \tau_c | n \rangle - \langle n | \tau_c | m \rangle
\langle m | \tau_b \bar{O} \delta_n | n \rangle ] \Bigr\} ,\end{aligned}$$ which contains symmetric and antisymmetric combinations of the two collective space operators $D_{ab}$ and $J_c$. On the other hand, it can be easily verified that $O^{(1)}_B$ term contains symmetric combination only : $$\begin{aligned}
O^{(1)}_B &=& M_N \,\frac{N_c}{4 I} \,
\,\frac{1}{2} \,\{ D_{ab}, J_c \}_{+} \nonumber \\
&\times& \Bigl\{ \sum_{m > 0, n \leq 0} \frac{1}{E_m - E_n} \,
[ \langle n | \tau_b \bar{O} \delta_n | m \rangle
\langle m | \tau_c | n \rangle + \langle n | \tau_c | m \rangle
\langle m | \tau_b \bar{O} \delta_n | n \rangle ] \nonumber \\
&-& \,\,\,\sum_{m \leq 0, n > 0} \frac{1}{E_m - E_n} \,
[ \langle n | \tau_b \bar{O} \delta_n | m \rangle
\langle m | \tau_c | n \rangle + \langle n | \tau_c | m \rangle
\langle m | \tau_b \bar{O} \delta_n | n \rangle ] \Bigr\} .\end{aligned}$$ Combining $O^{(1)}_A$ and $O^{(1)}_B$ terms, we then obtain for the isovector case $$O^{(1)}_A \ + \ O^{(1)}_B \ = \
O^{(1)}_{\{A,B \}} \ + \ O^{(1)}_{[A, B]} ,$$ where $$\begin{aligned}
&\,& O^{(1)}_{\{A, B \}} = M_N \,\frac{N_c}{4 I} \,
\,\frac{1}{2} \,\{ D_{ab}, J_c \}_{+} \,
\left( \sum_{m > 0, n \leq 0} - \sum_{m \leq 0, n > 0} +
\sum_{m \leq 0, n \leq 0} - \sum_{m > 0, n > 0}
\right) , \\
&\,& \hspace{33mm} \times \,
\frac{1}{E_m - E_n} \,
\left[ \langle n | \tau_b \bar{O} \delta_n | m \rangle
\langle m | \tau_c | n \rangle + \langle n | \tau_c | m \rangle
\langle m | \tau_b \bar{O} \delta_n | n \rangle
\right] \nonumber \\
&\,& O^{(1)}_{[A, B]} = M_N \,\frac{N_c}{4 I} \,
\,\frac{1}{2} \,
[ D_{ab}, J_c ] \,
\left(
\sum_{m > 0, n \leq 0} + \sum_{m \leq 0, n > 0}
\right) \nonumber \\
&\,& \hspace{33mm} \times \,
\frac{1}{E_m - E_n} \,
\left[ \langle n | \tau_b \bar{O} \delta_n | m \rangle
\langle m | \tau_c | n \rangle - \langle n | \tau_c | m \rangle
\langle m | \tau_b \bar{O} \delta_n | n \rangle
\right] .\end{aligned}$$ The situation is much simpler for isoscalar operators $O_a = \bar{O}$. Since $\tilde{O}_a = A^\dagger O_a A = A^\dagger \bar{O} A = \bar{O}$, we have only to replace both of $D_{ab}$ and $\tau_b$ by $1$ in the above manipulation, thereby leading to $$\begin{aligned}
O^{(1)}_{\{ A, B \}} &=& M_N \,\frac{N_c}{4 I} \,J_c
\left( \sum_{m > 0, n \leq 0} - \sum_{m \leq 0, n > 0} +
\sum_{m \leq 0, n \leq 0} - \sum_{m > 0, n > 0}
\right) \nonumber \\
&\,& \hspace{20mm} \times \,\frac{1}{E_m - E_n} \,
\left[
\langle n | \bar{O} \delta_n | m \rangle
\langle m | \tau_c | n \rangle + \langle n | \tau_c | m \rangle
\langle m | \bar{O} \delta_n | n \rangle
\right] , \\
O^{(1)}_{[ A, B ]} &=& 0 .\end{aligned}$$ One notices that only the symmetric combination of the matrix elements survives for this isoscalar case. This should be contrasted to the isovector case in which either of the symmetric part or the antisymmetric part survives, depending on the symmetry property of the relevant single quark matrix elements appearing in (78) and (79). As we shall discuss later, the symmetric part contributes to the isoscalar unpolarized distribution function $\Delta u(x) + \Delta d(x)$ and $\delta u(x) + \delta d(x)$ at the $O(\Omega^1)$, whereas the antisymmetric part plays an important role in the $O(\Omega^1)$ term of the isovector polarized distribution functions $\Delta u(x) - \Delta d(x)$ or $\delta u(x) - \delta d(x)$ \[24\].
Now we shall investigate the case of our interest in more detail for obtaining explicit formulas, which can be used for numerical calculation of polarized distribution functions of the nucleon.
$\Delta u(x) + \Delta d(x)$
---------------------------
The relevant operator in this case is $$\tilde{O}_a \ = \ A^\dagger (1 + \gamma^0 \gamma^3) \gamma_5 A
\ = \ (1 + \gamma^0 \gamma^3) \gamma^5 .$$ Since the $O (\Omega^0)$ contribution to $\Delta u(x) + \Delta d(x)$ vanishes due to the hedgehog symmetry, the leading contribution to this distribution function arises from the $O (\Omega^1)$ terms. Due to the symmetry property of the relevant single quark matrix elements, only the symmetric combination of $O^{(1)}_A + O^{(1)}_B$ survives. The total $O (\Omega^1)$ term therefore consists of three pieces, $O^{(1)}_{\{A, B\}}$, $O_{B^\prime}$ and $O^{(1)}_C$. Using the general formulas obtained so far, the contributions of these three terms to $\Delta u(x) + \Delta d(x)$ are given as $$\begin{aligned}
&\,& {[ \Delta u(x) + \Delta d(x) ]}^{(1)}_{\{ A, B \}} \ = \
{\langle J_3 \rangle}_{p \uparrow} \cdot M_N \,\frac{N_c}{4 I} \,
\,( \sum_{m = all, n \leq 0} -
\sum_{m = all, n > 0} ) \frac{1}{E_m - E_n} \nonumber \\
&\,& \hspace{20mm}
\times \,[ \langle n | (1 + \gamma^0 \gamma^3) \gamma_5 \delta_n
| m \rangle \langle m | \tau_3 | n \rangle +
\langle n | \tau_3 | m \rangle
\langle m | (1 + \gamma^0 \gamma^3) \gamma_5 \delta_n
| n \rangle ] , \\
&\,& {[ \Delta u(x) + \Delta d(x) ]}^{(1)}_{B^\prime} \ = \
{\langle 1 \rangle}_{p \uparrow} \cdot M_N \,\frac{N_c}{8 I} \,
\, ( \sum_{m \leq 0, n \leq 0} -
\sum_{m >, n > 0} ) \frac{1}{E_m - E_n} \nonumber \\
&\,& \hspace{56mm}
\times \,\langle n | \tau_c | m \rangle
\langle m | \tau_c (1 + \gamma^0 \gamma^3) \gamma_5
(\delta_n - \delta_m) | n \rangle , \\
&\,& {[ \Delta u(x) + \Delta d(x) ]}^{(1)}_{C} \ = \
{\langle J_3 \rangle}_{p \uparrow} \cdot \frac{d}{d x} \,
M_N \,\frac{N_c}{4 I} \,( \sum_{n \leq 0} - \sum_{n > 0} ) \,
\langle n | \tau_3 (1 + \gamma^0 \gamma^3) \gamma_5 \delta_n
| n \rangle .\end{aligned}$$ In the above equations, ${\langle {\cal O} \rangle}_{p \uparrow}$ denotes a matrix element of a collective space operator ${\cal O}$ with respect to the proton in the spin up state along the $z$-axis, i.e. $${\langle {\cal O} \rangle}_{p \uparrow} \ = \ \int
\Psi^{(\frac{1}{2})}_{\frac{1}{2} \frac{1}{2}} [\xi_A]
\,{\cal O} \,\Psi^{(\frac{1}{2})}_{\frac{1}{2} \frac{1}{2}}
[\xi_A] \,d \xi_A \ = \
\langle p, S_z = 1 / 2 | {\cal O} |
p, S_z = 1 / 2 \rangle .$$ In deriving (83), we have used the relation $$\begin{aligned}
{\langle {\{ \tilde{O}_a, \Omega \}}_{+} \rangle}_{p \uparrow} &=&
{\langle \,{\{ (1 + \gamma^0 \gamma^3) \gamma_5, \frac{1}{2}
J_c \tau_c \}}_{+} \,\rangle}_{p \uparrow}
\ = \ {\langle J_3 \rangle}_{p \uparrow} \cdot \tau_3 \,
(1 + \gamma^0 \gamma^3) \,\gamma_5 .\end{aligned}$$ One may notice that the collective space operator contained in the term ${[ \Delta u(x) + \Delta d(x)]}^{(1)}_{B^\prime}$ is $1$ and it is different from $J_3$ contained in other two terms. The appearance of this term seems to be inconsistent, since it does not change sign in contrast to the other two terms when the direction of the proton spin is reversed. Fortunately, it can be shown that this potentially dangerous term vanishes identically due to the symmetry of the double sum of the single quark matrix element : $${[ \Delta u(x) + \Delta d(x)]}^{(1)}_{B^\prime} = 0 .$$ We are then left with the two terms, i.e. ${[ \Delta u(x) + \Delta d(x)]}^{(1)}_{\{ A, B \}}$ and ${[ \Delta u(x) + \Delta d(x)]}^{(1)}_C$, which both have required state dependence. For the purpose of numerical calculation, it is convenient to rewrite the above two terms slightly further. Using the argument given in appendix of \[15\], we can prove the identity : $$\begin{aligned}
&\,& ( \sum_{m = all, n \leq 0} + \sum_{m = all, n > 0} )
\,\frac{1}{E_m - E_n} \nonumber \\
&\,& \hspace{1mm} \times \,\,\bigl[
\langle n | (1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | m \rangle
\langle m | \tau_3 | n \rangle +
\langle n | \tau_3 | m \rangle
\langle m | (1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | n \rangle
\bigr]
= 0 .\end{aligned}$$ As a consequence, ${[ \Delta u(x) + \Delta d(x) ]}^{(1)}_{\{A, B\}}$ can be expressed either of the following two forms : $$\begin{aligned}
[ \Delta u(x) \!\!\! &+& \!\!\!
\Delta d(x) ]^{(1)}_{\{A, B \}} \nonumber \\
&=& \ \,{\langle J_3 \rangle}_{p \uparrow} \cdot M_N \,\frac{N_c}{4 I}
\sum_{m = all, n \leq 0} \frac{1}{E_m - E_n} \nonumber \\
&\times& [
\langle n | (1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | m \rangle
\langle m | \tau_3 | n \rangle +
\langle n | \tau_3 | m \rangle
\langle m | (1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | n \rangle ]
\nonumber \\
&=& - \,{\langle J_3 \rangle}_{p \uparrow} \cdot M_N \,\frac{N_c}{4 I}
\sum_{m = all, n > 0} \frac{1}{E_m - E_n} \nonumber \\
&\times& [
\langle n | (1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | m \rangle
\langle m | \tau_3 | n \rangle +
\langle n | \tau_3 | m \rangle
\langle m | (1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | n \rangle ] .\end{aligned}$$ As advocated in \[20\], it is convenient to use the first expression given as a sum over the occupied states for the numerical calculation of distribution functions in the region $x > 0$, while to use the second one given as sum over the non-occupied states when $x < 0$, since one can thus avoid vacuum subtraction, i.e. subtraction of the corresponding sums over vacuum levels (with $U = 1$). Following \[20\], we also separate the $E_m = E_n$ contribution from the above sum over the single quark levels. This can be done by noting the identities, $$\begin{aligned}
&\,& \sum_{m \leq 0, n \leq 0} \frac{1}{E_m - E_n} \,
\langle n | (1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | m \rangle
\langle m | \tau_3 | n \rangle +
\langle n | \tau_3 | m \rangle
\langle m | (1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | n \rangle ]
\nonumber \\
&\,& \hspace{10mm} \ = \
\frac{1}{2} \sum_{m \leq 0, n \leq 0} \frac{1}{E_m - E_n} \,
\bigl[ \langle n | (1 + \gamma^0 \gamma^3) \gamma_5
(\delta_n - \delta_m)| m \rangle
\langle m | \tau_3 | n \rangle \nonumber \\
&\,& \hspace{47mm} \ + \
\langle n | \tau_3 | m \rangle
\langle m | (1 + \gamma^0 \gamma^3) \gamma_5
(\delta_n - \delta_m) | n \rangle ] ,\end{aligned}$$ and $$\begin{aligned}
\lim_{E_m \longrightarrow E_n} \!\!\!\!\!\! &\,& \!\!\!\!\!\!
\frac{\delta (x M_N - E_n - p^3) -
\delta (x M_N - E_m - p^3)}{E_m - E_n} \nonumber \\
&=& \delta^\prime (x M_N - E_n - p^3)
\ = \ \frac{1}{M_N} \,\frac{d}{d x} \,\delta (x M_N - E_n - p^3) .\end{aligned}$$ From (90), we can then readily obtain $$\begin{aligned}
&\,& \!\!\! {[ \Delta u(x) + \Delta d(x) ]}^{(1)}_{\{ A, B \}}
\nonumber \\
&=& {\langle J_3 \rangle}_{p \uparrow}
\cdot M_N \,\frac{N_c}{4 I} \,
\sum_{\stackrel{\scriptstyle m = all, n \leq 0}{(E_m \neq E_n)}} \,
\frac{1}{E_m - E_n} \nonumber \\
&\,& \hspace{5mm} \times \,
\left[ \langle n | (1 + \gamma^0 \gamma^3) \gamma_5
\delta_n | m \rangle
\langle m | \tau_3 | n \rangle +
\langle n | \tau_3 | m \rangle
\langle m | (1 + \gamma^0 \gamma^3) \gamma_5
\delta_n | n \rangle
\right] \nonumber \\
&+& {\langle J_3 \rangle}_{p \uparrow} \cdot
\frac{d}{d x} \,\frac{N_c}{4 I}
\sum_{\stackrel{\scriptstyle m \leq 0, n \leq 0}{(E_m = E_n)}} \,
\frac{1}{E_m - E_n} \nonumber \\
&\,& \!\!\!\!\!\! \times \,
\left[ \langle n | (1 + \gamma^0 \gamma^3) \gamma_5
(\delta_n - \delta_m)| m \rangle
\langle m | \tau_3 | n \rangle +
\langle n | \tau_3 | m \rangle
\langle m | (1 + \gamma^0 \gamma^3) \gamma_5
(\delta_n - \delta_m) | n \rangle
\right] ,\end{aligned}$$ and a corresponding expression given as sums over non-occupied levels. The remaining term ${[ \Delta u(x) + \Delta d(x)]}^{(1)}_C$ can similarly be expressed in either of the two equivalent forms as $$\begin{aligned}
{[ \Delta u(x) + \Delta d(x)]}^{(1)}_C &=&
\,\,\,\,\,\,{\langle J_3 \rangle}_{p \uparrow} \cdot \frac{d}{d x}
\,\frac{N_c}{4 I} \,\sum_{n \leq 0} \,
\langle n | \tau_3 (1 + \gamma^0 \gamma^3)
\gamma_5 \delta_n | n \rangle
\nonumber \\
&=& - \,{\langle J_3 \rangle}_{p \uparrow} \cdot \frac{d}{d x}
\,\frac{N_c}{4 I} \,\sum_{n > 0} \,
\langle n | \tau_3 (1 + \gamma^0 \gamma^3)
\gamma_5 \delta_n | n \rangle .\end{aligned}$$ Inserting the complete set of single quark states into the first expression and separating the $E_m \neq E_n$ and $E_m = E_n$ terms in this sum, we obtain $$\begin{aligned}
{[ \Delta u(x) + \Delta d(x)]}^{(1)}_C &=&
{\langle J_3 \rangle}_{p \uparrow} \cdot \frac{d}{d x}
\,\frac{N_c}{4 I} \,
\sum_{\stackrel{\scriptstyle m = all, n \leq 0}{(E_m \neq E_n)}}
\langle n | \tau_3 | m \rangle \langle m |
(1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | n \rangle
\nonumber \\
&+& {\langle J_3 \rangle}_{p \uparrow} \cdot \frac{d}{d x}
\,\frac{N_c}{4 I} \,\,
\sum_{\stackrel{\scriptstyle m \leq 0, n \leq 0}{(E_m = E_n)}}
\langle n | \tau_3 | m \rangle \langle m |
(1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | n \rangle ,\end{aligned}$$ and a corresponding expression given as sums over non-occupied states. Just as argued in \[15\] for the case of the unpolarized distribution function $u(x) - d(x)$, $E_m = E_n$ contribution in the double sums in ${[\Delta u(x) + \Delta d(x)]}^{(1)}_{\{A, B \}}$ and ${[ \Delta u(x) + \Delta d(x)]}^{(1)}_C$ precisely cancel each other. After regrouping the terms in such a way that this cancellation occurs at the level of analytical expressions, the $O (\Omega^1)$ contribution to the distribution function $\Delta u(x) + \Delta d(x)$ can finally be written in the following form : $$\begin{aligned}
&\,& {[ \Delta u(x) + \Delta d(x) ]}^{(1)}
\ = \
{[ \Delta u(x) + \Delta d(x) ]}^{(1)}_{\{ A, B \}^\prime}
\ + \
{[ \Delta u(x) + \Delta d(x) ]}^{(1)}_{C^\prime}\end{aligned}$$ where $$\begin{aligned}
&\,& \!\!\!
{[ \Delta u(x) + \Delta d(x) ]}^{(1)}_{\{ A, B \}^\prime}
\nonumber \\
&=& \,\,{\langle J_3 \rangle}_{p \uparrow} \cdot
M_N \,\frac{N_c}{2 I} \,
\sum_{\stackrel{\scriptstyle m = all, n \leq 0}{(E_m \neq E_n)}}
\frac{1}{E_m - E_n}
\langle n | (1 + \gamma^0 \gamma^3) \gamma_5
\delta_n | m \rangle
\langle m | \tau_3 | n \rangle , \\
&=& \!\!\! - \,{\langle J_3 \rangle}_{p \uparrow} \cdot
\,M_N \,\frac{N_c}{2 I} \,
\sum_{\stackrel{\scriptstyle m = all, n > 0}{(E_m \neq E_n)}}
\frac{1}{E_m - E_n}
\langle n | (1 + \gamma^0 \gamma^3) \gamma_5
\delta_n | m \rangle
\langle m | \tau_3 | n \rangle ,\end{aligned}$$ and $$\begin{aligned}
{[ \Delta u(x) + \Delta d(x)]}^{(1)}_{C^\prime}
&=& \,\,\,\,\,
{\langle J_3 \rangle}_{p \uparrow} \cdot \frac{d}{d x} \,
\frac{N_c}{4 I} \,
\sum_{\stackrel{\scriptstyle m = all, n \leq 0}{(E_m \neq E_n)}}
\langle n | \tau_3 | m \rangle \langle m |
(1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | n \rangle ,
\\
&=& - \,{\langle J_3 \rangle}_{p \uparrow} \cdot \frac{d}{d x} \,
\frac{N_c}{4 I} \,
\sum_{\stackrel{\scriptstyle m = all, n > 0}{(E_m \neq E_n)}}
\langle n | \tau_3 | m \rangle \langle m |
(1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | n \rangle .\end{aligned}$$ These expression will be used in the numerical calculation.
$\Delta u(x) - \Delta d(x)$
---------------------------
The relevant operator for the isovector longitudinally polarized distribution function is $$\tilde{O}_{a = 3} \ = \ A^\dagger \tau_3 (1 + \gamma^0 \gamma^3)
\gamma_5 A \ = \ D_{3 b} \,\tau_b \,
(1 + \gamma^0 \gamma^3) \gamma_5 .$$ The main contribution to this distribution function comes from the $0$th order term in $\Omega$. A simple manipulation gives $$\begin{aligned}
{[ \Delta u(x) - \Delta d(x)]}^{(0)} &=&
\,\,\,\,\,
{\langle D_{33} \rangle}_{p \uparrow} \cdot M_N \,N_c \,
\sum_{n \leq 0} \,\langle n | \tau_3 (1 + \gamma^0 \gamma^3)
\gamma_5 \delta_n | n \rangle \nonumber \\
&=& - \,{\langle D_{33} \rangle}_{p \uparrow} \cdot M_N \,N_c \,
\sum_{n > 0} \,\langle n | \tau_3 (1 + \gamma^0 \gamma^3)
\gamma_5 \delta_n | n \rangle .\end{aligned}$$ The $O (\Omega^1)$ contribution to $\Delta u(x) - \Delta d(x)$ is much far complicated. It generally consists of 4 terms, $O^{(1)}_A$, $O^{(1)}_B$, $O^{(1)}_{B^\prime}$ and $O^{(1)}_C$. As was already mentioned, the symmetric part of the sum of $O^{(1)}_A$ and $O^{(1)}_B$ vanishes for this particular operator, owing to the symmetry of the single quark matrix elements. Using the familiar commutation relation $$[ J_c, D_{3 b} ] \ = \ i \,\epsilon_{c b e} \,D_{3 e} ,$$ the antisymmetric part of $O^{(1)}_A + O^{(1)}_B$ becomes $$\begin{aligned}
&\,& \!\!\!\!\!\! {[\Delta u(x) - \Delta d(x)]}^{(1)}_{[A,B]}
\nonumber \\
&=& {\langle D_{33} \rangle}_{p \uparrow} \cdot M_N \,
\frac{N_c}{8 I} \,\,i \,\epsilon_{3 c b} \,\sum_{m > 0, n \leq 0}
\frac{1}{E_m - E_n} \nonumber \\
&\times& \!\!\!\! [ \langle n | \tau_c | m \rangle \langle m |
\tau_b (1 + \gamma^0 \gamma^3) \gamma_5 (\delta_n + \delta_m)
| n \rangle +
\langle n | \tau_b
(1 + \gamma^0 \gamma^3) \gamma_5 (\delta_n + \delta_m)
| m \rangle \langle m | \tau_c | n \rangle ] \ \ \ \nonumber \\
&=& - \,{\langle D_{33} \rangle}_{p \uparrow} \cdot M_N \,
\frac{1}{I} \,\frac{N_c}{2} \,\sum_{m > 0, n \leq 0}
\frac{1}{E_m - E_n} \nonumber \\
&\times& \!\!\!\! [ \langle n | \tau_{+1} | m \rangle \langle n |
\tau_{+1} (\gamma_5 + \Sigma_3) \frac{\delta_n + \delta_m}{2}
| m \rangle - \langle n | \tau_{-1} | m \rangle
\langle n | \tau_{-1} (\gamma_5 + \Sigma_3)
\frac{\delta_n + \delta_m}{2} | m \rangle ] ,\end{aligned}$$ with the standard definition $\tau_{\pm} = \mp (\tau_1 \pm
i \tau_2) / \sqrt{2}$ and $\Sigma_3 = \gamma^0 \gamma^3 \gamma_5$. Next, from (70) with the case of isovector operator, we find that $$\begin{aligned}
{[ \Delta u(x) - \Delta d(x) ]}_{B^\prime}^{(1)}
&=& {\langle D_{33} \rangle}_{p \uparrow} \cdot
M_N \,\frac{N_c}{8 I} \,\left( \sum_{m \leq 0, n \leq 0} -
\sum_{m > 0, n > 0} \right) \nonumber \\
&\,& \hspace{5mm} \times \,
\frac{1}{E_m - E_n} \,
\langle n | \tau_3 | m \rangle \langle m |
(1 + \gamma^0 \gamma^3) \gamma_5 (\delta_n - \delta_m)
| n \rangle .\end{aligned}$$ One should notice that the state dependence of this somewhat peculiar contribution is nothing different from that of the main term, which implies that there is no reason for this term to vanish. In fact, the single quark matrix element appearing in the above double sum is essentially the same as that appearing in the expression for ${[ \Delta u(x) + \Delta d(x)]}^{(1)}$.
The last but potentially important contribution comes from the nonlocality correction term $O^{(1)}_C$. First note that $$\begin{aligned}
{\{ \tilde{O}_{a=3}, \Omega \}}_{+} &=&
\Bigl\{ D_{3b} \tau_b (1 + \gamma^0 \gamma^3) \gamma_5,
\,\frac{1}{2 I} J_c \tau_c \Bigr\}_{+} \nonumber \\
&=& \frac{1}{2 I} \,
\left ( \,\frac{1}{2} \{ D_{3b}, J_c \}_{+}
\left[
\,\tau_b (1 + \gamma^0 \gamma^3) \gamma_5 \tau_c +
\,\tau_c \tau_b (1 + \gamma^0 \gamma^3) \gamma_5
\right] \right. \nonumber \\
&\,& \hspace{6mm} + \left. \frac{1}{2}
\,[ D_{3b}, J_c ] \,\,\,
\left[ \tau_b (1 + \gamma^0 \gamma^3) \gamma_5 \tau_c -
\tau_c \tau_b (1 + \gamma^0 \gamma^3) \gamma_5
\right] \right)
\nonumber \\
&=& \frac{1}{2 I}
\left( \{ D_{3c}, J_c \}_{+}
(1 + \gamma^0 \gamma^3) \gamma_5 \ + \
i \,\epsilon_{bce} \,[ D_{3b}, J_c ] \,
\tau_e \,(1 + \gamma^0 \gamma^3) \gamma_5
\right) .\end{aligned}$$ The first term of the above quation does not contribute, since $$\begin{aligned}
\sum_{n \leq 0} \,\langle n | (1 + \gamma^0 \gamma^3)
\gamma_5 \delta_n | n \rangle = \sum_{n > 0} \,
\langle n | (1 + \gamma^0 \gamma^3) \gamma_5 \delta_n |
n \rangle = 0 .\end{aligned}$$ Simplifying the second term by using the CR (103), we finally obtain $$\begin{aligned}
&\,& {[ \Delta u(x) - \Delta d(x)]}^{(1)}_C
= - \,{\langle D_{33} \rangle}^{(1)}_{p \uparrow}
\cdot \frac{d}{d x}
\,\frac{N_c}{4 I} \,\left( \sum_{n \leq 0} - \sum_{n > 0} \right)
\,\langle n | \tau_3 (1 + \gamma^0 \gamma^3) \gamma^5
\delta_n | n \rangle . \ \ \ \end{aligned}$$ For the same reason as before, it is convenient to consider these term in a combined way. To this end, we first rewrite (108) by inserting a complete set of single quark states and by separating the $E_m = E_n$ contributions from the resultant double sum. The result can be expressed in two alternative forms as $$\begin{aligned}
&\,& {[ \Delta u(x) - \Delta d(x)]}^{(1)}_C \nonumber \\
&=& \!\!\! - \,{\langle D_{33} \rangle}_{p \uparrow} \cdot
\frac{d}{dx} \, \frac{N_c}{4 I} \,
\left(
\sum_{\stackrel{\scriptstyle m = all, n \leq 0}{(E_m \neq E_n)}}
+ \sum_{\stackrel{\scriptstyle m \leq 0, n \leq 0}{(E_m = E_n)}}
\right)
\langle n | \tau_3 | m \rangle \langle m |
(1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | n \rangle
\nonumber \\
&=& \,{\langle D_{33} \rangle}_{p \uparrow} \cdot
\frac{d}{dx} \,\frac{N_c}{4 I}
\left(
\sum_{\stackrel{\scriptstyle m = all, n > 0}{E_m \neq E_n}}
+ \sum_{\stackrel{\scriptstyle m > 0, n > 0}{E_m = E_n}} \right)
\langle n | \tau_3 | m \rangle \langle m |
(1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | n \rangle .\end{aligned}$$ To rewrite the $B^\prime$ term, we first separate $E_m = E_n$ contributions in the double sum of (105) as $$\begin{aligned}
&\,& {[ \Delta u(x) - \Delta d(x)]}^{(1)}_{B^\prime}
\nonumber \\
&=& {\langle D_{33} \rangle}_{p \uparrow} \cdot
\frac{d}{dx} \,M_N \,\frac{N_c}{2 I} \,
\left(
\sum_{\stackrel{\scriptstyle m \leq 0, n \leq 0}{(E_m \neq E_n)}} -
\sum_{\stackrel{\scriptstyle m > 0, n > 0}{(E_m \neq E_n)}}
\right) \nonumber \\
&\,& \hspace{30mm} \times \,
\frac{1}{E_m - E_n} \,
\langle n | \tau_3 | m \rangle
\langle m |
(1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | n \rangle ,
\nonumber \\
&+& \!\!\! {\langle - D_{33} \rangle}_{p \uparrow} \cdot
\,\frac{d}{dx} \,\frac{N_c}{4 I} \,
\left(
\sum_{\stackrel{\scriptstyle m \leq 0, n \leq 0}{(E_m = E_n)}}
\! - \!
\sum_{\stackrel{\scriptstyle m > 0, n > 0}{(E_m = E_n)}} \right)
\langle n | \tau_3 | m \rangle \langle m |
(1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | n \rangle .\end{aligned}$$ Next, we notice the identity $$\begin{aligned}
0 &=& \sum_{\stackrel{\scriptstyle m = all, n = all}{(E_m \neq E_n)}}
\frac{1}{E_m - E_n} \langle n | \tau_3 | m \rangle
\langle m | (1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | n
\rangle \nonumber \\
&+& \frac{1}{2 M_N} \,\frac{d}{dx} \,
\sum_{\stackrel{\scriptstyle m = all, n = all}{(E_m = E_n)}}
\langle n | \tau_3 | m \rangle
\langle m | (1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | n
\rangle .\end{aligned}$$ Using this identity, (110) can be rewritten in either of the following two forms : $$\begin{aligned}
[ \Delta u(x) \!\!\! &-& \!\!\! \Delta d(x)]^{(1)}_{B^\prime}
\nonumber \\
&=& {\langle D_{33} \rangle}_{p \uparrow} \cdot
M_N \,\frac{N_c}{4 I} \,
\left( 2
\sum_{\stackrel{\scriptstyle m \leq 0, n \leq 0}{(E_m \neq E_n)}} +
\sum_{m > 0, n \leq 0} + \sum_{m \leq 0, n > 0} \right) \nonumber \\
&\,& \hspace{40mm} \times \frac{1}{E_m - E_n} \,
\langle n | \tau_3 | m \rangle \langle m |
(1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | n \rangle ,
\nonumber \\
&+& \,{\langle D_{33} \rangle}_{p \uparrow} \cdot
\frac{d}{dx} \,\frac{N_c}{4 I} \,
\sum_{\stackrel{\scriptstyle m \leq 0, n \leq 0}{(E_m = E_n)}}
\langle n | \tau_3 | m \rangle \langle m |
(1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | n \rangle ,
\nonumber \\
&=& - {\langle D_{33} \rangle}_{p \uparrow} \cdot
M_N \frac{N_c}{4 I}
\left( 2
\sum_{\stackrel{\scriptstyle m > 0, n > 0}{(E_m \neq E_n)}} +
\sum_{m > 0, n \leq 0} + \sum_{m \leq 0, n > 0}
\right) \nonumber \\
&\,& \hspace{40mm} \times \frac{1}{E_m - E_n}
\, \langle n | \tau_3 | m \rangle \langle m |
(1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | n \rangle ,
\nonumber \\
&\,& \hspace{3mm}
- \,{\langle D_{33} \rangle}_{p \uparrow} \cdot \frac{d}{dx}
\frac{N_c}{4 I}
\sum_{\stackrel{\scriptstyle m > 0, n > 0}{(E_m = E_n)}}
\langle n | \tau_3 | m \rangle \langle m |
(1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | n \rangle ,\end{aligned}$$ Comparing (109) and (112), one notices that the $E_m = E_n$ pieces in the double sums cancels precisely between $B^\prime$ and $C$ terms. (This is true for both of the occupied and nonoccupied expressions.) After some manipulation by taking care of this cancellation, the sum of these two terms can finally be expressed as $$\begin{aligned}
[ \Delta u(x) \!\!\! &-& \!\!\! \Delta d(x)]^{(1)}_{B^\prime + C}
\nonumber \\
&=& \,\,\,{\langle D_{33} \rangle}_{p \uparrow} \cdot
M_N \,\frac{N_c}{2 I} \nonumber \\
&\times& \Bigl\{
\sum_{\stackrel{\scriptstyle m = all, n \leq 0}{(E_m \neq E_n)}}
\frac{1}{E_m - E_n} \,
\langle n | \tau_3 | m \rangle \langle m |
(1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | n \rangle ,
\nonumber \\
&-& \frac{1}{2 M_N} \,\frac{d}{dx}
\sum_{\stackrel{\scriptstyle m = all, n \leq 0}{(E_m \neq E_n)}}
\langle n | \tau_3 | m \rangle \langle m |
(1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | n \rangle ,
\nonumber \\
&-& \sum_{m > 0, n \leq 0} \frac{1}{E_m - E_n} \,
\langle n | \tau_3 | m \rangle \langle m |
(1 + \gamma^0 \gamma^3) \gamma_5
\frac{\delta_n + \delta_m}{2} | n \rangle \Bigr\}
\nonumber \\
&=& - \,{\langle D_{33} \rangle}_{p \uparrow} \cdot
M_N \,\frac{N_c}{2 I} \nonumber \\
&\times& \Bigl\{
\sum_{\stackrel{\scriptstyle m = all, n > 0}{(E_m \neq E_n)}}
\frac{1}{E_m - E_n} \,
\langle n | \tau_3 | m \rangle \langle m |
(1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | n \rangle ,
\nonumber \\
&-& \frac{1}{2 M_N} \,\frac{d}{dx}
\sum_{\stackrel{\scriptstyle m = all, n > 0}{(E_m \neq E_n)}}
\langle n | \tau_3 | m \rangle \langle m |
(1 + \gamma^0 \gamma^3) \gamma_5 \delta_n | n \rangle ,
\nonumber \\
&-& \sum_{m \leq 0, n > 0} \frac{1}{E_m - E_n} \,
\langle n | \tau_3 | m \rangle \langle m |
(1 + \gamma^0 \gamma^3) \gamma_5
\frac{\delta_n + \delta_m}{2} | n \rangle \Bigr\} .\end{aligned}$$ For numerical calculation, we shall use the first form for $x > 0$, while the second form for $x < 0$.
$\delta u(x) + \delta d(x)$
---------------------------
Since the evaluation of the transversity distribution can be done in a completely parallel way as the longitudinally polarized distribution functions, we shall show below only the final results. The $O (\Omega^0)$ contribution to $\delta u(x) + \delta d(x)$ vanishes, i.e. $${[ \delta u(x) + \delta d(x)]}^{(0)} \ = \ 0 .$$ The $O (\Omega^1)$ contribution consists of two pieces as $${[ \delta u(x) + \delta d(x)]}^{(1)} \ = \
{[ \delta u(x) + \delta d(x)]}^{(0)}_{\{A, B\}^\prime} \ + \
{[ \delta u(x) + \delta d(x)]}^{(0)}_{C^\prime} ,$$ where $$\begin{aligned}
[ \delta u(x) \!\!\! &+& \!\!\!
\delta d(x) ]^{(1)}_{\{ A, B \}^\prime}
\nonumber \\
&=& \,\,{\langle J_x \rangle}_{p S_x} \cdot M_N \,\frac{N_c}{2 I} \,
\sum_{\stackrel{\scriptstyle m = all, n \leq 0}{(E_m \neq E_n)}}
\frac{1}{E_m - E_n}
\langle n | (\gamma^1 \gamma_5 - i \gamma^2)
\delta_n | m \rangle
\langle m | \tau_1 | n \rangle , \nonumber \\
&=& \!\!\! - \,{\langle J_x \rangle}_{p S_x} \cdot
M_N \,\frac{N_c}{2 I}
\sum_{\stackrel{\scriptstyle m = all, n > 0}{(E_m \neq E_n)}}
\frac{1}{E_m - E_n}
\langle n | (\gamma^1 \gamma_5 - i \gamma^2)
\delta_n | m \rangle
\langle m | \tau_1 | n \rangle , \ \ \end{aligned}$$ and $$\begin{aligned}
&\,& {[ \Delta u(x) + \Delta d(x)]}^{(1)}_{C^\prime} \nonumber \\
&=& \,\,\,\,\,\,
{\langle J_x \rangle}_{p S_x} \cdot \frac{d}{d x} \,
\frac{N_c}{4 I} \,
\sum_{\stackrel{\scriptstyle m = all, n \leq 0}{(E_m \neq E_n)}}
\langle n | \tau_1 | m \rangle \langle m |
(\gamma^1 \gamma_5 - i \gamma^2) \delta_n | n \rangle
\nonumber \\
&=& - \,{\langle J_x \rangle}_{p S_x} \cdot \frac{d}{d x} \,
\frac{N_c}{4 I}
\sum_{\stackrel{\scriptstyle m = all, n > 0}{(E_m \neq E_n)}}
\langle n | \tau_1 | m \rangle \langle m |
(\gamma^1 \gamma_5 - i \gamma^2) \delta_n | n \rangle\end{aligned}$$ Here ${\langle J_x \rangle}_{p S_x}$ is defined by $${\langle J_x \rangle}_{p S_x} \ = \ \langle p S_x | J_x |
p S_x \rangle .$$
$\delta u(x) - \delta d(x)$
---------------------------
The $O (\Omega^0)$ contribution is given by $$\begin{aligned}
{[ \delta u(x) - \delta d(x)]}^{(0)}
&=& \,\,\,\,\,{\langle D_{31} \rangle}_{p S_x} \cdot M_N \,N_c \,
\sum_{n \leq 0} \,\langle n | \tau_3
(\gamma^1 \gamma_5 - i \gamma^2) \delta_n | n \rangle \nonumber \\
&=& \! - \,{\langle D_{31} \rangle}_{p S_x} \cdot M_N \,N_c \,
\sum_{n > 0} \,\langle n | \tau_3
(\gamma^1 \gamma_5 - i \gamma^2) \delta_n | n \rangle .\end{aligned}$$ The $O (\Omega^1)$ contribution consists of two pieces as $${[ \delta u(x) - \delta d(x) ]}^{(1)} \ = \
{[ \delta u(x) - \delta d(x) ]}^{(1)}_{[A,B]} \ + \
{[ \delta u(x) - \delta d(x) ]}^{(1)}_{B^\prime + C}$$ where $$\begin{aligned}
&\,& \!\!\! {[ \delta u(x) - \delta d(x)]}^{(1)}_{[A,B]} \nonumber \\
&=& {\langle D_{31} \rangle}_{p S_x} \cdot M_N \,
\frac{N_c}{8 I} \,\,i \,\epsilon_{3 c b} \,\sum_{m > 0, n \leq 0}
\frac{1}{E_m - E_n} \nonumber \\
&\times& \!\!\!\! [ \langle n | \tau_c | m \rangle \langle m |
\tau_b (\gamma^1 \gamma_5 \! - \! i \gamma^2)
(\delta_n + \delta_m)
| n \rangle +
\langle n | \tau_b
(\gamma^1 \gamma_5 - i \gamma^2) (\delta_n + \delta_m)
| m \rangle \langle m | \tau_c | n \rangle ] \ \ \
\nonumber \\
&=& \!\! - \,{\langle D_{31} \rangle}_{p S_x} \cdot M_N \,
\frac{1}{I} \,\frac{N_c}{2} \,\sum_{m > 0, n \leq 0}
\frac{1}{E_m - E_n} \nonumber \\
&\times& \!\!\!\!\! [ \langle n | \tau_{+1} | m \rangle \langle n |
\tau_{+1} (\gamma^1 \gamma_5 \! - \! i \gamma^2)
\frac{\delta_n + \delta_m}{2}
| m \rangle - \langle n | \tau_{-1} | m \rangle
\langle n | \tau_{-1} (\gamma^1 \gamma_5 - i \gamma^2)
\frac{\delta_n + \delta_m}{2} | m \rangle ] , \ \ \ \end{aligned}$$ and $$\begin{aligned}
[ \delta u(x) \!\!\! &-& \!\!\!
\delta d(x)]^{(1)}_{B^\prime + C} \nonumber \\
&=& {\langle D_{31} \rangle}_{p S_x} \cdot
M_N \,\frac{N_c}{2 I} \nonumber \\
&\times& \Bigl\{
\sum_{\stackrel{\scriptstyle m = all, n \leq 0}{(E_m \neq E_n)}}
\frac{1}{E_m - E_n} \,
\langle n | \tau_3 | m \rangle \langle m |
(\gamma^1 \gamma_5 - i \gamma^2) \delta_n | n \rangle
\nonumber \\
&-& \frac{1}{2 M_N} \,\frac{d}{dx}
\sum_{\stackrel{\scriptstyle m = all, n \leq 0}{(E_m \neq E_n)}}
\langle n | \tau_3 | m \rangle \langle m |
(\gamma^1 \gamma_5 - i \gamma^2) \delta_n | n \rangle
\nonumber \\
&-& \sum_{m > 0, n \leq 0} \frac{1}{E_m - E_n} \,
\langle n | \tau_3 | m \rangle \langle m |
(\gamma^1 \gamma_5 - i \gamma^2)
\frac{\delta_n + \delta_m}{2} | n \rangle \Bigr\} ,
\nonumber \\
&=& - \,{\langle D_{31} \rangle}_{p S_x} \cdot
M_N \, \frac{N_c}{2 I} \nonumber \\
&\times& \Bigl\{
\sum_{\stackrel{\scriptstyle m = all, n > 0}{(E_m \neq E_n)}}
\frac{1}{E_m - E_n} \,
\langle n | \tau_3 | m \rangle \langle m |
(\gamma^1 \gamma_5 - i \gamma^2) \delta_n | n \rangle ,
\nonumber \\
&-& \frac{1}{2 M_N} \,\frac{d}{dx}
\sum_{\stackrel{\scriptstyle m = all, n > 0}{(E_m \neq E_n)}}
\langle n | \tau_3 | m \rangle \langle m |
(\gamma^1 \gamma_5 - i \gamma^2) \delta_n | n \rangle ,
\nonumber \\
&-& \sum_{m \leq 0, n > 0} \frac{1}{E_m - E_n} \,
\langle n | \tau_3 | m \rangle \langle m |
(\gamma^1 \gamma_5 - i \gamma^2)
\frac{\delta_n + \delta_m}{2} | n \rangle \Bigr\} .\end{aligned}$$
Numerical results and discussion
================================
Before showing the results of numerical calculations, we briefly discuss the parameters of our effective model specified by the lagrangian (26). Fixing $f_{\pi}$ to its physical value , i.e. $f_{\pi} = 93 \,\mbox{MeV}$, only one parameters of the model is the constituent quark mass M, which plays the role of the coupling constant between the pion and the effective quark fields. There is some argument based on the instanton picture of the QCD vacuum that the value of this mass parameter should not be extremely far from $350 \,\mbox{MeV}$ \[25\]. Phenomenological analyses of various static baryon observables based on this model prefer a slightly larger value of $M$ between $350 \,\mbox{MeV}$ and $425 \,\mbox{MeV}$ \[5,6\]. In the present analysis, we use the value $M = 375 \,\mbox{MeV}$ favored from analyses of various static observables of baryons. Actually the model contains ultraviolet divergences so that it must be regularized by introducing some physical cutoff. In the case of static nucleon observables, most frequently used regularization scheme is the one based on Schwinger’s proper-time representation \[5,6\]. Unfortunately, how to generalize this regularization scheme in the evaluation of nucleon structure functions is an open problem. For evaluating quark distribution functions, Diakonov et al. then proposed to use the so-called Pauli-Villars regularization scheme, which they claim has several nice properties as compared with the energy cutoff scheme like the proper-time regularization scheme \[14\]. The basic idea of this regularization scheme is very simple. Using the derivative (gradient) expansion, one can evaluate the effective meson action corresponding to the original effective quark lagrangian (26) as $$\begin{aligned}
S_{eff}^M [U] &=& - \,i \,N_c \,\mbox{Sp} [ i \not\!\partial -
M e^{i \gamma_5
\mbox{\boldmath $\tau$} \cdot \mbox{\boldmath $\pi$}/ f_{\pi}} ]
\nonumber \\
&=& \,\,\frac{4 N_c}{f_{\pi}^2} \,I_2 (M) \cdot \frac{1}{2} \,
( \partial_{\mu} \mbox{\boldmath $\pi$})^2 \ + \ \cdots \ .\end{aligned}$$ Here the coefficient of the pion kinetic term given by $$I_2 (M) \ \equiv \ i \int \frac{d^4 k}{(2 \pi)^2}
\frac{M^2}{(k^2 - M^2)^2},$$ contains logarithmic divergence. Clearly, this divergence can be removed by introducing a regularized action $S_{eff}^{reg}$ by $$S_{eff}^{reg} \ \equiv \ S_{eff}^M -
{\left( \frac{M}{M_{PV}}\right)}^2 \,S_{eff}^{M_{PV}} .$$ Here $S_{eff}^{M_{PV}}$ denotes the effective meson action obtained from $S_{eff}^M$ by replacing the dynamical quark mass M with the Pauli-Villars mass $M_{PV}$. In fact, this replaces $I_2 (M)$ with $$I_2^{reg} \ \equiv \ I_2 (M) -
{\left( \frac{M}{M_{PV}} \right)}^2 \,I_2 (M_{PV})
\ = \ \frac{M^2}{16 \pi^2} \log \frac{M_{PV}^2}{M^2} ,$$ which is clearly finite. Demanding further that the pion kinetic term in $S_{eff}^{reg}$ has the correct normalization, one obtain $$\frac{N_c}{4 \pi^2} \,M^2 \,\log \frac{M_{PV}^2}{M^2} = f_{\pi}^2 .$$ For $M = 375 \,\mbox{MeV}$, for instance, this gives $M_{PV} \simeq 562 \,\mbox{MeV}$. Other observables like quark distribution functions, which contains logarithmic divergence, can similarly be regularized as $${\langle O \rangle}^{reg} \ \equiv \
{\langle O \rangle}^M - {\left( \frac{M}{M_{PV}} \right)}^2
{\langle O \rangle}^{M_{PV}} .$$ For the sake of consistency, a soliton solution should also be obtained in the same regularization scheme. The startingpoint of soliton construction is the mean field equation $${\langle \bar{\psi} \psi \rangle}_r^{reg'} \sin F (r)
\ = \ {\langle \bar{\psi} \, i \,\gamma_5 \mbox{\boldmath $\tau$}
\psi \rangle}_r^{reg'} \cos F(r) ,$$ obtained under the assumption of the static hedgehog configuration $$\hat{\mbox{\boldmath $\pi$}} (\mbox{\boldmath $r$})
= f_{\pi} \,\hat{\mbox{\boldmath $r$}} \,F (r) .$$ Here ${\langle \bar{\psi} \psi \rangle}_r^{reg'}$ and ${\langle \bar{\psi} \, i \,\gamma_5 \mbox{\boldmath $\tau$}
\psi \rangle}_r^{reg'}$ are the regularized scalar and pseudoscalar densities in the Pauli-Villars subtraction scheme : $$\begin{aligned}
{\langle \bar{\psi} \psi \rangle}_r^{reg'} \ \ &\equiv& \ \
{\langle \bar{\psi} \psi \rangle}_r^M
\ - \ \left( \frac{M}{M_{PV}} \right)
{\langle \bar{\psi} \psi \rangle}_r^{M_{PV}} \\
{\langle \bar{\psi} \gamma_5 \mbox{\boldmath $\tau$}
\psi \rangle}_r^{reg'}
&\equiv& {\langle \bar{\psi} i \gamma_5
\mbox{\boldmath $\tau$} \psi \rangle}_r^M
\ - \ \left( \frac{M}{M_{PV}} \right) {\langle \bar{\psi} i \gamma_5
\mbox{\boldmath $\tau$} \psi \rangle}_r^{M_{PV}} .\end{aligned}$$ Recently, self-consistent solutions of this equation of motion has been obtained in \[26\] with use of the Kahana-Ripka basis \[27\]. (Essentially the same equation was solved in \[28\] within the framework of the Nambu-Jona Lasinio model with an [*ad hoc*]{} nonlinear constraint.) However, one should use this regularization scheme with some care. In fact, it is known that the single subtraction is not enough to get rid of linear divergences, for instance, contained in the expression of the vacuum quark condensate \[28\], which implies that ${<\bar{\psi} \psi>}_r^{reg'}$ and ${< \bar{\psi} i \gamma_5 \mbox{\boldmath $\tau$} \psi>}_r^{reg'}$ also contain convergences. Why could the authors of refs.\[26,28\] obtain self-consistent solutions then? The reason is in the way of solving the equation of motion (129) in the nonlinear model. Given an appropriate initial form of $F(r)$, one can evaluate ${\langle \bar{\psi} \psi \rangle}_r^{reg'}$ and ${\langle \bar{\psi} \gamma_5 \mbox{\boldmath $\tau$} \psi
\rangle}_r^{reg'}$ by using the Kahana-Ripka plane-wave basis as long as the box size D and the maximum momentum $k_{max}$ are finite. A new $F(r)$ can then be obtained from $$F(r) = \mbox{arctan}
\left( \frac{ {< \bar{\psi} i \gamma_5
\mbox{\boldmath $\tau$}
\psi > }_r^{reg'}}{{<\bar{\psi}
\psi >}_r^{reg'}}
\right) .$$ As $k_{max}$ increases, both of ${<\bar{\psi} \psi>}_r^{reg'}$ and ${<\bar{\psi} i \gamma_5 \mbox{\boldmath $\tau$}
\psi>}_r^{reg'}$ tend to diverge. We numerically find that both quantifies increases at the same rate as $k_{max}$ increases so that the resultant $F(r)$ is quite insensitive to the value of $k_{max}$ for large enough $k_{max}$. This is the reason why stable soliton solutions could be found in the above mentioned single-subtraction Pauli-Villars regularization scheme. The existence of finite energy soliton could also be inferred from the derivative expansion analysis of the nonlinear lagrangian (123) with vanishing current quark masses. Nevertheless, one should keep in mind that it is not a completely satisfactory scheme in the sense that its predictions for some special quantities like the vacuum quark condensate contain divergences. For obtaining satisfactory answers also for these special quantities, the single-substraction Pauli-Villars scheme is not enough. We found that more sophisticated Pauli-Villars scheme with three substraction meets this requirement, and that its self-consistent solutions are only slightly different from those of the naive single-subtraction scheme, except when discussing some special quantities as pointed out above. (This analysis will be reported elsewhere.) Considering the fact that the calculation of the structure functions are very time-consuming, we shall then use here the single-subtraction Pauli-Villars scheme, keeping in mind that some particular observables are out of the application of this regularization scheme. Finally, as for the nucleon mass, we prefer to using the theoretical soliton mass $M_N \simeq 1102 \,\mbox{MeV}$ rather than the physical mass, since it respects the energy-momentum sum rule at the energy scale of the model.
For evaluating quark distribution functions at the $O (\Omega^1)$, we must perform infinite double sums over all the single-quark orbitals which are eigenstates of the static Hamiltonian $H$ given by (31). As far as static nucleon observables, a numerical technique for carrying out such double sums was established in \[5\]. On the other hand, several new subtleties arising in the evaluation of quark distribution functions have been explained in \[17\]. In the actual numerical calculation, the expression of each physical quantity is divided into two pieces, i.e. the contribution of what we call the valence quark level (it is the lowest energy eigenstate of the static Dirac hamiltonian $H$, which emerges from the positive energy continuum) and that of the Dirac sea quarks (or the vacuum polarization contribution) as explained in \[17\]. The regularization is introduced into the latter part only.
Now we start to show the results of our numerical calculation for polarized quark distribution functions of the nucleon. Shown in Fig.1 are the $O(\Omega^0)$ contributions to the isovector longitudinally polarization distribution functions $\Delta u (x) - \Delta d (x)$ (solid curves) and $\Delta \bar{u} (x) - \Delta \bar{d} (x)$ (dashed curves), which was first calculated by Diakonov et al. \[14\]. Here Fig.1(a) represents the contributions of the discrete valence quark level, while Fig.1(b) is the vacuum polarization contribution to the same quantities. Sum of these two contributions are shown in Fig.1(c). As shown by Diakonov et al., the $O (\Omega^0)$ vacuum polarization contributions to $\Delta u(x) - \Delta d(x)$ and $\Delta \bar{u} (x) - \Delta \bar{d} (x)$ are fairly large. (The large and positive longitudinal polarization of the isovector combination of the antiquark distributions seems to be a characteristic prediction of the CQSM, which can in principle be tested by the improved phenomenological analyses of polarized parton distribution functions in the near future. More detailed discussion on this point will be given after finishing the evaluation of the $O (\Omega^1)$ contribution to the same distribution function as well as that of the isoscalar longitudinally polarized distribution functions.)
Next, we show in Fig.2 the $O (\Omega^1)$ contribution to the same distribution functions $\Delta u(x) - \Delta d(x)$ and $\Delta \bar{u} (x) - \Delta \bar{d} (x)$. Fig.2(a), 2(b) and 2(c) respectively stand for the $O (\Omega^1)$ contributions of the discrete valence quark level, those of the Dirac sea quarks (or the vacuum polarization contributions), and their sums. One sees that the $O (\Omega^1)$ contributions to the isovector longitudinally polarized distribution function are far from negligible as compared with the leading $O (\Omega^0)$ contributions. This could be expected since the first moment of this distribution functions gives the isovector axial coupling constant of the nucleon. $$g_A^{(3)} \ = \ \int_0^1 \,\{ [ \Delta u(x) - \Delta d(x)]
\ + \ [ \Delta \bar{u}(x) - \Delta \bar{d}(x)] \} \,d x ,$$ while we already know from the previous analyses that the $O (\Omega^1)$ contribution to $g_A^{(3)}$ is large enough to resolve the longstanding $g_A$ problem in the hedgehog soliton model \[29,30,24\]. Adding this $O(\Omega^1)$ contribution to the leading $O(\Omega^0)$ contribution, we obtain final answers for $\Delta u(x) - \Delta d(x)$ and $\Delta \bar{u} (x) - \Delta \bar{d} (x)$, which will be shown later together with the final answer for the isoscalar longitudinal distribution functions $\Delta u(x) + \Delta d(x)$ and $\Delta \bar{u} (x) + \Delta \bar{d}(x)$. Before showing those, we give in Fig.3 the result for the $O (\Omega^1)$ contributions to the isoscalar longitudinally polarized distribution functions, $\Delta u (x) + \Delta d(x)$ and $\Delta \bar{u}(x) + \Delta \bar{d}(x)$. (We recall that there is no $O (\Omega^0)$ contribution to these distribution functions.) Fig.3(a), 3(b) and 3(c) respectively stand for the contributions of the discrete valence quark level, those of the vacuum polarization contributions, and their sums. One sees that the vacuum polarization contributions to the distribution functions $\Delta u (x) +
\Delta d (x)$ and $\Delta \bar{u} (x) + \Delta \bar{d} (x)$ are much smaller than those of the corresponding isovector distributions $\Delta u (x) - \Delta d (x)$ and $\Delta \bar{u} (x) -
\Delta \bar{d} (x)$. Now we show in Fig.4 the final answers for $\Delta u(x) - \Delta d(x)$ and $\Delta \bar{u}(x) - \Delta \bar{d}(x)$, which are the sums of the $O (\Omega^0)$ and $O (\Omega^1)$ contributions, in comparison with the final answers for $\Delta u(x) + \Delta d(x)$ and $\Delta \bar{u}(x) + \Delta \bar{d}(x)$ arising from the $O (\Omega^1)$ terms alone. We observe quite a big difference between the isovector distributions and the isoscalar one. The overall magnitude of $\Delta u (x) +
\Delta d (x)$ is much smaller than that of $\Delta u (x) - \Delta d (x)$, which denotes that $u$-quark is positively polarized, while the $d$-quark is negatively polarized to the direction of proton spin.
At this stage, it may be interesting to compare our theoretical predictions for the longitudinally polarized quark distribution functions with some of the semi-phenomenological parametrization. The parametrization given by Glück, Reya, Stratmann and Vogelsang is especially convenient for the purpose of handy comparison \[31\], since the normalization point ($Q_{init}^2 \simeq 0.34 \,\mbox{GeV}^2$) of their parametrization is fairly close to the energy scale of our effective quark model ($M_{PV}^2 \simeq 0.32 \,\mbox{GeV}^2$. Fig.5 shows this comparison. The filled squares in Fig.5(a) and Fig.5(b) stand for the GRSV parametrizations for the quark distribution functions $x ( \Delta u(x) + \Delta \bar{u} (x) + \Delta d(x) +
\Delta \bar{d} (x))$ and $x ( \Delta u(x) + \Delta \bar{u} (x) -
\Delta d(x) - \Delta \bar{d} (x))$, respectively. Of the two theoretical curves in each figure, the solid curve is the answer of the present calculation, whereas the dashed curve is obtained by using the old treatment used in \[17\], which amounts to dropping some of the nonlocality effects in time. One observes that the nonlocality corrections newly introduced in the present analysis are quite important especially for the isoscalar distribution $x ( \Delta u(x) + \Delta \bar{u} (x) +
\Delta d(x) + \Delta \bar{d} (x))$, while it is less important for the isovector distribution $x ( \Delta u(x) + \Delta \bar{u} (x) -
\Delta d(x) - \Delta \bar{d} (x))$. (This is probably because the nonlocality correctios appearing at the $O (\Omega^1)$ are masked by the dominant $O (\Omega^0)$ contribution in the case of isovector polarized distribution functions.) By comparing the two theoretical curves for $x ( \Delta u(x) + \Delta \bar{u} (x) + \Delta d(x) + \Delta
\bar{d} (x))$ with the corresponding GRSV parametrization, one finds that the new treatment leads to a better agreement. Especially impressive is that the new treatment reproduces the negative sign of the GRSV distribution function in the smaller $x$ region, although one should not forget the fact that the GRSV parametrizations are not experimental data themselves. We point out that the most important factor leading to this qualitative difference between the old and new treatments of the quark distribution functions is the nonlocality correction arising from the second term of (38), i.e. the proper account of nonlocality in time of the operator $A^\dagger (0) O_a A(z_0)$.
Turning back to Fig.4, let us inspect the theoretical predictions for the antiquark distributions in more detail. An interesting feature is that, in most region of $x$, $\Delta \bar{u} (x) -
\Delta \bar{d} (x) > 0$ and $\Delta \bar{u} (x) +
\Delta \bar{d} (x) < 0$ with the relation $| \Delta \bar{u} (x) - \Delta \bar{d} (x) | \gg
| \Delta \bar{u} (x) + \Delta \bar{d} (x) |$. This denotes that $\bar{d}$ is strongly polarized in the opposite direction to the proton spin, while $\bar{u}$ is weakly polarized in the same direction to the proton spin. This appears to be a prominent prediction of the CQSM, which is worthy of special mention. In fact, it sharply contradicts the assumption of SU(2) symmetric sea quark polarization $\Delta \bar{u} (x) = \Delta \bar{d} (x)$, which is frequently used in semi-phenomenological analyses of parton distributions. The isospin symmetric polarization is also assumed in the analysis by Glück, Reya, Stirling and Vogt \[31\]. We compare in Fig.6 our prediction for the $x \Delta \bar{u} (x)$ and $x \Delta \bar{d} (x)$ with the GRSV parametrization, which assumes that $x \Delta \bar{u} (x) = x \Delta \bar{d} (x) \,
( \equiv x \Delta \bar{q} (x) )$. Naturally, one finds qualitative difference between the theoretical distributions and the GRSV parametrization. Still, it is interesting to see that the average of the two theoretical distributions $x \Delta \bar{u} (x)$ and $x \Delta \bar{d} (x)$ is not extremely different from the corresponding GRSV parametrization $x \Delta \bar{q} (x)$. As for the unpolarized distribution functions, the breakdown of the assumption of SU(2) symmetric sea has already been confirmed by the NMC measurement \[16\]. By the same token, there is no compelling reason to believe that the spin dependent antiquark (sea quark) distributions are isospin symmetric. In fact, our previous analyses based on the same model shows that the isospin asymmetry of the unpolarized sea quark distributions can be explained very naturally as combined effects of two ingredients, i.e. the apparently existing flavor asymmetry of valence quark numbers in the nucleon and the spontaneous chiral symmetry breaking of QCD vacuum \[17-19\]. It is just the same mechanism that is responsible for the opposite longitudinal polarization of the $\bar{u}$ and $\bar{d}$ quarks.
The above-mentioned fairly big difference between the isovector and isoscalar longitudinally polarized distribution functions manifests itself also in their first moments, i.e. the isovector and isoscalar axial charges given as $$\begin{aligned}
g_A^{(3)} &=& \int_0^1 \,\{ [ \Delta u(x) - \Delta d(x)]
\ + \ [ \Delta \bar{u}(x) - \Delta \bar{d}(x)] \} \,d x
\ \simeq \ 1.41 ,\\
g_A^{(0)} &=& \int_0^1 \,\{ [ \Delta u(x) + \Delta d(x) ]
\ + \ [ \Delta \bar{u}(x) + \Delta \bar{d}(x)] \} \,d x
\ \simeq \ 0.35 .\end{aligned}$$ The resultant large isovector axial charge and small isoscalar (flaver-singlet) one seem to be qualitatively consistent with the observation. Especially interesting here is the flavor-singlet axial charge identified with the quark spin content of the nucleon. In the context of the CQSM, this quantity was first investigated in \[5\] with use of the self-consistent soliton solution obtained in the proper-time regularization scheme. The value of $g_A^{(0)} = \langle \Sigma_3 \rangle$ obtained there ranges from $0.4 \sim 0.5$ corresponding to the variation of the dynamical quark mass $M$ from $425 \,\mbox{MeV}$ to $375 \,
\mbox{MeV}$. One may notice that the value $g_A^{(0)} \simeq
0.35$ obtained in the present calculation is a little smaller than the previous one. The cause of this difference can be traced back to the qualitative change of the self-consistent soliton solution obtained in the new regularization scheme. As a general trend, the Pauli-Villars regularization scheme cuts off high momentum components more weakly than the energy-cutoff scheme like the proper-time one, thereby leading to soliton solutions with stronger distortion. Incidentally, owing to the nucleon spin sum rule $\langle L_3 \rangle + \frac{1}{2} \langle \Sigma_3 \rangle
= \frac{1}{2}$ proved in \[5\], the rest of the nucleon spin is carried by the orbital angular momentum of the effective quark fields. (Naturally, this is true only at low $Q^2$ corresponding to the energy scale of our effective model. It will be shown later that an increasing portion of the nucleon spin is carried by gluons as $Q^2$ increases.) A soliton with stronger distortion gives larger orbital angular momentum, and consequently smaller quark spin fraction \[5\].
The characteristic feature of the above theoretical prediction, i.e. larger isovector charge and smaller isoscalar one seems also consistent with the idea of $N_c$ counting or $1 / N_c$ expansion of QCD. For understanding it, we just recall the fact that the collective angular velocity $\Omega$ scales as $1 /N_c$, so that the leading contributions to the isovector and isoscalar polarized distribution functions are respectively of the $O (N_c^1)$ and $O (N_c^0)$. The detailed comparison of the theoretical first moments with the corresponding experimental data will be given later after taking account of the scale dependence of them.
Now we show the results of our numerical calculation for transversity distributions. Fig.7 shows the $O(\Omega^0)$ contributions to the isovector transversity distribution functions $\delta u(x) - \delta d(x)$ (solid curves) and $\delta \bar{u}(x) - \delta \bar{d}(x)$ (dashed curves). Here Fig.7(a) stands for the contributions of the discrete valence level, while 7(b) is the vacuum polarization contributions to the same quantities. The sum of these two contributions are shown in Fig.7(c). One finds that the vacuum polarization contributions to these distribution functions are insignificant. Next, we show in Fig.8 the $O (\Omega^1)$ contributions to the same transversity distribution functions $\delta u(x) - \delta d(x)$ and $\delta \bar{u}(x) - \delta \bar{d}(x)$. Also for these $O (\Omega^1)$ terms, the vacuum polarization contributions are very small as compared with the contributions of the discrete valence quark level. However, we emphasized that the valence level contribution at the $O (\Omega^1)$ is far from small as compared with the leading $O (\Omega^0)$ contributions, and should not be discarded. Shown in Fig.9 are the theoretical isoscalar transversity distributions resulting at the $O (\Omega^1)$. One sees that vacuum polarization contributions to the isoscalar transversity distributions is also rather small. The final predictions of the CQSM for $\delta u(x) - \delta d(x)$ and $\delta \bar{u}(x) - \delta \bar{d}(x)$, which are the sums of the $O(\Omega^0)$ and $O(\Omega^1)$ contributions are shown in Fig.10, in comparison with the final answer for $\delta u(x) + \delta d(x)$ and $\delta \bar{u}(x) + \delta \bar{d}(x)$ arising from $O (\Omega^1)$ terms. One again sees that the magnitudes of the isoscalar distributions are much smaller than those of the isovector distributions. Remember the similar observation made for the longitudinally polarized distribution functions. To see it in more detail, we find that the ratio of the isoscalar to isovector distribution are much smaller for the longitudinally polarized distribution than for the transversity one. We shall come back later to this point when discussing the corresponding first moments of these spin dependent quark distribution functions.
Roughly speaking, the quark distribution functions evaluated here corresponds to the energy scale of the order of the Pauli-Villars cutoff mass $M_{PV} \simeq 0.56 \,\mbox{GeV}$. The $Q^2$ evolution must be taken into account in some way before comparing them with the observed nucleon structure functions at high $Q^2$. Recently, Saga group provided a Fortran program, which gives numerical solution of DGLAP evolution equations at the next-to-leading order (NLO) for the polarized as well as unpolarized structure functions of the nucleon \[32-34\]. We shall make use of their Fortran programs to evaluate the polarized distribution functions at large $Q^2$ \[33,34\]. The question here is what value we should take for the initial energy scale of this $Q^2$ evolution. Since the use of perturbative QCD below $1 \,\mbox{GeV}$ is anyhow questionable, one may take this initial energy scale $Q_{init}^2$ as an adjustable parameter, which would be fixed by adjusting the observed structure functions at high energy region. Here we have tried to see the effect of variation of $Q_{init}^2$ in a small range of $Q^2$ around the model energy scale of $M_{PV}^2
\simeq (0.56 \,\mbox{GeV})^2$. The value $Q_{init}^2 =
{(0.5 \,\mbox{GeV})}^2 = 0.25 \,\mbox{GeV}^2$ obtained from this analysis will be used throughout the following investigation. Before showing the results of $Q^2$ evolution, we want to make a short comment. One notices from the figures given so far, the distribution functions evaluated in our effective model have unphysical tails beyond $x > 1$, although they are not so significant. These unphysical tails of the theoretical distribution functions come from an approximate nature of our treatment of the soliton center-of-motion (as well as the collective rotational motion), which is essentially nonrelativistic. A simple procedure to remedy this defect was proposed by Jaffe based on the $(1 + 1)$ dimensional bag model \[35\] and recently reinvestigated by Gamberg et al. within the context of the NJL soliton model \[36\]. According to the latter authors \[36\], the effect of Lorentz contraction can effectively be taken into account by first evaluating the distribution functions in the soliton rest frame (as we are doing here) and then by using a simple analytical transformation that preserves first moments of distribution functions, as far as the $O (\Omega^0)$ contributions to the distribution functions are concerned. Such a simple relation may not be expected however if we consider the rotational motion of the soliton, which are anyhow three dimensional. In fact, a comparison with the corresponding phenomenological distribution functions seems to indicate that the above procedure based on the $(1+1)$ dimensional dynamics tends to overestimate the effect of Lorentz contraction. In the present investigation, we therefore decided not to use their procedure. Still we want distribution functions which vanish outside the range $0 < x < 1$ so that we can use the $Q^2$-evolution Fortran program provided by Saga group \[33,34\]. Since the unphysical tails of our theoretical distributions are rather small in magnitude, we are to use a simple cutoff procedure as follows. That is, we obtain modified distribution functions, which can be used as input distributions of the above Fortran program, from the original theoretical distribution functions by multiplying the $x$-dependent cutoff factor $(1 - x^{10})$. (This special cutoff factor is invented from the requirement that only the tails of the distribution functions are modified.) Fig.11 illustrates the effect of this tentative cutoff procedure. The solid curve here is the theoretical distribution function $\Delta u(x) - \Delta d(x)$ given as a sum of the $O (\Omega^0)$ and $O (\Omega^1)$ contributions. We point out that this distribution function is the worst case in the sense that the tail beyond $x = 1$ is most significant as compared with the other distributions. The dashed curve in the same figure is obtained by using the above cutoff procedure. One sees that it leaves the distribution function for $x \leq 0.7$ almost intact. Naturally, this cutoff procedure alters the values of integrals of the distribution functions, i.e. the first moments. However, it turns out that the reduction is less than $2 \,\%$ even in the above worst case. We therefore expect that the tentative nature of the above procedure hardly affects the following qualitative analyses of scale dependence of the quark distribution functions.
For the sake of comparison, we have carried out a similar evolution procedure also for the initial distributions given by the MIT bag model. The distribution functions of the (naive) MIT bag model are already known and they are given analytically as follows \[22\]. The isoscalar longitudinally polarized distribution functions is given by $$\begin{aligned}
\Delta u(x) + \Delta d(x) &=&
\frac{(M_N R) \,\omega_1}{2 \pi (\omega_1 - 1) j_0^2 (\omega_1)} \,
\biggl\{ \int_{y_{min}}^{\infty} d y y \,
\Bigl[\, t_0^2 (\omega_1, y) \nonumber \\
&+& 2 \,t_0 (\omega_1, y) \,t_1 (\omega_1, y)
\left( \frac{y_{min}}{y} \right)
+ t_1^2 (\omega_1, y) \left( 2
{\left( \frac{y_{min}}{y} \right)}^2 - 1 \right)
\Bigr] \biggr\} ,\end{aligned}$$ whereas the isoscalar transversity distribution functions is given as $$\begin{aligned}
\delta u(x) + \delta d(x) &=&
\frac{(M_NR) \,\omega_1}{2 \pi (\omega_1 - 1) j_0^2 (\omega_1)}
\biggl\{ \int_{y_{min}}^{\infty} dy y \,
\Bigl[ t_0^2 (\omega_1, y) \nonumber \\
&+& 2 \,t_0 (\omega_1, y) \,t_1 \left( \omega_1, y \right)
(\frac{y_{min}}{y})
+ t_1^2 \left( \omega_1, y \right)
(\frac{y_{min}}{y})^2 \,\Bigr] \biggr\} .\end{aligned}$$ On the other hand, the isovector distribution functions are simply related to the isoscalar ones as $$\begin{aligned}
\Delta u(x) - \Delta d(x) &=& \frac{5}{3} \,\,
[\, \Delta u(x) + \Delta d(x) \,] , \\
\delta u(x) - \delta d(x) &=& \frac{5}{3} \,\,
[\,\delta u(x) + \delta d(x) \,] .\end{aligned}$$ In eqs. (137) and (138), $M_N$ and $R$ respectively stard for the nucleon mass and the bag radius, while $\omega_n$ is the nth root of the bag eigenvalue equation as $$\tan \omega_n = - \,\frac{\omega_n}{\omega_n - 1} ,$$ and $y_{min} = | x M_N R - \omega_1|$. The function $t_l (\omega_n,y)$ is defined by $$t_l (\omega_n, y) = \int_0^1 \,j_l (u \omega_n) j_l (u y)
\,u^2 du .$$ The bag radius $R$ is only one free parameter of this simple model. In the numerical calculation, we adopt the value used by Jaffe and Ji \[22\], i.e. $$M_N R = 4.0 \,\omega_1 ,$$ where $\omega_1 \simeq 2.043$ is the lowest (dimensionless) eigenvalue of the bag equation.
To get a rough idea about the scale dependence, we show in Fig.12 and Fig.13 the theoretical polarized quark distribution functions before and after $Q^2$-evolution. Here $\Delta u(x)$ and $\delta u(x)$ in Fig.12(a) respectively stand for the longitudinal and transversity distributions for $u$-quark. In our model, the difference between the two distributions are sizable even at the initial low energy scale. A comparison with the existing and yet-to-be-obtained high energy data must be done with care, since the way of evolution of these two distributions are pretty different and the deference between the two becomes larger and larger as $Q^2$ increases. A general trend is a rapid growth of small $x$ component of the longitudinally polarized distribution due to the coupling with gluons. A similar tendency is also observed for the corresponding $d$-quark distributions shown in Fig.12(b). We can also give some predictions for the polarized antiquark distribution functions. As one can see in Fig.13, even the signs are different for the longitudinal and transversity distributions. (This is the case for both of $\bar{u}$ and $\bar{d}$ quarks.) The twist-2 spin dependent distribution functions were calculated by several authors based on various effective models of baryons \[38,39,12\]. As for the polarized quark distribution functions, the predictions of various models give more or less similar shape of distributions assuming that they take account of the dominant nature of the valence quark contribution as well as the effects of pion cloud in some effective way. The situation is quite different for the polarized [*antiquark*]{} distributions. The transversity distribution functions for the antiquarks have, for instance, been evaluated by Barone et al. within the chiral chromodielectric model \[38\]. Comparing their predictions for $\delta \bar{u} (x)$ and $\delta \bar{d} (x)$ with ours shown in Fig.13, we find that their model gives $\delta \bar{u} (x) > 0$, while ours does $\delta \bar{u} (x) < 0$. The shapes of $\delta \bar{u} (x)$ and $\delta \bar{d} (x)$ are also quite different in both models. In consideration of the fact that the polarized antiquark distributions are quite sensitive to the detailed dynamics of the model, it is very important to get precise phenomenological information for them.
Next we show in Fig.14(a) the theoretical predictions for the proton structure function $g_1^p (x,Q^2)$ at $Q^2 = 5 \,\mbox{GeV}^2$ in comparison with the corresponding experimental data given by E143 collaboration \[40\]. The theoretical curves are obtained as follows. Starting with the initial distributions $\Delta u(x) + \Delta d(x), \Delta \bar{u}(x) + \Delta \bar{d}(x)$ and $\Delta u(x) - \Delta d(x), \Delta \bar{u}(x) -
\Delta \bar{d}(x)$ or equivalently $\Delta u(x), \Delta \bar{u}(x)$ and $\Delta d(x), \Delta \bar{d}(x)$ given at $Q_{init}^2 \simeq 0.25 \,\mbox{GeV}^2$ (we assume $\Delta s(x) = \Delta \bar{s}(x) = 0$ and $\Delta g(x) = 0$ at this energy scale), we solve the NLO evolution equation to obtain the distribution functions at $Q^2 = 5 \,\mbox{GeV}^2$. These distribution functions are then convoluted with the relevant quark and gluon coefficient functions at the NLO within the framework of perturbative QCD. These procedures have been carried out for the initial distribution given by the CQSM and also by the MIT bag model. The solid and dashed curves in Fig.14(a) respectively stand for the prediction of the CQSM and that of the MIT bag model. A remarkable feature of the CQSM as compared with the MIT bag model is the enhancement of the structure function at small $x$ region, i.e. large sea quark components. One also observes that a clear peak of $g_1^p (x,Q^2)$ around $x \simeq 0.3$ predicted by the MIT bag model (a relativistic valence quark model) is not seen in the experimental structure function. On the other hand, one can say that the prediction of the CQSM reproduces qualitative feature of the observed structure function in the whole range of $x$. Fig.14(b) shows the theoretical prediction of the CQSM (solid curve) and that of the MIT bag model (dashed curve) for the neutron spin structure function $g_1^n (x,Q^2)$ in comparison with the E154 data \[41\]. One clearly sees that the neutron spin structure function $g_1^n (x,Q^2)$ predicted by the MIT bag model is negligibly small in magnitude even after evolution. We recall that at the initial energy scale the naive MIT bag model predict $g_1^n (x) = 0$, which is a necessary consequence of a model that does not properly incorporate chiral symmetry. On the other hand, the prediction of the CQSM for $g_1^n (x,Q^2)$ is seen to be large and negative especially in the small $x$ region in good agreement with the experimental observation. Then, this agreement may be regarded as a manifestation of the importance of chiral symmetry in the physics of high-energy deep-inelastic scattering.
As is widely known, the simplest but the most important quantities characterizing the quark distribution functions are the associated first moments. Here we are interested in the first moments of the longitudinally polarized distribution functions and of the transversity ones, which are respectively called the axial and tensor charges defined as $$\begin{aligned}
g_A^{(3)} &=& \int_0^1 \,\{ [ \Delta u(x) - \Delta d(x) ]
+ [ \Delta \bar{u}(x) - \Delta \bar{d}(x) ] \} \,dx ,\\
g_A^{(0)} &=& \int_0^1 \,\{ [ \Delta u(x) + \Delta d(x) ]
+ [ \Delta \bar{u}(x) + \Delta \bar{d}(x) ] \} \,dx ,\\
g_T^{(3)} &=& \int_0^1 \,\{ [ \delta u(x) - \delta d(x) ]
- [ \delta \bar{u}(x) - \delta \bar{d}(x) ] \} \,dx ,\\
g_T^{(0)} &=& \int_0^1 \,\{ [ \delta u(x) + \delta d(x) ]
- [ \delta \bar{u}(x) + \delta \bar{d}(x) ] \} \,dx .\end{aligned}$$ Before discussing the prediction of the CQSM for these quantities, it may be instructive to remember some basic properties of those. (We recall that the first calculation of the tensor charge in the CQSM was given in \[42\].) As emphasized by Jaffe and Ji \[22\], there is a remarkable difference between the axial and tensor charges originating from the charge conjugation properties of the relevant operators. For each flavor, the tensor charge counts the number of valence quarks (quarks [*minus*]{} antiquarks) of opposite transversity. Consequently, the sea quarks do not contribute to the tensor charge. (This does not necessarily means vanishing transverse polarization of antiquarks, however.) On the other hand, the axial charge counts the number of quarks [*plus*]{} antiquarks of opposite helicity. In fact, by rewriting (146) and (147) as $$\begin{aligned}
g_A^{(3)} &=& \int_0^1 \,\{ [ \Delta u(x) - \Delta d(x) ]
- [ \Delta \bar{u}(x) - \Delta \bar{d}(x) ] \} \,d x
+ 2 \int_0^1 [ \Delta \bar{u}(x) - \Delta \bar{d}(x) ] \,d x ,
\ \ \\
g_A^{(0)} &=& \int_0^1 \,\{ [ \Delta u(x) + \Delta d(x) ]
- [ \Delta \bar{u}(x) + \Delta \bar{d}(x) ] \} \,d x
+ 2 \int_0^1 [ \Delta \bar{u}(x) + \Delta \bar{d}(x) ] \,d x ,\end{aligned}$$ the first and the second terms of the above equation can respectively be interpreted as valence and sea quark contributions in the parton model. Since the sea quark degrees of freedom is absent in the nonrelativistic framework, the difference between the axial and tensor charges is purely relativistic. Still, one must clearly distinguish two types of relativistic effect. The one is dynamical effects, which generate sea quark polarization. The other is kinematical effects, which make a difference between the axial and tensor charges even though the sea quark degrees of freedom are totally neglected. The existence of this latter effect can readily be convinced by comparing the prediction of two “valence quark models”, i.e. the nonrelativistic (constituent) quark model and the MIT bag model. In fact, the non-relativistic quark model predicts $$\begin{aligned}
g_A^{(3)} &=& g_T^{(3)} \ \ = \ \ \frac{5}{3}, \\
g_A^{(0)} &=& g_T^{(0)} \ \ = \ \ 1,\end{aligned}$$ while the prediction of the MIT bag model is given by $$\begin{aligned}
g_A^{(3)} &=& \frac{5}{3} \cdot
\int (f^2 - \frac{1}{3} g^2) \,r^2 d r,
\ \ \ \ \ \ g_T^{(3)} = \frac{5}{3} \cdot
\int (f^2 + \frac{1}{3} g^2 ) \,r^2 d r ,\\
g_A^{(0)} &=& 1 \cdot \int (f^2 - \frac{1}{3} g^2) \,r^2 d r,
\ \ \ \ \ \ \,\,g_T^{(0)} = 1 \cdot
\int (f^2 + \frac{1}{3} g^2 ) \,r^2 d r ,\end{aligned}$$ where $f$ and $g$ are upper and lower components of the lowest energy quark wave functions. For a typical bag radius $R \simeq 4.0 \,\omega_1 / M_N$, which was used before, this gives $$\begin{aligned}
g_A^{(3)} &\simeq& 1.06, \ \ \ \ \ g_T^{(3)} \ \simeq \ 1.34 ,
\\
g_A^{(0)} &\simeq& 0.64, \ \ \ \ \ g_T^{(0)} \ \simeq \ 0.84 .\end{aligned}$$ As is obvious from eqs. (152) and (153), the splittings of the axial and tensor charges are due to the different sign of the lower component (p-wave) contributions \[22\]. One should however notice that there is one interesting feature shared by both of the nonrelativistic quark model and the MIT bag model. The predictions of the both models for the ratio of the isoscalar to isovector axial charges as well as the ratio of the isoscalar to isovector tensor charges are just the same : $$g_A^{(0)} / g_A^{(3)} \ = \ g_T^{(0)} / g_T^{(3)} \ = \ 3/5 .$$ Although there is no experimental information yet for the tensor charges, the above prediction for the ratio of the two axial charges obviously contradicts the EMC observation.
CQSM MIT-bag $\mbox{Lattice QCD}$ \[43\] Experiment
------------------------- --------- ------------ ----------------------------- -----------------------------
1.254 $\pm$ 0.006 \[44\]
($Q^2$-indep.)
0.31 $\pm$ 0.07 \[45\]
($Q^2$ = 10 $\mbox{GeV}^2$)
$g_T^{(3)}$ 1.22 1.34 1.07 –
$g_T^{(0)}$ 0.56 0.80 0.56 –
$g_A^{(0)} / g_A^{(3)}$ 0.25 0.60 0.18 0.24
$g_T^{(0)} / g_T^{(3)}$ 0.46 0.60 0.52 –
: The theoretical predictions for the isovector and isoscalar axial charges as well as the corresponding tesnsor charges. The predictions of the MIT bag model and those of the lattice QCD \[43\] are also shown together with some experimental data \[44,45\].
Now we shall argue that the above prediction may be interpreted as showing the limitation of simple valence quark models, which fail to properly incorporate chiral symmetry of QCD. To convince it, the predictions of the NRQM and the MIT bag model are compared with those of the CQSM in table 1, which maximally incorporate chiral symmetry. For the sake of reference, the predictions of the lattice QCD are also shown \[43\]. (Here we have omitted the errors of the lattice QCD calculation, for simplicity.) We first point out that the predictions of the CQSM for the above ratios, i.e. $$g_A^{(0)} / g_A^{(3)} \ \simeq \ 0.25 , \ \ \ \ \
g_T^{(0)} / g_T^{(3)} \ \simeq \ 0.46 ,$$ strongly deviate from the above predictions of the two valence quark models. What is remarkable here is that the CQSM predicts very small isoscalar axial charge in consistent with the EMC observation. (More meaningful comparison should be made after taking account of the scale dependence of this quantity.) Its prediction for the isovector axial charge is also qualitatively consistent with the experimental value determined from the neutron beta decay. (The deviation from the experimental value is only about 11 $\%$.) The lattice gauge theory also predicts a very small isoscalar axial charge $g_A^{(0)} \simeq 0.18$. However, this prediction may not be taken as a final one since it largely underestimates the isovector axial charge. At any rate, one can observe qualitative similarities between the predictions of the CQSM and those of the lattice QCD. Both predicts quite a small number for the ratio of the isoscalar to isovector axial charges as compared with the prediction $g_A^{(0)} / g_A^{(3)} = 0.6$ of the NRQM or the MIT bag model. On the other hand, the predictions of both models for the ratio of the isoscalar to isovector charges is not extremely different from the prediction $g_T^{(0)} / g_T^{(3)} = 0.6$ of the latter valence quark models. In our opinion, the observed deviation from the valence quark picture indicates an importance of chiral symmetry as a generator of “dynamical sea quark effect”, and the predicted feature is expected to be confirmed by future measurements of tensor charges.
To compare the theoretical first moments of the spin distribution functions with the existing data for the longitudinal case and with yet-to-be-observed ones for the transversity case, we must take account of the scale dependence of the relevant moments. As is well-known, the first moment of the isovector longitudinal distribution functions, i.e. the isovector axial charge is scale independent, i.e. it does not evolve : $g_A^{(3)} (Q^2) = g_A^{(3)} (Q_{init}^2)$. This is due to the conservation of the flavor nonsinglet axial-vector current \[51\]. This is not generally the case for the flavor singlet (isoscalar) axial charge owing to the so-called axial anomaly of QCD \[46,47\]. (Still, one can take a scheme called the chiral invariant factorization scheme in which the flavor singlet axial charge is independent of $Q^2$ \[48\]. Here, we take more standard gauge invariant factorization scheme \[49\].) In the singlet sector, the nth moments of the longitudinally polarized distribution functions are coupled with the corresponding gluon contributions. The evolution of these nth moments is governed by the anomalous dimension matrix $$\gamma^{(p)n} \equiv
\left(\begin{array}{cc}
\gamma_{qq}^{(p)n} & \gamma_{qg}^{(p)n} \\
\gamma_{gq}^{(p)n} & \gamma_{gg}^{(p)n}
\end{array} \right) .$$ where $\gamma^{(0)}$ and $\gamma^{(1)}$ are $1$- and $2$-loop contributions to the anomalous dimensions. An analytic solution to this coupled evolution equation of the NLO is given in the matrix form \[50,51\] : $$\Gamma^n (Q^2) =
\left(
\begin{array}{c}
\Delta \Sigma^n (Q^2) \\
\Delta G^n (Q^2)
\end{array}
\right) ,$$ $$\begin{aligned}
\Gamma^n (Q^2)
&=& \Biggl\{
{\left( \frac{\alpha_s (Q^2)}{\alpha_s (Q_{init}^2)}
\right)}^{\Lambda_-^n/2 \beta_0}
\Biggl[ P_{-}^n - \frac{1}{2 \beta_0}
\frac{\alpha_s (Q_{init}^2) - \alpha_s (Q^2)}{4 \pi}
\,P_{-}^n \gamma^n P_{-}^n \nonumber \\
&-& \left(
\frac{\alpha_s (Q_{init}^2)}{4 \pi} - \frac{\alpha_s (Q^2)}{4 \pi}
{\left(
\frac{\alpha_s (Q^2)}{\alpha_s (Q_{init}^2)}
\right)}^{(\lambda_+^n - \lambda_-^n)/2 \beta_0} \right)
\cdot \frac{P_-^n \gamma^n P_+^n}{2 \beta_0 +
\lambda_+^n - \lambda_-^n}
\Biggr] \nonumber \\
&+& \hspace{40mm} (+ \leftrightarrow - ) \hspace{40mm} \Biggr\}
\,\, \Gamma^n (Q_{init}^2) . \ \ \ \end{aligned}$$ Here $\alpha_s (Q^2)$ is the QCD running coupling constant at the next-to-leading order with $\overline{MS}$ scheme, $\beta_0$ and $\beta_1$ are the $1$- and $2$-loop QCD beta functions, respectively, and $$\gamma^n = \gamma^{(1)n} - \frac{\beta_1}{\beta_0} \,
\gamma^{(0)n} .$$ $P_{\pm}^n$ are $2 \times 2$ projection matrices defined by $$P_{\pm}^n = \pm \,(\gamma^{(0)n} - \lambda_{\mp}^n \,\hat{1})
/ (\lambda_+^n - \lambda_-^n) ,$$ with $\hat{1}$ being a $2 \times 2$ unit matrix and with $$\lambda_{\pm}^n = \frac{1}{2} \,\left[ \,\gamma_{qq}^{(0)n} +
\gamma_{gg}^{(0)n}
\pm \sqrt{(\gamma_{qq}^{(0)n} - \gamma_{gg}^{(0)n})^2
+ 4 \gamma_{qg}^{(0)n} \gamma_{gq}^{(0)n}} \,\,\right] ,$$ the eigenvalues of the $1-$loop anomalous dimension matrix $\gamma^{(0)n}$. Since the necessary anomalous dimension matrices are all given in \[51\], it is easy to calculate the $Q^2$ evolution of the first moment of the flavor singlet longitudinally polarized distribution functions, i.e. the isosinglet axial-charge.
Because of its chiral-odd nature, the moments of the transversity distributions do not couple with gluons, irrespective of the flavor quantum numbers, which especially means that isovector and isoscalar tensor charges follow the same evolution equation. The anomalous dimension of the transversity distribution at the leading $1-$loop order was first given by Artru and Mekhfi \[52\], while the corresponding $2-$loop contributions have recently been given by three groups independently \[53-55\]. Once the relevant anomalous dimensions are known, it is easy to obtain an analytical solution of the NLO evolution equation for the nth moment of transversity distribution. Here, we use the form given by Hayashigaki et al. \[54\] as $$\frac{\delta q_1^{(n)} (Q^2)}{\delta q_1^{(n)} (Q_{init}^2)}
= {\left(
\frac{\alpha_s (Q^2)}{\alpha_s (Q_{init}^2)}
\right)}^{\gamma_h^{(0)n} / 2
\beta_0}
{\left( \frac{\beta_0 + \beta_1 \frac{\alpha_s (Q^2)}{4 \pi}}
{\beta_0 + \beta_1
\frac{\alpha_s (Q_{init}^2)}{4 \pi}}\right)}^{\frac{1}{2}
\bigl( \frac{\gamma_h^{(1)n}}{\beta_1}
- \frac{\gamma_h^{(0)n}}{\beta_0} \bigr)}$$ where the relevant anomalous dimensions $\gamma_h^{(0)n}$ and $\gamma_h^{(1)n}$ are all given in \[51\]. Fig.15 show the calculated $Q^2$ dependence of the axial and tensor charges. For obtaining it, we start with the theoretical first moments given at the initial energy scale $Q_{init}^2 = 0.25 \mbox{GeV}$ : $$\begin{aligned}
g_A^{(3)} (Q_{init}^2) &=& 1.41 \\
g_A^{(0)} (Q_{init}^2) &\equiv& \Delta
\Sigma (Q_{init}^2) \ = \ 0.35 \\
\Delta G (Q_{init}^2) &=& 0 \\
g_T^{(3)} (Q_{init}^2) &=& 1.22 \\
g_T^{(0)} (Q_{init}^2) &=& 0.56\end{aligned}$$ One sees that the $Q^2$ dependence of the flavor singlet axial charge is very small (it is almost constant except in the very low $Q^2$ region). A characteristic prediction of the CQSM for the axial charges, i.e. large isovector charge and small isoscalar charge appears to be qualitatively consistent with the corresponding experimental data at the relevant energy scale. As was pointed out by many authors \[38,52-55\], the $Q^2$ dependence of the tensor charges are sizably large. Although there is no experimental information for these latter quantities, this $Q^2$ dependence must be taken seriously when comparing the theoretical prediction of low energy models with future experimental date. (Note however that that the ratio $g_T^{(0)} / g_T^{(3)}$ is $Q^2$ independent.)
Because of the coupling between the flavor singlet axial charge (the longitudinal quark polarization) and the gluon polarization in the evolution equation, non-zero gluon polarization appears at high $Q^2$ even if we have assumed $\Delta G = 0$ at the initial energy scale of $Q_0^2 = 0.25 \,\mbox{GeV}^2$. We show in Fig.16 the $Q^2$ evolution of $\Delta G$ in comparison with that of $\Delta \Sigma = g_A^{(0)}$. One sees that the gluon polarization rapidly grows with increasing $Q^2$. Already at $Q^2 \simeq 2 \,\mbox{GeV}^2$, $\Delta G$ is seen to be larger than $\Delta \Sigma$. As explained in \[49\], the growth of the gluon polarization with $Q^2$ can be traced back to the positive sign of the anomalous dimension $\gamma_{qg}^{(0)1}$ at the leading order ($\gamma_{qg}^{(0)1} = 2$). The positivity of this quantity means that a polarized quark is preferred to radiate a gluon with helicity parallel to the quark polarization. Since the net quark spin component in the proton is positive, it follows that $\Delta G > 0$ at least for the gluons perturbatively emitted from quarks \[49\]. It is hoped that the direct information on $\Delta g(x,Q^2)$ from the di-jet asymmetry analyses at HERA in conjunction with the precise NLO analyses of $g_1 (x,Q^2)$ will soon provide us with an accurate determination of the polarized gluon distribution as well as its first moment \[56\].
Summary
=======
In summary, we have shown that the CQSM naturally explains qualitative behavior of the experimentally measured longitudinally polarized structure functions of the proton and the neutron. As was shown in our previous papers, the model also reproduces an excess of $\bar{d}$ sea over the $\bar{u}$ sea in the proton very naturally \[17-19\]. Furthermore, it predicts qualitative difference between the transversity distribution functions and longitudinally polarized distribution functions. For example, in simple valence quark models like the NRQM or the MIT bag model, the ratios of the isoscalar to isovector charges are just the same for both of the axial charges and the tensor charges. On the contrary, in the CQSM or in the lattice gauge theory, this ratio turns out to be much smaller for the axial charges than for the tensor charges. In our viewpoint, what makes this difference is “dynamical sea quark effects” dictated by the spontaneous chiral symmetry breaking of the QCD vacuum. Another noteworthy prediction of the CQSM is the opposite (spin) polarization of the $\bar{u}$ and $\bar{d}$ sea quarks, thereby indicating SU(2) asymmetric sea quark polarization. These observations then indicates that nonperturbative QCD dynamics due to the spontaneous chiral symmetry breaking would [*survive*]{} and manifest itself in the isospin (or flavor) dependence of high energy spin observables, especially in that of the polarized (as well as unpolarized) [*antiquark*]{} distribution functions.
Acknowledgement {#acknowledgement .unnumbered}
===============
The authors would like to express their gratitude to T. Watabe at Ruhr Universität Bochum for useful discussion on the importance of nonlocality corrections in time. Numerical calculation was performed by using the workstations at the Laboratory of Nuclear Studies, and those at the Research Center for Nuclear Physics, Osaka University.
References {#references .unnumbered}
==========
[\[\]]{}
EMC Collab., J. Aschman et al., Phys. Lett. [**B206**]{}, 364 (1988) ;\
Nucl. Phys. [**B328**]{}, 1 (1989).
F.E Close, [*An Introduction to Quarks and Partons*]{} (Academic Press, London, 1979) ;\
T. Muta, [*Foundations of Quantum Chromodynamics*]{} (World Scientific, Singapore, 1987).
M. Gockeler, H. Oelrich, P.E.L. Rakow, G.Schierholz, R. Horsley, E.M. Ilgenfritz, H. Perlt and A. Schiller, J. Phys. [**G22**]{}, 703 (1996) ;\
M.Gockeler, R. Horsley, L. Mankiewicz, H. Perlt, P. Rakow, G. Schierholz and A. Schiller, Phys. Lett. [**B414**]{}, 340 (1997) ;\
C. Best, M. Gockeler, R. Horsley, L. Mankiewicz, H. Perlt, P. Rakow, A. Schafer, G. Schierholz, A. Schiller, S. Schramm and P. Stephenson, hep-ph / 9706502.
D.I. Diakonov, V.Yu. Petrov and P.V. Pobylista, Nucl. Phys. [**B306**]{}, 809 (1988).
M. Wakamatsu and H. Yoshiki, Nucl. Phys. [**A524**]{}, 561 (1991).
For reviews, see, M. Wakamatsu, Prog. Theor. Phys. Suppl. [**109**]{}, 115 (1992) ;\
Chr.V. Christov, A. Blotz, H.-C. Kim, P. Pobylitsa, T. Watabe, Th. Meissner,\
E. Ruiz Arriola and K. Goeke, Prog. Part. Nucl. Phys. [**37**]{}, 91 (1996) ;\
R. Alkofer, H.Reinhardt and H. Weigel, Phys. Rep. [**265**]{}, 139 (1996).
A. Bramon, Riazuddin and M.D. Scadron, J. Phys. [**G24**]{}, 1 (1998).
M. Wakamatsu, Phys. Lett. [**B300**]{}, 152 (1993).
J.D. Sullivan, Phys. Rev. [**D5**]{}, 1732 (1972).
E.M. Henley and G.A. Miller, Phys. Lett. [**B251**]{}, 453 (1990).
S. Kumano, Phys. Rev. [**D43**]{}, 59 (1991) ;\
S. Kumano and J.T. Londergan, Phys. Rev. [**D44**]{}, 717 (1991).
H. Weigel, L. Gamberg and H. Reinhardt, Mod. Phys. Lett. [**A11**]{}, 3021 (1996) ;\
Phys. Lett. [**B399**]{}, 287 (1997) ;\
L. Gamberg, H. Reinhardt and H. Weigel, Phys. Rev. [**D58**]{}, 054014 (1998).
D.I. Diakonov, V.Yu. Petrov and P.V. Pobylista, M.V. Polyakov and C. Weiss,\
Nucl. Phys. [**B480**]{}, 341 (1996).
D.I. Diakonov, V.Yu. Petrov and P.V. Pobylista, M.V. Polyakov and C. Weiss,\
Phys. Rev. [**D56**]{}, 4069 (1997).
K. Tanikawa and S. Saito, Nagoya Univ. preprint, DPNU-96-37 (1996).
NMC Collab., P. Amaudruz et al., Phys. Rev. Lett. [**66**]{}, 2712 (1991).
M. Wakamatsu and T. Kubota, Phys. Rev. [**D57**]{}, 5755 (1998).
M. Wakamatsu, Phys. Rev. [**D44**]{}, R2631 (1991) ; Phys. Lett. [**B269**]{}, 394 (1991) ;\
Phys. Rev. [**D46**]{}, 3762 (1992).
M. Wakamatsu, in [*Weak and Electromagnetic Interactions in Nucei (WEIN-92)*]{},\
Proceeding of the International Seminar, Dubna, Russia, 1992, edited by\
Ts.D. Vylov (World Scientific, Singapore, 1993).
P.V. Pobylitsa, M.V. Polyakov, K. Goeke, T. Watabe and C. Weiss, hep-ph / 9804436.
J.C. Collins and D.E. Soper, Nucl. Phys. [**B194**]{}, 445 (1982) ;\
J. Kogut and D. Soper, Phys. REv. [**D1**]{}, 2901 (1970).
R.L. Jaffe and X. Ji, Nucl. Phys. [**B375**]{}, 527 (1992).
R.L. Jafe, [*Spin, Twist, and Hadron Structure in Deep Inelastic Procceses*]{}, Lectures given at Ettore Majorana International School of Nucleon Structure (1995) : hep-ph / 9602236.
M. Wakamatsu, Prog. Theor. Phys. [**95**]{}, 143 (1996).
D.I. Diakonov and V.Yu. Petrov, Nucl. Phys. [**B272**]{}, 457 (1986).
C. Weiss and K. Goeke, hep-ph / 9712447.
S. Kahana and G. Ripka, Nucl. Phys. [**A429**]{}, 462 (1984) ;\
S. Kahana, G. Ripka and V. Soni, Nucl. Phys. [**A415**]{}, 351 (1984).
F. Döring, A. Blotz, C. Schüren, T. Meissner, E. Ruiz-Arriola and K. Goeke,\
Nucl. Phys. [**A415**]{}, 351 (1984).
M. Wakamatsu and T. Watabe, Phys. Lett. [**B312**]{}, 184 (1993).
Chr.V. Christov, A. Blotz, K. Goeke, P. Pobylitsa, V.Yu. Petrov, M. Wakamatsu\
and T. Watabe, Phys. Lett. [**B325**]{}, 467 (1994).
M. Glück, E. Reya, M. Stratmann and W. Vogelsang, Phys. Rev. [**D53**]{}, 4775 (1996).
M. Miyama and S. Kumano, Comput. Phys. Commun. [**94**]{}, 185 (1996).
M. Hirai, S. Kumano and M. Miyama, Comput. Phys. Commun. [**108**]{}, 38 (1998).
M. Hirai, S. Kumano and M. Miyama, Comput. Phys. Commun. [**111**]{}, 150 (1998).
R.L. Jaffe, Phys. Lett. [**B93**]{}, 313 (1980) ; Ann. Phys. (NY) [**132**]{}, 32 (1981).
L. Gamberg, H.Reinhardt and H.Weigel, hep-ph/9707352.
M. Gück, E. Reya and A. Vogt, Z. Phys. [**C67**]{}, 433 (1995).
V. Barone, T. Calarco and A. Drago, Phys. Lett. [**B390**]{}, 287 (1997).
K. Suzuki and T. Shigetani, Nucl. Phys. [**A626**]{}, 886 (1997) ;\
K. Suzuki and W. Weise, Nucl. Phys. [**A634**]{}, 141 (1998).
E143 Collab., K. Abe et al., hep-ph / 9802357.
E154 Collab., K. Abe et al., Phys. Rev. Lett. [**79**]{}, 26 (1997).
H.-C. Kim, M. Polyakov and K. Goeke, Phys. Lett. [**B387**]{}, 577 (1996).
Y. Kuramashi, Nucl. Phys. [**A629**]{}, 235c (1998) ;\
See, also, M. Fukugita, Y. Kuramashi, M. Okawa and A. Ukawa,\
Phys. Rev. Lett. [**75**]{}, 2092 (1995) ;\
S.-J. Dong, J.-F. Lagaë and K.-F. Liu, Phys. Rev. Lett. [**75**]{}, 2096 (1995) ;\
S. Aoki, M. Doui, T. Hatsuda and Y. Kuramashi, Phys. Rev. [**D56**]{}, 433 (1997).
C. Caso et al. (Particle Data Group), Europian Physical Journal [**C3**]{}, 1 (1998).
J. Ellis and M. Karliner, Phys. Lett. [**B341**]{}, 397 (1995).
G. Altarelli and G.G. Ross, Phys. Lett. [**B212**]{}, 391 (1988).
R.D. Carlitz, J.C. Collins and A.H. Mueller, Phys. Lett. [**B214**]{}, 229 (1988).
G.T. Bodwin and J. Qiu, Phys. Rev. [**D41**]{}, 2755 (1990).
For review, see, H.-Y. Cheng, [*Status of the Proton Spin Problem*]{}, Lectures given at\
Xth Spring School on Particle and Fields (1996) : hep-ph / 9607254.
W. Furmanski and R. Petronzio, Z. Phys. [**C11**]{}, 293 (1982).
M. Glück, E. Reya and A. Vogt, Z. Phys. [**C48**]{}, 471 (1990).
X. Artru and M. Mekhfi, Z. Phys. [**C45**]{}, 669 (1990).
S. Kumano and M. Miyama, Phys. Rev. [**D56**]{}, 2504 (1997).
A. Hayashigaki, Y. Kanazawa and Y. Koike, Phys. Rev. [**D56**]{}, 7350 (1997).
W. Vogelsang, Phys. Rev. [**D57**]{}, 1886 (1998).
A. De Roeck, A. Deshpande, V.W. Huges, J. Lichtenstadt and G. Rädel,\
hep-ph / 9801300.
**[Figure caption]{}\
**
\
Fig. 1.
The $O (\Omega^0)$ contributions to the isovector longitudinally polarized distribution functions $\Delta u(x) - \Delta d(x)$ (solid curves) and $\Delta \bar{u} (x) - \Delta \bar{d} (x)$ (dashed curves). Here the three figures (a), (b), and (c) correspond to the contributions of the discrete valence level, those of the Dirac sea quarks, and their sums, respectively.
\
\
Fig. 2.
The $O (\Omega^1)$ contributions to the isovector longitudinally polarized distribution functions $\Delta u(x) - \Delta d(x)$ (solid curves) and $\Delta \bar{u} (x) - \Delta \bar{d} (x)$ (dashed curves). The meaning of the three figures (a), (b), and (c) is the same as in Fig.1.
\
\
Fig. 3.
The $O (\Omega^1)$ contributions to the isoscalar longitudinally polarized distribution functions $\Delta u(x) + \Delta d(x)$ (solid curves) and $\Delta \bar{u} (x) + \Delta \bar{d} (x)$ (dashed curves). The meaning of the three figures (a), (b), and (c) is the same as in Fig.1.
\
\
Fig. 4.
The final predictions of the CQSM for the longitudinally polarized distribution functions $\Delta u(x) - \Delta d(x)$ and $\Delta \bar{u} (x) - \Delta \bar{d} (x)$ given as the sums of the $O (\Omega^0)$ and $O (\Omega^1)$ contributions, in comparison with those for the isoscalar longitudinally polarized distribution functions $\Delta u(x) + \Delta d(x)$ and $\Delta \bar{u} (x) + \Delta \bar{d} (x)$ comong from the $O (\Omega^1)$ terms.
\
\
Fig. 5.
The theoretical predictions for the longitudinally polarized distribution functions, $x (\Delta u(x) + \Delta \bar{u} (x) +
\Delta d(x) + \Delta \bar{d} (x))$ and $x (\Delta u(x) + \Delta
\bar{u} (x) - \Delta d(x) - \Delta \bar{d} (x))$, are compared with the corresponding semi-phenomenological parametrization given by Glück, Reya, Stratmann and Vogelsang \[31\]. Of the two theoretical curves in each figure, the solid curve is the answer of the present calculation, whereas the dashed curve is obtained by using the old treatment used in \[17\], which amounts to dropping some of the nonlocality effects in time.
\
\
Fig. 6.
The predictions of the CQSM for the polarized antiquark distributions $x \Delta \bar{u} (x)$ and $x \Delta \bar{d} (x)$ are compared with the corresponding GRSV parametrization, which assumes SU(2) symmetric sea quark polarization, $x \Delta \bar{u} (x) = x \Delta \bar{d} (x) \,\,
( \equiv x \Delta \bar{q} (x) )$.
\
\
Fig. 7.
The $O (\Omega^0)$ contributions to the isovector transversity distribution functions $\delta u(x) - \delta d(x)$ (solid curves) and $\delta \bar{u} (x) - \delta \bar{d} (x)$ (dashed curves). Here the three figures (a), (b), and (c) correspond to the contributions of the discrete valence level, those of the Dirac sea quarks, and their sums, respectively.
\
\
Fig. 8.
The $O (\Omega^1)$ contributions to the isovector transversity distribution functions $\delta u(x) - \delta d(x)$ (solid curves) and $\delta \bar{u} (x) - \delta \bar{d} (x)$ (dashed curves). The meaning of the three figures (a), (b), and (c) is the same as in Fig.1.
\
\
Fig. 9.
The $O (\Omega^1)$ contributions to the isoscalar transversity distribution functions $\delta u(x) + \delta d(x)$ (solid curves) and $\delta \bar{u} (x) + \delta \bar{d} (x)$ (dashed curves). The meaning of the three figures (a), (b), and (c) is the same as in Fig.1.
\
\
Fig. 10.
The final predictions of the CQSM for the transversity distribution functions $\delta u(x) - \delta d(x)$ and $\delta \bar{u} (x) - \delta \bar{d} (x)$ given as the sums of the $O (\Omega^0)$ and $O (\Omega^1)$ contributions, in comparison with those for the isoscalar transversity distribution functions $\delta u(x) + \delta d(x)$ and $\delta \bar{u} (x) + \delta \bar{d} (x)$ coming from the $O (\Omega^1)$ terms.
\
\
Fig. 11.
The solid curve represents the theoretical distribution functions $\Delta u(x) - \Delta d(x)$, whereas the dashed curve is a modified one obtained from it by multiplying a $x$-dependent cutoff factor $(1 - x^{10})$.
\
\
Fig. 12.
The theoretical predictions for the twist-2 spin dependent quark distribution functions before and after $Q^2$-evolution. Here $\Delta u(x)$ and $\delta u(x)$ (in (a)) respectively stand for the longitudinal and transversity distributions of $u$-quark, while $\Delta d(x)$ and $\delta d(x)$ (in (b)) are the corresponding quantities for $d$-quark.
\
\
Fig. 13.
The theoretical predictions for the twist-2 spin dependent antiquark distribution functions before and after $Q^2$-evolution. Here $\Delta \bar{u} (x)$ and $\delta \bar{u} (x)$ (in (a)) respectively stand for the longitudinal and transversity distributions of $\bar{u}$-quark, while $\Delta \bar{d} (x)$ and $\delta \bar{d} (x)$ (in (b)) are the corresponding quantities for $\bar{d}$-quark.
\
\
Fig. 14.
The theoretical predictions for the proton and neutron spin structure functions $g_1^p (x,Q^2)$ and $g_1^n (x,Q^2)$ at $Q^2 = 4 \,\mbox{GeV}^2$ in comparison with the corresponding SLAC data. The solid and dashed curves in (a) respectively stand for the prediction of the CQSM and that of the naive MIT bag model for $g_1^p (x,Q^2)$, whereas the black circles are the E143 data. The corresponding theoretical predictions for the $g_1^n (x,Q^2)$ are shown in (b) together with the E154 data.
\
\
Fig. 15.
The scale dependence of the axial and tensor charges. The evolution equations at the next-to-leading order are solved under the initial conditions $g_A^{(3)} (Q_{init}^2) = 1.41$, $g_A^{(0)} (Q_{init}^2) \equiv \Delta \Sigma (Q_{init}^2) = 0.35$, $g_T^{(3)} (Q_{init}^2) = 1.22$, $g_T^{(0)} (Q_{init}^2) = 0.56$, and $\Delta G (Q_{init}^2) = 0$ at $Q_{init}^2 = 0.25 \,\mbox{GeV}^2$.
\
\
Fig. 16.
The scale dependence of the flavor singlet axial charge (or the quark polarization) and the gluon polarization. The initial conditions for the evolution equation is the same as given in Fig.15.
[^1]: Email : [email protected]
[^2]: Email : [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We test the possible deviation of the cosmic distance duality relation $D_A(z)(1+z)^2/D_L(z)\equiv 1$ using the standard candles/rulers in a fully model-independent manner. Type-Ia supernovae are used as the standard candles to derive the luminosity distance $D_L(z)$, and ultra-compact radio sources are used as the standard rulers to obtain the angular diameter distance $D_A(z)$. We write the deviation of distance duality relation as $D_A(z)(1+z)^2/D_L(z)=\eta(z)$. Specifically, we use two parameterizations of $\eta(z)$, i.e. $\eta_1(z)=1+\eta_0 z$ and $\eta_2(z)=1+\eta_0 z/(1+z)$. The parameter $\eta_0$ is obtained using the Markov chain Monte Carlo methods by comparing $D_L(z)$ and $D_A(z)$ at the same redshift. The best-fitting results are $\eta_0=-0.06\pm 0.05$ and $-0.18\pm 0.16$ for the first and second parameterizations, respectively. Our results depend on neither the cosmological models, nor the matter contents or the curvature of the universe.'
date: 'Accepted xxxx; Received xxxx; in original form xxxx'
title: 'Testing the distance duality relation using type-Ia supernovae and ultra-compact radio sources'
---
\[firstpage\]
cosmological parameters – distance scale – supernovae: general
Introduction
============
In the standard cosmological model, there is a strict correlation between the luminosity distance $D_L(z)$ and the angular diameter distance $D_A(z)$ at the same redshift, i.e. $D_A(z)(1+z)^2/D_L(z)=1$ [@Etherington:1933; @Etherington:2007]. This is the so-called distance duality relation (DDR). The DDR holds true in any metric theory of gravity such as general relativity, as long as the photons travel along null geodesics and the photon number is conserved. Any deviation of DDR implies that there are new physics beyond the standard cosmological model. Therefore, testing the validity of DDR arouses great interests in recent years.
Many work have been devoted to testing the validity of DDR [@Bassett:2003vu; @Uzan:2004my; @Bernardis:2006; @Holanda:2010vb; @Piorkowska:2011nhd; @Fu:2011; @Yang:2013coa; @Costa:2015lja; @Lv:2016mmq; @Holanda:2016msr; @Ma:2016bjt; @Liao:2016uzb; @Chen:2016; @Holanda:2016zpz]. All the methods require the measurement of both luminosity distance and angular diameter distance at the same redshift. The luminosity distance $D_L$ can be obtained from type-Ia supernovae (SNe) with high precision. Type-Ia SNe have an approximately consistent absolute luminosity after correcting for stretch and color, and therefore are widely regarded as the standard candles in cosmology [@Riess:1998mnb; @Perlmutter:1998np]. However, the measurement of angular diameter distance $D_A$ is not as straightforward as that of $D_L$. One way to determine $D_A$ is using the Sunyaev–Zeldovich effect combined with the $X$-ray data from galaxy clusters [@Filippis:2005; @Bonamente:2006ct]. The obtained $D_A$ in this way depends on the mass model of galaxy clusters, which arouses large uncertainties. Many work have used the $D_L$ data from SNe and $D_A$ data from galaxy clusters to test the DDR, see e.g., @Yang:2013coa and the references therein. Due to the large uncertainty, most work found no evidence for the violation of DDR. A more accurate measurement of $D_A$ can be obtained from the baryon acoustic oscillations (BAO) [@Beutler:2011hx; @Anderson:2013zyy; @Kazin:2014qga; @Delubac:2014aqe]. However, the measurement of BAO requires the statistics of a large number of galaxies, and the number of measured BAO data points so far is very limited. @Ma:2016bjt used the $D_L$ from SNe and $D_A$ from BAO to test DDR and found 5% constraints in favor of its validity.
Recently, @Liao:2016uzb proposed to test the DDR using the angular diameter distance from strong gravitational lensing systems, in combination with the luminosity distance from type-Ia SNe. The angular diameter distance can be deduced from the Einstein radius, and the redshifts of lens and source. However, the gravitation lensing systems could only provide the information of distance ratio between lens and source, and between observer and source, i.e. $R_A\equiv D_{A,ls}/D_{A,s}$. To obtain the distance from observer to lens, a flat FLRW cosmology was assumed, which makes the test of DDR not completely model independent. What’s more, the Einstein radius depends on the mass profile of lens, which causes some uncertainty. The luminosity distance is from type-Ia SNe, which is independent of the gravitational lensing systems. The problem is that the SNe and gravitational lensing systems are usually located at different redshifts, making the direct comparison between $D_L$ and $D_A$ impossible. To solve this problem, the authors adopted a matching criterion that if the redshift different between the SNe and lens or source is no more than 0.005, them may be regarded as locating at the same redshift. With such a criterion, there are only 60 or so lensing systems which have matched SNe, although the total number of lensing systems is more than one hundred. The redshift of type-Ia SNe is limited to be $\lesssim 1.4$, much smaller than the redshift of strong gravitation lensing systems. @Holanda:2016msr combined the SNe, strong gravitational lensing systems and gamma-ray bursts to test the DDR at high redshift, and found that the DDR validity is verified within $1.5\sigma$.
The milliarcsecond ultra-compact radio sources (RSs) provide a unique tool to measure the angular diameter distance. @Kellermann:1993mki studied the angular size – redshift relation ($\theta-z$ relation) of 79 ultra-compact RSs associated with active galaxies and quasars observed by the very-long-baseline interferometry (VLBI), and showed that the $\theta-z$ relation can be naturally explained by geometrical effect (i.e. the further object has the smaller angular size) in the FLRW cosmology. This implies that the linear size of ultra-compact RSs, $d_0$, is approximately constant, free of evolution effects. From then on, much efforts have been done to standardize the ultra-compact RSs as the cosmological rulers [@Jackson:1996stj; @Jackson:1997ib; @Gurvits:1999hs; @Gurvits:1994fgs; @Jackson:2004jw; @Jackson:2006bg; @Jackson:2008ak; @Jackson:2012jt]. The possible cosmological evolution of $d_0$ with luminosity $L$, redshift $z$, and spectra index $\alpha$ has been investigated [@Gurvits:1999hs; @Gurvits:1994fgs; @Jackson:2004jw]. With appropriate cutoffs on $L$, $z$ and $\alpha$, the cosmological evolution of $d_0$ can be neglected [@Gurvits:1994fgs; @Gurvits:1999hs]. Assuming that there is no cosmological evolution of $d_0$, @Jackson:2006bg used a large sample of RSs data to give a very strict constraint on cosmological parameters.
In this paper, we use the angular diameter distance from ultra-compact RSs, and combine with the luminosity distance from type-Ia SNe, to test the validity of DDR. The main advantage of this method is that it is completely cosmological model-independent. The rest of the paper is organized as follows: In section \[sec:method\], we introduce the methodology of testing DDR using type-Ia SNe and ultra-compact RSs data in a model-independent way. In section \[sec:results\], we introduce the observational data samples that are used in the test of DDR and give the results. Finally, discussions and conclusions are given in section \[sec:conclusions\].
Methodology {#sec:method}
===========
In this section, we illustrate how to test the DDR using type-Ia SNe and ultra-compact RSs. Following @Holanda:2010vb, we write the possible violation of the standard DDR as $$\label{eq:modified_DDR}
\frac{D_A(z)(1+z)^2}{D_L(z)}=\eta(z).$$ We use two different parameterizations of $\eta(z)$, $$\eta_1(z)=1+\eta_0z,~~~~\eta_2(z)=1+\eta_0\frac{z}{1+z},$$ where $\eta_0$ is a free parameter representing the amplitude of DDR violation. There is no violation of DDR if $\eta_0=0$. By comparing $D_L(z)$ and $D_A(z)$ at the same redshift $z$, we can constrain the parameter $\eta_0$.
The luminosity distance can be derived from type-Ia SNe. Type-Ia SNe are widely used as the standard candles in cosmology due to their approximately consistent absolute luminosity [@Riess:1998mnb; @Perlmutter:1998np]. The distance modulus of SNe can be extracted from the light curves using the following empirical relation [@Tripp:1998; @Guy:2005; @Guy:2007; @Suzuki:2012dhd; @Betoule:2014frx] $$\label{eq:mu_sn}
\mu_{\rm sn}=m_B^*-M_B+\alpha X_1-\beta \mathcal{C}.$$ where $m_B^*$ is the apparent magnitude, $M_B$ is the absolute magnitude, $X_1$ and $\mathcal{C}$ are the stretch factor and color parameter, respectively. The two parameters $\alpha$ and $\beta$ are universal constants and can be fitted simultaneously with cosmological parameters, or more generally, can be marginalized over. The uncertainty of $\mu_{\rm sn}$ is propagated from the uncertainties of $m_B^*$, $X_1$ and $\mathcal{C}$ using the standard error propagation formula, $$\sigma_{\mu_{\rm sn}}=\sqrt{\sigma_{m_B^*}^2+\alpha^2\sigma_{X_1}^2+\beta^2\sigma_{\mathcal{C}}^2}.$$
The angular diameter distance is derived from the ultra-compact RSs. Due to the approximately constant linear size of RSs, the angular diameter distance can be easily obtained if the angular size $\theta$ is observed, $$\label{eq:DA}
D_{A,\rm rs}=\frac{d_0}{\theta},$$ where the linear size $d_0$ is not known a prior and is regarded as a free parameter. Combining Eqs.(\[eq:modified\_DDR\]) and (\[eq:DA\]), the luminosity distance of RSs can be written as $$\label{eq:DL_rs}
D_{L,\rm rs}=\frac{d_0}{\theta}\frac{(1+z)^2}{\eta(z)}.$$ The distance modulus of RSs is given by $$\label{eq:mu_rs}
\mu_{\rm rs}=5\log_{10}\frac{D_{L,{\rm rs}}}{\rm Mpc}+25.$$ The uncertainty of $\mu_{\rm rs}$ is propagated from that of $\theta$, $$\label{eq:mu_rs_error}
\sigma_{\mu_{\rm rs}}=\frac{5}{\ln 10}\frac{\sigma_{\theta}}{\theta},$$ where ‘ln’ is the natural logarithm.
Difficulties arise when we try to directly compare $\mu_{\rm sn}$ with $\mu_{\rm rs}$. This is because SNe and RSs usually locate at different redshift. For a specific SN, there is in general no RS locating at the same redshift, and vice visa. To solve this problem, we first reconstruct the $\theta-z$ relation using the Gaussian processes [@Seikel:2012uu]. Then the $\mu_{\rm rs}(z)$ function can be reconstructed using Eqs.(\[eq:DL\_rs\]) – (\[eq:mu\_rs\_error\]). Given $\mu_{\rm rs}(z)$ function, we can calculate the distance modulus of RSs at any desired redshift. We use the reconstructed $\mu_{\rm rs}(z)$ function to calculate the distance modulus of RS at the redshift of each SNe. Therefore, the comparison between $\mu_{\rm rs}$ and $\mu_{\rm sn}$ at the same redshift is possible.
We use the Markov chain Monte Carlo methods [@ForemanMackey:2012ig] to calculate the posterior probability distribution functions of free parameters. The likelihood is given by $$\label{eq:likelihood}
\mathcal{L(\mathbf p)}=\prod\frac{1}{\sqrt{2\pi}\sigma_\mu}\exp\left[-\frac{1}{2}\left(\frac{\mu_{\rm sn}-\mu_{\rm rs}}{\sigma_\mu}\right)^2\right],$$ where $$\label{eq:dmu_total}
\sigma_\mu=(\sigma_{\mu_{\rm sn}}^2+\sigma_{\mu_{\rm rs}}^2+\sigma_{\rm int}^2)^{1/2}$$ is the uncertainty of distance modulus of SNe and RSs, $\mathbf{p}=(d_0,\eta_0,\alpha,\beta, M_B, \sigma_{\rm int})$ is the set of free parameters, and the product runs over all the SN-RS pairs. We have added the intrinsic scatter term in Eq.(\[eq:dmu\_total\]) to account for any other uncertainties. Note that $d_0$ is degenerated with $M_B$, so they cannot be constrained simultaneously. One parameter should be fixed in order to constrain the other. We leave $d_0$ free and fix $M_B=-19.32$ [@Suzuki:2012dhd].
Data and Results {#sec:results}
================
The SNe sample is taken from @Suzuki:2012dhd, i.e. the Union2.1 sample. Union 2.1 consists of 580 type-Ia SNe with high light curve quality in the redshift range $[0.015,1.414]$. Each SN has well measured redshift $z$, apparent magnitude $m_B^*$, stretch factor $X_1$ and color factor $\mathcal{C}$. Note that @Suzuki:2012dhd also published the distance modulus of each SNe, which is calibrated in the $w$CDM model. We use the original light curve parameters rather than the published distance moduli in order to avoid the model dependence. The distance moduli are calculated using Eq.(\[eq:mu\_sn\]), with $\alpha$ and $\beta$ as free parameters.
The RSs sample is taken from @Jackson:2006bg. The sample is selected from a complication of VLBI survey of ultra-compact RSs at 2.29 GHz released by @Preston:1985bgd. The redshift and radio flux are updated according to the recent observations. The sample consists of 613 RSs in the redshift range $[0.0035,3.787]$. Among the sample, there are 468 RSs which have redshift larger than 0.5. Data with $z<0.5$ are not appropriate to use as standard rulers because they are more affected by cosmological evolutions [@Gurvits:1994fgs; @Jackson:2004jw; @Jackson:2006bg; @Jackson:2008ak; @Jackson:2012jt]. The 468 RSs data in the $\theta-z$ plane are plotted in Figure \[fig:logTheta\].
![[]{data-label="fig:logTheta"}](logTheta.eps){width="50.00000%"}
To show that the RSs data can be used to constrain the cosmological parameters, we fit the 468 RSs data points to the standard $\Lambda$CDM model. We get the best-fitting parameters $\Omega_M=0.29\pm 0.10$, $d_0=7.76\pm 0.57~h_0^{-1}$ pc. Here $h_0$ is the Hubble constant in unit of $100~{\rm km}~{\rm s}^{-1}~{\rm Mpc}^{-1}$. Although has a large uncertainty, the $\Omega_M$ value is well consistent with the much more precise result from CMB observations [@Ade:2015xua].
From Figure \[fig:logTheta\], we can see that the original data points show large scatter around the $\theta-z$ theoretical curve. Following Refs. [@Kellermann:1993mki; @Jackson:1996stj; @Jackson:1997ib; @Gurvits:1999hs; @Gurvits:1994fgs; @Jackson:2004jw; @Jackson:2006bg], we bin the raw data into 18 bins, with 26 data points in each bin. The central value and $1\sigma$ uncertainty are taken to be the mean and standard deviation of data points in each bin, respectively. The binned data are plotted in Figure \[fig:GP\_Theta\_z\].
![[]{data-label="fig:GP_Theta_z"}](GP_Theta_z.eps){width="50.00000%"}
Based on the binned data, we use the publicly available python package [@Seikel:2012uu] to reconstruct the $\theta(z)$ function in the redshift range $[0.5, 3.787]$. The Gaussian processes only depends on the covariance function. Here we adopt the most widely used squared exponential covariance function. The results are plotted in Figure \[fig:GP\_Theta\_z\]. The blue range stands for the $1\sigma$ uncertainty. In the reconstruction, we use $\log\theta$ instead of $\theta$. This is because in the likelihood function Eq.(\[eq:likelihood\]), we use the distance modulus rather than the luminosity distance, while the former is linearly correlated with $\log\theta$. In principle, the Gaussian processes can reconstruct function at any point. Beyond the data range, however, the reconstructed function has too large uncertainty. The reconstructed $\theta(z)$ function is used to calculate the distance modulus of RSs at the redshift of SNe according to the method described in the last section. SNe with $z<0.5$ are ignored. There are 167 SNe remain in the redshift range $0.5<z\leqslant 1.414$.
We use the publicly available python package [@ForemanMackey:2012ig] to do the Markov chain Monte Carlo analysis. The likelihood function is given by Eq.(\[eq:likelihood\]). We use a flat prior on the parameters: $P(d_0)=U[1,20]$ pc, $P(\eta_0)=U[-1,1]$, $P(\alpha)=U[0,1]$, $P(\beta)=U[0,10]$, and $P(\sigma_{\rm int})=U[0,1]$. The absolute magnitude $M_B$ is fixed to $-19.32$ due to the degeneracy between $d_0$ and $M_B$. The marginalized likelihood distributions and the 2-dimensional confidence regions for the parameters are plotted in Figure \[fig:triangle\] and Figure \[fig:triangle2\] for the first and second parameterizations, respectively. The best-fitting (mean) values and $1\sigma$ errors (standard deviations) of the parameters are listed in Table \[tab:parameter\].
![[]{data-label="fig:triangle"}](triangle.eps){width="50.00000%"}
![[]{data-label="fig:triangle2"}](triangle2.eps){width="50.00000%"}
$d_0$ \[pc\] $\eta_0$ $\alpha$ $\beta$ $\sigma_{\rm int}$
------------- ----------------- ----------------- ---------------- ---------------- --------------------
$\eta_1(z)$ $11.86\pm 0.54$ $-0.06\pm 0.05$ $0.07\pm 0.02$ $0.50\pm 0.20$ $<0.05$
$\eta_2(z)$ $11.46\pm 0.92$ $-0.18\pm 0.16$ $0.07\pm 0.02$ $0.50\pm 0.20$ $<0.05$
\[tab:parameter\]
Except for the intrinsic scatter $\sigma_{\rm int}$, all the other parameters can be well constrained. The $1\sigma$ upper limit of $\sigma_{\rm int}$ is $0.05$, which is much smaller than the errors of SNe and RSs, and therefore can be neglected. In fact, if we fix $\sigma_{\rm int}=0$, all the other parameters are almost unaffected. The best-fitting linear size is $d_0=11.86\pm 0.54$ pc in the first parametrization and $d_0=11.46\pm 0.92$ pc in the second parametrization. These are well consistent with the result of @Jackson:2012jt obtained in $\Lambda$CDM cosmology, i.e. $7.76~h_0^{-1}$ pc (assuming $h_0=0.67$ [@Ade:2013zuv; @Ade:2015xua]). The best-fitting $\eta_0$ is $-0.06\pm 0.05$ ($-0.18\pm 0.16$) in the first (second) parametrization. In both parameterizations, there is no strong evidence for the violation of DDR. Note that we make no assumption on the matter contents or the curvature of the universe, and therefore our results are completely model-independent.
Discussions and conclusions {#sec:conclusions}
===========================
The DDR, which relates the luminosity distance $D_L$ to the angular diameter distance $D_A$ at the same redshift $z$, plays an important role in modern cosmology and astronomy. It is a natural inference of Etherington reciprocity theorem. Any violation of DDR may imply new physics beyond the standard model. Thus, testing the validity of DDR is essential and has aroused great interests in recent years. The most ideal way to test DDR is to measure $D_L$ and $D_A$ to a specific object (hence $D_L$ and $D_A$ are measured at the same redshift, of course). However, at the present day it is difficult to directly measure both $D_L$ and $D_A$ to a celestial object. The most popular way is to measure $D_L$ from type-Ia SNe, and $D_A$ from other objects such as galaxy clusters, BAO, or strong gravitational lensing systems. Although $D_L$ from SNe is free of cosmological model and has high accuracy, $D_A$ from other objects is more or less model-dependent. For example, $D_A$ from galaxy clusters depends on the mass profile of clusters, $D_A$ from BAO depends on the cosmological models, and $D_A$ from strong gravitational lensing systems depends on the curvature of the universe and the mass profile of the lens. On the other hand, the uncertainty of $D_A$ is much larger compared to the uncertainty of $D_L$ from SNe. Due to the large uncertainty, previous tests show no strong evidence for the violation of DDR.
In this paper, we used the type-Ia SNe as the standard candles and ultra-compact RSs as the standard rulers to test the validity of DDR. The linear size of ultra-compact RSs was assumed to be approximately constant, so the angular diameters distance $D_A$ can be obtained from the measured angular size $\theta$. In order to compare $D_A$ from RSs to the luminosity distance $D_A$ from SNe at the same redshift, the Gaussian processes were used to reconstruct $D_A(z)$ function. We parameterized the violation of DDR in two different forms, and used the Markov chain Monte Carlo method to calculate the posterior probability density function of each parameters. The violation amplitudes in the first and second parameterizations are $\eta_0=-0.06\pm 0.05$ and $-0.18\pm 0.16$, respectively. These are consistent with the results of @Liao:2016uzb obtained from SNe and strong gravitational lensing systems (plus galaxy clusters), but the uncertainty was much reduced compared to the previous results. In both parameterizations, no strong evidence for the violation of DDR was found. The main advantage of our method is that it is completely independent of cosmological model and the curvature of the universe. What’s more, it is also independent of the mass profile of galaxies or clusters. In both parameterizations, $\eta_0$ is negative within $1\sigma$ uncertainty. The negative value of $\eta_0$ means that $\eta(z)$ is smaller than unity. This may happen if the photons from SNe are not conserved but some of them are absorbed by the intergalactic medium. This makes SNe seem to be dimmer than expected, which just looks as if the SNe locate at a further distance.
We note that the $\alpha$ and $\beta$ values obtained here are much smaller than that of @Suzuki:2012dhd. There are many reasons which can cause the discrepancy. First, our results are obtain from the combination of RSs and SNe, while @Suzuki:2012dhd obtained the $\alpha$ and $\beta$ values from the pure SNe. The uncertainty of RSs is much larger than that of SNe. Second, only SNe with $z>0.5$ are used in our paper, while in @Suzuki:2012dhd the full Union2.1 sample was used. Finally, biases from the selection effects in both the SNe and RSs samples may be the most important reason. Especially, the SNe sample used here is at high redshift $(z>0.5)$, where the biases are much worse. Dealing with biased samples is beyond the scope of our present work. The different nuisance parameters only cause the discrepancy of SNe distance by 0.1 mag, which is much smaller than the uncertainty of distance of RSs. The main conclusions of our manuscript, therefore, are not affected by the nuisance parameters.
The main shortcoming is that the low redshift ($z<0.5$) RSs couldn’t be used as the standard rulers, while most SNe has redshift $z<0.5$. In the redshift overlapping range ($0.5<z<1.414$), there are only 167 SNe, much smaller than the full Union2.1 sample. Nevertheless, this subsample of SNe is already much larger than the available strong gravitational lensing systems or the galaxy clusters. We may enlarge the sample by adding some high-redshift data such as gamma-ray bursts to SNe sample, or adding some low-redshift galaxy cluster to the RSs sample. The price, however, is that the uncertainty is also enlarged.
Acknowledgements {#acknowledgements .unnumbered}
================
This work has been supported by the National Natural Science Fund of China (Grant Nos. 11603005, 11775038 and 11647307), and the Fundamental Research Funds for the Central Universities (Grant No. 106112016CDJCR301206).
Ade P. A. R., et al. \[Planck Collaboration\], Astron. Astrophys. [**571**]{} (2014) A16 Ade P. A. R., et al. \[Planck Collaboration\], Astron. Astrophys. [**594**]{} (2016) A13 Anderson L., et al., Mon. Not. Roy. Astron. Soc. [**441**]{} (2014), 24 Bassett B. A. and Kunz M., Phys. Rev. D [**69**]{} (2004) 101305 Betoule M., et al. \[SDSS Collaboration\], Astron. Astrophys. [**568**]{} (2014) A22 Beutler F., et al., Mon. Not. Roy. Astron. Soc. [**416**]{} (2011) 3017 Bonamente M., Joy M. K., LaRoque S. J., Carlstrom J. E., Reese E. D. and Dawson K. S., Astrophys. J. [**647**]{} (2006) 25 Chen Z. X., et al., International Journal of Theoretical Physics [**55**]{} (2016) 1229 de Bernardis F., Giusarma E. and Melchiorri A., IJMPD [**15**]{} (2006) 759 De Filippis E., Sereno M., Bautz M. W. and Longo G., Astrophys. J. [**625**]{} (2005) 108 Delubac T., et al. \[BOSS Collaboration\], Astron. Astrophys. [**574**]{} (2015) A59 Etherington I. M. H., Philosophical Magazine [**15**]{} (1933) 761 Etherington I. M. H., General Relativity and Gravitation [**39**]{} (2007) 1055 Foreman-Mackey D., Hogg D. W., Lang D. and Goodman J., Publ. Astron. Soc. Pac. [**125**]{} (2013) 306 Fu X. Y., et al., Research in Astronomy and Astrophysics [**11**]{} (2011) 895 Gurvits L. I., Astrophys. J. [**425**]{} (1994) 442 Gurvits L. I., Kellermann K. I. and Frey S., Astron. Astrophys. [**342**]{} (1999) 378 Guy J., Astier P., Nobili S., Regnault N., and Pain R., A&A [**443**]{}, (2005) 781 Guy J., Astier P., Baumont S., et al., A&A [**466**]{} (2007) 11 Holanda R. F. L., Lima J. A. S. and Ribeiro M. B., Astrophys. J. [**722**]{} (2010) L233 Holanda R. F. L., Busti V. C. and Alcaniz J. S., JCAP [**1602**]{} (2016), 054 Holanda R. F. L., Busti V. C., Lima F. S. and Alcaniz J. S., arXiv:1611.09426 Jackson J. C. and Dodgson M., Mon. Not. Roy. Astron. Soc. [**278**]{} (1996) 603 Jackson J. C. and Dodgson M., Mon. Not. Roy. Astron. Soc. [**285**]{} (1997) 806 Jackson J. C., JCAP [**0411**]{} (2004) 007 Jackson J. C. and Jannetta A. L., JCAP [**0611**]{} (2006) 002 Jackson J. C., Mon. Not. Roy. Astron. Soc. [**390**]{} (2008) L1 Jackson J. C., Mon. Not. Roy. Astron. Soc. [**426**]{} (2012) 779 Kazin E. A., et al., Mon. Not. Roy. Astron. Soc. [**441**]{} (2014), 3524 Kellermann K. I., Nature [**361**]{} (1993) 134 Liao K., Li Z., Cao S., Biesiada M., Zheng X. and Zhu Z. H., Astrophys. J. [**822**]{} (2016) 74 Lv M. Z. and Xia J. Q., Phys. Dark Univ. [**13**]{} (2016) 139 Ma C. and Corasaniti P. S., arXiv:1604.04631 Perlmutter S., et al., Astrophys. J. [**517**]{} (1999) 565 Piorkowska A., et al., Acta Physica Polonica B [**42**]{} (2011) 2297 Preston R. A., Morabito D. D., Williams J. G., et al., Astron. J. [**90**]{} (1985) 1599 Riess A. G., et al., Astron. J. [**116**]{} (1998) 1009 Santos-da-Costa S., Busti V. C. and Holanda R. F. L., JCAP [**1510**]{} (2015) 061 Seikel M., Clarkson C. and Smith M., JCAP [**1206**]{} (2012) 036 Suzuki N., Rubin D., Lidman C., et al., Astrophys. J. [**746**]{} (2012) 85 Tripp R. 1998, A&A, 331, 815 Uzan J. P., Aghanim N. and Mellier Y., Phys. Rev. D [**70**]{} (2004) 083533 Yang X., Yu H. R., Zhang Z. S. and Zhang T. J., Astrophys. J. Lett. [**777**]{} (2013) L24
\[lastpage\]
|
{
"pile_set_name": "ArXiv"
}
|
[USITP 98-05]{}
FROM TOPOLOGICAL TO
PARAMETRIZED FIELD THEORY
\
[Nuno Barros e Sá]{}[^1]
[Ingemar Bengtsson]{}[^2]
*$^{1,2}$Fysikum, Stockholms Universitet, Box 6730,*
113 85 Stockholm, Sverige
$^1$DCTD, Universidade dos Açores, R. Mãe de Deus,
9500 Ponta Delgada, Açores, Portugal
[**Abstract**]{}\
\
It has been proposed to study the theory resulting from setting the gravitational constant to zero in the first order formalism for general relativity. In this letter we investigate this theory in the presence of matter fields, establish its equivalence with parametrized field theory on a flat background, and relate it to previous results in topological field theory (BF theory).
\
Over the past ten years a bewildering array of diffeomorphism invariant field theories has been studied. In many cases they contain a finite number of degrees of freedom only, and they carry topological information about the manifold on which they are defined. By contrast, the most interesting theory of this kind remains that of Einstein, in which the metric occurs as a dynamical variable carrying two degrees of freedom per spatial point. Einstein’s theory in 2+1 dimensions occupies the middle ground. It can be formulated like the 3+1 dimensional theory, but it can also be formulated as a topological gauge theory in which the metric—or more precisely the metrical triad—occurs as a gauge field [@Achucarro]. When matter couplings are included the first formulation turns out to be superior [@Blencowe]. One is left wondering about the conditions under which a set of variables in a diffeomorphism invariant theory can be meaningfully identified as a spacetime metric.
We will not try to answer the general question in this letter. Instead, we will point out that there is a set of ideas that can be used to provide a 3+1 dimensional illustration of the issues involved. First we will show that the Einstein-Hilbert action can be “short circuited” in a certain way so that the metric carries no degrees of freedom (as first observed by Smolin [@Smolin]). In effect we obtain a diffeomorphism invariant theory where the only solution is Minkowski space—also when matter fields are included. We then show that this model is closely related to the BF topological field theory defined by Horowitz [@Horowitz]. As in 2+1 gravity it is possible to reorganize the constraints of the model so that we obtain the constraint algebra familiar from general relativity. Our fourth and final point is that we can solve some of the constraints of our theory by means of a canonical transformation to gauge invariant variables. When this is done we recover precisely the parametrized field theories studied by Dirac and Kuchař [@Dirac].
We will assume that the metric that we define is non-degenerate. While rather foreign to topological field theory this assumption is natural in metrical theories. The same assumption has to be made in 2+1 gravity in order to show that the two formulations referred to above are indeed equivalent. In this as well as in our case there is a subtlety involved, and we will comment on this at the appropriate point.
Our starting point is the first order action for gravity built from tetrads $e_I$ and connections ${\omega}_{IJ}$,
$$S_E = \frac{1}{8}\int {\epsilon}^{{\alpha}
{\beta}{\gamma}{\delta}}{\epsilon}^{IJKL}e_{{\alpha}I}
e_{{\beta}J}R_{{\gamma}{\delta}KL} + S_m \ .$$
Here $I, J, ...$ are internal indices that can be raised and lowered with a Minkowski metric ${\eta}_{IJ}$, the curvature tensor is
$$R_{{\alpha}{\beta}IJ} = \partial_{\alpha}
{\omega}_{{\beta}IJ} - \partial_{\beta}{\omega}_{{\alpha}IJ} +
G{\omega}_{{\alpha}I}^{\ \ \ K}{\omega}_{{\beta}KJ} \ ,$$
$G$ is the gravitational coupling constant, and this form of the first order action may differ from that of other references by a redefinition $\omega_{IJ}\to G\omega_{IJ}$. $S_m$ is an action for matter fields that is independent of the connection and depends on the tetrad only through the metric tensor
$$g_{{\alpha}{\beta}} = e_{{\alpha}I}
e_{{\beta}J}{\eta}^{IJ}
\ .$$
For definiteness we may choose an action for a scalar field,
$$S_m = - \frac{1}{2}\int \sqrt{-g}(
g^{{\alpha}{\beta}}\partial_{\alpha}{\varphi}\partial_{\beta}
{\varphi} + m^2{\varphi}^2) \ .$$
Our results will however be general and will not depend on this particular choice of the matter action.
The model that we will consider is obtained by setting $G = 0$ in the above action;
$$S = \frac{1}{4}\int {\epsilon}^{{\alpha}
{\beta}{\gamma}{\delta}}{\epsilon}^{IJKL}e_{{\alpha}I}
e_{{\beta}J}\partial_{\gamma}{\omega}_{{\delta}KL} + S_m \ .$$
This is a drastic operation. Instead of an $SO(3,1)$ connection we now have a collection of six $U(1)$ connections, and the tetrad is a collection of four gauge invariant vector fields. Moreover we will see that this theory does not have any local degrees of freedom (in the absence of matter fields). The use of the matrix ${\eta}_{IJ}$ to build a spacetime metric may therefore seem completely ad hoc. However, we will see that—at least when matter fields are included—it is actually quite well motivated. We do need ${\eta}_{IJ}$ to build the matter action $S_m$.
The $G = 0$ “limit” was first studied by Smolin [@Smolin] in the context of Ashtekar’s variables. Then the starting point is the self-dual form of the action, and the detailed results are quite different from ours since the equivalence between the two forms of the action breaks down when $G = 0$. In particular the model studied by Smolin has the same number of degrees of freedom as Einstein’s theory (either Euclidean or complex Lorentzian).
Varying our action with respect to the connection yields an equation for the tetrad:
$${\epsilon}^{{\alpha}{\beta}{\gamma}{\delta}}
{\epsilon}^{IJKL}e_{{\gamma}K}e_{{\delta}L} = 0 \hspace{5mm}
\Leftrightarrow \hspace{5mm} \partial_{\alpha}e_{{\beta}I}
- \partial_{\beta}e_{{\alpha}I} = 0 \ .$$
(Varying Einstein’s action would yield an equation for the connection at this point—our action is indeed “short circuited”.) This means that the tetrads are closed forms and that the only solution is (locally) Minkowski spacetime:
$$e_{{\alpha}I} = \partial_{\alpha}f_I
\hspace{5mm} \Rightarrow \hspace{5mm} g_{{\alpha}{\beta}}
= \partial_{\alpha}f^I{\eta}_{IJ}\partial_{\beta}f^J \ .$$
Varying the tetrads one obtains a set of equations relating the connections to the tetrads and to the matter fields. These may be solved for the connection—they do not constrain either the tetrads or the matter fields. Finally, varying the matter fields leads to the usual field equations for matter propagating in a flat background. Hence we have a diffeomorphism invariant theory in which matter does not curve geometry.
One may wonder whether there is anything special about flat space here? From the present point of view there is; although one may linearize the first order action around any connection ${\omega}^{(0)}_{IJ}$ that solves Einstein’s equations and then proceed as above diffeomorphism invariance would be lost in the process. Unless ${\omega}^{(0)}_{IJ} = 0$ one ends up with an action that contains fixed functions of the coordinates.
We will now analyze our model in more detail using the Hamiltonian formalism. Until further notice we set $S_m = 0$, that is to say that we do not include the matter degrees of freedom. We start by defining a new set of variables
$$B_{{\alpha}{\beta}}^I = - \frac{1}{2}
{\epsilon}^{IJKL}e_{[{\alpha}J}{\omega}_{{\beta}]KL}
\ .$$
Provided that the tetrad is indeed non-degenerate this is a one-to-one transformation from the connections to the two forms $B^I$, having the inverse
$$\omega_{\alpha IJ}=\frac{1}{2}\epsilon_{IJKL}
e^{\beta K}\left( {B_{\alpha\beta}}^L-\frac{1}{2}
e^{\gamma L} e_{\alpha M} {B_{\beta\gamma}}^M\right) \ .$$
Hence we can perform this change of variables in the action. After a partial integration it becomes
$$S = \int {\epsilon}^{{\alpha}{\beta}
{\gamma}{\delta}}B_{{\alpha}{\beta}}^I\partial_{\gamma}
e_{{\delta}I} \ .$$
In this form the action is just four copies of the BF topological field theory studied by Horowitz [@Horowitz]. Hence we have shown that the $G = 0$ version of the Einstein action is equivalent to that sector of the abelian BF theory in which the tetrad is invertible.
The action has a large gauge invariance (larger than that of the Einstein action), namely
$$e_{{\alpha}I} \rightarrow e_{{\alpha}I} -
\partial_{\alpha}{\Lambda}_I \$$
$$B_{{\alpha}{\beta}}^I \rightarrow B_{{\alpha}{\beta}}^I
- \partial_{[{\alpha}}{\Lambda}_{{\beta}]}^I \ .$$
It follows that a non-degenerate tetrad can always be transformed so that it vanishes at any chosen point. The same difficulty occurs when one attempts to show the equivalence between the two possible formulations of 2+1 gravity referred to above. It is a moderately embarrassing difficulty; a possible attitude to take in both cases is that the true starting point is a suitable form of the phase space action, in which this problem does not appear. Therefore we proceed with the calculation.
It is straightforward to perform a 3+1 decomposition of the action in the form that we arrived at. We get
$$S = \int \dot{e}_{{\alpha}I}{\pi}^{aI}
+ e_{tI}\partial_a{\pi}^{aI} - {\lambda}_a^I{\epsilon}^{abc}
\partial_be_{bI} \ ,$$
where we renamed the components of the two form according to
$${\pi}^{aI} = {\epsilon}^{abc}B_{bc}^I
\hspace{1cm} {\lambda}_a^I = - 2B_{ta}^I \ .$$
Excluding the matter action there are 16 first class constraints of which 12 are independent, and there are 12 canonical variables per spatial point. Hence the model is devoid of dynamical degrees of freedom. The constraint algebra is abelian, which means that diffeomorphism invariance—which should be realized as a gauge freedom—is not manifest. This form of the phase space action is analogous to Witten’s form of the phase space action for 2+1 gravity [@Achucarro]. Non-degeneracy of the tetrad is not a gauge invariant property in this formulation, and inclusion of the matter fields in the constraints is problematic.
These problems can be avoided by a redefinition of the Lagrange multipliers. We wish to interpret the tetrad as giving rise to a metric, and we would therefore like to introduce the lapse function $N$ and the shift vector $N^a$ as multipliers. Any spacetime metric can be split into the spatial metric
$$q_{ab} = e_{aI}e_{bJ}{\eta}^{IJ}$$
induced on a spatial hypersurface at constant $t$, together with the lapse and shifts
$$N = \frac{1}{\sqrt{-qg^{tt}}} \hspace{1cm}
N^a = - \frac{g^{ta}}{g^{tt}} \ .$$
(Here $q$ is the determinant of the spatial metric, and we use a lapse function that is a tensor density of weight minus one.) These equations can be inverted so that the time component of the tetrad becomes
$$e_{tI} = \frac{N}{6}{\epsilon_I}^{JKL}
{\epsilon}^{abc}e_{aJ}e_{bK}e_{cL} + N^ae_{aI} \ . \label{17}$$
Our action does not single out a preferred metric, but if we trade the Lagrange multiplier $e_{tI}$ for the lapse and shifts it will acquire one, having a signature determined by the signature of the matrix ${\eta}_{IJ}$.
When we use this result in the phase space action we obtain
$$S = \int \dot{e}_{aI}{\pi}^{aI} - N{\cal H}
- N^a{\cal H} - {\lambda}^I{\epsilon}^{abc}\partial_be_{cI} \
.$$
The constraints are the Hamiltonian and vector constraints, together with the constraints ${\phi}^a_I = 0$ already encountered:
$${\cal H} = \frac{1}{3!}{\epsilon^{IJK}}_L
{\epsilon}^{abc}e_{aI}e_{bJ}e_{cK}\partial_d{\pi}^{dL}$$
$${\cal H}_a = - e_{aI}\partial_b{\pi}^{bI}$$
$${\phi}^a_I = {\epsilon}^{abc}\partial_be_{cI}\ .$$
In this formulation the spatial diffeomorphisms are manifest, or almost so. In fact there is a simple combination of the constraints that generates spatial diffeomorphisms;
$${\cal D}_a = {\cal H}_a + {\epsilon}_{abc}
{\pi}^{bI}{\phi}^c_I \ .$$
The constraint algebra has the following non-zero brackets:
$$\{{\cal H}[N], {\cal H}[M]\} = {\cal H}_a[(N\partial_aM
- M\partial_aN)qq^{ab}]$$
$$\{{\cal H}_a[N^a], {\cal H}[M]\} = {\cal H}
[{\cal L}_{\bar{N}}M] - {\phi}^a_I[{\epsilon^{IJK}}_L M N^b e_{bJ}
e_{aK} \partial_c{\pi}^{cL}] \ ,$$
$$\{{\cal H}_a[N^a], {\cal H}_b[M^b]\} = {\cal H}_a
[{\cal L}_{\bar{N}}M^a] - {\phi}^a_I[{\epsilon}_{abc}N^bM^c
\partial_d{\pi}^{dI}] \ ,$$
where ${\cal L}$ denotes the Lie derivative and the square bracket denote smearing with test functions. On the constraint surface ${\phi}^a_I = 0$ this is the usual constraint algebra of general relativity. It is the fingerprint of diffeomorphism invariance in a metric space in the Hamiltonian formulation, having the geometrical interpretation [@Hojman] as the algebra of deformations of spatial hypersurfaces in a Lorentzian spacetime. From this point of view the matrix ${\eta}_{IJ}$ is an object that is inserted in the phase space action precisely in order to make a geometrical interpretation of the solutions possible.
We may now adopt this first order action as a precise definition of our model. This is analogous to the ADM formulation of 2+1 gravity, and it has the double advantages that non-degeneracy of the metric can be consistently imposed, and that inclusion of matter degrees of freedom is straightforward. The latter will affect the form of the Hamiltonian and vector constraints, but they will not affect the constraint algebra. This follows from the assumption that only metrical couplings of the matter fields will be considered (that is the case for bosonic fields and also for the standard coupling of fermionic fields when setting $G=0$), together with the crucial bracket
$$\{q_{ab}, {\cal H}[N]\} = N\frac{1}{2}e_{aI}
{\epsilon}^{IJKL}{\epsilon}^{cde}\partial_b(e_{cJ}e_{dK}e_{eL})
\ + \ {\small (a \leftrightarrow b)} \ .$$
The point here is the absence of any derivatives acting on the lapse function. (For a full explanation of this point as well as of all other properties of our constraint algebra, see Hojman et al. [@Hojman].)
We now have a diffeomorphism invariant theory that describes matter propagating on a flat background. There are no local degrees of freedom attached to the geometry. One might expect that there should be a relation to parametrized field theory [@Dirac], which achieves the same goal at the expense of introducing special “embedding variables” into the action. This is in fact the case. To see this we observe that, restricting ourselves to simply connected space, we can solve some of our constraints explicitly:
$${\phi}^a_I = {\epsilon}^{abc}\partial_be_{cI} = 0
\hspace{.5cm} \Leftrightarrow \hspace{.5cm} e_{aI} = \partial_aX_I \ .$$
We can now use Hamilton-Jacobi theory to effect a (singular) canonical transformation from the tetrad and its momentum to a new set of canonical pairs $(X_I, P^J)$ that coordinatize the constraint surface ${\phi}^a_I = 0$ modulo the gauge transformations generated by this constraint [@Newman]. To this end we choose a generating functional that depends on the “old” momenta and the “new” coordinates,
$$S_{pQ} = - \int \partial_aX_I{\pi}^{aI} \ .$$
Then the canonical transformation is given by
$$e_{aI} = - \frac{{\delta}S_{pQ}}{{\delta}{\pi}^a_I}
= \partial_aX_I$$
$$P^I = - \frac{{\delta}S_{pQ}}{{\delta}X_I} =
- \partial_a{\pi}^{aI} \ .$$
Making use of this in the phase space action (in the first of the two forms given above) we obtain
$$S = \int \dot{X}_IP^I - e_{tI}P^I \ .$$
This is the action describing the kinematics of the vector $e_I$ that describes the deformations of spatial hypersurfaces in spacetime [@Dirac].
We can now trade the four vector $e_I$ for the lapse and shifts, just as we did above. This means that we write
$$e_{tI} = Nn_I + N^a\partial_aX_I \ ,$$
where the vector $n_I$, as defined in eq. (\[17\]), obeys
$$n^I N^a\partial_aX_I = 0 \ .$$
We can also add any matter action (with metrical couplings) to the phase space action. In this way we arrive at
$$\begin{aligned}
S = \int \dot{X}_IP^I + N\frac{1}{3!}{\epsilon^{IJK}}_L
{\epsilon}^{abc}\partial_aX_I\partial_bX_J\partial_cX_K P^L
- N^a\partial_aX_IP^I + S_m \ . \end{aligned}$$
This is the action of a parametrized field theory on a flat background [@Dirac]; the constraint algebra is the same as that of general relativity.
In conclusion, we have realised the equivalence between a sector of a BF theory and Smolin’s $G=0$ limit of Einstein’s gravity, successfully introduced matter terms in these models, and shown that they are equivalent to parametrized field theory on a flat background provided taht spacetime is simply connected.
[99]{}
A. Achúcarro and P.K. Townsend, Phys. Lett. (1986) 89.
E. Witten, Nucl. Phys. (1988) 46.
M.P. Blencowe, Nucl. Phys. (1990) 213.
L. Smolin, Class. Quant. Grav. (1992) 883.
G.T. Horowitz, Comm. Math. Phys. (1989) 417.
P.A.M. Dirac: Lectures on Quantum Mechanics, Belfer Graduate School of Science, New York 1964.
K. Kuchař, in C.J. Isham, R. Penrose and D.W. Sciama (eds.): Quantum Gravity 2—A Second Oxford Symposium, Oxford 1981.
S.A. Hojman, K. Kuchař and C. Teitelboim, Ann. Phys. (1976) 88.
E.T. Newman and C. Rovelli, Phys. Rev. Lett. (1992) 1300.
E.T. Newman and C. Rovelli, in: F. Colomo, L. Lusanna and G. Marmo: Constraint Theory and Quantization Methods, Singapore 1994
[^1]: Email address: [email protected]. Supported by grant PRODEP-Acção 5.2.
[^2]: Email address: [email protected]. Supported by NFR-project 07923.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We describe an adaptation and application of a search-based structured prediction algorithm “[<span style="font-variant:small-caps;">Searn</span>]{}” to unsupervised learning problems. We show that it is possible to reduce unsupervised learning to supervised learning and demonstrate a high-quality unsupervised shift-reduce parsing model. We additionally show a close connection between unsupervised [<span style="font-variant:small-caps;">Searn</span>]{} and expectation maximization. Finally, we demonstrate the efficacy of a semi-supervised extension. The key idea that enables this is an application of the *predict-self* idea for unsupervised learning.'
---
Introduction
============
A prevalent and useful version of unsupervised learning arises when both the observed data and the latent variables are structured. Examples range from hidden alignment variables in speech recognition [@rabiner89tutorial] and machine translation [@brownetal93; @vogel96], to latent trees in unsupervised parsing [@paskin01bigrams; @klein04induction; @smith05induction; @titov07latent], and to pose estimation in computer vision [@ramanan05pose]. These techniques are all based on probabilistic models. Their applicability hinges on the tractability of (approximately) computing latent variable expectations, thus enabling the use of EM [@dempster77em]. In this paper we show that a recently-developed *search-based* algorithm, [<span style="font-variant:small-caps;">Searn</span>]{} [@daume09searn] (see Section \[sec:searn\]), can be utilized for unsupervised structured prediction (Section \[sec:unsearn\]). We show: (1) that under an appropriate construction, [<span style="font-variant:small-caps;">Searn</span>]{} can imitate the expectation maximization (Section \[sec:em\]); (2) that unsupervised [<span style="font-variant:small-caps;">Searn</span>]{} can be used to obtain competitive performance on an unsupervised dependency parsing task (Section \[sec:gi\]); and (3) that unsupervised [<span style="font-variant:small-caps;">Searn</span>]{} naturally extends to a semi-supervised setting (Section \[sec:semi\]). The key insight that enables this work is that we can consider the prediction of the (observed) input to be, itself, a structured prediction problem.
Structured Prediction {#sec:sp}
=====================
The *supervised* structured prediction problem is the task of mapping inputs $x$ to complex structured outputs $y$ (e.g., sequences, trees, etc.). Formally, let $\cX$ be an arbitrary input space and $\cY$ be structure output space. $\cY$ is typically assumed to *decompose* over some smaller substructures (e.g., labels in a sequence). $\cY$ comes equipped with a loss function, often assumed to take the form of a Hamming loss over the substructures. Features are defined over pairs $(x,y)$ in such a way that they obey the substructures (e.g., one might have features over adjacent label pairs in a sequence). Under strong assumptions on the structures, the loss function and the features (essentially “locality” assumptions), a number of learning algorithms can be employed: for example, conditional random fields [@lafferty01crf] or max-margin Markov networks [@taskar05mmmn].
A key difficulty in structured prediction occurs when the output space $\cY$, the features, or the loss, does not decompose nicely. All of these issues can lead to intractable computations at either training or prediction time (often both). An attractive approach for dealing with this intractability is to employ a search-based algorithm. The key idea in search-based structured prediction is to first decompose the output $y$ into a sequence of (dependent) smaller predictions $y_1, \dots, y_T$. These may each be predicted in turn, with later predictions dependent of previous decisions.
Search-based Structured Prediction {#sec:sbsp}
----------------------------------
A recently proposed algorithm for solving the structured prediction problem is [<span style="font-variant:small-caps;">Searn</span>]{} [@daume09searn]. [<span style="font-variant:small-caps;">Searn</span>]{} operates by considering each substructure prediction $y_1, \dots, y_T$ as a classification problem. A classifier $h$ is trained so that at time $t$, given a feature vector, it predict the best value for $y_t$. The feature vector can be based on any part of the input $x$ and any previous decision $y_1, \dots, y_{t-1}$. This introduces a chicken-and-egg problem. $h$ should ideally be trained so that it makes the best decision for $y_t$ *given* that $h$ makes all past decisions $y_1, \dots, y_{t-1}$ and all future decisions $y_{t+1},
\dots, y_T$. Of course, at training time we do not have access to $h$ (we are trying to construct it). The solution is to use an iterative scheme.
[<span style="font-variant:small-caps;">Searn</span>]{} {#sec:searn}
-------------------------------------------------------
The presentation we give here differs slightly from the original presentation of the [<span style="font-variant:small-caps;">Searn</span>]{} algorithm. Our motivation for straying from the original formulation is because our presentation makes more clear the connection between our unsupervised variant of [<span style="font-variant:small-caps;">Searn</span>]{} and more standard unsupervised learning methods (such as standard algorithms on hidden Markov models).
Let $\cD^{\text{SP}}$ denote a distribution over pairs $(x,y)$ drawn from $\cX \times \cY$, and let $\ell(y,\hat y)$ be the loss associated with predicting $\hat y$ when the true answer is $y$. We assume that $y \in \cY$ can be decomposed into atomic predictions $y_1, \dots,
y_T$, where each $y_t$ is drawn from a discrete set $Y$. A *policy*, $\pi$, is a (possibly stochastic) function that maps tuples $(x, y_1, \dots, y_{t-1})$ to atomic predictions $y_t$.
The key ingredient in [<span style="font-variant:small-caps;">Searn</span>]{} is to use the loss function $\ell$ and a “current” policy $\pi$ to turn $\cD^{\text{SP}}$ into a distribution over cost-sensitive (multiclass) classification problems [@beygelzimer05reductions]. A cost-sensitive classification example is given by an input $x$ and a cost vector $\vec c = \langle
c_1, \dots, c_K \rangle$, where $c_k$ is the cost of predicting class $k$ on input $x$. Define by [<span style="font-variant:small-caps;">Searn</span>]{}$(\cD^{\text{SP}}, \ell, \pi)$ a distribution over cost-sensitive classification problems derived as follows. To sample from this induced distribution, we first sample an example $(x,y) \sim \cD^{\text{SP}}$. We then sample $t$ uniformly from $[1,T]$ and run $\pi$ for $t-1$ steps on $(x,y)$. This yields a partial prediction $(\hat y_1, \dots, \hat y_{t-1})$. The input for the cost sensitive classification problem is then the tuple $(x,\hat
y_1, \dots, \hat y_{t-1})$. The costs are derived as follows. For each possible choice $k$ of $\hat y_t$, we defined $c_k$ as the *expected* loss if $\pi$ were run, beginning at $(\hat y_1,
\dots, \hat y_{t-1}, k)$ on input $x$. Formally: $$c_k = \Ep_{\hat
y_{t+1}, \dots, \hat y_T \sim \pi} \ell(y, (\hat y_1, \dots, \hat
y_{t-1}, k, \hat y_{t+1}, \dots, \hat y_T))$$ [<span style="font-variant:small-caps;">Searn</span>]{} assumes access to an “initial policy” $\pi^*$ (sometimes called the “optimal policy”). Given an input $x$, a true output $y$ and a prefix of predictions $\hat y_1, \dots, \hat y_{t-1}$, $\pi^*$ produces a best next-action, $\hat y_t$. It should be constructed so that the choice $\hat y_t$ is optimal (or close to optimal) with respect to the problem-specific loss function. For example, if the loss function is Hamming loss, the $\pi^*$ will always produce $\hat
y_t = y_t$. For more complex loss functions, computing $\pi^*$ may be more involved.
Given these ingredients, [<span style="font-variant:small-caps;">Searn</span>]{} operates according the algorithm given in Figure \[fig:searn\]. Operationally, the sampling step is typically implemented by generating *every* example from a fixed structured prediction training set. The costs (expected losses) are computed by sampling with tied randomness [@ng00pegasus].
If $\be = 1/T^3$, one can show [@daume09searn] that after at most $2T^3 \ln T$ iterations, [<span style="font-variant:small-caps;">Searn</span>]{} is guaranteed to find a solution $\pi$ with structured prediction loss bounded as: $$\label{eq:searn}
L(\pi) \leq L(\pi^*) + 2 \ell_\textrm{avg} T \ln T + c (1+\ln T)/T$$ where $L(\pi^*)$ is the loss of the initial policy (typically zero), $T$ is the length of the longest example, $c$ is the worse-case per-step loss and $\ell_\textrm{avg}$ is the average multiclass classification loss. This shows that the structured prediction algorithm learned by [<span style="font-variant:small-caps;">Searn</span>]{} is guaranteed to be not-much-worse than that produced by the initial policy, *provided* that the created classification problems are easy (i.e., that $\ell_\textrm{avg}$ is small). Note that one can use *any* classification algorithm one likes.
Unsupervised [<span style="font-variant:small-caps;">Searn</span>]{} {#sec:unsearn}
====================================================================
In unsupervised structured prediction, we no longer receive an pair $(x,y)$ but instead observes only an input $x$. Our job is to construct a classifier that produces $y$, even though we have never observed it.
Reduction for Unsupervised to Supervised
----------------------------------------
The key idea—one that underlies much work in unsupervised learning—is that a good $y$ is one that enables us to easily recover $x$. This is precisely the intuition we build in to our model. The observation that makes this practical is that there is nothing in the theory or application of [<span style="font-variant:small-caps;">Searn</span>]{} that says that $\pi^*$ cannot be stochastic. Moreover, there is not requirement that the loss function depend on *all* components of the prediction. Our model will essentially first predict $y$ and then predict $x$ based on $y$. Importantly, the loss function is agnostic to $y$ (since we do not have true outputs).
The general construction is as follows. Let $\cD^{\text{unsup}}$ be a distribution over inputs $x \in \cX$ and let $\cY$ be the space of desired latent structures (e.g., trees). We define a distribution $\cD^{\text{sup}}$ over $\cX \times (\cY \times \cX)$ by defining a sampling procedure. To sample from $\cD^{\text{sup}}$, we first sample $x \sim \cD^{\text{unsup}}$. We then sample uniformly from the set of all $\cY$ that are valid structures for $x$. Finally, we return the pair $(x, (y, x))$. We define a loss function $L$ by $L((y, x), (\hat y, \hat x)) = L^{\text{input}}(x, \hat x)$ where $L^{\text{input}}$ is any loss function on the input space (e.g., Hamming loss). We apply [<span style="font-variant:small-caps;">Searn</span>]{} to the supervised structured prediction problem $\cD^{\text{sup}}$, and implicitly learn latent structures.
Sequence Labeling Example
-------------------------
To gain insight into the operation of [<span style="font-variant:small-caps;">Searn</span>]{} in the unsupervised setting, it is useful to consider a sequence labeling example. That is, our input $x$ is a sequence of length $T$ and we desire a label sequence $y$ of length $T$ drawn from a label space of size $K$. We convert this into a supervised learning problem by considering the “true” structured output to be a label sequence of length $2T$, with the first $T$ components drawn from the label space of size $K$ and the second $T$ components drawn from the input vocabulary. The loss function can then be anything that depends only on the last $T$ components. For simplicity, we can consider it to be Hamming loss. The construction of the optimal policy in this case is straightforward. For the first $T$ components, $\pi^*$ may behave arbitrarily (e.g., it may produce a uniform distribution over the $K$ labels). For the second $T$ components, $\pi^*$ always predicts the true label (which is known, because it is part of the input).
An important aspect of the model is the construction of the feature vectors. It is most useful to consider this construction as having two parts. The first part has to do with predicting the hidden structure (the first $T$ components). The second part has to do with predicting the observed structure (the second $T$ components). For the first part, we are free to use whatever features we desire, so long as they can be computed based on the input $x$ and a partial output. For instance, in the HMM case, we could use the two most recent label predictions and windowed features from $x$.
The construction of the features for the second part is, however, also crucial. For instance, if the feature vector corresponding to “predict the $t$th component of $x$” contains the $t$ component of $x$, then this learning problem is trivial—but also renders the latent structure useless. The goal of the designer of the feature space is to construct features for predicting $x_t$ that crucially depend on getting the latent structure $y$ correct. That is, the ideal feature set is one for which you can predict $x_t$ accurately *if an only if* we have found the correct latent structure (more on this in Section \[sec:guarantees\]). For instance, in the HMM case, we may predict $x_t$ based only on the corresponding label $y_t$, or maybe on the basis of $y_{t-1},y_t,y_{t+1}$. (Note that we are not limited to the Markov assumption, as in the case of HMMs.)
In the first iteration of [<span style="font-variant:small-caps;">Searn</span>]{}, all costs for the prediction of the latent structure are computed with respect to the initial policy. Recalling that the initial policy behaves randomly when predicting the latent labels and correctly when predicting the words, we can see that these costs are all *zero*. Thus, for the latent structure actions, [<span style="font-variant:small-caps;">Searn</span>]{} will not induce any classification examples (because the cost of all actions is equal). However, it will create example for predicting the $x$ component. For predicting the $x$s, the cost will be zero for the correct word and one for any incorrect word. These examples will have associated features: we will predict word $x_t$ based *exclusively* on $y_t$. Remember: $y_t$ was generated randomly by the initial policy.
In the *second* iteration, the behavior is different. [<span style="font-variant:small-caps;">Searn</span>]{}returns to creating examples for the latent structure components. However, in this iteration, since the current policy is not longer optimal, the future cost estimates may be non-zero. Consider generating an example corresponding to a (latent) state $y_t$. For some small percentage (as dictated by $\beta$) of the “generate $x$” decisions, the previously learned classifier will fire. If this learned classifier does well, then the associated cost will be low. However, if the learned classifier does poorly, the the associated cost will be high. Intuitively, the learned classifier will do well if and only if the action that labels $y_t$ is “good” (i.e., consistent with what was learned previously). This, in the second pass through the data, [<span style="font-variant:small-caps;">Searn</span>]{} *does* create classification examples specific to the latent decisions.
As [<span style="font-variant:small-caps;">Searn</span>]{} iterates, more and more of the latent prediction decisions are made according to the learned classifiers and not with respect to the random policy.
Comparison to EM {#sec:em}
================
In this section, we show an equivalence between expectation maximization in directed probabilistic structures and unsupervised [<span style="font-variant:small-caps;">Searn</span>]{}. We use mixture of multinomials as a motivating example (primarily for simplicity), but the results easily extend to more complicated models (e.g., HMMs: see Section \[sec:hmm\]).
EM for Mixture of Multinomials {#sec:em-mm}
------------------------------
In the mixture of multinomials problem, we are given $N$ documents $\vec d_1, \dots, \vec d_N$, where $\vec d_n$ is a vector of word counts over a vocabulary of size $V$; that is, $d_{n,v}$ is the number of times word $v$ appeared in document $n$. The mixture of multinomials is a probabilistic clustering model, where we assume an underlying set of $K$ clusters (multinomials) that generated the documents. Denote by $\th_k$ the multinomial parameter associated with cluster $k$, $\rho_k$ the prior probability of choosing cluster $k$, and let $\vec z_n$ be an indicator vector associating document $n$ with the unique cluster $k$ such that $z_{n,k} = 1$. The probabilistic model has the form: $$\label{eq:mmlik}
p(\vec d \| \vec \th, \vec \rho) =
\prod_n
\frac {(\sum_v d_{n,v})!}
{\prod_v d_{n,v}!}
\sum_{\vec z_n} \prod_k \left[
\rho_k
\prod_v \th_{k,v}^{d_{n,v}}
\right]^{z_{n,k}}$$ Expectation maximization in this model involves first computing expectations over the $\vec z$ vectors and then updating the model parameters $\vec \th$:
$$\begin{aligned}
\text{E-step:} &&
z_{n,k} &\varpropto \rho_k \prod_v \th_{k,v}^{d_{n,v}} \label{eq:em:z} \\
\text{M-step:} &&
\th_{k,v} &\varpropto \sum_n z_{n,k} d_{n,v} &
;~~~ \rho_k &\varpropto \sum_n z_{n,k} \label{eq:em:rho}\end{aligned}$$
In both cases, the constant of proportionality is chosen so that the variables sum to one over the last component. These updates are repeated until convergence of the incomplete data likelihood, Eq .
An Equivalent Model in Searn {#sec:em-searn}
----------------------------
Now, we show how to construct an instance of unsupervised [<span style="font-variant:small-caps;">Searn</span>]{} that effectively mimics the behavior of EM on the mixture of multinomials problem. The ingredients are as follows:
- The input space $\cX$ is the space of documents, represented as word count vectors.
- The (latent) output space $\cY$ is a single discrete variable in the range $[1,K]$ that specifies the cluster.
- The feature set for predicting $y$ (document counts).
- The feature set for predicting $x$ is the label $y$ and the total number of words in the document. The predictions for a document are estimated word probabilities, not the words themselves.
- The loss function ignores the prediction $y$ and returns the log loss of the true document $x$ under the word probabilities predicted.
- The cost-sensitive learning algorithm is different depending on whether the latent structure $y$ is being predicted or if the document $x$ is being predicted:
- Structure: The base classifier is a multinomial naïve Bayes classifier, parameterized by (say) $h^m$
- Document: The base classifier is a collection of independent maximum likelihood multinomial estimators for each cluster.
Consider the behavior of this setup. In particular, consider the distribution [<span style="font-variant:small-caps;">Searn</span>]{}$(\cD^{\text{SP}}, \ell, \pi)$. There are two “types” of examples drawn from this distribution: (1) latent structure examples and (2) document examples. The claim is that *both* classifiers learned are identical to the mixture of multinomials model from Section \[sec:em-mm\].
Consider the generation of a latent structure example. First, a document $n$ is sampled uniformly from the training set. Then, for each possible label $k$ of this document, a cost $\Ep_{\vec {\hat d}
\sim \pi} l((y,\vec d_n), (k,\vec {\hat d}))$ is computed. By definition, the $\vec {\hat d}$ that is computed is exactly the prediction according to the current multinomial estimator, $h^m$. Interpreting the multinomial estimator in terms of the EM parameters, the costs are *precisely* the $z_{n,k}$s from EM (see Eq ). These latent structure examples are fed in to the multinomial naïve Bayes classifier, which re-estimates a model exactly as per the M-step in EM (Eq ).
Next, consider the generation of the document examples. These examples are generated by $\pi$ first choosing a cluster according to the structure classifier. This cluster id is then used as the (only) feature to the “generate document” multinomial. As we saw before, the probability that $\pi$ will select label $k$ for document $n$ is precisely $z_{n,k}$ from Eq . Thus, the multinomial estimator will effectively receive weighted examples, weighted by these $z_{n,k}$s, thus making the maximum likelihood estimate exactly the same as the M-step from EM (Eq ).
Synthetic experiments {#sec:hmm}
---------------------
To demonstrate the advantages of the generality of [<span style="font-variant:small-caps;">Searn</span>]{}, we report here the result of some experiments on synthetic data. We generate synthetic data according to two different HMMs. The first HMM is a first-order model. The initial state probabilities, the transition probabilities, and the observation probabilities are all drawn uniformly. The second HMM is a second-order model, also will all probabilities drawn uniformly. The lengths of observations are given by a Poisson with a fixed mean.
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Model States Truth EM [<span style="font-variant:small-caps;">Searn</span>]{} -NB [<span style="font-variant:small-caps;">Searn</span>]{} -LR
--------------- -------- ----------------------- ------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------- -------------------------------------------------------------
1st order HMM K = 2 0.227 [$\pm 0.107$]{} *0.275 [$\pm 0.128$]{} &**0.287 [$\pm 0.138$]{} &*0.276 [$\pm 0.095$]{}\
1st order HMM &K = 5 &0.687 [$\pm 0.043$]{} &*0.678 [$\pm 0.026$]{} &*0.688 [$\pm 0.025$]{} &**0.672 [$\pm 0.022$]{}\
1st order HMM &K = 10 &0.806 [$\pm 0.035$]{} &*0.762 [$\pm 0.021$]{} &*0.771 [$\pm 0.019$]{} &**0.755 [$\pm 0.019$]{}\
2nd order HMM &K = 2 &0.294 [$\pm 0.072$]{} &0.396 [$\pm 0.057$]{} &0.408 [$\pm 0.056$]{} &**0.271 [$\pm 0.057$]{}\
2nd order HMM &K = 5 &0.651 [$\pm 0.068$]{} &0.695 [$\pm 0.027$]{} &0.710 [$\pm 0.016$]{} &**0.633 [$\pm 0.018$]{}\
2nd order HMM &K = 10 &0.815 [$\pm 0.032$]{} &0.764 [$\pm 0.021$]{} &0.771 [$\pm 0.015$]{} &**0.705 [$\pm 0.019$]{}\
******************
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
\[tab:hmm\]
In our experiments, we consider the following learning algorithms: EM, [<span style="font-variant:small-caps;">Searn</span>]{} with HMM features and a naïve Bayes classifier, and [<span style="font-variant:small-caps;">Searn</span>]{}with a logistic regression classifier (and an enhanced feature space: predicting $y_t$ depends on $x_{t-1:t+1}$. The first [<span style="font-variant:small-caps;">Searn</span>]{} should mimic EM, but by using sampling rather than exact expectation computations. The models are all first-order, regardless of the underlying process.
We run the following experiment. For a given number of states (which we will vary), we generate $10$ random data sets according to each model. Each data set consists of $5$ examples with mean example length of $40$ observations. The vocabulary size of the observed data is always $10$. We compute error rates by matching each predicted label to the best-matching true label and the compute Hamming loss. Forward-backward is initialized randomly. We run experiments with the number of latent states equal to $2$, $5$ and $10$.[^1]
The results of the experiments are shown in Table \[tab:hmm\]. The observations show two things. When the true model matches the model we attempt to learn (HMM1), there is essentially no statistically significant difference between any of the algorithms. Where once sees a difference is when the true model does not match the learned model (HMM2). In this case, we see that [<span style="font-variant:small-caps;">Searn</span>]{}-LR obtains a significant advantage over both EM and [<span style="font-variant:small-caps;">Searn</span>]{}-NB, due to its ability to employ a richer set of features. These results hold over all values of $K$. This is encouraging, since in the real world our model is rarely (if ever) right. The (not statistically significant) difference in error rates between EM and [<span style="font-variant:small-caps;">Searn</span>]{}-NB are due to a sampling versus exact computation of expectations. Many of the models outperform “truth” because likelihood and accuracy do not necessarily correlate [@liang08errors].
Analysis {#sec:guarantees}
========
There are two keys to success in unsupervised-[<span style="font-variant:small-caps;">Searn</span>]{}. The first key is that the features on the $\cY$-component of the output space be descriptive enough that it be learnable. One way of thinking of this constraint is that if we had labeled data, then we would be able to learn well. The second key is that the features on the $\cX$-component of the output space be intrinsically tied to the hidden component. Ideally, these features will be such that $\cX$ can be predicted with high accuracy if and only if $\cY$ is predicted accurately.
The general–though very trivial–result is that if we can guarantee that the loss on $\cY$ is bounded by some function $f$ of the loss on $\cX$, then the loss on $\cY$ is guaranteed after learning to be bounded by $f(L(\pi^*) + 2\ell_{\text{avg}}T_{\text{max}} ln
T_{\text{max}} + c(1 + ln T_{\text{max}})/T_{\text{max}})$, where all the constants now depend on the induced structured prediction problem; see Eq \[eq:searn\].
One can see the unsupervised [<span style="font-variant:small-caps;">Searn</span>]{} analysis as justifying a small variant on “Viterbi training”–the process of performing EM where the E-step is approximated with a delta function centered at the maximum. One significant issue with Viterbi training is that it is not guaranteed to converge. However, Viterbi training is recovered as a special case of unsupervised [<span style="font-variant:small-caps;">Searn</span>]{} where the interpolation parameter is fixed at $1$. While the [<span style="font-variant:small-caps;">Searn</span>]{} theorem no longer applies in this degenerate case, any algorithm that uses Viterbi training could easily be retrofitted to simply make some decisions randomly. In doing so, one would obtain an algorithm that does have theoretical guarantees.
Unsupervised Dependency Parsing {#sec:gi}
===============================
The dependency formalism is a practical and linguistically interesting model of syntactic structure. One can think of a dependency structure for a sentence of length $T$ as a directed tree over a graph over $T +
1$ nodes: one node for each word plus a unique root node. Edges point from heads to dependents. An example dependency structure for a $T =
7$ word sentence is shown in Figure \[fig:deptree\] . To date, unsupervised dependency parsing has only been viewed in the context of global probabilistic models specified over dependency pairs [@paskin01bigrams] or spanning trees [@klein04induction; @smith05induction]. However, there is an alternative, popular method for producing dependency trees in a supervised setting: shift-reduce parsing [@nivre03parsing; @sagae05shiftreduce].
![Dependency parse of a $T=7$ word sentence.[]{data-label="fig:deptree"}](deptree.pdf){width="45.00000%"}
Shift-reduce dependency parsing
-------------------------------
Shift-reduce dependency parsing [@nivre03parsing] is a left-to-right parsing algorithm that operates by maintaining three state variables: a stack $S$, a current position $i$ and a set of arcs $A$. The algorithm begins with $\langle S, i,A\rangle = \langle
\emptyset, 1, \emptyset \rangle$: the stack and arcset are empty and the current index is $1$ (the first word). The algorithm then proceeds through a series of actions until a final state is reached. A final state is one in which $i = T$, at which point the set $A$ contains all dependency edges for the parse. Denote by $i|I$ a stack with $i$ at the head and stack $I$ at the tail. There are four actions:
LeftArc:
: $\langle t|S, i, A \rangle \longrightarrow \langle S,
i, (i, t)|A \rangle$, so long as there does not exist an arc $(\cdot·, t) \in A$. (Adds a left dependency to the arc set between the word $t$ at the top of the stack and the word $i$ at the current index.)
RightArc:
: $\langle t|S, i, A \rangle \longrightarrow \langle
i|t|s, i + 1, (t, i)|A\rangle$, so long as there is no arc $(\cdot·,
i) \in A$. (Adds a right dependency between the top of the stack and the next input.)
Reduce:
: $\langle t|S, i,A \rangle \longrightarrow \langle S,
i,A \rangle$, so long as there does exist an arc $(\cdot·, t) \in
A$. (Removes a word from the stack.)
Shift:
: $\langle S, i, A\rangle \longrightarrow \langle n|S, i +
1,A \rangle$. (Place item on stack.)
This algorithm is guaranteed to terminate in at most $2T$ steps with a valid dependency tree [@nivre03parsing], unlike standard probabilistic algorithms that have a time-complexity that is cubic in $T$ [@mcdonald07complexity]. The advantage of the shift-reduce framework is that it fits nicely into [<span style="font-variant:small-caps;">Searn</span>]{}. However, until now, it has been an open question how to train a shift-reduce model in an unsupervised fashion. The techniques described in this paper give a solution to this problem.
Experimental setup
------------------
We follow the same experimental setup as [@smith05induction], using data from the WSJ10 corpus (sentences of length at most ten from the Penn Treebank [@marcus93treebank]). The data is stripped of punctuation and parsing depends on the part-of-speech tags, not the words. We use the same train/dev/test split as Smith and Eisner: $5301$ sentences of training data, $531$ sentences of development data and $530$ sentences of blind test data. All algorithm development and tuning was done on the development data.
We use a slight modification to SearnShell to facilitate the development of our algorithm together with a multilabel logistic regression classifier, MegaM.[^2] Our algorithm uses the following features for the tree-based decisions (inspired by [@hall06dependency]), where $t$ is the top of the stack and $i$ is the next token: the parts-of-speech within a window of $2$ around $t$ and $i$; the pair of tokens at $t$ and $i$; the distance (discretized) between $t$ and $i$; and the part-of-speech at the head (resp. tail) of any existing arc pointing to (resp. from) $t$ or $i$. For producing word $i$, we use the part of speech of $i$’s parent, grandparent, daughters and aunts.
We use [<span style="font-variant:small-caps;">Searn</span>]{} with a fixed $\be = 0.1$. One sample is used to approximate expected losses. The development set is used to tune the scale of the prior variances for the logistic regression (different variances are allowed for the “produce tree” and “produce words” features). The initial policy makes uniformly random decisions. Accuracy is directed arc accuracy.
Experimental results
--------------------
Algorithm Acc-Tr Acc-Tst \# Iter
---------------------------------------------------------------- -------------------- --------------------- ---------
Rand-Gen 23.5 [$\pm 0.9$]{} 23.5 [$\pm 1.3$ ]{}
Rand-[<span style="font-variant:small-caps;">Searn</span>]{} 21.3 [$\pm 0.2$]{} 21.0 [$\pm 0.6$]{}
K+M:Rand-Init 23.6 [$\pm 3.8$]{} 23.6 [$\pm 4.3$]{} 63.3
K+M:Smart-Init 35.2 [$\pm 6.6$]{} 35.2 [$\pm 6.0$]{} 64.1
S+E:Length 33.8 [$\pm 3.6$]{} 33.7 [$\pm 5.9$]{} 173.1
S+E:DelOrTrans1 47.3 [$\pm 6.0$]{} 47.1 [$\pm 5.9$]{} 132.2
S+E:Trans1 48.8 [$\pm 0.9$]{} 49.0 [$\pm 1.5$]{} 173.4
[<span style="font-variant:small-caps;">Searn</span>]{}: Unsup 45.8 [$\pm 1.6$]{} 45.4 [$\pm 2.2$]{} 27.6
S+E: Sup 79.9 [$\pm 0.2$]{} 78.6 [$\pm 0.8$]{} 350.5
[<span style="font-variant:small-caps;">Searn</span>]{}: Sup 81.0 [$\pm 0.3$]{} 81.6 [$\pm 0.4$]{} 24.4
: Accuracy on training and test data, plus number of iterations for a variety of dependency parsing algorithms (all unsupervised except for the last two rows).
\[tab:dep\]
The baseline systems are: two random baselines (one generative, one given by the [<span style="font-variant:small-caps;">Searn</span>]{} initial policy), Klein and Manning’s model [@klein04induction] EM-based model (with and without clever initialization), and three variants of Smith and Eisner’s model [@smith05induction] (with random initialization, which seems to be better for most of their models). We also report an “upper bound” performance based on supervised training, for both the probabilistic (Smith+Eisner model) as well as supervised [<span style="font-variant:small-caps;">Searn</span>]{}.
The results are reported in Table \[tab:dep\]: accuracy on the training data, accuracy on the test data and the number of iterations required. These are all averaged over $10$ runs; standard deviations are shown in small print. Many of the results (the non-[<span style="font-variant:small-caps;">Searn</span>]{}results) are copied from [@smith05induction]. The stopping criteria for the EM-based models is that the log likelihood changes by less than $10e-5$. For the [<span style="font-variant:small-caps;">Searn</span>]{}-based methods, the stopping criteria is that the development accuracy ceases to increase (on the individual classification tasks, not on the structured prediction task).
All learned algorithms outperform the random algorithms (except Klein+Manning with random inits). K+M with smart initialization does slightly better than the worst of the S+E models, though the difference is not statistically significant. It does so needing only about a third of the number of iterations (moreover, a single S+E iteration is slower than a single K+M iteration). The other two S+E models do roughly comparably in terms of performance (strictly dominating the previous methods). One of them (“DelOrTrans1”) requires about twice as many iterations as K+M; the other (“Trans1”) requires about three times (but has much high performance variance). Unsupervised [<span style="font-variant:small-caps;">Searn</span>]{} performs halfway between the best K+M model and the best S+E model (it is within the error bars for “DelOrTrans1” but not “Trans1”).
Nicely, it takes significantly fewer iterations to converge (roughly $15\%$). Moreover, each iteration is quite fast in comparison to the EM-based methods (a complete run took roughly $3$ hours on a 3.8GHz Opteron using SearnShell). Finally, we present results for the supervised case. Here, we see that the [<span style="font-variant:small-caps;">Searn</span>]{}-based method converges much more quickly to a better solution than the S+E model. Note that this comparison is unfair since the [<span style="font-variant:small-caps;">Searn</span>]{}-based model uses additional features (though it is a nice property of the [<span style="font-variant:small-caps;">Searn</span>]{}-based model that it *can* make use of additional features). Nevertheless we provide it so as to give a sense of a reasonable upper-bound. We imagine that including more features would shift the upper-bound and the unsupervised algorithm performance up.
A Semi-Supervised Version {#sec:semi}
=========================
![Parsing accuracy for semi-supervised, supervised and unsupervised [<span style="font-variant:small-caps;">Searn</span>]{}. X-axis is: (semi/sup) \# of labeled examples; (unsup) \# of unlabeled examples.[]{data-label="fig:semisup"}](semisup.pdf){width="40.00000%"}
The unsupervised learning algorithm described above naturally extends to the case where some labeled data is available. In fact, the only modification to the algorithm is to change the loss function. In the unsupervised case, the loss function completely ignores the latent structure, and returns a loss dependent only on the “predict self” task. In the semi-supervised version, one plugs in a natural loss function for the “latent” structure prediction for the labeled subset of the data.
In Figure \[fig:semisup\], we present results on dependency parsing. We show learning curves for unsupervised, fully supervised and semi-supervised models. The x-axis shows the number of examples used; in the unsupervised and supervised cases, this is the total number of examples; in the semi-supervised case, it is the number of labeled examples. Error bars are two standard deviations. Somewhat surprisingly, with only five labeled examples, the semi-supervised approach achieves an accuracy of over $70\%$, only about $10\%$ behind the fully supervised approach with $5182$ labeled examples. Eventually the supervised model catches up (at about $250$ labeled examples). The performance of the unsupervised model continues to grow as more examples are provided, but never reaches anywhere close to the supervised or semi-supervised models.
Conclusions
===========
We have described the application of a search-based structured prediction algorithm, [<span style="font-variant:small-caps;">Searn</span>]{}, to unsupervised learning. This answers positively an open question in the field of learning reductions [@beygelzimer05reductions]: can unsupervised learning be reduced to supervised learning? We have shown a near-equivalence between the resulting algorithm and the forward-backward algorithm in hidden Markov models. We have shown an application of this algorithm to unsupervised dependency parsing in a shift-reduce framework. This provides the first example of unsupervised learning for dependency parsing in a non-probabilistic model and shows that unsupervised shift-reduce parsing is possible. One obvious extension of this work is to structured prediction problems with additional latent structure, such as in machine translation. Instead of using the predict-self methodology, one could directly apply a predict-target methodology.
The view of “predict the input” for unsupervised learning is implicit in many unsupervised learning approaches, including standard models such as restricted Boltzmann machines and Markov random fields. This is made most precise in the wake-sleep algorithm [@hinton95wakesleep], which explicitly trains a neural network to reproduce its own input. The wake-sleep algorithm consists of two phases: the wake phase, where the latent layers are produced, and the sleep phase, where the input is (re-)produced. These two phases are analogous to the predict-structure phase and the predict-words phase in unsupervised [<span style="font-variant:small-caps;">Searn</span>]{}.
#### Acknowledgements.
Thanks for Ryan McDonald and Joakim Nivre for discussions related to dependency parsing algorithms. Comments from 5 (!) anonymous reviewers were incredibly helpful. This was partially supported by NSF grant IIS-0712764.
Beygelzimer, A., Dani, V., Hayes, T., Langford, J., & Zadrozny, B. (2005). Error limiting reductions between classification tasks. (pp. 49–56).
Brown, P., [Della Pietra]{}, S., [Della Pietra]{}, V., & Mercer, R. (1993). The mathematics of statistical machine translation: Parameter estimation. , [*19*]{}, 263–311.
, H., Langford, J., & Marcu, D. (2009 (to appear)). Search-based structured prediction. .
Dempster, A., Laird, N., & Rubin, D. (1977). Maximum likelihood from incomplete data via the [EM]{} algorithm. , [*B39*]{}, 1–38.
Hall, J., Nivre, J., & Nilsson, J. (2006). Discriminative classifiers for determining dependency parsing. (pp. 316–323).
Hinton, G., Dayan, P., Frey, B., & Neal, R. (1995). The wake-sleep algorithm for unsupervised neural networks. , [*26*]{}, 1158–1161.
Klein, D., & Manning, C. (2004). Corpus-based induction of syntactic structure: Models of dependency and constituency. (pp. 478–485).
Lafferty, J., McCallum, A., & Pereira, F. (2001). Conditional random fields: Probabilistic models for segmenting and labeling sequence data. (pp. 282–289).
Liang, P., & Klein, D. (2008). Analyzing the errors of unsupervised learning. (pp. 879–887).
Marcus, M., Marcinkiewicz, M. A., & Santorini, B. (1993). Building a large annotated corpus of [E]{}nglish: The [P]{}enn [T]{}reebank. , [*19*]{}, 313–330.
McDonald, R., & Satta, G. (2007). On the complexity of non-projective data-driven dependency parsing. (pp. 121–132).
Ng, A., & Jordan, M. (2000). : A policy search method for large [MDP]{}s and [POMDP]{}s. (pp. 406–415).
Nivre, J. (2003). An efficient algorithm for projective dependency parsing. (pp. 149–160).
Paskin, M. A. (2001). Grammatical bigrams. (pp. 91–97).
Rabiner, L. (1989). A tutorial on hidden [Markov]{} models and selected applications in speech recognition. (pp. 257–285).
Ramanan, D., Forsyth, D., & Zisserman, A. (2005). Strike a pose: Tracking people by finding stylized poses. (pp. 271–278).
Sagae, K., & Lavie, A. (2005). A classifier-based parser with linear run-time complexity. .
Smith, N. A., & Eisner, J. (2005). Guiding unsupervised grammar induction using contrastive estimation. (pp. 73–82).
Taskar, B., Chatalbashev, V., Koller, D., & Guestrin, C. (2005). Learning structured prediction models: A large margin approach. (pp. 897–904).
Titov, I., & Henderson, J. (2007). A latent variable model for generative dependency parsing. .
Vogel, S., Ney, H., & Tillmann, C. (1996). -based word alignment in statistical translation. (pp. 836–841).
[^1]: We ran experiments varying the number of samples [<span style="font-variant:small-caps;">Searn</span>]{} uses in $\{1, 2,
5\}$; there was no statistically significant difference. The results we report are based on $2$ samples.
[^2]: SearnShell and MegaM are available at <http://searn.hal3.name> and <http://hal3.name/megam>, respectively.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Light has shown up an incredibe capability in precision measurement based on opto-mechanic interaction in high vacuum by isolating environment noises. However, there are still obstructions, such as displacement and mass estimation error, highly hampering the improvement of absolute accuracy at the nanoscale. Here, we present a nonlinearity based metrology to precisely measure the position and mass of a nanoparticle with optical levitation under $10^{-5}$ mbar, 6-order of magnitude lower than the electrostatic-force and stochastic-force-based counterparts. By precisely controlling the amplitude of the levitated nanoparticle at the nonlinear regime, we realized a feasible sub-picometer-level position measurement with an uncertainty of $1.0\%$ without the prior information of mass, which can be further applied to weigh the femtogram-level mass with an uncertainty of $2.2\%$. It will also pave the way to construct a well-calibrated opto-mechanic platform in high vacuum for high sensitivity and accuracy measurement in force and acceleration at the nanoscale and the study in quantum superposition at the mesoscopic scale.'
author:
- Yu Zheng
- 'Lei-Ming Zhou'
- Yang Dong
- 'Cheng-Wei Qiu'
- 'Xiang-Dong Chen'
- 'Guang-Can Guo'
- 'Fang-Wen Sun'
title: 'Robust optical-levitation-based metrology of nanoparticle’s position and mass'
---
Light has been the most powerful tool for precision metrologies in time, frequency, and distance [@Brewer2019time; @Udem2002frequency; @Yuan2019ruler]. Based on opto-mechanic interaction in high vacuum, the gravitational wave has also been successfully detected [@LIGO2017]. Recently, the compact optical levitation in vacuum, which joins the fields of optomechanics [@Aspelmeyer2014cavity] and optical trapping [@Ashkin1980radiation; @Ashkin1986tweezer; @Gao2017qiuchengwei], is being put into the spotlight of high precision metrology for modern science, especially at the nanoscale. It extends the optical precision metrology in force [Ranjit2016force,Hempston2017force,Hebestreit2018freefall,Blakemore2019force]{}, mass [@Blakemore2019mass; @ricci2019mass], charge [@Moore2014charge] and acceleration [@Monteiro2017acc]. Such a system has also been considered as a promising platform for the investigation of quantum superposition at the mesoscopic scale [@romeroisart2010quantum; @Romero2011quantum; @Bateman2014nearfield], which may further enhance the performance of precision metrology due to the nature of quantum superposition and entanglement [@Giovannetti2011].
![Scheme of position measurement without the information of particle mass. The measured nonlinearity induced relative natural frequency shift ${\Delta\Omega}/{\Omega_{0}}$, which can be obtained from the spectrum analysis of the voltage signal of detector $V(t)$, is related to the actual oscillation amplitude $q_{A}$ as $q_{A}=w\sqrt{-4\Delta\Omega/(3{\Omega_{0}})}$. With the respective measured voltage signal amplitude $V_{A}$, the calibrated factor $c=q_{A}/V_{A}$ is got with only the information about $V(t)$. Then the position signal can be obtained as $q(t)=c\cdot V(t)$. []{data-label="fig:1"}](fig1.pdf){width="45.00000%"}
It had been regarded that the optical levitation has a high precision in position detection with state-of-the-art techniques, which required a priori knowledge of the particle mass [@hebestreit2018] or the assistance of stochastic [@berg2004power; @Hauer2013Calib; @hebestreit2018] or extra electrostatic [@hebestreit2018; @ricci2019mass] forces at a moderate vacuum to calibrate the experimental output. However, in the experiment, the mass of the nanoparticle and stochastic and electrostatic forces can be hardly estimated or measured precisely at the nanoscale [@millen2014temp; @Frangeskou2016diamond; @Hebestreit2018temp], which severely reduces the accuracy of the position and other metrologies with optical levitation system.
Here, we present an all-optical metrology for position and mass measurement with optical levitation in high vacuum. The geometry shape of the optical potential is regarded as a calibration gauge. The deviation from quadratic shape potential will induce a nonlinear natural frequency shift [@gieseler2013nonlinear], which acts as the ruler mark to precisely read out the position and the motion of the nanoparticle without a priori knowledge of mass. Furthermore, we are able to weigh the mass and measure the density of the optically levitated nanoparticle. This high accuracy position and mass measurement can help to construct a well-calibrated opto-mechanic platform for high sensitivity and accuracy measurement in force and acceleration.
![(a) Schematics of the parametric feedback locking. When the measured amplitude ($V_A$) of the trapped particle is larger (smaller) than the target amplitude ($V_{target}$), modulations will be applied with feedback cooling (heating) to reduce (increase) the amplitude. The solid curves represent the nonlinear Gaussian potential which deviate from harmonic potential (dashed curves). (b) Part of the recorded signal time trace from an amplitude locked particle at a pressure of $10^{-5}$ mbar. (c) Simplified experimental setup. A $1064$ nm laser was focused by an objective which was mounted in vacuum chamber to form the optical potential for particle trapping. The forward scattering light from optical trap was collected and sent to c.m. motion detection unit. The detector signal was processed to generate the feedback amplitude locking by modulating the trapping laser intensity with an acousto-optical modulator (AOM).[]{data-label="fig:2"}](fig2.pdf){width="45.00000%"}
An optically levitated sensor realizes its functions by analyzing the motions of the trapped particle. The optical potential is harmonic and the oscillation has a fixed natural frequency $\Omega_{0}$ (shown as the purple dashed potential shape in Fig. \[fig:1\]) when the particle is trapped very near the equilibrium point. However, in an optical trap built by a laser beam in Gaussian mode [gieseler2013nonlinear,novotny2012principles]{}, the optical potential will become anharmonic when the oscillator has a large amplitude and can move far from the equilibrium point (shown as the red solid potential shape in Fig. \[fig:1\]). This anharmonic potential features a Duffing nonlinearity [gieseler2013nonlinear]{}. Ignoring the interaction between different spatial degrees of freedom, the motion of the trapped particle with mass $m$ in this nonlinear regime can then be described as $$\ddot{q}+\Gamma_{0}\dot{q}+\Omega_{0}^{2}\left( q+\xi q^{3}\right)=F_{therm}/m%
\text{,} \label{motion}$$ in one degree of freedom, where $q(t)$ is the position of the trapped particle. $\Gamma_{0}$ is the damping rate induced by air. $\Omega_{0}$ is the natural angular frequency when the oscillation is in the linear regime. $F_{therm}$ is the Brownian stochastic force. $\xi$ describes the nonlinear coefficient of the trap. For a Gaussian distribution optical potential, $\xi=-{2}/{w^{2}}$, where $w$ is the $1/\mathrm{e}^{2}$ beam intensity radius. The position signal $q(t)=c\cdot V(t)$ is expected to be experimentally measured by a photodetector output voltage $V(t)$ and a calibration constant $c$.
It is noticed that we can get the calibration constant $c$ with only the voltage signal $V(t)$. Due to the nonlinearity, the oscillator has got a natural frequency shift $\Delta\Omega=-\frac{3}{4}{w}^{2}q_{A}^{2}\Omega_{0}$ according to the theoretical solution of Eq. (\[motion\]), when the amplitude of the oscillator is $q_{A}$. As both $\Omega_{0}$ and $\Delta\Omega$ can be obtained from the spectrum analysis of $V(t)$, we can get $q_{A}$ without other information. Along with the respective measured voltage signal amplitude $V_{A}$ , the calibration constant $c=q_{A}/V_{A}$ is determined. In more details, in experimental implementation to obtain the calibration constant $c$ with high accuracy, we can fit the dependence of the relative natural frequency shift on the voltage amplitude $$\frac{\Delta \Omega}{\Omega_{0}}=\alpha V_A^{2}\text{,} \label{alpha}$$with a nonlinear coefficient $\alpha$. Then the calibration constant $c=\sqrt{-\frac{4}{3}\alpha w^{2}}$, which can be figured out from $\alpha$ and $w$.
In order to measure $\alpha $ based on Eq. (\[alpha\]), we need to measure the frequency shift under different voltage amplitudes. Although some existing methods can be employed, such as picking out divided trajectory parts with different amplitudes of a thermally driven particle [@gieseler2013nonlinear], double-frequency parametric driving [gieseler2013nonlinear,Gieseler2014nonlinear,ricci2017bistable]{}, and optical tweezer phonon laser [@Pettit2019phonon], the imprecise amplitude control severely limits the performance and reliability of such methods for calibration.
{width="95.00000%"}
Here, we introduce a parametric feedback amplitude locking modulation to precisely control the oscillation amplitude, which corresponds to the center of mass (c.m.) motion temperature. As shown in Fig. \[fig:2\](a), when the amplitude is higher or lower than a target amplitude, a parametric feedback cooling [@Gieseler2012Cooling; @Vovrosh2017Cooling] or heating [zheng2019cooling,Vovrosh2017Cooling]{} process is applied to drive the amplitude to the target. In detail, assuming the short term time trace of the oscillation’s voltage signal shown in Fig. \[fig:2\](b) can be described as $u_{\sin }(t)=V_A\sin \left[
\Omega _{0}t+\varphi \right] $, where $\varphi $ is the phase of the oscillation. And a $\mathrm{\pi }/2$-phase-shifted signal is $u_{\cos
}(t)=V_A\cos \left[ \Omega _{0}t+\varphi \right] $. The applied trapping laser intensity for feedback amplitude locking modulation is $$I(t)=I_{0}\left\{ 1+A(t)\times 0.5\eta _{\text{m}}\times \mathrm{sign}\left[
u_{\sin }(t)u_{\cos }(t)\right] \right\} \text{,} \label{modulation}$$where $I_{0}$ is the laser intensity without modulation and $\eta _{\text{m}%
} $ is the modulation depth [@zheng2019cooling]. $A(t)=\mathrm{sign}( V_A^{2} -V_{\text{target}}^{2})$ is the amplitude criterion. When $V_A$ is larger (smaller) than the target voltage amplitude $V_{\text{%
target}}$, $A$ is positive (negative) and Eq. (\[modulation\]) will be a feedback cooling (heating) modulation to decrease (increase) the amplitude of the levitated oscillator [@zheng2019cooling].
In the experiment, as shown in Fig. \[fig:2\](c), an objective ($\mathrm{NA}=0.9$) was mounted in the vacuum chamber to create an optical potential by focusing the Gaussian trapping laser. To avoid the interaction between motional degrees of freedom when more than two axes of oscillation reach the nonlinear regime [@gieseler2013nonlinear; @Gieseler2014nonlinear], only one axis motion will be amplitude-locked at a time, and the motions of the other two axes will be cooled to sub-Kelvin [@zheng2019cooling]. The modulation depth $ \eta _{\text{m}} $ was set to be 0.5% during the experiment.
Figure \[fig:3\](a) shows the measured signal voltage time traces of a trapped particle under different locked amplitudes and feedback cooling along the Y axis at a pressure of $10^{-5} $ mbar. The relative standard deviation of the voltage amplitude for such an amplitude-locked oscillator was lower than $0.5\%$, as shown in Fig. \[fig:3\](b). And the power spectral densities (PSD) with different locked amplitudes confirmed that the natural frequency of the oscillator in an anharmonic optical potential decreased with increasing amplitude, as shown in Fig. \[fig:3\](c).
Based on the feedback amplitude locking technique at the nonlinear regime, the oscillator frequency was measured while the amplitude locking target was swept from $0.4$ $\mathrm{V}$ to $1.6$ . And the measurement results were averaged from tens of sweep cycles to eliminate the noise and low-frequency drift due to the system instability. This measurement procedure was applied to both $X$ and $Y$ axes to verify the reliability of the calibration process. The dependence of relative frequency shift on voltage amplitude is shown in Fig. \[fig:3\](d). By fitting it with Eq. (\[alpha\]), we can get that the nonlinear coefficients of $%
X $-axis and $Y$-axis are $\alpha _{\mathrm{X}}=-0.002978\text{ }\mathrm{V}%
^{-2} $ and $\alpha _{\mathrm{Y}}=-0.002669\text{ }\mathrm{V}^{-2}$, respectively.
To complete the high precision calibration, the intensity distribution in the $X-Y$ plane was simulated with Debye integral [gieseler2013nonlinear,novotny2012principles]{}, as shown in Figs. \[fig:3\](e) and (f). By fitting the intensity along $X$ and $Y$ axes with Gaussian function, we get the $1/\mathrm{e}^{2}$ intensity radius $w_{\mathrm{X}%
}=703\pm 7\text{ }\mathrm{nm}$ and $w_{\mathrm{Y}}=551\pm 22\text{ }\mathrm{%
nm}$, respectively. The errors come from the uncertainty of optical component parameters (see Supplementary Information for details [@SI]). With $\alpha _{\mathrm{X}}$, $\alpha _{\mathrm{Y}%
}$, $w_{\mathrm{X}}$, and $w_{\mathrm{Y}}$, we finally extracted the calibration constants of $c_{\mathrm{X}}=44.3\pm 0.5%
\text{ }\mathrm{nm}/\mathrm{V}$ along $X$ axis and $c_{\mathrm{Y}}=32.9\pm
1.3\text{ }\mathrm{nm}/\mathrm{V}$ along $Y$ axis. In Table.\[T1\], we compare the uncertainties of calibration constants with different calibration methods. It indicates that the uncertainty of position measurement was about $1.0\%$ at the sensitivity level of $0.44$ $\mathrm{pm}/\sqrt{\mathrm{Hz}}$ (X-axis), which was much lower than that from other methods [@hebestreit2018; @ricci2019mass]. Also, the operation pressure is $6$ orders of magnitude lower, which can be applied to enhance the sensitivity due to high environment noises isolation.
Calibration criterion Primary uncertainty source Operational pressure Relative uncertainty
----------------------------------------------------- ---------------------------- ---------------------- ----------------------
Thermal stochastic force [@hebestreit2018] Mass $10$ mbar to $1$ atm $15\%$
Electrostatic force [@hebestreit2018] Mass $<10$ mbar $30\%$
Stochastic and electrostatic force [@ricci2019mass] Electric-field strength $\sim50$ mbar $1.2\%$
Potential nonlinearity \[This work\] Potential geometry shape $10^{-5}$ mbar $1.0\%$
\
\[T1\]
Since we can get the calibrated trajectory of the trapped particle with high accuracy, it is possible to extract more information about the levitated oscillator from its thermal motion such as the calibration constants of other axes, mass and density of the trapped nanoparticle. For a nano-mechanical resonator sensor, the force measurement uncertainty highly depends on the accuracy of the mass of the trapped nanoparticle. Based on the calibration constant, we can obtain the mass of the nanoparticle from the mean square displacement along one axis in a thermal equilibrium, which follows $$m=\frac{k_{B}T}{\left\langle q_{\mathrm{X}%
}^{2}\right\rangle \Omega_{\mathrm{X0}}^{2}}\text{,} \label{thermal}$$where $k_{B}$ is the Boltzmann constant and $T$ is the temperature of surrounding environment.
![(a) Measured mass under different air pressures. The error bars came from the systematic errors. The dashed line is the average of the measurement results. (b) Measured damping coefficients $\Gamma_{0}$, which were averaged from X, Y, and Z axes as a function of air pressure. The solid line is a linear fitting. The error bars came from the variance of $\Gamma_{0}$ between different axes.[]{data-label="fig:4"}](fig4.pdf){width="45.00000%"}
In the experiment, the thermal motion properties, such as mean square signal voltage ($%
\left\langle V_{\mathrm{X}}^{2}\right\rangle =\left\langle q_{\mathrm{X}%
}^{2}\right\rangle /c_{\mathrm{X}}^{2}$), natural angular frequency ($\Omega_{\mathrm{%
X0}}^{2}$) and damping coefficient ($\Gamma_{0}$ along X, Y, and Z axes), were measured at a pressure between $20$ to $50$ by fitting the PSD. The pressure range we selected is to avoid the measurement errors due to the nonlinear distortion of PSD, effective bath temperature elevating under high vacuum and fitting error near overdamped regime. Before measurement, the vacuum chamber was exhausted below $0.1$ for a while and re-inflated to minimize the influence of particle properties varying at low gas pressure like the evaporation of adsorbed solvent. From the experiment measurement, we got the mass of the trapped nanoparticle was $m=(3.63\pm 0.08)\times 10^{-18}$ with Eq. ([thermal]{}) [@SI].
Moreover, the density of nanoparticle is an important parameter for its industrial and technological applications. Also the density of a nanoparticle is related to its refractive index [@jackson1962classical] which is one of the key parameters for optical levitation properties. However, it is difficult to measure the density of an individual nanoparticle. And for the most popular Stöber silica nanospheres, its density can be smaller than a bulk glass due to its porosity [@Parnell2016Porosity; @Blakemore2019mass]. As for the density measurement, when the Knudsen number $\mathrm{Kn}\gg1$, where $\mathrm { Kn } = \overline { l } / R$, $\overline { l }$ is the mean free path of air molecules, and $R$ is the radius of the nanosphere, the air damping coefficient of a nanosphere follows [@Epstein1924damping] $$\Gamma _{0}=\frac{8}{3m}\left(1+\frac{a \pi}{8}\right) \sqrt{\frac{2 \pi m_{g}}{ k_B T}} p R^{2} \text{,}
\label{pauli}$$where $p$ is the air pressure, $m_g$ is the mass of the air molecule, and $a$ is the momentum accommodation coefficient. With the measured damping coefficients in Fig. \[fig:4\] and Eq. (\[pauli\]), we can get that the radius of the trapped nanosphere was $%
R=75.4\pm 1.5\text{ }\mathrm{nm}$ and the density was $\rho =2.02\pm 0.10$ $%
\mathrm{g}/\mathrm{cm}^{3}$ [@SI]. This result agreed with the value provided by the manufacturer ($\rho_{\mathrm{Bangs}}\simeq2.0\text{ }\mathrm{g}/\mathrm{cm}^{3}$). As for a non-spherical nanoparticle, the dependence of the damping rate on the pressure can be obtained by the direct simulation with Monte Carlo method [@Ahn2018GHz].
In conclusion, we have introduced a nonlinear frequency shift based measurement of the position, mass and density of a nanoparticle with optical levitation in high vacuum. We are able to control the amplitude of an optically levitated oscillator with tiny deviation and make it possible to deploy a precise nonlinear frequency shift measurement. Such a method does not require the mass of nanoparticle or an assistance from an external force. The absolute precision is then mainly limited by the error from the tight focusing light field estimation, which can be further measured with high precision in experiment [@bauer2014light; @Neugebauer2015light; @bautista2016light]. Moreover, an amplitude locked nano-oscillator can be regarded as a nearly ideal harmonic oscillator for classical and quantum investigation. It is possible to transform a position or velocity depended static interaction, such as Casimir force, electric field gradient, and Lorentz force, into a harmonic force, since an optically levitated nanoparticle sensor has an incredible sensitivity for resonant force measurement. Such an amplitude locked optically levitated nanoparticle can also be applied in the study of nonequilibrium physics and thermodynamics at the nanoscale [@jones2015tweezer; @Gieseler2018therm].
Methods {#methods .unnumbered}
=======
[**Experimental setup.**]{} An objective ($\mathrm{NA}=0.9$) was mounted in the vacuum chamber to focus the $1064$ $\mathrm{nm}$ Gaussian beam laser ($\sim 200$ ). Before the objective, the beam diameter was $4.5$ , which was larger than the back pupil diameter ($3.6$ ) of the objective to make full use of the numerical aperture. An acousto-optic modulator (AOM) was mounted to modulate the laser intensity based on the feedback control signal. The forward scattering light was collected by an aspheric lens ($%
\mathrm{NA}=0.546$) and sent to three sets of balanced photodetectors to measure the positions of the trapped nanoparticle along three motional degrees (set as $X$, $Y$, $Z$ axis) of freedom. A silica nanosphere (nominal radius $82\pm10$ , Bangs labs Inc.) dispersed in ethanol was sent to the optical trap with a nebulizer. The position voltage signals were recorded by a digitizer on computer and simultaneously sent to a field programmable gate array (FPGA) board (Xilinx Spartan-6 XC6SLX16 Core Board) [@zheng2019cooling] to generate the feedback modulation signal that can control the oscillation amplitude along each axis.
Acknowledgment {#acknowledgment .unnumbered}
==============
This work is supported by the Science Challenge Project (No. TZ2018003), the National Natural Science Foundation of China (Nos. 91536219, 61522508, and 91850102), the Anhui Initiative in Quantum Information Technologies (No. AHY130000).
[10]{} url \#1[`#1`]{} urlprefix \[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
*et al.* . ** ****, ().
, & . . ** ****, ().
& . ** ****, ().
*et al.* . ** ****, ().
, & . ** ****, ().
. ** ****, ().
, , & . ** ****, ().
*et al.* . ** ****, ().
, , & . ** ****, ().
*et al.* . ** ****, ().
, , & . ** ****, ().
*et al.* . ** ****, ().
*et al.* . ** ****, ().
, , , & ** ****, ().
, & . ** ****, ().
, , & . ** ****, ().
, , & . ** ****, ().
*et al.* . ** ****, ().
, , & . ** ****, ().
, & . ** ****, ().
*et al.* ** ****, ().
& . ** ****, ().
, , & . ** ****, ().
, , & . ** ****, ().
*et al.* . ** ****, ().
, , & . ** ****, ().
, & . ** ****, ().
& ** (, ).
, , & . ** ****, ().
*et al.* . ** ****, ().
*et al.* . ** ****, ().
, , & . ** ****, ().
*et al.* . ** ****, ().
, & . ** ****, ().
, & ** (, ).
.
** (, ).
*et al.* . ** ****, ().
. ** ****, ().
*et al.* . ** ****, ().
, , , & . ** ****, ().
, , & . ** ****, ().
& . ** ****, ().
& . ** ****, ().
Experimental setup
==================
As shown in Fig. \[fig:S1\], a $1064$ nm laser was used to trap the nanoparticle in vacuum chamber. Its intensity was modulated with an acousto-optical modulator (AOM). The laser beam was expanded to fulfill the back pupil of the objective (Nikon CFI LU Plan Fluor EPI 100X). The forward scattering light from the optical trap was collected by an aspheric lens (NA$=0.546$) and sent to three sets of balanced photodetectors to measure the positions of the trapped nanoparticle along three motional degrees (set as X, Y, Z axis) of freedom.
{width="80.00000%"}
The position voltage signals were sent to a FPGA board to generate the feedback control signal. The digitized signal first passed a Kalman filter to suppress the out-band noise. And then, to eliminate the influence of feedback-loop time delay, two suitable delays were added to the voltage signal to generate a phase-matched signal $ u_{\mathrm{sin}} $ and a $\pi/2$ phase-shifted signal $ u_{\mathrm{cos}} $. $ u_{\mathrm{sin}} $ and $ u_{\mathrm{cos}} $ were sent to generate the feedback cooling signal with $ S_{cool}=\mathrm{sign}(u_{\mathrm{sin}}u_{\mathrm{cos}}) $ [@zheng2019cooling]. Simultaneously, the square of signal voltage amplitude $V_A^2$ was calculated with $ V_A^2=u^2_{\mathrm{sin}}+u^2_{\mathrm{cos}} $. The amplitude was processed in square form because it is not convenient to do square root with FPGA. Then $V_A^2$ was compared with the target amplitude $V_{\mathrm{target}}^2$. If $V_A^2\geq V_{\mathrm{target}}^2$, the cooling signal $ S_{cool} $ would not be changed. Else if $V_A^2<V_{\mathrm{target}}^2$, $ S_{cool} $ would be reversed and become a feedback heating signal. The feedback control signals along three axes were merged based on majority rule [@zheng2019cooling]. Finally, based on the modulation depth $ \eta_\mathrm{m} $, the merged three-dimensional control signal was processed and converted to an analog voltage signal to control the laser intensity with AOM.
{width="80.00000%"}
During the calibration, to obtain the nonlinear coefficient $\alpha$, the square of voltage amplitude was swept between $0.16\text{ }\mathrm{V}^2$ to $2.56\text{ }\mathrm{V}^2$. This voltage sweeping was applied to X and Y axis motions as shown in Fig. \[fig:S2\]. The frequency shift versus amplitude voltage was averaged from tens of sweeping cycles to reduce noises. Such calibration was not applied to Z-axis because the nonlinear property along Z-axis was different from X and Y axes due to the nonlinearity of the scattering force and tremendously affected by changes in equilibrium position. The equilibrium position was related to the radius and refractive index of the silica nanoparticle which were unable to obtain accurately.
Light field estimation
======================
The Debye integral [@novotny2012principles] was utilized to make the simulation of the tightly focused light field around trapping position. The simulation condition we used was that the input laser beam has a Gaussian intensity distribution with a beam diameter of $4.5$ mm which was measured with a CCD camera beam profiler (newport LBP2). The back pupil diameter of the objective was $3.6$ mm. The laser wavelength was $1064$ nm. And the laser was linearly polarized along X axis. the numerical aperture (NA) of the objective was $0.9$. The equilibrium position along Z-axis was $ z_{eq}\simeq100 $ nm. We fitted the intensity distribution at $ z_{eq}=100 $ nm with Gaussian function along X and Y axes and with $I(z)=A/[1+(z+z_0)^2] $ along Z axis. The fitted $1/\mathrm{e}^{2}$ intensity radius along X and Y axes were $w_{\mathrm{X}}=696.6\text{ }\mathrm{nm}$ and $w_{\mathrm{Y}}=530.0\text{ }\mathrm{nm}$, and the Rayleigh range was $z_0=1080.0\text{ }\mathrm{nm}$.
According to the optical force with Rayleigh approximation and the light field estimation result, we were able to calculate the natural frequency ratio between each axes. Based on the above light field estimation results, the natural frequency ratio between X, Y, and Z should be $1.31:1:2.93$, comparing with the experimental result $1.27:1:3.07$.
There were two possible reasons for the differences between the estimation and experimental results. First, the objective we used was designed for visible light. It introduced a slight deterioration of the focal length when it was used for near-infrared light. Second, the actual intensity distribution of the input laser was deviated from a perfect Gaussian distribution and there might be a slight misalign between the input laser and backpupil of the objective. It would introduce an error of the effective input laser diameter. The above reasons introduced simulation condition errors about the NA and the fill factor (input laser diameter/back pupil diameter). By tuning the simulation conditions, when the NA=$0.875$ and the input laser diameter was $4.2$ mm, the simulation result of the natural frequency ratio will be as same as the experimental result. Therefore, the simulated $1/\mathrm{e}^{2}$ intensity radius along X and Y axes were estimated to $w_{\mathrm{X}}=703\pm 7\text{ }\mathrm{nm}$ and $w_{\mathrm{Y}}=551\pm 22\text{ }\mathrm{nm}$. The errors demonstrated the difference between the experiment condition simulation and natural frequency ratio corrected simulation. Those errors can be further reduced in experiment [@bauer2014light; @Neugebauer2015light; @bautista2016light].
Error analysis of the mass, radius, and density measurement
===========================================================
To measure the mass, radius, and density of the trapped nanosphere, a total of $12$ data points at a pressure between $20$ to $50$ mbar are recorded. The recorded information of each data point was derived by fitting the power spectral density (PSD) of position signals, including the mean square signal voltage, natural frequency, and damping coefficient along X, Y, and Z axes.
Because the calibration error along X axis was less than Y axis, the mean square signal voltage ($\left\langle V_{\mathrm{X}}^{2}\right\rangle$) and natural angular frequency ($\Omega_{\mathrm{X0}}$) along X axis was utilized to obtain the mass. Here, $$m=\frac{k_{B}T}{c_\mathrm{X}^2\left\langle V_{\mathrm{X}%
}^{2}\right\rangle \Omega_{\mathrm{X0}}^{2}}\text{,}
\label{thermal}$$where $k_B$ is the Boltzmann constant, $T$ is the environment temperature, and $c_\mathrm{X}$ is the calibration factor along X. Based on the uncertainties shown in Table. \[density\] and the propagation of uncertainty, we can get that $m=(3.63\pm 0.08)\times 10^{-18}$ .
Quantity Value $z_i$ Error $\delta_{z_i}/z_i$
------------------------------------------------ ------------------------------------------------------- --------------------------
$c_\mathrm{X}$ 44.3 $\mathrm{nmV^{-1}}$ 0.010
$\Omega_{\mathrm{X0}}/(2\pi)$ 135450 Hz 0.001(stat)
$\left\langle V_{\mathrm{X}}^{2}\right\rangle$ 0.7980 $\mathrm{V^2}$ 0.005(stat)
$\Gamma_{0}$ $(0.942\sim2.42)\times 10^{5}\text{ }\mathrm{s^{-1}}$ 0.010
$p$ $(1.92\sim4.99)\times 10^{3}\text{ }\mathrm{Pa}$ 0.002
$a$ 0.9 0.111
$T$ 298 K 0.003
$m_g$ $4.8089\times 10^{-26}\text{ }\mathrm{kg}$ $-$
$k_B$ $1.3806\times 10^{-23}\text{ }\mathrm{JK^{-1}}$ $-$
$m$ $3.63\times 10^{-18}\text{ }\mathrm{kg}$ 0.022(sys), 0.004(stat)
$R$ $75.4\times 10^{-9}\text{ }\mathrm{m}$ 0.020(sys), 0.004(stat)
$\rho$ $2.02\text{ }\mathrm{g}/\mathrm{cm}^{3}$ 0.052(sys), 0.013(stat)
\
As for the radius, the most often used formula to estimate the damping coefficient in optically levitation is [@Beresnev_Chernyak_Fomyagin_1990; @Li2013] $$\Gamma_{0}=\frac{6 \pi \eta R}{m} \frac{0.619}{0.619+K n}\left(1+c_{K}\right) \text{,}
\label{knudsen}$$ where $R$ is the radius of the nanosphere, $c _ { K } = 0.31 \mathrm { Kn } / \left( 0.785 + 1.152 \mathrm { Kn } + \mathrm { Kn } ^ { 2 } \right)$, $\eta$ is the viscosity coefficient of air, and $\mathrm { Kn } = \overline { l } / R$ is the Knudsen number. $\overline { l } = k _ { B } T / \sqrt { 2 } \pi d ^ { 2 } p$ is the mean free path of air molecules, where $p$ is the air pressure and $d$ is the collision diameter of air molecules. And the correctness of Eq. (\[knudsen\]) is under the condition that the momentum accommodation coefficient $a$ of the target sphere is equal to $1$. Some experiments have recommended that $a\sim0.9$ [@Ewart2007accommodation]. So we turn to the Epstein’s work that the damping coefficient in the free-molecule regime $(\mathrm {Kn}\gg1)$ is [@Epstein1924damping] $$\Gamma _{0}=\frac{8}{3m}\left(1+\frac{\pi a}{8}\right) \sqrt{\frac{2 \pi m_{g}}{ k_B T}} p R^{2} \text{,}
\label{Epstein}$$where $m_g$ is the mass of the air molecule. We can obtain the radius of the nanosphere with $${R}=\left(\frac{\Gamma_{0}}{p}\right)^{\frac{1}{2}}\left(\frac{8+\pi a}{3 m}\right)^{-\frac{1}{2}}\left(\frac{2 \pi m_{g}}{k_{B} T}\right)^{-\frac{1}{4}} \text{.}
\label{radius}$$
Based on the uncertainties shown in Table. \[density\] and the propagation of uncertainty, the radius of the trapped nanosphere was estimated to be $R=75.4\pm1.5$ nm. As for the density, with Eq. (\[radius\]), we have $$\rho=\frac{3}{4 \pi}m^{-\frac{1}{2}}\left(\frac{\Gamma_{0}}{p}\right)^{-\frac{3}{2}} \left(\frac{8+\pi a}{3}\right)^{\frac{3}{2}}\left(\frac{2 \pi m_{g}}{k_{B} T}\right)^{\frac{3}{4}} \text{.}$$ We can get the density of the trapped nanoparticle was $\rho=2.02\pm0.10\text{ }\mathrm{g}/\mathrm{cm}^{3}$.
[1]{} url \#1[`#1`]{} urlprefix \[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, & . ** ****, ().
& ** (, ).
, , , & . ** ****, ().
, , & . ** ****, ().
& . ** ****, ().
, & . . ** ****, ().
** (, , ).
, , & . ** ****, ().
. ** ****, ().
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'While LDA+U method is well established for strongly correlated materials with well localized orbitals, its application to weakly correlated metals is questionable. By extending the LDA Stoner approach onto LDA+U, we show that LDA+U enhances the Stoner factor, while reducing the density of states. Arguably the most important correlation effects in metals, fluctuation-induced mass renormalization and suppression of the Stoner factor, are missing from LDA+U. On the other hand, for [*moderately*]{} correlated metals LDA+U may be useful. With this in mind, we derive a new version of LDA+U that is consistent with the Hohenberg-Kohn theorem and can be formulated as a constrained density functional theory. We illustrate all of the above on concrete examples, including the controversial case of magnetism in FeAl.'
author:
- 'I. I. Mazin'
- 'A. G. Petukhov'
- 'L. Chioncel'
- 'A. I. Lichtenstein'
bibliography:
- 'stoner.bib'
title: Correlated metals and the LDA+U method
---
One of the most influential, from practical point of view, developments in the Density Functional Theory (DFT) in the last two decades was the LDA+U method (see, *e.g.,* Ref. [@LDA+U]). This method includes the orbital dependence of the self-energy operators, missing from the Kohn-Sham potential, in a relatively crude, pseudo-atomic way, neglecting the fine details of the spatial variations of the Coulomb potential. On the contrary, the standard Local Density Approximation, LDA accounts for the spatial variation of the Hartree potential exceedingly well, but neglects the orbital dependence of the Coulomb interaction. There is one inherent ambiguity in the LDA+U method: In LDA, all electron-electron interactions have already been taken into account in a mean field way. The Hubbard Hamiltonian that represents the underlying physics of the LDA+U method also incorporates a large part of the total Coulomb energy of the system. Simple combination of the LDA and Hubbard Hamiltonian thus leads to a double counting (DC) of the Coulomb energy, so one may want to identify those parts of the DFT expression for the total energy that correspond to the interaction included in the Hubbard Hamiltonian and subtract them. However, since the DFT Hamiltonian is written in terms of the total density, and the Hubbard Hamiltonian in the orbital representation, one cannot build a direct link between the two. Second, even if it were possible, that would be undesirable. Spatial variation of the Hartree and the exchange-correlation potentials is very important. It would be unreasonable to subtract that out just because it has been already taken into account elsewhere in a primitive way (roughly speaking, $UN^{2}/2)$. Rather, one wants to identify the mean-field part of the Hubbard Hamiltonian, and subtract *that*, leaving only a *correction* to the LDA-type mean field solution.
This is not a uniquely defined procedure. Several recipes exist, and it has been appreciated lately [@Novak] that the results of LDA+U calculations may depend crucially on the choice of the DC recipe. It should be noticed that while for strongly correlated systems the LDA+U ideology is at least practically established, in a relatively new area of applying LDA+U to moderately-correlated, metallic systems [@Shick01; @Novak; @Yang], the situation is very far from clear.
In this Letter we analyze the effect of different DC prescription on the LDA+U results in correlated metals. We also present a systematic approach to the DC problem, of which the existing recipes are special cases. Finally, we discuss which problems associated with this class of materials can, in principle, be solved within LDA+U, and which cannot.
We use for our analysis the spherically averaged form of the rotationally-invariant LDA+U [@Liechtenstein], due to Dudarev *et al.* [@Dudarev]: $$\begin{aligned}
\Delta H_{LDA+U}^{0} =\frac{U}{2}\!\!\sum_{m\sigma \neq m^{\prime }\sigma
^{\prime }}\!\!n_{m\sigma }n_{m^{\prime }\sigma ^{\prime }}\!-\!\frac{J}{2}%
\!\!\sum_{m\neq m^{\prime },\sigma }\!\!n_{m\sigma }n_{m^{\prime }\sigma }
\nonumber \\
=\frac{1}{2}UN^{2}-\frac{1}{2}J\sum_{\sigma }N_{\sigma }^{2}-\frac{1}{2}%
(U-J)\sum_{\sigma }\mathrm{Tr}(\rho ^{\sigma }\cdot \rho ^{\sigma })
\label{U0-}\end{aligned}$$where $U$ and $J$ are spherically averaged Hubbard repulsion and intraatomic exchange for electrons with the given angular momentum $l$, $n_{m\sigma }$ is the occupation number of the $m$-th orbital, $\sigma =\pm 1$ is the spin index, and the superscript $0$ means that the double counting terms have not been subtracted yet. Here $\rho _{mm^{\prime }}^{\sigma }$ is the orbital occupation matrix, $N_{\sigma }=\mathrm{Tr}(\rho ^{\sigma })$ and $%
N=\sum_{\sigma }N_{\sigma }$.
To subtract from Eq. (\[U0-\]) the DC term, one naturally starts with the first two terms in Eq. (\[U0-\]), $i.e.,$ the Hartree and the Stoner energies. Both are explicit functionals of the spin density, and are likely to be better described by LDA. To identify the DC part of the last term of the Eq. (\[U0-\]), which explicitely depends on $n_{m\sigma },$ is less trivial; one needs to work out a mean field approximation to this term, that is, substitute $\mathrm{Tr}(\rho ^{\sigma }\cdot \rho ^{\sigma })$ by some quantity $x_{\sigma }$ that depends solely on total spin density. Czyżyk and Sawatzky [@ChyzhikSawatsky] suggested that $x_{\sigma }$ should be equal to $\mathrm{Tr}(\rho ^{\sigma
}\cdot \rho ^{\sigma })$ in the limit of the uniform occupancy, $\rho
_{mm^{\prime }}^{\sigma ,LDA}=\delta _{mm^{\prime }}n_{\sigma }$, and, consequently, $x_{\sigma }=(2l+1)n_{\sigma }^{2}$, where $n_{\sigma }=N_{\sigma }/(2l+1)$. This leads to the following corrections to the total energy and the effective potential:$$\begin{aligned}
\Delta E_{LDA+U}^{AMF} &=&-\frac{U-J}{2}\sum_{\sigma }\mathrm{Tr}%
(\delta\rho ^{\sigma }\cdot\delta \rho ^{\sigma })%-(2l+1)n_{\sigma }^{2}\right)
\nonumber \\
\Delta V_{LDA+U}^{AMF}(mm^{\prime }\sigma ) &=&-(U-J)\left( \rho
_{mm^{\prime }}^{\sigma }-n_{\sigma }\delta_{mm^\prime}\right) . \label{AMF}\end{aligned}$$Here AMF stands for Around Mean Field [@ChyzhikSawatsky] and $\delta\rho_{mm^\prime}^\sigma =\rho_{mm^\prime}^\sigma-n_\sigma\delta_{mm^\prime}$.
For strongly correlated systems the limit of the uniform occupancy is not correct (in fact, it is not correct even in weakly correlated systems, due to the crystal field). Thus, it is not surprising that for the systems with strongly localized electrons the AMF functional leads to rather unrealistic results. This observation led [@ChyzhikSawatsky; @Anisimov2] to another prescription, $x_{\sigma }=(2l+1)n_{\sigma }$, $$\begin{aligned}
\Delta E_{LDA+U}^{FLL} =-\frac{U-J}{2}\sum_{\sigma }\left( \mathrm{Tr}%
(\rho ^{\sigma }\cdot \rho ^{\sigma })-(2l+1)n_{\sigma }\right) \nonumber \\
\Delta V_{LDA+U}^{FLL}(mm^{\prime }\sigma ) =-(U-J)\left( \rho
_{mm^{\prime }}^{\sigma }-\frac{1}{2}\delta_{mm^\prime}\right) , \label{FLL}\end{aligned}$$which produces the correct behavior in the fully localized limit (FLL) where $n_{m\sigma }=0$ or 1. Most of the modern LDA+U calculations utilize one of these two functionals, although in real materials the occupation numbers lie between these two limits.
In the AMF the LDA+U correction to the electronic potential, Eq. \[AMF\], averaged over all occupied states, is zero. This is a possible way to define a mean field (cf. the Slater approximation to the Fock potential), but not the way used in the DFT. The latter is a mean field theory that produces the correct total energy, not the correct average potential. AMF and FLL represent the DFT mean field if all occupation numbers are all the same, or are all 0 or 1, respectively. It is easy to show that $%
(2l+1)n_{\sigma }^{2}\leq \mathrm{Tr}(\rho ^{\sigma }\cdot \rho ^{\sigma
})\leq (2l+1)n_{\sigma }$, so that AMF always gives a negative, and FLL a positive correction to the total energy, while the right (in the DFT sense) recipe should give zero correction to the total energy. That can be achieved by using a linear interpolation between the two extremes corresponding to AMF and FLL, $%
x_{\sigma }=(2l+1)\left( \alpha n_{\sigma }+(1-\alpha )n_{\sigma
}^{2}\right)$, where $0\leq \alpha\leq 1$, and $$\Delta E_{LDA+U}^{DFT}\!\!=\!\!-\frac{{U}\!\!-\!\!{J}}{2}\sum_{\sigma }\left[\mathrm{Tr}
(\delta\rho^\sigma\!\!\cdot\delta\rho^\sigma)
\!\!-\!\!(2l\!+\!1)\alpha n_\sigma(1-n_\sigma)\right]$$$$\Delta V_{LDA+U}^{DFT}(mm^{\prime }\sigma )\!=\!-({U}\!-\!{J})\!\!
\left[\rho_{mm^{\prime }}^{\sigma }\!-\!\left((1\!\!-\!\!\alpha)n_{\sigma }\!+\!\frac{\alpha}{2}
\right)\delta_{mm^\prime}\right]\!.
\label{MP}$$ In the spirit of the DFT, $\Delta E_{LDA+U}^{DFT}\!=\!0,$ so $$%0\leq
\alpha=\frac{\sum_\sigma\mathrm{Tr}(\delta \rho ^{\sigma }\cdot \delta
\rho ^{\sigma })}{(2l+1)\sum_\sigma n_{\sigma }(1-n_{\sigma })}
%\leq 1.
\label{alpha}$$
We emphasize that $\alpha $ is not adjustable, nor is it a formal functional of the charge density, but it is a material-dependent parameter (like, say, $U$ itself), defined by the *self-consistent* occupation matrix. However, in practical calculations it is better to recompute $\alpha $ after each iteration, as the current value of $\rho _{mm^{\prime }}^{\sigma } $ changes. Note that the total energy is given by the regular LDA expression that only implicitly depends on $U$ and $J$ *via* the changing density distribution; it is variational with respect to the charge density at a fixed $\alpha $, but *not* variational with respect to $\alpha
$ itself. The fact that this prescription is derived according to the DFT ideology allows one to formulate the proposed LDA+U functional (unlike the existing LDA+U functionals) as a *constrained DFT theory* [@Dederichs] at a given $%
\alpha$, with the constraint given by Eq. (\[alpha\]). $(U-J)/2$ appears then as a Lagrange multiplier. We have tested the proposed functional (\[MP\]) on NiO, a prototypical compound for LDA+U calculations (see, e.g., [Alluani]{}). Fig. \[NiO\] shows the band gap and the mangetic moment of NiO as a function of $U$ at $J=0.95$ eV for three different functionals (Eqs. \[AMF\], \[FLL\], and \[MP\]), calculated within the linear-muffin-tin orbital (LMTO) method in the atomic sphere approximation (ASA). The parameter $\alpha \simeq 0.5$ is almost independent on $U$. Accordingly, the results of our calculations based on Eq. \[MP\] for both band gaps and magnetic moments lie right between those for AMF and FLL calculations, and the effect of $U$ is reduced compared to the FLL calculations. This is in accord with a known observation [@Alluani] that in NiO the FLL LDA+U gives the best agreement with the experiment for $%
U\lesssim $6 eV, smaller than $U\simeq $ 8 eV calculated from the first principles[@Alluani; @Gunnarson].
Our next example is a weakly correlated metal FeAl. This paramagnetic material has attracted attention due to a recent suggestion by Mohn et al. [@Novak] that the short-range Coulomb correlations within the LDA+U may be responsible for suppression of ferromagnetism found in all LDA calculations. More specifically, they found in their AMF LDA+U calculations a reduction of the density of states (DOS) at the Fermi level, $D_{F},$ which was sufficient to make the Stoner criterion smaller than 1 and stabilize the paramagnetic state. To analyze this result, it is important to revisit the Stoner theory for the LDA+U case.
In DFT, the Stoner parameter $I$ is defined as $I=-2\partial
^{2}E_{xc}/\partial M^{2}$, the second derivative of the exchage-correlation energy with respect to the total magnetic moment. The paramagnetic ground state is unstable when $D_{F}I\geq 1.$ This can be derived from the force theorem, which states that the total energy for small magnetizations can be computed by assuming a rigid shift of the bands by $b=\pm M/2D_{F},$ so that the gain in the interaction energy, $-IM^{2}/4$, competes with the loss in the one-electron energy, $M^{2}/4D_{F}.$ In the LDA+U the criterion holds, but the product $D_{F}I$ changes, not only because $D_{F}$ changes, but also because the newly added interaction energy depends on $M$. Indeed, the force theorem calls for a change $\delta \rho _{mm^{\prime }}^{\sigma
}=b\sigma D_{mm^{\prime }}$, where $D_{mm^{\prime }}=-\pi ^{-1}\mathrm{Im}%
~G_{mm^{\prime }}(E_{F}).$ When applied to the functionals Eq. (\[AMF\]) - Eq. (\[MP\]), it generates a change in the interaction energy, which results in an additional contribution to the Stoner parameter, $$\Delta I(\alpha )=(U-J)\left( {\mathrm{Tr}}(D\cdot D)-\frac{(1-\alpha )({%
\mathrm{Tr}D})^{2}}{(2l+1)D_{F}^{2}}\right) \label{I}$$
In the limit of the uniform occupancy, Eq. (\[I\]) for the FLL case ($\alpha
=1$) reduces to $(U-J)/(2l+1)$. Given that the LDA Stoner parameter, $I,$ is of the same order as $J,$ we obtain for the total Stoner parameter $%
I_{FLL}\approx (U+2lJ)/(2l+1),$ which the well known expression for the Stoner factor in the atomic Hubbard model. On the contrary, $\Delta I_{AMF}$ ($\alpha =0$) in this limit is zero. In real metals $D_{mm^{\prime }}$ is complicated due to crystal field effects. Let us consider, for illustration, d-electrons in a cubic environment, and introduce the difference $\Delta
D=D_{eg}-D_{t2g}$, where $D_{eg}$ and $D_{t2g}$ are $e_{g}$ and $t_{2g}$ DOS per orbital at $E_{F}$, as a measure of the crystal field. This gives rise to a contribution to $\Delta I_{AMF}=\frac{5}{24}(U-J)(\Delta D/D_{F})^{2}$. However, when LDA+U reduces $D_{F}$, and $\Delta I_{AMF}$ is not large enough to overcome the decrease in $D_{F}$, LDA+U may stabilize the paramagnetic state(cf. Ref. [@dimitri]), as, for instance, observed in a very narrow range of large $U$’s for FeAl by Mohn *et al.* [Novak]{} (of course, only in the AMF functional; the FLL functional produces a large $\Delta I\approx (U-J)/5$, always increasing the tendency to magnetism).
With this in mind, we performed LMTO-ASA calculations for all three LDA+U functionals, using fixed $J=0.95$ eV. The results for $U$-dependence of the magnetic moment and $\alpha $ are shown in Fig. \[FeAl\] and compared with those by Mohn et al. [@Novak]. In our AMF calculations we also found a paramagnetic solution for $U=$ 4.85 eV, which however coexsits with a ferromagnetic high spin solution (Fig. \[FeAl\]). Note that for well localized orbitals there is no difference whether the $(U-J)$ term is applied inside the atomic sphere or only inside the MT sphere, as in Ref. [@Novak]; however, in less localized cases, where a noticeable part of the $d$-orbitals spills out of the MT sphere, the effect of the same $U$ is smaller when applied only inside the MT sphere. One can see in Fig. [FeAl]{} that, indeed, our calculations with large $U$ yield large $\alpha $’s and agree very well with Ref. [@Novak], while for small $U$ (small $%
\alpha $) the effect of $U$ in our ASA calculations is stronger than in Ref. [@Novak].
All LDA+U functionals shift unoccupied bands up and occupied bands down. Therefore LDA+U broadens the bands crossing the Fermi level. Because of this broadening, in FeAl for small $U$ the parameter $\alpha $ is initially decreasing (Fig. \[FeAl\]) with a minimum $\alpha =0.05$ at $U=2$ eV. The magnetic moment also decreases in this region. At larger $U,$ $\alpha $ starts growing again. At this point it is instructive to apply the logic of the constrained LDA approach in which for every fixed $\alpha $ the total LDA energy is minimized under the constraint $\sum_{\sigma }\mathrm{Tr}%
(\delta \rho ^{\sigma }\cdot \delta \rho ^{\sigma })/[(2l+1)\sum_{\sigma
}n_{\sigma }(1-n_{\sigma })]=\alpha $, $(U-J)/2$ being the Lagrange multiplier. For $\alpha \leq 0.087$ (Fig. \[FeAl\]) of the two possible solutions with $U<2$ eV and $U>2$ eV we should choose the one with lower energy (smaller $U).$ As a result, we find two admissible domains for $U:$ an AMF-like with $U<2$ eV and a FLL-like with $U\gtrsim 5$ eV. The latter is clearly unphysical. Both solutions are ferromagnetic. The solutions with intermediate values of $U$ and reduced magnetic moments are inadmissible in the framework of the constraint DFT formulation.
On the contrary, our explanation of the paramagnetism in FeAl is that the ferromagnetism instability is suppressed by the critical spin fluctuations. There are many other systems for which the fluctuations in the vicinity of a quantum critical point reduce the tendency to magnetism. Further examples include Sr$_{3}$Ru$_{2}$O$_{7}$ ($M_{LDA}\approx 0.8$ $\mu _{B},$ $M_{\exp
}=0),$ ZrZn$_{2}$ ($M_{LDA}\approx 0.7$ $\mu _{B},$ $M_{\exp }=0.2$ $\mu
_{B}),$ and other. The physics that is missing from both LDA and LDA+U equations in such systems can be described as exchange of virtual electronic excitations, roughly speaking, plasmons or (para)magnons. This leads to dressing of the one-particle excitations in the same way as the electron-phonon coupling dresses electrons near the Fermi surface, although in a correlated metal such mass renormalization effects occur on a large energy scale (of the order of $U$ or $J).$ LDA calculations cannot reproduce such a dressing, which has been observed in many different ways experimentally. For instance, LDA calculations do not explain large mass renormalizations in Sr$_{2}$RuO$_{4}$ [@mackenzie], and large specific heat renormalization in many correlated metals, produce too large plasma frequencies, $e.g.,$ in YBa$_{2}$Cu$_{3}$O$_{7},$ yield an optical absorption spectrum in CrO$_{2}$ shifted by about 20% to higher frequency, as compared with experiment [@Claudia], and overestimate the exchange splitting in Ni by a factor of 2 [@NiPES]. In all these cases the total width of the $d$-bands is *decreased*, as opposed to *broadening* inherent to LDA+U. Here the essential physics is missing from the LDA+U as well as in LDA, while the spatial variation of the mean-field Coulomb interaction is treated better by the LDA. The missing physics is associated, to a large degree, with dynamic fluctuations.
The dynamic version of the LDA+U method, the Dynamic Mean Field Theory (DMFT) [@Georges], which can account for some spin fluctuations [@spflex], resolves many of these problems. For instance, the mass renormalization in Sr$_{2}$RuO$_{4}$ is 3-4 [mackenzie]{}, far greater than possible renormalization due to the phonons. We applied all three flavor of LDA+U to Sr$_{2}$RuO$_{4}$ and found no mass renormalization compared to LDA. On the other hand, Eliashberg-type calculations [@mazinsingh] of the renormalization due to spin fluctuations, using a spectrum deduced from the LDA band structure, give mass renormalizations of the right order. Similarly, DMFT explicitely narrows the bands in Sr$_{2}$RuO$_{4}$ and enhances the electronic mass [LL]{}. With this in mind, we applied the DMFT with a realistic $U=2$ eV to FeAl and found the paramagnetic state to be perfectly stable, whith the bands *narrower* than in LDA, and the density of states practically the same (Fig. \[lulu\]). In other words, the spin fluctuations effectively reduce the Stoner factor $I$.
To conclude, we observe that no LDA+U functional correctly describes the essential physics of the weakly correlated metals: (i) reducing the band dispersion by dressing of the one-particle excitation, and (ii) spin fluctuations near the quantum critical point. One functional, labeled FLL here, correctly describes the important physics in the limit of well localized electrons, and can be recommended in this case. The other functional, labeled AMF, is exact in a hypothetical material with the uniform orbital occupancies. Although neither functional accounts for the fluctuation effects, LDA+U may be useful, if applied with a grain of salt, in moderately correlated metals. For this case, we propose a recipe that accounts for an incomplete localization and reduces to AMF or FLL in the appropriate limits. Finally, it is worth noting that in many correlated materials the spin-orbit interaction plays a key role. Since our $\alpha $ does not depend on spin, this prescription can be also formulated in terms of the full $(4l+2)\times(4l+2)$ occupation matrix $%
\rho $ and $n=\mathrm{Tr}(\rho)/(4l+2)$. Eq. (\[alpha\]) should be replaced with $\mathrm{Tr}(\delta \rho \cdot
\delta \rho )=$ $(4l+2)\alpha n(1-n)$. This formulation has another advantage in the case of a half-filled band, like in Gd, because in this limit it reduces to more physically meaningful in this case FLL, rather than to AMF as the nonrelativistic Eq. (\[MP\]).
This work is supported in part by ONR and by NSF (Grant No DMR-0071823).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Proton decay is usually discussed in the context of grand unified theories. However, as is well-known, in the standard model effective theory proton decay appears in the form of higher dimensional non-renormalizable operators. Here, we study systematically the 1-loop decomposition of the $d=6$ $B+L$ violating operators. We exhaustively list the possible 1-loop ultra-violet completions of these operators and discuss that, in general, two distinct classes of models appear. Models in the first class need an additional symmetry in order to avoid tree-level proton decay. These models necessarily contain a neutral particle, which could act as a dark matter candidate. For models in the second class the loop contribution dominates automatically over the tree-level proton decay, without the need for additional symmetries. We also discuss possible phenomenology of two example models, one from each class, and their possible connections to neutrino masses, LHC searches and dark matter.'
author:
- Juan Carlos Helo
- Martin Hirsch
- Toshihiko Ota
title: 'Proton decay at 1-loop'
---
Introduction {#sect:intro}
============
While searches for proton decay so far have yielded only lower bounds on the lifetime of various possible decay modes [@Sussman:2018ylo; @TheSuper-Kamiokande:2017tit; @Miura:2016krn; @Takhistov:2015fao; @TheKamLAND-Zen:2015eva; @Gustafson:2015qyo; @Takhistov:2014pfw; @Abe:2014mwa; @Abe:2013lua; @Tanabashi:2018oca], future large volume detectors, such as Hyper-Kamiokande [@Abe:2018uyc], DUNE [@Abi:2018dnh] and JUNO [@An:2015jdp], or more speculative multi-megaton proposals such as TITAND [@Suzuki:2001rb; @Kistler:2008us], MEMPHYS [@deBellefon:2006vq] or MICA [@Boser:2013oaa; @Cowen:2013zz] offer a good chance to finally discover this ultra-rare process. Although nucleon decay processes are usually discussed in the context of grand unified theories (GUTs), see e.g., Refs. [@Georgi:1974sy; @Langacker:1980js; @Sakai:1981pk; @Weinberg:1981wj; @Dimopoulos:1981dw; @Ellis:1981tv; @Nath:1985ub; @Hisano:1992jj; @Lucas:1996bc; @Goto:1998qg; @Murayama:2001ur; @Hisano:2013exa; @Ellis:2015rya], they can arise in many models. For a review on baryon number violation, see for example Ref. [@Babu:2013jba]. Motivated by the expected improvements in nucleon decay searches, here we study proton decay generated at the 1-loop level.
In the standard model baryon and lepton number violation arises at the non-renormalizable level. At the level of mass dimension five ($d=5$), there is only one operator, the famous Weinberg operator [@Weinberg:1979sa], corresponding to Majorana neutrino masses ($\Delta L=2$, $\Delta B=0$). At $d=6$ there are already five independent operators, which have $\Delta
B=\Delta L=1$ (but $\Delta (B-L)=0$) [@Weinberg:1979sa; @Wilczek:1979hc; @Abbott:1980zj]. All $d=6$ operators lead to two-body proton decays, such as $p \to \pi^0 + e^+$, $p \to \pi^+ + \bar{\nu}$ or $p \to K^+ + \bar{\nu}$.
GUT models predict proton decay to occur at tree-level [@Georgi:1974sy; @Langacker:1980js]. For coefficients of order ${\cal O}(1)$, the current experimental bounds then imply a lower limit on the scale of baryon number violation (for $d=6$ operators) of order $\Lambda \sim {\rm (few)}\hskip1mm{\cal O}(10^{15})$ GeV, which is far out of reach of any foreseeable accelerator experiment. This simple picture changes drastically, if proton decay is induced by higher dimensional operators and/or at loop level. The decay rate for a $k$-body $n$-loop proton/neutron decay induced by a $d$-dimensional operator can be very roughly estimated to be: $$\label{eq:pdec}
\frac{1}{\tau} \sim \frac{{\cal C}^{2}}{f[k]} \Big(\frac{1}{16 \pi^2}\Big)^{2 n}
\Big(\frac{m_p}{\Lambda}\Big)^{2(d-6)}\frac{m_p^5}{\Lambda^4}$$ Here, $f[k] \equiv4\left(4\pi\right)^{2k-3}\left(k-1\right)!
\left(k-2\right)!$ estimates the phase space volume available to the decay products for massless final state particles [@Fonseca:2018ehk]. The constant ${\cal C}$ is the coefficient of the effective interaction that induces the proton decay process, which contains products of couplings that appear in the ultra-violet models given at the scale $\Lambda$. Note that ${\cal C}$ can be small compared to one, depending on the model, see below. Obviously, to obtain decay rates within future experimental sensitivities much lower scales $\Lambda$ are needed for $k \gg 2$, $n \gg 0$ and/or $d \gg 6$.
Probably for this reason, not many studies on higher-dimensional proton decay operators can be found in the literature. For $d=7$ operators see, for example, Refs. [@Babu:2012iv; @Lehman:2014jma; @Bhattacharya:2015vja]. For operators with $d=9$ and higher see Refs. [@ODonnell:1993kdg; @Hambye:2017qix; @Fonseca:2018ehk]. In particular, Ref. [@Fonseca:2018ehk] discusses $\Delta L=3$ proton decay from operators up to $d=13$, where current experimental sensitivities correspond to new physics scales $\Lambda \lsim $TeV, even for couplings as large as order ${\cal O}(1)$.[^1] The authors of Ref. [@deGouvea:2014lva] listed the higher-mass-dimensional $B-L$-violating effective operators in a GUT model and discussed the relations between neutrino masses and the nucleon decays induced by the effective operators.
Even less work has been done so far for loop-induced proton decay. Perhaps the best-known example for it is supersymmetric (SUSY) GUTs, see for example the review [@Babu:2013jba]. Here, the importance of the loop stems from the fact that the decay amplitude is proportional to $(\Lambda_{\rm GUT}\Lambda_{\rm SUSY})^{-1}$ instead of $\Lambda_{\rm GUT}^{-2}$ as (for tree-level contributions) in non-SUSY GUT models.
In this paper, we exhaustively list the possible high-energy completions of the proton decay operators with $d=6$ at the 1-loop level. We also calculate group theoretical factors and define the 1-loop integrals, which appear in the reproduction of the proton decay operators from their decompositions. From these lists one can immediately estimate the rate of proton decay, once a (proto-)model is specified. For masses of the mediators at the TeV scale, we find that the couplings $Y$ entering the proton decay rate should be of order $Y < \mathcal{O}(10^{-6})$.[^2] This opens up the possibility that the charged/coloured mediator fields live long on the time scale of collider experiments, yielding particular signals at the LHC.
We divide the different models, found in our lists, into two sub-classes. Models in the first class require an additional symmetry to avoid tree-level proton decay. It is straightforward to introduce some extra symmetry in these cases, for example a $Z_2$, that guarantees that proton decay appears only at the 1-loop (and higher) level. In this class of models the lightest loop particle is then necessarily stable and thus can serve as a candidate for the dark matter. In the second class one finds models, in which the loop-induced $d=6$ decay is automatically the leading contribution to proton decay, despite the existence of tree-level decay modes. The reason for this counter-intuitive behaviour is simply that for models in the second class, tree-level proton decay appear only at the level of higher-dimensional effective operators.
We then discuss two example models, one from each model class, in more details. In Model-I, neutrino masses, dark matter and proton decay are all related. Majorana neutrino masses are generated using the scotogenic loop [@Ma:2006km] and the same $Z_2$ that stabilizes the dark matter guarantees that proton decay occurs only at the 1-loop level. The coloured mediators of proton decay, if at the TeV scale, can be produced at the LHC and will decay to jets, leptons and the dark matter candidate. These missing energy signals, possibly associated with charged tracks from heavy ionizing particles, are reminiscent of those discussed in the context of SUSY. Thus, one can use different existing searches at the LHC to derive constraints on the model. Also, since the model generates neutrino masses at 1-loop, one can constrain its parameters using searches for lepton flavour violation, such as $\mu\to e\gamma$ and others.
In our Model-II we do not impose any beyond the SM symmetry. Thus, there are no stable, heavy particles. Signals for searches at the LHC are therefore different from those discussed for Model-I. In particular, there are final states with no missing energy involved. For this model, we also show how tree-level proton decay will appear and is suppressed in models in this class. For the particular case of Model-II, the final state for proton decay is caused by a tree $d=12$ operator and is 5-body. The expected partial half-lives for these modes are therefore orders of magnitude larger than those of the 1-loop induced 2-body decays.
The rest of this paper is organized as follows. In Section II we will discuss the $d=6$ operators and their 1-loop decomposition. Section III then presents and discusses our two example models, before we conclude in Section IV. Some more technical aspects for the 1-loop decomposition are given in the appendix.
Proton decay operators at 1-loop
================================
The effective operators which lead to proton decay were already listed in Refs. [@Weinberg:1979sa; @Wilczek:1979hc; @Abbott:1980zj][^3]: $$\begin{aligned}
\mathcal{O}_{1} =& [du][QL],
\label{eq:O1}
\\
\mathcal{O}_{2} =& [QQ][ue],
\label{eq:O2}
\\
\mathcal{O}_{3} =& [QQ]_{\bf 1}[QL]_{\bf 1},
\label{eq:O3}
\\
\mathcal{O}_{4} =& [QQ]_{\bf 3}[QL]_{\bf 3},
\label{eq:O4}
\\
\mathcal{O}_{5} =& [du][ue],
\label{eq:O5}\end{aligned}$$ where the subscripts ${\bf 1}$ and ${\bf 3}$ in Eqs. and indicate the electroweak $SU(2)$ representation of the bilinears of the fermions. The contraction of all the indices on the operators is explicitly shown in Appendix.
(5.8,3) (0,0)[![Box diagram for decomposition of the $d=6$ proton decay operators. The direction of the arrows represents the flow of the particle number (not chirality). We put the lepton field $\ell$ always in the lower right corner $(Y_{3})$ in all decompositions, i.e., The $SU(3)$ structure is common in all decompositions.[]{data-label="Fig:Box"}](Fig1-crop.eps "fig:"){width="5.5cm"}]{}
We are particularly interested in the effective operators which arise from 1-loop diagrams. We decompose the effective operators Eqs. - and list the necessary mediators and the interactions. There are two types of topologies for $d=6$ 1-loop diagrams: triangle and box. However, the decompositions based on the triangle diagram allow to have the same effective operator at the tree level.[^4] Therefore, we concentrate on the decompositions with the box diagram. In the decomposition, we distribute the quarks and the lepton of an effective operator to the outer legs of the box diagram shown in Fig. \[Fig:Box\] in all possible ways, and identify the SM gauge charges of the mediator fields, $\psi$, $S$, $\psi'$, and $S'$. In the current study, we restrict ourselves to the decompositions with scalar and fermion mediators and do not introduce a vector mediator which may require an extension of the SM gauge symmetries and their spontaneous breaking to the SM. In short, we introduce the following Yukawa interactions, $$\begin{aligned}
\mathscr{L}
=&
Y_{1}
\overline{q^{c}}
\psi^{c}
S
+
Y_{2}
\overline{\psi^{c}} q'
S'^{\dagger}
+
Y_{3}
\overline{\psi'}
\ell
S'
+
Y_{4}
\overline{q''^{c}}
\psi'
S^{\dagger}
\nonumber
\\
&+
{\rm H.c.},
\label{eq:L-decom-general}\end{aligned}$$ where $q$, $q'$, and $q''$ are the quark fields ($Q$, $u$, or $d$), and $\ell$ is the lepton field ($L$ or $e$) in an effective operator. The mediator fields, $\psi$ and $\psi'$ for fermions and $S$ and $S'$ for scalars, are assigned as shown in Fig. \[Fig:Box\]. The mass terms for the mediator fields must also be included in the Lagrangian, although they are not explicitly written in Eq. . Later we will discuss the phenomenology of mediator fields, assuming that the masses $M_{\psi}$, $M_{S}$, $M_{\psi'}$, and $M_{S'}$ are at the TeV scale.
The colour $SU(3)$ structure of the box diagram Fig. \[Fig:Box\] is common in all decompositions, and the possible ways to assign the $SU(3)$ charges to the mediators are listed in Tab. \[Tab:SU3-decom\]. Here we assume that a mediator takes one of ${\bf 1}$, ${\bf 3}$, $\overline{\bf 3}$, ${\bf 6}$, $\overline{\bf 6}$, and ${\bf 8}$ representations under the $SU(3)$ transformation and do not pursue the possibility of decompositions with a mediator whose representation is higher than ${\bf 8}$.
------ -------------------- -------------------- -------------------- --------------------- ----------------
$\psi$ $S$ $\psi'$ $S'$ $SU(3)$ coeff.
\#1 ${\bf 1}$ $\overline{\bf 3}$ ${\bf 3}$ ${\bf 3}$ $-1$
\#2 ${\bf 3}$ ${\bf 1}$ $\overline{\bf 3}$ $\overline{\bf 3}$ $1$
\#3 ${\bf 3}$ ${\bf 8}$ $\overline{\bf 3}$ $\overline{\bf 3}$ $-\frac{8}{3}$
\#4 ${\bf 3}$ ${\bf 8}$ ${\bf 6}$ ${\bf 6}$ $4$
\#5 $\overline{\bf 3}$ ${\bf 3}$ ${\bf 1}$ ${\bf 1}$ $1$
\#6 $\overline{\bf 3}$ ${\bf 3}$ ${\bf 8}$ ${\bf 8}$ $-\frac{8}{3}$
\#7 $\overline{\bf 3}$ $\overline{\bf 6}$ ${\bf 8}$ ${\bf 8}$ $-4$
\#8 ${\bf 6}$ ${\bf 3}$ [**8**]{} ${\bf 8}$ $4$
\#9 $\overline{\bf 6}$ ${\bf 8}$ $\overline{\bf 3}$ $\overline{\bf $-4$
3}$
\#10 ${\bf 8}$ $\overline{\bf 3}$ ${\bf 3}$ $ {\bf 3}$ $\frac{8}{3}$
\#11 ${\bf 8}$ $\overline{\bf 3}$ $\overline{\bf 6}$ $ \overline{\bf 6}$ $-4$
\#12 ${\bf 8}$ ${\bf 6}$ ${\bf 3}$ ${\bf 3}$ $4$
------ -------------------- -------------------- -------------------- --------------------- ----------------
: Choices of the $SU(3)$ charges of the mediator fields and the $SU(3)$ coefficients which appear in the reordering of the $SU(3)$ indices to obtain the corresponding effective operator; see Eqs. and and the text. []{data-label="Tab:SU3-decom"}
In the column “$SU(3)$ coeff.”, we also list the coefficients appearing in the calculation, which we call [*operator projection*]{}, to derive the effective operators from the decompositions. In order to obtain the effective operators Eqs. - from the decomposition Eq. where each Yukawa interaction forms a $SU(3)$ singlet, we must rearrange the $SU(3)$ indices as $$\begin{aligned}
\mathscr{L}_{\text{eff}}
=&
\wick{1231}{
[Y_{1}
\overline{q^{c}}_{I}
<1\psi^{c} <3S]
[Y_{2} >1{\overline{\psi^{c}}} q'_{J} <4S'^{\dagger}]
[Y_{4} \overline{q''^{c}}_{K} <2\psi' >3S^{\dagger}]
[Y_{3} >2{\overline{\psi'}} \ell >4S']
}
\nonumber
\\
=&
Y_{1}Y_{2}Y_{3}Y_{4}
\times
\text{$SU(3)$ coeff.}
\times
\epsilon^{IJK}
\overline{q^{c}}_{I}
q'_{J}
\overline{q''^{c}}_{K}
\ell...
\label{eq:SU3-projection}
\end{aligned}$$ where $I$, $J$, and $K$ are the $SU(3)$ indices for ${\bf 3}$ representations, and $\epsilon^{IJK}$ is the total anti-symmetric tensor to form a singlet with three triplets. The part omitted from the second line of Eq. , which is expressed as “...”, represents all contents other than the Yukawa couplings ($Y_{1\text{-}4}$), the coefficient ($SU(3)$ coeff.) brought by the rearrangement of the $SU(3)$ indices, and the outer fermion field operators ($q$, $q'$, $q''$, and $\ell$), such as the propagators of the mediators and matrices with $SU(2)$ indices. We have not specified the quark fields at this stage and rearrange the $SU(3)$ indices by handling them as [**3**]{} representation field operators in general. Depending on the decomposition with a specific choice of the quark fields, an additional sign can show up in the further rearrangement of the $SU(3)$ indices, which will be taken into account after the full information of the decomposition, with which one can fully specify the ordering of the quark fields. The sign due to the ordering of the quarks will be given in Tabs. \[Tab:Summary-O1\]-\[Tab:Summary-O5\] (as “$SU(3)$ sign”). Note that the $SU(2)$ and the Lorentz indices have not been rearranged at this stage, and the rearrangement of them will bring other coefficients and factors. All the details of the method of decomposition and operator projection are given in Appendix, where we demonstrate the derivation of all the coefficients, signs, and factors, keeping all the indices on the field operators explicitly.
To proceed the operator projection onto the basis operators Eqs. -, we must specify the species of the outer fermion fields, determine the position of the quark fields on the box diagram, and identify the $SU(2)$ gauge charges of the mediator fields. In Tabs. \[Tab:Summary-O1\]-\[Tab:Summary-O5\], the ways of decomposition are given in the column “Decom”, where the given fermion fields correspond to $(qq')(q''\ell)$ in Fig. \[Fig:Box\] and Eq. . The electroweak charges of the mediator fields are listed at the column “Mediators $SU(2)_{U(1)}$”. We concentrate on ${\bf 1}$, ${\bf 2}$, and ${\bf 3}$ for the $SU(2)$ representation. Note that the sign that comes up in the rearrangement of the $SU(3)$ indices are also given in the column of “$SU(3)$ sign” in Tabs. \[Tab:Summary-O1\]-\[Tab:Summary-O5\], which cannot be included in Tab. \[Tab:SU3-decom\] because they depend on the ordering of the quark fields in a decomposition. We also list the factors and coefficients which come up in the process of the operator projection after Eq. : the coefficients and signs from the rearrangement of the $SU(2)$ indices (“$SU(2)$ coeff.”), the factors and signs from the Fierz transformations (rearrangement of the Lorentz indices), and the loop integral factors (“Fierz$\times$Loop factors”). The functions $I_{4}$ and $J_{4}$ for the loop integrals of the box diagrams are defined in Appendix. In short, once the decomposition (proto-model) is specified (one from Tab. \[Tab:SU3-decom\] and one from Tabs. \[Tab:Summary-O1\]-\[Tab:Summary-O5\] are chosen), the coefficient of the effective operator is given as $$\begin{aligned}
\mathscr{L}_{\text{eff}}
=&
\text{$SU(3)$ coeff.}
\times
\text{$SU(3)$ sign}
\times
\text{$SU(2)$ coeff.}
\nonumber
\\
&
\times
\text{Fierz factor}
\times
\text{Loop factor}
\times
Y_{1} Y_{2} Y_{3} Y_{4}
\nonumber
\\
&
\times
\text{effective op(s) $\mathcal{O}$
in Eqs.~\eqref{eq:O1}-\eqref{eq:O5}},
\label{eq:Leff}\end{aligned}$$ with which, and also with the help of the nucleon matrix elements calculated from lattice [@Yoo:2018fyn; @Aoki:2017puj; @Aoki:2013yxa; @Aoki:2006ib] and chiral perturbation theory [@Claudson:1981gh; @Chadha:1983sj; @Aoki:2008ku], one can directly calculate the rates of proton decay. The notations and the derivations of the coefficients, factors, and signs are given in Appendix.
------------------- -------------------- -------------------------------- -------------------------------- -------------------------------- -------- ---------------------------- ------
$\mathcal{O}_{1}$ $SU(2)$ Fierz$\times$Loop $SU(3)$
Decom. $\psi$ $S$ $\psi'$ $S'$ coeff. factors sign
$(du)(QL)$ ${\bf 1}_{\alpha}$ ${\bf 1}_{\alpha+\frac{1}{3}}$ ${\bf 2}_{\alpha+\frac{1}{6}}$ ${\bf 1}_{\alpha+\frac{2}{3}}$ $1$ $M_{\psi} M_{\psi'} I_{4}$ +
${\bf 2}$ [**2**]{} [**1**]{} [**2**]{} $-1$
${\bf 2}$ [**2**]{} [**3**]{} [**2**]{} $-3$
${\bf 3}$ [**3**]{} [**2**]{} [**3**]{} $3$
$(ud)(QL)$ ${\bf 1}_{\alpha}$ ${\bf 1}_{\alpha-\frac{2}{3}}$ ${\bf 2}_{\alpha-\frac{5}{6}}$ ${\bf 1}_{\alpha-\frac{1}{3}}$ $1$ $M_{\psi} M_{\psi'} I_{4}$ $-$
${\bf 2}$ [**2**]{} [**1**]{} [**2**]{} $-1$
${\bf 2}$ [**2**]{} [**3**]{} [**2**]{} $-3$
${\bf 3}$ [**3**]{} [**2**]{} [**3**]{} $3$
$(dQ)(uL)$ ${\bf 1}_{\alpha}$ ${\bf 1}_{\alpha+\frac{1}{3}}$ ${\bf 1}_{\alpha-\frac{1}{3}}$ ${\bf 2}_{\alpha+\frac{1}{6}}$ $-1$ $-\frac{1}{2} J_{4}$ $-$
[**2**]{} [**2**]{} [**2**]{} [**1**]{} $-1$
[**2**]{} [**2**]{} [**2**]{} [**3**]{} $3$
[**3**]{} [**3**]{} [**3**]{} [**2**]{} $-3$
$(Qd)(uL)$ ${\bf 1}_{\alpha}$ ${\bf 2}_{\alpha-\frac{1}{6}}$ ${\bf 2}_{\alpha-\frac{5}{6}}$ ${\bf 1}_{\alpha-\frac{1}{3}}$ $1$ $\frac{1}{2} J_{4}$ $+$
[**2**]{} [**1**]{} [**1**]{} [**2**]{} $-1$
[**2**]{} [**3**]{} [**3**]{} [**2**]{} $-3$
[**3**]{} [**2**]{} [**2**]{} [**3**]{} $3$
$(uQ)(dL)$ ${\bf 1}_{\alpha}$ ${\bf 1}_{\alpha-\frac{2}{3}}$ ${\bf 1}_{\alpha-\frac{1}{3}}$ ${\bf 2}_{\alpha+\frac{1}{6}}$ $-1$ $-\frac{1}{2} J_{4}$ $+$
[**2**]{} [**2**]{} [**2**]{} [**1**]{} $-1$
[**2**]{} [**2**]{} [**2**]{} [**3**]{} $3$
[**3**]{} [**3**]{} [**3**]{} [**2**]{} $-3$
$(Qu)(dL)$ ${\bf 1}_{\alpha}$ ${\bf 2}_{\alpha-\frac{1}{6}}$ ${\bf 2}_{\alpha+\frac{1}{6}}$ ${\bf 1}_{\alpha+\frac{2}{3}}$ $1$ $\frac{1}{2}J_{4}$ $-$
[**2**]{} [**1**]{} [**1**]{} [**2**]{} $-1$
[**2**]{} [**3**]{} [**3**]{} [**2**]{} $-3$
[**3**]{} [**2**]{} [**2**]{} [**3**]{} $3$
------------------- -------------------- -------------------------------- -------------------------------- -------------------------------- -------- ---------------------------- ------
: Decomposition (proto-models) of the $d=6$ effective operator $\mathcal{O}_{1}=[du][QL]$ which consists of $d$, $u$, $Q$ and $L$. The proto-models result in the same effective operator $\mathcal{O}_{1}$ but with different coefficients, factors, and signs.[]{data-label="Tab:Summary-O1"}
-------------------- -------------------- -------------------------------- -------------------------------- -------------------------------- -------- ---------------------------- ------
$\mathcal{O}_{2}$ $SU(2)$ Fierz$\times$Loop $SU(3)$
Decom. $\psi$ $S$ $\psi'$ $S'$ coeff. factors sign
$(Q_{1}Q_{2})(ue)$ ${\bf 1}_{\alpha}$ ${\bf 2}_{\alpha-\frac{1}{6}}$ ${\bf 2}_{\alpha-\frac{5}{6}}$ ${\bf 2}_{\alpha+\frac{1}{6}}$ $1$ $M_{\psi} M_{\psi'} I_{4}$ $+$
[**2**]{} [**1**]{} [**1**]{} [**1**]{} $1$
[**2**]{} [**3**]{} [**3**]{} [**3**]{} $-3$
[**3**]{} [**2**]{} [**2**]{} [**2**]{} $3$
$(uQ_{2})(Q_{1}e)$ ${\bf 1}_{\alpha}$ ${\bf 1}_{\alpha-\frac{2}{3}}$ ${\bf 2}_{\alpha-\frac{5}{6}}$ ${\bf 2}_{\alpha+\frac{1}{6}}$ $1$ $\frac{1}{2}J_{4}$ $-$
[**2**]{} [**2**]{} [**1**]{} [**1**]{} $1$
[**2**]{} [**2**]{} [**3**]{} [**3**]{} $-3$
[**3**]{} [**3**]{} [**2**]{} [**2**]{} $3$
$(Q_{1}u)(Q_{2}e)$ ${\bf 1}_{\alpha}$ ${\bf 2}_{\alpha-\frac{1}{6}}$ ${\bf 1}_{\alpha-\frac{1}{3}}$ ${\bf 1}_{\alpha+\frac{2}{3}}$ $1$ $-\frac{1}{2}J_{4}$ $-$
[**2**]{} [**1**]{} [**2**]{} [**2**]{} $-1$
[**2**]{} [**3**]{} [**2**]{} [**2**]{} $-3$
[**3**]{} [**2**]{} [**3**]{} [**3**]{} $3$
-------------------- -------------------- -------------------------------- -------------------------------- -------------------------------- -------- ---------------------------- ------
: Decomposition of $\mathcal{O}_{2}$, where the ordering of the two quark doublets in the effective operator is determined as $\mathcal{O}_{2}\equiv[Q_{1}Q_{2}][ue]$.[]{data-label="Tab:Summary-O2"}
------------------------ -------------------- -------------------------------- -------------------------------- -------------------------------- -------------------------------------------------------------- ---------------------------- ------
$\mathcal{O}_{3,4}$ $SU(2)$ Fierz$\times$Loop $SU(3)$
Decom. $\psi$ $S$ $\psi'$ $S'$ coeff. factors sign
$(Q_{1}Q_{2})(Q_{3}L)$ ${\bf 1}_{\alpha}$ ${\bf 2}_{\alpha-\frac{1}{6}}$ ${\bf 1}_{\alpha-\frac{1}{3}}$ ${\bf 2}_{\alpha+\frac{1}{6}}$ $-\frac{1}{2} \mathcal{O}_{3} + \frac{1}{2} \mathcal{O}_{4}$ $M_{\psi} M_{\psi'} I_{4}$ $+$
[**1**]{} [**2**]{} [**3**]{} [**2**]{} $ -\frac{3}{2} \mathcal{O}_{3} - \frac{1}{2}
\mathcal{O}_{4}$
[**2**]{} [**1**]{} [**2**]{} [**1**]{} $ \mathcal{O}_{3} $
[**2**]{} [**1**]{} [**2**]{} [**3**]{} $ \mathcal{O}_{4} $
[**2**]{} [**3**]{} [**2**]{} [**1**]{} $ -\mathcal{O}_{4} $
[**2**]{} [**3**]{} [**2**]{} [**3**]{} $ -3 \mathcal{O}_{3} + 2 \mathcal{O}_{4} $
[**3**]{} [**2**]{} [**1**]{} [**2**]{} $ - \frac{3}{2} \mathcal{O}_{3} - \frac{1}{2}
\mathcal{O}_{4}$
[**3**]{} [**2**]{} [**3**]{} [**2**]{} $ -\frac{9}{2} \mathcal{O}_{3}
+ \frac{1}{2} \mathcal{O}_{4}$
------------------------ -------------------- -------------------------------- -------------------------------- -------------------------------- -------------------------------------------------------------- ---------------------------- ------
: Decompositions of the effective operators with three $Q$s and a $L$. Each model results in a different combination of $\mathcal{O}_{3}$ and $\mathcal{O}_{4}$. Note that the ordering of the three $Q$s in the effective operators are fixed as $\mathcal{O}_{3/4}$ $\equiv [Q_{1}Q_{2}]_{{\bf 1}/{\bf
3}}[Q_{3}L]_{{\bf 1}/{\bf 3}}$.[]{data-label="Tab:Summary-O3O4"}
---------------------- -------------------- -------------------------------- -------------------------------- -------------------------------- -------- ---------------------------------------------------------------------------
$\mathcal{O}_{5}$ $SU(2)$ Fierz$\times$Loop factors
Decom. $\psi$ $S$ $\psi'$ $S'$ coeff. $\times SU(3)$ sign
$(du_{1})(u_{2}e)$ ${\bf 1}_{\alpha}$ ${\bf 1}_{\alpha+\frac{1}{3}}$ ${\bf 1}_{\alpha-\frac{1}{3}}$ ${\bf 1}_{\alpha+\frac{2}{3}}$ $1$ $M_{\psi}M_{\psi'}I_{4}\mathcal{O}_{5}$
[**2**]{} [**2**]{} [**2**]{} [**2**]{} $2$
[**3**]{} [**3**]{} [**3**]{} [**3**]{} $3$
$(u_{1} d)(u_{2} e)$ ${\bf 1}_{\alpha}$ ${\bf 1}_{\alpha-\frac{2}{3}}$ ${\bf 1}_{\alpha-\frac{4}{3}}$ ${\bf 1}_{\alpha-\frac{1}{3}}$ $1$ $-M_{\psi}M_{\psi'}I_{4}\mathcal{O}_{5}$
[**2**]{} [**2**]{} [**2**]{} [**2**]{} $2$
[**3**]{} [**3**]{} [**3**]{} [**3**]{} $3$
$(u_{2}u_{1})(de)$ ${\bf 1}_{\alpha}$ ${\bf 1}_{\alpha-\frac{2}{3}}$ ${\bf 1}_{\alpha-\frac{1}{3}}$ ${\bf 1}_{\alpha+\frac{2}{3}}$ $1$ $M_{\psi}M_{\psi'}I_{4} \left[ \mathcal{O}_{5} - \mathcal{O}_{5}'\right]$
[**2**]{} [**2**]{} [**2**]{} [**2**]{} $2$
[**3**]{} [**3**]{} [**3**]{} [**3**]{} $3$
---------------------- -------------------- -------------------------------- -------------------------------- -------------------------------- -------- ---------------------------------------------------------------------------
: Decompositions of $\mathcal{O}_{5}$. Here the ordering of the two $u$’s in the basis operators is determined as $\mathcal{O}_{5}\equiv[du_{1}][u_{2}e]$ and $\mathcal{O}_{5}'\equiv[du_{2}][u_{1}e]$.[]{data-label="Tab:Summary-O5"}
Here we show an example to demonstrate how to use the information of the tables. A famous dimension-five contribution to proton decay in SUSY-GUT models is found by taking \#1 from Tab. \[Tab:SU3-decom\] and the decomposition of the seventh row in Tab. \[Tab:Summary-O3O4\] with $\alpha=0$ for $U(1)$ hypercharge. The mediators are identified with the SUSY particles as $$\begin{gathered}
\psi({\bf 1},{\bf 3})_{0} = \widetilde{W},
\quad
S(\overline{\bf 3}, {\bf 2})_{-1/6} = \widetilde{Q}^{*}
\nonumber
\\
\psi'({\bf 3},{\bf 1})_{-1/3} = \widetilde{h}_{c},
\quad
S'({\bf 3},{\bf 2})_{+1/6} = \widetilde{Q},
\label{eq:model-for-d5SUSYGUT}\end{gathered}$$ where $\widetilde{W}$ is the wino, $\widetilde{h}_{c}$ is the coloured higgsino, and $\widetilde{Q}$ is the squark doublet. Using the information listed in the tables, we can reproduce the coefficient of the effective operator Eq. as $$\begin{aligned}
\mathscr{L}_{\text{eff}}
=&
\overbrace{
(-1)
}^{\text{SU(3) coeff.}}
\times
\overbrace{
(+)
}^{\text{$SU(3)$ sign}}
\times
\overbrace{
\left[
-\frac{3}{2}
\mathcal{O}_{3}
-
\frac{1}{2}
\mathcal{O}_{4}
\right]
}^{\text{$SU(2)$ coeff. and $\mathcal{O}$}}
\nonumber
\\
&
\times
\overbrace{
M_{\psi} M_{\psi'} I_{4}
}^{\text{Fierz$\times$Loop factors}}
\times
Y_{1} Y_{2} Y_{3} Y_{4}
\nonumber
\\
=&
\frac{1}{2}
Y_{1}Y_{2}Y_{3}Y_{4} M_{\widetilde{W}}
M_{\widetilde{h}_{c}}
I_{4}
\left[
3
\mathcal{O}_{3}
+
\mathcal{O}_{4}
\right].\end{aligned}$$ Note that $Y_{1}$ and $Y_{2}$ are given by the gauge coupling in $SU(5)$ SUSY-GUT models. The coupling $Y_{3}$ is identified with the coupling for the Yukawa interaction of ${\bf 10} \cdot \overline{\bf 5} \cdot H(\overline{\bf 5})$, and $Y_{4}$ is that for ${\bf 10} \cdot {\bf 10} \cdot H({\bf 5})$, where ${\bf 10}$ and $\overline{\bf 5}$ are the matter superfields and $H(\overline{\bf 5})$ and $H({\bf 5})$ are the Higgs superfields. Taking the decomposition of the first row in Tab. \[Tab:Summary-O3O4\], one can find the same diagram but with a bino $\widetilde{B}({\bf 1},{\bf 1})_{0}$ instead of the wino $\widetilde{W}$.
Models and phenomenology
========================
In this section we discuss the phenomenology of the different 1-loop models presented above. We will start with a brief discussion of the different model classes and an overview of commonalities that all these 1-loop models share. We then discuss two example models in some more detail.
General discussion\[sect:3.1\]
------------------------------
Our results listed in Tabs. \[Tab:SU3-decom\]-\[Tab:Summary-O5\] summarize the possible particle content that allows to construct models with 1-loop induced proton decay, we call this the “proto-models”. However, not all allowed choices of quantum numbers will automatically result in models, in which the 1-loop contribution to proton decay will be the dominant one. To see this in a simple example, consider decomposition \#1 from Tab. \[Tab:SU3-decom\] and the decomposition of the third row in Tab. \[Tab:Summary-O3O4\] with $\alpha=1/2$. In this case $S$ is identified with $S(\overline{\bf 3}, {\bf
1})_{1/3}$. The quantum numbers of this scalar allow to write down the following two interactions with standard model fermions: $QQS^{\dagger}$ and $QLS$. The product of these interactions, after integrating out $S$, generate ${\mathcal O}_3$ [*at tree-level*]{}.[^5] Thus, unless there is a strong hierarchy between the different Yukawa interactions, one expects that the tree-level contribution dominates the decay rate. One can eliminate such an unwanted hierarchy in couplings using additional symmetries. The simplest possibility is to just assign the particles running in the loop to be odd under a $Z_2$, while all the standard model particles are even. We classify models, which need such an additional symmetry to avoid unwanted tree-level proton decay, as class-I models. We discuss one example model from this class in Section \[sect:modI\].
In addition, there exist choices of quantum numbers, for which tree-level 2-body decays are not allowed, but higher multiplicity final states are generated at tree-level together with the 1-loop diagrams. For example, if $S$ and $S'$ are chosen to be $S({\bf 3},
{\bf 2})_{1/6}$ and $S'({\bf 1}, {\bf 2})_{1/2}$, which have the same charges as a scalar leptoquark and the SM Higgs field, the corresponding models will produce 3-body proton decays, such as $p\to \pi^+ \pi^+ e^-$ via an effective $d=9$ operator. In such cases, one expects in general that the $d=6$ 1-loop operator dominates over the $d=9$ tree-level operator for typical scales $\Lambda \gsim 1$ TeV, see Eq. . However, if there is some hierarchy in the Yukawa couplings, $(Y_3Y_4) \ll (Y_1Y_2)$, one can arrange the 3-body decays to dominate over the 2-body ones and one needs again a symmetry to assure that the loop dominates over the tree-level contribution. Our first example model is exactly of this type, see Section \[sect:modI\].
Finally, there are choices, where the particle content of the 1-loop model is such that tree-level proton decay can occur only at $d=12$ and higher (usually leading to proton decay with 5-body final states). In these cases the 1-loop $d=6$ decay will win over the tree-level decays for all practical choices of model parameters. We consider such models interesting and define these models as “class-II” models, since no symmetry is required to make the 1-loop $d=6$ decays dominant. We will discuss one concrete example model in Section \[sect:modII\].
Obviously, the main difference between models in class-I and class-II is that in class-I the lightest particle will be absolutely stable. This opens up the possibility to connect proton decay to dark matter, but requires an electrically neutral particle in the loop. We will come back to a more detailed discussion of this point in Section \[sect:modI\].
Let us now turn to a rough estimate of the proton decay half-life. Using results from lattice QCD calculations [@Yoo:2018fyn; @Aoki:2017puj; @Aoki:2013yxa; @Aoki:2006ib] and chiral perturbation theory [@Claudson:1981gh; @Chadha:1983sj; @Aoki:2008ku], the two-body proton decay half-life can be calculated as:[^6] $$\begin{aligned}
\tau
\simeq&
\frac{1}{\frac{m_{p}}{32\pi}
\left[
1 - \frac{m_{\text{meson}}^{2}}{m_{p}^{2}}
\right]^{2}
\left|
W \frac{Y_{1}Y_{2}Y_{3}Y_{4}}{16 \pi^{2} \cdot 6 \cdot M^{2}}
\right|^{2}}
\nonumber
\\
\sim &
10^{34}\text{[yrs]}
\left[
\frac{M}{1\text{TeV}}
\right]^{4}
\left[
\frac{3 \times 10^{-6}}{{\bar Y}}
\right]^{8}.
\label{eq:lifetime-estimation}\end{aligned}$$ Note that half-live estimates for different operators, ${\cal O}_1$-${\cal O}_5$, differ slightly due to the different possible final states. For the numerical estimate we use the charged pion mass for the mass $m_{\text{meson}}$ of the daughter meson. $W$ is the corresponding hadronic matrix element, $W \equiv \langle \text{meson}|(qq)q | p \rangle = -0.181$ \[GeV$^{2}$\]; the numerical value has recently been calculated in Ref. [@Aoki:2017puj]. The factor $1/6$ in the first equation above is due to the loop integral $I_{4}$, in the limit of equal masses. We have defined the mean coupling ${\bar Y}=(Y_{1}Y_{2}Y_{3}Y_{4})^{1/4}$, since proton decay is sensitive only to this product and used a mass scale of $1$ TeV, since we are interested in possible LHC phenomenology of these 1-loop models.
With couplings of the order of Eq. , the particles in the 1-loop diagrams can be rather long-lived. Depending on the choices of parameters, i.e. Yukawa couplings and mass hierarchies of the new particles, decay lengths can vary from unmeasurably short to many meters. The collider phenomenology of long-lived particles has recently attracted a lot of attention in the literature, see for example Refs. [@Helo:2013esa; @delaPuente:2015vja; @Liu:2018wte; @Lee:2018pag; @Belanger:2018sti; @Cottin:2018kmq; @Cottin:2018nms; @Cottin:2019drg]. There are also plans for several future experiments, dedicated to the search for ultra long-lived particles, see for example Refs. [@Chou:2016lxi; @Gligorov:2018vkc; @Dercks:2018wum; @Helo:2018qej; @Curtin:2018mvb]. For the current status of searches for long-lived particles at the LHC, see Refs. [@Sirunyan:2018vlw; @Aaboud:2018aqj; @Sirunyan:2018njd; @Aaboud:2018hdl; @Aaboud:2018jbr; @Sirunyan:2018pwn; @Sirunyan:2018ldc; @Sirunyan:2018vjp; @Sirunyan:2017sbs; @Aaboud:2017mpt; @Sirunyan:2017jdo; @Aaboud:2017iio; @Khachatryan:2016unx; @Aaboud:2016dgf; @Aad:2015rba; @CMS:2014hka; @CMS:2014wda; @Chatrchyan:2013oca]. We will come back to a more detailed discussion of the LHC phenomenology of our 1-loop models in Sections \[sect:modI\] and \[sect:modII\] below.
Model I\[sect:modI\]
--------------------
Here we discuss one example model, corresponding to the choices \#5 in Tab. \[Tab:SU3-decom\] for colour, the second row in Tab. \[Tab:Summary-O1\] for the decomposition of $\mathcal{O}_{1}$, and the parameter $\alpha$ for the electroweak $U(1)$ hypercharge to $-1/6$. The SM charges of the mediator fields are then determined as $$\begin{aligned}
\psi(\overline{\bf 3},{\bf 2})_{-1/6},
\quad
S({\bf 3},{\bf 2})_{+1/6},
\nonumber
\\
\psi'({\bf 1},{\bf 1})_{0},
\quad
S'({\bf 1},{\bf 2})_{+1/2}.\end{aligned}$$ Note that the scalar mediator field $S'$ has the same charges as the SM Higgs field $H$. The 1-loop diagram for proton decay is shown in Fig. \[Fig:Model-Nr1\].
(5.5,3.5) (0,0)[![An example, Model-I, for a one-loop decomposition of the proton decay operator $\mathcal{O}_{1}$. This model relates proton decay to neutrino masses, the dark matter, and possibly a long-lived coloured particle at the LHC, compare with Fig. \[Fig:Pheno-Model-Nr1\].[]{data-label="Fig:Model-Nr1"}](Fig2-crop.eps "fig:"){width="5.3cm"}]{}
As discussed above, the model allows for a $d=9$ 3-body decay of the proton, unless an additional symmetry is introduced. This can be seen easily as follows: Cutting the diagram in the scalar lines, one obtains the effective operator $d_R u_R S S'^{\dagger}$. Including the decays of the scalars produces then a $\Delta(B-L)=2$ $d=9$ operator: $(d_R u_R)(\overline{L} d_R)(\overline{u_R} Q)$. It is easy to forbid this operator, simply postulating $B-L$ conservation, since the original $d=6$ operator has $\Delta(B-L)=0$. More interesting phenomenologically, however, is to assign a new $Z_2$ to the model, under which all loop particles are odd, while the SM fermions are even. In this case, unless the $Z_2$ is spontaneously broken, the lightest of the particles in the loop is absolutely stable and can be therefore a candidate for the dark matter. In the following, we will discuss this variant of the model.
Combining the information listed in Tabs. \[Tab:SU3-decom\] and \[Tab:Summary-O1\], one can find the coefficient of the effective interaction of proton decay processes as $$\begin{aligned}
\mathcal{C}_{1}
=
-M_{\psi} M_{\psi'} I_{4}
Y_{1}Y_{2}Y_{3}Y_{4},
\label{eq:C1-Model-Nr1}\end{aligned}$$ with which the effective Lagrangian Eq. is given as $\mathscr{L}_{\text{eff}} \equiv \mathcal{C}_{1} \mathcal{O}_{1}$. The effective operator $\mathcal{O}_{1}$ causes decay of a proton in two modes, and the rates are calculated with the coefficient Eq. as[^7] $$\begin{aligned}
\Gamma (p \rightarrow \pi^{+} \bar{\nu}_{e}/\pi^{0} e^{+}) =&
\frac{m_{p}}{32 \pi}
\left[
1 - \frac{m_{\pi^{+/0}}^{2}}{m_{p}^{2}}
\right]^{2}
\left|
W_{0}
\mathcal{C}_{1}
\right|^{2},
\label{eq:Pdecayrate-formula}\end{aligned}$$ where the hadronic matrix elements $W_{0}$ are found in Ref. [@Aoki:2017puj]: $W_{0} = -0.186 (-0.131)$ \[GeV$^{2}$\] for the $\pi^{+}$ ($\pi^{0}$) mode. All decompositions in Tab. \[Tab:Summary-O1\], which result in the operator $\mathcal{O}_{1}$, predict roughly the same size of the rates for both decay modes. Therefore, if it turns out that the rates of the two modes are very different, models based on $\mathcal{O}_{1}$ will be disfavored. As we have already seen in Eq. , the mean of the couplings ${\bar Y}=(Y_{1}Y_{2}Y_{3}Y_{4})^{1/4}$ should be order few ${\cal O}(10^{-6})$ for masses accessible at the LHC.
The interaction $Y_{3} \overline{\psi'}({\bf 1},{\bf 1})_{0} L S'({\bf 1},{\bf 2})_{+1/2}$ in Fig. \[Fig:Model-Nr1\] can be identified with the corresponding interaction that appears in the scotogenic model [@Ma:2006km], since $\psi'({\bf 1},{\bf 1})_{0}$ can be interpreted as a $\nu_R$. Note that Model-I is not the only decomposition that contains such an interaction. Requiring the fields ($\psi'$ and $S'$) relevant for the radiative neutrino mass generation to be colour singlets, we have only one choice left for the assignment of the colour charges, which is \#5 in Tab. \[Tab:SU3-decom\]. Assuming the $\psi'$ to be a singlet under the electroweak $SU(2)$ as in the original scotogenic model, we have the choices \#2, 6, 9, 14, 17, 22 in Tab. \[Tab:Summary-O1\] for $\mathcal{O}_{1}$ and \#1 and 7 in Tab. \[Tab:Summary-O3O4\] for $\mathcal{O}_{3,4}$. In total, we have eight possibilities for loop-induced proton decays which can accommodate neutrino masses and dark matter with the scotogenic-type realization.[^8]
The phenomenology of the scotogenic model has been studied in many papers, see for example Refs. [@Kubo:2006yx; @Boehm:2006mi; @Sierra:2008wj; @Suematsu:2009ww; @Gelmini:2009xd; @Schmidt:2012yg; @Restrepo:2013aga; @Racker:2013lua; @Toma:2013zsa; @Molinaro:2014lfa; @Vicente:2014wga; @Hessler:2016kwm; @Hagedorn:2018spx]. We will therefore only briefly summarize the most important aspects of its phenomenology here and comment on the differences between our Model-I and the original scotogenic model.
(8,7) (-0.5,0.5)[![Phenomenology of Model-I: Majorana masses for neutrinos (upper left), example of a DM annihilation process for the freeze-out scenario (upper right), a charged lepton flavour violating process (lower left), and a decay chain of the coloured mediator field $\psi$ at the LHC (lower right). []{data-label="Fig:Pheno-Model-Nr1"}](Fig3-crop.eps "fig:"){width="9cm"}]{}
To generate a Majorana mass term for neutrinos from the 1-loop diagram (upper-left in Fig. \[Fig:Pheno-Model-Nr1\]) one introduces a scalar quartic interaction between the SM Higgs and the new scalar [@Ma:2006km]: $$\begin{aligned}
\mathscr{L}
\supset
\lambda_5
(S'^{\dagger} H)(S'^{\dagger} H)
+
{\rm H.c.}
\label{eq:L-scoto-H-Sdash-mixing}\end{aligned}$$ The flavour structure of the Majorana mass term for neutrinos in the scotogenic model can be expressed as $$\begin{aligned}
(m_{\nu})^{\alpha \beta} =
\sum_{i}
{(Y_{3}^{\sf T})^{\alpha}}_{i}
\Lambda_{i}
{(Y_{3})_{i}}^{\beta}.\end{aligned}$$ which shows that at least two $\psi'$ are necessary to reproduce the two mass squared differences measured by neutrino oscillation data. One can find the loop integral $\Lambda_{i}$ in Ref. [@Ma:2006km], which is $$\begin{aligned}
\Lambda_{i}
\equiv&
\frac{M_{\psi'_{i}}}{16 \pi^{2}}
\left[
f
(
M_{\psi'_{i}}^{2}/M_{\text{Re}S'^{0}}^{2}
)
-
f
(
M_{\psi'_{i}}^{2}/M_{\text{Im}S'^{0}}^{2}
)
\right]
\label{eq:scoto-Lambda}\end{aligned}$$ where the function $f(x)$ is defined as $f(x) = -\ln t/(1 - x)$. The splitting between the mass $M_{\text{Re}S'^{0}}$ of the real part of the neutral component of $S'$ and that $M_{\text{Im}S'^{0}}$ of the imaginary part is given by the scalar mixing term Eq. as $M_{\text{Re}S'^{0}}^{2} -
M_{\text{Im}S'^{0}}^{2} = 2 \lambda_{5} \langle H^{0} \rangle^{2}$. It is clear from Eq. and also the diagram in Fig. \[Fig:Pheno-Model-Nr1\] that the loop integral vanishes in the limit where the mass splitting, which is proportional to the scalar mixing, goes to zero. In short, the size of the neutrino masses is controlled by the scalar mixing coupling $\lambda_5$, the Yukawa interaction $Y_{3}$, and the masses of the mediators. Here we are interested in the phenomenology of the mediators with masses of the TeV scale. Setting the mediator masses to roughly a TeV, one finds that either $Y_{3}$ should be small, say of the order of ${\cal
O}(10^{-5})$ for $\lambda_5$ order ${\cal O}(1)$, or $\lambda_5$ should be order ${\cal O}(10^{-10}\text{-}10^{-8})$ for Yukawas order $0.1$-$1$ to obtain $m_{\nu} \sim \mathcal{O}(0.1)$ eV.
The $Y_{3}$ interaction also mediates charged lepton flavour violating (cLFV) processes. Studies with a parameter scan, e.g., Refs. [@Toma:2013zsa; @Vicente:2014wga], conclude that the $\ell_{\alpha} \rightarrow \ell_{\beta} \gamma$ processes [*currently*]{} places the most stringent constraints on the model parameters in wide area of the parameter space. A general formula for the rate of this cLFV process has been presented in Ref. [@Lavoura:2003xp]. It can be written as $$\begin{aligned}
\Gamma (\ell_{\alpha} \rightarrow \ell_{\beta} \gamma)
=
\frac{e^{2} m_{\ell_{\alpha}}^{5}}{16\pi}
\left|
\sum_{i}
(Y_{3}^{\dagger}{)_{\beta}}^{i}
(Y_{3}{)_{i}}^{\alpha}
\left[
-\overline{c}
+
\frac{3}{2}\overline{d}
\right]_{i}
\right|^{2}.
\label{eq:cLFV}\end{aligned}$$ The loop integral factor in Eq. is given as $$\begin{aligned}
\left[
-\overline{c}
+
\frac{3}{2} \overline{d}
\right]_{i}
=
\frac{{\rm i}}{16 \pi^{2}}
\frac{1}{M_{S'^{+}}^{2}}
\left[
\frac{2 t_{i}^{2} + 5t_{i} -1}{12 (t_{i}-1)^{3}}
-
\frac{t_{i}^{2} \ln t_{i}}{2 (t_{i}-1)^{4}}
\right]\end{aligned}$$ with $t_{i} \equiv M_{\psi'_{i}}^{2}/M_{S'^{+}}^{2}$. The non-observation of the $\mu \rightarrow e \gamma$ process [@TheMEG:2016wtm] suggests ${(Y_{3})_{i}}^{\alpha\in\{e,\mu\}} \lesssim \mathcal{O}(0.01\text{-}1)$ for mediators with masses of $\mathcal{O}(1)$ TeV [@Vicente:2014wga]. For future prospects of the experimental bounds to the cLFV processes are summarized in e.g., Ref. [@Calibbi:2017uvl].
We now turn to a brief discussion of dark matter. There are two possible candidates in Model-I. The scalar $S'({\bf 1},{\bf 2})_{+1/2}$ can be identified with the inert doublet, discussed many times in the literature. For inert doublet DM see, for example Refs. [@LopezHonorez:2006gr; @Arhrib:2013ela]. The second candidate is the neutral fermion. For a detailed study of singlet fermion DM in the scotogenic model see, for example Refs. [@Vicente:2014wga; @Hagedorn:2018spx].
Suppose the lightest Majorana fermion $\psi'_{1}$ is the DM field and was thermally produced and frozen out in the early Universe. In our model, there are two pair-annihilation modes for $\psi'_{1}$, which are $\psi'_{1} \psi'_{1} \rightarrow Q{\bar Q}$ and $L{\bar L}$. The $L{\bar L}$ mode, mediated by the $Y_{3}$ interaction, is the only mode in the original scotogenic model. For this diagram to be efficient enough, $Y_{3}$ should be large, which produces a mild tension between upper limits from cLFV and the minimal $Y_{3}$ required to reproduce the correct relic density [@Vicente:2014wga].[^9] However, the model we are discussing here also has the interaction $Y_{4}$ of the DM field $\psi'$ with $Q$ and can annihilate through the $\psi'\psi'\rightarrow Q{\bar Q}$ channel, see Fig. \[Fig:Pheno-Model-Nr1\]. The cross section for this pair-annihilation process can be roughly estimated as $$\begin{aligned}
&\sum_{q=u,d}
\langle
\sigma(\psi' \psi' \rightarrow q \bar{q})
v
\rangle
\sim
2
\times
\pi
\left[
\frac{|Y_{4}|^{2}}{4\pi}
\right]^{2}
\frac{1}{M^{2}}
\times
\frac{T}{M}
\nonumber
\\
&=
2 \cdot 10^{-26}
\text{[cm$^{3}$/s]}
\left[
\frac{Y_{4}}{1.0}
\right]^{4}
\left[
\frac{\text{TeV}}{M}
\right]^{2}\end{aligned}$$ where $T$ is the freeze-out temperature. The suppression factor $T/M \sim 1/20$ comes from the fact that the annihilation amplitude is $p$-wave, since the initial state consists of two Majorana fermions, cf. e.g., Ref. [@Sigl:2017wfx]. Note that, cross sections of order $2 \cdot 10^{-26}$ \[cm$^{3}$/s\] will reproduce the correct relic density. Again, as in the case of $Y_3$, much smaller values of this coupling would be sufficient, if $S({\bf 3},{\bf 2})_{+1/6}$ is not much heavier than $\psi'$, such that co-annihilation effects become important.
Finally, we will discuss the LHC phenomenology of Model-I. We will concentrate on the coloured states $S({\bf 3},{\bf 2})_{+1/6}$ and $\psi(\overline{\bf 3},{\bf 2})_{-1/6}$. Let us first consider $S$. This scalar will decay to a 2-body final state of jets ($j$) plus missing energy ($\esl$). The decay will be prompt, unless $Y_4$ is tiny, say $Y_4 \ll 10^{-7}$. Thus, limits from standard SUSY searches apply. For example, CMS has searched for scalar quarks decaying promptly to jets plus missing energy [@Sirunyan:2017cwe]. The limits for one generation of squarks reach up to 1 TeV for neutralino mass of $m_{\tilde{\chi}^{0}} \sim 100$ GeV and weaken to roughly 600 GeV for $m_{\tilde{\chi}^{0}} \sim 400$ GeV [@Sirunyan:2017cwe]. Similar numbers can be found in ATLAS searches, for example [@Aaboud:2017vwy].
The possible decay chains for the coloured fermion $\psi^{c}({\bf 3},{\bf 2})_{+1/6}$ lead to final states of either $jj\esl$ or $jl^{\pm}\esl$. Thus, even though $\psi$ resembles a vector-like quark (VLQ) from its quantum numbers, standard VLQ searches do not apply to this state. (For a summary of CMS searches for VLQs see, for example Ref. [@Beauceron:2651966].) On the other hand, pair production of $\psi$ will lead to final states that resemble again those for SUSY searches for squarks and gluinos. However, which of the LHC searches can be used to constrain the $\psi$ depends on whether its decays are prompt or not. This in turn depends on the mass hierarchy of the particles in the loop, see Fig. \[Fig:Pheno-Model-Nr1\]. If the mass $M_\psi$ of $\psi$ is larger than either of the masses of the scalars $S({\bf 3},{\bf 2})_{+1/6}$ or $S'({\bf 1},{\bf 2})_{+1/2}$, the decays of $\psi$ are 2-body and likely prompt, unless again the corresponding Yukawa is considerably smaller than the estimate for ${\bar Y}$ of ${\cal O}(10^{-6})$ discussed above from proton decay sensitivities. ATLAS gives lower limits on the gluino mass in simplified SUSY models of order $m_{\tilde{g}}\gsim 2$ TeV [@Aaboud:2017vwy]. However, the limits on $\psi$ will be weaker, since (a) the cross section for a colour triplet is smaller than for the gluino (octet) and (b) the $\psi$ can also decay to $jl^{\pm}\esl$ with an unknown branching ratio, so this 0-lepton search [@Aaboud:2017vwy] does not always directly apply. Masses $M_{\psi}$ below 1 TeV will, however, always be excluded since ATLAS leptoquark searches [@Aaboud:2016qeg] can be combined with the SUSY search [@Aaboud:2017vwy], as long as the decays of $\psi$ are prompt.
Assume now that $M_\psi$ is smaller than the mass $M_{S}$ of the scalar $S({\bf 3},{\bf 2})_{+1/6}$. The three-body decay rate of $\psi$ to $jj\esl$ can then be estimated as $$\label{eq:3bdecay}
\Gamma(\psi(\overline{\bf 3},{\bf 2})_{-1/6} \to jj\esl)
\simeq
\frac{\left| Y_1Y_4 \right|^2}{512 \pi^3}
\left[
\frac{M_\psi}{M_S}
\right]^4
M_\psi .$$ Note that the rate for the three-body final state $jl^{\pm}\esl$ is given by the same expression, simply replacing $|Y_1Y_4|$ with $|Y_2Y_3|$ and taking $M_S$ as the mass of $S'({\bf 1},{\bf 2})_{+1/2}$. From Eq. one can estimate that for $M_\psi \gsim
M_S \simeq 1$ TeV decay lengths will become larger than the order of millimetre for Yukawas smaller than $10^{-3}$. For Yukawas as small as $10^{-6}$, see Eq. , life-times exceed already 10 seconds. Thus, the $\psi$ will hadronize before decaying. ATLAS studied constraints on long-lived coloured particles [@Aaboud:2019trc], again in the context of a supersymmetric model. From Figs. 9 and 11 in Ref. [@Aaboud:2019trc] one can estimate that $\psi$ should be heavier than $M_\psi \gsim$ 1.8-1.9 TeV for $c\tau=$3-10 m. From Ref. [@Heinrich:2018pkj] one can estimate that similar numbers will apply also for quasi-stable $\psi$.
In summary, Model-I allows to connect proton decay, dark matter and neutrino masses. If the masses of the loop particles are of order of ${\cal O}
(1\text{-}2)$ TeV, one can have also a wide range of interesting signals at the LHC. We have discussed a few possible search strategies for the LHC for the coloured particles in this model.
Model II {#sect:modII}
--------
Let us consider now a model without additional discrete symmetries, which we categorized into the second class in Section \[sect:intro\]. The full new particle content of the model is $$\begin{gathered}
\label{PartContII}
\psi(\overline{\bf 3},{\bf 1})_{+4/3} = \psi_{+4/3} ,
\quad
S({\bf 3},{\bf 2})_{+7/6} = (S_{+2/3},S_{+5/3}),
\nonumber
\\
\psi'({\bf 8},{\bf 2})_{+1/2}=(\psi^{\prime}_ 0,\psi^{\prime}_{+1}),
\quad
S'({\bf 8},{\bf 2})_{+3/2}=(S^{\prime}_{ +1},S^{\prime}_{ +2}),\end{gathered}$$ which corresponds to choose the first row in Tab. \[Tab:Summary-O2\] of the decomposition of the $ \mathcal{O}_{2}$ operator with the parameter $\alpha=4/3$ for the electroweak $U(1)$ hypercharge. For the colour representation, we take \#6 in Tab. \[Tab:SU3-decom\]. The corresponding Feynman diagram of the 1-loop proton decay is shown in Fig. \[Fig:Model-1\]. The symmetries allow to have an additional interaction $$\begin{aligned}
\mathscr{L}_{2}
=
Y_{5}
\overline{u_{R}}
L
{\rm i}\tau^{2}
S
+
{\rm H.c.},
\label{eq:L-decom}\end{aligned}$$ which does not appear in the 1-loop proton decay diagram, shown in Fig. \[Fig:Model-1\]. With this interaction, a $d=12$ effective operator $Q Q u_{R} e_{R} \overline{u_{R}} L \overline{L} u_{R}$ appears at tree-level, as shown in Fig. \[fig:diagd12\], which causes 5-body proton decays such us $p \rightarrow e^+ e^- e^+ \pi^+
\pi^-$. However, the decay modes induced from the $d=12$ operator are sub-dominant; Using Eq. , for $Y_5 \sim 10^{-2}$, we can roughly estimate the contribution of these modes to the proton total decay width to be around 40 orders of magnitude smaller than the 2-body proton decay induced by the $d=6$ effective operator $\mathcal{O}_{2}$ given through the 1-loop diagram in Fig. \[Fig:Model-1\]. Combining the information listed in Tabs. \[Tab:SU3-decom\] and \[Tab:Summary-O2\], one can find the coefficient of the effective interaction $\mathcal{O}_{2}$ of proton decay as $$\begin{aligned}
\mathcal{C}_{2}
=
-\frac{8}{3}M_{\psi} M_{\psi'} I_{4}
Y_{1}Y_{2}Y_{3}Y_{4},
\label{eq:C1-Model-2}\end{aligned}$$ and the effective interaction causes only $p \rightarrow \pi^{0} e^{+}$. As we have seen in Section \[sect:3.1\], the experimental bounds on the proton decay rate require Yukawa couplings of order $Y < \mathcal{O}(10^{-6})$ for the masses of the mediators at the TeV scale.
Within this model it is possible to have signatures that violate the lepton number by 1 unit, $\Delta L = 1$. This can be seen in the pair production of the scalar $S^{\prime}_{ + 2} $ which has two possible decay modes $S^{\prime}_{+ 2} \rightarrow l^+ l^+ 2j$ and $S^{\prime}_{ + 2} \rightarrow l^+ 3j$, as shown in Fig. \[fig:LNV\]. Therefore the pair production of the colour-octet $S^{\prime}_{ +2}$ at the LHC might lead to the lepton number violating (LNV) signal 3 lepton plus 5 jets $(l^+ l^- l^\pm 5j)$. Observation of LNV through this process is only possible, if $\Gamma(S^{\prime}_{ + 2} \rightarrow l^+ l^+ 2j )$ is of similar order to $\Gamma(S^{\prime}_{ + 2} \rightarrow l^+ 3j)$, since both final states are needed to establish that LNV is indeed taking place. Similar order of these decay widths are possible if $Y_1 Y_2 \sim Y_{3} Y_4$. Observing this LNV process also requires to have short enough decays so the decays of the particles can be prompt.
This can be achieved for instance if we assume a mass hierarchy of the particles, $M_{S^{\prime}} \gtrsim M_{\psi} (M_{\psi^\prime}) \gtrsim
M_{S} $ such that the pair production of $S^{\prime}_{ +2}$ leads to a decay chain of 2-body decays as shown in Fig. \[fig:LNV\]. If this is the case, the two decay modes of $S^{\prime}_{ + 2}$ will be $S^{\prime}_{ + 2} \rightarrow \psi^{\prime}_{ +1} l^+$ and $S^{\prime}_{ + 2} \rightarrow \psi_{ + 4/3} j$, and its decay length can be estimated as $$\begin{aligned}
\label{decayL2}
L_0(S^{\prime}_{ +2})
\sim 10^{-2} \text{[m]}
\frac{\left[10^{-6}\right]^2}{|Y_3|^2+|Y_4|^2 }
\left[\frac{\text{TeV}}{M_{S^\prime}}\right].\end{aligned}$$ Here, the choice for the Yukawa couplings $Y_3$ and $Y_4$ being order $10^{-6}$ is motivated by the current proton decay experimental bounds.
It is also possible to have long-lived particles at the LHC, which are pair-produced. Let us assume for instance that the colour-octet fermion $\psi^\prime$ is slightly lighter than the scalar $S$. Then, the decay rates of the particles $\psi^{\prime}_{ +1}\rightarrow l^+ j j$ and $\psi^{\prime}_{
0}\rightarrow \nu_l j j$ can be estimated as (see Fig. \[fd-2\]) $$\begin{aligned}
\Gamma(\psi^\prime
\rightarrow
l (\nu) jj)
\sim
\frac{|Y_{4} Y_{5}|^{2}}{512\pi^3}
\left[
\frac{M_{\psi^\prime}}{M_{S}}
\right]^4
M_{\psi^\prime},\end{aligned}$$ which leads to the estimate of the decay length $$\begin{aligned}
\label{decayL}
L_0(\psi^\prime)
\sim 30 \text{[m]}
\left[\frac{10^{-6}}{|Y_{4}|}\right]^2
\left[\frac{10^{-2}}{|Y_{5}|}\right]^2
\left[\frac{\text{TeV}}{M_{\psi}}\right]
\left[\frac{M_{S}}{M_{\psi^\prime}}\right]^4.\end{aligned}$$ Here, the Yukawa coupling $Y_{5}$, which is not constrained by the proton decay, has been set to be of the order $10^{-2}$. Eq. (\[decayL\]) shows that $\psi^{\prime }$, after being pair produced at the LHC, will become a long lived particle. Since $\psi^{\prime }$ has also colour, one can use R-hadron searches at the LHC [@Alimena:2019zri; @Aaboud:2019trc; @Farrar:2010ps; @Buccella:1985cs; @CMS:2016ybj; @Liu:2015bma] to constrain it.
In summary, in Model-II depending on the mass hierarchies of the particles that appear in the proton decay diagram, one can have either long-lived coloured/charged particles or prompt decays. The latter would allow to establish experimentally the existence of LNV.
(5.5,3.5) (0,0)[![An example, Model II, for one-loop decomposition of the proton decay operator $\mathcal{O}_{2}$.[]{data-label="Fig:Model-1"}](Fig4-crop.eps "fig:"){width="5.3cm"}]{}
(5.5,5) (0,0)[![$d=12$ effective operator $Q Q u e \bar{u} L \bar{L} u$ at tree-level. []{data-label="fig:diagd12"}](Fig5-crop.eps "fig:"){width="5.3cm"}]{}
(7.8,5.5) (0,0)[![Pair production and corresponding possible decays modes of the colour-octet scalar $S^{\prime}_{+2}$ at the LHC.[]{data-label="fig:LNV"}](Fig6-crop.eps "fig:"){width="7.3cm"}]{}
(4.5,2) (0,0)[![ Decay channels of the colour-octet fermion $\psi^\prime$. []{data-label="fd-2"}](Fig7-crop.eps "fig:"){width="4.5cm"}]{}
Conclusions
===========
We have studied systematically the 1-loop decomposition of the $d = 6$ $B + L$ violating operators involving only SM fields. Our results are listed in tabular forms, from which all possible 1-loop ultra-violet completions of these operators involving fermions and scalars can be constructed. We have briefly discussed how to use the information provided to calculate all coefficients that enter the calculation of the proton decay rate.
We then discussed that all models, in general, can be divided into two distinct classes. Class-I models are those, for which the 1-loop proton decay exists, but is not necessarily the dominant diagram for a given model. Models in this class therefore need usually an additional symmetry, such that tree-level contributions to proton decay become forbidden. Class-II models are then simply those, for which the particle content guarantees that the 1-loop diagram is automatically the dominant contribution to proton decay.
We then turned to possible phenomenology of these models and discussed one example model from each class. Class-I models have an absolutely stable particle and thus proton decay can be connected to the dark matter in the universe. In the example we discussed, the same symmetry is responsible for 1-loop proton decay, dark matter and neutrino mass. The latter is also generated at 1-loop level, as in the scotogenic neutrino mass model. We have discussed existing constraints and possible LHC phenomenology of this model also briefly.
For the example model of class-II, we have discussed possible LHC phenomenology. It has been shown that depending on the mass hierarchy of the particles in the model, we can have particles that after being pair produced at the LHC can decay promptly, leading to LNV signals, or are long-lived coloured/charged particles. Signals without missing energy, such as the LNV signals discussed, do not appear in Model-I and so can be used to distinguish between these two classes of models.
M. H. was funded by Spanish grants FPA2017-90566-REDC (Red Consolider MultiDark), FPA2017-85216-P and SEV-2014-0398 (MINECO/AEI/FEDER, UE), as well as PROMETEO/2018/165 (Generalitat Valenciana). This work is supported in part by European Union’s Horizon 2020 research and innovation programme under the 777419-ESSnuSB, as well as by the COST Action CA15139 EuroNuNet. T. O. acknowledges the support of “Spanish Agencia Estatal de Investigación” (AEI) through the grant “IFT Centro de Excelencia Severo Ochoa SEV-2016-0597” and the EU “Fondo Europeo de Desarrollo Regional” (FEDER) through the project FPA2016-78645-P. J. C. H. is supported by Chile grant Fondecyt No. 1161463.
Factors in Tables
=================
Here we write down the basis operators given in Eqs. - with all the indices of the SM gauge groups and Lorentz spinors: $$\begin{aligned}
\mathcal{O}_{1}
\equiv&
\epsilon^{IJK}
[
(\overline{{d_{R}}^{c}})_{I\dot{a}}
(u_{R})_{J}^{\dot{a}}
]
[
(\overline{Q^{c}})_{Ki}^{a}
({\rm i}\tau^{2})^{ij}
(L)_{ja}
],
\nonumber
\\
\mathcal{O}_{2}
\equiv&
\epsilon^{IJK}
[
(\overline{Q^{c}})_{Ii}^{a}
({\rm i}\tau^{2})^{ij}
(Q)_{Jja}
]
[
(\overline{{u_{R}}^{c}})_{K \dot{a}}
(e_{R})^{\dot{a}}
],
\nonumber
\\
\mathcal{O}_{3}
\equiv&
\epsilon^{IJK}
[
(\overline{Q^{c}})_{Ii}^{a}
({\rm i}\tau^{2})^{ij}
(Q)_{Jja}
]
[
(\overline{Q^{c}})_{Kk}^{b}
({\rm i}\tau^{2})^{kl}
(L)_{l b}
],
\nonumber
\\
\mathcal{O}_{4}
\equiv&
\epsilon^{IJK}
[
(\overline{Q^{c}})_{Ii}^{a}
({\rm i}\tau^{2} \tau^{d})^{ij}
(Q)_{Jja}
]
[
(\overline{Q^{c}})_{Kk}^{b}
({\rm i}\tau^{2} \tau^{d})^{kl}
(L)_{l b}
],
\nonumber
\\
\mathcal{O}_{5}
\equiv&
\epsilon^{IJK}
[
(\overline{{d_{R}}^{c}})_{I \dot{a}}
(u_{R})_{J}^{\dot{a}}
]
[
(\overline{{u_{R}}^{c}})_{K \dot{b}}
(e_{R})^{\dot{b}}
],\end{aligned}$$ where the different indices are introduced to describe the different representations: $I,J,K \in \{1,2,3\}$ for a triplet under the colour $SU(3)$, $i,j,k,l \in \{1,2\}$ for a doublet under the electroweak $SU(2)$, $d \in \{1,2,3\}$ on the Pauli matrices $\tau^{d}$ for a triplet under the $SU(2)$, $a,b \in \{1,2\}$ for a left-handed 2-spinor and $\dot{a},\dot{b} \in \{\dot{1},\dot{2}\}$ for a right-handed 2-spinor. The position of the indices also depends on the representation of the field: A lower $I$ for ${\bf 3}_{I}$ and an upper $I$ for $\overline{\bf 3}^{I}$. The index on the $\overline{\bf 2}^{i}$ representation of $SU(2)$ can be lowered as ${\bf 2}_{i}$ with $({\rm i} \tau^{2})_{ij}$. On the position of the spinor indices, we follow the notation that is widely adopted in literature, e.g. Ref. [@Haber:1984rc]; the standard positions are determined as $(\psi_{L})_{a}$ and $(\psi_{R})^{\dot{a}}$ and the contraction is taken as $(\overline{\psi_{R}})^{a}(\psi_{L})_{a}$ and $(\overline{\psi_{L}})_{\dot{a}} (\psi_{R})^{\dot{a}}$ to form Lorentz scalars.
The ordering of the field operators in the decomposed interactions are determined as given in Eq. . In order to make the Yukawa interactions singlets under the colour $SU(3)$, we plug the total anti-symmetric tensors ($\epsilon^{IJK}$ and $\epsilon_{IJK}$), the Gell-Mann matrices (${(\lambda^{A})_{I}}^{J}$), and the Clebsch-Gordan (CG) coefficient matrices ($(T_{\bf 6})_{X}^{IJ}$ and $(T_{\overline{\bf 6}})^{X}_{IJ}$) into the interactions accordingly, where the index $A \in \{1 \cdots 8\}$ is for an octet, and a lower (upper) $X \in\{1\cdots 6\}$ is for ${\bf 6}$ ($\overline{\bf 6}$). For the contraction of the electroweak $SU(2)$ indices, we use the anti-symmetric tensors ($({\rm i}\tau^{2})^{ij}$ and $({\rm i}\tau^{2})_{ij}$) and the Pauli matrices (${(\tau^{d})_{i}}^{j}$). The CG matrices for the sextet representations are defined as $$\begin{gathered}
(T_{\bf 6})_{1}^{IJ} =
\begin{pmatrix}
1 & &
\\
& 0 &
\\
& & 0
\end{pmatrix}
,
\quad
(T_{\bf 6})_{2}^{IJ} =
\begin{pmatrix}
0 & \frac{1}{\sqrt{2}} &
\\
\frac{1}{\sqrt{2}} & 0 &
\\
& & 0
\end{pmatrix}
\nonumber
\\
(T_{\bf 6})_{3}^{IJ} =
\begin{pmatrix}
0 & &
\\
& 1 &
\\
& & 0
\end{pmatrix},
\quad
(T_{\bf 6})_{4}^{IJ} =
\begin{pmatrix}
0 & & \frac{1}{\sqrt{2}}
\\
& 0 &
\\
\frac{1}{\sqrt{2}} & & 0
\end{pmatrix}
\nonumber
\\
(T_{\bf 6})_{5}^{IJ} =
\begin{pmatrix}
0 & &
\\
& 0 & \frac{1}{\sqrt{2}}
\\
& \frac{1}{\sqrt{2}} & 0
\end{pmatrix},
\quad
(T_{\bf 6})_{6}^{IJ} =
\begin{pmatrix}
0 & &
\\
& 0 &
\\
& & 1
\end{pmatrix},\end{gathered}$$ and $T_{\overline{\bf 6}}$ are defined in the same manner.
Let us demonstrate the operator projection (=re-integrate out the mediator fields), keeping all the indices, i.e., we explicitly derive a basis operator(s) from a decomposition with all the coefficients, signs and factors. As an example, we take the basis operator $\mathcal{O}_{1}$ and decompose it with the mediators with \#4 in Tab. \[Tab:SU3-decom\] \#12 for Tab. \[Tab:Summary-O1\], i.e, $$\begin{gathered}
\psi({\bf 3}, {\bf 3})_{\alpha},
\quad
S({\bf 8},{\bf 3})_{\alpha+1/3},
\nonumber
\\
\psi'({\bf 6},{\bf 3})_{\alpha - 1/3}
\quad
S'({\bf 6}, {\bf 2})_{\alpha+1/6}.\end{gathered}$$ The Yukawa interactions are defined as $$\begin{aligned}
\mathscr{L}
=&
Y_{1}
(\overline{{d_{R}}^{c}})_{I \dot{a}}
{({\lambda^{\sf T}}^{A})^{I}}_{I'}
({\psi_{L}}^{c})^{I' d \dot{a}}
S^{A d}
\nonumber
\\
&+
Y_{2}
(\overline{{\psi_{L}}^{c}})_{J'}^{d' a}
(T_{\bf 6})^{J'J}_{X}
(Q)_{J i a}
{({\tau^{\sf T}}^{d'})^{i}}_{i'}
(S'^{\dagger})^{X i'}
\nonumber
\\
&
+
Y_{3}
(\overline{\psi'_{R}})^{X' f b}
(L)_{j b}
({\rm i} \tau^{2} \tau^{f})^{jj'}
S'_{X' j'}
\nonumber
\\
&+
Y_{4}
\epsilon^{KLM}
(\overline{{u_{R}}^{c}})_{K \dot{b}}
{(\lambda^{A'})_{L}}^{N}
(T_{\overline{\bf 6}})^{Y}_{NM}
(\psi'_{R})_{Y}^{f' \dot{b}}
(S^{\dagger})^{A' f'}
\nonumber
\\
&+
{\rm H.c.}\end{aligned}$$ The effective proton decay operator resulting from the box diagram mediated by them can be calculated as follows. First, the mediator fields are contracted, which give the propagators: $$\begin{aligned}
\mathscr{L}_{\text{eff}}
=&
Y_{1}Y_{2}Y_{3}Y_{4}
{({\lambda^{\sf T}}^{A})^{I}}_{I'}
(T_{\bf 6})^{J'J}_{X}
\epsilon^{KLM}
{(\lambda^{A'})_{L}}^{N}
(T_{\overline{\bf 6}})^{Y}_{NM}
\nonumber
\\
&\times
{({\tau^{\sf T}}^{d'})^{i}}_{i'}
({\rm i} \tau^{2} \tau^{f})^{jj'}
\Bigl\langle
S^{A d}
(S^{\dagger})^{A' f'}
\Bigr\rangle
\Bigl\langle
S'_{X' j'}
{(S'^{\dagger})}^{X i'}
\Bigr\rangle
\nonumber
\\
&\times
(\overline{{d_{R}}^{c}})_{I \dot{a}}
\Bigl\langle
({\psi_{L}}^{c})^{I' d \dot{a}}
(\overline{{\psi_{L}}^{c}})_{J'}^{d' a}
\Bigr\rangle
(Q)_{J i a}
\nonumber
\\
&\times
(\overline{{u_{R}}^{c}})_{K \dot{b}}
\Bigl\langle
(\psi'_{R})_{Y}^{f' \dot{b}}
(\overline{\psi'_{R}})^{X' f b}
\Bigr\rangle
(L)_{j b}
\nonumber
\\
=&
Y_{1}Y_{2}Y_{3}Y_{4}
{({\lambda^{\sf T}}^{A})^{I}}_{I'}
(T_{\bf 6})^{I'J}_{X}
\epsilon^{KLM}
{(\lambda^{A})_{L}}^{N}
(T_{\overline{\bf 6}})^{X}_{NM}
\nonumber
\\
&\times
({\rm i} \tau^{2} \tau^{d})^{ji'}
{(\tau^{d})_{i'}}^{i}
\int
\frac{{\rm d}^{d} p}{(2\pi)^{d} {\rm i}}
\frac{{\rm i}}{p^{2} - M_{S}^{2}}
\frac{{\rm i}}{p^{2} - M_{S'}^{2}}
\nonumber
\\
&\times
(\overline{{d_{R}}^{c}})_{I \dot{a}}
\frac{-{\rm i} p_{\rho} (\overline{\sigma}^{\rho})^{\dot{a}a}}
{p^{2} - M_{\psi}^{2}}
(Q)_{J i a}
(\overline{{u_{R}}^{c}})_{K \dot{b}}
\frac{{\rm i} p_{\sigma} (\overline{\sigma}^{\sigma})^{\dot{b}b}}
{p^{2} - M_{\psi'}^{2}}
(L)_{j b}.\end{aligned}$$ Next, the $SU(3)$ and the $SU(2)$ indices are rearranged: $$\begin{aligned}
\mathscr{L}_{\text{eff}}
=&
Y_{1}Y_{2}Y_{3}Y_{4}
\left[
4 \epsilon^{IJK}
\right]
\left[
- 3 ({\rm i} \tau^{2})^{ij}
\right]
\left[
-\frac{1}{4} J_{4}
\right]
\nonumber
\\
&\times
(\overline{{d_{R}}^{c}})_{I \dot{a}}
(\overline{\sigma}^{\rho})^{\dot{a}a}
(Q)_{J i a}
(\overline{{u_{R}}^{c}})_{K \dot{b}}
(\overline{\sigma}_{\rho})^{\dot{b}b}
(L)_{j b}.\end{aligned}$$ Here we arrived at the step shown in Eq. . If necessary, the $SU(3)$ indices are renamed so that they fit to the ordering in the corresponding basis operator. This step may give an additional sign (“$SU(3)$ sign” in the tables): $$\begin{aligned}
\mathscr{L}_{\text{eff}}
=&
Y_{1}Y_{2}Y_{3}Y_{4}
\left[
4 \epsilon^{IJK}
\right]
\left[
-1
\right]
\left[
- 3 ({\rm i} \tau^{2})^{ij}
\right]
\left[
-\frac{1}{4} J_{4}
\right]
\nonumber
\\
&\times
(\overline{{d_{R}}^{c}})_{I \dot{a}}
(\overline{\sigma}^{\rho})^{\dot{a}a}
(Q)_{K i a}
(\overline{{u_{R}}^{c}})_{J \dot{b}}
(\overline{\sigma}_{\rho})^{\dot{b}b}
(L)_{j b}.\end{aligned}$$ Finally, the Fierz transformation (rearrangement of the Lorentz indices) is carried out: $$\begin{aligned}
\mathscr{L}_{\text{eff}}
=&
Y_{1}Y_{2}Y_{3}Y_{4}
\left[
4
\right]
\left[
-1
\right]
\left[
- 3
\right]
\left[
-\frac{1}{4} J_{4}
\right]
\left[
2
\right]
\nonumber
\\
&\times
\epsilon^{IJK}
(\overline{{d_{R}}^{c}})_{I \dot{a}}
(u_{R})_{J}^{\dot{a}}
(\overline{Q^{c}})_{K i}^{a}
({\rm i} \tau^{2})^{ij}
(L)_{j a}
\nonumber
\\
=&
-6 Y_{1}Y_{2}Y_{3}Y_{4} J_{4} \mathcal{O}_{1}.
\label{eq:Op-Projection-example}
\end{aligned}$$
In this example, we obtain 4 for the $SU(3)$ coefficient, $-$ for the $SU(3)$ sign, $-3$ for the $SU(2)$ coefficient, $2$ for the factor of the Fierz transformation for Lorentz indices, and $-J_{4}/4$ for the loop integral factor. The loop integral factors $I_{4}$ and $J_{4}$ that appear in Tabs. \[Tab:Summary-O1\]-\[Tab:Summary-O5\] are defined as $$\begin{aligned}
I_{4} \equiv&
\int \frac{{\rm d}^{4} k}{(2\pi)^{4}{\rm i}}
\frac{1}{
(k^{2} - M_{\psi}^{2})
(k^{2} - M_{S}^{2})
(k^{2} - M_{\psi'}^{2})
(k^{2} - M_{S'}^{2})
}
\\
J_{4} \equiv&
\int \frac{{\rm d}^{4} k}{(2\pi)^{4}{\rm i}}
\frac{k^{2}}{
(k^{2} - M_{\psi}^{2})
(k^{2} - M_{S}^{2})
(k^{2} - M_{\psi'}^{2})
(k^{2} - M_{S'}^{2})
}\end{aligned}$$ In the limit where all the mediator masses are identical, the integrals converge to $$\begin{aligned}
I_{4} \rightarrow \frac{1}{16 \pi^{2}} \frac{1}{6} \frac{1}{M^{4}},
\quad
J_{4} \rightarrow -\frac{1}{16 \pi^{2}} \frac{1}{3} \frac{1}{M^{2}},\end{aligned}$$ where $M$ is the common value of the masses. All the information to reproduce the coefficient of the effective operator $\mathcal{O}_{1}$ in Eq. can be found in Tabs. \[Tab:SU3-decom\] and \[Tab:Summary-O1\].
[116]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
() (), .
(), ****, (), .
(), ****, (), .
(), ****, (), .
(), ****, (), .
(), ****, (), .
(), ****, (), .
(), ****, (), .
(), ****, (), .
(), ****, ().
() (), .
() (), .
(), ****, (), .
(), in ** (), pp. , .
, , , , , ****, (), .
(), .
, , , , , ****, (), .
, in ** (), [Link](https://indico.fnal.gov/contributionDisplay.py?contribId=118&sessionId=3&confId=6248).
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, , , ****, ().
, , , ****, ().
, , , ****, (), .
, ****, (), .
, ****, (), .
, ****, (), .
, , , , ****, (), .
, , , , , , ****, (), .
, in ** (), , [Link](http://www.slac.stanford.edu/econf/C1307292/docs/IntensityFrontier/BaryonNo-13.pdf).
, ****, ().
, ****, ().
, ****, ().
, , , ****, (), .
, ****, (), .
, ****, (), .
, ****, (), , .
, ****, (), .
, ****, (), .
, ****, (), .
, , , ****, (), .
, ****, (), .
, , , , , ****, (), .
, , , ****, (), .
, , , , ****, (), .
, , , (), .
, , , , ****, (), .
, , , ****, (), .
, , , , ****, (), .
, , , ****, ().
, ****, ().
, , , , , , , (), ****, (), .
, , , ****, (), , .
, ****, (), .
, , (), .
, , , (), .
(), .
, , , ****, (), .
, , , ****, (), .
, , , (), .
, , , ****, (), .
, , , , , ****, (), .
, , , (), .
, , , ****, (), .
(), .
(), (), .
(), (), .
(), (), .
(), ****, (), .
() (), .
(), ****, (), .
(), ****, (), .
(), ****, (), .
(), ****, (), .
(), ****, (), .
(), ****, (), .
(), ****, (), .
(), ****, (), .
(), ****, (), .
(), ****, (), .
(), ****, (), .
(), ****, (), .
(), ****, (), .
, , , ****, (), .
, , , , , ****, (), .
, , , , , ****, (), .
, , , ****, (), .
, , , ****, (), .
, , , ****, (), .
, , , ****, (), .
, ****, (), .
, ****, (), .
, , , ****, (), .
, ****, (), .
, , , , ****, (), .
, , , , ****, (), .
, ****, (), .
(), ****, (), .
, ****, (), .
, , , , ****, (), .
, , , , ****, (), .
, **, vol. of ** (, ), ISBN .
(), ****, (), .
(), ****, (), .
(), , , (), [Link](http://cds.cern.ch/record/2651966).
(), ****, (), .
() (), .
, Ph.D. thesis, (), [Link](https://www.etp.physik.uni-muenchen.de/publications/theses/download/phd_jheinrich.pdf).
(), .
, , , , ****, (), .
, , , ****, ().
() ().
, ****, (), .
, ****, ().
[^1]: The complete list of high-$d$ operators can easily be obtained with [Sym2Int]{} [@Fonseca:2017lem].
[^2]: The coefficient ${\cal C}$ in Eq. is ${\cal C} \propto Y^4$ in 1-loop $d=6$ models.
[^3]: See refs. [@Alonso:2014zka; @Helo:2018bgb] for the $d=6$ operators for proton decay with a SM singlet fermion (sterile neutrino, aka right-handed neutrino).
[^4]: Forbidding a Yukawa interaction that mediates a $d=6$ proton decay operator at the tree level by a symmetry and allowing the soft breaking of the symmetry by the mass term of a mediator field, one can induce a triangle diagram solely at the loop level. For more discussions on the realizations, see Ref. [@Bonnet:2012kz]. In this study, we do not pursue the possibility of such a setup with a symmetry and its soft breaking.
[^5]: $S(\overline{\bf 3}, {\bf 1})_{1/3}$ is not the only choice, that will lead to tree-level proton decay. The same argument applies to $S(\overline{\bf 3}, {\bf 3})_{1/3}$ and $S(\overline{\bf 3},
{\bf 1})_{4/3}$.
[^6]: The simple estimate, Eq. would give ${\bar Y}$ roughly a factor 2.5 smaller.
[^7]: In this estimate we do not take into account the effect of the renormalization group running [@Abbott:1980zj; @Alonso:2014zka] of the operators, and use the coefficient at the scale of the proton mass. This is sufficient for our rough estimates.
[^8]: The scotogenic type diagram for neutrino masses can be drawn with the Majorana fermion ($\psi'$) in the adjoint representations under the SM gauge symmetries. If we relax the requirements to include the adjoint representations, we have more possibilities: \#6, 7, and 8 in Tab. \[Tab:SU3-decom\] for colour, and \#3, 7, 12, 15, 20, and 23 in Tab. \[Tab:Summary-O1\] for $\mathcal{O}_{1}$ and \#2, and 8 in Tab. \[Tab:Summary-O3O4\] for $\mathcal{O}_{3,4}$.
[^9]: It is possible to solve this problem of the overabundance of DM field in scotogenic models in regions of parameter space where co-annihilation processes are sizeable, see for example Refs. [@Vicente:2014wga; @Hagedorn:2018spx].
|
{
"pile_set_name": "ArXiv"
}
|
0.3in
[Immigrated urn models - asymptotic properties and applications]{}\
L.X. Zhang$^{\mbox{\tiny a}}$, F. Hu$^{\mbox{\tiny b},*}$, S.H. Cheung$^{\mbox{\tiny c}}$, W.S. Chan$^{\mbox{\tiny d}}$\
[*$^{\mbox{\tiny a}}$Department of Mathematics, Zhejiang University, Hangzhou 310028, People’s Republic of China*]{}\
[*$^{\mbox{\tiny b}}$Department of Statistics, University of Virgina, Halsey Hall, Charlottesville, Virginia 200904-4135, USA*]{}\
[*$^{\mbox{\tiny c}}$Department of Statistics, The Chinese University of Hong Kong, Shatin, HongKong, People’s Republic of China*]{}\
[*$^{\mbox{\tiny d}}$Department of Finance, The Chinese University of Hong Kong, Shatin, HongKong, People’s Republic of China*]{}
0.2in [**Abstract**]{}
Urn models have been widely studied and applied in both scientific and social disciplines. In clinical studies, the adoption of urn models in treatment allocation schemes has been proved to be beneficial to both researchers, by providing more efficient clinical trials, and patients, by increasing the probability of receiving the better treatment. In this paper, we endeavor to derive a very general class of immigrated urn models that incorporates the immigration mechanism into the urn process. Important asymptotic properties are developed and illustrative examples are provided to demonstrate the applicability of our proposed class of urn models. In general, the immigrated urn model has smaller variability than the corresponding urn model. Therefore, it is more powerful when used in clinical trials.
0.2in [*Keywords:*]{} Adaptive designs, Asymptotic normality, Clinical trial, Urn model, Branching process with immigration, Birth and death urn, Drop-the-loser rule.
0.2in
0.3in
------------------------------------------------------------------------
0.15in $^{*}$Corresponding address: Department of Statistics, University of Virgina, Halsey Hall, Charlottesville, Virginia 200904-4135, USA. Tel: +1 434 924 3014; fax +1 434 924 3076.
[*E-mail addresses:*]{} [email protected] (L.X. Zhang), [email protected] (F. Hu), [email protected] (S.H. Cheung), [email protected] (W.S. Chan)
Introduction
============
1.1. [*Urn Models and their applications*]{}
0.1in Urn models have long been considered powerful mathematical instruments in many areas, including the physical sciences, biological sciences, social sciences, and engineering (Johnson and Kotz, 1977; Kotz and Balakrishnan, 1997). For example, in medical science, Knoblauch, Neitz, and Neitz (2006) apply an urn model to study cone ratios in human and macaque retinas. In population genetics, Benaïm, Schreiber, and Tarrés (2004) make use of a class of generalized Pólya urn models to scrutinize evolutionary processes. In economics, the model is employed to capture the mechanism of reinforcement learning by Erev and Roth (1998) and Beggs (2005). In business, a coupled Pólya urn model is utilized by Windrum (2004) to examine the recent browser war between Netscape and Microsoft. In decision science, McCabe-Dansted and Slinko (2006) employ the urn model to investigate the social choice rules that are influential factors which affect the outcomes of elections. In addition, numerous examples of applications of urn models in the areas of physics, communication theory, and computer science are also provided by Milenkovic and Compton (2004).
In recent decades, the application of urn models in clinical trials has attracted great attention from biomedical researchers. Urn models have played an important role in the advancement of adaptive design methodologies that provide treatment allocation schemes which depend on the history of treatment outcomes. The principle of adaptive designs is to skew the probability of treatment allocation so that fewer patients receive inferior treatment, amid the accumulation of information on treatment effectiveness as the clinical study progresses.
In adaptive design, an early influential urn model is the generalized Pólya urn (GPU) that was discussed in Athreya and Karlin (1968), Wei and Durham (1978), and Wei (1979). The GPU is based on the Pólya urn model which was originally constructed to model contagious diseases (Eggenberger and Pólya, 1923). For example, the GPU model of Wei (1979) can be described as following: Assume that each incoming patient is assigned to one of $K$ treatments, and the GPU is updated at each stage by adding one additional ball of the same treatment type if the treatment is successful, and by adding an additional $\frac{1}{K-1}$ ball of the other $K -1$ treatment types if the treatment is unsuccessful. By embedding the urn process in a branching process, asymptotic properties of the GPU model have been obtained in Janson (2004).
Another class of urn models with applications to clinical trials is the ternary urn model (Ivanova and Flournoy, 2001). A special case is the randomized Pólya urn (RPU) proposed by Durham, Flournoy, and Li (1998). Unlike the GPU of Wei (1979), the RPU only rewards balls to successful treatments. Another example is the birth and death urn (Ivanova et al., 2000; Ivanova and Flournoy, 2001). The ternary urn model induces diagonal generating matrices. The asymptotic properties of ternary urn model can be obtained by embedding the urn process in a branching process.
However, there are several serious drawbacks of the classical urn models: (i) it is only for binary (multinomial) responses; (ii) it has a predetermined limit that usually does not have any formal optimal properties (for example, the randomized play-the-winer rule (Wei and Durham, 1978), see Hu and Rosenberger, 2006 for details); (iii) it usually has higher variability than other procedures (Hu and Rosenberger, 2003), and therefore is less powerful.
0.2in 1.2. [*Objectives and organization of the paper*]{}
0.1in In this paper, we develop a more general class of urn models that incorporate immigration, called the immigrated urn (IMU) model hereafter. Both the GPU model and the ternary urn model can be viewed as special cases of the IMU model. An example of the IMU model is the the drop-the-loser (DL) rule proposed by Ivanova (2003). But the DL rule is limited to binary (multinomial) responses and it has a predetermined limit that does not have any formal optimal properties (Hu and Rosenberger, 2006). The IMU encompasses a much wider spectrum of urn models and incorporates the immigration process, which offers a greater flexibility in the choice of appropriate urn models in applications. The main features of the proposed IMU model are: (i) It includes other classes of urn model as special cases; (ii) It can be used to target any given allocation proportion (optimal allocations in clinical trials, see examples in Section 4); and (iii) in the class of IMU models, one can design his own IMU model that yields a smaller variability than other urn procedures.
The exploration of asymptotic properties also helps the reader to evaluate the merits of different models in this class. In the literature, asymptotic properties of urn models are usually proven by using Athreya and Ney’s (1972) technique of embedding the urn process in a continuous-time branching process. This technique does not apply to the IMU due to the possible nonhomogeneous immigration process of the IMU. It is worth to point out that the asymptotic properties of the GPU model are obtained under specific conditions of the generating matrix. However, these conditions are usually not satisfied in framework of the IMU models discussed in this article.
In this paper, we represent the IMU process by an approximated Wiener process and then obtain the asymptotic properties of the IMU. The main contributions of this paper are:
1. to formulate a general class of urn models with the immigration process (IMU model); and
2. to derive important asymptotic properties, including strong consistency and asymptotic normality of treatment proportions;
3. to obtain some useful and desirable IMU models based on the asymptotic properties.
In Section 2, the IMU process is outlined. Asymptotic properties are examined in Section 3. Illustrative examples that are related to the applications of the IMU are given in Section 4. Concluding remarks are provided in Section 5. Finally, the technical proofs are stated in the Appendix.
0.2in
The immigrated urn model
========================
Suppose subjects arrive sequentially to be randomized to one of $K$ treatments and respond immediately. An IMU model is defined as follows. Consider an urn that contains balls of $K+1$ types. Balls of types $1,\ldots, K$ represent treatments, and balls of type $0$ are the immigration balls.
Initially, there are $Z_{0,i}(>0)$ balls of type $i$, $i=0,\ldots,
K$. Let $\bm Z_0=(Z_{0,0},\ldots,Z_{0,K})$ be the initial urn composition. When a patient arrives, a ball is drawn at random and the corresponding treatment will be assigned to this person. Now, right before the $m$-th subject arrives to be randomized to a treatment, let $\bm Z_{m-1}=(Z_{m-1,0},\ldots,Z_{m-1,K})$ denote the composition of the urn. To avoid a negative likelihood of selecting a treatment, we adopt a slight adjustment to $\bm Z_{m-1}$ and let $Z_{m-1,i}^+=\max(0, Z_{m-1,i})$, $i=1,\ldots, K$, and $\bm
Z_{m-1}^+=(Z_{m-1,0}^+,\ldots,Z_{m-1,K}^+)$.
To randomize the $m$-th subject, a ball is drawn at random without replacement. The probability of selecting a ball of type $i$ is $Z_{m-1,i}^+/|\bm Z_{m-1}^+|$, $i=0,1,\ldots, K$. Here, $|\bm
Z_{m-1}^+|=\sum_{j=0}^K Z_{m-1,j}^+$. Therefore, the balls with a negative value in $\bm Z_{m-1}$ will have no chance of being selected.
If the drawn ball is of type $0$ (i.e, an immigration ball), no subject is treated and the ball is returned to the urn. Furthermore, $A_{m-1}=a_{m-1,1}+\ldots+a_{m-1,K}$ additional balls, $a_{m-1,k}(\ge 0)$ of treatment type $k$, $k=1,\ldots, K$ are added to the urn. If the immigration ball is selected $l$ times before a treatment ball is drawn, the urn composition $\bm Z_{m-1}$ is updated to $(Z_{m-1,0},Z_{m-1,1}+la_{m-1,1},\ldots,Z_{m-1,K}+la_{m-1,K})$ and the $\bm Z_{m-1}^+$ is updated to $(Z_{m-1,0},(Z_{m-1,1}+la_{m-1,1})^+,\ldots,(Z_{m-1,K}+la_{m-1,K})^+).$
If a treatment ball is drawn (say, of type $k$, for some $k=1,\ldots,K$), the $m$-th subject is given treatment $k$ and the outcome $\xi_{m,k}$ of this subject on treatment $k$ is observed. The ball is not replaced. Instead, $D_{m,kj}=D_{kj}(\xi_{m,k})$ balls of type $j$ will be added to the urn, $j=1,\ldots,K$. $D_{m,kj}<0$ signifies the removal of balls.
With the IMU, the number of immigration balls remains unchanged and a treatment ball is dropped when it is drawn. The number of treatment balls that is added to the urn depends on
1. the value of $a_{m,k}$ when an immigration ball is drawn from the urn, and
2. the value of $D_{m,kj}$ when a ball of treatment type $k$ is selected.
Here, $a_{m,k}$s represents the immigration rates and $D_{m,kj}$s represents the adding rules. When $a_{m,k}=0$, we obtain the GPU models if one chooses the adding rule $D_{m,kj}$ as in Section 4.1 in Hu and Rosenberger (2006). When $a_{m,k}=0$, one can obtain the ternary urn model by choosing the $D_{m,kj}$ as in the Section 4.4 in Hu and Rosenberger (2006). For two-treatment case, one may choose $a_{m,k}=1$, and suitable $D_{m,kj}$ (see Example 2) to obtain the drop-the-loser rule. In general, we can select suitable $a_{m,k}$ and $D_{m,kj}$ to obtain the desirable IMU model. Examples can be found in Section 4.
It is important to note that both $a_{m,k}$ and $D_{m,kj}$ depend on $m$. This allows both the immigration rates and the adding rules to depend on the responses of clinical trial. Consequently, it enables us to search for the desirable IMU model that can be used to target any given allocation proportion. Since both $a_{m,k}$ and $D_{m,kj}$ depend on $m$, it is impossible to use Atheya and Ney’s (1972) technique of embedding the urn process in a continuous-time branching process.
After the treatment assignment of the $n$-th subject, $N_{n,k}$ is used to denote the number of subjects who are being assigned to treatment $k$, $k=1,\ldots, K$. In clinical studies, the proportion $N_{n,k}/n$, $k=1,\ldots, K$ of patients being assigned to various treatments is useful information. The distributional properties and statistical behavior of $N_{n,k}/n$ are discussed in Section 3.
0.3in
Asymptotic Properties {#section3}
=====================
3.1. [*Notation and assumptions*]{}
0.1in Important asymptotic properties for the immigrated urn model are derived in this section. First, we need to introduce some basic notation and the required assumptions. Let $\theta_k$ be an unknown parameter from the distribution of the response $\xi_{m,k}$, and write $\bm \theta=(\theta_1,\ldots,\theta_K)$. Assume that the immigration rates $$\label{eqamk}
a_{m-1,k}=a_k(\widehat{\bm\theta}_{m-1}), \;\; k=1,\ldots, K,$$ where $a_k(\cdot)$s are continuous functions, and $\widehat{\bm\theta}_{m-1}=(\widehat{\theta}_{m-1,1},\ldots,\widehat{\theta}_{m-1,k})$ and $\widehat{\theta}_{m-1,k}$ are the current estimates of $\theta_k$, $k=1,\ldots, K$. The immigration process that is defined in (\[eqamk\]) has two functions. The first is to prevent the possibility of the extinction of a particular type of treatment ball. The second is to adjust the treatment allocation proportions according to the current estimates of $\bm \theta$.
Without loss of generality, we assume that the unknown parameter $\theta_k$ is the mean of the outcome $\xi_{m,k}$ and we take the sample mean as its estimate. Write $\bm
\xi_m=(\xi_{m,1},\ldots,\xi_{m,K})$. For the adding rules, let $\bm
D_m=(D_{m,kj}; k,j=1,\ldots,K)$, $\bm
D_m^{(k)}=(D_{m,k1},\ldots,D_{m,kK})$, $k=1,\ldots, K$, and $\bm
H=(h_{kj})=\ep \bm D_n$. Let $\widehat{\theta}_{m-1,k}$ be the sample mean of the outcomes $$\label{eqestimate} \widehat{\theta}_{m-1,k}=\frac{\alpha+S_{m-1,k}}{\beta+
N_{m-1,k} },$$ where $S_{m-1,k}$ is the sum of the outcomes on treatment $k$ of the previous $m-1$ subjects. Here, $\alpha,\beta>0$ are used to avoid the nonsense case of $0/0$. In general, many estimators, for example, the MLE, can be written in the form of (\[eqestimate\]) with $S_{m-1,k}$ being replaced by a sum of functions of the treatment outcomes plus a negligible remainder.
\[Asp0\] Functions $a_k(\cdot)>0$ are continuous and twice differentiable at $\bm\theta$.
\[Asp2\] $\{(\xi_{m,k}, D_{m,k1},\ldots, D_{m,kK});m\ge 1\}$, $k=1,\ldots,
K$, are $K$ sequences of i.i.d. random variables with $\sup_m\ep|D_{m,kj}|^{2+\delta}<\infty$, and $\sup_m\ep|\xi_{m,k}|^{2+\delta}<\infty$ for some $0<\delta\le 2$, $k=1,\ldots, K$. Further, assume that $D_{m,kk}\ge -C$ for some $C$, $k=1,\ldots, K$, and also, either $h_{kj}> 0$ or $D_{m,kj}\ge 0$ for $k\ne j$.
\[Asp3\] $\bm H\bm 1^{\prime}<\bm 1^{\prime}$, i.e., $\sum_{j=1}^Kh_{kj}<1 $ for all $k=1,\ldots, K$. Here, $\bm
1=(1,\ldots,1)$.
Assumption \[Asp3\] means that the average number of added balls at each step according to the outcome of a treatment is less than the number (one) of dropped balls. Thus, the urn is updated mainly by immigration. When the number of added balls exceeds one, the total number of balls in the urn will monotonically increase to infinity, and as a result, the probability of drawing an immigration ball will monotonically decrease to zero. This model is equivalent to the generalized Pólya urn model without immigration, which has been well studied (c.f., Hu and Rosenberger, 2006).
For the immigration rates, write $a_k=a_k(\bm \theta)$. Let $\bm
a=(a_1,\ldots, a_K)$, $\bm u=\bm a(\bm I-\bm H)^{-1}$, $s=\bm a(\bm
I-\bm H)^{-1}\bm 1^{\prime}=\sum_{k=1}^Ku_k$ and $\bm v=\bm u/s$. Further, denote $\bm \Sigma_k=\Var\{\bm D_1^{(k)}\}$, $\bm
\Sigma_{11}=\sum_{k=1}^K v_k \bm \Sigma_k$, $\bm
\Sigma_{12}=(\Cov\{D_{1,kj},\xi_k\}; j,k=1,\ldots, K)$, $\bm
\Sigma_{22}=diag(\Var\{\xi_{1,1},\ldots, \Var\{\xi_{1,K}\})$, and $$\label{eqLambda}
\bm \Lambda=
\begin{pmatrix} \bm \Lambda_{11} & \bm \Lambda_{12} \\
\bm \Lambda_{12}^{\prime} & \bm \Lambda_{22} \end{pmatrix}
=\begin{pmatrix} \bm \Sigma_{11} & \bm \Sigma_{12}\; diag(\bm v) \\
diag(\bm v)\; \bm \Sigma_{12}^{\prime} & \bm \Sigma_{22}\; diag(\bm
v)\end{pmatrix}.$$
Notice that $h_{ij}\ge 0$ for $i\ne j$. The existence of $(\bm I-\bm H)^{-1}$ is implied by Assumption \[Asp3\]. Also, this assumption can be replaced by a more general assumption in which there is a vector $\bm e= (e_1,\ldots,e_K)$ such that $\bm
H\bm e^{\prime}<\bm e^{\prime}$ and $e_i>0$, $i=1,\ldots, K$.
0.1in 3.2. [*Main asymptotic results*]{}
\[th3\] Suppose Assumptions \[Asp0\]-\[Asp3\] are satisfied. Denote $\bm N_n=(N_{n,1},\ldots,N_{n,k})$ and $\bm v=\bm v(\bm\theta)$. One can define a $2K$-dimensional Wiener processes $(\bm W(t), \bm B(t))$ such that $$\label{eqth3.1}
\Var\{(\bm W(t),
\bm B(t))\}=t\bm \Lambda$$ and $$\begin{aligned}
\label{eqth3.2}
\bm N_n-n\bm v=\bm W(n)\bm A
+\int_0^{n}\frac{\bm B(x)}{x}dx\;diag\Big(\frac{1}{\bm v} \Big)
\frac{\partial \bm v(\bm\theta)}{\partial\bm\theta}+o(n^{1/2-\epsilon})\;\; a.s.\end{aligned}$$ for some $\epsilon>0$, where $\bm A=(\bm I-\bm H)^{-1}(\bm I-\bm 1^{\prime}\bm v)$, $$\bm v(\bm\theta)=\frac{a(\bm \theta)(\bm I-\bm H)^{-1}}{a(\bm \theta)(\bm I-\bm H)^{-1}\bm 1^{\prime}}
\;\; \text{ and } \;\;\frac{\partial \bm
v(\bm\theta)}{\partial\bm\theta}=\left(\frac{\partial
v_k(\bm\theta)}{\partial \theta_j};j,k=1,\ldots,K\right).$$ Here, $1/\bm v=(1/v_1,\ldots, 1/v_K)$.
The proof of Theorem \[th3\] is given in the Appendix. Using Equation (\[eqth3.2\]), the consistency property and asymptotic normality can be derived.
\[normal\]Under the Assumptions in Theorem \[th3\], $$\label{eqnormal1.1}
\bm N_n - n\bm v =O(\sqrt{n\log\log n})\; a.s. \; \text{ and } \;
\sqrt{n}\big(\frac{\bm N_n}{n}-\bm v\big)\overset{\mathscr{D}}\to N(\bm
0,\bm\Sigma),$$ where $\bm\Sigma=\bm\Sigma_{D}+2\bm\Sigma_{\xi}+\bm\Sigma_{D\xi}+\bm\Sigma_{D\xi}^{\prime}$, and $$\begin{aligned}
&\bm\Sigma_D=\bm A^{\prime}\bm\Sigma_{11}
\bm A,\quad \bm\Sigma_{D\xi}=\bm A^{\prime} \bm\Sigma_{12} \frac{\partial \bm v(\bm\theta)}{\partial\bm\theta}\\
&\bm\Sigma_{\xi}=\Big(\frac{\partial \bm v(\bm\theta
)}{\partial\bm\theta}\Big)^{\prime} diag\Big(\frac{\sigma_{\xi
1}^2}{v_1},\ldots,\frac{\sigma_{\xi K}^2}{ v_K}\Big)\frac{\partial
\bm v(\bm\theta)}{\partial\bm\theta}.\end{aligned}$$ In particular, if $\bm D_m \equiv const$, then $$\sqrt{n}\big(\frac{\bm N_n}{n}-\bm v\big)\overset{\mathscr D}\to N(\bm 0,2\bm\Sigma_{\xi});$$ if $a_{m,k}\equiv a_k$, $k=1,\ldots,K$, do not depend on the estimates, then $$\sqrt{n}\big(\frac{\bm N_n}{n}-\bm v\big)\overset{\mathscr D}\to N(\bm 0,\bm\Sigma_{D}).$$
[**Proof**]{}. $~$ Notice $(\bm W(n),\int_0^n\frac{\bm B(x)}{x}dx)$ is a centered Gaussian vector with $$\bm W(n)=O(\sqrt{n\log\log n})\;\; a.s.,$$ $$\int_0^n\frac{\bm B(x)}{x}dx=O(1)+
\int_e^n\frac{O(\sqrt{x\log\log x})}{x}dx=O(\sqrt{n\log\log n})\;\; a.s.,$$ $$\Var\{\bm W(n)\}=n \bm\Sigma_{11},$$ $$\begin{aligned}
\Var\left\{\int_0^n\frac{\bm B(x)}{x}dx\right\}=
\bm\Sigma_{22} diag(\bm v)\int_0^n\int_0^n\frac{x\wedge y}{xy}dxdy =2n
\;\bm\Sigma_{22} diag(\bm v)\end{aligned}$$ and $$\Cov\left\{\bm W(n),\int_0^n\frac{\bm B(x)}{x}dx\right\}=
\bm\Sigma_{12}diag(\bm v)\int_0^n\frac{x\wedge n}{x}dx=n\;\bm\Sigma_{12}diag(\bm v).$$ (\[eqnormal1.1\]) follows from (\[eqth3.2\]) immediately. $\Box$
In practice, the outcomes in clinical trials are not available immediately prior to the treatment allocation of the next subject. The parameters can be estimated and the urn can be updated only according to the observed responses. In the delayed response case, we let $\mu_k(m,l)$ be the probability that the outcome of the $m$-th subject on treatment $k$ occurs after at least another $l$ subjects arrive. If $\mu_k(m,l)\le C l^{-\gamma}$ for some $\gamma\ge 2$, we can show that the total sum of unobserved outcomes up to the $n$-th assignment is a high order of $\sqrt{n}$ and so the conclusion in Theorem \[th3\] remains true. It has been shown that the delay machine does not effect the asymptotic properties for many response-adaptive designs if the delay decays with a power rate (c.f., Bai, Hu, and Rosenberger, 2002; Hu and Zhang, 2004; Zhang et al., 2007).
Next, we consider a special case that includes the generalized drop-the-loser rule of Zhang et al. (2007) in which $\bm D_m$ is a diagonal matrix.
\[gdl\]Suppose Assumptions \[Asp0\]-\[Asp2\] are satisfied, $D_{m,kj}=0$ for $j\ne k$, and $h_k=1-\ep D_{1,kk}>0$. Write $\bm h=(h_1,\ldots, h_K)$, $$v_k(\bm\theta,\bm h)=\frac{a_k(\bm\theta)/h_k}{\sum_{j=1}^Ka_j(\bm\theta)/h_j}\;\;
k=1,\ldots, K,$$ $\bm v=\bm v(\bm\theta,\bm h)=(v_1(\bm\theta,\bm h),\ldots,v_K(\bm\theta,\bm h))$, and $$\begin{aligned}
\frac{\partial
\bm v(\bm\theta,\bm h)}{\partial\bm\theta} =&\left(\frac{\partial
v_k(\bm\theta,\bm h)}{\partial \theta_j};j,k=1,\ldots,K\right),
\quad \\
\frac{\partial \bm v(\bm\theta,\bm h)}{\partial\bm h}
=&\left(\frac{\partial v_k(\bm\theta,\bm h)}{\partial
h_j};j,k=1,\ldots,K\right).\end{aligned}$$ Then, $$\label{eqCor1.2}
\frac{\bm N_n}{n}\to \bm v \quad a.s. \; \text{ and } \;
\sqrt{n}\big(\frac{\bm N_n}{n}-\bm v\big)\overset{\mathscr D}\to N(\bm
0,\bm\Sigma),$$ where $\bm\Sigma=\bm\Sigma_{D}+2\bm\Sigma_{\xi}+\bm\Sigma_{D\xi}+\bm\Sigma_{D\xi}^{\prime}$, $$\begin{aligned}
&\bm\Sigma_D=\Big(\frac{\partial \bm v(\bm\theta,\bm h)}{\partial\bm h}\Big)^{\prime} diag\Big(\frac{\sigma_{D
1}^2}{v_1},\ldots,\frac{\sigma_{D K}^2}{ v_K}\Big)\frac{\partial
\bm v(\bm\theta,\bm h)}{\partial\bm h},\\
&\bm\Sigma_{\xi}=\Big(\frac{\partial \bm v(\bm\theta,\bm
h)}{\partial\bm\theta}\Big)^{\prime} diag\Big(\frac{\sigma_{\xi
1}^2}{v_1},\ldots,\frac{\sigma_{\xi K}^2}{ v_K}\Big)\frac{\partial
\bm v(\bm\theta,\bm h)}{\partial\bm\theta},
\\
&\bm\Sigma_{D\xi}=-\Big(\frac{\partial \bm v(\bm\theta,\bm
h)}{\partial\bm h}\Big)^{\prime} diag\Big(\frac{\sigma_{D\xi
1}^2}{v_1},\ldots,\frac{\sigma_{D\xi K}^2}{ v_K}\Big)\frac{\partial
\bm v(\bm\theta,\bm h)}{\partial\bm\theta},\end{aligned}$$ and $\sigma_{Dk}^2=\Var\{D_{1,kk}\}$, $\sigma_{\xi k}^2=\Var\{\xi_{1,k}\}$, $\sigma_{\xi Dk}^2=
\Cov\{D_{1,kk}, \xi_{k,1}\}$, $k=1,2,\ldots,K$.
[**Proof.**]{} It is easy to check that $\bm \Sigma_{11}=diag(\sigma_{D1}^2v_1,\ldots,\sigma_{DK}^2v_K)$, $ \bm \Sigma_{12}=
diag(\sigma_{\xi D1}^2,\ldots,\sigma_{\xi DK}^2)$, $ \bm \Sigma_{22}=
diag(\sigma_{\xi 1}^2,\ldots,\sigma_{\xi K}^2)$, $\bm A=diag(1/\bm h)(\bm I-\bm 1^{\prime}\bm v)$, and $ \partial \bm v(\bm\theta,\bm h)/ \partial\bm h=-diag(\bm v)\bm A$. The results follow from Corollary \[normal\]. $\Box$
If Assumption \[Asp3\] is not satisfied and $\bm H\bm
1^{\prime}=\bm 1^{\prime}$, the following theorem can be used to yield the consistency property of the allocation proportion.
\[th4\] Suppose Assumptions \[Asp0\] and \[Asp2\] are satisfied, and $\bm H\bm 1^{\prime}=\bm 1^{\prime}$. Suppose further, $1$ is a single eigenvalue of $\bm H$. Then $$\bm N_n-n\bm v=O(\sqrt{n\log\log n}) \;\; a.s.\;\; \text{ and } \;\; \bm N_n-n\bm v=O_P(\sqrt{n}),$$ where $\bm v$ is the left eigenvalue vector of $\bm H$ that corresponds to the eigenvalue $1$.
The asymptotic normality of $\bm N_n$ is still unknown when $\bm
H\bm 1^{\prime}=\bm 1^{\prime}$. The proof of the above theorem is given in the Appendix.
Hu, Rosenberger and Zhang (2006) studied the variability of a randomization procedure that targets any given allocation proportion. They obtained a lower bound of the variability. For the IMU model in this paper, we have following results. When the variance of IMU model attains the lower bound, we can use the Crámer-Rao formula to compute the variance. In general, we have
\[theorem4.1\] If each $D_{m,kj}$ is a linear function of a random $\eta_{m,k}$, $j=1,\ldots, K$. Then, we have $$\begin{aligned}
\bm \Sigma_D=&\left(\frac{\partial \bm v}{\partial \bm d}\right)^{\prime}
diag\left(\frac{\Var\{\eta_{1,1}\}}{v_1},\ldots,\frac{\Var\{\eta_{1,K}\}}{v_K}\right)
\frac{\partial \bm v}{\partial \bm d} \label{eqlowboundV1}\\
\bm \Sigma_{D\xi}=&\left(\frac{\partial \bm v}{\partial \bm
d}\right)^{\prime}
diag\left(\frac{\Cov\{\eta_{1,1},\xi_{1,1}\}}{v_1},\ldots,\frac{\Var\{\eta_{1,K},\xi_{1,K}\}}{v_K}\right)
\frac{\partial \bm v}{\partial \bm \theta}, \label{eqlowboundV2}\end{aligned}$$ where $\bm d=(d_1,\ldots,d_K)=(\ep\eta_{1,1},\ldots,\ep\eta_{1,K})$. Further, if $\bm a(\cdot)=const$ and $\Var\{\eta_{1,k}\}$ is the inverse of the Fisher information of $d_k$, then the asymptotic variance-covariance matrix of $\bm N_n/\sqrt{n}$ attains the following lower bound, $$\label{lowerb}
\left(\frac{\partial \bm v}{\partial \bm d}\right)^{\prime}
diag\left((v_1I_1)^{-1},\ldots,(v_K I_K)^{-1}\right)
\left(\frac{\partial \bm v}{\partial \bm d}\right),$$ where $I_k$ is the Fisher information function of parameter $d_k$.
[**Proof.**]{} In fact, if we write $\bm D_1^{(k)}=\bm
\alpha_k+\bm\beta_k\eta_{1,k}$ and $\bm K=\begin{pmatrix} \bm
\beta_1 \\ \cdots \\\bm \beta_K\end{pmatrix}$, then $$\begin{aligned}
\bm\Lambda_{11}=&\sum_{k=1}^K v_k\Var\{\eta_{1,k}\} \bm\beta_k^{\prime}\bm\beta_k \\
=&(diag(\bm v)\bm K)^{\prime}
diag\left(\frac{\Var\{\eta_{1,1}\}}{v_1},\ldots,\frac{\Var\{\eta_{1,K}\}}{v_K}\right)diag(\bm
v)\bm K\end{aligned}$$ and $$\begin{aligned}
\bm\Sigma_{12}= (diag(\bm v)\bm K)^{\prime}
diag\left(\frac{\Cov\{\eta_{1,1},\xi_{1,1}\}}{v_1},\ldots,\frac{\Cov\{\eta_{1,K},\xi_{1,K}\}}{v_K}\right).\end{aligned}$$ However, $\partial \bm H/\partial d_k=diag(\bm 1_k)\bm K$, where $\bm 1_k$ has zero elements except the $k$-th one which is $1$. Also, $$\frac{\partial (\bm I-\bm H)^{-1}}
{\partial d_k}=(\bm I-\bm H)^{-1}\frac{\partial \bm H}{\partial
d_k}(\bm I-\bm H)^{-1} =(\bm I-\bm H)^{-1} diag(\bm 1_k)\bm K (\bm
I-\bm H)^{-1}.$$ It follows that $$\begin{aligned}
\frac{\partial\bm v}{\partial d_k}=&\frac{\partial \bm a(\bm I-\bm
H)^{-1}/\partial d_k}{\bm a(\bm I-\bm H)^{-1}\bm 1^{\prime}}
-\frac{\partial\bm a (\bm I-\bm H)^{-1}/\partial d_k}{(\bm a(\bm
I-\bm H)^{-1}\bm 1^{\prime})^2}
\bm 1^{\prime} \bm a(\bm I-\bm H)^{-1} \\
= & \bm v diag(\bm 1_k)\bm K(\bk I-\bm H)^{-1}(\bm I-\bm
1^{\prime}\bm v)=\bm v diag(\bm 1_k)\bm K\bm A,\end{aligned}$$ i.e., $\partial \bm v/\partial \bm d=diag(\bm v)\bm K\bm A$. (\[eqlowboundV1\]) and (\[eqlowboundV2\]) are proved. $\Box$
In Theorem \[theorem4.1\], we assume that the parameter $d_k$ is a one-dimensional parameter that corresponds to treatment $k$ for simplicity of notation. The results can be rewritten to a vector parameter $d_k$ without any further assumptions.
Applications
============
In this section, we apply the general asymptotic results in Section 3 to selected IMU models for illustrative purposes.
0.15in [*Example 1*]{}: The birth and death urn (BDU)
0.1in Ivanova et al. (2000) proposed a birth and death urn (BDU) model to allocate treatments in clinical trials. In the BDU model, if a ball of treatment $k$ is drawn and a subject is assigned to this treatment, the ball is not replaced when the treatment is a failure. When the treatment is a success, the ball is replaced and one additional ball of the same type is added to the urn. If an immigration ball is drawn, the ball is replaced and in addition, one ball of each treatment type is added. The BDU is a special case of the IMU with $a_{m,k}\equiv 1$, $D_{m,kj}=0$ for $j\ne k$, and
1. $D_{m,kk}=2$ if the treatment outcome of the $m$-th subject on treatment $k$ is a success, and
2. $D_{m,kk}=0$ if the treatment outcome of the $m$-th subject on treatment $k$ is a failure.
For this example, the asymptotic variance attains the lower bound. It is easily seen that $\bm H=diag(2p_1,\ldots, 2p_K)$, where $p_k$ ($q_k$) is the success (failure) probability for treatment $k$, $k=1,\ldots, K$. By Corollary \[normal\] and Theorem \[theorem4.1\], $$\frac{N_{n,k}}{n}\to v_k=\frac{1/(1-2p_k)}{\sum_{j=1}^K 1/(1-2p_j)}\; a.s.,\;\;
\sqrt{n}\left(\frac{\bm N_n}{n}-\bm v\right)\overset{\mathscr D}\to
N(\bm 0,\bm\Sigma)$$ if $p_k<1/2$, $k=1,\ldots, K$, where $\bm
v=(v_1,\ldots, v_K)$, $\bm p=(p_1,\ldots,p_K)$ and $$\bm \Sigma=\bm \Sigma_D=\left(\frac{\partial \bm v}{\partial \bm p}\right)^{\prime}
diag\left(\frac{p_1q_1}{v_1},\ldots,\frac{p_Kq_K}{v_K}\right)\frac{\partial
\bm v}{\partial \bm p}.$$
0.2in [*Example 2*]{}: Drop-the-loser (DL) rule and Modified DL (MDL) rule
0.1in The urn process of the DL rule is basically the same as that of the BDU. The major difference is that when the treatment is a success, the treatment ball is replaced without the addition of extra treatment balls. The DL rule is also a member of the IMU with $a_{m,k}\equiv 1$, $D_{m,kj}=0$ for $j\ne k$, and
1. $D_{m,kk}=1$ if the treatment outcome of the $m$-th subject on treatment $k$ is a success, and
2. $D_{m,kk}=0$ if the treatment outcome of the $m$-th subject on treatment $k$ is a failure.
A modified version of the DL rule (MDL) is given as follows. When a treatment ball is drawn, this ball is replaced only when the treatment outcome is a success. However, when an immigration ball is drawn, instead of adding an equal number of treatment balls to the urn, we add $C\widehat{p}_k$ balls of type $k$, $k=1,\ldots,K$, where $\widehat{p}_k$ is the current estimate of the successful probability $p_k$ of treatment $k$, and $C$ is a constant. With this model, more balls are immigrated to treatments with higher success rates, and subsequently, the limit proportions will be higher for better treatments. In fact, as shown below, the limit proportions are proportional to the odds $p_k/q_k$. For the DL rule, the limit proportions are proportional to $1/q_k$ (Ivanova, 2003).
This design can be compared to the generalized Pólya urn model that was proposed by Bai, Hu, and Shen (2002). For the latter, the asymptotic normality was obtained by Zhang, Hu, and Cheung (2006). Unlike the generalized Pólya urn models without immigration in which the asymptotic normality holds only when a very strict condition on eigenvalues of a generating matrix is satisfied (c.f., Bai and Hu, 2005; Janson, 2004; Zhang, Hu, and Cheung, 2006), the MDL rule and the IMU in general, as we will show, allow asymptotic normality under very mild conditions.
Regarding the asymptotic variance, it is easily seen that $\bm
a=(p_1C,\ldots,p_KC)$ and $\bm H=diag(p_1,\ldots,p_K)$. The conditions in Corollary \[gdl\] are satisfied for all cases with $0<p_k<1$ and $k=1,\ldots, K$. Therefore, by Corollary \[gdl\], the limit proportions are $$v_k=\frac{p_k/q_k}{\sum_{j=1}^Kp_j/q_j}, \;\; k=1,\ldots, K.$$ The asymptotic variance-covariance can be derived by the formulae in Corollary \[gdl\] in which $\bm\theta=(p_1,\ldots,p_K)$, $\bm
h=(q_1,\ldots,q_K)$, and $\sigma_{Dk}^2=\sigma_{\xi
k}^2=\sigma_{D\xi k}^2=p_kq_k$, $k=1,\ldots, K$. For the two-treatment case, $$\frac{N_{n,1}}{n}\to v_1=\frac{p_1/q_1}{p_1/q_1+p_2/q_2}\; a.s.\;\;\text{ and }\;
\; \sqrt{n}( N_{n,1}/n- v_1)\overset{\mathscr D}\to N(0,\sigma^2),$$ where $
\sigma^2=q_1q_2[p_1^2(1+q_2^2)+p_2^2(1+q_1^2)]/(p_2q_1+p_1q_2)^3. $ However, the lower bound of the asymptotic variance is $q_1q_2(p_1^2 +p_2^2 )/(p_2q_1+p_1q_2)^3$.
0.2in [*Example 3*]{}: Optimal proportions for two treatments with dichotomous outcomes.
0.1in To compare two treatments with dichotomous outcomes, success or failure, Rosneberger et al. (2001) suggested an optimal proportion $v_1=\sqrt{p_1}/(\sqrt{p_1}+\sqrt{p_2})$, where $p_1$ ($p_2$) is the successful probability of treatment $1$ ($2$). To target this proportion, Zhang et al. (2007) proposed a generalized drop-the-loser urn, a special case of IMU with $a_{m,k}=C \sqrt{p_k}$, $k=1,2$, and $\bm D_m\equiv 0$. By Corollary \[normal\] or Corollary \[gdl\], the asymptotic variance of the (normalized) sample proportion is $$\frac{1}{2(\sqrt{p_1}+\sqrt{p_2})^3}\left(\frac{p_2q_1}{\sqrt{p_1}}+\frac{p_1q_2}{\sqrt{p_2}}\right),$$ which is twice the lower bound. Zhang et al. (2006) propose to use the GPU without immigration to target this proportion (c.f., their Example 2). The corresponding asymptotic variance is $$\frac{\sqrt{p_1p_2}}{(\sqrt{p_1}+\sqrt{p_2})^2}+\frac{3}{2(\sqrt{p_1}+\sqrt{p_2})^3}\left(\frac{p_2q_1}
{\sqrt{p_1}}+\frac{p_1q_2}{\sqrt{p_2}}\right),$$ which is much larger.
0.2in
Conclusions {#section5}
===========
In this paper, we have generated a general class of urn models that incorporates immigration. As illustrated in the examples, the IMU class comprises of many useful urn models. For explicatory purposes, we focus our discussion of these applications to the clinical environment, but these urn models can be applied to other areas for which the setting is deemed appropriate.
Important asymptotic properties are derived in this paper for the IMU class. With the asymptotic normality formula, the models can be evaluated in terms of the distributions of treatment allocation proportions. Under very mild conditions, the IMU models always yield relatively small asymptotic variances. In many cases, the asymptotic variance attains the lower bound. Thus, the IMU model usually has smaller variability than the corresponding urn model.
When the IMU models are applied in clinical trials, the responses may not be available immediately and the population could be heterogenous. There is no logistical difficulty incorporating delayed responses into the IMU models. One can update the urn when responses become available. A moderate delay in response (see Hu and Zhang, 2004) will not affect the asymptotic properties of the IMU. One can modify the proof in the Appendix to incorporate delayed responses. It can also be shown that the IMU model can accommodate heterogeneity under the conditions that are discussed in Bai and Hu (2005).
In this article, we have focused on the asymptotic properties of the IMU itself. It is important to assess the validity of the statistical inferences after using the IMU models. By applying the results of Hu, Rosenberger, and Zhang (2006), Theorem 3.1 and Corollary 3.1 ensure that the estimators have the same asymptotic properties as they would if a fixed design were used for the same allocation proportion. This provides a solid foundation for both the IMU model and its related statistical inferences. The applications of IMU models have been discussed mainly in the clinical context, but the IMU could play an important role in other areas because of its desirable properties. Clearly, the application of the IMU in other areas requires further research effort.
0.3in
Appendix. Proofs {#appendix.-proofs .unnumbered}
================
We prove Theorem \[th3\] first. Recall that $\bm Z_{m-1}=(Z_{m-1,0}$, $Z_{m-1,1},$ $\ldots, Z_{m-1,K})$ are the numbers of balls when the $m$-th subject arrives to be randomized, $\bm Z_{m-1}^+=(Z_{m-1,0}^+,Z_{m-1,1}^+,$ $\ldots,$ $Z_{m-1,K}^+)$ are the non-negative numbers, and $|\bm
Z_{m-1}^+|=Z_{m-1,0}^+ + Z_{m-1,1}^+ +\ldots+ Z_{m-1,K}^+$. Write $\widetilde{\bm Z}_{m-1}=(Z_{m-1,1},\ldots,Z_{m-1,K})$. Because every immigration ball is replaced, $Z_{m-1,0}^+=Z_{m-1,0}=Z_{0,0}$ for all $m$. Let $\bm X_m$ be the result of the $m$-th assignment, where $X_{m,k}=1$ if the $m$-th subject is assigned to treatment $k$ and $0$ otherwise, $k=1,\ldots, K$. Then, $\bm N_n=(N_{n,1},\ldots,
N_{n,K})=\sum_{m=1}^n \bm X_m$. Further, we denote $\bm
a_m=(a_{m,1},\ldots,a_{m,K})$, and $u_m$ to be the number of draws of type $0$ balls between the $(m-1)$-th assignment and the $m$-th assignment.
Notice that between the $(m-1)$-th assignment and the $m$-th assignment, we have drawn $u_m$ balls of type $0$. Accordingly, we have added $a_{m-1,k}u_m$ balls of type $k$ to the urn. However, when a ball of type $k$ is drawn, this ball is not replaced and another $D_{m,kj}$ balls of type $j$ are added to the urn. So, the change in the number of balls after the $m$-th assignment is $$\label{eqPF1.3}
\widetilde{\bm Z}_m-\widetilde{\bm Z}_{m-1}=\bm a_{m-1} u_m +\bm X_m (\bm D_m-\bm I).$$ It follows that $$\begin{aligned}
\label{eqPF1.2}
&\widetilde{\bm Z}_n-\widetilde{\bm Z}_0=\sum_{m=1}^n\bm a_{m-1} u_m +\sum_{m=1}^n\bm X_m
(\bm D_m-\bm I)\nonumber\\
=&\sum_{m=1}^n\bm a_{m-1} u_m -\bm N_n(\bm I-\bm H)+
\sum_{m=1}^n\bm X_m
(\bm D_m-\ep[\bm D_m])\nonumber\\
=&\bm a N_{n,0}+\sum_{m=1}^n(\bm a_{m-1}-\bm a) u_m
-\bm N_n(\bm I -\bm H)+ \bm M_n,\end{aligned}$$ where $N_{n,0}=\sum_{m=1}^nu_m$ is total number of draws of type $0$ balls after the $n$-th assignment, and $\bm M_n =\sum_{m=1}^n \bm X_m (\bm D_m- \ep[\bm D_m])$ is a martingale.
To prove Theorem \[th3\] we need two lemmas the proofs of which are stated later.
\[lem2\] Suppose Assumptions \[Asp0\]-\[Asp3\] are satisfied. Then, $$\label{eqlem2.4}
Z_{n,k}=o(n^{1/2-\delta_0}) \quad a.s., \quad k=1,\ldots, K$$ for $0<\delta_0<\frac{1}{2}-\frac{1}{2+\delta}$.
\[lem3\] Suppose Assumptions \[Asp0\]-\[Asp3\] are satisfied. Then, $$\label{eqLILNast0}
N_{n,0}=n/s+O(\sqrt{n\log\log n})\;\; a.s.,$$ $$\label{eqLILNastk}
N_{n,k}=nv_k+O(\sqrt{n\log\log n}) \;\; a.s.,\;\;
k=1,\ldots, K,$$ where $s=\bm a(\bm I-\bm H)^{-1}\bm 1^{\prime}$. Also, for each $k=1,\ldots,K$, $$\label{eqlem3.1}\widehat{\theta}_{n,k}\to \theta_k
\quad a.s.$$ and $$\label{eqapptheta}
\widehat{\theta}_{n,k}-\theta_k=
\frac{Q_{n,k}}{nv_k}+o(n^{-1/2-\delta_0})\quad a.s.,$$ where $Q_{n,k}=\sum_{m=1}^nX_{m,k}(\xi_{m,k}- \ep\xi_{m,k})$ is a martingale and $\bm Q_n=(Q_{n,1},\ldots, Q_{n,K})$.
Now we begin the proof of Theorem \[th3\]. Consider the $2K$-dimensional martingale $\{(\bm M_n, \bm Q_n),\Cal A_n; n\ge
1\}$, where $\Cal A_n=\sigma(\bm X_1,\ldots,\bm X_n, \bm
\xi_1,\ldots, \bm \xi_{n+1})$. According to (\[eqLILNastk\]) we have $$\begin{aligned}
&\sum_{i=1}^n\ep[ (\Delta \bm M_i)^{\prime}\Delta\bm M_i | \Cal A_{i-1}]
=\sum_{k=1}^K N_{n,k}\bm \Sigma_k \nonumber\\
& \qquad = n \bm \Sigma_{11} +O(\sqrt{n\log\log n})\quad a.s.,\label{eqvarianceofM1} \\
&\sum_{i=1}^n\ep[ ( \Delta \bm Q_i)^{\prime} \Delta \bm Q_i | \Cal A_{i-1}]
=\bm \Sigma_{22}diag(\bm N_n)\nonumber\\
&\qquad = n \bm \Lambda_{22}+O(\sqrt{n\log\log n})\quad a.s., \label{eqvarianceofM2}\\
&\sum_{i=1}^n\ep[ (\Delta \bm M_i)^{\prime} \Delta \bm Q_i | \Cal
A_{i-1}] = \bm \Sigma_{12} diag(\bm N_n)\nonumber\\
&\qquad = n \bm\Lambda_{12} +O(\sqrt{n\log\log n})\quad a.s.\label{eqvarianceofM3}\end{aligned}$$ By Corollary 1.1 of Zhang (2004), we can define the $2K$-two-dimensional Wiener processes $(\bm W(t), \bm B(t))$ with variance-covariance matrix $\bm\Lambda$ such that for some $\epsilon>0$, $$\label{eqappMQ}
\bm M_n=\bm W(n) +o(n^{1/2-\epsilon})\;\; a.s., \;\; \bm Q_n =\bm
B(n)+o(n^{1/2-\epsilon})\;\; a.s.$$ Without loss of generality, we assume $\epsilon\le \delta_0$, where $\delta_0$ is defined as it is in Lemma \[lem2\]. Next, we need to show that $(\bm W(t), \bm B(t))$ satisfies (\[eqth3.2\]). Combining (\[eqPF1.2\]) and (\[eqlem2.4\]) yields $$\begin{aligned}
\label{eqLILofN}
\bm N_n (\bm I-\bm H)-\bm a N_{n,0}
=\bm M_n+\sum_{m=1}^n(\bm a_m-\bm a)u_m+o(n^{1/2-\delta_0})\;\;
a.s.
\end{aligned}$$
Recall $\bm A=(\bm I-\bm H)^{-1}(\bm I-\bm 1^{\prime}\bm v)$, $\bm v=\bm a(\bm I-\bm H)^{-1}/(\bm a(\bm I-\bm H)^{-1}\bm 1^{\prime})$ and notice $\bm N_n\bm 1^{\prime}=n$. According to (\[eqLILofN\]), $$\label{eqproofth3.10}
\bm N_n-n\bm v=
\big(\bm M_n+\sum_{m=1}^n(\bm a_{m-1}-\bm a)u_m\big)\bm A+o(n^{1/2-\delta_0})\;\; a.s.$$
For $\bm a_m-\bm a$, due to (\[eqapptheta\]) and (\[eqappMQ\]), $$\begin{aligned}
\bm a_m-\bm a=&(\widehat{\bm\theta}_{m}-\bm\theta)\frac{\partial \bm a(\bm\theta)}{\partial\bm\theta}
+O\big(\|\widehat{\bm\theta}_{m}-\bm\theta\|^2\big)
\nonumber\\
=&
\frac{\bm Q_m}{m}diag\big(\frac{1}{\bm v} \big)\frac{\partial \bm a(\bm\theta)}{\partial\bm\theta}
+o(m^{-1/2-\delta_0}) \label{eqappofAbyQ}\\
=&
\frac{\bm B(m)}{m}diag\big(\frac{1}{\bm v}\big)\frac{\partial \bm
a(\bm\theta)}{\partial\bm\theta}+o(m^{-1/2-\epsilon}). \nonumber
\end{aligned}$$ Note that immigration happens only when a type $0$ ball is drawn. Let $\tau_m$ be the total number of draws when the $m$-th type $0$ ball is drawn. At this time, $\tau_m-m$ subjects are assigned and the $(\tau_m-m+1)$-th subject arrives to be randomized. So, we add $a_{(\tau_m-m+1)-1,k}$ balls of type $k$ to the urn, $k=1,\ldots,
K$. It follows that $$\sum_{j=1}^na_{j-1,k}\cdot u_j=\sum_{m=1}^{N_{n,0}}a_{\tau_m-m,k},$$ $$i.e., \quad \sum_{m=1}^n(\bm a_{m-1}-\bm a)u_m=\sum_{m=1}^{N_{n,0}}(\bm a_{\tau_m-m}-\bm a).$$ It is easily seen that $\tau_m=\min\{n:N_{n,0}\ge m\}+m$. Due to (\[eqLILNast0\]), $$\tau_m-m=\min\{n:N_{n,0}\ge m\}=sm+O(\sqrt{m\log\log m}) \;\; a.s.$$ It follows that $$\begin{aligned}
\bm a_{\tau_m-m}-\bm a=&\frac{\bm B(\tau_m-m)}{\tau_m-m}diag\big(\frac{1}{\bm v}
\big)\frac{\partial \bm a(\bm\theta)}{\partial\bm\theta}+o(m^{-1/2-\epsilon}) \\
=&\frac{\bm B(sm)}{sm}diag\big(\frac{1}{\bm v}\big)\frac{\partial \bm
a(\bm\theta)}{\partial\bm\theta}+o(m^{-1/2-\epsilon})\;\;a.s.
\end{aligned}$$ Noticing (\[eqLILNast0\]), we conclude that $$\begin{aligned}
\label{eqproofth3.12}
&\sum_{m=1}^n(\bm a_{m-1}-\bm a)u_m
=\sum_{m=1}^{N_{n,0}}\left(\frac{\bm
B(sm)}{sm}diag\big(\frac{1}{\bm v} \big)\frac{\partial \bm
a(\bm\theta)}{\partial\bm\theta}+o(m^{-1/2-\epsilon})\right)
\nonumber\\
&=\int_0^{n/s}\frac{\bm
B(sx)}{sx}dx\;diag\big(\frac{1}{\bm v} \big)\frac{\partial \bm
a(\bm\theta)}{\partial\bm\theta}+o(n^{1/2-\epsilon})
\nonumber\\
&=\int_0^{n}\frac{\bm
B(x)}{x}dx\;diag\big(\frac{1}{\bm v} \big)\frac{1}{s}\frac{\partial \bm
a(\bm\theta)}{\partial\bm\theta}+o(n^{1/2-\epsilon})\;\; a.s.\end{aligned}$$ However, it is easily checked that $$\label{eqproofth3.13} \frac{1}{s}\frac{\partial \bm
a(\bm\theta)}{\partial\bm\theta}\bm A = \frac{\partial \bm
v(\bm\theta)}{\partial\bm\theta}.$$ Combining (\[eqappMQ\])-(\[eqproofth3.13\]) finishes the proof of (\[eqth3.2\]). $\Box$
Now, we begin the proofs of Lemmas \[lem2\] and \[lem3\]. We need three more lemmas.
\[lem4\] Under Assumption \[Asp2\], we have $$Z_{n,k}^-=O(1)\quad a.s., \;\; k=1,\ldots,K.$$
[**Proof**]{}. $~$ Notice that the balls with a negative number have no chance to be drawn. If we let $S_n=\max\{1\le j\le n:
Z_{j,k}>0\}$, then according to (\[eqPF1.3\]), $$\begin{aligned}
Z_{n,k}\ge & Z_{n-1,k}+\sum_{j\ne k} X_{n,j}D_{n,kj}\ge \ldots\\
\ge & Z_{S_n,k}+X_{S_{n+1},k}(D_{S_n+1,kk}-1)+\sum_{j\ne k} \sum_{m=S_n+1}^n X_{m,j}D_{m,kj}\\
\ge &-C-1+ \sum_{j\ne k} \sum_{m=S_n+1}^n X_{m,j}D_{m,kj}.\end{aligned}$$ If $D_{m,kj}\ge 0$, then $\sum_{m=S_n+1}^n X_{m,j}D_{m,kj}\ge 0$. When $h_{kj}>0$, it is easily seen that $$\begin{aligned}
\sum_{m=S_n+1}^n X_{m,j}D_{m,kj}=&\sum_{m=S_n+1}^n X_{m,j}(D_{m,kj}-\ep D_{m,kj})+\sum_{m=S_n+1}^n X_{m,j}h_{ij} \\
&\to +\infty \;\; a.s.\end{aligned}$$ on $\{\sum_{m=S_n+1}^n X_{m,j}\to \infty\}$. The lemma is now proved. $\Box$
\[lem5\] Let $\Cal F_n=\sigma(\bm X_1,\ldots,\bm X_n,\bm Z_1,\ldots, \bm Z_n)$ be the history sigma field, and recall $A_m=\sum_{k=1}^K a_{m,k}$. Suppose Assumption \[Asp2\] is satisfied. Then, $\underline{A}:=\min_m A_m>0$ implies $$\label{eqlem5.1} \ep[u_n^p|\Cal F_{n-1}]\le c_p
\Big(\big(\sum_{k=1}^K Z_{n-1,k}\big)^-/\underline{A}\Big)^{p+1}\frac{Z_{0,0}}{|\bm
Z_{n-1}^+|}\;\; a.s., \;\; \forall\; p\ge 1,$$ where $c_p>0$ is a random variable that is a function of $Z_{0,0}$ and $\min_mA_m$. Particularly, $$\label{eqlem5.2}
\min_m A_m>0 \; \text{ implies } \;\ep[u_n^p|\Cal F_{n-1}]=O(1)\quad a.s.$$
[**Proof**]{}. Write $\underline{A}=\min_m A_m$. The event $\{u_n=l\}$ means that when the $n$-th subject is assigned, we have drawn $l+1$ balls continuously in which the first $l$ balls is of type $0$ and the last one is not. So, for $l=1,2,\ldots$, $$\begin{aligned}
\pr\big(u_n=l|\Cal F_{n-1}\big)
=\frac{Z_{0,0}}{|\bm
Z_{n-1}^+|}\prod_{j=1}^{l-1}\frac{Z_{0,0}}{|(\bm Z_{n-1}+\bm j\bm
a_{n-1})^+|}\Big(1-\frac{Z_{0,0}}{|(\bm Z_{n-1}+\bm
a_{n-1})^+|}\Big).\end{aligned}$$ Obviously, $
\pr\big(u_n=l|\Cal F_{n-1}\big) \le Z_{0,0}/|\bm
Z_{n-1}^+|$, $l\ge 1$. Notice $$|(\bm Z_{n-1}+\bm j\bm
a_{n-1})^+|=Z_{0,0}+\sum_{k=1}^K(Z_{n-1,k}+ja_{n-1,k})^+ \ge
Z_{0,0}+\sum_{k=1}^K Z_{n-1,k}+jA_{n-1}.$$ It follows that $\underline{A}>0$ and $\sum_{k=1}^K Z_{n-1,k}\ge
-L\underline{A}$ imply $$\begin{aligned}
\label{eqprooflem5.1}
\pr\big(u_n=l|\Cal F_{n-1}\big)\le \frac{Z_{0,0}}{|\bm
Z_{n-1}^+|}\prod_{j=L}^{l-1}\frac{Z_{0,0}}{Z_{0,0}+(j-L)\underline{A}}
\le c_0\frac{Z_{0,0}}{|\bm Z_{n-1}^+|}e^{-2(l-L)}, \;\; l\ge L,\end{aligned}$$ where $c_0>0$ depends only on $\underline{A}$ and $Z_{0,0}$. So $$\begin{aligned}
\ep[u_n^p|\Cal F_{n-1}]\le \sum_{l=1}^Ll^p \frac{Z_{0,0}}{|\bm
Z_{n-1}^+|}+\sum_{l=L+1}^{\infty}l^pc_0\frac{Z_{0,0}}{|\bm
Z_{n-1}^+|}e^{-2(l-L)}
\le c_p L^{p+1}\frac{Z_{0,0}}{|\bm
Z_{n-1}^+|}.\end{aligned}$$ Taking $L=\big[(\sum_{k=1}^K
Z_{n-1,k})^-/\underline{A}\big]+1$ completes the proof of (\[eqlem5.1\]). (\[eqlem5.2\]) follows from (\[eqlem5.1\]) and Lemma \[lem4\]. $\Box$
\[lem6\] Suppose Assumptions \[Asp0\]-\[Asp2\] are satisfied. Then $$\label{eqboundofa}
\min_{m,k} a_{m,k}>0 \;\; \text{ and } \max_{m,k}a_{m,k}<\infty
\;\; a.s.$$
[**Proof**]{}. $~$ It suffices to show that $$\label{eqprooftheta2}
N_{n,k}\to \infty \;\; \text{ implies } \widehat{\theta}_{n,k}\to
\theta_k\;\; a.s.,\;\; k=1,\ldots,K.$$ In fact, if (\[eqprooftheta2\]) is true, then $a_k(\bm y)>0$ for any $\bm y$ on $$\text{cl}\{\widehat{\bm\theta}_m;m=1,2,\ldots\}
=\bigotimes_{k=1}^K\{\theta_k,\widehat{\theta}_{m,k};m=1,2,\ldots\}.$$ By the continuity of $a_k(\cdot)$, $ 0<\min_{m,k}a_k(\widehat{\bm\theta}_m)\le a_{m,k}\le
\max_{m,k}a_k(\widehat{\bm\theta}_m)<\infty\;\; a.s. $ (\[eqboundofa\]) is satisfied. Now, notice $\{Q_{n,k},\Cal A_n\}$ is a martingale with $$\sum_{m=1}^n\ep[(\Delta Q_{m,k})^2|\Cal
A_{m-1}]=N_{n,k} \Var\{\xi_{1,k}\}\le C N_{n,k} .$$ By Theorem 3.3.10 of Stout (1974) (where $f(x)=x$), $$\label{eqprooflem3.29}
N_{n,k}\to \infty \;\; \text{ implies }\;\;
\frac{Q_{n,k}}{N_{n,k}}\to 0\;\; a.s.$$ By noticing $\widehat{\theta}_{n,k}=(\alpha+N_{n,k}\theta_k+Q_{n,k})/(N_{n,k}+\beta)$, (\[eqprooftheta2\]) is proved. $\Box$
[**Proof of Lemma \[lem2\]**]{}. By Lemma \[lem6\], $$\label{eqboundofAA}\underline{A}=:\min_m A_m>0\;\; \text{ and
}\;\;\overline{A}=:\max_m A_m<\infty$$ Notice $\widetilde{\bm Z}_n\bm 1^{\prime}=\sum_{k=1}^KZ_{n-1,k}$. By (\[eqlem5.1\]) and Lemma \[lem4\], $\ep[u_n|\Cal F_{n-1}]\le C_0
Z_{0,0}/|\bm Z_{n-1}^+|$. So, according to (\[eqPF1.3\]) or (\[eqPF1.2\]), we have $$\begin{aligned}
\label{eqprooflem2.2}
\widetilde{\bm Z}_n\bm 1^{\prime}=
&\widetilde{\bm Z}_{n-1}\bm 1^{\prime}+u_n A_{n-1}-\bm X_n(\bm I-\bm H)\bm 1^{\prime}+\Delta \bm M_n\bm 1^{\prime}
\nonumber\\
\le &\widetilde{\bm Z}_{n-1}\bm 1^{\prime}+A_{n-1}\ep[u_n|\Cal
F_{n-1}]-\underline{h}
+ A_{n-1}(u_n-\ep[u_n|\Cal F_{n-1}])+\Delta \bm M_n\bm 1^{\prime}\nonumber\\
\le & \widetilde{\bm Z}_{n-1}\bm 1^{\prime}+C_0\overline{A}\frac{Z_{0,0}}{|\bm
Z_{n-1}^+|}-\underline{h}
+\Delta U_n \nonumber\\
\le & \widetilde{\bm Z}_{n-1}\bm 1^{\prime}+\Delta U_n-\underline{h}\; /2, \; \text{
if } \widetilde{\bm Z}_{n-1}\bm 1^{\prime}\ge 2C_0\overline{A}
Z_{0,0}/\underline{h},\end{aligned}$$ where $\underline{h}=\min_k(1-\sum_{j=1}^Kh_{kj})>0$. Here, $U_n=\sum_{m=1}^nA_{m-1}(u_m-\ep[u_m|\Cal F_{n-1}])+ \bm M_n\bm
1^{\prime}$ is a real martingale. Let $S_n=\max\{1\le j\le n: \widetilde{\bm Z}_j\bm 1^{\prime}<2C_0\overline{A} Z_{0,0}/\underline{h}
\;\}$, where $\max(\emptyset)=0$. Then, according to (\[eqprooflem2.2\]), $$\begin{aligned}
\label{eqbeforelem2.24}
&\widetilde{\bm Z}_n\bm 1^{\prime}\le \widetilde{\bm Z}_{n-1}\bm 1^{\prime}+\Delta U_n -\underline{h}\;/2
\le \ldots \nonumber \\
\le& \widetilde{\bm Z}_{S_n}\bm 1^{\prime}+\Delta U_{S_n+1}+\ldots+\Delta
U_n-(n-S_n)\underline{h}\;/2
\nonumber \\
\le& |\bm Z_0|\vee \big(2C_0\overline{A} Z_{0,0}/\underline{h}\big) + U_n-U_{S_n}-(n-S_n)\underline{h}\;/2.\end{aligned}$$ For the martingale $\{U_n,\Cal F_n;n=1,2,\ldots\}$, we have $$\ep[(\Delta
U_n)^2|\Cal F_{n-1} ]\le C+C\max_jA_j^2=O(1) ,$$ $$\ep[|\Delta U_n|^{2+\delta}|\Cal F_{n-1}]\le
C+C\max_jA_j^{2+\delta}=O(1)$$ due to Assumption \[Asp2\] and (\[eqlem5.2\]). Accordingly, we can show that $$\label{eqLILofU}
U_n =O\big(\sqrt{n\log\log n}\big)\;\; a.s.$$ and $$\label{eqincofU}
\max_{m\le \sqrt{n\log n}}|U_{n-[\sqrt{n\log
n}]+m} -U_{n-[\sqrt{n\log
n}]} |=o(n^{\frac{1}{2+\delta}}\log n )\;\;a.s.$$ If $n-S_n\ge \sqrt{n\log n}$, then for $n$ large enough $$\begin{aligned}
U_n -U_{S_n} -(n-S_n)\underline{h}\;/2\le
O\big(\sqrt{n\log\log n}\big)- \underline{h}\sqrt{n\log n}/ 2<0\end{aligned}$$ due to (\[eqLILofU\]). Notice $n\ge S_n$. If $n-S_n<\sqrt{n\log
n}$, then $$\begin{aligned}
U_n -U_{S_n} -(n-S_n)\underline{h}\;/2
\le & 2\max_{m\le \sqrt{n\log
n}}|U_{n-[\sqrt{n\log n}]+m} -U_{n-[\sqrt{n\log
n}]} |\\
=&o(n^{\frac{1}{2+\delta}}\log n )\;\;a.s.\end{aligned}$$ by (\[eqincofU\]). It follows that $ \sum_{k=1}^K Z_{n,k}\le o(n^{1/2-\delta_0})$ a.s. due to (\[eqbeforelem2.24\]). However, $Z_{n,k}^-=O(1)$ a.s. by Lemma \[lem4\]. (\[eqlem2.4\]) is proved. $\Box$
. Recall $Q_{n,k}=\sum_{m=1}^n
X_{m,k}(\xi_{m,k}-\theta_k)$ $k=1,\ldots, K$, and both $\{
M_{n,k},\Cal A_n; n\ge 1\}$ and $\{ Q_{n,k},\Cal A_n; n\ge 1\}$ are martingales. According to the law of the iterated logarithm for martingales, we have $$\label{eqLILofM}
M_{n,k}=O(\sqrt{n\log\log n}) \;\;\text{ and }\;\; Q_{n,k}=O(\sqrt{n\log\log n})
\quad a.s.$$
However, for each $k=1,\ldots, K$, $$\label{eqprooftheta1}\widehat{\theta}_{n,k}-\theta_k=
\frac{Q_{n,k}+O(1)}{N_{n,k} +\beta}\;\;
a.s.$$
(\[eqLILofN\]) remains true by Lemmas \[lem2\]. By (\[eqLILofN\]) and (\[eqLILofM\]) we have $$\label{eqlowboundofN} \bm N_n(\bm I-\bm H)= \sum_{m=1}^n \bm a_{m-1}u_m +o(n)
\;\;a.s.$$ Notice $\bm H\bm 1^{\prime} < \bm
1^{\prime}$. It follows that $$n=\bm N_n\bm 1^{\prime}\le \sum_{m=1}^n A_{m-1}u_m/\underline{h} +o(n)\le
\big( K\overline{a}/\underline{h}\big)N_{n,0}+o(n), a.s.$$ where $\underline{h}=\min_k(1-\sum_{j=1}^Kh_{kj})>0$. That is, $$\label{eqlowboundofN0} \liminf_{n\to \infty}\frac{N_{n,0} }{n}\ge
\frac{\underline{h}}{K\overline{a}}>0 \quad a.s.$$ Notice $h_{kj}\ge 0$, $k\ne j$. Then, by (\[eqlowboundofN\]) and (\[eqlowboundofN0\]), $$(1-h_{kk})\liminf_{n\to \infty}\frac{N_{n,k} }{n}\ge \liminf_{n\to \infty}\frac{1}{n}
\sum_{m=1}^n a_{m,k}X_{m,0} \ge \underline{a}
\liminf_{n\to \infty}\frac{N_{n,0} }{n} >0\;\;a.s..$$ So, $\liminf_{n\to \infty} N_{n,k} / n>0$ a.s., which, together with (\[eqprooftheta1\]) and (\[eqLILofM\]), implies $$\widehat{\theta}_{n,k}-\theta_k=
O\Big(\frac{Q_{m,k}+O(1)}{n}\Big)=
O\Big(\sqrt{\frac{\log\log n}{n}}\Big)\to 0\;\; a.s.$$ (\[eqlem3.1\]) is proved and also $$\label{eqLILofak} a_{m,k}-a_k=a_k(\widehat{\bm\theta}_m)-a_k(\bm\theta)
=O(\|\widehat{\bm\theta}_m-\bm\theta\|)=O\Big(\sqrt{\frac{\log\log
m}{m}}\Big) \quad a.s.$$ Notice that $\sum_{m=1}^n(a_{m-1,k}-a_k)(u_m-\ep[u_m|\Cal F_{m-1}])$ is a martingale and $$\begin{aligned}
&\sum_{m=1}^{\infty}\ep\Big[\Big(\frac{(a_{m-1,k}-a_k)(u_m-\ep[u_m|\Cal F_{m-1}])}{\sqrt m}\Big)^2\Big|\Cal F_{m-1}\Big]\\
\le &\sum_{m=1}^{\infty}\frac{(a_{m-1,k}-a_k)^2}{m}\ep[u_m^2|\Cal F_{m-1}]\le C\sum_{m=1}^{\infty}
\frac{\log\log
m}{m^2}<\infty \;\; a.s.\end{aligned}$$ due to (\[eqlem5.2\]) and (\[eqLILofak\]). By the law of large numbers, it follows that $$\begin{aligned}
\label{eqLILofaku}
&\sum_{m=1}^n(a_{m-1,k}-a_k)u_m=\sum_{m=1}^n(a_{m-1,k}-a_k)\ep[u_m|\Cal F_{m-1}]+o(\sqrt{n}) \nonumber\\
=&\sum_{m=1}^n O\Big(\sqrt{\frac{\log\log
m}{m}}\Big)O(1)+o(\sqrt{n})=O(\sqrt{n\log\log n}) \;\; a.s.\end{aligned}$$ by (\[eqlem5.2\]) and (\[eqLILofak\]) again. Combining (\[eqLILofN\]), (\[eqLILofM\]), and (\[eqLILofaku\]) yields $$\bm N_n- N_{n,0}\bm a(\bm I-\bm H)^{-1} =O( \sqrt{n\log\log n}) \quad a.s.,$$ which, together with $\bm N_n \bm 1^{\prime}=n$, implies (\[eqLILNast0\]) and (\[eqLILNastk\]). Then combining (\[eqLILNastk\]), (\[eqLILofM\]), and (\[eqprooftheta1\]) yields $$\widehat{\theta}_{n,k}-\theta_k=
\frac{Q_{n,k}+O(1) }{nv_k
+O(\sqrt{n\log\log n})}=\frac{Q_{n,k}}{nv_k}
+o(n^{1/2-\delta_0}) \;\; a.s.$$ (\[eqapptheta\]) is proved, and the proof of Theorem \[th3\] is completed. $\Box$
. Suppose Assumptions \[Asp0\] and \[Asp2\] are satisfied. Notice $\bm H\bm 1^{\prime}=\bm
1^{\prime}$. Similar to (\[eqprooflem2.2\]), $$\begin{aligned}
\widetilde{\bm Z}_n\bm 1^{\prime}= &\widetilde{\bm Z}_{n-1}\bm
1^{\prime}+u_n A_{n-1}+\Delta \bm
M_n\bm 1^{\prime}
\le \widetilde{\bm Z}_{n-1}\bm 1^{\prime}+C_0\overline{A}\frac{Z_{0,0}}{|\bm
Z_{n-1}^+|}
+\Delta U_n \nonumber\\
\le & \widetilde{\bm Z}_{n-1}\bm 1^{\prime}+C_0\overline{A}/
\sqrt{n}+\Delta U_n, \; \text{
if } \widetilde{\bm Z}_{n-1}\bm 1^{\prime}\ge Z_{0,0}\sqrt{n}.\end{aligned}$$ It follows that $$\begin{aligned}
\widetilde{\bm Z}_n\bm 1^{\prime} \le& \widetilde{\bm Z}_{S_n}\bm
1^{\prime}+\Delta U_{S_n+1}+\ldots+\Delta
U_n+C_0\overline{A}(n-S_n)/\sqrt{n}
\nonumber \\
\le& 2C_0\overline{A}\sqrt{n} + U_n-U_{S_n}
\le 2C_0\overline{A}\sqrt{n} + 2\max_{m\le n}|U_m|,\end{aligned}$$ where $S_n=\max\{1\le j\le n:
\widetilde{\bm Z}_j\bm 1^{\prime}<
Z_{0,0}/\sqrt{n}
\;\}$ and $\max(\emptyset)=0$. Hence, $$\widetilde{\bm Z}_n= O(\sqrt{n\log\log n})\; a.s. \;\; \text{ and } \;\;
=O_P(\sqrt{n}),$$ by the properties of a martingale. So, by (\[eqPF1.2\]) and the law of the iterated logarithm of martingales, it follows that $$\begin{aligned}
\bm N_n (\bm I-\bm H)
=&\bm M_n+\sum_{m=1}^n\bm a_{m-1}u_m-\widetilde{\bm Z}_n+\widetilde{\bm Z}_0\\
=&\sum_{m=1}^n\bm a_{m-1}u_m+O(\sqrt{n\log\log n})\;\;
a.s.
\end{aligned}$$ Multiplying $\bm 1^{\prime}$ yields $\sum_{m=1}^nu_m A_{m-1}= O(\sqrt{n\log\log n})\;\;
a.s.$, and then $N_{n,0}= O(\sqrt{n\log\log n})$ a.s. and $\sum_{m=1}^n\bm a_{m-1}u_m = O(\sqrt{n\log\log n})$ a.s. by (\[eqboundofa\]). So, $$\begin{aligned}
(\bm N_n -n\bm v)(\bm I-(\bm H-\bm 1^{\prime}\bm v))=\bm N_n(\bm I-\bm H)
=O(\sqrt{n\log\log n})\;\;
a.s.
\end{aligned}$$ It follows that $\bm N_n-n\bm v=O(\sqrt{n\log\log n})\;\; a.s.$ because $(\bm I-(\bm H-\bm 1^{\prime}\bm v))$ is reversible. The proof of $\bm N_n-n\bm v=O_P(\sqrt{n})$ is similar. $\Box$
0.3in [**Acknowledgments**]{}
The first author’s research was supported by a grant from the National Natural Science Foundation of China (10771192). The second author’s research was supported by grant DMS-0349048 from the National Science Foundation (USA)
[**References**]{}
1. [Athreya, K. B.]{} and [Karlin, S.]{} (1968). Embedding of urn schemes into continuous time branching processes and related limit theorems. [*Ann. Math. Statist.*]{} [**39**]{} 1801–1817.
2. [Athreya, K. B.]{} and [Ney, P. E.]{} (1972). [*Branching Processes*]{}. Spring-Verlag, Berlin.
3. [Bai, Z. D.]{} and [Hu, F.]{} (2005). Strong consistency and asymptotic normality for urn models. [*Ann. Appl. Probab.*]{} [**15**]{} 914–940.
4. [Bai, Z. D., Hu, F.]{} and [Shen, L.]{} (2002). An adaptive design for multi-arm clinical trials. [*J. Multi. Anal.*]{} [**81**]{} 1–18.
5. [Bai, Z. D., Hu, F.]{} and [Rosenberger, W. F. ]{} (2002). Asymptotic properties of adaptive designs for clinical trials with delayed response. [*The Annals of Statistics*]{} [**30**]{} 122–139.
6. [Beggs]{}, A. W. (2005). On the convergence of reinforcement learning. [*Journal of Economic Theory*]{} [**122**]{} 1–36.
7. [Benaïm]{} M., [Schreiber]{} S. J. and [Tarrés]{}, P. (2004). Generalized urn models of evolutionary processes. [*Ann. Appl. Probab.*]{} [**14**]{} 1455–1478.
8. [Durham, S. D., Flournoy, N.]{} and [Li, W.]{} (1998), Sequential designs for maximizing the probability of a favorable response. [*Canadian Journal of Statistics*]{} [**3**]{} 479–495.
9. [Eggenberger]{}, F. and [Pólya]{}, G. (1923). Über die statistik verketetter vorgänge. [*Zeitschrift für Angewandte Mathematik und Mechanik*]{} [**1**]{} 279–289.
10. [Erev]{}, I. and [Roth]{}, A (1998). Predicting how people play games: reinforcement learning in experimental games with unique, mixed strategy equilibria. [*Amer. Econ. Rev.*]{} [**88**]{} 848–881.
11. [Hu, F.]{} and [Rosenberger, W. F.]{} (2006). [*The Theory of Response-Adaptive Randomization in Clinical Trials*]{}. John Wiley and Sons. Wiley Series in Probability and Statistics.
12. [Hu, F.]{} and [Rosenberger, W. F.]{} (2003). Optimality, variability, power: evaluating response-adaptive randomization procedures for treatment comparisons. [*J. Amer. Statist. Assoc.*]{} [**98**]{} 671–678.
13. [Hu, F., Rosenberger, W. F.]{} and [Zhang, L.-X.]{} (2006). Asymptotically best response-adaptive randomization procedures. [*J. Statist. Plann. Inf.*]{} [**136**]{} 1911–1922.
14. [Hu, F.]{} and [Zhang, L.-X.]{} (2004). Asymptotic normality of urn models for clinical trials with delayed response. [*Bernoulli*]{} [**10**]{} 447–463.
15. [Ivanova, A.]{} (2003). A play-the-winner type urn model with reduced variability. [*Metrika*]{} [**58**]{} 1–13.
16. [Ivanova, A.]{} and [Flournoy, N.]{} (2001). A birth and death urn for ternary outcomes: stochastic processes applied to urn models. In [*Probability and Statistical Models with Applications*]{} (Charalambides, Ch. A., Koutras, M. V. and Balakrishnan, N., Eds.). Chapman & Hall/CRC, 583–600.
17. [Ivanova, A., Rosenberger, W. F., Durham, S. D. ]{} and [Flournoy, N.]{} (2000). A birth and death urn for randomized clinical trials: asymptotic methods. [*Sankhya B*]{} [**62**]{} 104–118.
18. [Janson, S.]{} (2004). Functional limit theorems for multitype branching processes and generalized Pólya urns. [*Stochastic Process. Appl.*]{} [**110**]{} 177–245.
19. [Johnson]{}, N. L. and [Kotz]{}, S. (1977). [*Urn Models and Their Applications*]{}. Wiley, New York.
20. [Kotz]{}, S. and [Balakrishnan]{}, N. (1997). Advances in urn models during the past two decades, In [*Advances in Combinatorial Methods and Applications to Probability and Statistics*]{} (Balakrishnan, N., Ed.) Birkhäuser, Boston.
21. [Knoblauch, K., Neitz, M.]{} and [Neitz, J.]{} (2006). An urn model of the development of L/M cone ratios in human and macaque retinas. [*Visual Neuroscience*]{} [**23**]{} 387–394.
22. [McCabe-Dansted, J. C.]{} and [Slinko, A.]{} (2006). Exploratory analysis of similarities between social choice rules. [*Group Decision and Negotiation*]{} [**15**]{} 77–107.
23. [Milenkovic, O.]{} and [Compton, K. J.]{} (2004). Probabilistic transforms for combinatorial urn models. [*Combinatorics, Probability and Computing*]{} [**13**]{} 645–675.
24. [Rosenberger, W. F., Stallard, N., Ivanova, A., Harper, C. N.]{} and [Ricks, M. L.]{} (2001). Optimal adaptive designs for binary response trials. [*Biometrics*]{} [**57**]{} 909–913.
25. [Stout, W. F.]{} (1974). [*Almost sure convergence*]{}. Academic Press, New York.
26. [Wei, L. J.]{} (1979). The generalized Polya’s urn design for sequential medical trials. [*Ann. Statist.*]{} [**7**]{} 291–296.
27. [Wei, L. J.]{} and [Durham, S. D.]{} (1978). The randomized play-the-winner rule in medical trials. [*J. Amer. Statist. Assoc.*]{} [**73**]{}, 840–843.
28. [Windrum]{}, P. (2004). Leveraging technological externalities in complex technologies: Microsoft’s exploitation of standards, in the browser wars. [*Research Policy*]{} [**33**]{} 385–394.
29. [Zhang, L.-X.]{} (2004). Strong approximations of martingale vectors and their applications in Markov-chain adaptive designs. [*Acta Math. Appl. Sinica, English Series*]{} [**20**]{}(2) 337–352.
30. [Zhang, L.-X., Chan, W. S., Cheung. S. H.]{} and [Hu, F.]{} (2007). A generalized urn model for clinical trials with delayed responses. [*Statistica Sinica*]{} [**17**]{} 387–409..
31. [Zhang, L.-X., Hu, F.]{} and [Cheung. S. H.]{} (2006). Asymptotic theorems of sequential estimation-adjusted urn models. [*Ann. Appl. Probab.*]{} [**16**]{} 340–369.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The main features of nuclear reactions with participation of baryon resonances ($\Delta$(1232), S(1535) and $\Lambda$(1405)), which are interpreted as a manifestation of the bound states of nucleus and corresponding meson are discussed.'
author:
- 'O.D.Dalkarov'
date: 'Lebedev Physical Institute, Moscow, Russia'
title: On the nuclear reactions with participation of nearthreshold baryon resonances
---
Pion-nuclear interaction in resonance energy region
===================================================
In the first experimental works for $\pi$-meson interaction with nuclei in the energy diapason of the order of 100 MeV and higher the specific behaviour caused by the resonance nature of $\pi$-meson interaction with nuclear nucleon or with nuclei as whole were revealed. From the experimental point of view the reactions with simplest nucleus - deutron, in particular, $\pi^+$+ d$\to$p + p, $\pi^+$ + d $\to$ $\pi^+$ + p + n, $\pi$d-elastic scattering and others were most investigated. The reactions enumerated above are followed to common peculiarity: the behaviour of total cross-sections and others characteristics are corresponded to resonanced nature of $\pi$d - interaction at the energy about 150 MeV ( in the c.m.s. ). In particular, first experimental data obtained by M.G.Mescheryakov et al. [@1] about behaviour of total cross - sections of $\pi^+$d$\to$pp reaction and inverse one show distinquishable maximum in this energy region ( with the width of the order of 120 MeV ). The analysis of angular distribution of $\pi^+$-meson for p + p$\to$ d + $\pi^+$ reaction points to P-wave meson production and besides that the angular distribution is the same with the increasing of $\pi$-meson energy by the factor two. This fact can indicate either on the equal ratio between $^{1}S_{0}\to^{3}S_{1}$ and $^{1}D_{2}\to^{3}S_{1}$ transitions to whole considered energy diapason and amplitudes for these transitions should be equally depend on the energy or the dominant role plays one of these transitions.
Experimental data could be qualitativaly explained if one take into account that the $\pi$-meson interaction has a resonance behaviour in the state with spin and isospin are equal 3/2 (isobar $\Delta_{33}$) In this case the main contribution to P-wave $\pi$-meson production will be caused by $^{1}D_{2}\to^{3}S_{1}$ transition. There are two possibilities for explanation of resonance maximum in the reactions $pp\to d\pi^{+}$ and $\pi^{+}d\to pp$ : either two-nucleon interaction in$^{1}D_{2}$-state has resonance nature, i.e. so called “dinucleon resonance” is formed (in other terminology - simplest “isonucleus” or “$\pi$- mesonucleus”) or this process takes place throw the virtual production of $\Delta_{33}$-isobar.
The theoretical consideration of these reactions in the assumption of virtual isobar-nucleon state was carried out in [@2], where triangle Feynman diagram with virtual isobar was calculated and it was shown that the energetical behaviour near complex singularity of considered diagram could immitate resonance behaviour of total cross - section for $\pi^{+}d\to pp$ reaction. In this case angular distribution of $\pi$-meson is caused by dominant contribution of S-wave in the relative motion of $\Delta_{33}$ and nucleon. By the analogy resonance behaviour of $\pi^{+}d\to pp$ reaction in the energy region near 2.84 GeV could be explained by the virtual production of $\Delta(1920)$ [@3].
As for the experimental investigation of $\pi$d elastic scattering in resonance region i.e. at the $\pi$-meson energies of the order of 200 MeV( in the l.s. ) it was also revealed the existence of distinquishable maximum in the total cross - section. Observed maximum in the total cross - section and ratio $\left(\sigma(\pi^{-}d)-\sigma(\pi{-}p)\right)/\sigma(\pi^{-}p)\approx
3$ indicate on virtual $\Delta_{33}$-nucleon production.
On the language of Feynman diagrams this mechanism is corresponded to the calculation of nonrelativistic squared diagram. The total and differential cross - sections of $\pi$d elastic scattering were calculated with this diagram [@4]. Resonance behaviour is caused by singularities of this diagram which lie near physical energy region. The comparison with experimental data show rather well aggreement ( some differences between the calculations correspond to the different deutron form-factors at the relatively large ($ >$ 150 MeV/c ) relative momenta). Precision investigation of differential cross - section could be used as a criterium between two possible explanations for observed resonance maximum: a) formation of two - nucleon resonance ( “isonucleus” or “mesonucleus” ) with B = 2 and b) enchancement mechanism due to singularities of Feynman diagrams just near physical region of kinematical variables.
Thus the exact conclusion about the existence of simplest “isonucleus” ( or “mesonucleus” ) is not possible to make from the contemporary experimental data and theoretical works since it is necessary to exclude the contribution of peripheral mechanisms which are corresponded to $\Delta_{33}$-nucleon virtual state.
As for the interaction of $\pi$-meson with complex nuclei the interesting experimental observation was made for the reaction $\pi^{-}{^{12}C} \to \pi^{-}n^{11}C$ where the resonance maximum in the total cross - section at the kinetic $\pi$-meson energy of the order of 200 MeV ( in the l.s. ) was observed [@5]. Using the well known theory of direct nuclear reactions the analysis of this reaction was made [@6]. The behaviour of total cross-section at $\pi$-meson energies more than 200 MeV was well quantitatively described by the single nucleon mechanism ( pole approximation ) for the production of $\Delta_{33}$ and its following decay to $\pi$-nucleon channel. However in the low energy region ( $< 200$ MeV ) there is significant discrepance between experimental data and calculations with pole approximation. Finally, maximum in the total cross-section for the $^{12}C{\pi^{-} \to \pi^{-}n}^{11}C$ reaction was found to be wider and shifted in comparison with pole approximation.
The reason for this discrepance (as it was shown [@7]) could be consist in the assumption that not only pole diagram is distinquished in this energy region. Namely, one can be noted that the triangle diagram, which corresponds to elastic scattering of $\Delta_{33}$ on $^{11}C$, has the singularity just near physical region of kinematical variables. In this case we deal with the complex singularity at the energy ( in the c.m.s. ) is equal to the sum of the masses of $\Delta_{33}$ and $^{11}C$ nucleus. Taking into account the contribution of pole and triangle diagrams the rather well description of the experimental data was obtained. In particular, the broadering of resonance maximum and its shift (of the order of 30 MeV ) to the region of smaller energy of incident $\pi$-mesons were explained. Note, that the contribution of diagram with rescattering of $\Delta$-isobar on intermediate nucleus looks like as interaction of “stopped” isobar with $^{11}C$ nucleus.
Thus the present experimental data and theoretical considerations leave the question open on the existence of “isonuclei”. In each concrete case it is necessary to carry out the complex of experimental and theoretical investigations to separate the effects of nuclear reaction mechanism from the formation of “isonuclei” i.e. bound state of isobar ( or - $\pi$-meson ) with nucleus. First of all it is needed to investigate the angular distributions and polarized characteristics.
$\eta$-nuclei
==============
At present there are considerable experimental data and theoretical investigations on the possible existence of $\eta$-meson-nucleus bound states or “$\eta$-nuclei”. Detailed review of present situation was done in [@8]. Difficulties in interpretation of experimental data consist in the production type mechanism for $\eta$-meson on free nucleon which has threshold character and is determined by the production of $S_{11}$(1535) baryon resonance and its following decay to $\pi$N or $\eta$N with equal probabilities. Most convinced data on the photoproduction of $S_{11}$(1535) on nuclei were obtained using the beam of $\gamma$-quanta from electron sinchrotron (FIAN).The spectrum of correlated $\pi^{+}n$ pairs arising from $\gamma+^{12}C\to\pi^{+}nX$ and flying transversely to photon beam have been observed [@9]. The main result consists in observation of the shift for resonance maximum in ($\pi^{+}n$) system in comparison with Table value for $S_{11}$(1535).
In view of these experimental data one could takes notice of the following: a. observed shift in the position of $S_{11}$(1535) in the reaction $\pi^{+}+^{12}C\to \pi^{+}nX$ looks like as the same phenomenon for production of $\Delta_{33}$(1232) in the reaction $\pi+^{12}C\to \pi$NX. In the last case observed shift could be explained by the possible nuclear reaction mechanism for production of $\Delta_{33}$ on nucleus. This mechanism includes triangle Feynman diagram with rescattering $\Delta_{33}$ on intermediate nucleus besides pole diagram corresponded to the production on the bound nucleon. Taking into account these two diagrams it was turn out to explain the broadering and shift of $\Delta_{33}$ in production on nuclei. Note that in rescattering processes as a propagater for virtual $\Delta$ it is necessary to use a pole position in the $\pi$N amplitude which corresponds to smaller mass in comparison with the value on mass shell ( Re M($\Delta_{33}$)= 1210 MeV for $\Delta_{33}$(1232)). The same pole position for $S_{11}$(1535) is equal to Re M($S_{11}$) = 1500 MeV. At the same time the position of complex singularity is determined by the sum of Re M(resonance) and mass of nucleus. To clear up a nuclear reaction mechanism for photoproduction of $S_{11}$(1535) on nuclei the additional experimental and theoretical efforts are needed. As for the one nucleon mechanism (pole diagram ) this question was considered in details in [@10]. In particular an investigation of small transvers momentum region for ($\pi^{+}n$) pairs gives possibility to use the Treyman - Yang criterium [@11].
In view of considered above the investigation of peripheral mechanisms for production of $S_{11}$(1535) on nuclei can give additional information about the phenomena connected with the formation of $\eta$-nuclei on the background of peculiarities of the nuclear reaction mechanism.
K - nuclei
============
By the analogy with the cases considered above the existence of nearthreshold $\Lambda$(1405) baryon resonance determining low energy KN - interaction as input for hypothesis about possible formation of nearthreshold K - nucleus bound states was used.
In the work [@12] phenomenological potential model for K-meson interaction with light nuclei was proposed. For elementary KN-interaction a potential approximation with simplest Gaussion form for interaction potential was used. The parameters of KN - potential were choosed so to obtain correct KN scattering lengths known from the shifts and widths of $K^{-}p$ and $K^{-}d$ atoms and to reproduce the mass of $\Lambda$(1405) as a bound state of K-meson and nucleon. Using obtained so two-body KN - potential the optical potential for K - nucleus interaction was calculated. With this optical potential the existence of $^{3}_{K}H$(I=0,1) and $^{4}_{K}H$(I=1/2) bound states with binding energies from 20 MeV up to 100 MeV and widths of the order several tens MeV were predicted.
The proton and neutron energy distributions from He(stopped $K^{-}{,}N$) reactions were measured in [@13]. In the missing mass spectrum two mono-energetic peaks were observed. These peaks peaks were interpreted as the formation of tribaryon $S^{0}$(3115) and $S^{+}$(3140) with isospins I=0 and 1 and strangeness S=-1, correspondingly. In the potential model considered in [@12] these states should correspond to the bound $^{3}_{K}H$(I=0) and $^{3}_{K}H$(I=1) states. Binding energies for these states ($~$200 MeV) turned out to be larger and widths smaller ( $<$20 MeV ) than theoretically predicted. Nevertheless taking into account statistical significance ( 13$\sigma$ for S(3115) and 3.7$\sigma$ for S(3140)) these experimental results really indicate on the possible existence of three-body K-nuclei. Another interpretations of these data are not so convincing ( for instance, the interpretation as the $\Sigma$NN hypernuclei is less probable since excitation energies for these hupernuclei should be of the order 50 MeV or higher ).
As the direct test to prove the existence of bound K-nuclei it will be useful to observe the discrete spectra of $\gamma$-quanta or $\pi$-mesons injected due to transitions from states of $K^{-}$-He atom to one of the states of three-body K-nucleus. Simple estimations give for the probabilities of radiative transitions the value of the order of $10^{-3}-10^{-4}$ and in the case of probabilities for $\pi$-meson transitions will be $10^{-1}-10^{-2}$, correspondingly.
In conclusion, it would be note that the question on the existence of new kind of nuclei with meson as a constituent is a great of interest. However the interpretation of present experimental data is not so simple. From our point of view the reasons are either in the necessity of experimental confirmation for obtained data ( for instance, in the case of K -nuclei ) or in the additional experimental and theoretical analysis. In particular the question about possible nuclear reaction mechanism should be especially checked. Here we are not considered the predictions of quarks model in which the rich spectrum of multiquark states are predicted [@14]since to speak about quark nature of the states considered above without identification even if one of ground states looks like as prematurely.
This work is supported by the RFBR grant 05-02-17482-a.
[100]{} M.G.Mescheryakov, B.S.Neganov. Dokl.Acad.Nauk SSSR 100(1955)677 O.D.Dalkarov. Yad. Fiz. 6(1967)71 G.Cocconi et al. Phys.Lett. 7(1963)222 O.D.Dalkarov. Yad. Fiz. 4(1966)728 P.L.Reeder, S.S.Markowitz. Phys.Rev. B639(1964)133 V.M.Kolybasov. Yad. Fiz. 2(1965) O.D.Dalkarov. Yad. Fiz. 7(1968)820 G.A.Sokol. Preprint FIAN-5 (2005) G.A.Sokol et al. ArXiv:nucl-ex/0011005 I.S.Shapiro, V.M.Kolybasov and G.A.Leksin.Usp.Fiz.Nauk SSSR 113(1974)239 I.S.Shapiro, V.M.Kolybasov and R.G.August. Nucl.Phys. 61(1965)353 Y.Akaishi and T.Yamazaki. Phys.Rev.C 65(2002)044005 T.Suzuki et al. Phys.Lett. B597(2004)263 Y.Maezawa, T.Hatsuda and S.Sasaki. Prog.Theor.Phys. 114(2005)317
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
Victoria M. KASPI\
[*Department of Physics and Center for Space Research, Massachusetts Institute of Technology, Cambridge, MA, 02139, USA, [email protected]*]{}
nocite:
- '[@sr68]'
- '[@shh84]'
- '[@kmj+92]'
- '[@cks+92]'
- '[@fk91]'
- '[@mkj+91]'
- '[@wcd91]'
- '[@kcb+88]'
- '[@kpha93]'
- '[@klm+93]'
- '[@fkv93]'
- '[@car93]'
- '[@jml+95]'
- '[@cl86]'
- '[@gj95a]'
- '[@po93]'
- '[@kmj+96]'
- '[@rv88]'
- '[@dls72]'
- '[@tc94]'
- '[@rtj+96]'
- '[@lv72]'
- '[@dth78]'
- '[@lr91]'
- '[@kbm+97]'
- '[@acj+96]'
- '[@sh82]'
- '[@fsp96]'
title: 'NEUTRON STAR/SUPERNOVA REMNANT ASSOCIATIONS'
---
34.5pc
Abstract {#abstract .unnumbered}
========
We summarize the current observational evidence for associations between supernova remnants and neutron stars, including radio pulsars, proposed “inactive” young neutron stars, “anomalous” X-ray pulsars, and soft gamma-ray repeaters, and argue that the paradigm that all young neutron stars are Crab-like requires reconsideration.
Introduction
============
When Baade & Zwicky proposed in 1934 that collapsed stars composed of neutrons could be formed in supernova explosions, they had little notion of how such creatures would manifest themselves observationally. The surprise discovery of pulsars, and in particular, of the Crab and Vela pulsars in their respective supernova remnants, heralded the first visual image of the isolated neutron star: that of a compact, highly magnetized (surface field $\sim
10^{12}$ G) star, spinning down slowly due to magnetic braking, emitting a collimated beacon, and exciting its surroundings via the injection of ultra-relativistic particles. However, even the simplest follow-up questions, such as whether all neutron stars form in supernovae, what fraction of supernovae produce neutron stars, and in particular, whether all young pulsars are born with properties like those of the Crab and Vela pulsars remain to this day naggingly unanswered.
Here I summarize observations of neutron stars plausibly associated with supernova remnants. The last similar review was published by Helfand & Becker (1984). Their interesting synthesis of the observational data is beyond the scope of this paper. More recent reviews including only radio pulsar/supernova remnant associations can be found elsewhere (Kaspi 1996; Kaspi 1998).
Rotation-Powered Neutron Star/Supernova Remnant Associations
============================================================
Associations between remnants and neutron stars have traditionally been identified with detections of radio pulsars. The Helfand & Becker (1984) review included only three such objects: the Crab, Vela, and PSR B1509$-$58. Since then, the number of associations has blossomed to between 7 and 23, depending on the one’s criteria for certainty. The great increase is for several reasons. Many young pulsars were found in radio pulsar surveys of the Galactic plane at observing frequencies near 1400 MHz (Clifton & Lyne 1986; Johnston et al. 1992), where dispersion-measure and scatter-broadening of radio pulses are reduced relative to traditional pulsar survey frequencies, like 430 MHz. Also some, but not all, radio pulsar searches specifically targeting remnants have been successful (Manchester, D’Amico & Tuohy 1985; Kaspi et al. 1996; Gorham et al. 1996; Lorimer, Lyne & Camilo 1998). Finally major advances in the capabilities of X-ray observatories have led to several recent exciting discoveries, discussed further below (§\[sec:psr\])
Proposed associations may be merely a result of coincidental projection of the pulsar and remnant on the sky; the probability for chance alignment is generally significant, particularly for the inner Galactic plane. The burden of proof therefore rests on the observer. Evidence comes from consideration of whether independent distance and age estimates agree, whether there is an interaction between the pulsar and remnant, whether the transverse velocity implied by the angular displacement of the pulsar from the approximate remnant geometrical center, its (perhaps naively) assumed birthplace, is consistent with the known pulsar velocity distribution, and, even better, if the measured pulsar velocity’s magnitude and direction is as predicted. For most proposed associations, few of these questions have clear answers (see Kaspi 1998 for more detail).
Table 1 summarizes proposed rotation-powered pulsar/supernova remnant associations. In the Table, PSR and SNR are the pulsar and remnant names, T is the remnant type (S is shell, P is plerion, C is composite), $\tau$ and $d$ are the best age and distance estimates for PSR/SNR, $\beta$ is the angular displacement of the pulsar from the apparent remnant center in remnant radii, $v_t$ is the best-guess transverse velocity assuming the pulsar characteristic age and the most reliable distance estimate, with $v_t$ in bold actual measurements. ${\cal E}$ is a rough figure of merit, where ${\cal E}
= 1$ is secure, and ${\cal E} = 5$ dubious. Note that often ${\cal
E}$ is high for lack of supporting, rather than conclusively damaging, evidence.
Conclusions to be drawn from Table 1 were discussed by Kaspi (1998). The main points are that observable young pulsars have larger magnetic fields, shorter spin periods, but not always larger radio luminosities, than the typical pulsar. The evidence from associations for high pulsar velocities should as yet be considered inconclusive; proper motion measurements for young pulsars are crucial for deciding whether many of the associations listed in Table 1 are genuine. The detection of bow-shock nebulae via high-resolution radio and X-ray observations may provide another way to test associations.
Recent X-ray Discoveries {#sec:psr}
------------------------
[**N157B:**]{} Marshall et al. (1998) have just discovered a 16 ms X-ray pulsar in the Crab-like supernova remnant N157B in the Large Magellanic Cloud using the [*Rossi X-ray Timing Explorer*]{}. The surprisingly short period makes it the most rapidly rotating neutron star known that has not been spun up. The measured pulsar characteristic age is $\sim$5 kyr old, making it the fourth youngest pulsar known, and suggesting that neutron stars can be born spinning significantly faster than previously thought, possibly as fast as a few milliseconds.
[**RCW 103:**]{} Torii et al. (1998) have discovered 69 ms X-ray pulsations from a source $7^{\prime}$ north of the center of the well-studied, young shell supernova remnant RCW 103 using [*ASCA*]{}, finally confirming an earlier claim by Aoki, Dotani & Mitsuda (1992) who used [*GINGA*]{}. The spin-down rate implied by the difference in periods indicates that the pulsar is very young, having age $\sim$8 kyr. The existence of so young a pulsar near RCW 103 suggests the two might be related, casting doubt on the famous central point source being the stellar remnant (see §\[sec:inactive\]) An association requires a transverse velocity of $\sim$800 km s$^{-1}$ for a distance of 3.3 kpc, deduced from HI absorption. The radio-pulsar counterpart has recently been discovered at the Parkes observatory (Kaspi et al. in preparation). The dispersion measure suggests a distance to the pulsar of $\sim$4.5–7 kpc. Thus the association can be considered only tentative, and the nature of the central object remains uncertain.
[**PSR J1105$-$6107:**]{} The 63 ms radio pulsar PSR J1105$-$6107 lies almost three remnant radii from the approximate center of the supernova remnant G290.1$-$0.8, also called MSH 11$-$61A (Kaspi et al. 1997). For an association, under standard assumptions, the pulsar transverse velocity must be $\sim$650 km s$^{-1}$. The recent detection of the pulsar at X-ray energies (Gotthelf & Kaspi 1998) is best explained as arising from a pulsar wind nebula confined by ram-pressure. High-resolution X-ray imaging can test the association if a bow-shock morphology is found.
[**G11.2$-$0.3:**]{} The supernova remnant G11.2$-$0.3 is the possible counterpart of the event recorded by the Chinese in AD 386 (Strom 1994). The recent detection of evidence for 65 ms X-ray pulsations (Torii et al. 1997) is exciting, as after the Crab, this is the only pulsar associated with an historic event. From the single [*ASCA*]{} observation, $\dot{P} < 8 \times 10^{-13}$. If the pulsar was born in AD 386, and assuming a short birth spin period, the implied $\dot{P} = 6.4 \times 10^{-13}$, consistent with the upper limit. This implies a spin-down luminosity $\dot{E} < 9 \times
10^{37}$ erg s$^{-1}$, and surface magnetic field $B < 1 \times
10^{13}$ G, reasonable for a young pulsar. Confirmation measurement of $\dot{P}$ are top priorities for the future.
“Inactive” Neutron Star/Supernova Remnant Associations {#sec:inactive}
======================================================
Several unidentified X-ray point sources observed in supernova remnants have been hypothesized as being neutron stars emitting X-rays from their initial cooling after formation. To date, the sources that fall in this category are: 1E 1207.4$-$5209 in G296.5+10.0 (Mereghetti, Bignami & Caraveo 1996), 1E 1613$-$5055 in RCW 103 (Tuohy & Garmire 1980; but see also §\[sec:psr\] and Torii et al. 1998), RX J0002+6246 in G117.7+0.6 (Hailey & Craig 1995), and a point source in Puppis A (Petre, Becker & Winkler 1996). The observational properties of these sources that support the neutron star identification are: they are X-ray point sources to within available spatial resolutions, they are coincident with supernova remnants, they have high X-ray to optical luminosity ratios, their spectra are roughly consistent with blackbody emission with temperatures and emitting surface areas roughly as expected for neutron stars, and their X-ray emission does not vary long-term. However they are clearly different from pulsars like the Crab because no X-ray pulsations are seen, no associated radio or $\gamma$-ray emission, pulsed or unpulsed, has been detected, and no evidence for extended pulsar-powered synchrotron nebulae is observed. These objects challenge the conventional paradigm that young neutron stars are born as pulsars, energetic and spinning fast, exciting synchrotron nebulae in their immediate surroundings. An unambiguous conclusion regarding the neutron-star nature of these objects could come from the detection of low-pulsed-fraction X-ray pulsations, expected for magnetized, thermally cooling sources, or through high-resolution spectral X-ray observations, which could reveal absorption features predicted in models of neutron star atmospheres (see Pavlov, this volume).
“Anomalous” X-ray Pulsar/Supernova Remnant Associations {#sec:magnetar}
=======================================================
The class of objects collectively known as “anomalous” X-ray pulsars is characterized by pulsations long in duration compared with those of the radio pulsar population (5-9 s), and slow but steady spin-down. The members of this class are 4U 0142+61, RX J0720.4$-$3125, 1E 1048.1$-$5937, 4U 1626$-$67, RX J1838.4$-$0301 and 1E 2259+586. Mereghetti & Stella (1995) summarize the properties of these systems and argue that they are binaries in which the neutron star accretes from so low-mass a companion that the orbital Doppler shift is not detectable. No optical counterpart has been found for any of the sources except 4U 1626$-$67, which is certainly an X-ray binary (Middleditch et al. 1981).
It is intriguing that two of the sources, 1E 2259+586 (Fahlman & Gregory 1981) and RX J1838.4$-$0301 (Schwentker 1994) are associated with supernova remnants. These associations are unexplained in the binary model, although they would not represent the first known X-ray binaries associated with remnants. That honor goes to SS 433, a well-known unusual radio-emitting X-ray binary associated with the supernova remnant W50 (Margon 1984).
The anomalous X-ray pulsars in supernova remnants could be young, isolated, and rotation-powered. However their X-ray luminosities far exceed the implied spin-down luminosities. This demands the introduction of a physical mechanism for the production of X-rays not previously observed in neutron stars. Furthermore, the implied surface magnetic fields are enormous, $\sim 10^{14}$ G. This has led to them being dubbed “magnetars,” and suggests that the origin of the observed X-rays could lie, for example, in the decay of the magnetic field by diffusive processes (e.g. Thompson & Duncan 1996).
[**Kes 73:**]{} The recent discovery of a 12 s X-ray pulsar in the young supernova remnant Kes 73 (Vasisht & Gotthelf 1997) adds renewed strength to the magnetar hypothesis. The strategic location of the pulsar near the center of the remnant makes a chance coincidence improbable. A weak archival [*ROSAT*]{} detection of the periodicity implies a spin-down rate and hence $\dot{E}$ and $B$ that are consistent with the magnetar model. Confirmation of the spin-down rate and long-term timing are important for establishing the nature of this source.
Soft Gamma-Ray Repeaters
========================
The class of objects known as “soft gamma-ray repeaters” includes three sources: SGRs 0526$-$66, 1806$-$20, and 1900+14. They are characterized by recurrent episodes of soft-spectrum, short-duration $\gamma$-ray emission (see Kouveliotou 1996 for a review). Multi-wavelength studies suggest that they represent yet another facet of the young neutron star population, since two of the three (SGRs 0526$-$66 and 1806$-$20) are apparently associated with supernova remnants (Cline et al 1982; Kulkarni & Frail 1993). The absence of an obvious remnant host to SGR 1900+14 remains problematic (Vasisht et al. 1996). Thompson & Duncan (1996) discuss SGRs in the context of a magnetar model, suggesting that the SGR phenomenon may represent a phase of evolution in the life of a magnetar, with X-ray pulsations (§\[sec:magnetar\]) characterizing a different stage. SGRs are discussed by Kulkarni in more detail elsewhere in this volume.
Conclusions
===========
The most striking conclusion following examination of the associations between neutron stars and supernova remnants is that the paradigm that every young neutron star is like the Crab pulsar requires reconsideration. Neutron stars, it seems, come in many flavors, perhaps some yet to be discovered. The continued analysis of valuable archival high-energy data, the upcoming launch of the [*Advanced X-ray Astrophysics Facility*]{}, as well as the ongoing Parkes multi-beam and Nançay Galactic plane survey should ensure significant progress, and even perhaps more surprises in this field in the near future.
References
==========
Anderson, S. et al., 1996, ApJ, 468, L55 Aoki, T., Dotani, T., Mitsuda, K. 1992, IAUC 5588 Caraveo, P. A., 1993, ApJ, 415, L111 Caswell,J., Kesteven,M., Stewart,R., Milne,D., Haynes,R., 1992, ApJ, 399, L151 Clifton, T. R., Lyne, A. G., 1986, Nature, 320, 43 Cline, T., et al., 1982, ApJ, 255, L45 Damashek, M., Taylor, J. H., Hulse, R. A., 1978, ApJ, 225, [L]{}31 Davies, J. G., Lyne, A. G., Seiradakis J. H., 1972, Nature, 240, 229 Fahlman, G. G., Gregory, P. C., 1981, Nature, 293, 202 Finley, J. P., Ögelman, H., 1994, ApJ, 434, L25 Finley, J. P., Srinivasan, R. & Park, S., 1996, ApJ, 466, 938 Frail, D. A., Kulkarni, S. R., 1991, Nature, 352, 785 Frail, D. A., Kulkarni, S. R., Vasisht, G., 1993, Nature, 365, 136 Frail, D. A., Goss, W. M., Whiteoak, J. B. Z., 1994, ApJ, 437, 781 Gaensler, B. M., Johnston, S., 1995, MNRAS, 275, 73P Gorham, P., Ray, P., Anderson S., Kulkarni, S., Prince, T., 1996, ApJ, 458, 257 Gotthelf, E. V., Kaspi, V. M., 1998, ApJL, in press Hailey, C. J., Craig, W. W., 1995, ApJ, 455, L151 Helfand, D. J., Becker, R. H. 1984, Nature, 307, 215 Johnston, S., et al., 1992, MNRAS, 255, 401 Johnston,S.,Manchester,R.,Lyne,A.,Kaspi,V.,D’Amico,N., 1995, A&A, 293, 795 Kaspi,V.,Manchester,R.,Johnston,S.,Lyne,A.,D’Amico,N., 1992, ApJ, 399, L155 Kaspi, V. M. et al. 1993, ApJ, 409, L57 Kaspi, V. M., 1996, [*IAU Colloquium 160: Pulsars: Problems and Progress*]{}, ASP Conference Series, 105, 375 Kaspi,V., Manchester,R., Johnston,S., Lyne,A., D’Amico,N., 1996, AJ, 111, 2028 Kaspi, V. M. et al. 1997, ApJ, 485, 820 Kaspi, V. M., 1998, Adv. Space. Res., 21, in press Kassim, N. E., Weiler, K. W., 1990, Nature, 343, 146 Kouveliotou, C., 1996, [*IAU Symp. 165: Compact Stars in Binaries*]{}, Kluwer, Dordrecht, 477 Kulkarni, S. R. et al., 1988, Nature, 331, 50 Kulkarni, S., Predehl, P., Hasinger, G., Aschenbach, B., 1993, Nature, 362, 135 Kulkarni, S. R., Frail, D. A., 1993, Nature, 365, 33 Large, M. I., Vaughan, A. E. & Mills, B. Y., 1968, Nature, 220, 340 Large, M. I., Vaughan, A. E., 1972, Nature Phys. Sci., 236, 117 Leahy, D. A., Roger, R. S., 1991, Astron. J., 101, 1033 Lorimer, D. R., Lyne, A. G., Camilo, F. 1998, A&A, 331, 1002 Lyne, A. G., Lorimer, D. R., 1994, Nature, 369, 127 Manchester, R. N., D’Amico, N., Tuohy, I. R., 1985, MNRAS, 212, 975 Manchester,R.,Kaspi,V.,Johnston,S.,Lyne,A.,D’Amico,N., 1991, MNRAS, 253, 7P Margon, B. 1984, ARAA, 22, 507 Marshall, F., Middleditch, J. Zhang, W., Gotthelf, E., 1998, IAUC 6810 McAdam, W. B., Osborne, J. L., Parkinson, M. L., 1993, Nature, 361, 516 Mereghetti, S., Stella, L. 1995, 442, L17 Mereghetti, S., Bignami, G. F., Caraveo, P. A., 1996, ApJ, 464, 842 Middleditch, J., Mason, K. O., Nelson, J. E., White, N. E. 1981, ApJ, 244, 1001 Nicastro, L., Johnston, S. & Koribalski, B. 1996, A&A, 306, 49 Petre R., Becker C. M., Winkler P. F., 1996, ApJ, 465, L43 Phillips, J. A., Onello, J. S., 1993, [*Massive Stars: Their Lives in the Interstellar Medium*]{}, ASP, 35, 419 Ray, P. S. et al., 1996, ApJ, 470, 1103 Routledge, D., Vaneldik, J. F., 1988, ApJ, 326, 751 Schwentker, O. 1994, A&A, 286, L47 Seward, F. D., Harnden, F. R., 1982, ApJ, 256, L45 Seward, F. D., Harnden, F. R., Helfand D. J., 1984, ApJ, 287, L19 Staelin, D. H., Reifenstein, [III]{} E. C., 1968, Science, 162, 1481 Strom, R. G. 1994, A&A, 288, L1 Thompson, R. J., Córdova F. A., 1994, ApJ, 421, L13 Thompson, C., Duncan, R. C. 1996, ApJ, 473, 322 Torii, K., Tsunemi, H., Dotani, T., Mitsuda, K. 1997, ApJ, 489, L145 Torii, K. et al 1998, ApJ, 494, L207 Tuohy, I., Garmire, G., 1980, ApJ, 239, 107 Vasisht, G., Kulkarni, S. R., Frail, D. A., Greiner, J. 1996, ApJ, 456, L59 Vasisht, G., Gotthelf, E. V. 1997, ApJ, 486, L129 Wolszczan, A., Cordes, J. M., Dewey, R. J., 1991, ApJ, 372, L99
---------------- --------------- ------ ------------- ------------- ----------- ---------------- ---------- ----------
SNR Type $\beta$ $\cal E$ Refs
B0531+21 Crab P 1.3/0.9 2/2 $\sim$0 [**125**]{} 1 1
B0540$-$69 SNR0540$-$693 P 1.7/0.6 50/50 $\sim$0 $\sim$0 1 2
J0537$-$6910 N157B C 5/4 50/50 $\sim$0.5 $\sim$600 1 38
B0833$-$45 Vela C 11/18 0.6/0.5 0.3 [**170**]{} 1 32
B1509$-$58 MSH 15–52 C 1.7/10 5.7/4.2 0.2 3000? 2 27
B1757$-$24 G5.4$-$1.2 C 16/14 4.6/5 1.2 1600 2 5,6
B1853+01 W44 S 20/$\sim$10 3/3.1 0.6 250 2 7
B1951+32 CTB 80 C 107/96 2.4/3 $\sim$0 300 2 8
J1341$-$6220 G308.8$-$0.1 C 12/32 8.7/7 0.35 600 3 3,4
B1800$-$21 G8.7$-$0.1 S 16/15-28 4/3.2-4.3 $\sim$0 $\sim$0 3 33,34
B1823$-$13 X-ray nebula P 21/? 2.5/? $\sim$0 $\sim$0 3 28
J0538+2817 S147 S 600/80-200 1.8/0.8–1.6 0.4 30 3 26
B1643$-$43 G341.2+0.9 C 33/- 6.9/8.3-9.7 0.7 475 3 29
B2334+61 G114.3+0.3 C 41/10-100 2.4/1.8 0.1 $<$50 3 9
B1758$-$23 W28 C 58/35-150 13.5/2 1.0 200 3 10,11
J1811.5$-$1926 G11.2$-$0.3 C -/2 -/$>$5 $<$1 ? 4 35
B1610$-$50 Kes 32 S 7.5/5 7/3-7 1.5 1600 4 12,13
J1617$-$5055 RCW 103 S 8/4 7/3.3 1.5 800 4 36,37
B1727$-$33 G354.1+0.1 ? 26/- 4.2/- $\sim$0 460 4 29
J1105$-$6107 G290.1$-$0.8 S 63/? 7/7 2.9 650 4 25
B1830$-$08 W41 S 148/$<$50 4-5/4.8 1.6 200 4 14,15
B1855+02 G35.6$-$0.5 ? 160/- 9/4 or 12 0.4 100 4 16
J1627$-$4845 G335.2+0.1 S 2700/- 6.8/6.5 0.4 70 4 17
B1706$-$44 G343.1$-$2.3 C 17.5/- 2.4-3.2/3 1.0 [**$<$50**]{} 4 29,30,31
B1930+22 G57.3+1.2 ? 40/- 9.6/4.5 0.5 750 5 18
B0611+22 IC 443 S 89/65 4.7/1.5 1.7 [**110**]{} 5 19
B0656+14 Monogem S? 110/60-90 0.8/0.3 0.5 [**250**]{} 5 20
B1832$-$06 G24.7+0.6 C 120/12 6.3/4.4 1.6 360 5 15
J2043+2740 Cygnus Loop S 1200/20 1.1/0.6 2.5 1500 5 21
B1154$-$62 G296.8$-$0.3 S 1600/25 10/4 1.4 550 5 22
B0458+46 G160.9+2.6 S 1800/30-100 1.8/1-4 0.3 [**$<$150**]{} 5 23,24
---------------- --------------- ------ ------------- ------------- ----------- ---------------- ---------- ----------
: Proposed Rotation-Powered Pulsar/Supernova Remnant Associations.
Refs: \[1\] Staelin & Reifenstein (1968) \[2\] Seward et al. (1984) \[3\] Kaspi et al. (1992), \[4\] Caswell et al. (1992) \[5\] Frail & Kulkarni (1991) \[6\] Manchester et al. (1991) \[7\] Wolszczan et al. (1991) \[8\] Kulkarni et al. (1988) \[9\] Kulkarni et al. (1993) \[10\] Kaspi et al. (1993) \[11\] Frail et al. (1993) \[12\] Caraveo (1993) \[13\] Johnston et al. (1995) \[14\] Clifton & Lyne (1986) \[15\] Gaensler & Johnston (1995a) \[16\] Phillips & Onello (1992) \[17\] Kaspi et al. (1996) \[18\] Routledge & Vaneldik (1988) \[19\] Davies, Lyne & Seiradakis (1972) \[20\] Thompson & Cordova (1994) \[21\] Ray et al. (1996) \[22\] Large & Vaughan (1972) \[23\] Damashek et al. (1978) \[24\] Leahy & Roger (1991) \[25\] Kaspi et al. (1997) \[26\] Anderson et al. (1996) \[27\] Seward & Harnden (1982) \[28\] Finley, Srinivasan & Park (1996) \[29\] Frail, Goss, & Whiteoak (1994) \[30\] McAdam, Osborne & Parkinson (1993) \[31\] Nicastro et al. (1996) \[32\] Large, Vaughan & Mills (1968) \[33\] Kassim & Weiler (1990) \[34\] Finley & Ögelman (1994) \[35\] Torii et al. (1997) \[36\] Torii et al. (1998) \[37\] Kaspi et al. (1998) \[38\] Marshall et al. (1998)
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
\
E-mail:
- 'Richard Armstrong$^{1,2}$, Paul Groot$^3$, Vanessa McBride$^{2,4}$, James Miller-Jones$^5$, Kunal Mooley$^{1}$, Ben Stappers$^6$, Ralph Wijers$^7$'
- |
Michael Bietenholz$^{8,9}$, Sarah Blyth$^2$, Markus Bottcher$^{10}$, David Buckley$^{4,11}$, Phil Charles$^{1,12}$, Laura Chomiuk$^{13}$, Deanne Coppejans$^{14}$, Stéphane Corbel$^{15,16}$, Mickael Coriat$^{17}$, Frederic Daigne$^{18}$, Erwin de Blok$^2$, Heino Falcke$^3$, Julien Girard$^{15}$, Ian Heywood$^{19}$, Assaf Horesh$^{20}$, Jasper Horrell$^{21}$, Peter Jonker$^{3,22}$, Tana Joseph$^4$, Atish Kamble$^{23}$, Christian Knigge$^{12}$, Elmar Körding$^3$, Marissa Kotze$^4$, Chryssa Kouveliotou$^{24}$, Christine Lynch$^{25}$, Tom Maccarone$^{26}$, Pieter Meintjes$^{27}$, Simone Migliari$^{28}$, Tara Murphy$^{25}$, Takahiro Nagayama$^{29}$, Gijs Nelemans$^3$, George Nicholson$^8$, Tim O’Brien$^6$, Alida Oodendaal$^{27}$, Nadeem Oozeer$^{21}$, Julian Osborne$^{30}$, Miguel Perez-Torres$^{31}$, Simon Ratcliffe$^{21}$, Valerio Ribeiro$^{32}$, Evert Rol$^6$, Anthony Rushton$^1$, Anna Scaife$^6$, Matthew Schurch$^2$, Greg Sivakoff$^{33}$, Tim Staley$^1$, Danny Steeghs$^{34}$, Ian Stewart$^{35}$, John Swinbank$^{36}$, Kurt van der Heyden$^2$, Alexander van der Horst$^{24}$, Brian van Soelen$^{27}$, Susanna Vergani$^{37}$, Brian Warner$^2$, Klaas Wiersema$^{30}$\
\
$^1$ Astrophysics, Department of Physics, University of Oxford, UK\
$^2$ Department of Astronomy, University of Cape Town, South Africa\
$^3$ Department of Astrophysics, Radboud University, Nijmegen, the Netherlands\
$^4$ South African Astronomical Observatory, Cape Town, South Africa\
$^5$ ICRAR, Curtin University, Perth, Australia\
$^6$ University of Manchester, UK\
$^7$ Anton Pannekoek Institute for Astronomy, University of Amsterdam, the Netherlands\
$^8$ Hartebeeshoek Radio Observatory, South Africa\
$^9$ Department of Physics and Astronomy, York University, Toronto, Canada\
$^{10}$ North-West University, Potchefstroom campus, South Africa\
$^{11}$ Southern African Large Telescope, Cape Town, South Africa\
$^{12}$ School of Physics and Astronomy, Southampton University, UK\
$^{13}$ Department of Physics and Astronomy, Michigan State University, USA\
$^{14}$ Northwestern University, USA\
$^{15}$ Laboratoire AIM (CEA/IRFU - CNRS/INSU - Univ. Paris Diderot), Paris, France\
$^{16}$ Station de Radioastronomie de Nancay, Observatoir de Paris, France\
$^{17}$ University of Toulouse, UPS-OMP, France
- |
\
$^{18}$ Institute d’Astrophysique de Paris, France\
$^{19}$ CSIRO Astronomy and Space Science, Epping, Australia\
$^{20}$ Benoziyo Center fo Astrophysics, Weizmann Institute of Science, Israel\
$^{21}$ SKA South Africa, Pinelands, South Africa\
$^{22}$ SRON, Netherlands Institute for Space Research, Utrecht, the Netherlands\
$^{23}$ Harvard-Smithsonian Center for Astrophysics, Cambridge, USA\
$^{24}$ Department of Physics, The George Washington University, Washington USA\
$^{25}$ School of Physics, University of Sydney, Australia\
$^{26}$ Department of Physics, Texas Tech University, USA\
$^{27}$ Department of Physics, University of the Free State, Bloemfontein, South Africa\
$^{28}$ European Space Agency, Madrid, Spain\
$^{29}$ Department of Physics and Astronomy, University of Kagoshima, Japan\
$^{30}$ Department of Physics and Astronomy, University of Leicester, UK\
$^{31}$ Insituto de Astrofisica de Andalucia, Spain\
$^{32}$ Botswana International University of Science and Technology, Botswana\
$^{33}$ Department of Physics, University of Alberta, Canada\
$^{34}$ Department of Physics, University of Warwick, UK\
$^{35}$ Sterrewacht Leiden, University of Leiden, the Netherlands\
$^{36}$ Department of Astrophysical Sciences, Princeton University, USA\
$^{37}$ GEPI, Observatoire de Paris, France
bibliography:
- 'ThunderKAT2016\_PW17.bib'
title: 'ThunderKAT: The MeerKAT Large Survey Project for Image-Plane Radio Transients'
---
Introduction
============
Transient radio emission is associated with essentially all explosive phenomena and high-energy astrophysics in the universe. It acts as a locator for such events, and a calorimeter of their kinetic feedback to the local environment. Because of this, radio transients are invaluable probes for subjects as diverse as stellar evolution, relativistic astrophysics and cosmology. ThunderKAT is the programme to observe such phenomena in the image plane with MeerKAT. It is also the umbrella project behind the MeerLICHT optical telescope [@2016SPIE.9906E..64B], which will provide simultaneous optical images for all night-time MeerKAT observations.
Accretion physics
=================
Accretion phenomena on all scales and across all environments are associated with the generation of outflows, which reveal themselves via synchrotron radiation in the radio band. Observations of these outflows allow us to test how much kinetic energy is being fed back from the accretion process, and in some cases to directly resolve the fastest-moving bulk phenomena in the universe.
Relativistic accretion
----------------------
### Black holes and neutron stars in X-ray binaries
Stellar-mass black holes (BHs) in X-ray binary systems provide our best sample with which to understand how varying accretion states affect the power and type of kinetic feedback from BH via winds and jets. The neutron star (NS) X-ray binaries provide the best possible ’control sample’ to study which aspects of this behaviour are unique to BH and may therefore be dependent on the existence of an event horizon or ergosphere.
![The radio-X-ray plane for black hole and neutron star X-ray binaries [@2011MNRAS.414..677C]. Quasi-simultaneous observations allow us to map out the long-term connection between accretion and ejection. The empirical patterns observed can be scaled to supermassive black holes in AGN, and yet throw up many mysteries. A comprehensive monitoring programme with MeerKAT will increase the number of measurements in this plane by a factor of up to 5, notably probing for the first time the mapping for the neutron star binaries, which are radio-quiet relative to the black holes.[]{data-label="fig1"}](coriat-f5.png){width="100.00000%"}
The current fundamental tool for understanding the global connection between accretion and ejection in X-ray binaries is the radio-X-ray plane (Fig 1). Currently the plane is populated by around 200 quasi-simultaneous radio and X-ray observations of X-ray binaries and has led to some key results and even more unanswered questions:
- [a strong correlation exists between accretion rate and jet power, which can be extended with the addition of a mass term to include BH of all masses, from X-ray binaries to supermassive black holes in AGN.]{}
- [the jet is suppressed in some high-luminosity states (analogues of radio-quiet AGN)]{}
- [some neutron stars (the transitional millisecond pulsars) behave like black holes in this plane, whereas others show the steeper correlation expected for radiatively efficient accretion in the presence of a hard surface]{}
- [there is a mysterious ’radio quiet’ branch where some BH produce less powerful jets under the same accretion conditions (as measured in X-rays) as others.]{}
Note that for sources in bright accretion modes, X-ray monitors (e.g. ASTROSAT, HXMT) will provide good enough measurements to place the MeerKAT data in the plane. For fainter sources we will request e.g. Swift/Chandra/XMM observations quasi-simultaneously with MeerKAT. Because the neutron star X-ray binaries are significantly fainter in the radio band than the black holes [@2006MNRAS.366...79M], it is only with the advent of the SKA precursors that we finally have the potential to map out the radio-X-ray plane for these sources (while the X-rays are just as bright so do not need new facilities). On any given day there will be 1-10 sources of interest to monitor, typically at $0.1 - 10$ mJy flux densities, i.e. readily achievable within an hour (some sources will be much fainter however and require more time). In addition to the slow evolution of sources in the radio-X-ray plane, X-ray binaries occasionally go into major outbursts, which are associated with radio flaring [@2004MNRAS.355.1105F]. These flaring states are associated with powerful, intermittent jet production and, when matched with high-sensitivity X-ray observations and VLBI, are our best hope for understanding the details of the connection between accretion and jet formation. We note that because the sources are typically bright but rapidly varying during these phases, and that it is the comparison with the X-ray data which provides the physical insights, cadence can trump raw sensitivity. In fact, such observations could sometimes be undertaken with a minor subarray of MeerKAT, or even with a refurbished KAT-7.
### Tidal disruption events
Tidal Disruption Events (TDEs) result from the catastrophic destruction and subsequent accretion of a star by the gravitational field of a supermassive black hole. They provide our best insight into the response of a supermassive black hole to a sudden increase in mass accretion rate. It now seems that this response may be quite different to the ’steadier’ modes that are analogous to X-ray binaries, and yet this could have been the dominant mode during the phases of strongest AGN activity at redshifts 2-3. The current rate of discovery of good TDE candidates is quite low, probably less than 5 per year. The recent detection of weak radio emission from the ’thermal’ TDE ASASSN-14li [@2016Sci...351...62V] suggests that all TDEs produce radio emission, but that there may be a range of kinetic modes and powers. As with neutron star (above) and white dwarf (below) accretion, the huge sensitivity leap with MeerKAT will allow us to probe a significant number of TDE radio afterglows for the first time (currently five in total).
### Ultraluminous X-ray sources
ULX are off-centre X-ray sources in galaxies that have luminosities in excess (in some cases) of the maximum for stellar-mass black holes. They may be our best route to finding the population of intermediate-mass black holes (IMBH) which constitute the path from stellar remnants to supermassive black holes (see also the recent aLIGO results). The best accreting IMBH candidate to date, HLX X-1, has revealed radio flaring indicating transient jet production [@2012Sci...337..554W]. A second flaring ‘HLX’ candidate [@2015MNRAS.454L..26H] has since been discovered. Flaring transient ULXs in nearby galaxies also provide a way to probe Eddington-rate accretion, showing how jets couple to the accretion flow in a radiation-pressure dominated environment, as was likely required for the rapid growth of the first quasars.
White dwarf accretion
---------------------
### Outflows of accretion-powered outbursts of white dwarfs
Cataclysmic variables (CVs) show violent outbursts similar to those outlined above for XRBs, but have as an accretor a white dwarf. It is assumed that outbursts in both classes of objects are triggered by the same disc instability. The similarity between XRBs and CVs is most prominent for dwarf novae (DNe). These non-magnetic, disc-accreting white dwarfs show fairly regular outbursts that are thought to be powered solely by accretion.
Radio emission from CVs had been reported previously but was usually not reproducible. It had thus been suggested that the radio emission is correlated with the optical outburst. This has been confirmed by [@2008Sci...320.1318K] with a detailed radio light-curve of SS Cyg using MERLIN and VLA. Following the repeatable behaviour of SS Cyg, a subsequent survey by South African PhD student Deanne Coppejans with the JVLA using triggered observations of 5 DNe in outburst yielded a 100% detection rate. The triggers have been provided by amateur astronomers from the AAVSO. Most of these nearby sources (100-260 pc), show flux densities around 15-60 $\mu$Jy [@2016MNRAS.463.2229C]. In addition to these transient CVs, persistent CVs of both magnetic and non-magnetic nature show radio emission [@2015MNRAS.451.3801C].
MeerLICHT will greatly enhance the study of accreting white dwarfs. Radio observations will trace the outflowing material of CVs while the optical observations will provide us real time information on the current state of the accretion disc. MeerLICHT’s 1 minute cadence will yield optical fluxes and colours as well as variability information. Unlike in XRBs, the accretion disc in CVs is visible at optical and UV wavelengths. Thus, any night-time radio observation of a CV will allow us to study both the outflow and the accretion flow for the first time truly simultaneously. A MeerKAT in-depth study of a large sample of southern CVs, sampling accreting white dwarfs with a range of different donor stars (from giants to white dwarfs) will determine the radio properties of the population of accreting white dwarfs and will allow for a direct comparison of the outflow properties of non-relativistic accretors with the relativistic black hole and neutron star XRBs.
### Outflows in thermonuclear eruptions on white dwarfs
Radio observations can provide the best estimates of the fundamental nova parameters, e.g. distance (via expansion parallax from high resolution imaging), ejected mass and kinetic energy (via modelling radio light curves) and ejecta geometry (via imaging and light curve modelling). The ENOVA program on the JVLA has demonstrated that low frequency radio observations like those that will be made with the initial L-band of MeerKAT, will still provide new insights into the outflow properties of novae as shown by the recent results on T Pyx [@2014ApJ...785...78N] and V1723 Aql [@2016MNRAS.457..887W]. It is now well understood that in most novae the mass is not ejected in spherically symmetric fashion, but instead shaped in high velocity, strongly collimated bipolar outflows [@2009ApJ...706..738W; @2014Natur.514..339C].
The need for a southern program can be demonstrated by the recent nova V1369 Cen (2013), which has not been properly covered in the radio, even though it was detected by Fermi/LAT and even visible to the naked eye. KAT-7 observations were attempted with an upper limit of 6 mJy at 1.8 GHz (3 sigma). Through a dedicated ToO programme of Galactic novae with MeerKAT - in combination with simultaneous X-ray and optical/near-infrared observations - we can make critical advances in our understanding of nova ejecta (mass, energetics, nature of the outflow).
Cosmic explosions
-----------------
The endpoints of the lives of the most massive stars in the universe are the core collapse supernovae (CCSNe) and gamma-ray bursts (GRBs), the most powerful explosions in the universe. In recent years a picture is beginning to emerge of a continuum of kinetic powers and Lorentz factors, with relativistic, engine-powered CCSNe bridging the gap between classical CCSNe and GRBs (Fig \[fig2\], left panel). Our targetted programme is augmented by commensal studies.
![Left panel: The zoo of radio light curves of stellar explosions from classical supernovae (bottom of panel) to GRB afterglows (top of panel), representing a progression from sub- to highly-relativistic outflows. From [@2014ApJ...797..107M]. Right panel: The extraordinary radio light curve of SN2014 (AMI-LA data). Three, possibly four, peaks have been observed in the radio light curve of this source so far, indicating interactions of the explosive ejecta with a number of discrete shells, providing a unique probe of the last decades (SN ejecta are an order of magnitude faster than the winds which created the shells) of pre-supernova evolution and mass loss of the most massive stars [@2017MNRAS.466.3648A].[]{data-label="fig2"}](Margutti-cut.png "fig:"){width="40.00000%"} ![Left panel: The zoo of radio light curves of stellar explosions from classical supernovae (bottom of panel) to GRB afterglows (top of panel), representing a progression from sub- to highly-relativistic outflows. From [@2014ApJ...797..107M]. Right panel: The extraordinary radio light curve of SN2014 (AMI-LA data). Three, possibly four, peaks have been observed in the radio light curve of this source so far, indicating interactions of the explosive ejecta with a number of discrete shells, providing a unique probe of the last decades (SN ejecta are an order of magnitude faster than the winds which created the shells) of pre-supernova evolution and mass loss of the most massive stars [@2017MNRAS.466.3648A].[]{data-label="fig2"}](sn2014c.png "fig:"){width="60.00000%"}
### Gamma-ray bursts
At all redshifts (currently to $z \sim 8$) half of GRB afterglows are detected in radio with present techniques. This means we can use them as probes of the distant universe, and the star formation process through cosmic history. The special class of short GRBs (duration under 2 s) are thought to result from compact-object mergers and thus also the prime candidates for EM counterparts to gravitational wave events (Sect 2.5).
The connection between (long) GRBs and SNe is still uncertain, and needs to be further explored. According to the best available hypothesis, what makes a GRB different is that it signals the birth of a black hole from a fast-rotating star, which unlike a SN causes significant energy to be emitted as ultra-relativistic jets, giving rise to the spectacularly bright flashes of gamma rays and afterglows lasting days (X rays) to weeks (optical) to years (radio). GRBs are exceedingly rare: only about 1/Myr/galaxy, 10,000 times rarer than SNe. This predicts, among other things, that for every GRB we see, there are 100 whose collimated outflow is not aimed at us. These should be visible as SNe of type Ib/c, for which at late times, when the GRB afterglow becomes non-relativistic and unbeamed, we see a GRB-like radio afterglow. A few recent type Ib/c SNe have shown that a (mildly) relativistic component o f the SN ejecta can sometimes already be observed in the radio a few days after the initial stellar explosion, while a gamma-ray signal has not been detected.
Targeted MeerKAT follow-up of GRBs will help us constrain the GRB physics, determine the true total kinetic energy released, and understand relativistic shocks and particle acceleration in them. Especially for short GRBs, where less than 5 radio afterglows have been seen, this will help greatly in improving our physical understanding of them. Similar follow-up of SNe (Sect. 2.3.2) will allow us to compare this with the non-relativistic version of the same process, and possibly also find the GRB-SN connection through detecting GRBs that were not initially collimated towards us.
Commensal observations in deep fields, such as the LADUMA, MHONGOOSE and Fornax fields, will allow the detection of a more complete sample of radio events, possibly including highly obscured ones, and better establish the population ratios. Tying down the flux and spectrum of GRBs and SNe in the MeerKAT radio bands will allow us to determine the complete physical parameters of the explosion, which has only been possible for a few out of many hundreds of GRBs with afterglows so far. Once we can compare the behaviour in the L band with possible other bands of MeerKAT, this will become all the more powerful.
### Core collapse supernovae
Many CCSNe are not observed at radio wavelengths or, if observed, only followed at radio if there is a detection early after its explosion or optical detection (a few days to a few weeks). The sensitivity of MeerKAT will allow us to follow a statistically meaningful number of exploding CCSNe per year within 40 Mpc with a modest amount of MeerKAT time, permitting us to also obtain the radio luminosity function of CCSNe. Some individual CCSNe result in unique views of the final decades of evolution of the most massive stars (e.g. SN2014C, right panel of Fig 2).
The ideal cadence for the commensal detections of CCSNe is that the same fields are visited 3-5 times per year (LADUMA, MHONGOOSE). Assuming a maximum effective distance of 150 Mpc to discover CCSNe, we estimate that ThunderKAT could yield between 20 and 50 blind CCSNe discoveries, every year.
### Type Ia supernovae with MeerKAT
Type Ia SNe, the end-products of white dwarfs, have been used to make the essential discovery of the accelerating expansion of the universe, yet we do not know what their progenitors are: a C+O white dwarf and a non-degenerate star (single-degenerate scenario), or two white dwarfs (double-degenerate scenario). Radio observations can discriminate between the progenitor models of SNe Ia. For example, in all scenarios with mass transfer from a companion, a significant amount of circumstellar gas is expected, and therefore a shock is bound to form when the supernova ejecta are expelled. The situation would then be very similar to circumstellar interaction in core-collapse SNe (see Sect. 2.3.1), where the interaction of the blast wave from the supernova with its circumstellar medium results in strong radio and X-ray emission. On the other hand, the double-degenerate scenario will not give rise to any circumstellar medium close to the progenitor system, and hence essentially no prompt radio emission is expected.
We aim to perform deep ToO observations of up to two nearby (D $<$ 20 Mpc) Type Ia SNe per year with ThunderKAT, aimed at making the first ever discovery of radio emission from a Type Ia SN (the estimated rate for SN Ia within this volume is 1 $\pm$ 1 per year). We remark that a detection would represent a breakthrough in our understanding of Type Ia SN progenitors, given the existing contraints on SN2011fe [@2012ApJ...746...21H] and on SN2014J [@2014ApJ...792...38P], and would validate the single-degenerate scenario. On the other hand, systematic non-detections will be so constraining that it will rule out essentially most, or all, single-degenerate scenarios, leaving the double-degenerate scenario as the only viable one.
We need the observations to be scheduled as soon as the SN candidate is confirmed as a type Ia SN, and to be done at the highest available frequency of the array, aiming at a rms of around or better than 1 $\mu$Jy/beam.
Fast and coherent transients
----------------------------
### Fast radio bursts
The nature of Fast Radio Bursts (FRBs) remains uncertain, and the largest fraction of these discovered and studied with MeerKAT are going to be done so via the TRAPUM/MeerTRAP projects. Nevertheless, for a telescope as sensitive as MeerKAT, it is entirely possible to discover a very bright FRB in our commensal short timescale imaging, and we anticipate finding and localising a significant number of FRBs this way via our commensal observations. Fast imaging of observations commensal with TRAPUM will be a joint TRAPUM-ThunderKAT collaboration.
However, as well as the short coherent burst there remains the possibility that there is a synchrotron afterglow of the event [@2016Natur.530..453K], and subsequent literature, and/or that some events repeat [@2016Natur.531..202S] and may be associated with faint continuous sources.
Gravitational wave sources
--------------------------
### Gravitational wave events and electro-magnetic counterparts
The recent direct detections of gravitational waves (GWs) from extragalactic binary black hole mergers by the advanced LIGO (aLIGO) detector has opened a new window into the Universe [@2016PhRvL.116f1102A]. This discovery, together with the expected increase in the aLIGO sensitivity, and the addition of VIRGO detectors, suggests that the coming years will see many compact binary star mergers. The current localization of these GW sources is several hundred square degrees (90% containment probability), but this will likely improve to 300 deg$^2$ and 100 deg$^2$ by 2018 and by 2022, respectively. Finding the electromagnetic (EM) counterparts of GW sources is crucial for increasing the science yield from the aLIGO/VIRGO detections, such as 1) lifting degeneracies associated with the binary parameters inferred from the GW signal, 2) reducing the SNR for a significant GW detection by aLIGO/VIRGO, and 3) providing arcsec localization and identifying the precise merger redshift, thus setting/testing the energy and distance scales.
Compact binary mergers containing neutron stars, i.e. NS-NS/NS-BH systems, are expected to produce EM counterparts. With MeerKAT, the goal will be to study the radio afterglows, which probe the post-merger outflows and the (shocked) circum-merger environments. These afterglows are expected to be weak, 10s to 100s of $\mu$Jy, and expected to peak on weeks – months timescales after the merger and detection of the GW signal. Although faint, such afterglows will probably be the only means of finding the EM counterparts of mergers located in dusty galaxies, where the optical kilonova could be highly extincted. With MeerLICHT, we will hunt for the blue (red) kilonova signature that is expected on day (week) timescales. We envisage the MeerKAT follow up observations to be carried out in four epochs logarithmically spaced between 1 day and 1 year post-merger. Subsequent monitoring/follow up will be necessary only for the radio transients identified in these epochs.
Commensal transient searches
============================
In parallel with the targetted programmes outlined above, we will search all incoming data streams in the image plane for transients. We will aim to do this on timescales from $\sim 1$ sec upwards, at approximately logarithmic time intervals. For example, for a ’typical’ 4hr observation we may search for transients at timescales of 1, 10, 100 and 1000 seconds. These searches will be carried out in near real time and will probably operate alongside the beam-formed transient searches associated with the TRAPUM and MeerTRAP projects. By ’piggybacking’ on the other approved LSP surveys, from wide to very deep fields, we will perform the most extensive blind image-plane radio search for transients ever undertaken. In some of our core areas, e.g. SNe, we can confidently predict that the commensal results will augment the science, but the true value of the integrated commensal survey lies in providing a comprehensive understanding of the radio sky variability at all timescales and flux levels. We note that for L-band and lower frequency observations of transients it would be our intention to search for (inter)galactic HI absorption features.
We would like to stress that this aspect of the programme is important in that it is the major ‘exploration of the unknown’ aspect of the MeerKAT LSP, and a very important preparatory study for the SKA. However, the clearer knowledge we have gathered over the past half a decade shows that it is with targetted programmes such as those we have outlined above, that we will get the guaranteed high-profile science.
Comments on observing strategy
==============================
Within the ThunderKAT programme, we will need to use a range of different observing modes, including different speeds of ToO (in terms of response time) interleaved with regular observations (some of which may be regular monitoring of a significant number of point sources).
Commensality with other Large Survey Projects
---------------------------------------------
Following the original MeerKAT LSP awards, we agreed with the other LSPs that we could search their data commensally for transients (with some appropriate ’reward’ in terms of authorship). For the revised 2016 LSP plans we note:
- [MeerTIME / TRAPUM / MeerGAL - These projects will search extensively in the galactic plane and globular clusters, which is good for us for searches for galactic transients;]{}
- [Fornax - This project will observe over two 4.5-hr epochs per pointing across the Fornax survey area;]{}
- [MHONGOOSE - This LSP will provide $6 - 8$ pointings per deep field, providing an excellent resource to look for faint (extragalactic) transients over time scales of weeks to months;]{}
- [LADUMA - This project provides the longest baseline for transient detections on a single pointing over time scales of weeks to months to years;]{}
- [MIGHTEE - We will provide a time-variable sky model to MIGHTEE.]{}
MeerLICHT and the need for some night time observations
-------------------------------------------------------
All of the ThunderKAT science benefits from having simultaneous optical data during the MeerKAT observations. For transients discovered in commensal searches, the simultaneous optical flux and radio flux density measurements of an astrophysical transient allow an initial classification (Stewart et al. 2016, in prep) which will assist in allocating and prioritising multi-wavelength follow-up resources. In this regard, ThunderKAT is already part of a consolidated multi-year South African-led program for rapid follow up of various classes of transients on the Southern African Large Telescope (SALT).
Acknowledgements {#acknowledgements .unnumbered}
================
This proceeding is an adapted version of the 2016 updated ThunderKAT project plan submitted to SKA South Africa in June 2016, ahead of the first light images of MeerKAT-16 released on 30 June 2016. The authorship of this manuscript reflects the full ThunderKAT membership at the time of submission of these proceedings, and are presented in the following order: the two principal investigators of ThunderKAT (Fender and Woudt), and alphabetical list of the work group leaders, followed by a second alphabetical list of ThunderKAT co-investigators. RF acknowleges partial support from the European Research Council. PAW acknowledges the University of Cape Town and the South African National Research Foundation for financial support.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Quantitative isoperimetric inequalities are shown for anisotropic surface energies where the isoperimetric deficit controls both the Fraenkel asymmetry and a measure of the oscillation of the boundary with respect to the boundary of the corresponding Wulff shape.'
address: 'Department of Mathematics, University of Texas at Austin, Austin, TX, USA'
author:
- Robin Neumayer
bibliography:
- 'references2.bib'
title: A strong form of the quantitative Wulff inequality
---
Introduction
============
The Wulff inequality and stability
----------------------------------
For $n\geq 2$, the well-known isoperimetric inequality states that $$P(E) \geq n|B|^{1/n}|E|^{1/n'},$$ $n' = n/(n-1)$, with equality if and only if $E$ is a translation or dilation of $B=\{ x \in {\mathbb{R}^n}: |x| <1\}$, the Euclidean unit ball in ${\mathbb{R}^n}$. The isoperimetric inequality holds for all sets of finite perimeter $E \subset {\mathbb{R}^n}$, with the perimeter of $E$ equal to $ P(E) = {\mathcal{H}^{n-1}}({\partial}^* E). $ Here, ${\partial}^*E$ is the reduced boundary of $E$; see Section \[SOFP\], or [@maggi2012sets] for a more complete overview.
The anisotropic surface energy is a natural generalization of the notion of perimeter and has applications in modeling of equilibrium configurations for solid crystals (see [@Wulff1901; @Herring1951; @taylor1978]) and of phase transitions (see [@Gurtin1985]). We introduce a *surface tension* $f: {\mathbb{R}^n}\to [0, + \infty)$ to be a convex positively $1$-homogeneous function that is positive on $S^{n-1}.$ The corresponding (anisotropic) *surface energy* of a set of finite perimeter $E \subset {\mathbb{R}^n}$ is defined by $${{\Phi}}(E) = \int_{\partial^* E} f(\nu_E(x)) \,d{\mathcal{H}^{n-1}}(x).$$ Here, $\nu_E$ is the measure theoretic outer unit normal; see Section \[SOFP\]. Just as the ball minimizes perimeter among sets of the same volume, as expressed by the isoperimetric inequality, the surface energy is uniquely minimized among sets of a given volume by translations and dilations of a fixed convex set $K$ determined by the surface tension. This set $K$ is given by $$K = \underset{\nu \in S^{n-1}}\bigcap \{ x \in {\mathbb{R}^n}: x\cdot \nu <f(\nu)\} ,$$ and is known as the *Wulff shape* of ${{\Phi}}$. The minimality of the Wulff shape is expressed by the *Wulff inequality*: $${{\Phi}}(E) \geq {{\Phi}}(K)\bigg(\frac{|E|}{|K|}\bigg)^{1/n'}= n |K|^{1/n}|E|^{1/n'},$$ with equality if and only if $E$ is a translation or dilation of $K$; see [@taylor1978; @milman1986asymptotic; @fonseca1991wulff; @fonseca1991uniqueness; @dacorogna1992wulff; @DGS92; @brothers1994]. In the case where the surface tension $f$ is constantly equal to one, ${{\Phi}}(E)$ reduces to the perimeter $P(E)$ and the Wulff inequality reduces to the isoperimetric inequality.
Given a surface tension $f$, one may define the *gauge function* $f_*:{\mathbb{R}^n}\to [0,+\infty)$ by $$f_*(x) = \sup\{ x \cdot \nu : f(\nu) \leq 1\}.$$ The gauge function provides another characterization of the Wulff shape: $K = \{ x : f_*(x) <1\}.$
To quantify how far a set is from achieving equality in the Wulff inequality, we introduce the [*anisotropic isoperimetric deficit*]{}, or simply the [*deficit*]{}, of a set $E$, defined by $$\delta_{{{\Phi}}}(E) = \frac{{{\Phi}}(E) }{n|K|^{1/n}|E|^{1/n'}} -1.$$ The deficit equals zero if and only if, up to a set of measure zero, $E= x +rK$ for some $x\in{\mathbb{R}^n}$ and $r>0$. This quantity is invariant under translations and dilations of $E$, as well as modifications of $E$ by sets of measure zero.
The optimality of the Wulff shape in the Wulff inequality naturally gives rise to the question of stability: does the deficit control the distance of a set from the Wulff shape? In other words, given a distance $d$ from the family $\{x+ rK : x\in {\mathbb{R}^n}, r>0 \}$, one wants to find inequalities of the form $$\label{stability}
\delta_{{{\Phi}}}(E) \geq \omega( d(E)) ,$$ where $\omega $ is a (possibly explicit) function such that $\omega(d(E)) \to 0^+$ as $d(E) \to 0^+$. Ideally, one hopes to find the function $\omega$ that provides the *sharp* rate of decay. Such an inequality can be viewed as a *quantitative* form of the Wulff inequality: by rearranging the deficit, becomes $${{\Phi}}(E) \geq {{\Phi}}(K) \left(\frac{|E|}{|K|}\right)^{1/n'} + \omega( d(E)) n|K|^{1/n}|E|^{1/n'}.$$ In this way, $\omega(d(E))$ serves as a remainder term in the Wulff inequality.
A well studied distance is the *asymmetry index*, $\alpha_{{{\Phi}}}(E)$, defined by $$\label{asym}
\alpha_{{{\Phi}}}(E) = \min_{y \in {\mathbb{R}^n}} \left\{ \frac{|E\Delta (rK+y)|}{|E|} :\ |rK| = |E|\right\},$$ where $E\Delta F =(E\setminus F) \cup( F \setminus E)$ is the symmetric difference of $E$ and $F$. For the case $f$ constantly equal to one, the asymmetry index is known as the *Fraenkel asymmetry*. The quantitative isoperimetric inequality with the Fraenkel asymmetry was proven in sharp form by Fusco, Maggi, and Pratelli in [@FMP08]. Using symmetrization techniques, they showed that if $E$ is a set of finite perimeter with $0<|E|<\infty$, then $$\label{FMPStatement}
\alpha_1(E)^2 \leq C(n) \delta_1 (E).$$ Here and in the future, we use the notation $\delta_1$ and $\alpha_1$ for the deficit and asymmetry index corresponding to the perimeter.
Before this full proof of was given, several partial results were shown; see [@fuglede1989; @Hall1992; @HHW]. Another proof of was given in [@CiLe12], introducing a technique known as the selection principle, where a penalization technique and the regularity theory for almost-minimizers of perimeter reduce the problem to the case shown in [@fuglede1989].
Stability of the Wulff inequality was first addressed in [@Esposito2005], without the sharp exponent. Figalli, Maggi, and Pratelli later proved the sharp analogue of for the Wulff inequality in [@FiMP10], using techniques from optimal transport and Gromov’s proof of the Wulff inequality. They showed that there exists a constant $C(n)$, independent of $f$, such that $$\label{FiMPStatement}
\alpha_{{{\Phi}}}(E)^2 \leq C(n) \delta_{{{\Phi}}}(E)$$ for any set of finite perimeter $E$ with $0<|E| < \infty$. In both (\[FMPStatement\]) and (\[FiMPStatement\]), the power $2$ is sharp. The constant $C(n)$ in is explicit, a feature that is not shared with any other proofs of , , or any of the quantitative inequalities obtained in this paper.
In [@fuscojulin11], Fusco and Julin proved a strong form of the quantitative isoperimetric inequality, improving by showing $$\label{FuscoJulin}
\alpha_1 (E)^2 + \beta_1(E)^2 \leq C(n) \delta_1(E)$$ for any set of finite perimeter $E$ with $0<|E| < \infty$, where the *oscillation index* $\beta_1(E)$ is defined by $$\begin{aligned}
\beta_1(E)& =
\min_{y\in {\mathbb{R}^n}}\bigg\{
\bigg(
\frac{1}{2n|B|^{1/n}|E|^{1/n'} }
\int_{\partial^* E } \Big|\nu_E(x) -\nu_{B_r(y)}\Big(y + r\frac{x-y}{|x-y|}\Big)\Big|^2\, d{\mathcal{H}^{n-1}}(x)\bigg)^{1/2} :
|B_r| = |E|\bigg\} \nonumber\\
\label{betaB}
&= \min_{y\in {\mathbb{R}^n}}
\left( \frac{1}{n|B|^{1/n}|E|^{1/n'} }
\int_{\partial^* E }
1 - \frac{x-y}{|x-y|} \cdot \nu_E(x) \, d{\mathcal{H}^{n-1}}(x)\right)^{1/2} .\end{aligned}$$ While the asymmetry index $\alpha_1(E)$ is an $L^1$ distance between $E$ and $B$, the oscillation index $\beta_1(E)$ quantifies the oscillation of ${\partial}^*E$ with respect to ${\partial}B$. In [@fuscojulin11 Proposition 1.2], $\beta_1(E)$ is shown to control $\alpha_1(E)$; see Proposition \[poincare\] for the analogous statement in the anisotropic case. Once again, the power $2$ in (\[FuscoJulin\]) is sharp for both $\alpha_1(E)$ and $\beta_1(E)$.
Since [@fuscojulin11], analogous strong form quantitative inequalities have been studied in several settings: in Gauss space [@Eldan2015; @BarBraJul14], on the sphere [@BDuzFus13], and in hyperbolic $n$-space [@BDS2015].
Statements of the main theorems
-------------------------------
The goal of this paper is to address the question of what form the inequality $(\ref{FuscoJulin})$ takes in the case of the anisotropic surface tension ${{\Phi}}$. In other words, we prove a *strong form* of the quantitative Wulff inequality, improving by adding a term to the left hand side that quantifies the oscillation of ${\partial}^*E$ with respect to ${\partial}K$. We define the *${{\Phi}}$-oscillation index* by $$\label{bpp}
{\beta_{{{\Phi}}}}(E) = \min_{y\in {\mathbb{R}^n}}
\left( \frac{1}{n|K|^{1/n}|E|^{1/n'}}
\int_{\partial^* E } f(\nu_E(x)) -\nu_E(x) \cdot \frac{x-y}{f_*(x-y) }\, d{\mathcal{H}^{n-1}}(x)\right)^{1/2}.$$ The following theorem is a strong form of the quantitative Wulff inequality that holds for an arbitrary surface energy.
\[th1\] There exists a constant $C$ depending only on $n$ such that $$\label{statement1}
\alpha_{{{\Phi}}}(E)^2 + {\beta_{{{\Phi}}}}(E)^{4n/(n+1)}\leq C \delta_{{{\Phi}}}(E)$$ for every set of finite perimeter $E$ with $0<|E|<\infty$.
As in (\[FiMPStatement\]), the constant is independent of $f$. We expect that, as in (\[FuscoJulin\]), the sharp exponent for ${\beta_{{{\Phi}}}}(E)$ in should be $2$. With additional assumptions on the surface tension $f$, we prove the stability inequality in sharp form for two special cases.
\[lellipt\] A surface tension $f$ is *$\lambda$-elliptic*, $\lambda>0$, if $f\in C^2({\mathbb{R}^n}\setminus \{0\})$ and $$(\nabla^2 f(\nu)\tau)\cdot \tau \geq \frac{\lambda}{|\nu| } \left| \tau - \Big(\tau\cdot \frac{\nu}{|\nu|}\Big)\frac{\nu}{|\nu|} \right|^2$$ for $\nu, \tau \in {\mathbb{R}^n}$ with $\nu\neq 0$.
This is a uniform ellipticity assumption for $\nabla^2 f(\nu)$ in the tangential directions to $\nu$. If $f$ is $\lambda$-elliptic, then the corresponding Wulff shape $K$ is of class $C^2 $ and uniformly convex (see [@schneider2013convex], page $111$). When ${{\Phi}}$ is a surface energy corresponding to a $\lambda$-elliptic surface tension, the following sharp result holds. The constant depends on ${m_{{{\Phi}}}}$ and ${M_{{{\Phi}}}}$, a pair of constants defined in that describe how much $f$ stretches and shrinks unit-length vectors.
\[Smooth\] Suppose $f$ is a $\lambda$-elliptic surface tension with corresponding surface energy ${{\Phi}}$. There exists a constant $C$ depending on $n, \lambda, {m_{{{\Phi}}}}/ {M_{{{\Phi}}}},$ and $\|\nabla^2 f\|_{C^0({\partial}K)}$ such that $$\label{smooth12}\alpha_{{{\Phi}}}(E)^2 +{\beta_{{{\Phi}}}}(E)^2 \leq C \delta_{{{\Phi}}}(E)$$ for any set of finite perimeter $E$ with $0<|E|<\infty$.
The second case where we obtain the strong form quantitative Wulff inequality with the sharp power is the case of a crystalline surface tension.
\[crystals\] A surface tension $f$ is *crystalline* if it is the maximum of finitely many linear functions, in other words, if there exists a finite set $\{x_j\}_{j=1}^N \subset {\mathbb{R}^n}\setminus \{0\}, N\in \mathbb{N}$, such that $$f(\nu) = \underset{1\leq j\leq N}{\max} \{ x_j \cdot \nu\}\ \ \mbox{ } \text{ for all } \nu \in S^{n-1}.$$
If $f$ is a crystalline surface tension, then the corresponding Wulff shape $K$ is a convex polyhedron. In dimension two, when $f$ is a crystalline surface tension, we prove the following sharp quantitative Wulff inequality.
\[dim2\] Let $n=2$ and suppose $f$ is a crystalline surface tension with corresponding surface energy ${{\Phi}}$. There exists a constant $C$ depending on $f$ such that $$\alpha_{{{\Phi}}}(E)^2 +{\beta_{{{\Phi}}}}(E)^2 \leq C \delta_{{{\Phi}}}(E)$$ for any set of finite perimeter $E$ with $0<|E|<\infty$.
Some remarks about the definition of the ${{\Phi}}$-oscillation index $\beta_{{{\Phi}}}$ in are in order. The oscillation index $\beta_1(E)$ in measures oscillation of the reduced boundary of a set $E$ with respect to the boundary of the ball. Indeed, the quantity $\beta_1(E)$ is the integral over $\partial^*E$ of the Cauchy-Schwarz deficit $1 - \frac{x}{|x|}\cdot\nu_E(x)$, which quantifies in a Euclidean sense how closely $\nu_E(x)$ aligns with $\frac{x}{|x|}$.
To understand , we remark that $f$ and $f_*$ are dual in the sense that they yield a Cauchy-Schwarz-type inequality called the Fenchel inequality, which states that $$\nu_E(x) \cdot \frac{x}{f_*(x)} \leq f(\nu_E(x)).$$ Just as the oscillation index $\beta_1(E)$ quantifies the overall Cauchy-Schwarz deficit between $\frac{x}{|x|}$ and $\nu_E(x)$, the term ${\beta_{{{\Phi}}}}(E)$ is an integral along ${\partial}^*E$ of the deficit in the Fenchel inequality. In Section \[Subsection:anisotropies\], we show that $f(\nu_E(x)) = y \cdot \nu_E(x)$ for $y\in\partial K$ if and only if $y$ is a point on $\partial K$ where $\nu_E(x)$ is normal to a supporting hyperplane of $K$ at $y$. In this way, ${\beta_{{{\Phi}}}}(E)$ quantifies how much normal vectors of $E$ align with corresponding normal vectors of $K$, and therefore provides a measure of the oscillation of the reduced boundary of $E$ with respect to the boundary of $K$. Note that in the case $f$ constantly equal to one, ${\beta_{{{\Phi}}}}$ agrees with $\beta_1$.
It is not immediately clear that is the appropriate analogue of in the anisotropic case. Noting that $x \mapsto (x-y)/f_*(x-y)$ is the radial projection of ${\mathbb{R}^n}\setminus \{0\}$ onto $\partial K + y$, one may initially want to consider the term $$\begin{aligned}
{\beta_{{{\Phi}}}}^*(E)&
= \min_{y \in R^n}
\bigg( \frac{1}{2n |K|^{1/n}|E|^{1/n'} }\int_{\partial^* E}
\Big|\nu_E(x) - \nu_{K}\Big(\frac{x-y}{f_*(x-y)}\Big) \Big|^2\, d{\mathcal{H}^{n-1}}(x) \bigg)^{1/2}\nonumber \\
& = \min_{y \in R^n}
\bigg( \frac{1}{n|K|^{1/n}|E|^{1/n'} }\int_{\partial^* E}
1-\nu_E(x) \cdot \nu_K\Big(\frac{x-y}{f_*(x-y)}\Big)\, d{\mathcal{H}^{n-1}}(x) \bigg)^{1/2}.\label{betastar}\end{aligned}$$ However, in Section \[other\] we see that such a term does not admit any stability result for general ${{\Phi}}$. Indeed, in Example \[crystalexample\], we construct a sequence of crystalline surface tensions that show that there does not exist a power $\sigma$ such that $$\label{noway1}
\beta^*_{{{\Phi}}}(E)^{\sigma} \leq C(n,f) \delta_{{{\Phi}}}(E)$$ for all sets $E$ of finite perimeter with $0<|E|<\infty$ and for all ${{\Phi}}$. Furthermore, Example \[pexample\] shows that even if we restrict our attention to surface energies which are ${\gamma}$-$\lambda$ convex, a weaker notion of $\lambda$-ellipticity introduced in Definition \[gammalambda\], an inequality of the form cannot hold with an exponent less than $\sigma = 4.$ The examples in Section \[other\] illustrate the fact that, in the anisotropic case, measuring the alignment of normal vectors in a Euclidean sense is not suitable for obtaining a stability inequality for general ${{\Phi}}$; it is essential to account for the anisotropy in this measurement. The ${{\Phi}}$-oscillation index $\beta_{{{\Phi}}}(E)$ in does exactly this.
In the positive direction, when the surface tension $f$ is ${\gamma}$-$\lambda$ convex, $\beta^*_{{{\Phi}}}(E)$ is controlled by ${\beta_{{{\Phi}}}}(E)$. As one expects from Example \[pexample\], the exponent in this bound depends on the ${\gamma}$-$\lambda$ convexity of $f$. We now define ${\gamma}$-$\lambda$ convexity.
\[gammalambda\] Let $f:{\mathbb{R}^n}\to \mathbb{R}$ be a nonnegative, convex, positively one-homogeneous function. Then we say that $f$ is *${\gamma}$-$\lambda$ convex* for ${\gamma}\geq 0 , \lambda>0$ if $$\label{gamma lambda}f(\nu + \tau ) + f(\nu- \tau) -2f(\nu)
\geq \frac{\lambda}{|\nu|}
\left| \tau -\Big(\tau \cdot \frac{\nu}{|\nu|}\Big)\frac{\nu}{|\nu|}\right|^{2+{\gamma}}$$ for all $\nu, \tau \in {\mathbb{R}^n}$ such that $\nu \neq 0$.
Dividing by $\tau^2$, the left hand side gives a second difference quotient of $f$. While $\lambda$-ellipticity assumes that $f\in C^2({\mathbb{R}^n}\setminus \{0\})$ and that its second derivatives in directions $\tau$ that are orthogonal to $\nu$ are bounded from below, $\gamma$-$\lambda$ convexity only assumes that the second difference quotients in these directions have a bound from below that degenerates as $\tau$ goes to $0$. Of course, a $0$-$\lambda$ convex surface tension $f$ with $f\in C^2({\mathbb{R}^n}\setminus \{0\})$ is $\lambda$-elliptic. The $\ell^p$ norms $f_p (x) = (\sum_{i=1}^n |x_i|^p)^{1/p}$ for $p \in (1, \infty)$ are examples of $\gamma$-$\lambda$ convex surface tensions; see Section \[other\]. When $f$ is a ${\gamma}$-$\lambda$ convex surface tension, the following theorem shows that ${\beta_{{{\Phi}}}}$ controls $\beta^*_{{{\Phi}}}$.
\[th2\] Let $f$ be a ${\gamma}$-$\lambda$ convex surface tension. Then there exists a constant $C$ depending on $ {\gamma}, \lambda,$ and $m_{{{\Phi}}}/M_{{{\Phi}}}$ such that $${\beta_{{{\Phi}}}}^*(E)^{(2+{\gamma})/2} \leq C \left(\frac{P(E)}{n|K|^{1/n}|E|^{1/n'}}\right)^{{\gamma}/4} \beta_{{{\Phi}}}(E).$$ for any set of finite perimeter $E$ with $0<|E|<\infty$.
As in Theorem \[Smooth\], the constant depends on ${m_{{{\Phi}}}}$ and ${M_{{{\Phi}}}}$ which are defined in . As an immediate consequence of Theorem \[th2\], Theorem \[th1\], and Theorem \[Smooth\], we have the following result.
\[cor1\] If $f$ is a ${\gamma}$-$\lambda$ convex surface tension, then there exists a constant $C$ depending on $n, {\gamma}, \lambda,$ and $ m_{{{\Phi}}}/M_{{{\Phi}}}$ such that $$\alpha_{{{\Phi}}}(E)^2 + {\beta_{{{\Phi}}}}^*(E)^{\sigma} \leq C \left(\frac{P(E)}{n|K|^{1/n}|E|^{1/n'}}\right)^{{\gamma}n/(n+1)} \delta_{{{\Phi}}}(E)$$ for any set of finite perimeter $E$ with $0<|E|<\infty$, where $\sigma=2n(2+{\gamma})/(n+1)$.
If $f$ is a $\lambda$-elliptic surface tension, then there exists a constant $C$ depending on $n, {\gamma}, \lambda, m_{{{\Phi}}}/M_{{{\Phi}}},$ and $\| \nabla^2 f\|_{C^0({\partial}K)}$ such that $$\alpha_{{{\Phi}}}(E)^2 + {\beta_{{{\Phi}}}}^*(E)^{2} \leq C \delta_{{{\Phi}}}(E)$$ for any set of finite perimeter $E$ with $0<|E|<\infty$.
Discussion of the proofs
------------------------
At the core of the proof of are a selection principle argument, the regularity theory of almost-minimizers of perimeter, and an analysis of the second variation of perimeter. Indeed, with a selection principle argument in the spirit of the proof of (\[FMPStatement\]) by Cicalese and Leonardi in [@CiLe12], Fusco and Julin reduce to a sequence $\{F_j\}$ such that each $F_j$ is a $(\Lambda, r_0)$-minimizer of perimeter (Definition \[pm\]) and $F_j \to B$ in $L^1$. Then, by the standard regularity theory, each set $F_j$ has boundary given by a small $C^1$ perturbation of the boundary of the ball. This case is handled by a theorem of Fuglede in [@fuglede1989], which says the following: Let $E$ be a *nearly spherical set*, i.e., a set with barycenter ${{\rm bar }\,}E = |E|^{-1}\int_E x\,dx$ at the origin such that $|E| = |B|$ and $$\partial E = \{ x + u(x) x : x \in \partial B\}$$ for $u: \partial B \to \mathbb{R}$ with $u\in C^1(\partial B)$. There exist $C$ and ${\epsilon}$ depending on $n$ such that if $\| u \|_{C^1(\partial B)} \leq{\epsilon}$, then $$\label{fugst}
\|u\|_{H^1(\partial B)}^2 \leq C \delta_1(E).$$ The proof of makes explicit use of spherical harmonics to provide a lower bound for the second variation of perimeter. It is then easily shown that $\alpha_1(E) + \beta_1(E) \leq C \|u\|_{H^1(\partial B)},$ and therefore (\[fugst\]) implies in the case of nearly spherical sets. Indeed, $\alpha_1(E) \leq C \beta_1(E)$ as shown in Proposition \[prop1\], and in the case of nearly spherical sets, the oscillation index $\beta_1$ is essentially an $L^2$ distance of gradients: if $w(x) = x+ u(x)x$, then $$\nu_E(w(x)) = \frac{ x(1+ u(x)) + \nabla u(x)}{\sqrt{(1+u)^2 + |\nabla u|^2}},$$ where the $\nabla u$ is the tangential gradient of $u$. Then $$\begin{aligned}
n|K|\beta_1(E)^2 \leq \int_{\partial E} 1 - \nu_E(w) \cdot \frac{w}{|w|}\, d {\mathcal{H}^{n-1}}&= \int_{\partial B} \sqrt{ (1+u)^2 + |\nabla u|^2} - (1+u) \,d {\mathcal{H}^{n-1}}\\
&= \int_{\partial B} \frac{1}{2}| \nabla u|^2 + O(|\nabla u |^2)\, d{\mathcal{H}^{n-1}}\leq \|u\|_{H^1(\partial B)}^2.\end{aligned}$$
In each of Theorems \[th1\], \[Smooth\], and \[dim2\], at least one of the three key ingredients of the proof of Fusco and Julin is missing. The proof of Theorem \[th1\] uses a selection principle to reduce to a sequence of $(\Lambda, r_0)$-minimizers of ${{\Phi}}$ converging in $L^1$ to $K$. However, for an arbitrary surface tension, uniform density estimates (Lemma \[UDE\]) are the strongest regularity property that one can hope to extract. We pair these estimates with to obtain the result.
The proof of Theorem \[Smooth\] follows a strategy similar to that of the proof of (\[FuscoJulin\]) in [@fuscojulin11]. If $f$ is a $\lambda$-elliptic surface tension, then $(\Lambda, r_0)$-minimizers of the corresponding surface energy ${{\Phi}}$ enjoy strong regularity properties. Using a selection principle argument and the regularity theory, we reduce to the case where ${\partial}E$ is a small $C^1$ perturbation of ${\partial}K$. The difficulty arises, however, in showing the following analogue of Fuglede’s result (\[fugst\]) in the setting of the anisotropic surface energy.
\[Fug-type\] Let $f$ be a $\lambda$-elliptic surface tension with corresponding surface energy ${{\Phi}}$ and Wulff shape $K$. Let $E$ be a set such that $|E| = |K|$ and ${{\rm bar }\,}E = {{\rm bar }\,}K$, where ${{\rm bar }\,}E =|E|^{-1} \int_E x\, dx$ denotes the barycenter of $E$. Suppose $$\partial E = \{ x + u(x) \nu_K(x) : x \in \partial K\}$$ where $u: \partial K \to \mathbb{R}$ is in $ C^1(\partial K)$. There exist $C$ and ${\epsilon}_1$ depending on $n, \lambda,$ and ${m_{{{\Phi}}}}/{M_{{{\Phi}}}}$ such that if $\| u \|_{C^1(\partial K)}\leq {\epsilon}_1$, then $$\label{fugnew}\|u\|_{H^1(\partial K)}^2 \leq C\delta_{{{\Phi}}}(E).$$
Again, ${m_{{{\Phi}}}}$ and ${M_{{{\Phi}}}}$ are defined in . To prove , Fuglede shows that, due to the volume and barycenter constraints respectively, the function $u$ is orthogonal to the first and second eigenspaces of the Laplace operator on the sphere. This implies that, thanks to a gap in the spectrum of this operator, functions satisfying these constraints satisfy a Poincaré inequality with a larger constant than the Poincaré inequality that holds for $w \in H^{1}({\partial}B)$ with mean zero (i.e., satisfying only the volume constraint). Fuglede’s reasoning uses that fact that the eigenvalues and eigenfunctions of the Laplacian on the sphere are *explicitly known*.
The analogous operator on ${\partial}K$ arising in the second variation of ${{\Phi}}$ also has a discrete spectrum, but one cannot expect to understand its spectrum explicitly. Instead, to prove , we exploit in order to obtain a Poincaré inequality with a larger constant for functions $u\in H^{1}({\partial}K)$ satisfying the volume and barycenter constraints.
Then, as in the isotropic case, one shows that $ \alpha_{{{\Phi}}}(E) + {\beta_{{{\Phi}}}}(E) \leq C \|u \|_{H^1(\partial K)}$ for a constant $C = C(n, \| \nabla^2 f\|_{C^0({\partial}K)})$, and therefore implies for small $C^1$ perturbations. Indeed, Proposition \[prop1\] implies that $\alpha_{{{\Phi}}}(E) \leq C(n) {\beta_{{{\Phi}}}}(E)$, and the fact that $ {\beta_{{{\Phi}}}}(E) \leq C \|u\|_{H^1(\partial K)}$ is a consequence of a Taylor expansion and a change of coordinates. The computation is postponed until (\[betaexp\]) as it relies on notation introduced in Section \[smoothsection\].
The proof of Theorem \[dim2\] also uses a selection principle-type argument to reduce to a sequence of almost-minimizers of ${{\Phi}}$ converging in $L^1$ to the Wulff shape. In this case, a rigidity result of Figalli and Maggi in [@figallimaggi11] allows us reduce to the case where $E$ is a convex polygon whose set of normal vectors is equal to the set of normal vectors of $K$. From here, an explicit computation (Proposition \[crystalprop\]) shows the result.\
The paper is organized as follows. In Section \[preliminaries\], we introduce some necessary preliminaries for our main objects of study. Section \[generalsection\] is dedicated to the proof of Theorem \[th1\], while in Sections \[smoothsection\] and \[crystalsection\] we prove Theorems \[Smooth\] and \[dim2\] respectively. In Section \[other\], we consider the term $\beta_{{{\Phi}}}^*(E)$ defined in , providing two examples that show that one cannot expect stability with a power independent of the regularity of $f$ and proving Theorem \[th2\].\
[**[Acknowledgments:]{}**]{} The author would like to thank Alessio Figalli and Francesco Maggi for their mentorship, guidance, and many helpful discussions. Further thanks are due to a thorough referee for providing several useful remarks. This research was supported by the NSF Graduate Research Fellowship under Grant No. DGE-$1110007$.
Preliminaries
=============
Let us introduce a few key properties about sets of finite perimeter, the anisotropic surface energy, and the ${{\Phi}}$-oscillation index ${\beta_{{{\Phi}}}}$.
Sets of finite perimeter {#SOFP}
------------------------
Given an ${\mathbb{R}^n}$-valued Borel measure $\mu$ on ${\mathbb{R}^n}$, the *total variation* $|\mu|$ of $\mu$ on a Borel set $E$ is defined by $$|\mu|(E) = \sup \bigg\{ \sum_{j\in\mathbb{N}} |\mu(E_j) | : \ E_j\cap E_i = \emptyset, \ \bigcup_{j\in\mathbb{N}} E_j \subset E\bigg\}.$$ A measurable set $E\subset {\mathbb{R}^n}$ is said to be a *set of finite perimeter* if the distributional gradient ${\rm{D}}\chi_E$ of the characteristic function of $E$ is an ${\mathbb{R}^n}$-valued Borel measure on ${\mathbb{R}^n}$ with $|{\rm{D}}\chi_E|({\mathbb{R}^n}) < \infty.$
For a set of finite perimeter $E$, the *reduced boundary* $\partial^*E$ is the set of points $x\in{\mathbb{R}^n}$ such that $|{\rm{D}}\chi_E|(B_r(x))>0$ for all $r>0$ and $$\label{reducedbd}
\underset{r\to 0^+}{\lim} \frac{{\rm{D}}\chi_E (B_r(x))}{|{\rm{D}}\chi_E| (B_r(x) )} \quad \text{ exists and belongs to }S^{n-1}.$$ If $x \in \partial^*E$, then we let $-\nu_E$ denote the limit in . We then call $\nu_E: \partial^*E \to S^{n-1} $ the *measure theoretic outer unit normal to E*. Up to modifying $E$ on a set of Lebesgue measure zero, one may assume that the topological boundary ${\partial}E$ is the closure of the reduced boundary ${\partial}^*E.$ For the remainder of the paper, we make this assumption.
The surface tension and the gauge function {#Subsection:anisotropies}
------------------------------------------
Throughout the paper, we let $$\label{mM}
m_{{{\Phi}}} = \underset{\nu \in S^{n-1} }{\inf}f(\nu), \qquad M_{{{\Phi}}}= \underset{\nu \in S^{n-1} }{\sup}f(\nu).$$ It follows that $$\frac{1}{M_{{{\Phi}}}} = \underset{x \in S^{n-1}}{\inf}f_*(x), \qquad \frac{1}{m_{{{\Phi}}} } =\underset{x\in S^{n-1}}{\sup}f_*(x).$$ One easily shows that $f(\nu) = \sup \{ x\cdot \nu : x \in K \} $ and $f_*(x) = \inf \{ \lambda: \frac{x}{\lambda} \in K\}.$ This also implies that $B_{m_{{{\Phi}}}} \subset K \subset B_{M_{{{\Phi}}}}$, and so if $|K| =1,$ then ${m_{{{\Phi}}}}^n |B| \leq 1 \leq {M_{{{\Phi}}}}^n |B|$. As mentioned in the introduction, the surface tension $f$ and gauge function $f_*$ are dual in the sense that they satisfy a Cauchy-Schwarz-type inequality, called the *Fenchel inequality*: $$x\cdot \nu \leq f_*(x) f(\nu)$$ for all $x,\nu \in {\mathbb{R}^n}$. We may characterize the equality cases in the Fenchel inequality: for any $\nu$, ${x}\cdot \nu={f_*(x)}f(\nu) $ if and only if $\nu$ is normal to a supporting hyperplane of $K$ at the point $\frac{x}{f_*(x)}\in \partial K$. Indeed, $\nu$ is normal to a supporting hyperplane of $K$ at $x \in \partial K$ if and only if $\nu \cdot (y- x) \leq 0$ (so $ \nu \cdot y \leq \nu \cdot x$) for all $y \in K.$ This holds if and only if $\nu \cdot x = \sup\{ y\cdot \nu : y\in K\} = f(\nu)$. In particular, if $x\in \partial^* K$, then $f_*(x) = 1$ and $$\label{bdK}
f(\nu_K(x)) = x\cdot \nu_K(x).$$
We may compute the gradient of $f_*$ at points of differentiability using the Fenchel inequality. The gauge function $f_*$ is differentiable at $x_0\in {\mathbb{R}^n}$ if there is a unique supporting hyperplane to $K$ at $\frac{x_0}{f_*(x_0)}\in \partial K$. For such an $x_0, $ let $\nu_0 = \nu_K(\frac{x}{f_*(x)}) \in {\mathbb{R}^n}$ be normal to the supporting hyperplane to $K$ at $\frac{x_0}{f_*(x_0)}$, so $\frac{x_0}{f_*(x_0)}\cdot\nu_0 = f(\nu_0)$ by . We define the Fenchel deficit functional by $G(x) = f(\nu_0)f_*(x) - x\cdot \nu_0.$ By the Fenchel inequality, $G(x) \geq 0$ for all $x$ and $G(x_0) =0$, so $G$ has a local minimum at $x_0$ and thus $$0 =\nabla G(x_0)
= f(\nu_0)\nabla f_*(x_0) -\nu_0.$$ Rearranging, we obtain $\nabla f_*(x_0) = \frac{\nu_0}{ f(\nu_0)}.$ The $1$-homogeneity of $f$ then implies that $$\label{equal1}
f(\nabla f_*(x)) =1.$$ Furthermore, this implies that $$\label{gradprop}
x \cdot \nabla f_*(x) = x \cdot \nu_K\Big(\frac{x}{f_*(x)}\Big) = f_*(x)$$ (alternatively, this follows from Euler’s identity for homogeneous functions). An analogous argument ensures that $$\label{gradf}\nabla f(\nu_K(x)) = x$$ for $x\in\partial^* K$. Furthermore, we compute $$\label{div}
{{\rm div \,}}\frac{x}{f_*(x)} = \frac{n-1}{f_*(x)}.$$ Indeed, $$\begin{aligned}
{{\rm div \,}}\frac{x}{f_*(x)}
& = \frac{ {{\rm tr}}\, \nabla x}{f_*(x)} + x\cdot \nabla\Big( \frac{1}{f_*(x)}\Big) =\frac{n}{f_*(x)} -\frac{x \cdot \nabla f_*(x) }{f_*(x)^2} = \frac{n-1}{f_*(x)},\end{aligned}$$ where the final equality follows from .
Properties of $\alpha_{{{\Phi}}}$, ${\beta_{{{\Phi}}}}$, and ${\gamma}_{{{\Phi}}}$ {#abg}
----------------------------------------------------------------------------------
Using the divergence theorem, by approximation and the dominated convergence theorem, and , we find that for any $y\in{\mathbb{R}^n}$, $$\int_{\partial^* E } \frac{x-y}{f_*(x-y) }\cdot \nu_{E}(x) \,d{\mathcal{H}^{n-1}}= (n-1)\int_E \frac{dx}{f_*(x-y)} .$$ We may then write $$\label{betaform}{\beta_{{{\Phi}}}}(E)^2 = \frac{{{\Phi}}(E) -(n-1) {\gamma}_{{{\Phi}}} (E)}{n|K|^{1/n}|E|^{1/n'}},$$ where ${\gamma}_{{{\Phi}}}(E)$ is defined by $$\label{gamma}
{\gamma}_{{{\Phi}}} (E) = \underset{y\in{\mathbb{R}^n}}{\sup} \int_{E }
\frac{dx}{f_*(x-y) }.$$ The supremum in (\[gamma\]) is attained, though perhaps not uniquely. If $y\in {\mathbb{R}^n}$ is a point such that $${\gamma}_{{{\Phi}}}(E) = \int_E \frac{dx}{f_*(x-y)},$$ then we call $y$ a *center of $E$*, and we denote by $y_E$ a generic center of $E$. The Wulff shape $K$ has unique center $y_K = 0$. Indeed, take any $y \in {\mathbb{R}^n},$ $y\neq 0$, and recall that $K=
\{f_*(x) < 1\} $. Then $$\begin{aligned}
\int_{K} \frac{dx }{f_*(x)}& - \int_K \frac{dx}{f_*(x-y)}
= \int_K \frac{dx}{f_*(x)} - \int_{K+y} \frac{dx}{f_*(x)}\\
& = \int_{K \setminus (K+y)} \frac{dx}{f_*(x)} - \int_{(K+y)\setminus K} \frac{dx}{f_*(x) } > \int_{K \setminus (K+y)}1dx - \int_{(K+y ) \setminus K } 1dx =0.
\end{aligned}$$ A similar argument verifies that if $|E| = |K|$, then $$\label{maxgamma}
{\gamma}_{{{\Phi}}}(E) \leq {\gamma}_{{{\Phi}}}(K).$$ Moreover, $(n-1) {\gamma}_{{{\Phi}}}(K) = {{\Phi}}(K) = n|K|.$
The following continuity properties of ${{\Phi}}$ and ${\gamma}_{{{\Phi}}}$ will be useful.
\[convergence\] Suppose that $\{E_j\}$ is a sequence of sets converging in $L^1$ to a set $E$, and suppose that $\{f^j\}$ is a sequence of surface tensions converging locally uniformly to $f$, with corresponding surface energies $\{{{\Phi}}_j\}$ and ${{\Phi}}$.
1. The following lower semicontinuity property holds: $${{\Phi}}(E) \leq \underset{j\to\infty}{\liminf}\, {{\Phi}}_j(E_j).$$
2. The function ${\gamma}_{{{\Phi}}}$ defined in (\[gamma\]) is Hölder continuous with respect to $L^1$ convergence of sets with Hölder exponent equal to $1/n'$. In particular, $$|{\gamma}_{{{\Phi}}}(E) - {\gamma}_{{{\Phi}}}(F) | \leq \frac{n|K|}{n-1} |E\Delta F|^{1/n'}$$ for any two sets of finite perimeter $E, F\subset {\mathbb{R}^n}$. Moreover, $$\label{gammacts}
\underset{j \to \infty}\lim {\gamma}_{{{\Phi}}_j}(E_j) = {\gamma}_{{{\Phi}}}(E).$$
*Proof of $(1)$:* From the divergence theorem and the characterization $f(\nu) = \sup \{ x\cdot \nu : f_*(x) \leq 1\}$, one finds that the surface energy of a set $E$ is the anisotropic total variation of its characteristic function $\chi_E$: $$\label{TV}
{{\Phi}}_j(E_j) = TV_{f_{j}}(\chi_{E_j}) : = \sup \Big\{ \int_{E_j} {{\rm div \,}}T\, dx\ \big{|} \ \ T\in C^1_c({\mathbb{R}^n}, {\mathbb{R}^n}),\
f_*^j(T) \leq 1 \Big\}.$$ Let $T \in
C^1_c({\mathbb{R}^n}, {\mathbb{R}^n})$ be a vector field such that $f_*(T) \leq 1$ for all $x\in {\mathbb{R}^n}$. Then we have $$\begin{aligned}
\int_E {{\rm div \,}}T \,dx & = \lim_{j\to \infty} \int_{E_j} {{\rm div \,}}T\, dx = \lim_{j \to \infty} \|f_*^j (T)\|_{L^{\infty}({\mathbb{R}^n})} \int_{E_j} {{\rm div \,}}S_j\, dx \leq \underset{j\to\infty}{\liminf} \mbox{ }{{\Phi}}_j(E_j),\end{aligned}$$ where we take $S_j =
T/
\|f_*^j(T)\|_{L^{\infty}({\mathbb{R}^n})}.$ Taking the supremum over $\{T\in C^1_c({\mathbb{R}^n},{\mathbb{R}^n}) : f_*(T) \leq 1\}$, we obtain the result.\
*Proof of $(2)$:* By , $$\begin{aligned}
{\gamma}_{{{\Phi}}}(E) - {\gamma}_{{{\Phi}}}(F)& \leq \int_E \frac{dx}{f_*(x-y_E)} - \int_F \frac{dx}{f_*(x-y_E)} \leq \int_{E\Delta F } \frac{dx}{f_*(x-y_E)} .\end{aligned}$$ Letting $r$ be such that $|rK| = |E\Delta F|$ and recalling , we have $$\label{tozero}
\int_{E\Delta F } \frac{dx}{f_*(x-y_E)} \leq \int_{rK} \frac{dx}{f_*(x)} ={\gamma}_{{{\Phi}}}(rK) = \frac{{{\Phi}}(rK)}{n-1} = \frac{n|K|r^{n-1}}{n-1} = \frac{n|K|}{n-1} |E\Delta F|^{1/n'}.$$ Thus ${\gamma}_{{{\Phi}}}(E) - {\gamma}_{{{\Phi}}}(F) \leq \frac{n|K|}{n-1} |E\Delta F|^{1/n'}$. The analogous argument holds for ${\gamma}_{{{\Phi}}}(F) - {\gamma}_{{{\Phi}}}(E),$ implying the Hölder continuity of ${\gamma}_{{{\Phi}}}$.
For the second equation, we note that if $f^j \to f$ locally uniformly, then $f_*^j \to f_* $ locally uniformly and $M_{{{\Phi}}_j} \to M_{{{\Phi}}}$. The triangle inequality gives $$\begin{aligned}
|{\gamma}_{{{\Phi}}_j}(E_j) - {\gamma}_{{{\Phi}}}(E) | \leq |{\gamma}_{{{\Phi}}_j}(E_j) - {\gamma}_{{{\Phi}}}(E_j) | + |{\gamma}_{{{\Phi}}}(E_j) - {\gamma}_{{{\Phi}}}(E)|.
\end{aligned}$$ The second term goes to zero by the Hölder continuity that we have just shown. To bound the first term, let $y_{E_j}$ be a center of $E_j$ with respect to the surface energy ${{\Phi}}_j$. If ${\gamma}_{{{\Phi}}_j}(E_j) \geq {\gamma}_{{{\Phi}}}(E_j) $, then $$\begin{aligned}
0\leq {\gamma}_{{{\Phi}}_j}(E_j) - {\gamma}_{{{\Phi}}}(E_j) &\leq \int_{E_j} \frac{1}{f_*^j (x-y_{E_j})} - \frac{1}{f_*(x-y_{E_j})} dx
= \int_{E_j + y_{E_j}} \frac{1}{f_*^j(x)} - \frac{1}{f_*(x) } dx\\
&= \int_{{\mathbb{R}^n}}\chi_{(E_j + y_{E_j})\setminus B_{{\epsilon}}(0)} \bigg(\frac{1}{f_*^j(x)} - \frac{1}{f_*(x) } \bigg)dx + \int_{B_{{\epsilon}}(0)} \frac{1}{f_*^j(x)} - \frac{1}{f_*(x) } dx .
\end{aligned}$$ For ${\epsilon}>0$ fixed, the first integral goes to zero as $j\to \infty$. For the second integral, we have $$\int_{B_{{\epsilon}}(0)} \frac{1}{f_*^j(x) } + \frac{1}{f_*(x)} \,dx \leq \int_{B_{{\epsilon}}(0)} \frac{M_{{{\Phi}}_j} + M_{{{\Phi}}}}{|x|}\, dx \leq C {\epsilon}^{n-1}.$$ Taking ${\epsilon}\to 0$, we conclude that $ {\gamma}_{{{\Phi}}_j}(E_j) - {\gamma}_{{{\Phi}}}(E_j) \to0$ as $j \to \infty$. The case where ${\gamma}_{{{\Phi}}_j}(E_j) \leq {\gamma}_{{{\Phi}}}(E_j) $ is analogous.
\[RMK: LSC\][With sequences as in the hypothesis of Proposition \[convergence\] above, ${\beta_{{{\Phi}}}}$ has the following lower semicontinuity property: $${\beta_{{{\Phi}}}}(E)\leq \underset{j\to\infty}{\liminf}\,\beta_{{{\Phi}}_j}(E_j).$$ This follows immediately from parts $(1)$ and $(2)$ of Proposition \[convergence\] and the decomposition in .]{}
\[3.1\] For every ${\epsilon}>0$, there exists $\eta>0$ such that if $|F\Delta K | \leq \eta$, then $|y_F|<{\epsilon}$ for any center $y_F$ of $F$.
Suppose $|K \Delta F_j| \to 0$. By the triangle inequality, $$\begin{aligned}
\int_K \frac{dx}{f_*(x)}
& \leq \bigg|\int_{K} \frac{dx}{f_*(x)} - \int_{F_j} \frac{dx}{f_*(x-y_{F_j})} \bigg|+\bigg| \int_{F_j} \frac{dx}{f_*(x -y_{F_j})} - \int_{K} \frac{dx}{f_*(x -y_{F_j})} \bigg|
+\int_{K} \frac{dx}{f_*(x -y_{F_j})}.
\end{aligned}$$ By , the first two terms on the right hand side go to zero as $j\to \infty,$ implying that $$\int_K \frac{dx}{f_*(x)}\leq \underset{j \to \infty}{\lim} \int_{K} \frac{dx}{f_*(x -y_{F_j})}.$$ Because $K$ has unique center $y_K = 0$, we conclude that $|y_{F_j} |\to 0.$
We now introduce the relative surface energy and the anisotropic coarea formula. Given an open set $A$ and a set of finite perimeter $E$, the anisotropic surface energy of $E$ *relative to $A$* is defined by $${{\Phi}}(E;A) = \underset{\partial^* E \cap A}{\int} f(\nu_E(x))\, d{\mathcal{H}^{n-1}}(x).$$ For a Lipschitz function $u:{\mathbb{R}^n}\to \mathbb{R}$ and an open set $E$, the anisotropic coarea formula states that $$\int_E f(-\nabla u(x))\,dx = \int_0^{\infty}{{\Phi}}(\{ u>r\};E)\, dr.$$ The anisotropic coarea formula is proved in the same way as the coarea formula (see, for instance, [@maggi2012sets Theorem 13.1]), replacing the Euclidean norm with $f$ and $f_*$ and using . When $u$ is bounded by a constant $C$ on $E$, then applying the anisotropic coarea formula to $w = C-u$ yields $$\begin{aligned}
\int_E f(\nabla u(x))\,dx&= \int_E f(-\nabla w(x))\, dx =\int_0^C {{\Phi}}(\{C-u>r\}; E) \,dt\\
&= \int_0^C {{\Phi}}(\{u <C-r\} ; E) \,dr = \int_0^C {{\Phi}}(\{u<r\} ; E) \, dr\end{aligned}$$ Moreover, approximating by simple functions, we may produce a weighted version: $$\int_E f(\nabla u(x)) g(f_*(x)) \,dx= \int_0^{\infty} {{\Phi}}(\{u<r\};E) g(r) \,dr$$ whenever $g: \mathbb{R} \to[0,\infty] $ is a Borel function. We will frequently use this weighted version with $u(x) = f_*(x)$, $E$ a bounded set, and $g(r) = \frac{1}{r},$ which, using (\[equal1\]), gives $$\label{awcoarea}
\int_E \frac{dx}{f_*(x)} = \int_0^{\infty} \frac{{{\Phi}}( \{f_*(x) <r\} ; E )}{r}\, dr= \int_0^{\infty} \frac{{{\Phi}}( rK; E )}{r}\, dr.$$
We conclude this section with the following Poincaré-type inequality, which shows that ${\beta_{{{\Phi}}}}(E)$ controls $\alpha_{{{\Phi}}}(E)$ for all sets $E$ of finite perimeter.
\[prop1\] \[poincare\] There exists a constant $C(n)$ such that if $E$ is a set of finite perimeter with $0<|E|<\infty$, then $$\label{bd}
\alpha_{{{\Phi}}}(E) + \delta_{{{\Phi}}}(E)^{1/2} \leq C(n) {\beta_{{{\Phi}}}}(E).$$
We follow the proof of the analogous result for the perimeter in [@fuscojulin11]. Due to the scaling and translation invariance of $\alpha_{{{\Phi}}}, {\beta_{{{\Phi}}}}$, and $\delta_{{{\Phi}}}$, we may assume that $|E| = |K|=1$ and that $E$ has center zero. We have $${\gamma}_{{{\Phi}}}(K) - {\gamma}_{{{\Phi}}}(E)= \int_K \frac{dx}{f_*(x)}-\int_E \frac{dx}{f_*(x) }
=\int_{K\setminus E} \frac{dx}{f_*(x)}-\int_{E\setminus K} \frac{dx}{f_*(x)} .$$ Therefore, adding and subtracting ${{\Phi}}(K)/n=(n-1){\gamma}_{{{\Phi}}}(K)/n$ in , we have $$\begin{aligned}
{\beta_{{{\Phi}}}}(E)^2 &
=\delta_{{{\Phi}}}(E)
+\frac{n-1}{n}\left( \int_{K\setminus E} \frac{dx}{f_*(x)} -
\int_{E\setminus K} \frac{dx}{f_*(x)} \right).\end{aligned}$$ We want to bound the final two integrals from below by $\alpha_{{{\Phi}}}(E)^2$. To this end, we let $a:= |E\setminus K| = |K\setminus E|$ and define the $K$-annuli $A_{R,1} = K_R\setminus K$ and $A_{1,r}=K\setminus K_r$, where $R> 1 >r$ are chosen such that $|A_{R,1} | =|A_{1,r} | =a$. In particular, $R = (1+a)^{1/n} $ and $r= (1-a)^{1/n}.$ By (\[maxgamma\]) and (\[awcoarea\]), $$\begin{aligned}
\int_{K\setminus E} \frac{dx}{f_*(x) }
\geq \int_{A_{1,r}} \frac{dx}{f_*(x) }
= \int_r^1 \frac{{{\Phi}}(sK)}{s}\,ds
= \int_r^1 ns^{n-2}\, ds = \frac{n}{n-1} [1-r^{n-1}]\end{aligned}$$ and $$\begin{aligned}
\int_{E\setminus K} \frac{dx}{f_*(x) }
\leq \int_{A_{R,1}} \frac{dx}{f_*(x)}
= \int_1^R\frac{{{\Phi}}(sK)}{s} \,ds
= \int_1^R ns^{n-2} \,ds
= \frac{n}{n-1}[R^{n-1} - 1].\end{aligned}$$ Subtracting the second from the first, we have $$\frac{n-1}{n} \left(\int_{K\setminus E} \frac{dx}{f_*(x)} -\int_{E\setminus K} \frac{dx}{f_*(x)}\right)\geq 2 - r^{n-1} - R^{n-1}.$$
The function $g(t) = (1+t)^{1/n'}$ is function is strictly concave, with $\frac{1}{2}(g(t) + g(s) ) \leq g(\frac{t}{2} + \frac{s}{2}) - C|t-s|^2$, and therefore $2 - [ (1+a )^{1/n'} + (1-a)^{1/n'}] \geq 8C|a|^2.$ Thus $$\begin{aligned}
{\beta_{{{\Phi}}}}(E)^2 \geq \delta_{{{\Phi}}}(E) + [ 2 - (1-a)^{1/n'} - (1+a)^{1/n'}] &\geq \delta_{{{\Phi}}}(E) + 8C|a|^2 = \delta_{{{\Phi}}} +2C \left(|E\setminus K| + |K \setminus E|\right)^2\\
& = \delta_{{{\Phi}}}(E) +2C |K\Delta E|^2\geq \delta_{{{\Phi}}}(E) +2C\alpha_{{{\Phi}}}(E)^2.\end{aligned}$$
Stability for General Anisotropic Surface Energy ${{\Phi}}$ {#generalsection}
===========================================================
In this section, we prove Theorem \[th1\]. We begin by introducing a few lemmas that are needed the proof. The first allows us to reduce the problem to sets contained in some fixed ball.
\[compact\] There exist constants $R_0>0$ and $C>0$ depending only on $n$ and $M_{{{\Phi}}}$ such that, given a set of finite perimeter $E$ with $|E|= |K|$, we may find a set $E'$ such that $|E' | = |K|, \ E'\subset B_{R_0}$, and $$\label{BOUND}{\beta_{{{\Phi}}}}(E)^2
\leq {\beta_{{{\Phi}}}}(E')^2 + C \delta_{{{\Phi}}}(E), \mbox{ } \qquad
\delta_{{{\Phi}}}(E') \leq C \delta_{{{\Phi}}}(E).$$
A simple adaptation of the proof of [@maggi2008some Theorem $4.1$] ensures that we may find constants $\delta_0, C_0,$ $C_1$, and $\tilde{R}_0$ depending on $n$ and $M_{{{\Phi}}}$ such that $C_0\delta_0<1/2$ and the following holds: if $\delta_{{{\Phi}}}(E) \leq \delta_0,$ then there exists a set $\tilde{E}\subset E$ such that $\tilde{E} \subset B_{\tilde{R}_0}$ and $$\label{cut}
|\tilde{E}| \geq |K|( 1 - C_1 \delta_{{{\Phi}}}(E)), \qquad {{\Phi}}(\tilde{E}) \leq {{\Phi}}(E) + C_0\delta_{{{\Phi}}}(E) |E|^{1/n'}.$$
If $\delta_{{{\Phi}}}(E) > \delta_0$, then $${\beta_{{{\Phi}}}}^2 (E)
\leq \frac{{{\Phi}}(E)}{n|K|}
= \delta_{{{\Phi}}}(E)+1
\leq \frac{1+\delta_0}{\delta_0} \delta_{{{\Phi}}}(E).$$ Simply taking $E' = K$, we have $\delta_{{{\Phi}}}(E')
\leq \delta_{{{\Phi}}}(E)$ and ${\beta_{{{\Phi}}}}(E)^2 \leq \frac{1+\delta_0}{\delta_0} \delta_{{{\Phi}}}(E)$, proving .
On the other hand, if $\delta_{{{\Phi}}}(E) \leq \delta_0$, let $E' = r \tilde{E}$ with $r\geq 1$ such that $|E'|=|r\tilde{E}|=|E|.$ By , $$\label{Bounded set}
\begin{split}
{\beta_{{{\Phi}}}}(E)^2 -{\beta_{{{\Phi}}}}(E')^2
& = \frac{{{\Phi}}(E) - {{\Phi}}(E')}{n|K|}
+ \frac{n-1}{n|K|} \left( {\gamma}_{{{\Phi}}}(E' ) - {\gamma}_{{{\Phi}}}(E) \right) \\
&\leq \delta_{{{\Phi}}}(E)
+ \frac{n-1}{n|K|} \left(r^{n-1} {\gamma}_{{{\Phi}}}(\tilde{E}) - {\gamma}_{{{\Phi}}}(E) \right).
\end{split}$$ Since $\tilde{E}\subset E$, ${\gamma}_{{{\Phi}}}(\tilde{E} ) \leq {\gamma}_{{{\Phi}}}(E)$, which implies that $$\begin{aligned}
\frac{n-1}{n|K|} \left(r^{n-1} {\gamma}_{{{\Phi}}}(\tilde{E}) - {\gamma}_{{{\Phi}}}(E) \right)\leq
\frac{n-1}{n|K|} (r^{n-1} -1) {\gamma}_{{{\Phi}}}(E).\end{aligned}$$ By and the fact that ${\gamma}_{{{\Phi}}} (K)= n|K|/(n-1)$, $$\begin{aligned}
\frac{n-1}{n|K|} (r^{n-1} -1) {\gamma}_{{{\Phi}}}(E)& \leq \frac{n-1}{n|K|}(r^{n-1} -1) {\gamma}_{{{\Phi}}}(K)
=r^{n-1} -1,\end{aligned}$$ and since $r\geq 1,$ $$r^{n-1} -1\leq r^n-1=\frac{ |E| - |\tilde{E}|}{|\tilde{E}|}.$$ The first part of $(\ref{cut})$ implies that $$\frac{|E| - |\tilde{E}| }{|\tilde{E}|} \leq \frac{C_1 \delta_{{{\Phi}}}(E)}{1 -C_1 \delta_{{{\Phi}}}(E)}\leq \frac{C_1}{1-C_1\delta_0} \delta_{{{\Phi}}}(E).$$ We have therefore shown that $$\frac{n-1}{n|K|} \left(r^{n-1} {\gamma}_{{{\Phi}}}(\tilde{E}) - {\gamma}_{{{\Phi}}}(E) \right) \leq \frac{C_1}{1-C_1\delta_0} \delta_{{{\Phi}}}(E);$$ this together with concludes the proof of the first claim in .
In the direction of the second claim in , the first and second parts of respectively imply that $${{\Phi}}(E') = r^{n-1} {{\Phi}}(\tilde{E})
\leq \frac{{{\Phi}}(\tilde{E}) }{(1-C_1 \delta_{{{\Phi}}}(E))^{1/n'}} \leq \frac{{{\Phi}}(E)+ C_0 \delta_{{{\Phi}}}(E) |E|^{1/n'}}{(1-C_1 \delta_{{{\Phi}}}(E))^{1/n'}}.$$ A Taylor expansion in $\delta_{{{\Phi}}}(E)$ of the right hand side shows that $$\begin{aligned}
{{\Phi}}(E') & \leq {{\Phi}}(E)+ C_0 \delta_{{{\Phi}}}(E) |E|^{1/n'} +\frac{n-1}{n}C_1 \delta_{{{\Phi}}}(E){{\Phi}}(E)+ O(\delta_{{{\Phi}}}(E)^2)
\\
&
\leq {{\Phi}}(E) + C \delta_{{{\Phi}}}(E) {{\Phi}}(E)\end{aligned}$$ for $\delta_0$ chosen sufficiently small. Thus $$\delta_{{{\Phi}}}(E') = \frac{{{\Phi}}(E') - {{\Phi}}(K)}{n|K|} \leq
\frac{{{\Phi}}(E') - {{\Phi}}(E)}{n|K|}
\leq \frac{C{{\Phi}}(E) \delta_{{{\Phi}}}(E)}{n|K|} \leq C \delta_{{{\Phi}}}(E),$$ since ${{\Phi}}(E) \leq {{\Phi}}(K) +n|K| \delta_0$. Finally, since $\tilde{E} \subset B_{\tilde{R}_0}$ and $E' = r \tilde{E}$ with $r \leq 1/(1-C_1\delta_0)^{1/n}$, we have $E' \subset B_{R_0}$ for $R_0 = r \tilde{R}_0$.
Let us now consider the functional $$\label{Q}Q(E) = {{\Phi}}(E) + \frac{|K|m_{{{\Phi}}}}{8M_{{{\Phi}}}}\left| {\beta_{{{\Phi}}}}(E)^2 - {\epsilon}^2\right| + \Lambda\big| |E| - |K| \big| ,$$ with $0<{\epsilon}<1$ and $\Lambda>0$.
\[LSC\] A minimizer exists for the problem $$\label{Minimization}\min \left\{ Q(E) \, : \, E \subset B_{R_0}\right\}$$ for $\Lambda > 4n$ and ${\epsilon}>0 $ sufficiently small. Moreover, any minimizer $F$ satisfies $$\label{Perimeter Bounds}
|F| \geq \frac{|K|}{2}, \qquad {{\Phi}}(F) \leq 2n|K|.$$
Let $\overline{Q} = \inf\{Q(E) : E \subset B_{R_0} \}$, and let $\{F_j\}$ be a sequence such that $Q(F_j) \to \overline{Q}$. Since $F_j \subset B_{R_0}$ and ${{\Phi}}(F_j)<2\overline{Q}$ for $j$ large enough, up to a subsequence, $F_j \to F$ in $L^1$ for some $F \subset B_{R_0}$. The lower semicontinuity of ${{\Phi}}$ (Proposition \[convergence\](1)) ensures that ${{\Phi}}(F) <\infty$.
We first show that $|F| \geq\frac{|K|}{2}$. For any $\eta>0$, $Q(F_j) \leq \overline{Q} + |K|\eta$ for $j$ sufficiently large. Furthermore, $\overline{Q} \leq Q(K) = {{\Phi}}(K) + \frac{{\epsilon}^2 |K|{m_{{{\Phi}}}}}{8{M_{{{\Phi}}}}}$, so $$\big| |F_j| - |K| \big| \leq \frac{1}{\Lambda}\left( {{\Phi}}(K) +|K|\eta +\frac{{\epsilon}^2 |K|{m_{{{\Phi}}}}}{8{M_{{{\Phi}}}}} \right)
= \frac{|K|}{\Lambda} \left(n + \eta +\frac{{\epsilon}^2{m_{{{\Phi}}}}}{8{M_{{{\Phi}}}}} \right)\leq \frac{|K|}{2}$$ for ${\epsilon}$ and $\eta$ sufficiently small. Therefore $|F_j|\geq \frac{|K|}{2},$ implying that $|F| \geq \frac{|K|}{2}$ as well.
We now show that $\liminf Q(F_j) \geq Q(F),$ so $F$ is a minimizer. Recalling (\[betaform\]), we have $$\begin{aligned}
Q(F_j)
&= {{\Phi}}(F_j) + \frac{|K|{m_{{{\Phi}}}}}{8{M_{{{\Phi}}}}} \bigg| \frac{{{\Phi}}(F_j)-(n-1){\gamma}_{{{\Phi}}}(F_j) }{n|K|^{1/n}|F_j|^{1/n'}} - {\epsilon}^2 \bigg| + \Lambda \big| |F_j| - |K|\big|
\\
& \geq
{{\Phi}}(F_j ) +\frac{|K|{m_{{{\Phi}}}}}{8{M_{{{\Phi}}}}} \bigg| \frac{{{\Phi}}(F)-(n-1){\gamma}_{{{\Phi}}}(F_j)}{n|K|^{1/n}|F_j|^{1/n'}} - {\epsilon}^2 \bigg|
-\frac{|K|{m_{{{\Phi}}}}}{8{M_{{{\Phi}}}}} \bigg|\frac{ {{\Phi}}(F_j) - {{\Phi}}(F)}{n|K|^{1/n}|F_j|^{1/n'}} \bigg| + \Lambda \big| |F_j|- |K|\big|.
\end{aligned}$$ Let $a={\liminf}\mbox{ }{{\Phi}}(F_j)$. Up to a subsequence, we may take this limit infimum to be a limit. By the lower semicontinuity of ${{\Phi}}$, $a \geq {{\Phi}}(F)$. Furthermore, ${\gamma}_{{{\Phi}}}$ is continuous by Proposition \[convergence\]$(2)$, so $$\begin{aligned}
\underset{j\to \infty}{\liminf}\mbox{ } Q(F_j)& \geq Q(F) + \left(a-{{\Phi}}(F)\right) - \frac{|K|^{1/n'}{m_{{{\Phi}}}}}{8n|F|^{1/n'}{M_{{{\Phi}}}}}\left|a - {{\Phi}}(F)\right| \\
& = Q(F) + \left(a-{{\Phi}}(F)\right)\Big( 1 - \frac{|K|^{1/n'}{m_{{{\Phi}}}}}{8n|F|^{1/n'}{M_{{{\Phi}}}}}\Big)\\
&\geq Q(F) + \left(a-{{\Phi}}(F)\right)\Big(1 - \frac{2^{1/n'}{m_{{{\Phi}}}}}{8n{M_{{{\Phi}}}}}\Big)
\geq Q(F).\end{aligned}$$ Finally, ${\epsilon}<1$ and therefore ${{\Phi}}(F) \leq Q(F) \leq Q(K) \leq 2n|K|.$
The following lemma shows that a minimizer of satisfies uniform density estimates.
\[UDE\] Suppose $F$ is a minimizer of $Q(E)$ as defined in (\[Q\]) among all sets $E\subset B_{R_0}$. Then there exist $r_0>0$ depending on $n, \Lambda,$ and $|K|$ and $0<c_0<1/2$ depending on $n$ and $\Lambda$ such that for any $x \in \partial^* F$ and for any $r<r_0$, $$\label{UDES}
\frac{c_0{m_{{{\Phi}}}}^n}{{M_{{{\Phi}}}}^n} \, \omega_n r^n \leq |B_r(x) \cap F| \leq \bigg(1-\frac{c_0 {m_{{{\Phi}}}}^n}{ {M_{{{\Phi}}}}^n} \bigg)\, \omega_n r^n.$$
We follow the standard argument for proving uniform density estimates for minimizers of perimeter functionals; see, for example, [@maggi2012sets Theorem $16.14$]. The only difficulty arises when handling the term $\frac{|K|{m_{{{\Phi}}}}}{8{M_{{{\Phi}}}}} |{\beta_{{{\Phi}}}}(E)^2 - {\epsilon}^2 |$ in $Q(E)$, as it scales like the surface energy.
For any $x_0 \in \partial^* F$, let $r<r_0$, where $r_0$ is to be chosen later in the proof and $r$ is chosen such that $$\label{lineup}{\mathcal{H}^{n-1}}(\partial^* F \cap \partial B_r(x_0)) =0.$$ This holds for almost every $r>0$. Note that if holds for almost every $r<r_0$, then it must hold for all $r<r_0$ by continuity; it is therefore enough to consider $r$ such that holds. Let $G = F \setminus B_r(x_0)$. For simplicity, we will use the notation $B_r$ for $B_r(x_0).$ Because $F$ minimizes $Q$, $${{\Phi}}(F) + \frac{|K|{m_{{{\Phi}}}}}{8{M_{{{\Phi}}}}}
\left|{\beta_{{{\Phi}}}}(F)^2 - {\epsilon}^2 \right|
+ \Lambda \big| |F| - |K| \big|
\leq {{\Phi}}(G)
+\frac{|K|{m_{{{\Phi}}}}}{8{M_{{{\Phi}}}}} \left|{\beta_{{{\Phi}}}}(G)^2 - {\epsilon}^2\right|
+ \Lambda \big| |G| - |K| \big| ,$$ and so rearranging and using the triangle inequality, we have $${{\Phi}}(F) \leq {{\Phi}}(G)
+ \frac{|K|{m_{{{\Phi}}}}}{8{M_{{{\Phi}}}}} \left| {\beta_{{{\Phi}}}}(F)^2 - {\beta_{{{\Phi}}}}(G)^2 \right|
+ \Lambda |F \cap B_r|.$$ We subtract ${{\Phi}}(F; {\mathbb{R}^n}\setminus B_r)$ from both sides; this is the portion of the surface energy where $\partial^*F$ and $\partial^*G$ agree. We obtain $$\begin{aligned}
\label{udeB}
{{\Phi}}(F ; B_r) &
\leq
\int_{\partial B_r \cap F} f(\nu_{B_r})\, d{\mathcal{H}^{n-1}}+ \frac{|K|{m_{{{\Phi}}}}}{8{M_{{{\Phi}}}}} \left|{\beta_{{{\Phi}}}}(F)^2-{\beta_{{{\Phi}}}}(G)^2 \right|
+ \Lambda |F \cap B_r|.
\end{aligned}$$ Indeed, this holds because implies that $${{\Phi}}(G) = {{\Phi}}(F; {\mathbb{R}^n}\setminus B_r) + \int_{\partial B_r \cap F} f(\nu_{B_r}) \,d{\mathcal{H}^{n-1}}.$$
We must control the term $\frac{|K|{m_{{{\Phi}}}}}{8{M_{{{\Phi}}}}} \left|{\beta_{{{\Phi}}}}(F)^2-{\beta_{{{\Phi}}}}(G)^2 \right|$ and require a sharper bound than the one obtained using Hölder continuity of ${\gamma}_{{{\Phi}}}$ shown in Proposition \[convergence\](2). Indeed, we must show that the only contributions of this term are perimeter terms that match those in and terms that scale like the volume and thus behave as higher order perturbations. We have $$\begin{aligned}
\left| {\beta_{{{\Phi}}}}(F)^2 - {\beta_{{{\Phi}}}}(G)^2 \right| & = \frac{1}{n|K|^{1/n}} \left| \frac{{{\Phi}}(F) - (n-1) {\gamma}_{{{\Phi}}}(F)}{|F|^{1/n'} } - \frac{ {{\Phi}}(G) - (n-1) {\gamma}_{{{\Phi}}}(G)}{|G|^{1/n'}} \right| \\
& \leq \frac{2{{\Phi}}(F)}{n|K|^{1/n}} \big||F|^{-1/n'} -|G|^{-1/n'} \big|
+ \frac{\left| {{\Phi}}(F) - {{\Phi}}(G) \right| +(n-1) \left| {\gamma}_{{{\Phi}}}(F) - {\gamma}_{{{\Phi}}}(G) \right|}{n|K|^{1/n}|G|^{1/n'}}.\end{aligned}$$ The function $v(z)= 1 - (1-z)^{1/n'}$ is convex and increasing with $v(1) = 1$, hence $v(z) \leq z$ for $z\in[0,1]$. Thus, as $|G| = |F| - |F\cap B_r|,$ $$\label{aaa}\big| |F|^{-1/n'} - |G|^{-1/n'} \big| = |G|^{-1/n'}\left(1 - \left( 1- \frac{|F\cap B_r|}{|F|}\right)^{1/n'}\right) \leq \frac{|F\cap B_r|}{|G|^{1/n'}|F| }.$$ Since $2|F| \geq |K|$ by , $4|G| \geq |K| $ for $r_0$ sufficiently small depending on $n$, so the right hand side of is bounded by $8|K|^{-1-1/n'} |F\cap B_r| .$ The coefficient $\frac{2{{\Phi}}(F)}{n|K|^{1/n}}$ is bounded by $4|K|^{1/n'}$ thanks to , so $$\label{D1}\frac{2{{\Phi}}(F)}{n|K|^{1/n}} \big||F|^{-1/n'} -|G|^{-1/n'} \big|
\leq 32 |K|^{-1}|F \cap B_r|.$$ Therefore, by and again using the facts that $4|G| \geq |K|$, $2|F|\geq |K|$, and ${m_{{{\Phi}}}}/{M_{{{\Phi}}}}\leq 1,$ we have shown that $$\label{bbound}
\frac{|K| {m_{{{\Phi}}}}}{8{M_{{{\Phi}}}}}| {\beta_{{{\Phi}}}}(F)^2 - {\beta_{{{\Phi}}}}(G)^2| \leq 4|F\cap B_r|+
\frac{|{{\Phi}}(F) -{{\Phi}}(G)|}{2n} + \frac{{m_{{{\Phi}}}}}{{M_{{{\Phi}}}}}\frac{n-1}{2n} |{\gamma}_{{{\Phi}}}(F) - {\gamma}_{{{\Phi}}}(G)|.$$ For the term $\left| {{\Phi}}(F) - {{\Phi}}(G) \right| $, using , we have $$\label{D2}\left| {{\Phi}}(F) - {{\Phi}}(G) \right| = \left| \int_{{\partial}^* F} f(\nu_F) \,d{\mathcal{H}^{n-1}}- \int_{{\partial}^* G} f(\nu_G) \,d {\mathcal{H}^{n-1}}\right|\leq {{\Phi}}(F; B_r) + \int_{{\partial}B_r \cap F} f(\nu_{B_r})\, d{\mathcal{H}^{n-1}},$$ using and the fact that ${\partial}^*F$ and ${\partial}^*G$ agree outside of $B_r$. Similarly, for the term $\left| {\gamma}_{{{\Phi}}}(F) - {\gamma}_{{{\Phi}}}(G) \right|$, when ${\gamma}_{{{\Phi}}}(F) \geq {\gamma}_{{{\Phi}}}(G)$, thanks to we have $$\begin{aligned}
{\gamma}_{{{\Phi}}}(F) - {\gamma}_{{{\Phi}}}(G) &
\leq \int_{{\partial}^* F } \frac{(x-y_F)\cdot \nu_F(x) }{f_*(x-y_F)}\, d{\mathcal{H}^{n-1}}- \int_{{\partial}^* G} \frac{(x-y_F)\cdot \nu_G(x)}{f_*(x-y_F)}\, d{\mathcal{H}^{n-1}}\\
&\leq \frac{M_{{{\Phi}}}}{m_{{{\Phi}}}} \bigg( {{\Phi}}(F; B_r) + \int_{{\partial}B_r \cap F} f(\nu_{B_r})\, d{\mathcal{H}^{n-1}}\bigg).\end{aligned}$$ The analogous inequality holds when ${\gamma}_{{{\Phi}}}(G) \geq {\gamma}_{{{\Phi}}}(F)$, so $$\label{D3}
| {\gamma}_{{{\Phi}}}(F) - {\gamma}_{{{\Phi}}}(G)|\leq \frac{M_{{{\Phi}}}}{m_{{{\Phi}}}} \bigg( {{\Phi}}(F; B_r) + \int_{{\partial}B_r \cap F} f(\nu_{B_r})\, d{\mathcal{H}^{n-1}}\bigg).$$ Combining , , and , we have shown $$\label{udeC}
\frac{|K| {m_{{{\Phi}}}}}{8{M_{{{\Phi}}}}}
\left| {\beta_{{{\Phi}}}}(F)^2 - {\beta_{{{\Phi}}}}(G)^2 \right| \leq
4 |F \cap B_r|+ \frac{1}{2} \bigg( {{\Phi}}(F; B_r) + \int_{{\partial}B_r \cap F} f(\nu_{B_r})\, d{\mathcal{H}^{n-1}}\bigg).$$ Combining (\[udeB\]) and (\[udeC\]) and rearranging, we have $$\begin{aligned}
\frac{1}{2} {{\Phi}}( F; B_r) &
\leq \frac{3}{2} \int_{\partial B_r \cap F} f(\nu_{B_r})\,d{\mathcal{H}^{n-1}}+ (4+ \Lambda )\left| F \cap B_r\right|.\end{aligned}$$ Proceeding in the standard way, we add the term $\frac{1}{2}\int_{\partial B_r \cap F} f(\nu_{B_r} )\,d{\mathcal{H}^{n-1}}$ to both sides, which gives $$\begin{aligned}
\frac{1}{2} {{\Phi}}(F\cap B_r) &
\leq 2 \int_{\partial B_r\cap F} f(\nu_{B_r })\,d{\mathcal{H}^{n-1}}+(4+ \Lambda) \left|F \cap B_r\right|.\end{aligned}$$ By the Wulff inequality, ${{\Phi}}(F\cap B_r) \geq n|K|^{1/n} |F\cap B_r|^{1/n'}$, and for $r_0$ small enough depending on $n, \Lambda$, and $|K|$, we may absorb the last term on the right hand side to obtain $$\begin{aligned}
\label{density}
\frac{n|K|^{1/n}|F\cap B_r|^{1/n'}}{4}
\leq 2 \int_{\partial B_r\cap F} f(\nu_{B_r })\,d{\mathcal{H}^{n-1}}.\end{aligned}$$ Let $u(r) = |F \cap B_r|$, and thus $u'(r) = {\mathcal{H}^{n-1}}(\partial B_r \cap F)$, so the right hand side above is bounded by $2 M_{{{\Phi}}} u'(r)$. Furthermore, $|K|^{1/n}\geq {m_{{{\Phi}}}}$, so yields the differential inequality $$\frac{ n{m_{{{\Phi}}}}}{8{M_{{{\Phi}}}}} \leq u'(r) u(r)^{-1/n'} = n (u^{1/n})'.$$ Integrating these quantities over the interval $[0,r]$, we get $$\frac{{m_{{{\Phi}}}}r }{8{M_{{{\Phi}}}}}
\leq u(r)^{1/n} = |B_r \cap F|^{1/n},$$ and taking the power $n$ of both sides yields the lower density estimate. The upper density estimate is obtained by applying an analogous argument, using $G = F\cup B_r(x_0)$ as a comparison set for $x_0 \in \partial^* F$ and $r<r_0$ satisfying .
The following lemma is a classical argument showing that a set that is close to $K $ in $L^1$ and satisfies uniform density estimates is close to $K$ in an $L^{\infty}$ sense.
\[bound\] Suppose that $F$ satisfies uniform density estimates as in . Then there exists $C$ depending on ${m_{{{\Phi}}}}/{M_{{{\Phi}}}}$, $n$, and $\Lambda$ such that $${\rm{hd}}({\partial}F , {\partial}K)^n \leq C |F\Delta K|,$$ where ${\rm{hd}}(\cdot, \cdot) $ is the Hausdorff distance between sets. In particular, for any $\eta>0$, there exists ${\epsilon}>0$ such that if $|F\Delta K | < {\epsilon}$, then $K_{1-\eta} \subset F \subset K_{1+\eta},$ where $K_{a} = aK.$
Let $d = {\rm{hd}}({\partial}F , {\partial}K).$ Then there is some $x \in {\partial}F$ such that either $B_d(x) $ is contained entirely in the complement of $K$ or $B_d(x)$ is entirely contained in $K$. If the first holds, then the lower density estimate in implies that $$|F\Delta K| \geq |F\cap B_d(x) | \geq \frac{c_0{m_{{{\Phi}}}}^n}{{M_{{{\Phi}}}}^n} d^n,$$ while if the second holds, then the upper density estimate in implies that $$|F \Delta K | \geq |B_d(x) \setminus F| \geq \frac{c_0{m_{{{\Phi}}}}^n}{{M_{{{\Phi}}}}^n} d^n.$$
We will make use of the following form of the Wulff inequality without a volume constraint.
\[isop\] Let $R_0>$ diam$(K)$ and $\Lambda >n$. Up to translation, the Wulff shape $K$ is the unique minimizer of the functional $${{\Phi}}(F) + \Lambda\big| |F| - |K|\big|$$ among all sets $F\subset B_{R_0}$.
Let $E$ be a minimizer of ${{\Phi}}(F) + \Lambda\big| |F| - |K|\big| $ among all sets of finite perimeter $F\subset B_{R_0}$; this functional is lower semicontinuous so such a set exists. Comparing with $K$, we find that $$\label{compare no vc}
{{\Phi}}(E) + \Lambda \big| |E| - |K|\big| \leq {{\Phi}}(K) = n|K|.$$ The Wulff inequality implies that $|E| \leq |K|$, and so ${{\Phi}}(E) \geq n|E|^{1/n'} |K|^{1/n} \geq n|E|$. Thus implies that $\Lambda \left( |K| - |E|\right) \leq n \left(|K| - |E|\right).$ Since $\Lambda >n$, it follows that $|E| = |K|$. It follows that $E$ must be a translation of $K$, the unique (up to translation) equality case in the Wulff inequality.
We are now ready to prove Theorem \[th1\].
By , we need only to show that there exists a constant $C= C(n)$ such that $$\label{new1}{\beta_{{{\Phi}}}}(E)^{4n/(n+1)} \leq C\delta_{{{\Phi}}}(E),$$ for any set of finite perimeter $E$ with $0<|E|<\infty.$ By Lemma \[compact\], it suffices to consider sets contained in $B_{R_0}$. Let us introduce the set $$\mathcal{F}_{{\mathcal{N}}} = \Big\{f \ : \ \frac{{M_{{{\Phi}}}}}{{m_{{{\Phi}}}}} \leq \mathcal{N} \Big\}$$ for ${\mathcal{N}}\geq 1$, recalling ${M_{{{\Phi}}}}$ and ${m_{{{\Phi}}}}$ defined in . In Steps $1$-$4$, we prove that, for every ${\mathcal{N}}\geq 1$, there exists a constant $C= C(n,{\mathcal{N}})$ such that holds for any surface energy ${{\Phi}}$ corresponding to a surface tension $f \in \mathcal{F}_{{\mathcal{N}}}$. In Step $5$, we remove the dependence of the constant on ${\mathcal{N}}$.\
*Step 1:* *Set-up.*\
Suppose for the sake of contradiction that (\[new1\]) is false for some ${\mathcal{N}}$. We may then find a sequence of sets $\{E_j\}$ with $E_j \subset B_{R_0}$ and a sequence of surface energies $\{{{\Phi}}_{j}\}$, each ${{\Phi}}_j$ with corresponding surface tension $f^j\in \mathcal{F}_{{\mathcal{N}}}$, Wulff shape $K_j$, and support function $f_*^j$, such that the following holds: $$\begin{aligned}
|E_j | &= |K_j|=1, \nonumber \\
{{\Phi}}_j(E_j) &- {{\Phi}}_j(K_j)\to 0, \nonumber\\
\label{contra}{{\Phi}}_j(E_j)& < {{\Phi}}_j(K_j) +c_1 \beta_{{{\Phi}}_j}(E_j)^{4n/(n+1)},
\end{aligned}$$ where $c_1 = c_1({\mathcal{N}},n)$ is a constant to be chosen later in the proof.
Each $f^j $ is in $ \mathcal{F}_{{\mathcal{N}}}$ and is normalized to make $|K_j| =1$ implying that $\{f^j\}$ is locally uniformly bounded above, and hence, by convexity, locally uniformly Lipschitz. By the Arzela-Ascoli theorem, up to a subsequence, $f^j \to f^{\infty}$ locally uniformly. The uniform convergence ensures that this limit function $f^{\infty}$ is a surface tension in $\mathcal{F}_{{\mathcal{N}}}$. We denote the corresponding surface energy by ${{\Phi}}_{\infty}$, Wulff shape by $K_{\infty}$ and support function by $f_*^{\infty}$. Note that $|K_{\infty}| =1.$
There exists $c({\mathcal{N}})$ such that ${{\Phi}}_j(E) \geq c({\mathcal{N}})P(E)$ for any set of finite perimeter $E$, again thanks to $f^j \in\mathcal{F}_{{\mathcal{N}}}$ and $|K_j|=1$. Then, since ${{\Phi}}_j(E_j) \to n$ (as ${{\Phi}}_j(K_j)=n$), the perimeters are uniformly bounded. Furthermore, $E_j \subset B_{R_0}$, so up to a subsequence, $E_j\to E_{\infty}$ in $L^1$ with $|E_{\infty}| =1$.
Proposition \[convergence\](1) implies that ${{\Phi}}_{\infty}(E_{\infty}) \leq
{\lim} \ {{\Phi}}_j(E_j)= n$, so by the Wulff inequality, $E_{\infty} = K_{\infty}$ up to translation. Furthermore, Proposition \[convergence\](2) then ensures that ${\lim}\ {\gamma}_{{{\Phi}}_j}(E_j) = {\gamma}_{{{\Phi}}_{\infty}}(K_{\infty}) =\frac{n}{n-1}, $ and therefore, by (\[betaform\]), $$\underset{j \to \infty}{\lim} \ \beta_{{{\Phi}}_j}(E_j)^2 =\underset{j \to \infty}{\lim} \ \frac{1}{n} \left({{\Phi}}_j(E_j ) -(n-1) {\gamma}_{{{\Phi}}_j}(E_j ) \right)= 0.$$\
*Step 2:* *Replace each $E_j$ with a minimizer $F_j$.*\
As in [@fuscojulin11], the idea is to replace each $E_j$ with a set $F_j$ for which we can say more about the regularity. We let ${\epsilon}_j = \beta_{{{\Phi}}_j}(E_j)$ and let $F_j$ be a minimizer to the problem $$\text{min} \left\{Q_j(F)= {{\Phi}}_j(F)
+ \frac{m_{{{\Phi}}_j}}{8M_{{{\Phi}}_j}} |\beta_{{{\Phi}}_j}(F)^2 -{\epsilon}_j^2|
+ \Lambda \big| |F| - 1 \big| \ :\ F \subset B_{R_0} \right\}$$ for a fixed $\Lambda > 4n$. Lemma \[LSC\] ensures that such a minimizer exists. As before, ${{\Phi}}_j(F_j) \geq c(\mathcal{N})P(F_j)$. Pairing this with provides a uniform bound on $P(F_j)$, so by compactness, $F_j \to F_{\infty} $ in $L^1$ up to a subsequence for some $F_{\infty} \subset B_{R_0}$.
For each $j$, we use the fact that $F_j$ minimizes $Q_j$, choosing $E_j$ as a comparison set. This, combined with (\[contra\]) and Lemma \[isop\], yields $$\begin{aligned}
{{\Phi}}_j(F_j) &
+ \frac{1}{8{\mathcal{N}}} |\beta_{{{\Phi}}_j}(F_j)^2 - {\epsilon}_j^2|
+ \Lambda \big| |F_j| - 1 \big|
\leq Q_j(F_j) \leq {{\Phi}}_j(E_j)\nonumber \\
\label{fine}
&\leq {{\Phi}}_j(K_j) + c_1 {\epsilon}_j^{4n/(n+1)} \leq
{{\Phi}}_j(F_j) + \Lambda \big| |F_j| - 1 \big| + c_1 {\epsilon}_j^{4n/(n+1)}.\end{aligned}$$ It follows that $\frac{1}{8{\mathcal{N}}} \big|\beta_{{{\Phi}}_j}(F_j)^2 -{\epsilon}_j^2 \big| \leq c_1 {\epsilon}_j^{4n/(n+1)}$, immediately implying that $\beta_{{{\Phi}}_j}(F_j) \to 0$. Moreover, rearranging and using the fact that ${\epsilon}_j \to 0$ and $\frac{4n}{n+1}>2$, we have $$\frac{{\epsilon}_j^2}{2^{(n+1)/2n} }\leq {\epsilon}_j^2 - 8{\mathcal{N}}c_1{\epsilon}_j^{4n/(n+1)} \leq \beta_{{{\Phi}}_j}(F_j)^2,$$ where the exponent $(n+1)/2n$ is chosen so that, taking the power $2n/(n+1),$ we obtain $$\label{B}
{\epsilon}_j^{4n/(n+1)} \leq 2\beta_{{{\Phi}}_j}(F_j)^{4n/(n+1)}.$$ In the last inequality in , if we replace $F_j$ with arbitrary set of finite perimeter $E\subset B_{R_0}$, then we obtain $${{\Phi}}_j(F_j)+ \Lambda\big| |F_j| -1\big| \leq {{\Phi}}_j(E) + \Lambda\big| |E| - 1\big| + c_1 {\epsilon}_j^{4n/(n+1)},$$ again using Lemma \[isop\]. Taking the limit inferior as $j \to \infty$, this implies that $F_{\infty}$ is a minimizer of the problem $$\min \left\{ {{\Phi}}_{\infty}(F) + \Lambda||F|-1| \ :\ F \subset B_{R_0}\right\},$$ and so $F_{\infty } = K_{\infty}$ up to a translation by Lemma \[isop\]. With no loss of generality, we translate each $F_j$ such that $\inf\{ |(F_j+z) \Delta K_{\infty} | :z\in{\mathbb{R}^n}\} = |F_j \Delta K_{\infty}|.$\
*Step 3:* *For $j$ sufficiently large, $\frac{1}{2}K_j \subset F_j \subset 2K_j$ and $|F_j | = 1$.*\
Lemma \[UDE\] implies that each $F_j$ satisfies uniform density estimates, and thus for $j$ sufficiently large, Lemma \[bound\] ensures that $\frac{1}{2} K_j \subset F_j\subset 2 K_j,$ as $|K_j \Delta F_j| \leq |K_j \Delta K_{\infty}| + |K_{\infty} \Delta F_j|$ and both terms on the right hand side go to zero.
Let $r_j>0$ be such that $|r_jF_j| =1$. We may take $r_jF_j$ as a comparison set for $F_j$; $r_j\leq2$ by Lemma \[LSC\], so $r_jF_j \subset 4K_j\subset B_{R_0}$ as long as $R_0> 4{M_{{{\Phi}}}}> C\mathcal{N},$ the second inequality following from $|K_j|=1$. Since $\beta_{{{\Phi}}_j}$ is invariant under scaling, $Q_j(F_j)\leq Q_j(r_jF_j)$ yields $$\begin{aligned}
\label{scale1} {{\Phi}}_j(F_j)+ \Lambda|1- |F_j||\leq r_j^{n-1} {{\Phi}}_j(F_j) .\end{aligned}$$ This immediately implies that $r_j \geq 1$ for all $j$, in other words, $|F_j| \leq 1$. Furthermore, $r_j \to 1$ because $F_j \to K_{\infty}$ in $L^1$ and $|K_{\infty}|=1$. Suppose that, for some subsequence, $r_j>1$. Then, using $|F_j| = 1/r_j^{n}$, implies $$\label{etasmall}
\Lambda \leq \bigg(\frac{r_j^n(r_j^{n-1} - 1)}{r_j^n -1 } \bigg){{\Phi}}_j(F_j).$$ For any $0<\eta<\frac{1}{n}$ and for $j$ sufficiently large, the right hand side is bounded by $ (1-\eta){{\Phi}}_j(F_j)$, as $\underset{r \to 1^+}{\lim}\ \frac{r^n(r^{n-1}-1)}{r^n -1} = \frac{n-1}{n}$. Furthermore, ${{\Phi}}_j(F_j) \leq n + {\epsilon}_j^2$ since $Q_j(F_j) \leq Q_j(K_j)$, so implies that $$\Lambda \leq(1-\eta) {{\Phi}}_j(F_j) \leq (1-\eta) \left(n + {\epsilon}_j^2\right) \leq n$$ for $j$ sufficiently large. Since $n<\Lambda$, we reach a contradiction, concluding that $|F_j| = 1$ for $j$ sufficiently large.\
*Step 4:* *Derive a contradiction to .*\
We will show that $\beta_{{{\Phi}}_j}(F_j)^{4n/(n+1)} \leq C \delta_{{{\Phi}}_j}(F_j)$, which in turn will be used to contradict . Adding and subtracting the term ${{\Phi}}_{j}(K_j)/n = (n-1) {\gamma}_{{{\Phi}}_j}(K_j)/n$ to , we have $$\begin{split}
\beta_{{{\Phi}}_j}(F_j)^2 &
\leq \frac{ {{\Phi}}_j(F_j)}{n} - \frac{n-1}{n} \int_{F_j} \frac{dx}{f^j_*(x)} =\delta_{{{\Phi}}_j}(F_j) + \frac{n-1}{n}\bigg(\int_{K_j} \frac{dx}{f^j_*(x) } - \int_{F_j} \frac{dx}{f^j_*(x)}\bigg) \\
&=\delta_{{{\Phi}}_j}(F_j) + \frac{n-1}{n}\bigg( \int_{F_j\setminus K_j} 1 - \frac{1}{f^j_*(x) } \ dx + \int_{K_j\setminus F_j} \frac{1}{f^j_*(x) } - 1 \ dx
\bigg).
\end{split}$$ We now control the last term in terms of $\delta_{{{\Phi}}_j}(F_j).$ Note the following: since $\frac{1}{2} K_j \subset F_j \subset 2K_j$, the last term above is bounded by $C|F_j\Delta K_j| \leq \delta_{{{\Phi}}_j}(F_j)^{1/2}.$ This could establish with the exponent $4$. However, with the following argument, we obtain the improved exponent $4n/(n+1)$.
As noted before, Lemma \[UDE\] implies that each $F_j$ satisfies uniform density estimates (\[UDES\]) with $m_{{{\Phi}}_j}/M_{{{\Phi}}_j}\geq 1/{{\mathcal{N}}}$. The lower density estimate provides information about how far $f^j_*(x)$ can deviate from $1$ for $x \in F_j\setminus K_j$, thus bounding the first integrand. Indeed, arguing as in the proof of Lemma \[bound\], for any $x \in F_j \setminus K_j$, let $d = f^j_*(x) -1$. The intersection $K_j\cap B_{d} (x) $ is empty by the definition of $f_*^j$, and thus $F_j\cap B_{d}(x) \subset F_j \setminus K_j$. Therefore, for $x \in \partial^*F_j \setminus K_j$, $$\frac{c_0}{ {{\mathcal{N}}}^n} d^n \leq |B_{d} (x) \cap F_j |
\leq |F_j \Delta K_j| \leq C\delta_{{{\Phi}}_j}(F_j)^{1/2}$$ by the lower density estimate in and the quantitative Wulff inequality as in . In fact, this bound holds for any $x \in F_j \setminus K_j$; since $F_j$ is bounded, for any $x \in F_j \setminus K_j$, there is some $y \in \partial^*F_j \setminus K_j$ such that $f_*^j(x) \leq f^*_j (y).$ Therefore, $f^j_*(x) - 1 \leq C\delta_{{{\Phi}}_j}(F_j)^{1/2n}$ for all $x \in F_j \setminus K_j$, and so $$\label{aa}
\begin{split}\int_{F_j \setminus K_j } 1- \frac{1}{f^j_*(x) }dx &\leq \int_{F_j \setminus K_j} f_*(x) - 1 \ dx
\leq \int_{F_j \setminus K_j} C\delta_{{{\Phi}}_j}(F_j)^{1/2n} \ dx\\
& = C |F_j\Delta K_j | \delta_{{{\Phi}}_j}(F_j)^{1/2n} \leq C \delta_{{{\Phi}}_j}(F_j)^{1/2+1/2n},
\end{split}$$ where $C = C({{\mathcal{N}}},n)$ and the final inequality uses once more. The analogous argument using the upper density estimate in , paired with the fact that eventually $\frac{1}{2}K_j \subset F_j$, provides an upper bound for the size of $1-f^j_*(x) $ for $x \in K_j\setminus F_j$, giving $$\label{bb}
\begin{split}
\int_{K_j\setminus F_j} \frac{1}{f^j_*(x)} - 1 \ dx \leq 2 \int_{K_j \setminus F_j }1 -f^j_*(x) \ dx
\leq C\delta_{{{\Phi}}_j}(F_j)^{1/2 + 1/2n}.
\end{split}$$ Combining (\[aa\]) and (\[bb\]), we conclude that $$\label{feb}\beta_{{{\Phi}}_j}(F_j)^{4n/(n+1)} \leq C_1\delta_{{{\Phi}}_j}(F_j)$$ where $C_1= C_1 ({\mathcal{N}}, n)$.
We now use the minimality of $F_j$, comparing against $E_j$, along with (\[contra\]) and (\[B\]) to obtain $$\begin{aligned}
{{\Phi}}_j(F_j) & \leq {{\Phi}}_j(E_j)
\leq {{\Phi}}_j(K_j) + c_1 {\epsilon}_j^{4n/(n+1)} \leq {{\Phi}}_j(K_j) +2c_1 \beta_{{{\Phi}}_j}(F_j)^{4n/(n+1)}.
\end{aligned}$$ By , $\beta_{{{\Phi}}_j}(F_j)$ is positive, so by choosing $c_1< n/2C_1$, this contradicts , thus proving for the class $\mathcal{F}_{\mathcal{N}}$ with the constant $C$ depending on $n$ and ${\mathcal{N}}$.\
*Step 5:* *Remove the dependence on ${{\mathcal{N}}}$ of the constant in (\[new1\]).*\
We argue as in [@FiMP10]. We will use the following notation: ${{\Phi}}_{K}$ is the surface energy with Wulff shape $K$, surface tension $f^{K}$, and support function $f_*^{K}$. We use $\delta_K, \beta_K$, and ${\gamma}_K$ to denote $\delta_{{{\Phi}}_K}, \beta_{{{\Phi}}_K}$, and ${\gamma}_{{{\Phi}}_K}$ respectively.
By John’s Lemma ([@John48 Theorem III]), for any convex set $K \subset {\mathbb{R}^n}$, there exists an affine transformation $L$ such that $\det L >0$ and $B_1 \subset L(K) \subset B_n.$ This implies that $M_{L(K)} / m_{L(K)} \leq n$ and so $f^{L(K)} \in \mathcal{F}_n$. Our goal is therefore to show that $\beta_{{{\Phi}}}(E)$ and $\delta_{{{\Phi}}}(E)$ are invariant under affine transformations. Indeed, once we verify that $\beta_K(E) = \beta_{L(K)}(L(E))$ and $ \delta_{K}(E) = \delta_{L(K)}(L(E)),$ we have $$\beta_K(E)^{4n/(n+1)} = \beta_{L(K)}(L(E))^{4n/(n+1)} \leq C(n) \delta_{L(K)}(L(E)) =
C(n) \delta_{K}(E) ,$$ and is proven with a constant depending only on $n$.
Suppose $E$ is a smooth, open, bounded set. Then $${{\Phi}}_K(E) =
\underset{{\epsilon}\to 0 }{\lim}\ \frac{ |E + {\epsilon}K | - |E| }{{\epsilon}};$$ this is shown by applying the anisotropic coarea formula to the function $$d^K(x, {\partial}E) :=\begin{cases}
\inf \{ f_*(x-y) : y \in {\partial}E\} & {\rm if }\ x \in E^c\\
-\inf \{ f_*(x-y) : y \in {\partial}E\} & {\rm if }\ x \in E
\end{cases}$$ and noting that $(E+{\epsilon}K) \setminus E = \{ x: 0\leq d^K(x, {\partial}E) <{\epsilon}\}.$
Since $L$ is affine, $ |L(E + {\epsilon}K)| - |L(E)| = \det L \left(|E + {\epsilon}K | - |E|\right)$, and so $${{\Phi}}_K(E) = \underset{{\epsilon}\to 0}{\lim} \ \frac{ |L(E + {\epsilon}K)| - |L(E)|}{{\epsilon}\det L}
=\frac{{{\Phi}}_{L(K)}(L(E))}{\det L}.$$ Since $|E| =|L(E)|/\det L$, we have $$\begin{aligned}
\delta_K(E) =
\frac{ {{\Phi}}_K(E)}{n|K|^{1/n}|E|^{1/n'} } -1
& = \frac{ {{\Phi}}_{L(K)}(L(E))}{n|L(K)|^{1/n} |L(E)|^{1/n'} } -1 = \delta_{L(K)}(L(E)), \end{aligned}$$ and thus $\delta_K(E)$ is invariant. Similarly, $$\begin{aligned}
f_*^K(L^{-1}z - y )& =
\inf \Big\{ \lambda \ : \ \frac{L^{-1}(z) -y }{\lambda} \in K \Big\}
= \inf \Big\{ \lambda \ : \ \frac{z - L(y)}{\lambda} \in L(K) \Big\} = f_*^{L(K)}\left(z-L(y)\right),\end{aligned}$$ and thus $$\int_E \frac{dx}{f_*^K(x-y)}= \int_{L(E) } \frac{dz}{f_*^K\left(L^{-1}(z) - y\right) \det L} = \int_{L(E)} \frac{dz}{f_*^{L(K)}\left(z-L(y)\right)\det L}.$$ Taking the supremum over $y\in {\mathbb{R}^n}$ of both sides, we have $${\gamma}_{K} (E)= \frac{{\gamma}_{L(K)}(L(E))}{\det L} .$$ From $(\ref{betaform})$, $$\beta_K(E) =\bigg(\frac{{{\Phi}}_K(E) -(n-1) {\gamma}_K(E)}{n|K|^{1/n} |E|^{1/n'} } \bigg)^{1/2} .$$ We have just shown that, for the denominator, $$\bigg(\frac{1}{n|K|^{1/n} |E|^{1/n'} } \bigg)^{1/2} = \bigg(\frac{\det L }{n|L(K)|^{1/n} |L(E)|^{1/n' }}\bigg)^{1/2}
,$$ and for the numerator, $$\bigg({{\Phi}}_K(E) -(n-1) {\gamma}_K(E) \bigg)^{1/2} =\bigg( \frac{ {{\Phi}}_{L(K) } (L(E)) - (n-1){\gamma}_{L(K)}(L(E))}{\det L } \bigg)^{1/2}.$$ The term $\det L$ cancels, yielding $$\beta_K(E) = \bigg(\frac{{{\Phi}}_{L(K) } (L(E)) - (n-1){\gamma}_{L(K)}}{n|L(K)|^{1/n} |L(E)|^{1/n'} }\bigg)^{1/2}
=\beta_{L(K)}(L(E)),$$ showing that $\beta_K(E)$ too is invariant.
The case of $\lambda$-elliptic Surface Tension {#smoothsection}
==============================================
In this section, we prove Theorem \[Smooth\]. This proof closely follows the proof of $(\ref{FuscoJulin})$ in [@fuscojulin11]. Using a selection principle argument and the regularity theory for $(\Lambda, r_0)$-minimizers of ${{\Phi}}$, we reduce to the case of sets that are small $C^1$ perturbations of the Wulff shape $K$. In [@fuscojulin11], this argument brings Fusco and Julin to the case of nearly spherical sets, at which point they call upon $(\ref{fugst})$, where Fuglede proved precisely this case in [@fuglede1989].
We therefore prove in Proposition \[Fug-type\] an analogue of $(\ref{fugst})$ in the case of the anisotropic surface energy ${{\Phi}}$ when $f$ is a $\lambda$-elliptic surface tension. The following lemma shows that if $E$ is a small $C^1$ perturbation of the Wulff shape $K$ with $|E| =|K|$, then the Taylor expansion of the surface energy vanishes at first order and takes the form . We then use the quantitative Wulff inequality as in $(\ref{FiMPStatement})$ and the barycenter constraint along with to prove Proposition \[Fug-type\].
\[expansionlem\] Suppose that ${{\Phi}}$ is a surface energy corresponding to a $\lambda$-elliptic surface tension $f$, and $E$ is a set such that $|E| = |K|$ and $$\partial E = \{ x + u(x) \nu_K(x) \, :\, x \in \partial K \}$$ where $u: {\partial}K \to \mathbb{R}$ and $\| u \|_{C^1(\partial K)} = {\epsilon}.$ There exists ${\epsilon}_0>0$ depending on $\lambda$ and $n$ such that if ${\epsilon}<{\epsilon}_0$, $$\begin{aligned}
\label{theexpansion}
{{\Phi}}(E) = {{\Phi}}(K) + &\frac{1}{2}\int_{\partial K }(\nabla u)^{\rm{T}} \nabla^2 f(\nu_K)\nabla u \,d{\mathcal{H}^{n-1}}- \frac{1}{2} \int_{\partial K }H_K u^2 \,d{\mathcal{H}^{n-1}}+{\epsilon}\ O (\|u\|_{H^1(\partial K)}^2),\end{aligned}$$ where $H_K$ is the mean curvature of $K$ and all derivatives are restricted to the tangential directions.
[The second fundamental form $\rm{II}_K$ of $K$ satisfies $$\nabla^2 f(\nu_K(x)) \rm{II}_K(x) = \rm{Id}_{T_x{\partial}K} \ \ \text{ for all } x \in {\partial}K.$$ Therefore, $H_K= \text{tr}({\rm{II}_K})$ is equal to $\text{tr}(\nabla^2 f\, \rm{II}_K^2)$ and thus agrees with, for example, [@clarenz2004surfaces Corollary $4.2$].]{}
For a point $x\in \partial K$, let $\{\tau_1, \hdots, \tau_{n-1}\}$ be normalized eigenvectors of $\nabla \nu_K$, where each $\tau_i $ corresponds to the eigenvalue $\lambda_i$. This set is an orthonormal basis for $T_xK$, and thus $\{\tau_1,\hdots, \tau_{n-1}, \nu_K\}$ is an orthonormal basis for ${\mathbb{R}^n}.$ A basis for $T_{x+u\nu_K}E$ is given by the set $\{g_1 , \hdots, g_{n-1}\}$, where, adopting the notation $u_i = \partial_{\tau_i}u,$ $$g_i = \partial_{\tau_i}[x + u\nu_K] = (1+\lambda_i u)\tau_i + u_i \nu_K.$$
We make the standard identification of an $(n-1)$-vector with a vector in ${\mathbb{R}^n}$ in the following way. The norm of an $(n-1)$-vector $v_1\wedge\cdots\wedge v_{n-1}$ is given by $|v_1 \wedge \hdots \wedge v_{n-1}| = | \det (v_1, \hdots , v_{n-1})|.$ If $|v_1 \wedge \hdots \wedge v_{n-1}| \neq 0$, then the vectors $v_1,\hdots , v_{n-1} $ are linearly independent and we may consider the $n-1$ dimensional hyperplane $\Pi$ spanned by $v_1, \cdots , v_{n-1}.$ Letting $\nu$ be a normal vector to $\Pi$, we make the identification $$v_1 \wedge \hdots \wedge v_{n-1} = \pm |v_1 \wedge \hdots \wedge v_{n-1}|\, \nu,$$ where the sign is chosen such that $\det (v_1, \hdots , v_{n-1} , \pm\nu) >0.$ In particular, we make the identifications $$\tau_1 \wedge \hdots \wedge \tau_{n-1} = \nu_K, \qquad\frac{ g_1 \wedge \hdots \wedge g_{n-1}}{|g_1 \wedge \hdots \wedge g_{n-1}|} = \nu_E
, \quad \text{and} \quad\tau_1 \wedge \hdots \wedge \nu_K \wedge \hdots \wedge \tau_{n-1} = - \tau_i.$$ The sign is negative in the third identification because $$\begin{aligned}
\det (\tau_1 ,\hdots, \nu_K , \hdots, \tau_{n-1}, -\tau_{i})
= -\det (\tau_1 ,\hdots,-\tau_{i},\hdots, \tau_{n-1}, \nu_K ) =\det (\tau_1 ,\hdots,\tau_{i},\hdots, \tau_{n-1}, \nu_K ) = 1.\end{aligned}$$ We let $w:=g_1 \wedge \hdots \wedge g_{n-1}$, and so $$\begin{aligned}
w &= [ (1+\lambda_1 u )\tau_i + u_1 \nu_K] \wedge \hdots \wedge [(1+\lambda_{n-1}u )\tau_{n-1} + u_{n-1} \nu_K] \nonumber \\
& = \prod_{i=1}^{n-1} (1+ \lambda_i u) \nu_K - \sum_{i=1}^{n-1} u_i \prod_{i\neq j} (1+ \lambda_ju)\tau_i \nonumber \\
&\label{ggg} = \Big[1+ H_K u + \sum_{i<j } \lambda_i \lambda_j u^2 \Big]\nu_K
-\sum_{i=1}^{n-1}u_i \Big[ 1 + \sum_{j\neq i } \lambda_j u\Big] \tau_i +{\epsilon}\, O( |u|^2 + |\nabla u|^2).\end{aligned}$$
In order to show (\[theexpansion\]), the volume constraint is used to show that the first order terms in the Taylor expansion of the surface tension vanish. We achieve this by expanding the volume in two different ways.
The divergence theorem implies that $$\begin{aligned}
n|E| = \int_{\partial E} x \cdot \nu_E\, d{\mathcal{H}^{n-1}}& =
\int_{\partial K} (x+ u\, \nu_K) \cdot \left(\frac{w}{|w|}\right)|w|\,d{\mathcal{H}^{n-1}}= \int_{\partial K} (x+ u\,\nu_K) \cdot w\,d{\mathcal{H}^{n-1}}. \end{aligned}$$ Adding and subtracting $\nu_K = \tau_1 \wedge \hdots \wedge \tau_{n-1}$, and using and the fact that $\nu_K \cdot \tau_i = 0$, we have $$\begin{aligned}
n|E| = \int_{\partial K } x\cdot \nu_K\, d{\mathcal{H}^{n-1}}+\int_{\partial K} u + x \cdot (w &- \nu_K )+H_K u^2\, d{\mathcal{H}^{n-1}}+{\epsilon}\, O( \|u\|_{H^1(\partial K)}^2
).\end{aligned}$$ Since $ \int_{\partial K } x\cdot \nu_K \,d{\mathcal{H}^{n-1}}= n|K|,$ the volume constraint $|E| = |K|$ implies that $$\label{vol1}
\begin{split} \int_{\partial K} x \cdot (w &- \nu_K )\, d{\mathcal{H}^{n-1}}= -\int_{\partial K} u +H_K u^2 \, d{\mathcal{H}^{n-1}}+{\epsilon}\, O( \|u\|_{H^1(\partial K)}^2).
\end{split}$$
Now we expand the volume in a different way. Because $f$ is a $\lambda$-elliptic surface tension, the Wulff shape $K$ is $C^2$ with mean curvature depending on $\lambda$ and $n$. Therefore, there exists $t_0=t_0(\lambda, n)>0$ such that the neighborhood $$D = \{ x+t\nu_K(x) :x\in \partial K , t\in (-t_0, t_0) \}$$ satisfies the following property: for each $y\in D,$ there is a unique projection $\pi:D \to \partial K$ such that $\pi(y) =x$ if and only if $y = x + t\nu_K(x)$ for some $t\in (-t_0, t_0)$. In this way, we extend the normal vector field $\nu_K$ to a vector field $N_K$ defined on $D$ by letting $N_K:D\to{\mathbb{R}^n}$ be defined by $N_K(y)= \nu_K(\pi(y)).$ We also extend $u$ to be defined on $D$ by letting $u(y) = u(\pi(y))$ for all $y \in D$. Therefore, if ${\epsilon}_0<t_0$, $\partial E$ may be realized as the time $t=1$ image of $\partial K$ under the flow defined by $$\frac{d}{dt} \psi_t(x) = uN_K(\psi_t(x)), \qquad \psi_0(x) = x.$$ Such a flow is given by $\psi_t(x) = x + tuN_K,$ and so $\nabla \psi_t(x) = {\rm{Id}} + tA$ where $A= \nabla (uN_K).$ A small adaptation of the proof of [@maggi2012sets Lemma $17.4$] gives $$\label{Jacobian}
J\psi_t = 1 + t \text{ tr}(A) +\frac{t^2}{2} (\text{tr}(A)^2 - \text{tr}(A^2)) + {\epsilon}\, O(|u|^2 + |\nabla u|^2).$$ Integrating by parts, it is easily verified that $$\begin{aligned}
\int_{K} \text{tr} (A)^2 - \text{tr} (A^2)\, dx
& = \int_{K } {{\rm div \,}}(uN_K\,{{\rm div \,}}(uN_K))\, dx - \int_{\partial K } \sum_{i,j=1}^n
(uN_K)^{(i)}\partial_i(uN_K)^{(j)} \nu_K^{(j)}\,d{\mathcal{H}^{n-1}}\\
& = \int_{K} \text{div} (uN_K\,{{\rm div \,}}(uN_K)) \,dx - \int_{\partial K }
u \nabla u \cdot \nu_K\, d{\mathcal{H}^{n-1}}.\end{aligned}$$ The second equality is clear by choosing the basis $\tau_1, \hdots, \tau_{n-1}, \tau_{n}$, where $\tau_n = \nu_K$. Furthermore, the divergence theorem implies that $$\int_{K} {{\rm div \,}}(uN_K\, {{\rm div \,}}(uN_K))\, dx = \int_{\partial K } u \ {{\rm div \,}}(uN_K) \, d{\mathcal{H}^{n-1}}=
\int_{\partial K}u \nabla u \cdot \nu_K + H_K u^2 \,d{\mathcal{H}^{n-1}},$$ so that $$\int_{K} \text{tr} (A)^2 - \text{tr} (A^2)\, dx=
\int_{\partial K} H_K u^2 \,d{\mathcal{H}^{n-1}}.$$ With this and in hand, we have the following expansion of the volume: $$\begin{aligned}
|\psi_t(K) | = \int_K J\psi_t \,dx
&= |K| + t \int_{\partial K } u \,d{\mathcal{H}^{n-1}}+ \frac{t^2}{2} \int_{\partial K} H_K u^2\,d{\mathcal{H}^{n-1}}+ t^3 {\epsilon}\, O(\|u\|_{H^1(\partial K)}^2).\end{aligned}$$ Therefore, the volume constraint $|K| =|E|= |\psi_1(K)|$ implies that $$\label{vol2} \int_{\partial K } u \, d{\mathcal{H}^{n-1}}=- \frac{1}{2}\int_{\partial K} H_K u^2\, d{\mathcal{H}^{n-1}}+ {\epsilon}\,O(\|u\|_{H^1(\partial K)}^2).$$ Combining $(\ref{vol1})$ and $(\ref{vol2})$, we conclude that $$\label{volume}
\int_{\partial K} x \cdot (w-\nu_K )\, d{\mathcal{H}^{n-1}}=-\frac{1}{2}\int_{\partial K } u^2H_K\, d{\mathcal{H}^{n-1}}+ {\epsilon}\, O(\|u\|_{H^1(\partial K)}^2).$$
We now proceed with a Taylor expansion of the surface energy of $E$: $$\begin{aligned}
{{\Phi}}(E) =& \int_{\partial^* E} f(\nu_E) \,d{\mathcal{H}^{n-1}}=
\int_{\partial K} f\Big(\frac{w}{|w|}\Big) |w|\,d{\mathcal{H}^{n-1}}= \int_{\partial K} f(w)\,d{\mathcal{H}^{n-1}}\\
= &\int_{\partial K } f(\nu_K)\, d{\mathcal{H}^{n-1}}+ \int_{\partial K } \nabla f(\nu_K)\cdot (w - \nu_K )\,d{\mathcal{H}^{n-1}}\\
&+ \frac{1}{2}\int_{\partial K } [w-\nu_K ]^{\text{T}} \nabla^2 f(\nu_K) [w - \nu_K ] \,d{\mathcal{H}^{n-1}}+{\epsilon}\, O( \|u\|_{H^1(\partial K)}^2),\end{aligned}$$ so, recalling that $\nabla f(\nu_K(x)) = x$ by , $${{\Phi}}(E) ={{\Phi}}(K)
+ \int_{\partial K }x\cdot (w- \nu_K )\,d{\mathcal{H}^{n-1}}+\frac{1}{2}\int_{\partial K } \sum_{i,j = 1}^{n-1} u_iu_j(\tau_i^{\text{T}} \nabla^2 f(\nu_K) \tau_j) \,d{\mathcal{H}^{n-1}}+ {\epsilon}\, O( \|u\|_{H^1(\partial K)}^2).$$ Applying $(\ref{volume})$ yields $(\ref{theexpansion})$, completing the proof.
We now prove Proposition \[Fug-type\], using $(\ref{theexpansion})$ as a major tool.
Suppose $E$ is a set as in the hypothesis of the proposition, i.e., $|E| = |K|,$ ${{\rm bar }\,}E = {{\rm bar }\,}K$, and $${\partial}E = \{ x + u(x) \nu_K(x): x \in {\partial}K \},$$ where $u: {\partial}K \to \mathbb{R}$ is a function such that $u \in C^1({\partial}K)$ and $\|u\|_{{C^1}({\partial}K)}= {\epsilon}\leq {\epsilon}_1$ with ${\epsilon}_1$ to be fixed during the proof. Up to multiplying $f$ by a constant, which changes $\lambda$ by the same factor and leaves ${m_{{{\Phi}}}}/{M_{{{\Phi}}}}$ unchanged, we may assume that $|K| =1$. Let $$B(u) = \frac{1}{2}\int_{{\partial}K }(\nabla u)^{\text{T}} \nabla^2 f(\nu_K)\nabla u\, d{\mathcal{H}^{n-1}}-\frac{1}{2} \int_{{\partial}K }H_K u^2 \,d{\mathcal{H}^{n-1}},$$ so that, by , $$\label{deficit bilinear form}
\delta_{{{\Phi}}}(E) = \frac{1}{n} B(u) +{\epsilon}\,O (\|u\|_{H^1({\partial}K)}^2)$$ as long as ${\epsilon}_1 \leq {\epsilon}_0$ for ${\epsilon}_0$ from Lemma \[expansionlem\].
[*[Step 1:]{}*]{} [*There exists $C = C(n,\lambda, {m_{{{\Phi}}}}/{M_{{{\Phi}}}})$ such that, for ${\epsilon}_1$ small enough depending on $ {m_{{{\Phi}}}}/{M_{{{\Phi}}}}$ and $\lambda$,*]{} $$\label{L1squared} \Big( \int_{{\partial}K} |u|\, d {\mathcal{H}^{n-1}}\Big)^2 \leq C\delta_{{{\Phi}}}(E).$$
[*[Step 1(a):]{}*]{} [*There exists $C= C(n,{m_{{{\Phi}}}}/{M_{{{\Phi}}}})$ such that, for ${\epsilon}_1= {\epsilon}_1({m_{{{\Phi}}}}/{M_{{{\Phi}}}})$ small enough,*]{} $$\label{okok}|E \Delta K| \leq C\delta_{{{\Phi}}}(E)^{1/2}.$$ The quantitative Wulff inequality in the form states that $ |E \Delta(K+ x_0)|\leq C(n)\delta_{{{\Phi}}}(E)^{1/2}$ for some $x_0 \in {\mathbb{R}^n}$, so by the triangle inequality, $$\label{bound1}
|E \Delta K| \leq C(n) \delta_{{{\Phi}}}(E)^{1/2} + |(K+x_0) \Delta K| .$$ It therefore suffices to show that $|(K+x_0) \Delta K| \leq C\delta_{{{\Phi}}}(E)^{1/2}.$ By [@maggi2012sets Lemma $17.9$], $$\label{center}
|K\Delta (K+x_0)| \leq 2 |x_0| P(K) \leq\frac{2n}{{m_{{{\Phi}}}}} |x_0| .$$ Furthermore, the barycenter constraint ${{\rm bar }\,}E= {{\rm bar }\,}K$ implies that $$x_0 = \int_{K} x_0\, dx = \int_{E} x\, dx - \int_K x - x_0\, dx
= \int_{E} x\,dx - \int_{K + x_0} x\,dx.$$ For ${\epsilon}_1$ small enough depending on ${M_{{{\Phi}}}}/{m_{{{\Phi}}}}$, $E, K+x_0 \subset B_{2{M_{{{\Phi}}}}}$, a fact that is verified geometrically since $|x_0| \to 0$ as ${\epsilon}\to 0$ and thus $|x_0|$ may be taken as small as needed. Therefore, $$\begin{aligned}
|x_0| =\bigg| \int_{E} x\,dx - \int_{K + x_0} x\,dx\bigg|
\leq 2{M_{{{\Phi}}}}|E \Delta (K+x_0)| \leq {M_{{{\Phi}}}}C(n)\delta_{{{\Phi}}}(E)^{1/2},\end{aligned}$$ where the second inequality comes from . This, , and prove .\
[*[Step 1(b):]{}*]{} [ *For ${\epsilon}_1$ sufficiently small depending on $\lambda$ and $n$, $$\label{koko} \int_{{\partial}K} | u| \, d{\mathcal{H}^{n-1}}\leq 2 |E \Delta K | .$$* ]{} Let $d_K(x) = \text{dist} (x,{\partial}K)$. As in the proof of Lemma \[expansionlem\], there exists $t_0 = t_0(\lambda, n)$ such that for all $t<t_0$, $\{ d_K = t\} = \{x + t\nu_K(x) \}$. Take ${\epsilon}_1<t_0$ and let $G_{t} =\{d_K = t\} \cap (E\setminus K)$. Then $$\begin{aligned}
E\setminus K &=& \{x+ t \nu_K\, : \, x\in \{ x \in {\partial}K : u(x)>0\},\, t \in (0, u(x) ) \},\\
G_{t}& =& \{ x + t \nu_K : \, x\in \{ x \in {\partial}K : u(x) >t\}\}.
\end{aligned}$$ The coarea formula and the area formula imply that $$\begin{aligned}
|E\setminus K | &= \int_{E\setminus K } |\nabla d_K| \,dx = \int_0^{\infty} dt \int_{G_{t}}\, d{\mathcal{H}^{n-1}}= \int_0^{\infty} \,dt \int_{\{u>t\}} J (\text{Id} + t\nu_K)\, d{\mathcal{H}^{n-1}},
\end{aligned}$$ so $$\begin{aligned}
|E\setminus K | & \geq \frac{1}{2} \int_0^{\infty} dt \int_{\{u>t\} } \,d{\mathcal{H}^{n-1}}= \frac{1}{2} \int_0^{\infty} |\{u>t\} |\,dt
= \frac{1}{2} \int_{{\partial}K} u^+\, d {\mathcal{H}^{n-1}}.
\end{aligned}$$ The analogous argument yields $|K\setminus E| \geq \frac{1}{2} \int_{{\partial}K} u^-\,d{\mathcal{H}^{n-1}}$, and is shown. Combining (\[okok\]) and (\[koko\]) implies (\[L1squared\]).
[*[Step 2:]{}*]{} [*There exists $C=C(n, \lambda,{m_{{{\Phi}}}}/{M_{{{\Phi}}}})$ such that, for ${\epsilon}_1={\epsilon}_1(n, \lambda, {m_{{{\Phi}}}}/{M_{{{\Phi}}}})$ small enough,*]{} $$\label{L2bound1} \|u\|_{H^1({\partial}K)}^2 \leq C \delta_{{{\Phi}}}(u).$$ The $\lambda$-ellipticity of $f$ implies $$\begin{aligned}
\int_{{\partial}K} |\nabla u|^2\, d{\mathcal{H}^{n-1}}&
\leq
\frac{1}{\lambda} \int_{{\partial}K} (\nabla u)^{\text{T}} \nabla^2 f(\nu_K ) (\nabla u)\, d{\mathcal{H}^{n-1}}=
\frac{1}{\lambda}\Big(2B(u) + \int_{{\partial}K} H_K |u|^2\,d{\mathcal{H}^{n-1}}\Big).\end{aligned}$$ The Wulff shape $K$ is bounded and $C^2$, so $H_K$ is bounded by a constant $C= C(n, \lambda)$. Therefore, $$\label{bound on gradient}
\int_{{\partial}K} |\nabla u|^2\, d{\mathcal{H}^{n-1}}\leq
\frac{2}{\lambda}B(u) +C \int_{{\partial}K} |u|^2\,d{\mathcal{H}^{n-1}}.$$
As pointed out in [@de2014sharp proof of Theorem 4], one may produce a version of Nash’s inequality on ${\partial}K$ that takes the form $$\begin{aligned}
\label{nash}
\int_{{\partial}K} |u|^2\, d {\mathcal{H}^{n-1}}\leq c \eta^{(n+2)/n} \int_{{\partial}K} |\nabla u|^2 \,d{\mathcal{H}^{n-1}}+ \frac{c}{\eta^{(n+2)/2}}\Big( \int_{{\partial}K} |u| \,d {\mathcal{H}^{n-1}}\Big)^2
\end{aligned}$$ for all $\eta>0$, where $c$ is a constant depending on $H_{K}$ (and therefore on $\lambda$ and $n$) and ${M_{{{\Phi}}}}/{m_{{{\Phi}}}}$. Indeed, as in ${\mathbb{R}^n}$, this form of Nash’s inequality is a consequence of the Sobolev inequality on ${\partial}K$, shown in [@simon1984lectures Section 18]. We pair with and to obtain $$\begin{aligned}
\int_{{\partial}K} |\nabla u|^2\, d{\mathcal{H}^{n-1}}& \leq
\frac{2}{\lambda}B(u) +
C \eta^{(n+2)/n} \int_{{\partial}K} |\nabla u|^2\, d{\mathcal{H}^{n-1}}+ \frac{C}{\eta^{(n+2)/2}} \delta_{{{\Phi}}}(E).
\end{aligned}$$ For $\eta$ small enough, we absorb the middle term into the left hand side. Then, recalling , we have $$\begin{aligned}
\frac{1}{2}\int_{{\partial}K} |\nabla u|^2\, d{\mathcal{H}^{n-1}}\leq
C \delta_{{{\Phi}}}(E) + {\epsilon}\,O \big(\|u\|_{H^1({\partial}K)}^2\big).\end{aligned}$$ Combining this estimate with and , we find that $\int_{{\partial}K} |u|^2 \, d{\mathcal{H}^{n-1}}$ is also bounded by $C \delta_{{{\Phi}}}(E) + {\epsilon}\,O \big(\|u\|_{H^1({\partial}K)}^2\big)$. Therefore, $$\begin{aligned}
\| u \|_{H^1({\partial}K)}^2 \leq
C \delta_{{{\Phi}}}(E) + {\epsilon}\,O \big(\|u\|_{H^1({\partial}K)}^2\big).\end{aligned}$$ Finally, taking ${\epsilon}_1$ small enough, we absorb the second term on the right, proving .
We now show that if $\partial E = \{ x+ u\nu_K: x \in \partial K\}$ with $\|u\|_{C^1({\partial}K)}$ small, then ${\beta_{{{\Phi}}}}(E)$ is controlled by $\|u\|_{H^1(\partial K)}.$ With the notation from the proof of Lemma \[expansionlem\], $$\begin{aligned}
n|K|{\beta_{{{\Phi}}}}(E)^2 \leq& \int_{\partial E} f(\nu_E) - \frac{x}{f_*(x)}\cdot \nu_{E} \, d{\mathcal{H}^{n-1}}= \int_{\partial K } f(w) - x\cdot w \, d{\mathcal{H}^{n-1}}.\end{aligned}$$ From the expansion of ${{\Phi}}$ in the proof of Lemma \[expansionlem\] and the fact that $x \cdot \nu_K = f(\nu_K)$ by , the right hand side is equal to $$\begin{aligned}
\frac{1}{2}\int_{\partial K } (\nabla u)^{\text{T}} \nabla^2 f(\nu_K) \nabla u\, d{\mathcal{H}^{n-1}}+ {\epsilon}\, O( \|u\|_{H^1(\partial K)}^2) \leq C \| u\|_{H^1({\partial}K)}^2+ {\epsilon}\, O( \|u\|_{H^1(\partial K)}^2),\end{aligned}$$ where $C= \| \nabla^2 f\|_{C^{0}({\partial}K)}$. For ${\epsilon}$ sufficiently small, we absorb the term ${\epsilon}\, O( \|u\|_{H^1(\partial K)}^2)$ and have $$\label{betaexp}
{\beta_{{{\Phi}}}}(E)^2 \leq \frac{C}{n|K|} \|u\|_{H^1(\partial K)}^2.$$
[ This is the first point at which we use the upper bound on the Hessian of $f$. In other words, Proposition \[Fug-type\] still holds for surface tensions $f \in C^{1,1}({\mathbb{R}^n}\setminus \{0\})$ that satisfy the lower bound on the Hessian in the definition of $\lambda$-ellipticity. ]{}
Next, we prove Theorem \[Smooth\], for which we need the following definition.
\[pm\] A set of finite perimeter $E$ is a $(\Lambda, r_0)$-minimizer of ${{\Phi}}$, for some $0\leq \Lambda <\infty $ and $r_0>0$, if $${{\Phi}}(E ; B(x,r) ) \leq {{\Phi}}(F; B(x,r)) + \Lambda|E\Delta F|$$ for $E\Delta F \subset \subset B(x,r) $ and $r<r_0$.
Proposition \[poincare\] implies that the proof reduces to showing $$\label{betaonly}{\beta_{{{\Phi}}}}(E)^2 \leq C\delta_{{{\Phi}}}(E).$$ where $C = C(n ,\lambda, \| \nabla^2 f\|_{C^0({\partial}K)} , {m_{{{\Phi}}}}/{M_{{{\Phi}}}})$. Suppose for contradiction that fails. There exists a sequence $\{E_j\}$ such that $|E_j| = |K|$ for all $j$, $\delta_{{{\Phi}}}(E_j) \to 0$, and $$\begin{aligned}
\label{contrastatement}
{{\Phi}}(E_j) \leq {{\Phi}}(K) + c_2 {\beta_{{{\Phi}}}}(E_j)^2\end{aligned}$$ for $c_2$ to be chosen at the end of this proof. Arguing as in the proof of Theorem \[th1\], we determine that, up to a subsequence, $\{E_j\}$ converges in $L^1$ to a translation of $K$. As in the proof of Theorem \[th1\] (and as in [@fuscojulin11]), we replace the sequence $\{E_j\}$ with a new sequence $\{F_j\}$, where each $F_j$ is a minimizer of the problem $$\min \left\{ Q_j (E)= {{\Phi}}(E) + \frac{|K|{m_{{{\Phi}}}}}{8{M_{{{\Phi}}}}} \left| {\beta_{{{\Phi}}}}(E)^2 - {\epsilon}_j^2\right|+ \Lambda \big| |E | - |K|\big|
\ \ : \ E \subset B_{R_0} \right\}$$ with ${\epsilon}_j = {\beta_{{{\Phi}}}}(E_j)$; existence for this problem is shown in Lemma \[LSC\]. Continuing as in the proof of Theorem \[th1\], we determine that $$\label{epsbeta2}
{\epsilon}_j^2 \leq 2 {\beta_{{{\Phi}}}}(F_j)^2,$$ that up to a subsequence and translation, $F_j \to K$ in $L^1$, and that $|F_j| = |K|$ for $j$ sufficiently large. By Lemma \[UDE\], each $F_j$ satisfies uniform density estimates, and so by Lemma \[bound\], for any $\eta>0$, we may choose $j$ sufficiently large such that $K_{1-\eta} \subset F_j \subset K_{1+\eta}.$
Arguing as in [@fuscojulin11], we show that $F_j$ is a $(\Lambda, r_0)$-minimizer of ${{\Phi}}$ for $j$ large enough, where $\Lambda$ and $r_0$ are uniform in $j$. Let $G$ such that $G\Delta F_j \subset \subset B_r(x_0)$ for $x_0 \in F_j$ and for $r<r_0$, where $r_0$ is to be fixed during the proof. For any $\eta>0$, if $B_r(x_0) \subset K_{1-\eta}$, then trivially ${{\Phi}}(G) \geq {{\Phi}}(F_j)$. If $B_r(x_0) \not\subset K_{1-\eta},$ then for $\eta$ sufficiently small, Lemma \[3.1\] implies that $|y_{F_j}| \leq 1/4 $ and $|y_G|\leq 1/4$. Furthermore, by choosing $\eta$ and $r_0$ sufficiently small, we may take $B_r(x_0) \cap K_{1/2} = \emptyset.$ The minimality of $F_j$ implies $Q(F_j) \leq Q(G)$; after rearranging and applying the triangle inequality, this implies that $$\begin{aligned}
\label{aboveok}
{{\Phi}}(F_j)& \leq {{\Phi}}(G) + \Lambda | F_j \Delta G | + \frac{|K|{m_{{{\Phi}}}}}{8{M_{{{\Phi}}}}}\left| {\beta_{{{\Phi}}}}(G)^2 - {\beta_{{{\Phi}}}}(F_j)^2 \right|.
\end{aligned}$$ As in in the proof of Lemma \[UDE\], $$\frac{|K|{m_{{{\Phi}}}}}{8{M_{{{\Phi}}}}}\left| {\beta_{{{\Phi}}}}(F)^2 - {\beta_{{{\Phi}}}}(G)^2 \right|
\leq
\ \frac{\left| {{\Phi}}(F_j) - {{\Phi}}(G) \right|}{2} + \frac{\left| {\gamma}_{{{\Phi}}}(F_j) - {\gamma}_{{{\Phi}}}(G) \right|}{2} +4|F_j \Delta G|$$ for $r_0$ small enough depending on $n$. If ${{\Phi}}(F_j) \leq {{\Phi}}(G)$, then the $(\Lambda, r_0)$-minimizer condition is automatically satisfied. Otherwise, subtracting $\frac{1}{2}{{\Phi}}(F_j)$ from both sides of and renormalizing, we have $$\begin{aligned}
\label{january}
{{\Phi}}(F_j) \leq {{\Phi}}(G) + |{\gamma}_{{{\Phi}}}(G) - {\gamma}_{{{\Phi}}}(F_j)|+(8+ 2\Lambda)|F_j \Delta G|.\end{aligned}$$ To control $|{\gamma}_{{{\Phi}}}(G) - {\gamma}_{{{\Phi}}}(F_j)|$, we need something sharper than the Hölder modulus of continuity of ${\gamma}_{{{\Phi}}}$ given in Proposition \[convergence\](2). Indeed, ${\gamma}_{{{\Phi}}}$ is Lipschitz continuous for sets whose intersection contains a ball around their centers: $${\gamma}_{{{\Phi}}}(F_j) - {\gamma}_{{{\Phi}}}(G)
\leq \int_{F_j} \frac{dx}{f_*(x-y_{F_j})} - \int_G \frac{dx}{f_*(x-y_{F_j})}
= \int_{F_j \Delta G} \frac{dx}{f_*(x-y_{F_j})},$$ and analogously, $${\gamma}_{{{\Phi}}}(G) - {\gamma}_{{{\Phi}}}(F_j ) \leq \int_{F_j \Delta G} \frac{dx}{f_*(x-y_G)}.$$ Since $B_r \cap K_{1/2} = \emptyset$, $|y_{F_j}| \leq 1/4$, and $|y_G| \leq 1/4$, we know that $1/f_*(x-y_{F_j}) \geq 4/m_{{{\Phi}}}$ and $1/f_*(x-y_G) \geq 4/m_{{{\Phi}}}$ for any $x \in F_j \Delta G$, implying that $$|{\gamma}_{{{\Phi}}}(F_j) - {\gamma}_{{{\Phi}}}(G) | \leq \frac{4}{m_{{{\Phi}}}}|F_j \Delta G|.$$ Therefore, becomes $${{\Phi}}(F_j) \leq {{\Phi}}(G) + \Lambda_0 \left|F_j \Delta G\right|,$$ where $\Lambda_0= 8 + 2\Lambda + 4/{m_{{{\Phi}}}},$ and so $F_j$ is a $(\Lambda_0, r_0)$-minimizer for $j$ large enough.
We now exploit some regularity theorems for sets $F_j$ that are $(\Lambda, r_0)$-minimizers that converge in $L^1$ to a $C^2$ set. First, let us introduce a bit of notation. For $x \in {\mathbb{R}^n}$, $r>0$, and $\nu \in S^{n-1}$, we define $$\begin{aligned}
\textbf{C}_{\nu}(x,r) = \{ y\in \mathbb{R}^{n} : |p_{\nu} (y-x)|<r, |q_{\nu}(y-x) < r\},\\
\textbf{D}_{\nu} (x,r) = \{ y \in {\mathbb{R}^n}: |p_{\nu}(y-x)| <r , |q_{\nu}(y-x) | = 0\},\end{aligned}$$ where $q_{\nu}(y) = y \cdot \nu$ and $p_{\nu}(y) = y - (y \cdot \nu) y.$ We then define the *cylindrical excess* of $E$ at $x$ in direction $\nu$ at scale $r$ to be $$\textbf{exc} (E, x ,r,\nu ) = \frac{1}{r^{n-1}} \int_{\textbf{C}_{\nu}(x,r) \cap \partial^* E} \frac{|\nu_E - \nu|^2}{2} \,d{\mathcal{H}^{n-1}}$$
The following regularity theorem for almost minimizers of an elliptic integrand is the translation in the language of sets of finite perimeter of a classical result in the theory of currents, see [@alm66; @SSA77; @bomb82; @DuzaarSteffen02]. For a closer statement to ours, see Lemma 3.1 in [@DePMag14].
\[epsreg\] Let $f$ be a $\lambda$-elliptic surface tension with corresponding surface energy ${{\Phi}}$. Suppose $E$ is a $(\Lambda, r_0)$-minimizer of ${{\Phi}}$. For all $\alpha<1$ there exist constants ${\epsilon}$ and $C_1$ depending on $n, \lambda$ and $\alpha$ such that if $${\rm{\bf{exc}}}(E, x,r,\nu) + \Lambda r <{\epsilon}$$ then there exists $u\in C^{1,\alpha}(\textbf{D}_{\nu}(x,r))$ with $u(x) =0$ such that $$\begin{aligned}
\textbf{C}_{\nu}(x,r/2) \cap \partial^*E & = ({\rm{Id}} + u \nu)({\rm{\bf{D}}}_{\nu}(x,r/2)),\\
\|u\|_{C^0({\rm{\bf{D}}}_\nu(x_0, r/2))}& < C_1r\, {\rm{\bf{exc}}}(E,x,r,\nu)^{1/(2n-2)},\\
\|\nabla u\|_{C^0({\rm{\bf{D}}}_{\nu}(x_0, r/2))}
&< C_1\, {\rm{\bf{exc}}}(E,x,r,\nu)^{1/(2n-2)},\\
\text{and} \qquad r^{\alpha} [\nabla u ]_{C^{0,\alpha}({\rm{\bf{D}}}_{\nu}(x,r/2))}& < C_1\, {\rm{\bf{exc}}}(E,x,r,\nu)^{1/2}.\end{aligned}$$
Applying Theorem \[epsreg\] as in [@CiLe12], we come to prove the following statement.
\[improvedconvergence\] Let $f$ be $\lambda$-elliptic with corresponding surface energy ${{\Phi}}$ and let $\{E_j\}$ be a sequence of $(\Lambda, r_0)$-minimizers such that $E_j \to E$ in $L^1$, with $\partial E \in C^2$. Then there exist functions $\psi_j \in C^1(\partial E)$ such that $$\partial E_j =({\rm{Id}} + \psi_j \nu_E)(\partial E),$$ and $\|\psi_j\|_{C^1(\partial E)} \to 0.$
Theorem \[improvedconvergence\] implies that we may express $\partial F_j$ as $$\partial F_j = \{ x+ \psi_j \nu_K : x \in \partial K\},$$ where $\|\psi_j\|_{C^1(\partial K)} \to 0$. Moreover, ${{\rm bar }\,}F_j = {{\rm bar }\,}K$ and $|F_j| = |K|$, so Proposition \[Fug-type\] and imply that $$\label{here} C \delta_{{{\Phi}}}(F_j)\geq \| \psi_j \|_{H^1(\partial K)}^2 \geq c {\beta_{{{\Phi}}}}(F_j)^2.$$ On the other hand, $F_j$ minimizes $Q_j$, so choosing $E_j$ as a comparison set and using (\[contrastatement\]) and (\[epsbeta2\]), we have $${{\Phi}}(F_j) \leq {{\Phi}}(E_j) \leq {{\Phi}}(K) + c_2{\epsilon}_j^2 \leq {{\Phi}}(K) + 2c_2{\beta_{{{\Phi}}}}(F_j)^2.$$ By , ${\beta_{{{\Phi}}}}(F_j)>0,$. Then, using and choosing $c_2$ sufficiently small, we reach a contradiction.
The case of crystalline surface tension in dimension $2$ {#crystalsection}
========================================================
In this section, we prove Theorem \[dim2\]. As in the previous section, we begin by showing the result in a special case, and then use a selection principle argument paired with specific regularity properties to reduce to this case.
Let $n=2$ and suppose that $f$ is a crystalline surface tension as defined in Definition \[crystals\], with ${{\Phi}}$ the corresponding anisotropic surface energy. The corresponding Wulff shape $K \subset \mathbb{R}^2 $ is a convex polygon with normal vectors $\{\nu_i\}_{i=1}^N$. Let us fix some notation to describe $K$, illustrated in Figure $1$. Denote by $s_i$ the side of $K$ with normal vector $\nu_i$, choosing the indices such that $s_i$ is adjacent to $s_{i+1}$ and $s_{i-1}$. Let $\theta_i \in (0,\pi)$ be the angle between $s_i$ and $s_{i+1}$, adopting the convention that $s_{n+1} = s_1$. Let $H_i$ be the distance from the origin to the side $s_i$. By construction, $$\label{Hi}
f(\nu_i) = H_i.$$
We say that a set $E\subset \mathbb{R}^2$ is *parallel to $K$* if $E$ is an open convex polygon with $\{\nu_E\}= \{\nu_i\}_{i=1}^N$, that is, $\nu_E(x)\in \{\nu_i\}_{i=1}^N$ for all $ x \in \partial^*E$, and for each $i \in \{1,\hdots, N\}$, there exists $x \in {\partial}^*E$ with $\nu_E(x) = \nu_i$. For a set $E$ that is parallel to $K$, we denote by $\sigma_i$ the side of $E$ with normal vector $\nu_i$, and $h_i$ the distance between the origin and $\sigma_i$; again see Figure $1$. We define ${\epsilon}_i= h_i - H_i$. Notice that ${\epsilon}_i$ has a sign, with ${\epsilon}_i \geq 0$ when dist$(0, s_i) \leq $ dist$(0, \sigma_i)$ and ${\epsilon}_i \leq 0 $ when dist$(0, s_i) \geq $ dist$(0, \sigma_i)$. For simplicity of notation, we let $|s | = \mathcal{H}^1(s)$ for any line segment $s$.
The following proposition proves strong form stability for sets $E$ that are parallel to $K$ such that $|E|=|K|$ and $|E\Delta K| ={\inf} \{|E\Delta (K+y)| : y\in \mathbb{R}^2\}$. Then, by a selection principle-type argument and a rigidity result, we will reduce to this case.
\[crystalprop\] Let $E\subset \mathbb{R}^2$ be parallel to $K$ such that $|E| = |K|$ and $|E\Delta K| ={\inf} \{ |E\Delta (K+y)|: y\in \mathbb{R}^2\}$. Then there exists a constant $C$ depending on $f$ such that $${\beta_{{{\Phi}}}}(E)^2 \leq C\delta_{{{\Phi}}}(E).$$
Let $E$ be as in the hypothesis of the proposition. By (\[Hi\]), we have $$\begin{aligned}
{{\Phi}}(E) =\sum_{i=1}^N H_i |\sigma_i|,\qquad
{{\Phi}}(K) =\sum_{i=1}^N H_i |s_i|, \qquad
|E| =
\sum_{i=1}^N \frac{h_i |\sigma_i|}{2}, \qquad
|K| = \sum_{i=1}^N \frac{ H_i |s_i|}{2}.
\end{aligned}$$
Recalling that ${\epsilon}_i = h_i - H_i$, we may express the volume constraint $|E| = |K|$ as $$\sum_{i=1}^N \frac{ H_i |s_i|}{2} = |K| = |E| = \sum_{i=1}^N \frac{ H_i |\sigma_i| }{2} + \sum_{i=1}^N \frac{ {\epsilon}_i |\sigma_i|}{2}.$$ Furthermore, $$\label{deficit} 2|K|\delta_{{{\Phi}}}(E) ={{\Phi}}(E) -{{\Phi}}(K)=\sum_{i=1}^N H_i (|\sigma_i| - |s_i| )= -\sum_{i=1}^N {\epsilon}_i |\sigma_i|.$$ Note that $\sum_{i=1}^N |{\epsilon}_i| \leq C |E\Delta K|$ for some constant $C=C(f)$, and so by (\[FiMPStatement\]), $$\label{epsilon}\bigg( \sum_{i=1}^N |{\epsilon}_i|\bigg)^2 \leq C\delta_{{{\Phi}}}(E),$$ and in particular, $|{\epsilon}_i|^2 \leq C\delta_{{{\Phi}}}(E)$ for each $i$.\
*Step 1:* We use (\[betaform\]) and add and subtract $\frac{{{\Phi}}(K)}{2|K|} = \frac{{\gamma}_{{{\Phi}}}(K)}{2|K|}$ to obtain $$\begin{aligned}
{\beta_{{{\Phi}}}}(E)^2
\leq \frac{1}{2|K|}\Big({{\Phi}}(E) -\int_E \frac{dx}{f_*(x) }\Big) &=\delta_{{{\Phi}}}(E) + \frac{1}{2|K|} \Big(\int_{K \setminus E} \frac{dx}{f_*(x)}
- \int_{E \setminus K} \frac{dx}{f_*(x)}\Big).
\end{aligned}$$ Thus we need only to control the term $A-B$ linearly by the deficit, where $$A = \int_{K \setminus E} \frac{dx}{f_*(x) } , \qquad \qquad
B= \int_{E\setminus K} \frac{dx}{f_*(x) }.$$ To bound the term $A-B$ from above, we bound $A$ from above and bound $B$ from below. Our main tool is the anisotropic coarea formula in the form given in .
First, we consider the term $A$, where (\[awcoarea\]) yields $$\label{Coarea A}
A = \int_{K\setminus E} \frac{dx}{f_*(x)} = \int_0^\infty \frac{{{\Phi}}(rK; K\setminus E)}{r}dr
=\int_0^1 \frac{{{\Phi}}(rK; K\setminus E)}{r}dr.$$ We introduce the notation $$I^- = \{ i \in \{1, \hdots N\} : {\epsilon}_i<0\}, \qquad
I^+ = \{1, \hdots N\} \setminus I^-.$$ From , we obtain an upper bound on $A$ by integrating over $r$, for each $i \in I^-$, the part of the perimeter of $rK$ that lies between $\sigma_i$ and $s_i$. This means that for each $r$, we pick up the part of $\partial^*(rK)$ that is parallel to $\sigma_i$ and $s_i$, as well as part of the adjacent sides: $$\label{seepic}{{\Phi}}(rK; K\setminus E) \leq \sum_{I^-}\left[ H_i r |s_i| + H_{i-1} \frac{(rH_{i} - h_i)}{\sin(\theta_{i-1})}
+ H_{i+1} \frac{(rH_i - h_i )}{\sin( \theta_i)} \right];$$ see Figure 2 and recall .
\[seeline\]
This and imply that $$\label{5}
A
\leq \sum_{I^-}
\int_{h_i/H_i}^1
\left[ H_i r |s_i| + H_{i-1} \frac{(rH_{i} - h_i)}{\sin(\theta_{i-1})}
+ H_{i+1} \frac{(rH_i - h_i )}{\sin( \theta_i)} \right] \frac{dr}{r},$$
Now we add and subtract the term $\int_{h_i/H_i}^1 H_i |\sigma_i| \frac{dr}{r}.$ The idea is that $H_i |\sigma_i|$ gives a rough estimate of the term in brackets on the right hand side of $(\ref{5})$. Indeed, for each $r$, the part of $\partial^*(r K)$ between $\sigma_i$ and $s_i$ has length roughly equal to $H_i|\sigma_i|$. We will see that this estimate is not too rough; the error can be controlled by the deficit. Thus we rewrite as $$\begin{aligned}
A& \leq \sum_{i \in I^-} \int_{h_i/H_i}^1 \frac{H_i |\sigma_i|}{r}\,dr
+\sum_{i \in I^-} \int_{h_i/H_i}^1 H_i |s_i| + \left[ H_i - \frac{h_i}{r} \right] \left( \frac{H_{i-1}}{\sin (\theta_{i-1}) } + \frac{H_{i+1}}{\sin(\theta_i)} \right) - \frac{H_i |\sigma_i|}{r} \,dr.\end{aligned}$$ Noting that $H_i/\sin(\theta_j) \leq C =C(f)$ for each $i, j$, the right hand side is bounded by $A_1 + A_2$, where $$A_1 =
\sum_{i \in I^-} \int_{h_i/H_i}^1 \frac{H_i |\sigma_i|}{r}dr , \qquad A_2 = \sum_{i \in I^- } \int_{h_i/H_i}^1 H_i |s_i| +C \left[ H_i - \frac{h_i}{r} \right] - \frac{H_i |\sigma_i|}{r}dr.$$ The term $A_2$ is the error term that we will show is controlled by the deficit in Step $2$.
First, we perform an analogous computation for $B$, and show how, once the error terms are taken care of, the proof is complete. Again, by (\[awcoarea\]), we have $$B = \int_{E\setminus K} \frac{dx}{f_*(x) }
=
\int_0^{\infty} \frac{{{\Phi}}(rK ; E\setminus K)}{r}\, dr=
\int_1^{\infty} \frac{{{\Phi}}(rK ; E\setminus K)}{r} \,dr.$$ To bound $B$ from below, we integrate, for each $i\in I^+$, only the part of $\partial^*(rK)$ that is parallel to $s_i$ and $\sigma_i$ and lies between $s_i$ and $\sigma_i$ . We call this segment $\ell_i^r := E\setminus K \cap \{e_i + rx_i\}$, where $e_i$ is the vector parallel to the sides $\sigma_i $ and $s_i$, $x_i \in s_i$, and $r \in [1, h_i/H_i]$.
Thus, letting $s_i^r$ be the side of $rK$ parallel to $s_i$ and recalling , we have $$\int_1^{\infty} \frac{{{\Phi}}(rK; E\setminus K)}{r} \,dr
\geq \sum_{i \in I^+} \int_1^{h_i/H_i}\frac{ H_i|s_i^r \cap \ell_i^r| }{r}\,dr.$$ Once again, a rough estimate for $H_i|s_i^r \cap \ell_i^r |$ is given by $H_i |\sigma_i|.$ We will again show that this estimate is not too rough, specifically, that the error between these integrals is controlled by the deficit. So we continue: $$B\geq \sum_{i \in I^+} \int_1^{h_i/H_i} \frac{H_i |\sigma_i| }{r} \,dr
+ \sum_{i \in I^+} \int_1^{h_i/H_i} \frac{H_i|s_i^r \cap \ell_i^r|}{r} - \frac{H_i |\sigma_i|}{r} dr =B_1 +B_2,$$ where $$\begin{aligned}
B_1 = \sum_{i \in I^+} \int_1^{h_i/H_i} \frac{H_i |\sigma_i| }{r} \,dr,\qquad B_2& = \sum_{i \in I^+} \int_1^{h_i/H_i} \frac{H_i|s^r_i \cap \ell_i^r|}{r} - \frac{H_i |\sigma_i|}{r}\, dr.
\end{aligned}$$ Like $A_2$, $B_2$ is an error term that we will show is controlled by the deficit in Step $2$.
Before bounding $|A_2|$ and $|B_2|$ by the deficit, let us see how this will conclude the proof. As we saw, ${\beta_{{{\Phi}}}}(E)^2 \leq \delta_{{{\Phi}}}(E) + \frac{1}{2|K|}(A-B). $ Recalling that $h_i = H_i + {\epsilon}_i$, $$\begin{aligned}
A-B &= \sum_{i \in I^-} \int_{h_i/H_i}^1 \frac{H_i |\sigma_i|}{r} dr
- \sum_{i \in I^+ } \int_1^{h_i/H_i} \frac{H_i |\sigma_i|}{r} dr+ A_2 -B_2\\
& = -\sum_{i \in I^-} H_i |\sigma_i| \log \Big(\frac{h_i}{H_i} \Big)
- \sum_{i \in I^+} H_i |\sigma_i| \log\Big(\frac{h_i}{H_i} \Big) + A_2- B_2 \\
& =- \sum_{i =1}^N H_i |\sigma_i|\Big(\frac{{\epsilon}_i}{H_i} + O({\epsilon}_i^2)\Big)
+ A_2 - B_2 = - \sum_{i=1}^{N} {\epsilon}_i |\sigma_i| + \sum_{i=1}^N O({\epsilon}_i^2) +A_2 - B_2 .\end{aligned}$$ The first term is precisely equal to $2|K| \delta_{{{\Phi}}}(E)$ by (\[deficit\]), while $\sum_{i} O({\epsilon}_i^2) \leq C\delta_{{{\Phi}}}(E)$ by $(\ref{epsilon})$. Therefore, once we show that $|A_2|$ and $|B_2|$ are controlled linearly by the deficit, our proof is complete.\
*Step 2:* In this step we bound the error terms. We show that $|A_2| \leq C \delta_{{{\Phi}}}(E)$; the proof that $|B_2| \leq C \delta_{{{\Phi}}}(E)$ is analogous. The main idea for estimating the integral $A_2$ is to show that the contribution of the adjacent sides is small, and then estimate the rest of integrand slice by slice. Recalling $A_2,$ the triangle inequality gives $$\label{a2} |A_2|\leq \bigg| \sum_{i \in I^-} \int_{h_i/H_i}^1 \frac{H_i}{r} (r|s_i| - |\sigma_i|) dr \bigg|
+ C\sum_{i \in I^-}\bigg| \int_{h_i/H_i}^1 \Big[H_i - \frac{h_i}{r}\Big] dr\bigg|.$$ The second term in (\[a2\]) corresponds to the contribution of adjacent sides. By $h_i = H_i +{\epsilon}_i$, $$\begin{aligned}
&C \sum_{ i \in I^-}\bigg| \int_{h_i / H_i}^1
\left[H_i - \frac{h_i}{r}\right]
dr \bigg| = C\sum_{i \in I^-} \bigg| (H_i - h_i) + h_i \log\Big(\frac{h_i}{H_i}\Big)\bigg| \\
=& C\sum_{i \in I^-}
\Big| -{\epsilon}_i + h_i \frac{{\epsilon}_i}{H_i} + O({\epsilon}_i^2)\Big| =C \sum_{i \in I^-}
\Big| \frac{{\epsilon}_i^2 }{H_i} + O({\epsilon}_i^2) \Big| =C \sum_{I^-} O({\epsilon}_i^2) \leq C \delta_{{{\Phi}}}(E).\end{aligned}$$ To bound the first term in , we will show that $\big| r|s_i| - |\sigma_i| \big| \leq C
\max \{ |{\epsilon}_{i-1}| \}$ for $r \in [h_i/H_i, 1]$, where the constant $C$ depends on ${{\Phi}}$, and then obtain our bound by integrating. To this end, we rotate our coordinates such that $\nu_i = e_2$, so the side $s_i$ has endpoints $(a, H_i)$ and $(b, H_i)$ for some $a<b$. We compute explicitly the endpoints of $\sigma_i$; it has, respectively, left and right endpoints $$\Big( a + \tan\left(\theta_{i-1} - \pi/2 \right) {\epsilon}_i - \frac{{\epsilon}_{i-1}}{\sin(\theta_{i-1})} ,\ h_i \Big) \qquad \text{and} \qquad
\Big(b - \tan\left(\theta_i -\pi/2 \right){\epsilon}_i + \frac{{\epsilon}_{i+1}}{\sin(\theta_{i})}, \ h_i \Big).$$ Thus $$|\sigma_i| = \Big|b
- \tan \left(\theta_i -\pi/2 \right) {\epsilon}_i
+ \frac{{\epsilon}_{i+1}}{\sin(\theta_{i})}
-\Big(a + \tan\left(\sigma_{i-1} - \pi/2\right)
{\epsilon}_i
- \frac{{\epsilon}_{i-1}}{\sin(\theta_{i-1})}\Big)\Big| .$$ and so $$\|\sigma_i| - |b-a| | \leq C( |{\epsilon}_i| +|{\epsilon}_{i+1}| +|{\epsilon}_{i-1}|),$$ where $C$ depends on ${{\Phi}}$. Therefore, recalling that $|b-a| = |s_i|$, $$\begin{aligned}
\Big| r|s_i| - |\sigma_i| \Big|
& \leq (1-r)|s_i| +C( |{\epsilon}_i| +|{\epsilon}_{i+1}| +|{\epsilon}_{i-1}|) \leq
\frac{|{\epsilon}_i|}{H_i} |s_i| + C \max \{ |{\epsilon}_j |\}\leq C \max \{ |{\epsilon}_j |\}.\end{aligned}$$ Given this estimate on slices, we integrate over $r$: $$\begin{aligned}
\sum_{i \in I^-}& \int_{h_i / H_i}^1 \frac{H_i}{r} \left(r|s_i| - |\sigma_i|\right)dr \leq C\max \{ |{\epsilon}_j |\} \sum_{i \in I^-} \int_{h_i / H_i}^1 \frac{H_i}{r} dr=C \max \{ |{\epsilon}_j |\}\sum_{i \in I^-} H_i
\Big| \log\Big(\frac{h_i}{H_i}\Big)\Big| \\
&= C \max \{ |{\epsilon}_j |\}\sum_{i \in I^-} ({\epsilon}_i + O({\epsilon}_i^2)) = O(\max\{ |{\epsilon}_j|^2\}) \leq C({{\Phi}}) \delta_{{{\Phi}}}(E),\end{aligned}$$ where the last inequality follows from .
We prove Theorem \[dim2\] after introducing the following definition that we will need in the proof.
\[special\] A set $E$ is a *volume constrained* $({\epsilon}, \eta_0)$-*minimizer of ${{\Phi}}$* if $${{\Phi}}(E) \leq {{\Phi}}(F) + {\epsilon}|E \Delta F|$$ for all $F$ such that $|E| = |F|$ and $(1-\eta_0)E \subset F\subset (1+\eta_0) E .$
By Proposition \[poincare\], we need only to show that there exists some $C$ depending on $f$ such that $$\label{betaonly2}{\beta_{{{\Phi}}}}(E)^2 \leq C\delta_{{{\Phi}}}(E).$$ for all sets $E$ of finite perimeter with $0<|E|<\infty.$ Suppose for contradiction that (\[betaonly2\]) does not hold. There exists a sequence $\{E_j\}$ such that $|E_j| = |K|$, $\delta_{{{\Phi}}}(E_j) \to 0$, and $$\begin{aligned}
\label{contra2}
{{\Phi}}(E_j) \leq {{\Phi}}(K) + c_3 {\beta_{{{\Phi}}}}(E_j)^2\end{aligned}$$ for $c_3$ to be chosen at the end of this proof. By an argument identical to the one given in the proof of Theorem \[Smooth\], we obtain a new sequence $\{F_j\}$ with $F_j \subset B_{R_0}$ for all $j$ such that the following properties hold:
- each $F_j $ is a minimizer of $Q_j(E) = {{\Phi}}(E) + \frac{|K|{m_{{{\Phi}}}}}{8{M_{{{\Phi}}}}}| {\beta_{{{\Phi}}}}(E)^2 - {\epsilon}_j^2| + \Lambda\big| |E| - |K|\big|$ among all sets $E \subset B_{R_0},$ where ${\epsilon}_j = {\beta_{{{\Phi}}}}(E_j)$;
- $F_j$ converges in $L^1$ to a translation of $K$;
- $|F_j | = |K| \text{ for $j$ sufficiently large}$;
- the following lower bound holds for ${\beta_{{{\Phi}}}}(F_j):$ $$\label{epsbeta} {\epsilon}_j^2 \leq 2 {\beta_{{{\Phi}}}}(F_j)^2.$$
Translate each $F_j$ such that $|F_j \Delta K | = {\inf}\{|F_j \Delta (K+y)|: y\in\mathbb{R}^2\}$. We claim that for all ${\epsilon}>0$, there exists $\eta_0>0$ such that $F_j$ is a volume constrained $({\epsilon}, \eta_0)$-minimizer of ${{\Phi}}$ (Definition \[special\]) for $j$ large enough. Indeed, fix ${\epsilon}>0$ and let $\eta_1 = c_1 {\epsilon}$, where $c_1 = c_1(f)$ will be chosen later. By Lemma \[3.1\], there exists $\eta_2 $ such that if $(1- \eta_2)K \subset E \subset (1+ \eta_2)K,$ then $|y_{E}| <\eta_1$. Let $\eta_0 =\min \{ \eta_1, \eta_2\}/2$.
By Lemma \[UDE\], each $F_j$ satisfies uniform density estimates, and so Lemma \[bound\] implies that, for $j$ large, $(1- \eta_0)K \subset F_j \subset (1+ \eta_0)K$ and thus $|y_{F_j}|<\eta_1$. Let $E$ be such that $|E| = |F_j|$ and $(1-\eta_0)F_j \subset E \subset (1+\eta_0)F_j$. Then $|y_E| <\eta_1$ and $$\begin{aligned}
(1- \eta_1)K \subset F_j \subset (1+\eta_1)K, \qquad (1- \eta_1)K \subset &E \subset (1+ \eta_1)K.
$$ Because $F_j$ minimizes $Q_j$, $${{\Phi}}(F_j) + \frac{|K|{m_{{{\Phi}}}}}{4{M_{{{\Phi}}}}} |{\beta_{{{\Phi}}}}(F_j)^2 -{\epsilon}_j^2| \leq {{\Phi}}(E) + \frac{|K|{m_{{{\Phi}}}}}{4{M_{{{\Phi}}}}} |{\beta_{{{\Phi}}}}(E)^2 - {\epsilon}_j^2|$$ and so by the triangle inequality and since ${m_{{{\Phi}}}}\leq {M_{{{\Phi}}}}$, $${{\Phi}}(F_j) \leq {{\Phi}}(E) + \frac{|K|}{4} |{\beta_{{{\Phi}}}}(E)^2 - {\beta_{{{\Phi}}}}(F_j)^2| .$$ If ${{\Phi}}(F_j) \leq {{\Phi}}(E)$, then the volume constrained minimality condition holds trivially. Otherwise, with a bound as in , we have $${{\Phi}}(F_j) \leq {{\Phi}}(E) + \frac{{{\Phi}}(F_j) - {{\Phi}}(E)}{2} + \frac{|{\gamma}_{{{\Phi}}}(E) - {\gamma}_{{{\Phi}}}(F_j)|}{2} .$$ and so $${{\Phi}}(F_j) \leq {{\Phi}}(E) + |{\gamma}_{{{\Phi}}}(E) - {\gamma}_{{{\Phi}}}(F_j)| .$$
As in the proof of Theorem \[Smooth\], the Hölder modulus of continuity for ${\gamma}_{{{\Phi}}}$ shown in Proposition \[convergence\](2) does not provide a sharp enough bound on the term $ |{\gamma}_{{{\Phi}}}(E) - {\gamma}_{{{\Phi}}}(F_j)| $; we must show that ${\gamma}_{{{\Phi}}}$ is Lipschitz when the centers of $E$ and $F_j$ are bounded away from their symmetric difference. In this case, we must be more careful and show that the Lipschitz constant is small when $|E| = |F|$ and $E$ and $F_j$ are $L^{\infty}$ close. If ${\gamma}_{{{\Phi}}}(E) \geq {\gamma}_{{{\Phi}}}(F_j)$, then using , we have $$\begin{aligned}
{\gamma}_{{{\Phi}}} (E ) - {\gamma}_{{{\Phi}}}(F_j) & \leq \int_E \frac{dx}{f_*(x-y_E)} - \int_{F_j} \frac{dx}{f_*(x-y_E)}= \int_{E\setminus F_j} \frac{dx}{f_*(x-y_E)} - \int_{F_j\setminus E} \frac{dx}{f_*(x-y_E)}.\end{aligned}$$ One easily shows from the definition that for any $x, y \in {\mathbb{R}^n}$, $$f_*(x) - \frac{1}{m_{{{\Phi}}}} |y| \leq f_*(x-y) \leq f_*(x) + \frac{1}{{m_{{{\Phi}}}}}|y|.$$ Therefore, since $(1- \eta_1) K \subset E\Delta F_j \subset (1+ \eta_1) K$ and $|y_E| \leq \eta_1,$ $$1- \eta_1 (1 +1/{m_{{{\Phi}}}}) \leq f_*(x-y_E) \leq 1+ \eta_1 (1 +1/{m_{{{\Phi}}}})$$ for $x \in E\Delta F_j$, implying that $$\begin{aligned}
{\gamma}_{{{\Phi}}}(E) - {\gamma}_{{{\Phi}}}(F_j)
& \leq \int_{E\setminus F_j} \frac{dx}{1-\eta_1 (1+ 1/{m_{{{\Phi}}}})}
- \int_{F_j\setminus E}\frac{dx}{1+\eta_1(1+1/{m_{{{\Phi}}}})}
\leq C \eta_1 |E\Delta F_j|.\end{aligned}$$ where $C= 1+ 1/{m_{{{\Phi}}}}$. The analogous argument holds if ${\gamma}_{{{\Phi}}}(E) \leq {\gamma}_{{{\Phi}}}(F_j),$ and so $${{\Phi}}(F_j) \leq {{\Phi}}(E) +C \eta_1 |E\Delta F_j|.$$ Letting $c_1 =1/C$, we conclude that $F_j$ is a volume constrained $({\epsilon}, \eta_0)$-minimizer of surface energy, and for $j$ large enough, $(1- \eta_0/2)K \subset F_j \subset (1+ \eta_0/2)K$ by Lemma \[bound\]. Therefore, Theorem \[figmag\] below implies that, for $j$ sufficiently large, $F_j$ is a convex polygon with $\nu_{F_j}(x) \in \{v_i\}_{i=1}^N$ for $\mathcal{H}^{1}\text{-a.e. } x \in \partial F_j.$ Moreover, for any $\eta,$ $(1- \eta)K \subset F_j \subset (1+ \eta)K$ for $j$ large enough, so actually $\{v_{F_j}\} = \{v_i\}_{i=1}^N$ for $j$ sufficiently large. In other words, for $j$ large enough, $F_j$ is parallel to $K$, so Proposition \[crystalprop\] implies that $${\beta_{{{\Phi}}}}(F_j)^2 \leq C_1 \delta_{{{\Phi}}}(F_j),$$ where $C_1$ depends on $f$. On the other hand, $F_j$ minimizes $Q_j$, so comparing against $E_j$ and using (\[contra2\]) and (\[epsbeta\]) implies $$\begin{aligned}
{{\Phi}}(F_j) & \leq {{\Phi}}(E_j)
\leq {{\Phi}}(K) + c_3 {\epsilon}_j^2 \leq {{\Phi}}(K) +2c_3{\beta_{{{\Phi}}}}(F_j)^{2}.
\end{aligned}$$ By , ${\beta_{{{\Phi}}}}(F_j)>0$, so choosing $c_3$ small enough such that $c_3 < |K|/C_1$, we reach a contradiction.
[@figallimaggi11 Theorem 7]\[figmag\] Let $n=2$ and let $f$ be a crystalline surface tension. There exists a constant ${\epsilon}_0$ such that if, for some $\eta>0$ and some $0<{\epsilon}<{\epsilon}_0$ , $(1-\eta/2)K \subset E \subset (1+\eta/2)K$ and $E$ is a volume constrained $({\epsilon}, \eta)$-minimizer, then $E$ is a convex polygon with $$\nu_E(x) \in \{v_i\}_{i=1}^N \qquad \text{for }\mathcal{H}^{1}\text{-a.e. } x \in \partial E.$$
[In [@figallimaggi11 Theorem 7], Figalli and Maggi assume that $E$ is a volume constrained $({\epsilon}, 3)$-minimizer (and actually, their notion of $({\epsilon}, 3)$-minimality is slightly stronger than ours). However, by adding the additional assumption that $(1-\eta/2) K \subset E \subset (1+\eta/2)K$, it suffices to take $E$ to be a volume constrained-$({\epsilon}, \eta)$ minimizer (with the definition given here) with $\eta$ as small as needed. Indeed, if $(1-\eta/2) K \subset E \subset (1+\eta/2)K$, then $(1-\eta)E \subset$ co$(E) \subset (1+\eta)E$ where co$(E)$ is the convex hull of $E$. Then, in the proof of [@figallimaggi11 Theorem 7], the only sets $F$ used as comparison sets are such that $|E| = |F|$ and $(1-\eta)E\subset F\subset (1+\eta)E$. ]{}
Another form of the oscillation index {#other}
=====================================
The oscillation index ${\beta_{{{\Phi}}}}(E)$ is the natural way to quantify the oscillation of the boundary of a set $E$ relative to the Wulff shape $K$ for a given surface energy ${{\Phi}}$, as it admits the stability inequality with a power that is independent of $f$. One may wonder if it would be suitable to quantify the oscillation of $E$ by looking at the Euclidean distance between normal vectors of $E$ and corresponding normal vectors of $K$. While such a quantity may be useful in some settings, in this section we show that it does not admit a stability result with a power independent of $f$. This section examines the term ${\beta_{{{\Phi}}}}^*(E)$ defined in and gives two examples showing a failure of stability. We then give a relation between ${\beta_{{{\Phi}}}}$ and ${\beta_{{{\Phi}}}}^*$ for ${\gamma}$-$\lambda$ convex surface tensions. As a consequence of Theorem \[th1\], this implies a stability result for ${\beta_{{{\Phi}}}}^*$, though, as the examples show, there is a necessary dependence on the ${\gamma}$-$\lambda$ convexity of $f$.
The following example illustrates that there does not exist a power $\sigma$ such that $$\label{noway}
\beta^*_{{{\Phi}}}(E)^{\sigma} \leq C(n,f) \delta_{{{\Phi}}}(E)$$ for all sets $E$ of finite perimeter with $0<|E|<\infty$ and for all surface energies ${{\Phi}}$.
\[crystalexample\]
In dimension $n=2$, we construct a sequence of Wulff shapes $K_{\theta}$ (equivalently, a sequence of surface energies ${{\Phi}}_{\theta}$) and a sequence of sets $E_{\theta}$ such that $\delta_{{\theta}} (E_{\theta}) \to0$ but $\beta^*_{{\theta}}(E_{\theta}) \to \infty$ as $\theta \to 0$. We use the notation $ \delta_{\theta}=\delta_{{{\Phi}}_{\theta}}$ and $ \beta^*_{\theta}=\beta^*_{{{\Phi}}_{\theta}}$.
We let $K_{\theta}$ be a unit area rhombus where one pair of opposing vertices has angle $\theta<\frac{\pi}{4} $ and the other has angle $\frac{\pi}{2} - \theta$. The length of each side of $K_{\theta}$ is proportional to $\theta^{-1/2}$. Let $L = \theta^{-1/4}$. We then construct the sets $E_{\theta}$ by cutting away a triangle with a zigzag base and with height $L$ from both corners of $K_{\theta}$ with vertex of angle $\theta$ (see Figure $3$). We choose the zigzag so that each edge in the zigzag is parallel to one of the adjacent edges of $K_{\theta}$. By taking each segment in the zigzag to be as small as we wish, we may make the area of each of the two zigzag triangles arbitrarily close to the area of the triangle with a straight base, which is $$A = L^2\tan(\theta/2) = \theta^{-1/2}\tan(\theta/2)\approx \theta^{1/2},$$ as this triangle has base $2L \tan(\theta/2).$ Both of the other two sides of the triangle have length $m= L/\cos(\theta/2).$
Let us now compute the deficit $\delta_{\theta}$ and the Euclidean oscillation index $\beta_{{\theta}}^*$ of $E_{\theta}$. By construction, ${{\Phi}}_{\theta}(E_{\theta}) = {{\Phi}}_{\theta}(K_{\theta}) = 2$, and therefore $$\delta_{{\theta}}(E_{\theta}) = \frac{2}{2(1- A)^{1/2}} - 1 = \frac{1}{(1- A)^{1/2}} -1 = \theta^{1/2} + o(\theta^{1/2}).$$ To compute $\beta^*_{\theta} (E_{\theta})^2$, we cannot characterize the point $y$ for which the minimum in is attained in general. However, something may be said for an $n$-symmetric set, i.e., a set $E$ that for which there exist $n$ orthogonal hyperplanes such that $E$ is invariant under reflection with respect to each of them. The intersection of these orthogonal hyperplanes is called the *center of symmetry of $E$*. Indeed, a slight variation in the proof of [@maggi2008some Lemma $5.2$] shows that $$\label{3beta}3{\beta_{{{\Phi}}}}^*(E) \geq \bigg( \frac{1}{n|K|^{1/n}|E|^{1/n'} }\int_{\partial^* E} 1-\nu_E(x) \cdot \nu_K\Big(\frac{x-z}{f_*(x-z)}\Big)\, d{\mathcal{H}^{n-1}}(x) \bigg)^{1/2}.$$ where $z$ is the center of symmetry of $E$. By construction, $E_{\theta}$ is a $2$-symmetric set with center of symmetry $0$, so $$\begin{aligned}
9\beta^*_{{\theta}} (E_{\theta})^2
&\geq
\frac{1}{2(1-A)^{1/2}} \int_{Z} 1 -\nu_{E_{\theta}} (x) \cdot \nu_{K_{\theta}}\Big(\frac{x}{f_*(x)}\Big)\, d \mathcal{H}^{1} \geq
\frac{1}{2} \int_{Z} 1 -\nu_{E_{\theta}} (x) \cdot \nu_{K_{\theta}}\Big(\frac{x}{f_*(x)}\Big)\, d \mathcal{H}^{1},
\end{aligned}$$ where $Z$ denotes the union of the two zigzags. By construction, $\mathcal{H}^1(Z)$ is exactly equal to $\mathcal{H}^{1}(\partial K_{\theta} \setminus \partial E_{\theta}) = 4m$. Moreover, because the edges of $E_{\theta}$ are parallel to those of $K_{\theta}$, we find that $$1 -\nu_{E_{\theta}} (x) \cdot \nu_{K_{\theta}}\Big(\frac{x}{f_*(x)}\Big) = \begin{cases} 0 & x \in Z_1\\
1-\cos(\pi - \theta) & x \in Z_2
\end{cases}$$ where $Z_1$ is the set of $x\in Z$ where $\nu_{E_{\theta}}(x)$ is equal to $\nu_{K_{\theta}}(\frac{x}{f_*(x)})$ and $Z_2$ is the set of $x \in Z$ where $\nu_{E_{\theta}}(x)$ is equal to the normal vector to the other side of $K_{\theta}.$ Moreover, we have constructed $E_{\theta}$ so that $\mathcal{H}^1(Z_1) = \mathcal{H}^1(Z_2) = 2 m.$ Thus, as $\theta<\frac{\pi}{4},$ $$\begin{aligned}
\beta^*_{{\theta}} (E_{\theta})^2 & \geq \frac{1}{2} \int_{Z_2} 1 - \cos (\pi - \theta) \ d \mathcal{H}^{1} \geq \frac{\mathcal{H}^1(Z_2)}{2} = m =1/(\theta^{1/4}\cos(\theta/2))\to \infty\end{aligned}$$ as $\theta \to 0$. Therefore, for any exponent $\sigma$, the inequality (\[noway\]) fails to hold; we may choose $\theta$ sufficiently small such that $E_{\theta}$ is a counterexample.
The next example shows that even if we restrict our attention to surface energies that are ${\gamma}$-$\lambda$ convex (Definition \[gammalambda\]), an inequality of the form in (\[noway\]) cannot hold with an exponent smaller than $\sigma = 4.$ The example is presented in dimension $n=2$ for convenience, though the analogous example in higher dimension also holds.
\[pexample\]
Fix $p>2$ and define the surface tension $f_p(x) = \left( |x_1|^p+ |x_2|^p \right)^{1/p}$ to be the $\ell^p$ norm in $\mathbb{R}^2$. We show below that $f_p$ is a ${\gamma}$-$\lambda$ convex surface tension. Hölder’s inequality ensures that the support function $f_*$ is given by $f_q$, in the notation above, where $q$ is the Hölder conjugate of $p$. The Wulff shape $K= \{f_q(x) <1\}$ is therefore the $\ell^q$ unit ball. We let ${{\Phi}}_p$ denote the surface energy corresponding to the surface tension $f_p$.
We build a sequence of sets $\{E_{r}\}$ depending on $p$ such that, for any $\sigma<4$, we may choose $p$ large enough so that $\delta_{p}(E_{r})/\beta^*_{p}(E_{r})^{\sigma}\to 0$ as $r \to 0$. Here we use the notation $\beta^*_{p}(E) = \beta^*_{{{\Phi}}_p}(E) $ and $\delta_p(E) = \delta_{{{\Phi}}_p}(E).$ We may locally parameterize $K$ near $(0,1)$ as the subgraph of the function $v_q(x_1) = \left(1 - |x_1|^q\right)^{1/q}.$ Thus $v_q'(x_1) = -|x_1|^{q-2}x_1/{(1- |x_1|^q )^{1/p}}$ and $$\label{nuKK}
\nu_K((x_1, \ v_q(x_1))) = \frac{\Big(\frac{|x_1|^{q-2}x_1}{(1- |x_1|^q )^{1/p}}, \ 1\Big) }{\sqrt{1+
\frac{|x_1|^{2q-2}}{(1-|x_1|^q )^{2/p}} }} =\frac{(|x_1|^{q-2}x_1 + O(|x_1|^{2q-1}) , \ 1) }{\sqrt{1+|x_1|^{2q-2} + O(|x_1|^{3q-2} ) }}$$ The sets $E_r$ are formed by replacing the top and bottom of $K$ with cones. More precisely, let $\textbf{C}_r = (-r,r) \times \mathbb{R}$. We form $E_r$ by replacing $\partial K \cap \textbf{C}_r$ with the graphs of the functions $w$ and $-w$, where $w_1: (-r, r) \to \mathbb{R}$ is defined by $w (x_1) = -r^{q-1} |x_1|/{(1-r^q)^{1/p}} + C_0,$ with $C_0= (1-r^q)^{1/q } + r^q/{(1-r^q)^{1/p}}.$ The constant $C_0$ is chosen so that $w(r) = v_q(r)$ and $w(-r) = v_q(-r)$.
For $x_1 \in (-r, r)$ for $r<1$, we have $w'(x_1) = - r^{q-1}\text{sgn}(x_1)/(1-r^q)^{1/p}$ and $$\label{nuCC}
\nu_E\left((x_1, \ w(x_1))\right) = \frac{ \Big( \text{sgn}(x_1) \frac{ r^{q-1}}{(1-r^q)^{1/p}} ,\ 1 \Big) }{\sqrt{ 1 +
\frac{ r^{2q-2}}{(1-r^q)^{2/p}}}}
=\frac{( \text{sgn}(x_1) r^{q-1} + O(r^{2q-1}), \ 1) }{\sqrt{ 1 + r^{2q-2} +O(r^{3q-2})}}.$$ Now, ${{\Phi}}_p(E_r ) = {{\Phi}}_p(K) + {{\Phi}}_p(E_r; \textbf{C}_r) - {{\Phi}}_p(K; \textbf{C}_r)$, so $$\begin{aligned}
{{\Phi}}_p(E_r) - {{\Phi}}_p(K)&
= \int_{-r}^r \Big( \frac{r^q}{1-r^q} +1\Big)^{1/p } - \Big( \frac{|x_1|^q}{1-|x_1|^q} +1\Big)^{1/p } dx_1\\
&=\frac{1}{p} \int_{-r}^r r^q - |x_1|^q + O(r^{2q}) \ dx_1 =Cr^{q+1} + o(r^{q+1} ).
\end{aligned}$$ The graph of $w$ lies above the graph of $v_q$ for all $|x_1|<r$, so $|E_r| >|K|.$ This implies that $$\delta_{p}(E_r) \leq \frac{{{\Phi}}_p(E_r) - {{\Phi}}_p(K)}{2|K|} =Cr^{q+1} + o(r^{q+1} ).$$
Next we compute $\beta_{p}^*(E_r)$ in several steps. As in Example \[crystalexample\], $E_r$ is a $2$-symmetric set with center of symmetry $0$, thus it is enough to compute the right hand side of . First, the Taylor expansions in and imply that, for $x\in {\bf{C}}_r \cap \partial^* E$, $ \nu_E(x) \cdot \nu_K\big( \frac{x}{f_*(x)} \big)$ is given by $$\begin{aligned}
&\frac{(|x_1|^{q-2}x_1 + O(|x_1|^{2q-1}) , \ 1) }{\sqrt{1+|x_1|^{2q-2}+ O(|x_1|^{3q-2}) }}\cdot \frac{( \text{sgn}(x_1) r^{q-1} + O(r^{2q-1}), \ 1)}{\sqrt{ 1 +
r^{2q-2} + O(r^{3q-2)}}} = \frac{ 1 + |x_1|^{q-1}r^{q-1} + O(r^{3q-2} )}{\sqrt{(1+|x_1|^{2q-2}+ r^{2q-2} + O( r^{4q-4}))}}\\
&= 1 + |x_1|^{q-1}r^{q-1} -\frac{1}{2} (|x_1|^{2q-2}+ r^{2q-2} ) + O( r^{3q-2})=1 -\frac{1}{2}( |x_1|^{q-1}-r^{q-1})^2+O( r^{3q-2} ).\end{aligned}$$ For $x \in \partial^* E \setminus {\bf{C}}_r$, $ \nu_E(x) \cdot \nu_K\big( \frac{x}{f_*(x)} \big) = 0$. Hence, $$\begin{aligned}
\bigg( \frac{1}{2} \int_{\partial^*E}& \left|\nu_E(x) - \nu_K\Big(\frac{x}{f_*(x)}\Big)\right|^2\, d\mathcal{H}^{1}\bigg)^{1/2} =\bigg( \int_{\partial^*E \cap {\bf{C}}_r} 1 - \nu_E \cdot \nu_K\big(\frac{x}{f_*(x)} \big)\,d\mathcal{H}^{1}\bigg)^{1/2}\\
& =\bigg( \int_{-r}^r \frac{1}{2} ( |x_1|^{q-1}-r^{q-1})^2 \sqrt{ 1 + r^{2q-2} +O(r^{3q-1})} + O(r^{3q-2} )
\, dx_1\bigg)^{1/2} \\
&=\bigg( \int_{-r}^r \frac{1}{2} ( |x_1|^{q-1}-r^{q-1})^2 + O(r^{3q-2} )
\, dx_1\bigg)^{1/2}
=Cr^{q-1/2} + o(r^{q-1/2}).\end{aligned}$$ Furthermore, $|E| = |K| + o(1)$, so $\sqrt{2}|K|^{-1/4} |E|^{-1/4}= \sqrt{2}|K|^{-1/2} +o(1)$, and so $$\begin{aligned}
\beta_p^*(E_r)&
= \frac{1}{\sqrt{2}|K|^{1/4} |E|^{1/4}}
\bigg( \frac{1}{2} \int_{\partial^* E} \Big|\nu_E (x) - \nu_K\Big(\frac{x}{f_*(x)} \Big)\Big|^2 d \mathcal{H}^1\bigg)^{1/2} =C r^{q -1/2} + o(r^{q -1/2}).\end{aligned}$$ Therefore, $$\frac{\delta_{p}(E_r) }{\beta_p^*(E_r)^{\sigma} } \approx \frac{r^{(q+1)}}{r^{\sigma(q - 1/2)}} = r^{ q+1 - \sigma q +\sigma/2}.$$ This quantity goes to $0$ as $r$ goes to zero if and only if $q+1 -\sigma q+ \sigma/2 >0,$ or, equivalently, if and only if $\frac{2 +\sigma}{2(\sigma-1)}>q$. For any $\sigma<4$ we may find $1<q <\frac{2 + \sigma}{2(\sigma-1)}.$ Therefore, for any $\sigma<4,$ there exists a $\gamma$-$\lambda$ convex surface tension $f$ such that a bound of the form $\delta_{{{\Phi}}}(E) \geq C{\beta_{{{\Phi}}}}^*(E)^{\sigma}$ fails.
When $f$ is ${\gamma}$-$\lambda$ convex (recall Definition \[gammalambda\]), we can control $\beta^*_{{{\Phi}}}(E)$ by ${\beta_{{{\Phi}}}}(E)$. As one expects after the previous example, the exponent in this bound depends on the ${\gamma}$-$\lambda$ convexity of ${{\Phi}}$. Indeed, this is the content of Theorem \[th2\]. First, we show that the $\ell_p$ norms $f_p$ as defined in the previous example are ${\gamma}$-$\lambda$ convex for each $p\in(1, \infty)$. In the case where $1<p\leq 2$, $f_p$ is actually *uniformly* convex in tangential directions, so it is ${\gamma}$-$\lambda$ convex with ${\gamma}=0.$ Indeed, $f_p(\nu+\tau) = f_p(\nu) + \nabla f_p(\nu)\tau + \frac{1}{2}\int_0^1 \nabla^2 f_p(\nu + s\tau)[\tau , \tau] ds,$ and thus $$f_p(\nu+\tau) +f_p(\nu-\tau) - 2 f_p(\nu) = \frac{1}{2} \int_{-1}^{1} \nabla^2f_p(\nu+ s\tau) [\tau, \tau] ds.$$ We can bound the integrand from below pointwise. We compute $$\partial_{ii} f_p(\nu ) = (p-1) \Big( \frac{ |\nu_i|^{p-2}}{f_p(\nu)^{p-1}} - \frac{ |\nu_i|^{2p-2}}{f_p(\nu)^{2p-1}} \Big) , \qquad
\partial_{ij}f_p(\nu) =(1-p) \frac{| \nu_i|^{p-2}\nu_i |\nu_j|^{p-2} \nu_j}{ f_p(\nu)^{2p-1}}.$$ Therefore, if $f_p(\nu) = 1,$ then $$\nabla^2 f_p(\nu) = (p-1) \sum_{i=1}^{n} |\nu_i|^{p-2} e_i \otimes e_i - (p-1) \sum_{i,j=1}^n
| \nu_i|^{p-2}\nu_i |\nu_j|^{p-2} \nu_j e_i \otimes e_j$$ and so $$\nabla^2f_p(\nu)[\tau, \tau] =
(p-1)\sum_{i=1}^{n} |\nu_i|^{p-2} \tau_i^2 - (p-1) \Big(\sum_{i=1}^n
| \nu_i|^{p-2}\nu_i \tau_i \Big)^2 .$$ It is enough to consider $\tau$ such that $\tau$ is tangent to $K_p=\{f_p<1\}$ at $\nu$, as $f_p$ is positive $1$-homogeneous and the span of $\nu$ and $T_{\nu}K_p$ is all of ${\mathbb{R}^n}$. Observe that $\nabla f_p(\nu) = \sum_{i=1}^n |\nu_i|^{p-2} \nu_i e_i$; this is verified by the fact that the support function of $f_p$ is $f_q$, and that $\nabla f_p(\nu) = \frac{x}{f_q(x)} $ such that $\frac{x}{f_q(x)} \cdot \nu = f_p(\nu) = 1.$ Thus $\tau$ is tangent to $K_p$ at $\nu$ if and only if $\tau \cdot \nabla f_p(\nu) = \sum_{i=1}^n
| \nu_i|^{p-2}\nu_i \tau_i = 0.$ Therefore, for such $\tau$, $$\nabla^2f_p(\nu)[\tau, \tau] =
(p-1) \sum_{i=1}^{n} |\nu_i|^{p-2} \tau_i^2 \geq (p-1) |\tau|^2.$$ In the case where $p\geq 2$, we use Clarkson’s inequality, which states that for $p \geq 2$, $$f_p\Big(\frac{x+y}{2}\Big)^p +f_p\Big(\frac{x-y}{2}\Big)^p \leq \frac{f_p(x)^p}{2} +\frac{ f_p(y)^p}{2}.$$ For $\nu$ such that $f_p(\nu) =1$ and $\tau $ tangent to $K_p $ at $\nu$ with $f_p(\tau)=1,$ Clarkson’s inequality with $x = \nu + {\epsilon}\tau$ and $y = \nu - {\epsilon}\tau$ implies $$2{\epsilon}^p \leq f_p(\nu + {\epsilon}\tau)^p + f_p(\nu - {\epsilon}\tau) - 2.$$ This is almost the condition we need, except we have $f_p^p$ instead of $f_p$ for the terms on the right hand side. Note that both $f_p(\nu + {\epsilon}\tau)$ and $f_p(\nu - {\epsilon}\tau)$ are greater than $1$, as moving in the tangent direction to $K_p= \{f_p <1\} $ increases $f_p$. The function $z^p$ is convex with derivative $pz^{p-1}$, so $z^p \leq 2^{p-1}pz + (2^{p-1}p-1)$ for all $z\in [1,2]$. Applying this to $z_1 = f_p(\nu+{\epsilon}\tau)$ and $z_2 = f_p(\nu-{\epsilon}\tau)$ yields $$2{\epsilon}^p \leq 2^{p-1}p f_p(\nu+{\epsilon}\tau)^p +2^{p-1}pf_p(\nu - {\epsilon}\tau)^p -2(2^{p-1}p).$$ Thus $f_p$ is ${\gamma}$-$\lambda$ convex with ${\gamma}= p-2$ and $\lambda =1/(2^{p-2}p).$
The following lemma about ${\gamma}$-$\lambda$ convexity condition will be used in the proof of Theorem \[th2\].
\[lem5\] Assume that $f$ is ${\gamma}$-$\lambda$ convex. Then for all $\nu, \tau \in {\mathbb{R}^n}$ such that $\nu \neq 0$, $$\label{frombelow}f(\nu + \tau ) \geq \frac{\lambda}{2^{2+{\gamma}}|\nu|} \Big|\tau -\Big(\tau \cdot\frac{\nu}{|\nu|}\Big) \frac{\nu}{|\nu|}\Big|^{2+{\gamma}}
+f(\nu) +\nabla f(\nu)\cdot \tau,$$
Note that if $f$ is ${\gamma}$-$\lambda$ convex, then $f$ is convex. To see that holds for given $\nu_0$ and $\tau_0$, we let $\tilde{f}(\nu) = f(\nu) - f(\nu_0) - \nabla f(\nu_0)\cdot(\nu - \nu_0).$ At the midpoint $\nu_0 + \frac{\tau_0}{2}$, the ${\gamma}$-$\lambda$ convexity condition gives us the following: $$\tilde{f}(\nu_0) + \tilde{f}(\nu_0 + \tau_0) - 2\tilde{f}(\nu_0 +\frac{\tau_0}{2} ) \geq \frac{\lambda}{|\nu_0|} \Big|\frac{\tau_0}{2} - \Big(\frac{\tau_0}{2}\cdot \frac{\nu_0}{|\nu_0|}\Big) \frac{\nu_0}{|\nu_0|}\Big|^{2+{\gamma}}.$$ Convexity implies that $\tilde{f}(\nu_0 + \frac{\tau_0}{2}) \geq 0$, and $\tilde{f}(\nu_0) = 0$ by definition of $\tilde{f}$, implying .
Finally, we prove Theorem \[th2\].
The quantity $\beta^*_{{{\Phi}}}(E)$ measures the overall size of the Cauchy-Schwarz deficit on the boundary of $E$, while ${\beta_{{{\Phi}}}}(E)$ measures the overall deficit in the Fenchel inequality. Our aim is to obtain a pointwise bound of the Cauchy-Schwarz deficit functional by the Fenchel deficit functional, and then integrate over the reduced boundary of $E$. Without loss of generality, we may assume that $|E|= |K| = 1 $ and $E$ has center zero in the sense defined in Section \[abg\].
We fix $x\in \partial^*E$ and consider the Fenchel deficit functional $G(\nu) = f(\nu) -\nu \cdot \frac{x}{f_*(x)}$, which possesses the properties that $G(\nu)\geq 0$ and $G(\nu )= 0$ if and only if $\nu =c\, \nabla f_*(x)$ for some $c>0$.
Let $w=\frac{\nabla f_*(x)}{|\nabla f_*(x)|} = \nu_K(\frac{x}{f_*(x)}).$ Lemma \[lem5\], with $\nu = w $ and $\tau = \nu_E -w$, implies that $$f(\nu_E )
\geq \frac{\lambda}{2^{2+{\gamma}}} |(\nu_E - w) -((\nu_E - w) \cdot w) w|^{2+{\gamma}}
+f(w) +\nabla f(w)\cdot (\nu_E - w).$$ Therefore, since $\nabla f(w) = \frac{x}{f_*(x)} $ and $f(w) = \nabla f(w) \cdot w$, $$\begin{aligned}
G(\nu_E) \geq \frac{\lambda}{2^{2+{\gamma}}} \left|(\nu_E - w ) - ((\nu_E - w)\cdot w) w\right|^{2+{\gamma}}& = \frac{\lambda}{2^{2+{\gamma}}}
(1-(\nu_E\cdot w)^2)^{(2+{\gamma})/2}\\
&= \frac{\lambda}{2^{2+{\gamma}}}((1-\nu_E\cdot w)(1+\nu_E\cdot w))^{(2+{\gamma})/2}.\end{aligned}$$ We want to show that there exists some $c_1$ such that $$\label{lastone}G(\nu_E) \geq c_1 (1 - \nu_E \cdot w)^{(2+{\gamma})/2}.$$ When $ w \cdot \nu_E \geq -c_0$ for some fixed $0<c_0<1,$ then $G(\nu_E) \geq \frac{\lambda}{2^{2+{\gamma}}}(1-c_0)^{(2+{\gamma})/2} (1-\nu_E\cdot w)^{(2+{\gamma})/2}$ and holds. On the other hand, when $ w \cdot \nu_E < -c_0$ for $c_0$ small, we expect that $\frac{x}{f_*(x)}\cdot \nu_E$ must also be small and so $G(\nu_E)$ is not too small. Indeed, $$m_{{{\Phi}}} \leq f( w) =\frac{x}{f_*(x)} \cdot w = \frac{|x|}{f_*(x)} \cos (\theta_1)\leq M_{{{\Phi}}} \cos( \theta_1),$$ where $\theta_1$ is the angle between $ w$ and $\frac{x}{f_*(x)}.$ Similarly, $$-c_0 \geq \nu_E \cdot w = \cos( \theta_2),$$ where $\theta_1$ is the angle between $ w$ and $\nu_E$. Noting that $0<m_{{{\Phi}}}/M_{{{\Phi}}}<1$, so $\cos^{-1} (m_{{{\Phi}}}/M_{{{\Phi}}} )\in (0, \pi/2), $ we let $\theta_0 =2 \cos^{-1 }(m_{{{\Phi}}}/M_{{{\Phi}}}) + {\epsilon},$ where ${\epsilon}>0$ is chosen small enough so that $\theta_0<\pi.$ Letting $c_0 = - \cos(\theta_0),$ we deduce that $\theta_1 \leq \cos^{-1} (m_{{{\Phi}}}/M_{{{\Phi}}})$ and $\theta_2 \geq \theta_0.$ Then $$\frac{x}{f_*(x)} \cdot \nu_E \leq \frac{|x|}{f_*(x)} \cos (\theta_2 - \theta_1) \leq M_{{{\Phi}}} \cos\left( \cos^{-1}(m_{{{\Phi}}}/M_{{{\Phi}}})+ {\epsilon}\right) \leq m_{{{\Phi}}} - M_{{{\Phi}}} c_{{\epsilon}},$$ for a constant $c_{{\epsilon}}>0$. Since $f(\nu_E) \geq m_{{{\Phi}}}$, we have $G(\nu_E)\geq M_{{{\Phi}}} c_{{\epsilon}}$, implying because $1 -\nu_E \cdot w\leq 2$.
Hölder’s inequality and imply $$\begin{aligned}
\int_{\partial^* E} 1 - \nu_E \cdot w \ d{\mathcal{H}^{n-1}}&
\leq
{\mathcal{H}^{n-1}}(\partial^*E)^{{\gamma}/(2+{\gamma})} \Big( \int_{\partial^* E} (1-\nu_E\cdot w)^{(2+{\gamma})/2}d{\mathcal{H}^{n-1}}\Big)^{2/(2+{\gamma})}
\\
&=
c_1^{-2/(2+{\gamma})} P(E)^{{\gamma}/(2+{\gamma})}
\Big( \int_{\partial^* E}c_1 (1-\nu_E\cdot w)^{(2+{\gamma})/2}d{\mathcal{H}^{n-1}}\Big)^{2/(2+{\gamma})} \\
&\leq
c_1^{-2/(2+{\gamma})} P(E)^{{\gamma}/(2+{\gamma})} \Big( \int_{\partial^* E} G(\nu_E)d{\mathcal{H}^{n-1}}\Big)^{2/(2+{\gamma})}.\end{aligned}$$ Dividing by $n|K|^{1/n}|E|^{1/n'}$ and taking the square root, we obtain $${\beta_{{{\Phi}}}}^*(E) \leq c_1^{-1/(2+{\gamma})} \Big(\frac{P(E)}{n|K|^{1/n}|E|^{1/n'}}\Big)^{{\gamma}/2(2+{\gamma})} \beta_{{{\Phi}}}(E)^{2/(2+{\gamma})}.$$
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present here our approach to the GermEval 2019 Task 1 - Shared Task on hierarchical classification of German blurbs. We achieved first place in the hierarchical subtask B and second place on the root node, flat classification subtask A. In subtask A, we applied a simple multi-feature TF-IDF extraction method using different n-gram range and stopword removal, on each feature extraction module. The classifier on top was a standard linear SVM. For the hierarchical classification, we used a local approach, which was more light-weighted but was similar to the one used in subtask A. The key point of our approach was the application of a post-processing to cope with the multi-label aspect of the task, increasing the recall but not surpassing the precision measure score.'
author:
- |
Fernando Benites\
` [email protected]`\
Zurich University of Applied Sciences,\
Switzerland\
title: 'TwistBytes - Hierarchical Classification at GermEval 2019: walking the fine line (of recall and precision)'
---
Introduction
============
Hierarchical Multi-label Classification (HMC) is an important task in Natural Language Processing (NLP). Several NLP problems can be formulated in this way, such as patent, news articles, books and movie genres classification (as well as many other classification tasks like diseases, gene function prediction). Also, many tasks can be formulated as hierarchical problem in order to cope with a large amount of labels to assign to the sample, in a divide and conquer manner (with pseudo meta-labels). A theoretical survey exists [@silla2011survey] discussing on how the task can be engaged, several approaches and the prediction quality measures. Basically, the task in HMC is to assign a sample to one or many nodes of a Directed Acyclic Graph (DAG) (in special cases a tree) based on features extracted from the sample. In the case of possible multiple parent, the evaluation of the prediction complicates heavily, for once since several paths can be taken, but only in a joining node must be considered.
The GermEval 2019 Task 1 - Shared Task on hierarchical classification of German blurbs focus on the concrete challenge of classifying short descriptive texts of books into the root nodes (subtask A) or into the entire hierarchy (subtask B). The hierarchy can be described as a tree and consisted of 343 nodes, in which there are 8 root nodes. With about 21k samples it was not clear if deep learning methods or traditional NLP methods would perform better. Especially, in the subtask A, since for subtask B some classes had only a few examples. Although an ensemble of traditional and deep learning methods could profit in this area, it is difficult to design good heterogeneous ensembles.
Our approach was a traditional NLP one, since we employed them successfully in several projects [@BenitesdeAzevedoeSouza2017Multi; @benites2017ST; @benites2019twistbytes], with even more samples and larger hierarchies. We compared also new libraries and our own implementation, but focused on the post-processing of the multi-labels, since this aspect seemed to be the most promising improvement to our matured toolkit for this task. This means but also, to push recall up and hope to not overshot much over precision.
Related Work
============
The dataset released by [@lewis2004ran] enabled a major boost in HMC on text. This was a seminating dataset since not only was very large (800k documents) but the hierarchies were large (103 and 364). Many different versions were used in thousands of papers. Further, the label density [@Tsoumakas07] was considerably high allowing also to be treated as multi-label, but not too high as to be disregarded as a common real-world task. Some other datasets were also proposed ([@partalas2015lshtc], [@MenciaF10]), which were far more difficult to classify. This means consequently that a larger mature and varied collection of methods were developed, from which we cannot cover much in this paper.
An overview of hierarchical classification was given in [@silla2011survey] covering many aspects of the challenge. Especially, there are local approaches which focus on only part of the hierarchy when classifying in contrast to the global (big bang) approaches.
A difficult to answer question is about which hierarchical quality prediction measure to use since there are dozens of. An overview with a specific problem is given in [@brucker2011empirical]. An approach which was usually taken was to select several measures, and use a vote, although many measures inspect the same aspect and therefore correlate, creating a bias. The GermEval competition did not take that into account and concentrates only on the flat micro F-1 measures[^1].
Still, a less considered problem in HMC is the number of predicted labels, especially regarding the post-processing of the predictions[^2]. We discussed this thoroughly in [@BenitesdeAzevedoeSouza2017Multi]. The main two promising approaches were proposed by [@Yang99] and [@read2009classifier]. The former focuses on column and row based methods for estimating the appropriate threshold to convert a prediction confidence into a label prediction. [@read2009classifier] used the label cardinality ([@Tsoumakas07]), which is the mean average label per sample, of the training set and change the threshold globally so that the test set achieved similar label cardinality.
Data and Methodology
====================
Task Definition and Data Description
------------------------------------
The shared task aimed at Hierarchical Multi-label Classification (HMC) of Blurbs. Blurbs are short texts consisting of some German sentences. Therefore, a standard framework of word vectorization can be applied. There were 14548 training, 2079 development, and 4157 test samples.
The hierarchy can be considered as an ontology, but for the sake of simplicity, we regard it as a simple tree, each child node having only on single parent node, with 4 levels of depth, 343 labels from which 8 are root nodes, namely: ’Literatur & Unterhaltung’, ’Ratgeber’, ’Kinderbuch & Jugendbuch’, ’Sachbuch’, ’Ganzheitliches Bewusstsein’, ’Glaube & Ethik’, and ’Künste, Architektur & Garten’.
The label cardinality of the training dataset was about 1.070 (train: 1.069, dev: 1.072) in the root nodes, pointing to a clearly low multi-label problem, although there were samples with up to 4 root nodes assigned. This means that the traditional machine learning systems would promote single label predictions. Subtask B has a label cardinality of 3.107 (train: 3.106, dev: 3.114), with 1 up to 14 labels assigned per sample. Table \[tab:td-ns\] shows a short dataset summary by task.
Task samples labels cardinality density
----------- --------- -------- ------------- ---------
subtask A 20,784 8 1.069 0.1336
subtask B 20,784 343 3.11 0.0091
: Specs for dataset for subtasks A and B []{data-label="tab:td-ns"}
System Definition
-----------------
We used two different approaches for each subtask. In subtask A, we used a heavier feature extraction method and a linear Support-Vector-Machine (SVM) whereas for subtask B we used a more light-weighted feature extraction with same SVM but in a local hierarchical classification fashion, i.e. for each parent node such a base classifier was used. We describe in the following the approaches in detail. They were designed to be light and fast, to work almost out of the box, and to easily generalise.
### Classifiers {#sec:gdi_clf}
#### Base Classifier
For subtask A, we use the one depicted in Fig. \[fig:mc\], for subtask B, a similar more light-weight approach was employed as base classifier (described later). As can be seen, several vectorizer based on different n-grams (word and character) with a maximum of 100k features and preprocessing, such as using or not stopwords, were applied to the blurbs. The term frequency obtained were then weighted with inverse document frequency (TF-IDF). The results of five different feature extraction and weighting modules were given as input for a vanilla SVM classifier (parameter C=1.5) which was trained in a one-versus-all fashion.
### Hierarchical Classifier
We use a local parent node strategy, which means the parent node decides which of its children will be assigned to the sample. This creates also the necessity of a virtual root node. For each node the same base classifier is trained independently of the other nodes. We also adapt each feature extraction with the classifier in each single node much like [@paes2014exploring]. As base classifier, a similar one to Fig. \[fig:mc\] was used, where only one 1-7 word n-gram, one 1-3 word n-gram with German stopwords removal and one char 2-3 n-gram feature extraction were employed, all with maximum 70k features. We used two implementations achieving very similar results. In the following, we give a description of both approaches.
#### Recursive Grid Search Parent Node
Our implementation is light-weighted and optimized for a short pipeline, however for large amount of data, saving each local parent node model to the disk. However, it does not conforms the way scikit-learn is designed. Further, in contrast to the Scikit Learn Hierarchical, we give the possibility to optimize with a grid search each feature extraction and classifier per node. This can be quite time consuming, but can also be heavily parallelized. In the final phase of the competition, we did not employ it because of time constrains[^3] and the amount of experiments performed in the Experiments Section was only possible with a light-weighted implementation.
#### Scikit Learn Hierarchical
Scikit Learn Hierarchical[^4] (Hsklearn) was forked and improved to deal with multi-labels for the task[^5], especially, allowing each node to perform its own preprocessing. This guaranteed that the performance of our own implementation was surpassed and that a contribution for the community was made. This ensured as well that the results are easily reproducible.
### Post-processing: Threshold
Many classifiers can predict a score or confidence about the prediction. Turning this score into the prediction is usually performed by setting a threshold, such as 0 and 0.5, so labels which have a score assigned greater than that are assigned to the sample. This might be not the optimal threshold in the multi-label classification setup and there are many approaches to set it ([@Yang99]). Although these methods concentrate in the sample or label, we have had good results with a much more general approach.
As described in [@BenitesdeAzevedoeSouza2017Multi], Read and Pfahringer [@read2009classifier] introduce a method (referred hereinafter to as LCA) to estimate the threshold globally. Their method chooses the threshold that minimizes the difference between the label cardinality of the training set and the predicted set. $$t = \underset{t \in [0,1]}{argmin} |LCard(D_T ) - LCard(H_t(D_S))|$$ where $LCard(D_T)$ denotes the label cardinality of training set and $LCard(H_t(D_S))$ the label cardinality of the predictions on test set if $t$ was applied as the threshold. For that the predictions need to be normalized to unity[^6]. We also tested this method not for the label cardinality over all samples and labels but only labelwise. In our implementation, the scores of the SVM were not normalized, which produced slightly different results from a normalized approach.
For the HMC subtask B, we used a simple threshold based on the results obtained for LCA. Especially, using multiple models per node could cause a different scaling.
Alternative approaches
----------------------
We also experimented with other different approaches. The results of the first two were left out (they did not perform better), for the sake of conciseness.
- Meta Crossvalidation Classifier: [@benites2019twistbytes]
- Semi-Supervised Learning: [@jauhiainen2018heli; @benites2019twistbytes]
- Flair: Flair [@akbik2018coling] with different embeddings (BERT (out of memory)[^7], Flair embeddings (forward and backward German)). Such sophisticated language models require much more computational power and many examples per label. This was the case for the subtask A but subtask B was not.
Experiments {#sec:results}
===========
We divide this Section in two parts, in first we conduct experiments on the development set and in the second on the test set, discussing there the competition results.
Preliminary Experiments on Development Set
------------------------------------------
The experiments with alternative approaches, such as Flair, meta-classifier and semi-supervised learning[^8] yielded discouraging results, so we will concentrate in the SVM-TF-IDF methods. Especially, semi-supervised proved in other setups very valuable, here it worsened the prediction quality, so we could assume the same “distribution” of samples were in the training and development set (and so we concluded in the test set).
In Table \[tab:results-dev-a\], the results of various steps towards the final model can be seen. An SVM-TF-IDF model with word unigram already performed very well. Adding more n-grams did not improve, on the contrary using n-grams 1-7 decreased the performance. Only when removing stopwords it improved again, but then substantially. Nonetheless, a character 2-3 n-gram performed best between these simple models. This is interesting, since this points much more to not which words were used, but more on the phonetics[^9].
Using the ensemble feature model produced the best results without post-processing. The simple use of a low threshold yielded also astonishingly good results. This indicates that the SVM’s score production was very good, yet the threshold 0 was too cautious.
-------------------------------------------------------------------------------- -- --
**Nr.& **Method& **micro F-1\
1&SVM-TF-IDF, word unigram&0.7965\
2&SVM-TF-IDF, word unigram, t=-0.25&0.8234\
3&SVM-TF-IDF, word n-gram (1-7)& 0.7875\
4&SVM-TF-IDF, word n-gram (1-7), t=-0.25&0.8152\
5&SVM-TF-IDF, word n-gram (1-3), stopwords& 0.8075\
6&SVM-TF-IDF, word n-gram (1-3), stopwords, t=-0.25 & 0.8240\
7&SVM-TF-IDF, char n-gram (2-3)& 0.8205\
8&SVM-TF-IDF, char n-gram (2-3), t=-0.25& 0.8332\
9&SVM-TF-IDF, feat. ensemble& 0.8414\
10&SVM-TF-IDF, feat. ensemble, threshold LCA& 0.8545\
10&SVM-TF-IDF, feat. ensemble, threshold LCA normed & 0.8534\
11&SVM-TF-IDF, feat. ensemble, threshold LCA-labelwise& **0.8603\
12&SVM-TF-IDF, feat. ensemble, threshold -0.25&0.8540\
13&SVM-TF-IDF, feat. ensemble, threshold -0.2& **0.8557\
14&Flair Embeddings German (forward,backward), 60 epochs& 0.8151\
15&SVM-TF-IDF, feat. ensemble, threshold LCA, fixing null& **0.8577\
16&SVM-TF-IDF, feat. ensemble, threshold LCA-labelwise, fixing null& **0.8623\
**************
-------------------------------------------------------------------------------- -- --
-0.5cm ![Threshold/micro F-1 dependency[]{data-label="fig:threshold"}](threshold_graph_2a.png "fig:"){width="9cm"}
In Fig. \[fig:threshold\], a graph showing the dependency between the threshold set and the micro F-1 score achieved in the development set is depicted. The curve fitted was $a*x^2+b*x+c$ which has the maximum at approx. -0.2. We chose -0.25 in the expectation that the test set would not be exactly as the development set and based on our previous experience with other multi-label datasets (such as the RCv1-v2) which have an optimal threshold at -0.3. Also as we will see, the results proved us right achieving the best recall, yet not surpassing the precision score. This is a crucial aspect of the F-1 measure, as it is the harmonic mean it will push stronger and not linearly the result towards the lower end, so if decreasing the threshold, increases the recall linearly and decreases also the precision linearly, balancing both will consequently yield a better F-1 score.
Although in Fig. \[fig:threshold\], the curve fitted is parabolic, in the interval between -0.2 and 0, the score is almost linear (and strongly monotone decreasing) giving a good indication that at least -0.2 should be a good threshold to produce a higher F-1 score without any loss.
Even with such a low threshold as -0.25, there were samples without any prediction. We did not assign any labels to them, as such post-process could be hurtful in the test set, although in the development it yielded the best result (fixing null).
---------------------------- ------- ----- ----- ------ ------- ----- ----- ------ ------- ----- ----- ------
tn fp fn tp tn fp fn tp tn fp fn tp
Architektur & Garten 2062 0 4 13 2061 1 2 15 2061 1 2 15
Ganzheitliches Bewusstsein 1959 8 45 67 1951 16 29 83 1951 16 30 82
Glaube & Ethik 1986 3 31 59 1983 6 23 67 1984 5 24 66
Kinderbuch & Jugendbuch 1783 8 80 208 1759 32 50 238 1762 29 51 237
Künste 2061 0 6 12 2061 0 4 14 2061 0 4 14
Literatur & Unterhaltung 874 98 58 1049 801 171 31 1076 807 165 31 1076
Ratgeber 1799 20 110 150 1781 38 75 185 1785 34 77 183
Sachbuch 1701 40 148 190 1672 69 106 232 1674 67 111 227
Total 14225 177 482 1748 14069 333 320 1910 14085 317 330 1900
---------------------------- ------- ----- ----- ------ ------- ----- ----- ------ ------- ----- ----- ------
In Table \[tab:results-confsubtaska\], the results of the one-vs-all approach regarding the true negative, false positives, false negatives and true positives for the different threshold 0, -0.25 and LCA are shown. Applying other threshold than 0 caused the number of true positives to increase without much hurting the number of true negatives. In fact, the number of false positives and false negatives became much more similar for -0.25 and LCA than for 0. This results in the score of recall and precision being also similar, in a way that the micro F-1 is increased. Also, the threshold -0.25 resulted that the number of false positive is greater than the number of false negatives, than for example -0.2. LCA produced similar results, but was more conservative having a lower false positive and higher true negative and false negative score.
We also noticed that the results produced by subtask A were better than that of subtask B for the root nodes, so that a possible crossover between the methods (flat and hierarchical) would be better, however we did not have the time to implement it. Although having a heavier feature extraction for the root nodes could also perform similar (and decreasing complexity for lower nodes). We use a more simple model for the subtask B so that it would be more difficult to overfit.
--------------------------------------- --
**Method& **micro F-1\
Hsklearn & 0.6544\
Hsklearn, t=-0.25 &**0.6758\
Hsklearn, t=-0.2 & 0.6749\
Hsklearn, LCA normalized & 0.6645\
Hsklearn, LCA & 0.6717\
Hsklearn extended & 0.6589\
Hsklearn extended, t=-0.25 &**0.6750\
Hsklearn extended, t=-0.2 & **0.6765\
own imp. & 0.6541\
own imp., t=-0.25 & 0.6704\
own imp., t=-0.2 & 0.6715\
**********
--------------------------------------- --
: Preliminary experiments on subtask B, best three values marked in bold[]{data-label="tab:dev-b"}
Table \[tab:dev-b\] shows the comparison of the different examined approaches in subtask B in the preliminary phase. Both implementations, Hsklearn and our own produced very similar results, so for the sake of reproducibility, we chose to continue with Hsklearn. We can see here, in contrary to the subtask A, that -0.25 achieved for one configuration better results, indicating that -0.2 could be overfitted on subtask A and a value diverging from that could also perform better. The extended approach means that an extra feature extraction module was added (having 3 instead of only 2) with n-gram 1-2 and stopwords removal. The LCA approach yielded here a worse score in the normalized but almost comparable in the non-normalized. However, the simple threshold approach performed better and therefore more promising.
Subtask A
---------
[|l|c|c|c|c|]{} **Rank& **Teamname& **precision& **recall& **micro F-1\
**********
1& EricssonResearch&0.8923 &0.8432& 0.8670\
- &twistbytes LCA fixing null& 0.8536& 0.8790& 0.8661\
- &twistbytes LCA-labelwise fixing null& 0.8536 & 0.8763 &0.8648\
2& twistbytes&0.8650&0.8617 &0.8634\
3& DFKI-SLT&0.8760& 0.8472 &0.8614\
4& Raghavan&0.8777& 0.8383 &0.8575\
5& knowcup&0.8525& 0.8362 &0.8443\
6& fosil-hsmw&0.8427& 0.832& 0.8373\
7& Averbis&0.8609& 0.8083 &0.8337\
8& HSHL1&0.8244& 0.8159 &0.8201\
9& Comtravo-DS&0.8144& 0.8255& 0.8199\
10& HUIU&0.8063& 0.8072 &0.8067\
11& LT-UHH&0.8601& 0.7481 &0.8002\
In Table \[tab:results-a\], the best results by team regarding micro F-1 are shown. Our approach reached second place. The difference between the first four places were mostly 0.005 between each, showing that only a minimal change could lead to a place switching. Also depicted are not null improvements results, i.e. in a following post-processing, starting from the predictions, the highest score label is predicted for each sample, even though the score was too low. It is worth-noting that the all but our approaches had much higher precision compared to the achieved recall.
Despite the very high the scores, it will be difficult to achieve even higher scores with simple NLP scores. Especially, the n-gram TF-IDF with SVM could not resolve descriptions which are science fiction, but are written as non-fiction book[^10], where context over multiple sentences and word groups are important for the prediction.
Subtask B
---------
--------------------------------------------------------- -- -- -- --
**Rank& **Teamname& **precision& **recall& **micro F-1\
1&twistbytes&0.7072& 0.6487&0.6767\
2&EricssonResearch&0.7377& 0.6174&0.6722\
3&knowcup&0.7507& 0.5808&0.6549\
4&Averbis&0.677& 0.614&0.644\
5&DFKI-SLT&0.7777& 0.5151&0.6197\
6&HSHL1&0.7216& 0.5375&0.6161\
7&Comtravo-DS&0.7042& 0.5274&0.6031\
8<-UHH&0.8496& 0.3892&0.5339\
9&NoTeam&0.4166& 0.276&0.332\
10&DexieDuo&0.0108& 0.0034&0.0052\
**********
--------------------------------------------------------- -- -- -- --
The best results by team of subtask B are depicted in Table \[tab:results-b\]. We achieved the highest micro F-1 score and the highest recall. Setting the threshold so low was still too high for this subtask, so precision was still much higher than recall, even in our approach. We used many parameters from subtask A, such as C parameter of SVM and threshold. However, the problem is much more complicated and a grid search over the nodes did not complete in time, so many parameters were not optimised. Moreover, although it is paramount to predict the parent nodes right, so that a false prediction path is not chosen, and so causing a domino effect, we did not use all parameters of the classifier of subtask A, despite the fact it could yield better results. It could as well have not generalized so good.
The threshold set to -0.25 shown also to produce better results with micro F-1, in contrast to the simple average between recall and precision. This can be seen also by checking the average value between recall and precision, by checking the sum, our approach produced 0.7072+0.6487 = 1.3559 whereas the second team had 0.7377+0.6174 = 1.3551, so the harmonic mean gave us a more comfortable winning marge.
Conclusion
==========
We achieved first place in the most difficult setting of the shared Task, and second on the “easier” subtask. We achieved the highest recall and this score was still lower as our achieved precision (indicating a good balance). We could reuse much of the work performed in other projects building a solid feature extraction and classification pipeline. We demonstrated the need for post-processing measures and how the traditional methods performed against new methods with this problem. Further, we improve a hierarchical classification open source library to be easily used in the multi-label setup achieving state-of-the-art performance with a simple implementation.
The high scoring of such traditional and light-weighted methods is an indication that this dataset has not enough amount of data to use deep learning methods. Nonetheless, the amount of such datasets will probably increase, enabling more deep learning methods to perform better.
Many small improvements were not performed, such as elimination of empty predictions and using label names as features. This will be performed in future work.
Acknowledgements
================
We thank Mark Cieliebak and Pius von Däniken for the fruitful discussions. We also thank the organizers of the GermEval 2019 Task 1.
[14]{} natexlab\#1[\#1]{}
Alan Akbik, Duncan Blythe, and Roland Vollgraf. 2018. Contextual string embeddings for sequence labeling. In *[COLING]{} 2018, 27th International Conference on Computational Linguistics*, pages 1638–1649.
Fernando Benites. 2017. *[Multi-label Classification with Multiple Class Ontologies]{}*. Ph.D. thesis, University of Konstanz, Konstanz.
Fernando Benites and Mark Cieliebak. 2017. Hierarchical classification for news articles. In *SwissText 2017 : 2nd Swiss Text Analytics Conference*.
Fernando Benites, Pius von D[ä]{}niken, and Mark Cieliebak. 2019. Twistbytes-identification of cuneiform languages and german dialects at vardial 2019. In *Proceedings of the Sixth Workshop on NLP for Similar Languages, Varieties and Dialects*, pages 194–201.
Florian Brucker, Fernando Benites, and Elena Sapozhnikova. 2011. An empirical comparison of flat and hierarchical performance measures for multi-label classification with hierarchy extraction. In *International Conference on Knowledge-Based and Intelligent Information and Engineering Systems*, pages 579–589. Springer.
Tommi Jauhiainen, Heidi Jauhiainen, and Krister Lind[é]{}n. 2018. . In *Proceedings of the Fifth Workshop on NLP for Similar Languages, Varieties and Dialects (VarDial 2018)*, pages 254–262.
David D. Lewis, Yiming Yang, Tony G. Rose, and Fan Li. 2004. Rcv1: A new benchmark collection for text categorization research. *J. Mach. Learn. Res.*, 5:361–397.
Eneldo Loza Mencía and Johannes F[ü]{}rnkranz. 2010. Efficient multilabel classification algorithms for large-scale problems in the legal domain. In *Semantic Processing of Legal Texts*, pages 192–215.
Bruno C Paes, Alexandre Plastino, and Alex Alves Freitas. 2014. Exploring attribute selection in hierarchical classification.
Ioannis Partalas, Aris Kosmopoulos, Nicolas Baskiotis, Thierry Artieres, George Paliouras, Eric Gaussier, Ion Androutsopoulos, Massih-Reza Amini, and Patrick Galinari. 2015. Lshtc: A benchmark for large-scale text classification. *arXiv preprint arXiv:1503.08581*.
Jesse Read, Bernhard Pfahringer, Geoff Holmes, and Eibe Frank. 2009. . In *Joint European Conference on Machine Learning and Knowledge Discovery in Databases*, pages 254–269. Springer.
Carlos N Silla and Alex A Freitas. 2011. A survey of hierarchical classification across different application domains. *Data Mining and Knowledge Discovery*, 22(1-2):31–72.
Grigorios Tsoumakas and Ioannis Katakis. 2007. Multi-label classification: An overview. *Int J Data Warehousing and Mining*, 2007:1–13.
Y. Yang. 1999. . *Information retrieval*, 1(1):69–90.
[^1]: The harmonic mean between micro recall and precision gives more weight for the predominantely label. Many new tasks consider the macro averaged F-1 since it gives equal weights for all labels which can be interesting for a large amount of labels (or samples to come).
[^2]: This is especially important if macro F-1 is used as quality prediction measure, in order to predict as many labels as possible.
[^3]: The system was trained on a Intel Xeon 32 cores and 100 Gb RAM.
[^4]: https://github.com/globality-corp/sklearn-hierarchical-classification/
[^5]: <https://github.com/fbenites/sklearn-hierarchical-classification/>
[^6]: Although a sample wise normalization can be applied, we used a normalization over all predicted samples.
[^7]: The system was trained in a Intel icore 7 CPU with 32 Gb RAM with a NVIDIA GeForce 1060 6Gb GPU.
[^8]: The training, dev and test set seems to come from the same distribution, so the quality prediction of using a semi-supervised method was worse than without.
[^9]: For the sake of conciseness, we will not discuss it here.
[^10]: Exemplary are books describing dystopias which from a n-gram perspective have very much the same vocabulary of a non-fiction book. Here, more aspects of the language need to be captured, such as a focus to constructions like “in a future New York City”.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
In papers[@js; @jsa], the amplitudes of continuous-time quantum walk on graphs possessing quantum decomposition (QD graphs) have been calculated by a new method based on spectral distribution associated to their adjacency matrix. Here in this paper, it is shown that the continuous-time quantum walk on any arbitrary graph can be investigated by spectral distribution method, simply by using Krylov subspace-Lanczos algorithm to generate orthonormal bases of Hilbert space of quantum walk isomorphic to orthogonal polynomials. Also new type of graphs possessing generalized quantum decomposition have been introduced, where this is achieved simply by relaxing some of the constrains imposed on QD graphs and it is shown that both in QD and GQD graphs, the unit vectors of strata are identical with the orthonormal basis produced by Lanczos algorithm. Moreover, it is shown that probability amplitude of observing walk at a given vertex is proportional to its coefficient in the corresponding unit vector of its stratum, and it can be written in terms of the amplitude of its stratum. Finally the capability of Lanczos-based algorithm for evaluation of walk on arbitrary graphs ( GQD or non-QD types), has been tested by calculating the probability amplitudes of quantum walk on some interesting finite (infinite) graph of GQD type and finite (infinite) path graph of non-GQD type, where the asymptotic behavior of the probability amplitudes at infinite limit of number of vertices, are in agreement with those of central limit theorem of Ref.[@nko].
[**Keywords: Continuous-time quantum walk, Spectral distribution, Graph, Krylov subspace, Lanczos algorithm.**]{}
[**PACs Index: 03.65.Ud** ]{}
author:
- |
M. A. Jafarizadeh$^{a,b,c}$ [^1] , S. Salimi$^{a,b}$ [^2] , R. Sufiani$^{a,b}$ [^3]\
$^a$[Department of Theoretical Physics and Astrophysics, Tabriz University, Tabriz 51664, Iran.]{}\
$^b$[Institute for Studies in Theoretical Physics and Mathematics, Tehran 19395-1795, Iran.]{}\
$^c$[Research Institute for Fundamental Sciences, Tabriz 51664, Iran.]{}
bibliography:
- 'plain.bib'
title: '**Investigation of continuous-time quantum walk by using Krylov subspace-Lanczos algorithm** '
---
-0.75in
\[thm\][Corollary]{} \[thm\][Lemma]{} \[thm\][Proposition]{} \[thm\][Definition]{} \[thm\][Remark]{}
Introduction
============
Random walks on graphs are the basis of a number of classical algorithms. Examples include 2-SAT (satisfiability for certain types of Boolean formulas), graph connectivity, and finding satisfying assignments for Boolean formulas. It is this success of random walks that motivated the study of their quantum analogs in order to explore whether they might extend the set of quantum algorithms.
Recently, the quantum analogue of classical random walks has been studied in a flurry of works [@fg; @cfg; @abnvw; @aakv; @mr; @k]. The works of Moore and Russell [@mr] and Kempe [@k] showed faster bounds on instantaneous mixing and hitting times for discrete and continuous quantum walks on a hypercube (compared to the classical walk).
A study of quantum walks on simple graph is well known in physics(see [@fls]). Recent studies of quantum walks on more general graphs were described in [@fg; @cfg; @aakv; @adz; @ccdfgs]. Some of these works studies the problem in the important context of algorithmic problems on graphs and suggests that quantum walks is a promising algorithmic technique for designing future quantum algorithms. One approach for investigation of continuous-time quantum walk on graphs is using the spectral distribution associated with the adjacency matrix of graphs. Authors in [@js; @jsa] have introduced a new method for calculating the probability amplitudes of quantum walk based on spectral distribution, where a canonical relation between the Hilbert space of stratification corresponding to the graph and a system of orthogonal polynomials has been established, which leads to the notion of quantum decomposition (QD) introduced in [@nob; @obah] for the adjacency matrix of graph. Also it is shown in [@js] that by using spectral distribution one can approximate long time behavior of continuous-time quantum walk on infinite graphs with finite ones and vice versa. In [@js; @jsa], only the particular graphs of QD type have been studied .
Here in this work we try to investigate continuous-time quantum walk on arbitrary graphs by spectral distribution method. To this aim, first by turning the graphs into a metric space based on distance function, we have been able to generalize the stratification and quantum decomposition introduced in [@nob], such that the basis of Hilbert space of quantum walk consist of superposition of quantum kets of vertices belonging to the same stratum, but with different coefficients, while the coefficients are the same in QD case, therefore QD graphs introduced in [@js; @nob] are particular kind of graphs possessing generalized quantum decomposition (GQD). Then we show that both in QD and GQD graphs, the unit vectors of strata are identical with the orthonormal basis produced by Lanczos algorithm. Also, in the case of GQD graphs we show that probability amplitude of observing walk at a given vertex is proportional to its coefficient in corresponding unit vector of its stratum, and it can be written in terms of the amplitude of its stratum. For more general graphs, the Lanczos algorithm transforms the adjacency matrix into a tridiagonal form (quantum decomposition) iteratively, where we use this fact for studying non-QD type graphs. Indeed, the Lanczos algorithm gives a three-term recursion structure to the graph, so the spectral distribution associated to adjacency matrix can be determined by Stieltjes transform. In order to see the power of Lanczos-based algorithm in the investigation of continuous-time quantum walk on arbitrary graphs ( GQD or non-QD types), we have calculated the amplitudes of quantum walk on some interesting finite (infinite) graph of GQD type and finite (infinite) path graph of non-GQD type.
The organization of the paper is as follows. In section 2, we review the Krylov subspace methods and Lanczos algorithm. In Section $3$, we give a brief outline of some of the main features of graphs and introduce generalized stratification. Section $4$ is concerned with the Hilbert space of generalized stratification. In Section $5$, we review the Stieltjes transform method for obtaining spectral distribution $\mu$, and establish an isometry between orthogonal polynomials and Hilbert space of generalized stratification. Section $6$ is devoted to the method for computing amplitudes of continuous-time quantum walk, through spectral distribution $\mu$ of the adjacency matrix $A$. In section $7$ we calculate the amplitudes of quantum walk on some interesting finite (infinite) graph of GQD type and finite (infinite) path graph of non-GQD type. At the end we study the asymptotic behavior the probability amplitudes at infinite limit of number of vertices, where the results thus obtained are in agreement with those of central limit theorem of Ref.[@nko]. Paper is ended with a brief conclusion together with two appendices.
Krylov subspace-Lanczos algorithm
=================================
In this section we give a brief review of some of the main features of Krylov subspace projection methods and Lanczos algorithm and more details are referred to [@bpa; @jwi; @ltdb; @jcrw].
Krylov subspace projection methods (KSPM) are probably the most important class of projection methods for linear systems and for eigenvalue problems. In KSPM, approximations to the desired eigenpairs of an $n\times n$ matrix $A$ are extracted from a $d$-dimensional Krylov subspace $$K_d({|\phi_0\rangle},A) = span\{{|\phi_0\rangle},A{|\phi_0\rangle}, \cdots,A^{d-1}{|\phi_0\rangle}\},$$ where ${|\phi_0\rangle}$ is often a randomly chosen starting vector called reference state and $d \ll n$. In practice, the retrieval of desired spectral information is accomplished by constructing an orthonormal basis $V_d \in R^{n\times d}$ of $K_d({|\phi_0\rangle},A)$ and computing eigenvalues and eigenvectors of the $d$ by $d$ projected matrix $H_d = {P_{V_d}}^TAP_{V_d}$, where $P_{V_d}$ is projection operator to $d$-dimensional subspace spanned by the basis $V_d$.
The most popular algorithm for finding an orthonormal basis for the Krylov subspace, is Lanczos algorithm. The Lanczos algorithm transforms a Hermitian matrix $A$ into a tridiagonal form iteratively, i.e., the matrix $A$ will be of tridiagonal form in the $d$-dimensional projected subspace $H_d$. In fact, the Lanczos algorithm is deeply rooted in the theory of orthogonal polynomials, which builds an orthonormal sequence of vectors $\{{|\phi_0\rangle},{|\phi_1\rangle},...,{|\phi_{d-1}\rangle}\}$ and satisfy the following three-term recursion relations $$\label{trt}
A{|\phi_i\rangle}=\beta_{i+1}{|\phi_{i+1}\rangle}+\alpha_i{|\phi_i\rangle}+\beta_i{|\phi_{i-1}\rangle}.$$
The vectors ${|\phi_i\rangle}, i=0,1,...,d-1$ form an orthonormal basis for the Krylov subspace $K_d({|\phi_0\rangle},A)$. In these basis, the matrix $A$ is projected to the following symmetric tridiagonal matrix: $$L_j=\left(
\begin{array}{ccccc}
\alpha_0 & \beta_1 & 0 & ... &... \\
\beta_1 & \alpha_1 & \beta_2 & 0 &... \\
0 & \beta_2 & \alpha_3 & \beta_3 & ... \\
... & ... &...& ... &... \\
...& ... &0 & \beta_{d-1} & \alpha_{d-1}\\
\end{array}
\right),$$ where the scalars $\beta_{i+1}$ and $\alpha_i$ are computed to satisfy two requirements, namely that ${|\phi_{i+1}\rangle}$ be orthogonal to ${|\phi_i\rangle}$ and that $\|{|\phi_{i+1}\rangle}\|= 1$.
In fact, the Lanczos algorithm is a modified version of the classical Gram-Schmidt orthogonalization process. As it can be seen, at its heart is an efficient three-term recursion relation which arises because the matrix $A$ is real and symmetric.\
If we define the Krylov matrix $K$ such that the columns of $K$ are Krylov basis $\{A^i\phi_0 ; i=0,...,d-1\}$ as: $$K:=({|\phi_0\rangle}, A{|\phi_0\rangle}, ... , A^{d-1}{|\phi_0\rangle}),$$ the application of the orthonormalization process to the Krylov matrix is equivalent to the construction of an upper triangular matrix $P$ such that the resulting sequence $\Phi=KP$ satisfies $\Phi^\dag \Phi=1$. We denote by ${|\phi_j\rangle}$ and $P_j$ respectively the $j$-th column of $\Phi$ and $P$. Then we have $${\langle \phi_0|P_i^{\dagger}(A)P_j(A)|\phi_0\rangle}={\langle KP_i|KP_j\rangle}={\langle \phi_i|\phi_j\rangle},$$ where $P_i=a_0+a_1A+...+a_iA^i$ is a polynomial of degree $i$ in indeterminate $A$.
In the remaining part of this section we give an algorithmic outline of the Lanczos algorithm, where it will be used in calculation of amplitudes of continuous-time quantum walk.
**Lanczos algorithm**\
Input: Matrix $A\in R^{n\times n}$, starting vector ${|\phi_0\rangle}$, $\|{|\phi_0\rangle}\|=1$, scalar $d$\
Output: Orthogonal basis $\{{|\phi_0\rangle},...,{|\phi_{d-1}\rangle}\}$ of Krylov subspace $K_d({|\phi_0\rangle},A)$ $$\beta_0=0, {|\phi_0\rangle}={|\phi\rangle}/\|{|\phi\rangle}\|$$ $$for\;\;\ i=0,1,2,...$$ $${|\upsilon_i\rangle}=A{|\phi_i\rangle}$$ $$\alpha_i={\langle \phi_i|\upsilon_i\rangle}$$ $${|\upsilon_{i+1}\rangle}={|\upsilon_i\rangle}-\beta_{i}{|\phi_{i-1}\rangle}-\alpha_i{|\phi_i\rangle}$$ $$\beta_{i+1}=\|{|\upsilon_{i+1}\rangle}\|$$ $$if$$ $$\beta_{i+1}\neq 0$$ $${|\phi_{i+1}\rangle}={|\upsilon_{i+1}\rangle}/\beta_{i+1}$$ $$else$$ $${|\phi_{i+1}\rangle}=0.$$
Graphs, adjacency matrix and generalized stratification
=======================================================
In this section we give a brief outline of some of the main features of graphs such as adjacency matrix, distance function and then by turning the graphs into a metric space based on distance function, we have been able to generalize the stratification introduced in [@nob]. A graph is a pair $\Gamma=(V,E)$, where $V$ is a non-empty set and $E$ is a subset of $\{(\alpha,\beta);\alpha,\beta\in V,\alpha\neq \beta\}$. Elements of $V$ and of $E$ are called *vertices* and *edges*, respectively. Two vertices $\alpha,\beta\in V$ are called *adjacent* if $(\alpha,\beta)\in E$, and in that case we write $\alpha\sim \beta$. For a graph $\Gamma=(V,E)$ we define the adjacency matrix $A$ by
$$A_{\alpha\beta} = \left\{
\begin{array}{ll}
1 & \mbox{if $ \alpha\sim \beta$}\\
0 & \mbox{otherwise.}
\end{array}
\right.$$ Obviously, (i) $A$ is symmetric; (ii) an element of $A$ takes a value in $\{0, 1\}$; (iii) a diagonal element of $A$ vanishes. Conversely, for a non-empty set $V$, a graph structure is uniquely determined by such a matrix indexed by $V$.
The *degree* or *valency* of a vertex $\alpha\in V$ is defined by $$\kappa(\alpha)=|\{\beta\in V; \alpha\sim \beta\}|,$$ where $\mid.\mid$ denotes the cardinality and $\kappa(\alpha)$ is finite for all $\alpha\in V$ (local boundedness). A finite sequence $\alpha_0, \alpha_1, . . . , \alpha_n \in V$ is called a walk of length $n$ (or of $n$ steps) if $\alpha_{k-1}\sim
\alpha_k$ for all $k=1, 2, . . . , n$. For $\alpha\neq \beta$ let $\partial(\alpha, \beta)$ be the length of the shortest walk connecting $\alpha$ and $\beta$. By definition $\partial(\alpha,
\alpha)=0$ for all $\alpha\in V$ and $\partial(\alpha, \beta)=1$ if and only if $\alpha\sim \beta$. Therefore, graphs become metric space with respect to above defined distance function $\partial$.
Now, in the remaining part of this section we try to define generalized stratification based on distance function. To this aim, similar to association scheme [@bailey] we define a partition (called distance partition) on $V\times V$, i.e., $V\times V=\bigcup_{i}\Gamma_i$ based on distance function $\partial$, where the subset $\Gamma_i$ are defined by $$\Gamma_i=\{(\alpha,\beta)\in V\times V
|\partial(\alpha,\beta)=i\}.$$ Using above distance partition one can define the set $\Gamma_i(\alpha)$ ($i$-th neighborhood of vertex $\alpha$) as $$\label{dr}
\Gamma_i(\alpha)=\{\beta\in V | (\alpha, \beta)\in \Gamma_i\}.$$ Obviously the class of subsets $\Gamma_i(\alpha)$ defined above partition $V$ as $$\label{v1}
V=\bigcup_{i}\Gamma_{i}(\alpha),$$ (see Fig.1). As we see the graph is stratified into a disjoint union of strata, hence we call it the generalized stratification based on distance function with respect to vertex $\alpha$, where the vertex $\alpha$ is referred to as a reference state(see Fig.2).
In this stratification for any connected graph $\Gamma$ , we have $$\Gamma_1(\beta)\subseteq \Gamma_{i-1}(\alpha)\cup
\Gamma_i(\alpha)\cup \Gamma_{i+1}(\alpha),$$ for each $\beta\in \Gamma_i(\alpha)$.
Obviously above relations are similar to those of distance regular graphs [@bailey], where in the later case the sets $\Gamma_i$ form an association scheme and the stratification $\Gamma_i(\alpha)$ is independent of reference state $\alpha$, but in an arbitrary graph the generalized stratification depends on the choice of reference state.
In order to study continuous-time quantum walk on a given graph via stratification, we define in the following section a Hilbert space which is suitable for Lanczos algorithm.
Hilbert space for the generalized stratification
================================================
Hereafter, we fix a point $o\in V$ as a reference state of the graph then with each stratum $\Gamma_k(o)$ we associate a unit vector ${|\phi_{k}\rangle}$ in $l^2(V)$ called unit vector of $k$-th stratum. The closed subspace of $l^2(V)$ spanned by $\{{|\phi_{k}\rangle}\}$ is denoted by $\Lambda(\Gamma)$. In section $6$, we will deal with continuous-time quantum walk, where $\Lambda(\Gamma)$ will be referred as walk space denoted by $V_{walk}$, i.e., the strata $\{{|\phi_{k}\rangle}\}$ span a closed subspace, where the quantum walk remains on it forever.
Since $\{{|\phi_{k}\rangle}\}$ become a complete orthonormal basis of $\Lambda(\Gamma)$, we often write $$\Lambda(\Gamma)=\sum_{k}\oplus \textbf{C}{|\phi_{k}\rangle}.$$ Then for each stratum $\Gamma_k(o)$ of generalized stratification, the unit vector associated to $k$-th stratum is defined by $$\label{unitv1}
{|\phi_{k}\rangle}=\frac{1}{\sqrt{\sum_{\alpha}g_{k,\alpha}^2}}\sum_{\alpha\in
\Gamma_{k}(o)}g_{k,\alpha}{|k,\alpha\rangle},$$ where, ${|k,\alpha\rangle}$ denotes the eigenket of $\alpha$-th vertex at the stratum $k$ and integers $g_{k,\alpha}\geq 1$ for each $\alpha\in \Gamma_{k}(o)$. We refer to a graph as GQD graph if the coefficients $g_{k,\alpha}$ satisfy conditions appearing in appendix $A$ (the conditions (\[A1\]) through (\[A3\])).
By choosing $g_{k,\alpha}=1$ for each $\alpha\in \Gamma_{k}(o)$, Eq.(\[unitv1\]) reduces to $$\label{unitv}
{|\phi_{k}\rangle}=\frac{1}{\sqrt{|\Gamma_{k}(o)}|}\sum_{\alpha\in
\Gamma_{k}(o)}{|k,\alpha\rangle},$$ where, ${|\phi_{k}\rangle}$, $k=0,1,2,...$ correspond to unit vectors of QD graphs of Ref.[@js].
In the following we show that, for the QD type graphs the unit vectors of strata given in Eq.(\[unitv\]), are the same as the orthonormal basis produced via Lanczos algorithm (this is true for GQD graphs too, where its proof is referred to appendix $A$.). To do so, let us consider the action of adjacency matrix $A$ over ${|\phi_{k}\rangle}$ as $$A{|\phi_{k}\rangle}=\frac{1}{\sqrt{|\Gamma_{k}(o)|}}\sum_{\alpha\in
\Gamma_{k}(o)}A{|k,\alpha\rangle}$$$$=\frac{1}{\sqrt{|\Gamma_{k}(o)|}}\sum_{\alpha\in
\Gamma_{k}(o)}\sum_{\nu\in \Gamma_{k+1}(o),\nu\sim
\alpha}{|k+1,\nu\rangle}+\frac{1}{\sqrt{|\Gamma_{k}(o)|}}\sum_{\alpha\in
\Gamma_{k}(o)}\sum_{\nu\in \Gamma_{k}(o),\nu\sim
\alpha}{|k,\nu\rangle}$$ $$+\frac{1}{\sqrt{|\Gamma_{k}(o)|}}\sum_{\alpha\in
\Gamma_{k}(o)}\sum_{\nu\in \Gamma_{k-1}(o),\nu\sim
\alpha}{|k-1,\nu\rangle}=\sqrt{\frac{|\Gamma_{k+1}(o)|}{|\Gamma_{k}(o)|}}
\frac{1}{\sqrt{|\Gamma_{k+1}(o)|}}\sum_{\nu\in
\Gamma_{k+1}(o)}\lambda_{k+1}(\nu){|k+1,\nu\rangle}+$$ $$\label{d1}
\frac{1}{\sqrt{|\Gamma_{k}(o)|}}\sum_{\nu\in
\Gamma_{k}(o)}\alpha_{k}(\nu){|k,\nu\rangle}+\sqrt{\frac{|\Gamma_{k-1}(o)|}{|\Gamma_{k}(o)|}}
\frac{1}{\sqrt{|\Gamma_{k-1}(o)|}}\sum_{\nu\in
\Gamma_{k-1}(o)}\frac{\lambda_{k}(\nu)|\Gamma_{k}(o)|}{|\Gamma_{k-1}(o)|}{|k-1,\nu\rangle}.$$
By defining $\beta_{k}=\frac{|\Gamma_{k}|^{1/2}}{|\Gamma_{k-1}|^{1/2}}\lambda_{k}(\nu)$, $\lambda_{k}(\nu)=|\{\alpha\in \Gamma_{k-1}(o);\alpha\sim\nu \}|$ and $\alpha_{k}=|\{\nu\in \Gamma_k;\nu\sim \alpha\}|$ for $\alpha,
\nu\in \Gamma_k(o)$, the three-term recursion relations (\[d1\]) reduce to those given in (\[trt\]).
Therefore, the adjacency matrix takes a tridiagonal form in the basis ${|\phi_k\rangle}$ (orthonormal basis associated with strata), consequently these basis are identical with the orthonormal basis produced by Lanczos algorithm.
Spectral distribution of the adjacency matrix $A$
=================================================
It is well known that, for any pair $(A,{|\phi_0\rangle})$ of a matrix $A$ and a vector ${|\phi_0\rangle}$, it can be assigned a measure $\mu$ as follows $$\label{sp1}
\mu(x)={\langle \phi_0|E(x)|\phi_0\rangle},$$ where $E(x)=\sum_i|u_i\rangle\langle u_i|$ is the operator of projection onto the eigenspace of $A$ corresponding to eigenvalue $x$, i.e., $$A=\int x E(x)dx.$$ It is easy to see that, for any polynomial $P(A)$ we have $$\label{sp2}
P(A)=\int P(x)E(x)dx,$$ where for discrete spectrum the above integrals are replaced by summation.
Actually the spectral analysis of operators is an important issue in quantum mechanics, operator theory and mathematical physics [@simon; @Hislop]. As an example $\mu(dx)=|\psi(x)|^2dx$ ($\mu(dp)=|\widetilde{\psi}(p)|^2dp$) is a spectral distribution which is assigned to the position (momentum) operator $\hat{X}(\hat{P})$. Moreover, in general quasi-distributions are the assigned spectral distributions of two hermitian non-commuting operators with a prescribed ordering. For example the Wigner distribution in phase space is the assigned spectral distribution for two non-commuting operators $\hat{X}$ (shift operator) and $\hat{P}$ (momentum operator) with Wyle-ordering among them [@Kim; @Hai].
Here in this paper we are concerned with spectral distribution of adjacency matrices of graphs, since the spectrum of a given graph can be determined by spectral distribution of its adjacency matrix $A$.
Therefore, using the relations (\[sp1\]) and (\[sp2\]), the expectation value of powers of adjacency matrix $A$ over starting site ${|\phi_0\rangle}$ can be written as $$\label{v2}
{\langle \phi_{0}|A^m|\phi_0\rangle}=\int_{R}x^m\mu(dx), \;\;\;\;\
m=0,1,2,....$$ The existence of a spectral distribution satisfying (\[v2\]) is a consequence of Hamburger’s theorem, see e.g., Shohat and Tamarkin \[[@st], Theorem 1.2\].
Obviously relation (\[v2\]) implies an isomorphism from the Hilbert space of generalized stratification onto the closed linear span of the orthogonal polynomials with respect to the measure $\mu$. Since, from the orthogonality of vectors ${|\phi_j\rangle}$ ( Hilbert space of generalized stratification) produced from Lanczos algorithm process we have, $$\delta_{ij}={\langle \phi_i|\phi_j\rangle}={\langle \phi_0|P_i^{\dagger}(A)P_j(A)|\phi_0\rangle}$$ $$=\int P_i^*(x)P_j(x)\mu(x)dx=(P_i,P_j)_{\mu}.$$ Conversely if $P_0,...,P_{n-1}$ is the system of orthonormal polynomials for the measure $\mu$ then the vectors $$\label{poly}
{|\phi_j\rangle}= P_j(A){|\phi_0\rangle},$$ will coincide with the sequence of orthonormal vectors produced by the Lanczos algorithm applied to $(A,{|\phi_0\rangle})$.
Now, substituting (\[poly\]) in (\[trt\]), we get three term recursion relations between polynomials $P_j(A)$, which leads to the following three term recursion between polynomials $P_j(x)$ $$\beta_{k+1}P_{k+1}(x)=(x-\alpha_k)P_{k}(x)-\beta_kP_{k-1}(x)$$ for $k=0,...,n-1$.\
Multiplying by $\beta_1...\beta_k$ we obtain $$\beta_1...\beta_{k+1}P_{k+1}(x)=(x-\alpha_k)\beta_1...\beta_kP_{k}(x)-\beta_k^2.\beta_1...\beta_{k-1}P_{k-1}(x).$$ By rescaling $P_k$ as $P'_k=\beta_1...\beta_kP_k$, the spectral distribution $\mu$ under question is characterized by the property of orthonormal polynomials $\{P'_n\}$ defined recurrently by $$P'_0(x)=1, \;\;\;\;\;\
P'_1(x)=x,$$ $$\label{op}
xP'_k(x)=P'_{k+1}(x)+\alpha_{k}P'_k(x)+\beta_k^2P'_{k-1}(x),$$ for $k\geq 1$.
If such a spectral distribution is unique, the spectral distribution $\mu$ is determined by the identity: $$\label{v3}
G_{\mu}(z)=\int_{R}\frac{\mu(dx)}{z-x}=\frac{1}{z-\alpha_0-\frac{\beta_1^2}{z-\alpha_1-\frac{\beta_2^2}
{z-\alpha_2-\frac{\beta_3^2}{z-\alpha_3-\cdots}}}}=\frac{Q_{n-1}^{(1)}(z)}{P'_{n}(z)}=\sum_{l=1}^{n}
\frac{A_l}{z-x_l},$$ where $G_{\mu}(z)$ is called the Stieltjes transform of spectral distribution $\mu$ and polynomials $\{Q_{k}^{(1)}\}$ are defined recurrently as $$Q_{0}^{(1)}(x)=1, \;\;\;\;\;\
Q_{1}^{(1)}(x)=x-\alpha_1$$, $$\label{oq}
xQ_{k}^{(1)}(x)=Q_{k+1}^{(1)}(x)+\alpha_{k+1}Q_{k}^{(1)}(x)+\beta_{k+1}^2Q_{k-1}^{(1)}(x),$$ for $k\geq 1$. The coefficients $A_l$ appearing in (\[v3\]) are the same Guass quadrature constants which are calculated as $$\label{Gauss}
A_l=\lim_{z\rightarrow x_l}(z-x_l)G_{\mu}(z),$$ where, $x_l$ are the roots of polynomial $P'_{n}(x)$.
Now let $G_{\mu}(z)$ is known, then the spectral distribution $\mu$ can be recovered from $G_{\mu}(z)$ by means of the Stieltjes inversion formula as $$\label{m1}
\mu(y)-\mu(x)=-\frac{1}{\pi}\lim_{v\longrightarrow
0^+}\int_{x}^{y}Im\{G_{\mu}(u+iv)\}du.$$ Substituting the right hand side of (\[v3\]) in (\[m1\]), the spectral distribution can be determined in terms of $x_l,
l=1,2,...$ and Guass quadrature constants $A_l, l=1,2,... $ as $$\label{m}
\mu=\sum_l A_l\delta(x-x_l)$$ ( for more details see Ref. [@obah; @st; @tsc; @obh]).
Finally, using the relation (\[poly\]) and the recursion relations (\[op\]) of polynomial $P'_k(x)$, the other matrix elements $\label{cw1} {\langle \phi_{k}|A^m\mid \phi_0\rangle}$ can be calculated as $$\label{cw1}
{\langle \phi_{k}|A^m\mid \phi_0\rangle}=\frac{1}{\beta_1\beta_2\cdots
\beta_k }\int_{R}x^{m}P'_{k}(x)\mu(dx), \;\;\;\;\ m=0,1,2,....$$
Investigation of Continuous-time quantum walk on an arbitrary graph via spectral distribution of its adjacency matrix
=====================================================================================================================
Our main goal in this paper is the evaluation of probability amplitudes for continuous-time quantum walk by using Eq.(\[cw1\]), such that we have $$\label{v4}
q_{k}(t)={\langle \phi_{k}|e^{-iAt}\mid
\phi_0\rangle}=\frac{1}{\beta_1\beta_2\cdots \beta_k
}\int_{R}e^{-ixt}P'_{k}(x)\mu(dx),$$ where $q_{k}(t)$ is the amplitude of observing the walk at stratum $k$ at time $t$. The conservation of probability $\sum_{k=0}{\mid
q_{k}(t)\mid}^2=1$ follows immediately from Eq.(\[v4\]), simply by using the completeness relation of orthogonal polynomials $P'_k(x)$.
Investigation of continuous-time quantum walk via spectral distribution method, pave the way to approximate infinite graphs with finite ones and vice versa, simply via Gauss quadrature formula, where in cases of infinite graphs, one can study asymptotic behavior of walk at large enough times by using the method of stationary phase approximation (for more details see [@js]).
One should note that, the spectral distribution is Fourier transform of the amplitude of observing the walk at starting site at time $t$, i.e., $$q_0(t)=\int e^{-ixt}\mu(x)dx \;\;\ \longmapsto \;\;\
\mu(x)=\frac{1}{2\pi}\int e^{ixt}q_0(t)dt.$$ Above relations imply that $$q_k(t)=\frac{1}{\beta_1\beta_2\cdots \beta_k }\int
P'_k(x)e^{-ixt}\mu(x)dx=\frac{1}{2\pi\beta_1\beta_2\cdots \beta_k
}\int P'_k(x)q_0(t')e^{-ix(t-t')}dt'dx,$$ therefore, the amplitudes $q_k(t)$ can be written in terms of the amplitude $q_0(t)$.
Obviously for finite graphs, the formula (\[v4\]) yields $$\label{fin}
q_{k}(t)=\frac{1}{\beta_1\beta_2\cdots \beta_k
}\sum_{l}A_le^{-ix_lt}P'_{k}(x_l),$$ where by straightforward calculation one can evaluate the average probability for the finite graphs as $$P(k)=\lim_{T\rightarrow \infty}\frac{1}{T}\int_{0}^{T}\mid
q_{k}(t)\mid ^2dt=\frac{1}{\beta_1\beta_2\cdots \beta_k
}\sum_{l}A_l^2{P'}^2_{k}(x_l).$$ In appendix I of Ref.[@js] it is proved that for QD graphs the amplitudes on the vertices belonging to the one stratum is the same, hence the probability of observing the walk at a site belonging to stratum $k$ is equal to $\frac{|q_{k}(t)|^2}{|\Gamma_k(o)|}$. Unfortunately for non-QD graphs the lemma appearing in appendix I of Ref.[@js] is not true any more, consequently the probability amplitudes of observing walk at sites can not be obtained from those of strata in a simple way and reader can follow the details of calculation of amplitudes in appendix $B$.
Examples
========
Generalized QD graphs
---------------------
Here in this subsection we give examples of GQD type graphs. These graphs look like kite and they are embedded in $Z^k,
k=2,3,...$ lattices and defined as follows: Let K(k,n) be an k-dimensional lattice graph with $n+1$ generalized strata, which consists of vertices $\underbrace{(0,0,...,0)}_k,(l,0,...,0),
(0,l,0,...,0),..., (0,0,...0,l)$ and $\underbrace{(l,l,...,l)}_k$ only for odd values of $l$, where $l=0,1,...,n$. The vertex $(0,0,...,0)$ is connected to vertices $(1,0,...,0),(0,1,...,0),...,(0,0,...,1)$ , the vertex $(0, ...,
0, \underbrace{l}_i, 0,..., 0)$ is connected to vertices $(0,
..., 0, \underbrace{l-1}_i, 0,..., 0)$ and $(0,
..., 0, \underbrace{l+1}_i, 0, ..., 0)$ for each $i=1,...,k$, but for odd values of $l$, there is an extra connection between $(0,...,0,
\underbrace{l}_i, 0,..., 0)$ and $(l, l, ..., l)$ (see Fig.3 for $k=2$, $n=6$).
Now, we define unit vectors of generalized strata in such a way that, they coincide with the orthonormal basis produced by lanczos algorithm (see appendix $A$) $${|\phi_0\rangle}={|\underbrace{0, 0, ..., 0}_k\rangle}$$ $${|\phi_1\rangle}=\frac{1}{\sqrt{k}}\sum_{perm.}{|1, 0, ..., 0\rangle}$$ $${|\phi_2\rangle}=\frac{1}{\sqrt{k(k+1)}}(\sum_{perm.}{|2,
0,...,0\rangle}+k{|1, 1, ..., 1\rangle})$$ $$\vdots$$ $${|\phi_{2l-1}\rangle}=\frac{1}{\sqrt{k}}\sum_{perm.}{|2l-1,
0,...,0\rangle}$$ $${|\phi_{2l}\rangle}=\frac{1}{\sqrt{k(k+1)}}(\sum_{perm.}{|2l,
0,..., 0\rangle}+k{|2l-1, 2l-1,..., 2l-1\rangle}),$$ where, the summations are taken over all possible permutations. Using the relations (\[A1\])-(\[A3\]), one can show that the coefficients $\beta_i$ and $\alpha_i$ are $$\label{ab}
\beta_1^2=k, \;\;\;\;\ \beta_2^2=\beta_3^2=....=k+1 \;\;\
\mbox{and} \;\;\ \alpha_i=0 , \;\ i=1,2,....$$ Now, one can study quantum continuous time walk on these graphs for finite values of $n$ simply by following the general prescriptions, but here we restrict ourselves to infinite $n$. Substituting coefficients $\beta_i$ and $\alpha_i$ in (\[v3\]) , the Stieltjes transform $G_{\mu}(z)$ of spectral distribution $\mu$ takes the following form $$G_{\mu}(z)=\frac{1}{z-\frac{k}{z-\frac{k+1}
{z-\frac{k+1}{z-\cdots}}}}.$$ In order to evaluate above infinite continued fraction, we need first to evaluate the following infinite continued fraction defined as $$\label{st.10}
\tilde{G}(z)=\frac{1}{z-\frac{k+1}{z-\frac{k+1}
{z-\frac{k+1}{z-\cdots}}}}= \frac{1}{z-(k+1)\tilde{G}(z)},$$ where by solving above equation, we get $$\label{st.1}
\tilde{G}(z)=\frac{z-\sqrt{z^2-4(k+1)}}{2(k+1)}.$$ Inserting (\[st.1\]) in (\[st.10\]), we get $$\label{st.3}
G_{\mu}(z)=\frac{1}{z-kG'_{\mu}(z)},$$ then substituting (\[st.1\]) in (\[st.3\]), we obtain the following expression for Stieltjes transform of $\mu$ $$\label{st.2}
G_{\mu}(z)=\frac{(k+2)z-k\sqrt{z^2-4(k+1)}}{2(k^2+z^2)},$$ finally by applying Stieltjes inversion formula, we get the absolutely continuous part of spectral distribution $\mu$ as follows $$\mu(x)=\frac{k}{2\pi}\frac{\sqrt{4(k+1)-x^2}}{k^2+x^2}, \;\;\;\;\
|x|\leq 2\sqrt{k+1} .$$ Now, we study the probability amplitudes of walk at time $t$ in the limit of large $k$ i.e., $$q_l(t)=\lim_{k\rightarrow\infty}{\langle \phi_l|e^{\frac{-iAt}{\sqrt{k}}}\mid\phi_0\rangle}=
\lim_{k\rightarrow\infty}\frac{1}{\sqrt{k(k+1)^l}}
\int_{-2\sqrt{k+1}}^{2\sqrt{k+1}}e^{\frac{-ixt}{\sqrt{k}}}P'_l(x)
\frac{k}{2\pi}\frac{\sqrt{4(k+1)-x^2}}{k^2+x^2}dx$$ $$\label{cent}
=\lim_{k\rightarrow\infty}\frac{1}{2\pi\sqrt{k(k+1)^l}}
\int_{-2\sqrt{(k+1)/k}}^{2\sqrt{(k+1)/k}}e^{-ixt}P'_l(\sqrt{k}x)\frac{\sqrt{4(k+1)/k-x^2}}{1+x^2/k}dx$$ $$=\frac{1}{2\pi}\int_{-2}^{2}e^{-ixt}P'_{l,\infty}(x)\sqrt{4-x^2}dx
=\frac{2}{\pi}\int_{-1}^{1}e^{-i2xt}P'_{l,\infty}(2x)\sqrt{1-x^2}dx,$$
where the polynomial $P'_{l,\infty}(x)$ is defined by $$\label{Eq3}
P'_{l,\infty}(x)=\lim_{k\rightarrow\infty}\frac{1}{\sqrt{k(k+1)^{l-1}}}P'_l(\sqrt{k}x).$$ Now, substituting $\beta_i$ and $\alpha_i$ from (\[ab\]) in three-term recursion relations (\[op\]), we obtain the following relations for polynomials $P'_l(x)$ $$P'_{0}(\sqrt{k}x)=1,$$ $$P'_{1}(\sqrt{k}x)=\sqrt{k}x,$$ $$P'_{2}(\sqrt{k}x)=kx^2-k,$$ $$\label{Eq4}
\sqrt{k}xP'_{l}(\sqrt{k}x)=P'_{l+1}(\sqrt{k}x)+(k+1)P'_{l-1}(\sqrt{k}x),
\;\;\ l=3,4,...\;\ .$$ Then dividing left and right hand sides of the recursion relations in (\[Eq4\]) by $\sqrt{k(k+1)^{l-1}}$ and taking the limit at $k\rightarrow\infty$, one can obtain the following recursion relations for $P'_{l,\infty}(x)$ $$P'_{0,\infty}(x)=1,$$ $$P'_{1,\infty}(x)=\lim_{k\rightarrow\infty}\frac{\sqrt{k}x}{\sqrt{k}}=x,$$ $$P'_{2,\infty}(x)=\lim_{k\rightarrow\infty}\frac{kx^2-k}{\sqrt{k(k+1)}}=x^2-1,$$ $$\label{Eq5}
xP'_{l,\infty}(x)=P'_{l+1,\infty}(x)+P'_{l-1,\infty}(x),
\;\;\l=3,4,...\;\ .$$ By comparing the recursion relations (\[Eq5\]) of $P'_{l,\infty}(x)$ with those of Tchebichef polynomials of second kind, we conclude that $$P'_{l,\infty}(x)=U_l(x/2),$$ where, $U_l(x)$’s are Tchebichef polynomials of second kind. Therefore the probability amplitudes in Eq.(\[cent\]) can be rewritten as $$\label{Eq1}
q_l(t)=\frac{2}{\pi}\int_{-1}^1
e^{-2ixt}U_l(x)\sqrt{1-x^2}dx=\frac{2}{\pi}\int_{-1}^1
e^{-2ixt}\sin((l+1)\cos^{-1}x)dx.$$ Now, by doing the change the variable $x=\cos\theta$, the integral(\[Eq1\]) can be written as $$\label{Eq2}
q_l(t)=\frac{2}{\pi}\int_{0}^{\pi}
e^{-2it\cos{\theta}}\sin((l+1)\theta)\sin\theta d\theta.$$ Then, using the following integral representation of Bessel polynomials $$J_l(x)=\frac{i^{-l}}{\pi}\int_0^{\pi}e^{-ix\cos\theta}\cos\theta
d\theta,$$ the integral in (\[Eq2\]) can be written as $$q_l(t)=i^l(J_l(2t)+J_{l+2}(2t)).$$ Now, from the recursion relations for Bessel polynomials, i.e., $$J_{l+1}(x)=\frac{2l}{x}J_l(x)-J_{l-1}(x),$$ we obtain the following expression for the probability amplitudes of walk in the limit of large $k$ $$q_l(t)=(l+1)i^l\frac{J_{l+1}(2t)}{t},$$ where the results are in agreement with the corresponding quantum central limit theorem of Ref.[@nko].
Non-GQD type graphs
-------------------
In this subsection we study an example of non-GQD type graphs, those graphs that do not possess three term recursion property. In order to obtain spectral distribution of adjacency matrix of a give non-GQD graph, we need to find the basis in which the adjacency matrix has tridiagonal form. To this aim we have to choose starting site of walk as a reference state and then apply Lanczos algorithm to its adjacency matrix.Then by using spectral distribution, we will be able to calculate the amplitudes of walk as will be explained in the following example.
### Walk on finite path graph with second vertex as the starting site of the walk
Finite path graph $\textsf{P}_n=\{1,2,... \}$ is a $n$- vertex graph with $n-1$ edges all on a single open path [@js]. For this graph, the stratification depends on the choice of starting site of walk. If we choose the second vertex as starting site of the walk, as it is shown in Fig.2, the graph does not satisfy a three term recursion relations, i.e., the adjacency matrix has not tridiagonal form.
Therefore, in order to find the basis in which the adjacency matrix has tridiagonal form, we have to apply Lanczos algorithm to the adjacency matrix $A$ of the graph $\textsf{P}_n$, where starting site $|\phi_0\rangle={|1\rangle}$ is chosen as a reference state. Also, the Lanczos algorithm provides the coefficients $\alpha$ and $\beta$ from which the Stieltjes transform $G_\mu(z)$ of $\mu$, Eq.(\[v3\]) can be calculated.
Hence, following the prescription of Lanczos algorithm given in section $2$, we get the following results for $\textsf{P}_n$, which are different for even and odd values of $n$.\
**A.** $n=2k$\
$\alpha_i=0,\;\;\;\;\ i=0,1,...,2k-1,$\
$\beta_{2i}=\sqrt{\frac{i}{i+1}}, \;\;\;\ \\
\beta_{2i-1}=\sqrt{\frac{i+1}{i}},\;\;\;\;\ i=1,...,k-1,$\
$\beta_{2k-1}=\frac{1}{\sqrt{k}}.$\
**B.** $n=2k+1$\
$\alpha_i=0,\;\;\;\;\ i=0,1,...,2k-1,$\
$\beta_{2i}=\sqrt{\frac{i}{i+1}},\;\;\;\ i=1,..., k-1 \\
\beta_{2i-1}=\sqrt{\frac{i+1}{i}},\;\;\;\;\ i=1,...,k,$ respectively.
Substituting the coefficients $\alpha_i$ and $\beta_i$ in (\[op\]) and (\[oq\]), and using (\[v3\]), we get the following closed form of the Stieltjes transform of $\mu$ $$G_\mu(z)=\frac{zU_{n-2}(z/2)}{U_n(z/2)}$$ where, $U_n$’s are Tchebichef polynomials of second kind. Therefore, the roots $x_l$ appearing in (\[v3\]) are roots of Tchebichef polynomials of second kind, i.e., $x_l=2\cos(\frac{l\pi}{n+1})$. Also, using (\[Gauss\]) we get the following expression for the coefficients $A_l$ $$A_l=\frac{2}{n+1}\sin^2(\frac{2l\pi}{n+1}).$$ Thus, spectral distribution is given by $$\mu=\frac{2}{n+1}\sum_{l=1}^{n}\sin^2(\frac{2l\pi}{n+1})\delta(x-2\cos(\frac{l\pi}{n+1})).$$ Then the probability amplitude of the walk at starting site at time $t$ is $$q_0(t)
=\frac{1}{n+1}\sum_{l=1}^{n}\sin^2(\frac{2l\pi}{n+1})e^{-2it\cos{l\pi/(n+1)}},$$ again one can calculate the other amplitudes by using Eq.(\[fin\]).
It should be noticed that, for odd $n$ the Lanczos algorithm produces $n-1$ orthonormal basis, therefore for calculating the amplitudes on vertices we need to construct an extra vector orthogonal to the walk space $V_{walk}$.
Finally, in the limit of large $n$, the continuous part of spectral distribution $\mu(x)$ is obtained as follows $$\mu(x)=\frac{2}{\pi}\int_{0}^{\pi}
dy\sin^2(2y)\delta(x-2\cos(y))$$ $$=\frac{2}{\pi}\int_{0}^{\pi}
dy\frac{\sin^2(2y)\delta(y-\arccos(x/2))}{2\sin(y)}$$ $$=\frac{4}{\pi}\int_{0}^{\pi}
dy\sin(y)\cos^2(y)\delta(y-\arccos(x/2))$$ $$=\frac{1}{2\pi}x^2\sqrt{4-x^2}, \;\;\;\;\;\;\ -2\leq x\leq 2,$$ therefore, the probability amplitude of the walk at starting site at time $t$ is $$q_0(t)=\frac{1}{2\pi}\int_{-2}^{2}e^{-ixt}x^2\sqrt{4-x^2}dx=\frac{4J_1(2t)}{t}-\frac{6J_2(2t)}{t^2},$$ where above result is obtained by making the change of variable $x=\cos\theta$, and using the integral representation of Bessel polynomials given in (7-47). Similarly, other amplitudes of walk can be calculated by using Eq.(\[v4\]).
Conclusion
==========
By turning the graphs into a metric space based on distance function, we have been able to generalize the stratification and quantum decomposition introduced in [@nob]. Then the continuous-time quantum walk on arbitrary graphs are investigated by spectral distribution method based on Krylov subspace-Lanczos algorithm. We have showed that both in QD and GQD graphs, the unit vectors of strata are identical with the orthonormal basis produced by Lanczos algorithm. For more general graphs, we have used the Lanczos algorithm to get a basis in which the adjacency matrix has tridiagonal form, where it is necessary for determination of spectral distribution of adjacency matrix by using inverse Stieltjes transform. We believe that the introduced algorithm is a powerful and general tool to investigate the continuous-time quantum walk on any arbitrary graph.
In this appendix we show that in the case of GQD graphs the unit vectors of strata (i.e., Eq.(\[unitv1\])), are the same as the orthonormal basis produced via Lanczos algorithm. To do so, let us consider the action of adjacency matrix $A$ over ${|\phi_{k}\rangle}$ as $$A{|\phi_{k}\rangle}=\frac{1}{\sqrt{\sum_{\alpha}g_{k,\alpha}^2}}\sum_{\alpha\in
\Gamma_{k}(o)}g_{k,\alpha}A{|k,\alpha\rangle}
=\frac{1}{\sqrt{\sum_{\alpha}g_{k,\alpha}^2}}\sum_{\alpha\in
\Gamma_{k}(o)}g_{k,\alpha}\sum_{\nu\in
\Gamma_{k+1}(o),{\alpha\sim\nu}}{|k+1,\nu\rangle}$$ $$\label{tree1}
+\frac{1}{\sqrt{\sum_{\alpha}g_{k,\alpha}^2}}\sum_{\alpha\in
\Gamma_{k}(o)}g_{k,\alpha}\sum_{\nu\in
\Gamma_{k}(o),{\alpha\sim\nu}}{|k,\nu\rangle}
+\frac{1}{\sqrt{\sum_{\alpha}g_{k,\alpha}^2}}\sum_{\alpha\in
\Gamma_{k}(o)}g_{k,\alpha}\sum_{\nu\in
\Gamma_{k-1}(o),{\alpha\sim\nu}}{|k-1,\nu\rangle},$$ now in order to have a GQD graph the coefficients $g_{k,\alpha}$ should satisfy the following conditions $$\label{A1}
\sum_{\alpha\in\Gamma_{k}(o)}g_{k,\alpha}=\gamma_{k+1}g_{k+1,\nu},$$ for all $\nu\in\Gamma_{k+1}(o)$ and $\alpha\sim\nu$, $$\label{A2}
\sum_{\alpha\in\Gamma_{k}(o)}g_{k,\alpha}=\eta_k g_{k,\nu},$$ for all $\nu\in\Gamma_{k}(o)$ and $\alpha\sim\nu$, $$\label{A3}
\sum_{\alpha\in\Gamma_{k}(o)}g_{k,\alpha}=
\gamma_{k}(\frac{\sum_{\alpha\in\Gamma_{k}(o)}g_{k,\alpha}^2}
{\sum_{\xi\in\Gamma_{k-1}(o)}g_{k-1,\xi}^2})g_{k-1,\nu},$$ for all $\nu\in\Gamma_{k-1}(o)$ and $\alpha\sim\nu$. One should note that the constants $\gamma_{k}$ and $\eta_k$ depend only on strata number.
Then by defining $
\beta_{k}=\gamma_{k}\sqrt{\frac{\sum_{\alpha\in\Gamma_{k}(o)}g_{k,\alpha}^2}
{\sum_{\xi\in\Gamma_{k-1}(o)}g_{k-1,\xi}^2}} $ and $\alpha_{k}=\eta_k$ for all $\alpha\in\Gamma_{k}(o),\xi\in\Gamma_{k-1}(o)$ and $\xi\sim\alpha$, the three-term recursion relations (\[tree1\]) reduce to those given in (\[trt\]).
Therefore similar to the QD case, the adjacency matrix takes a tridiagonal form in the basis ${|\phi_k\rangle}$ (orthonormal basis associated with strata of GQD graphs), consequently these basis are identical with the orthonormal basis produced by Lanczos algorithm.
Here in this appendix we first prove that in GQD graphs, the ratio of amplitude of a vertex in a given stratum to its coefficient appearing in (\[unitv1\]) is constant, i.e., $\frac{\phi_{k,\alpha}}{g_{k,\alpha}}$ is independent of $\alpha\in\Gamma_{k}(o)$. To do so, let us consider the eigenket ${|\phi_{k}\rangle}$ given in (\[unitv1\]), it is straightforward to see that, the eigenket ${|\phi_{k}\rangle}$ together with the following set of states $${|\phi_{k,l}^\perp\rangle}=\frac{1}{\sqrt{\sum_{\nu\in\Gamma_{k}(o)}}\frac{1}{|g_{k,\nu}|^2}}\sum_{\alpha\in\Gamma_{k}(o)}
\frac{\omega^{l\alpha}}{g_{k,\alpha}}{|k,\alpha\rangle}, \;\;\;\
l=1,2,...,|\Gamma_{k}(o)|-1,$$
form a set of orthonormal basis for a complex space formed by linear span of eigenkets belonging to stratum $k$ where $\omega=e^{-\frac{2\pi i}{|\Gamma_{k}(o)|}}$.
The above given states are actually orthogonal to all states of walk space ($V_w$), since the eigenket of other stratum do not contain any of ${|k,\alpha\rangle}, \alpha\in\Gamma_{k}(o)$. Therefore, $e^{-iAt}{|\phi_o\rangle}$ is orthogonal to set of orthogonal vectors ${|\phi_{k,l}^\perp\rangle}$, for all $l=1,2,...,|\Gamma_{k}(o)|-1; k=0,1,...,d$ since it is a state which remains in $V_w$ for all $t$. Now, substituting (\[unitv1\]) in (\[v4\]) and ${\langle \phi_{k,l}^\perp|e^{-iAt}|\phi_0\rangle}=0,
l=1,2,...,|\Gamma_{k}(o)|-1$, we get the following set of equations for amplitudes of vertices belonging to stratum $k$, $$\label{eq1}
q_{k}(t)=\frac{1}{\sqrt{\sum_{\nu\in\Gamma_{k}(o)}g_{k,\nu}^2}}
\sum_{\alpha\in\Gamma_{k}(o)}g_{k,\alpha}q_{k,\alpha}(t),$$ $$\label{eq2}
0=\frac{1}{\sqrt{\sum_{\nu\in\Gamma_{k}(o)}\frac{1}{g_{k,\nu}^2}}}\sum_{\alpha\in\Gamma_{k}(o)}
\frac{\omega^{-l\alpha}}{g_{k,\alpha}} q_{k,\alpha}(t), \;\;\;\;\
l=1,2,...,|\Gamma_{k}(o)|-1,$$ where $q_{k,\alpha}(t)$ denotes the amplitude of vertex $\alpha\in\Gamma_{k}(o)$. To solve equations (\[eq1\]) and (\[eq2\]), first we multiply equations (\[eq2\]) by $\omega^{l\nu}$ and sum over $l=1,2,...,|\Gamma_{k}(o)|-1$, where by using the identity $\sum_{l=0}^{|\Gamma_{k}(o)|-1}\omega^{l(\nu-\alpha)}=|\Gamma_{k}(o)|\delta_{\alpha\nu}$, we get for $\nu\neq\alpha$ $$\frac{q_{k,\alpha}(t)}{g_{k,\alpha}}=\frac{1}{|\Gamma_{k}(o)|}\sum_{\nu\neq\alpha}\frac{q_{k,\nu}(t)}{g_{k,\nu}},
\;\;\;\;\ \mbox{for all} \;\ \alpha\in\Gamma_{k}(o).$$
Above equations imply that $\frac{q_{k,\alpha}(t)}{g_{k,\alpha}}=\frac{q_{k,\xi}(t)}{g_{k,\xi}}=B_k$ for all $\alpha,\xi\in\Gamma_{k}(o)$ where, $B_k$ is some constant independent of vertices of stratum $k$, and it can be determined by substituting $q_{k,\alpha}(t)=B_k g_{k,\alpha}$ in (\[eq1\]) as $$B_k=\frac{1}{\sqrt{\sum_{\nu\in\Gamma_{k}(o)}g_{k,\nu}^2}}q_{k}(t).$$ Therefore, probability amplitude of observing walk at given vertex is proportional to its coefficient $g_{k,\alpha}$ and it can be written in terms of amplitude of the $k$-th stratum $q_{k}(t)$ as $$q_{k,\alpha}(t)=\frac{g_{k,\alpha}}{\sqrt{\sum_{\nu\in\Gamma_{k}(o)}g_{k,\nu}^2}}q_{k}(t).$$
In QD graphs we have $g_{k,\alpha}=1$, for all $\alpha\in\Gamma_{k}(o)$, hence vertices belonging to the same stratum, have the same amplitude which is in agreement with the result of appendix I of Ref.[@js].
In non-GQD type graphs, the coefficients of unit vectors ${|\phi_k\rangle}$ do not satisfy the conditions (\[A1\])-(\[A3\]) , and we can not obtain vectors orthogonal to $V_{walk}$ by the above explained prescription of GQD graphs. Therefore, one should use Lanczos algorithm for obtaining $n$ independent linear equations, where the amplitudes of vertices of the graph can be deteremined by solving them. Let the Krylov subspace generated by the adjacency matrix $A$ and starting site ${|\phi_0\rangle}$ has dimension $d$, then we will have $d$ unit vectors of strata produced from Lanczos algorithm applied to the pair $(A,{|\phi_0\rangle})$ (one should note that the walk space $V_{walk}$ is generated by applying the Lanczos algorithm to adjacency matrix and starting site of the walk ). In the most cases, the dimension of $V_{walk}$ is less than the number of vertices ($d<n$), i.e., the Lanczos algorithm applied to the pair $(A,{|\phi_0\rangle})$, dose not produce the enough basis, therefore for obtaining remaining equations we choose new reference states orthogonal to walk space $V_{walk}$ and then we apply the Lanczos algorithm to the adjacency matrix with new reference states, respectively. In the following, we explain the procedure in details for the following example.
**Example**
We consider tree graph of Fig.4, with six vertices and complete orthonormal basis $$\{{|1\rangle},{|2\rangle},{|3\rangle},{|4\rangle},{|5\rangle},{|6\rangle}\},$$ where vertex ${|1\rangle}$ is considered as starting site of the walk. We apply the Lanczos algorithm to adjacency matrix $A$ and starting site ${|\phi_0\rangle}={|1\rangle}$, where orthonormal basis and coefficients $\alpha_k, \beta_k$ produced from Lanczos algorithm are $${|\phi_0\rangle}={|1\rangle}, \;\;\;\;\;\
{|\phi_1\rangle}=\frac{1}{\sqrt{3}}({|2\rangle}+{|3\rangle}+{|4\rangle})$$ $${|\phi_2\rangle}=\frac{1}{\sqrt{2}}({|5\rangle}+{|6\rangle}), \;\;\;\;\;\;\
{|\phi_3\rangle}=\frac{1}{\sqrt{6}}(-2{|2\rangle}+{|3\rangle}+{|4\rangle}),$$ $$\beta_1=\sqrt{3}, \;\;\;\;\ \beta_2=\sqrt{2/3}, \;\;\;\;\
\beta_3=\sqrt{1/3}; \;\;\;\;\
\alpha_1=\alpha_2=\alpha_3=\alpha_4=0,$$ respectively. One can straightforwardly show that the corresponding Stieltjes transform of $\mu$ and spectral distribution are $$G_{\mu}(z)=\frac{z^3-(1+\sqrt{2})z/\sqrt{3}}{z^4-(4+\sqrt{2})z^2/\sqrt{3}+1},$$ $$\mu=0.2851952676(\delta(x-1.662563892)+\delta(x+1.662563892))$$ $$+0.2148047323(\delta(x-0.6014806445)+\delta(x+0.6014806445),$$ respectively, which yield the following probability amplitudes of walk at $k$-th stratum at time $t$, for $k=0,1,2,3$ $$q_0(t)=\int_{R}e^{-ixt}\mu(dx)= 0.2851952676\cos(1.662563892t)
+0.2148047323\cos(0.6014806445t).$$ $$q_1(t)=\frac{1}{\sqrt{3}}\int_{R}e^{-ixt}P'_1(x)\mu(dx)=\frac{1}{\sqrt{3}}\int_{R}xe^{-ixt}\mu(dx)$$ $$=\frac{2}{i\sqrt{3}}[0.2851952676\sin(1.662563892t)+
0.2148047323\sin(0.6014806445t)],$$ $$q_2(t)=\frac{1}{\sqrt{2}}\int_{R}e^{-ixt}P'_2(x)\mu(dx)=\frac{1}{\sqrt{2}}\int_{R}(x^2-\sqrt{3})e^{-ixt}\mu(dx)$$$$=\frac{1}{\sqrt{2}}[0.2943408772\cos(1.662563892t)
-.2943408762\cos(0.6014806445t)],$$ $$q_3(t)=\frac{\sqrt{3}}{\sqrt{2}}\int_{R}e^{-ixt}P'_3(x)\mu(dx)=
\frac{\sqrt{3}}{\sqrt{2}}\int_{R}(x^3-\frac{3+\sqrt{2}}{\sqrt{3}}x)e^{-ixt}\mu(dx)$$ $$\label{tree}
= \frac{6}{i\sqrt{6}}[0.102214289\sin(1.662563892t)-
0.2825324240\sin(0.6014806445t)].$$ Obviously, we need two extra equations for obtaining amplitudes on sites of the graph. According to the above explained prescription we can consider $${|\psi_0\rangle}=\frac{1}{\sqrt{2}}({|5\rangle}-{|6\rangle}),$$ as new reference state (${|\psi_0\rangle}\in V_{walk}^{\perp}$) and then by applying Lanczos algorithm to the pair ($A, {|\psi_0\rangle}$) we obtain $${|\psi_1\rangle}=\frac{1}{\sqrt{2}}({|3\rangle}-{|4\rangle}),$$ which leads to two following extra equations $${\langle \psi_0|e^{-itA}|\phi_0\rangle}=0,$$ $${\langle \psi_1|e^{-itA}|\phi_0\rangle}=0.$$ Now, by solving the above six equations, one can obtain amplitudes of continuous-time quantum walk on vertices of the graph as $${\langle 1|e^{-iAt}\mid\phi_0\rangle}=q_0(t),$$ $${\langle 2|e^{-iAt}\mid\phi_0\rangle}=\frac{1}{\sqrt{3}}(q_1(t)-\sqrt{2}q_3(t)),$$ $${\langle 3|e^{-iAt}\mid\phi_0\rangle}={\langle 4|e^{-iAt}\mid\phi_0\rangle}=
\frac{1}{\sqrt{3}}(q_1(t)+\frac{1}{\sqrt{2}}q_3(t)),$$ $${\langle 5|e^{-iAt}\mid\phi_0\rangle}={\langle 6|e^{-iAt}\mid\phi_0\rangle}=\frac{1}{\sqrt{2}}q_2(t),$$ where $q_0(t), q_1(t), q_2(t)$ and $q_3(t)$ have been given in Eq.(\[tree\]).
[99]{} M. A. Jafarizadeh, S. Salimi, e-print quan-ph/0510174. M. A. Jafarizadeh, S. Salimi, e-print quan-ph/0603139. Norio Konno, e-print quant-ph/0602213. E. Farhi and S. Gutmann, Phys. Rev. A **58**, 915 (1998). E. Farhi, M. Childs, and S. Gutmann, Quantum Information Process. **1**, 35 (2002). A. Ambainis, E. Bach, A. Nayak, A. Viswanath, and J. Watrous (2001), [*One-Dimensional Quantum Walks*]{}, in Proceedings of the 33rd ACM Annual Symposium on Theory Computing (ACM Press), p. 60. D. Aharonov, A. Ambainis, J. Kempe, and U. Vazirani, in Proceedings of the 33rd ACM Annual Symposium on Theory Computing (ACM Press, New York, 2001). C. Moore and A. Russell, in Proceedings of the 6th Int. Workshop on Randomization and Approximation in Computer Science (RANDOM’02)(2002). J. Kempe (2003), [*Discrete Quantum Random Walks Hit Exponentially Faster*]{}, Proceedings of 7th International Workshop on Randomization and Approximation Techniques in Computer Science (RANDOM’03), p. 354-69. R. Feynman, R. Leighton, and M. Sands(1965), [*The Feynman Lectures on Physics*]{}, Volume 3, Addison-Wesley. Y. Aharonov, L. Davidovich, and N.Zagury, Phy. Rev. A **48**, 1687 (1993). A. Childs, E. Deotto, R. Cleve, E. Farhi, S. Gutmann, D. Spielman , in Proc. $35$th Ann. Symp. Theory of Computing (ACM Press), 59 (2003). N. Obata, Interdisciplinary Information Sciences, **10**, 41 (2004). A. Hora, and N. Obata, Quantum Information V, World Scientific, Singapore (2002). B. Parlett (1980), [*The Symmetric Eigenvalue Problem*]{}, Prentice-Hall Inc., Englewood Cliffs, N.J. J. Wilkkinson (1965), [*The Algebraic Eigenvalue Problem*]{}, Clarendon Press, Oxford. L. Trefethen and D. Bau (1997), [*Numerical Linear Algebra*]{}, Society for Industrial and Applied Mathematics (SIMA), Philadelphia, PA. J. Cullum and R. Willoughby, [*L$\acute{a}$nczos Algorithems for Large Symmetric Eigenvalue Computations*]{}, Vol. I, Birkh$\ddot{a}$user Boston Inc, Boston, MA, 1985, Theory. R. A. Bailey, [*Association Schemes: Designed Experiments, Algebra and Combinatorics*]{} ( Cambridge University Press, Cambridge, 2004). H.Cycon, R.Forese, W.Kirsch and B.Simon [*Schrodinger operators*]{} (Springer-Verlag, 1987). P.D.Hislop and I.M.Sigal, [*Introduction to spectral theory: With applications to schrodinger operators*]{} (1995). Y.S.Kim, [*Phase space picture of quantum mechanics:group theoretical approach*]{}, (Science, 1991). H.W.Lee, Physics.Report, **259**, 147, 1995. J. A. Shohat, and J. D. Tamarkin, [*The Problem of Moments, American Mathematical Society*]{}, Providence, RI (1943). T. S. Chihara (1978), [*An Introduction to Orthogonal Polynomials*]{}, Gordon and Breach, Science Publishers Inc. A. Hora, and N. Obata, Fundamental Problems in Quantum Physics, World Scientific, **284**(2003).
[**Figure.1:**]{} Shows the stratification with respect to distance function.
[**Figure.2:**]{} Shows the finite path graph of $\textsf{P}_n$, where walk starts at vertex $2$.
[**Figure.3:**]{} Kite graph with $k=2$ and $n=6$, where walk starts at vertex (0,0). All vertices lying on a given vertical dashed line belong to the same statum
[**Figure.4:**]{} Shows the tree graph, where walk starts at vertex $1$.
[^1]: E-mail:[email protected]
[^2]: E-mail:[email protected]
[^3]: E-mail:[email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This paper investigates numerically a phenomenon which can be used to transport a single q-bit down a J1-J2 Heisenberg spin chain using a quantum adiabatic process. The motivation for investigating such processes comes from the idea that this method of transport could potentially be used as a means of sending data to various parts of a quantum computer made of artificial spins, and that this method could take advantage of the easily prepared ground state at the so called Majumdar-Ghosh point. We examine several annealing protocols for this process and find similar results for all of them. The annealing process works well up to a critical frustration threshold. There is also a brief section examining what other models this protocol could be used for, examining its use in the XXZ and XYZ models.'
author:
- 'Nick Chancellor$^{1}$ and Stephan Haas$^{1}$'
title: 'Using the J1-J2 Quantum Spin Chain as an Adiabatic Quantum Data Bus'
---
1 Department of Physics and Astronomy and Center for Quantum Information Science & Technology, University of Southern California, Los Angeles, California 90089-0484, USA
Introduction {#introduction .unnumbered}
============
The ability to send data from one part of a computer to another accurately and quickly is an essential feature in virtually any design. The use of artificial spin clusters in quantum computing has been of growing interest. There is an implementation which has been demonstrated using superconducting flux q-bits[@Johnson2011; @Harris2010-1; @Perdomo2008; @vanderPloeg2006; @Harris2010-2]. This paper demonstrates an effective and scalable way of sending arbitrary q-bit states along a spin chain with Heisenberg type coupling using quantum annealing. Assuming one could implement a Hamiltonian which follows this model, for example using the methods proposed in [@Chen2010] using coupled cavities, this system design could be used for a data bus which transports quantum states to different sections of a quantum computer system. For instance, the protocols discussed in this paper could potentially be used to move states from memory to a system of quantum gates in an implementation of the circuit model.
There has already been significant work done on the subject of quantum data buses using spin chains, [@Banchi2010; @Banchi2011; @Apollaro2012; @Banchi2011-2]. However these works differ significantly from the method proposed in this paper in that the encoded q-bit is not transmitted through a degenerate ground state manifold, but through excitations of the Hamiltonian.
This paper investigates a method of using q-bits as an intermediate bus for the transfer of quantum information. This method can be compared to another method which is that of pulses [@Fitzsimmons2006], where a Hamiltonian is applied to a system for a period of time to perform a given operation. In the case of information transfer this operation is usually a swap. Unlike the method of using pulses, this method of using q-bits does not require precise timing to insure that the correct operation is performed. The method of using a spin chain Hamiltonian as a data bus also means that one does not need to either be able to address any pair of q-bits in the system or perform multiple operations to transfer an arbitrary q-bit. The pulse method does have the advantage that every intermediate spin can be used as quantum memory. However this is at the cost of the increased complexity of using dynamic quantum evolution in excited states, and the requirement of precise timing.
The adiabatic quantum bus method also has the advantage that, as in any adiabatic quantum process, only the lowest energy parts of Hamiltonian need to be faithfully realised by the implementation method. For example, a Hamiltonian which actually has an infinite number of excited states on each “spin”, but where only the low energy states which act like a spin $\frac{1}{2}$ Heisenberg system, contribute to the ground state would be perfectly acceptable to use as an adiabatic quantum bus without modification. But the higher energy states may cause issues using a method such as pulses. This general feature of adiabatic quantum processes such as the one illustrated in this paper makes them more versatile than their non-adiabatic counterparts.
The effect we will examine exploits the SU(2) symmetry of the Heisenberg Hamiltonian and uses the ground state degeneracy created by this symmetry in a chain with an odd number of spins. It has already been demonstrated [@Chancellor2011] that disturbances can be sent an unlimited distance along such chains because of their degenerate ground state. This paper goes a step further and actually demonstrates how a specific state can be transported across the chain using a quantum annealing protocol. Further investigation will also be provided into application of this method to systems such as the XYZ spin chain which only have a $\mathbb{Z}_{2}$symmetry.
Setup\[sub:Setup\] {#setupsubsetup .unnumbered}
------------------
The model we consider is the J1-J2 Heisenberg spin chain with open boundaries,
$$H=\sum_{n=1}^{N-1}J_{1}^{n}\vec{\sigma}_{n}\cdot\vec{\sigma}_{n+1}+\sum_{n=1}^{N-2}J_{2}^{n}\vec{\sigma}_{n}\cdot\vec{\sigma}_{n+2}.\label{eq:J1-J2}$$
This model has SU(2) symmetry, which is expressed by the Hamiltonian being block diagonal, such that there are N+1 blocks each with $\binom{N}{k}$ states. Each block represents all of the states with a given number, k, of up spins. If the number of spins in the model is odd, then the additional symmetry under a flip of the spins in the z direction, i.e. $\sigma^{z}\rightarrow-\sigma^{z}$ implies that all states of the Hamiltonian have a twofold energy degeneracy. In the anti-ferromagnetic case, ( $J_{1},\: J_{2}>0$ ) the ground state manifold consists of one state from the $\textrm{k=floor(}\frac{N}{2}\textrm{)}$ and one from the $\textrm{k=ceil(}\frac{N}{2}\textrm{)}$ sector. A simple example of this would be taking a system with 5 spins, the ground state would be twofold degenerate and would span the k=2 and k=3 sectors. One can now consider an initial Hamiltonian of the form of Eq. \[eq:J1-J2\] where the couplings are the ones given by
$$\begin{aligned}
J_{1}^{n} & = & \begin{cases}
J_{1}^{n,init} & n<N-1\\
0 & n=N-1
\end{cases},\label{eq:J1 init}\\
J_{2}^{n} & = & \begin{cases}
J_{2}^{n,init} & n<N-2\\
0 & n=N-2
\end{cases}.\label{eq:J2 init}\end{aligned}$$
The general condition on $J_{1}^{n,init}$ and $J_{2}^{n,init}$ is that the coupling is predominantly anti-ferromagnetic everywhere and that each spin is coupled to the others by at least one non zero J. For simplicity this paper considers only $J_{1}^{n,init}=1$ and $J_{2}^{n,init}=J_{2}^{init}$. This ground state manifold consists of the tensor product of the (unique) ground state of the chain of length N-1 with the Nth spin in an arbitrary state, a state in this manifold is of the from given by
$${|\Psi^{init} \rangle}={|\Psi_{0}^{N-1} \rangle}\times{|\psi^{init} \rangle},\label{eq:gs. init}$$
where ${|\Psi_{0}^{N-1} \rangle}$ is the ground state of the spin chain of length N-1 and ${|\psi^{init} \rangle}$is an arbitrary single spin state. One can now consider the same Hamiltonian, but with $n\rightarrow(N-n)+1$ . This Hamiltonian also has the form of Eq. \[eq:J1-J2\], but with couplings
$$\begin{aligned}
J_{1}^{n} & = & \begin{cases}
J_{1}^{n,final} & n>1\\
0 & n=1
\end{cases},\label{eq:J1 fin}\\
J_{2}^{n} & = & \begin{cases}
J_{2}^{n,final} & n>2\\
0 & n=2
\end{cases}.\label{eq:J2 fin}\end{aligned}$$
The general condition on $J_{1}^{n,final}$ and $J_{2}^{n,final}$ is that the coupling is predominantly anti-ferromagnetic everywhere and that each spin is coupled to the others by at least one non-zero J. For simplicity this paper considers only $J_{1}^{n,final}=1$ and $J_{2}^{n,final}=J_{2}^{final}$. A state in the ground state manifold is now given by
$${|\Psi^{final} \rangle}={|\psi^{final} \rangle}\times{|\Psi_{0}^{N-1} \rangle},\label{eq:gs. fin}$$
where ${|\psi^{final} \rangle}$ is an arbitrary single spin state. One can now consider a quantum annealing process with described by
$$\textrm{\textrm{H}(t;}\tau\textrm{)}=\textrm{\textrm{\textrm{A}(t;}\ensuremath{\tau}\textrm{)} H}^{\textrm{init}}+\textrm{\textrm{\textrm{B}(t;}\ensuremath{\tau}\textrm{)} H}^{\textrm{final}},\label{eq:H anneal}$$
where $\textrm{H}^{\textrm{init}}$ is \[eq:J1-J2\] with the conditions given in \[eq:J1 init\] and \[eq:J2 init\] and $\textrm{H}^{\textrm{final}}$ is \[eq:J1-J2\] with the conditions given in \[eq:J1 fin\] and \[eq:J2 fin\]. Also A and B follow the conditions $$\begin{aligned}
\textrm{\textrm{A}(}t\leq0\textrm{;}\tau\textrm{)} & = & 1,\label{eq:anneal conds. begin}\\
\textrm{\textrm{B}(}t\leq0\textrm{;}\tau\textrm{)} & = & 0,\\
\textrm{\textrm{A}(}t\geq\tau\textrm{;}\tau\textrm{)} & = & 0,\\
\textrm{\textrm{B}(}t\geq\tau\textrm{;}\tau\textrm{)} & = & 1.\label{eq:anneal conds. end}\end{aligned}$$
For all values of A and B the SU(2) symmetry is preserved. Therefore the Hamiltonian remains block diagonal at all times. The symmetry of the Hamiltonian under $\sigma^{z}\rightarrow-\sigma^{z}$ is also preserved at all times. This implies that the ground-state degeneracy (as well as the twofold degeneracy of all states) is preserved. The block diagonal structure implies that there will be no exchange of amplitude between spin sectors during the annealing process, while the degeneracy implies that no relative phase can be acquired between the states in the $\textrm{k=floor(}\frac{N}{2}\textrm{)}$ and the $\textrm{k=ceil(}\frac{N}{2}\textrm{)}$ sector. From the combination of these two conditions one can see that as long as one anneals slowly enough with $\textrm{\textrm{H}(t;}\tau\textrm{)}$ [^1] one can start with a state of the form given in Eq. \[eq:gs. init\] and reach a final state in the form Eq. \[eq:gs. fin\] where ${|\psi^{fin} \rangle}=\exp(\imath\varphi){|\psi^{init} \rangle}$, and $\varphi$ is an irrelevant phase. One specific example of such an annealing protocol to transport a spin is given in Fig. \[fig:transport cartoon\].
![Cartoon representation of a process where a spin is joined to the chain, then the spin on the opposite end is removed. Note that this is only one specific example of many possible processes for transporting a q-bit.\[fig:transport cartoon\]](figure1)
Advantages {#advantages .unnumbered}
==========
The use of the J1-J2 Heisenberg chain for transport by quantum annealing has several advantages. First the model with uniform coupling is gapped for $\frac{J_{2}}{J_{1}}\gtrsim0.25$ [@Chitra1995]. This suggests that within the adiabatic evolution process, at least locally, the system should behave as a gapped system in this regime, as long as global effects such as odd length frustration do not cause problems.
It is important to note that even the largest system size considered here is far from the thermodynamic limit. One should note, however, that given the connectivity schemes of adiabatic quantum chips already in existence [@Harris2010-2], one may only need to transport a q-bit state a few spins to get it to any part of the system.
Further evidence of favorable scaling comes from [@Chancellor2011] which demonstrates that disturbances can travel an unlimited distance in the presence of a degenerate ground state, even in a gapped system. Furthermore, [@Chancellor2011] suggests that these disturbances can carry entanglement, polarization, and quantum information. The transport by annealing given here is a specific example of how this effect can be taken advantage of.
Another advantage of the use of the J1-J2 Heisenberg Hamiltonian is the existence of the so called Majumdar-Ghosh point [@Majumdar1970] ($\frac{J_{2}}{J_{1}}=0.5$). At this point the ground state (with an even number of spins) has the simple form of a matrix product of singlets. Due to this fact the system should be relatively easy to prepare. The system is also gapped at the Majumdar-Ghosh point, making the Majumdar-Ghosh Heisenberg Hamiltonian, an ideal system for transport by quantum annealing and the ideal candidate for building an adiabatic quantum data bus.
Although this paper focuses on the J1-J2 Heisenberg model, it should be noted that this same annealing scheme should work with any pattern of coupling in the intermediate spins (i.e. J1-J2-J3)[^2]. One would also expect this scheme to work in models where the SU(2) symmetry is broken but there is a remaining $\mathbb{Z}_{2}$symmetry such as the XYZ or XY model, again with arbitrary patterns of coupling. Note however that this mehhod will not work in the Ising model, because although there is a $\mathbb{Z}_{2}$ symmetry, the Hamiltonian lacks terms to exchange q-bits between sites because it is diagonal in the computational basis.
Proof of Principle
==================
None of the arguments so far have given much illumination to the difficulty or ease of annealing within the sector. While we have discussed that transport of a q-bit state is possible in principle by annealing, we have not yet shown that the annealing process is fast enough to be practical. For this we turn to numerics. For the purposes of this paper we will consider the annealing time, $\tau$ , required to reach a given fixed fidelity, $F(\tau)$, with the true final ground state,
$$F(\tau)=|{\left \langle \Psi^{fin} \mid \intop_{0}^{\tau}dt\, H(t,\tau) \mid \Psi^{init} \right\rangle}|.\label{eq:anneal time}$$
The J1-J2 Heisenberg model is not an analytically solved model, at least for finite values of $J_{2}$, so numerical methods must be used in this calculation. One can first consider one part of the annealing process, in which a single spin is joined to a even length J1-J2 spin chain, using both $J_{1}$ and $J_{2}$ couplings which are linearly increased to equal values of those used in the rest of the chain [^3],
$$H(t,\tau)=\sum_{n=1}^{N-2}J_{1}\vec{\sigma}_{n}\cdot\vec{\sigma}_{n+1}+\sum_{n=1}^{N-3}J_{2}\vec{\sigma}_{n}\cdot\vec{\sigma}_{n+2}+\lambda(t,\tau)(J_{1}\vec{\sigma}_{N-1}\cdot\vec{\sigma}_{N}+J_{2}\vec{\sigma}_{N-2}\cdot\vec{\sigma}_{N}),\label{eq:join}$$
$$\lambda(t,\tau)=\begin{cases}
0 & t\leq0\\
\frac{t}{\tau} & 0<t<\tau\\
1 & t\geq\tau
\end{cases}.$$
![Coupling constant $\lambda(t,\tau)$ from Eq.\[eq:join\] and Eq. \[eq:teleport H\] versus $\frac{t}{\tau}$ .](figure2)
![Annealing time required to reach a 90% fidelity with the true ground state within one of the two largest spin sectors of the Hamiltonian vs. $J_{2}$, with $J_{1}$ set to unity. One can see that for larger values of $J_{2}$ the annealing time behaves unpredictably. The annealing time also scales poorly with system size close to the Majumdar-Ghosh point.\[fig:T\_join\]](figure3)
As shown in Fig. \[fig:T\_join\], the annealing time required becomes large and highly sensitive to small variations for larger values of $J_{2}$. Also the behavior seems to get worse in this regime as system size is increased, and is poor at the Majumdar-Ghosh point [^4].
[00.00.0000]{}
As a further demonstration of the scaling with annealing time versus $J_{2}$, one can plot the annealing time versus system size, as we have done in Fig. \[fig:time\_scaling\]. This figure shows polynomial or even sub polynomial scaling for small values of $J_{2}$, but than shows strongly non-monotonic behavior for stronger coupling. It is important to note however that even the longest chain length considered here is probably far from the infinite system limit, and this data may not be trustworthy for making predictions for scaling as the chain length approaches the infinite system limit.
![\[fig:join\_gap\] Plots of gap for joining a single spin to an even length J1-J2 Heisenberg spin chain. For density plots lighter colors indicate larger gap. a) gap versus $\lambda$ in Eq. \[eq:join\] and $J_{2}$ for 15 total spins d) Gap versus $J_{2}$ with $\lambda=1$](figure5)
By examining the gap one can hope to gain insight into the underlying cause of the behaviour of annealing time curves. As Figs. \[fig:join\_gap\](a) and (b) show, the behavior of the annealing time curves is reflected by the presence of what appear to be true crossings [^5] for the odd length spin chain with uniform coupling. Fig. \[fig:join\_gap\](b) shows the gap for an odd length spin chain and seems to confirm the presence of points with very small gap with uniform coupling for $J_{2}$ above 0.5. Figs. \[fig:T\_join\] and \[fig:join\_gap\] together show that, at least at the length scales considered here, there are good annealing paths for joining a single spin to an even length chain. However, the simplest method of taking advantage of the simple ground-state wavefunction at the Majumdar-Ghosh point is not optimal. Fortunately there are many other possible options to take advantage of the easily prepared ground state and hopefully avoid the regions of small gap found here.
Dynamically Tuning J2
=====================
One method to avoid regions of small gap while still taking advantage of the Majumda-Ghosh point would be to start at the Majumdar-Ghosh point and then dynamically reduce the value of $J_{2}$ during the annealing process, a simple way of doing this would be to use the Hamiltonian in Eq. \[eq:dynamic J2\].
$$H(t,\tau)=\sum_{n=1}^{N-2}J_{1}\vec{\sigma}_{n}\cdot\vec{\sigma}_{n+1}+\sum_{n=1}^{N-3}J_{2}(t,\tau)\vec{\sigma}_{n}\cdot\vec{\sigma}_{n+2}+\lambda(t,\tau)(J_{1}\vec{\sigma}_{N-1}\cdot\vec{\sigma}_{N}+J_{2}(t,\tau)\vec{\sigma}_{N-2}\cdot\vec{\sigma}_{N}),\label{eq:dynamic J2}$$
$$\lambda(t,\tau)=\begin{cases}
0 & t\leq0\\
\frac{t}{\tau} & 0<t<\tau\\
1 & t\geq\tau
\end{cases},$$
$$J_{2}(t,\tau)=\begin{cases}
0.5 & t\leq0\\
0.5+\frac{t}{\tau}(J_{2f}-0.5) & 0<t<\tau\\
J_{2f} & t\geq\tau
\end{cases}.$$
![\[fig:MDG\_time\] Annealing time required to reach a 90% fidelity with the true ground state within one of the two largest spin sectors of the Hamiltonian with dynamical coupling starting at $J_{2}$=0.5 and linearly changing to $J_{2f}$ while also joining a spin to the chain, with $J_{1}$ set to unity throughout the process. Notice that this figure is qualitatively and quantitatively very similar to Fig. \[fig:T\_join\]. ](figure7)
Fig. \[fig:MDG\_time\] shows that taking advantage of the easily prepared ground state at the Majumdar-Ghosh point does in fact work, and the curves in this figure are strikingly similar to those in Fig. \[fig:T\_join\]. This similarity is to be expected because Fig. \[fig:join\_gap\] demonstrates that the gap is the smallest where the spin is completely joined. Hence this part of the process should dominate the annealing time.
It is reasonable to argue that because the regions of phase space which are visited are the same in the uncoupling process as coupling, the behavior of the system during the uncoupling process is determined by the gaps shown in Fig. \[fig:join\_gap\], and therefore the annealing times for the uncoupling process should be at least qualitatively similar to those given in Fig. \[fig:T\_join\]. One advantage to the uncoupling process is that unlike the coupling process, the need is not as strong to end in an easily prepared state. The only reason one may have to want to end in the Majumdar-Ghosh point is as an error check. The spins in the chain can be measured after the end of the process to ensure that no error has occurred [^6].
![\[fig:uncoupling\_time\]Annealing time required to reach a 90% Fidelity with the true ground state for uncoupling process within one of the two largest spin sectors of the Hamiltonian vs. $J_{2}$ with $J_{1}$ set to unity. One can see that this figure is very similar to Fig. \[fig:T\_join\] as one would expect because it is simply the time reversed version of that process.](figure8)
Fig. \[fig:uncoupling\_time\] shows the time required to uncouple a spin from the chain, not surprisingly this figure looks very similar to Fig. \[fig:T\_join\] which is the coupling process. Note that in this system the Hamiltonian is simply Eq. \[eq:join\] with $\frac{t}{\tau}\rightarrow(1-\frac{t}{\tau})$ .
![\[fig:uncoupling\_MDG\]Annealing time required to reach a 90% fidelity with the true ground state for uncoupling process within one of the two largest spin sectors of the Hamiltonian vs. initial $J_{2i}$ with a final $J_{2}$ at the Majumdar-Ghosh point with $J_{1}$ set to unity. This figure is very similar to Fig. \[fig:MDG\_time\] as one would expect, because it is simply the time reversed version of that process. ](figure9)
As expected, except for one curve where a numerical error made some points unable to plot one can see from Fig. \[fig:uncoupling\_MDG\] that the uncoupling process also requires roughly the same time as the coupling process for dynamically tuned $J_{2}$. Note that the Hamiltonian for this process is simply Eq. \[eq:dynamic J2\] with $\frac{t}{\tau}\rightarrow(1-\frac{t}{\tau})$ and $J_{2f}\rightarrow J_{2i}$.
Simultaneous Uncoupling and Coupling
====================================
Because many of the issues encountered with the coupling protocol seem to relate to odd-spin frustration, it may be reasonable to consider simultaneously coupling one q-bit to the chain while uncoupling the other. The Hamiltonian in this case is given in Eq. \[eq:teleport H\].
$$H(t,\tau)=\sum_{n=1}^{N-2}J_{1}\vec{\sigma}_{n}\cdot\vec{\sigma}_{n+1}+\label{eq:teleport H}$$
$$\sum_{n=1}^{N-3}J_{2}(t,\tau)\vec{\sigma}_{n}\cdot\vec{\sigma}_{n+2}+\lambda(t,\tau)((J_{1}\vec{\sigma}_{N-1}\cdot\vec{\sigma}_{N}+J_{2}\vec{\sigma}_{N-2}\cdot\vec{\sigma}_{N})-(J_{1}\vec{\sigma}_{1}\cdot\vec{\sigma}_{2}+J_{2}\vec{\sigma}_{1}\cdot\vec{\sigma}_{3})),$$
$$\lambda(t,\tau)=\begin{cases}
0 & t\leq0\\
\frac{t}{\tau} & 0<t<\tau\\
1 & t\geq\tau
\end{cases}.$$
![\[fig:teleport\_gaps\]Plots of gap for simultaneously joining a single spin to an even length J1-J2 Heisenberg spin chain and unjoining a spin from the other end. For density plots lighter colors indicate larger gap. a) gap versus $\lambda$ from Eq. \[eq:teleport H\] and J2 for 17 total spins b) Gap versus $J_{2}$ with $\lambda=0.5$.](figure10)
Fig. \[fig:teleport\_gaps\] shows the gaps for various system sizes for the process where the couplings are turned on and off simultaneously. This process does not seem to avoid the area of low gap for $J_{2}\gtrsim0.5$ seen in Fig. \[fig:join\_gap\]. However by comparing Fig. \[fig:teleport\_gaps\] d) and Fig. \[fig:join\_gap\] d) one can see that it appears that the process of simultaneous uncoupling and coupling is characterized by avoided crossings rather than true crossings [^7].
![\[fig:teleport time\]Annealing time required to reach a 90% Fidelity with the true ground state for combined coupling and uncoupling process within one of the two largest spin sectors of the Hamiltonian vs. $J_{2}$ with $J_{1}$ set to unity. ](figure11)
Fig. \[fig:teleport time\] shows the time required for annealing processes with for the combined coupling and uncoupling process, the results are consistent with what one would expect from looking at Fig. \[fig:teleport\_gaps\], and confirm that the annealing time also tends to be very long and vary a lot for larger values of $J_{2}$.
Requirements for use as an Adiabatic Quantum Bus
================================================
It is now useful to consider a broader class of models that may be used as adiabatic quantum buses, as in general the full SU(2) symmetry of the Heisenberg Hamiltonian is not required.
The requirements for a spin chain (or network) Hamiltonain to be usable as an adiabatic quantum bus are as follow:
1. The ground state must be at least 2 fold degenerate, and the ground state manifold must be able to encode a q-bit. In this paper this is achieved by having at least a $\mathbb{Z}_{2}$symmetry, and an odd number of spins, but there may be other ways.
2. The Hamiltonian (or at least the low energy states) must be predominantly anti-ferromagnetic in nature. This guarantees that the encoded q-bit will be excluded from the larger spin chain (or network) when a single q-bit is removed.
3. The Hamiltonian must contain terms which perform exchanges between sites. This excludes models such as the Ising model which, although it has the required symmetry, cannot be used a quantum bus because its Hamiltonian is diagonal in the computational basis
4. One must be able to slowly couple in a spin with an arbitrary state on one end of the chain (network) and also to slowly remove coupling on the other end. More control may improve performance, but is not necessary.
5. Annealing paths in parameter space must not contain true crossings. This is a general requirement for adiabatic quantum computing.
XXZ and XYZ model
=================
As previously mentioned, the full SU(2) symmetry of the Heisenberg Hamiltonian is not required. The Hamiltonian must only have a $\mathbb{Z}_{2}$symmetry to encode and transport one q-bit of information. In this section we will briefly examine two other possibilities: the XXZ model, where the SU(2) symmetry is broken, but the block diagonal structure imparted by this symmetry remains, and the XYZ model where only the block diagonal structure of a $\mathbb{Z}_{2}$ symmetry is present.
![\[fig:XXZ\] Annealing time to reach 90% fidelity on using the adiabatic quantum bus protocol on an XXZ spin chain versus the ratio of X and Z coupling strengths note that Z/X=0 is an XX model while Z/X=1 is a J1 Heisenberg spin chain. This data was obtained with joining and disconnecting of spins occurring simultaneously.](figure12)
As one can see from Fig. \[fig:XXZ\], the XXZ model can be used as an adiabatic quantum data bus. There is a regime where this system outperforms the XXX Heisenberg model for Z/X between 1 and roughly 2. This is to be expected because adding additional coupling in the z direction may serve to open the gap between the the ground-state manifold and the next excited state. The increasing time as the z coupling is increased further can be explained because the system would behave like an Ising model in the limit of $\frac{Z}{X}\gg1$ .
One can further examine the behavior of an XYZ model as an adiabatic quantum spin bus. For this purpose we consider the quantum bus protocol performed on the following normalized XYZ Hamiltonian
$$H_{XYZ}(\Delta;N)=C_{\Delta}\sum_{i=1}^{N-1}\sigma_{i}^{x}\sigma_{i+1}^{x}+(1+\Delta)\sigma_{i}^{y}\sigma_{i+1}^{y}+(1+2\Delta)\sigma_{i}^{z}\sigma_{i+1}^{z},\label{eq:XYZ_h}$$
where the normalization is
$$C_{\Delta}=\frac{\sqrt{3}}{\sqrt{1+(1+\Delta)^{2}+(1+2\Delta)^{2}}}.$$
One can now examine the performance of this Hamiltonian for different values of $\Delta$, noting that $H_{XYZ}(0;N)$ is simply the J1 Heisenberg spin chain of length N.
As Fig. \[fig:XYZ\_plot\] shows, a slight advantage can be gained by using an XYZ model rather than a simple Heisenberg chain. Fig. \[fig:XYZ\_plot\] also seems to suggest that the benefit gained is relatively independent of chain length.
![\[fig:XYZ\_plot\] Plot of fractional difference from annealing time for an chain with small $\Delta$ (Heisenberg chain). This data is for the adiabatic quantum bus protocol performed on a chain of the form eq. \[eq:XYZ\_h\] with spins being attached and removed simultaneously.](figure13)
Other Protocols
===============
So far we have only investigated a small subset of the possible annealing protocols which meet the criteria given in the introduction. For example the XY spin chain should also have and easily prepared ground state and may be easier to experimentally realize [@Johnson2011]. One could also try to examine the case of dynamically tuning the y and or z direction coupling and starting out at the Majumdar-Ghosh point but using modified coupling in the y and z directions with an XYZ model to avoid low gap regions.
One could also try to change the coupling scheme to avoid the low gap region, by either randomly or systematically modifying the coupling between intermediate spins, if this is done dynamically, one can still take advantage of the Majumdar-Ghosh point. This technique could also be used in conjunction with any of the ideas in the previous paragraph.
This paper is intended only to provide proof of principle for this method and is by no means an exhaustive search of all possible protocols.
Conclusions
===========
We have demonstrated how a J1-J2 Heisenberg spin chain can be used to transport a q-bit state adiabatically. We have also shown that many extensions of this Hamiltonian; such as different coupling schemes or the XY or XYZ model which have only a $\mathbb{Z}_{2}$ symmetry, will also be able to be used to transport a q-bit [^8]. We have found that for values of high frustration, transport by quantum annealing does not work very well. We have also demonstrated that this does not prevent us from exploiting the easily prepared ground state at the Majumdar-Ghosh point. We have given some examples of possible annealing protocols in this paper, but have really only investigated a very small section of a vast space of possible protocols for transportation of quantum states by annealing.
Acknowledgements {#acknowledgements .unnumbered}
================
The numerical computations were carried out on the University of Southern California high performance supercomputer cluster. This research is partially supported by the ARO MURI grant W911NF-11-1-0268.
[10]{} M. W. Johnson et al. Nature 473, 194-198 (12 May 2011)
R. Harris et al. Physical Review B 81, 134510 (2010)
A. Perdomo et al. Physical Review A 78, 012320 (2008)
S. H. W. van der Ploeg et al. IEEE Trans. App. Supercond. 17, 113 (2006)
R. Harris et al. Phys. Rev. B 82, 024511 (2010)
Z. Chen, Z. Zhou, X. Zhou, X. Zhou, and G. Guo, Phys. Rev. A 81, 022303 (2010)
L. Banchi, T. J. G. Apollaro, A. Cuccoli, R. Vaia, and P. Verrucchi, Phys. Rev. A 82, 052321 (2010)
L Banchi, T J G Apollaro, A Cuccoli, R Vaia, and P Verrucchi, New Journal of Physics 13 (2011) 123006
T. J. G. Apollaro, L. Banchi, A. Cuccoli, R. Vaia, P. Verrucchi, arXiv:1203.5516v1 [\[]{}quant-ph[\]]{} 2012
L. Banchi, A. Bayat, Paola Verrucchi, and Sougato Bose, Phys. Rev. Lett. 106, 140501 (2011)
J. Fitzsimmons and J. Twamley, Phys. Rev. Lett. 97, 090502 (2006)
N. Chancellor and S. Haas, Phys. Rev. B 84, 035130 (2011)
Phys. Rev. B 52, 6581 - 6587 (1995)
C K Majumdar, J. Phys. C: Solid State Phys. 3 911 (1970)
[^1]: Technically one must give the additional condition that there is no true crossing within the spin sectors on the annealing path.
[^2]: At least this should work for small systems. In the continuum limit many of these systems may become gapless, so that quantum annealing cannot be effectively performed. Also one may be able to construct certain pathological cases with paths which pass through true crossings.
[^3]: Note that this Hamiltonian (and all other annealing Hamiltonians in this paper) can be rewritten in the form of \[eq:H anneal\]. However it is much more compact not to write the unchanging parts of the Hamiltonian twice.
[^4]: At least for fixed coupling, the case of dynamically changing coupling will be considered later.
[^5]: Strictly speaking nothing in this paper has demonstrated them to be true crossings, they could just be close avoided crossings, it does not matter for the purposes of this paper.
[^6]: For example if two spins which should be in a singlet together ended up being measured to be facing in the same direction than the annealing process would have failed.
[^7]: This statement is based on the fact that the gap does not have a cusp when plotted on a log scale. Strictly speaking this just shows that there is not a true crossing at the line where the two couplings are equal.
[^8]: Assuming there is not a true crossing along the annealing path, the coupling must also be (at least predominately) anti-ferromagnetic so that the excess spin does not become trapped in the larger spin chain.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A non-Hermitian random matrix model proposed a few years ago has a remarkably intricate spectrum. Various attempts have been made to understand the spectrum, but even its dimension is not known. Using the Dyson-Schmidt equation, we show that the spectrum consists of a non-denumerable set of lines in the complex plane. Each line is the support of the spectrum of a periodic Hamiltonian, obtained by the infinite repetition of any finite sequence of the disorder variables. Our approach is based on the “theory of words.” We make a complete study of all 4-letter words. The spectrum is complicated because our matrix contains everything that will ever be written in the history of the universe, including this particular paper.'
author:
- |
D.E. Holz$^1$, H. Orland$^{1,2}$, and A. Zee$^{1,3}$\
$^{1}$Institute for Theoretical Physics,\
University of California, Santa Barbara, CA 93106, USA\
$^{2}$Service de Physique Théorique, CEA-Saclay,\
91191 Gif-sur-Yvette Cedex, France\
$^{3}$Center for Advanced Study, Tsinghua University,\
Beijing 10084, People’s Republic of China
title: 'On the Remarkable Spectrum of a Non-Hermitian Random Matrix Model'
---
A Class of Non-Hermitian Random Matrix Models
=============================================
Some time ago, Feinberg and one of us (in a paper to be referred to as FZ [@FZ]) proposed the study of the equation $$\psi _{k+1}+r_{k-1}\psi _{k-1}=E\psi _k \label{Seq}$$ where the real numbers $r_k$ are generated from some random distribution. Two particularly simple models were studied: (A) the $r_k$’s are equal to $%
\pm 1$ with equal probability, and (B) $r_k=e^{i\theta _k}$ with the angle $%
\theta _k$ uniformly distributed between $0$ and $2\pi .$
Imposing the boundary condition $\psi _{0}=\psi _{N+2}=0$ on (\[Seq\]) we can write the set of equations as the eigenvalue equation $$H_{N}\psi =E\psi$$ with $\psi $ the column eigenvector with components $\{\psi _{1},$ $\psi
_{2},$ $\cdots ,$ $\psi _{N},$ $\psi _{N+1}\}$ and $H_{N}$ the $(N+1)$ by $%
(N+1)$ non-Hermitian random matrix $$H_{N}=\left(
\begin{array}{llllll}
0 & 1 & 0 & & & \\
r_{1} & 0 & 1 & 0 & & \\
0 & r_{2} & 0 & 1 & 0 & \\
& 0 & \ddots & \ddots & \ddots & 0 \\
& & 0 & r_{N-1} & 0 & 1 \\
& & & 0 & r_{N} & 0
\end{array}
\right)$$
While quantum mechanics is of course Hermitian it is convenient to think of $%
H$ as a Hamiltonian and (\[Seq\]) as the non-Hermitian Schrödinger equation describing the propagation of a particle hopping on a 1-dimensional lattice.
Some applications of non-Hermitian random Hamiltonians include vortex line pinning in superconductors [@1; @2; @fz2] and growth models in population biology [@4]. A genuine localization transition can occur for random non-Hermitian Schrödinger Hamiltonians [@5; @6; @7; @8; @9; @10; @11; @12] in one dimension.
As mentioned in FZ, with the open chain boundary condition $\psi _{0}=\psi
_{N+2}=0$ the more general equation $$s_{k+1}\psi _{k+1}+r_{k-1}\psi _{k-1}=E\psi _{k} \label{schrody}$$ can always be reduced to (\[Seq\]) by an appropriate “gauge” transformation $\psi _{k}\rightarrow \lambda _{k}\psi _{k}.$ Furthermore, applying the transformation $\psi _k\rightarrow u^{-k}\psi _k$ to (\[Seq\]) we see that if we change $r_k\rightarrow u^2r_k$ then the spectrum changes by $E\rightarrow uE.$ Thus, scaling the magnitude of the $%
r_k$’s merely stretches the spectrum, and flipping the sign of all the $r_k$’s corresponds to rotating the spectrum by $90^0.$
It is also useful to formulate the problem in the transfer matrix formalism. Write (\[Seq\]) as $$\left(
\begin{array}{l}
\psi _{k+1} \\
\psi _{k}
\end{array}
\right) =T_{k-1}\left(
\begin{array}{l}
\psi _{k} \\
\psi _{k-1}
\end{array}
\right)$$ where the transfer matrix $T_{k}$ is defined as the 2 by 2 matrix $$T_{k}=\left(
\begin{array}{ll}
E & -r_{k} \\
1 & 0
\end{array}
\right)$$ Define $P\equiv T_{N}T_{N-1}\cdots T_{2}T_{1}.$ Then the boundary condition implies $$EP_{11}+P_{12}=0$$ The solution of this polynomial equation in $E$ determines the spectrum.
Since $H$ is non-Hermitian the eigenvalues invade the complex plane. For model B, the spectrum has an obvious rotational symmetry and forms a disk (see Fig. \[fig1\], which displays the support of the density of states).
An expansion of the density of eigenvalues around $E=0$ to very high orders in $E$ has been given by Derrida et al. [@derrida]. This analytic expansion however cannot predict singularities in the density of states.
In contrast, for model A the spectrum enjoys only a rectangular $%
Z_{2}\otimes Z_{2}$ symmetry. The first $Z_{2}$ corresponds to $E\rightarrow
E^{\ast }$ obtained by complex conjugating the eigenvalue equation $H\psi
=E\psi .$ The second $Z_{2}$ corresponds to $E\rightarrow -E$ obtained by the bipartite transformation $\psi _{k}\rightarrow (-)^{k}\psi _{k}.$ Remarkably, FZ found that the spectrum has an enormously complicated fractal-like form. In Fig. \[fig2a\] we plot the support of the density of eigenvalues in the complex plane for a $4000\times 4000$ matrix, for a specific realization of the disorder.
In Fig. \[fig2b\] we plot the support of the density of states in the complex plane for a $1000\times1000$ matrix, averaged over 100 realizations of the disorder.
Contrasting Figs. \[fig2a\] and \[fig2b\] with Fig. \[fig1\], one can see why it has been a challenge for mathematical physicists to understand the nature of the spectrum.
Basic Formalism
===============
In general, for random non-Hermitian matrices $H$, the density of eigenvalues can be obtained by $$\rho (x,y)=\frac{1}{\pi }\frac{\partial }{\partial z^{\ast }}G(z,z^{\ast })
\label{rho}$$ where the Green’s function is defined by $$G(z,z^{\ast })=\left\langle{\frac{1}{N}\mathrm{tr}\frac{1}{z-H}}\right\rangle$$ with the bracket denoting averaging and $z=x+iy$ (see, for example, ref. [@8] for a proof of these relations). Equation (\[rho\]) follows from the identity $$\frac{\partial }{\partial z^{\ast }}\frac{\partial }{\partial z}\mathrm{log}%
\text{ }z=\frac{\partial }{\partial z^{\ast }}\frac{1}{z}=\pi\,\delta
(x)\,\delta (y)\label{Cauchy}$$ where $z=x+iy$. Expanding $G(z,z^{\ast })=\frac{1}{z}\sum_{k=0}^{\infty }{z^{-k}}%
\left\langle{\frac{1}{N}\mathrm{tr}H^{k}}\right\rangle $ we see that $\left\langle{\frac{1}{N}
\mathrm{tr}H^{k}}\right\rangle $ counts the number of paths of a particle returning to the origin in $k$ steps.
Evidently, for model A each link has to be traversed 4 times. The spectrum of model A was studied by Cicuta et al. [@cicuta] and by Gross and one of us [@GZ] by counting paths. In particular, Cicuta et al. gave an explicit expression for the number of paths. Recently, the more general problem of even-visiting random walks has been studied extensively (see ref. [@Bau]). In addition, exact analytic results for the Lyapunov exponent have been found by Sire and Krapivsky [@sire].
In this paper we propose a different approach based on the theory of words. We will focus on model A although some of our results apply to the general class of models described by (\[schrody\]).
One important issue is the dimensionality of the spectrum of model A. In general, the spectrum of non-Hermitian random matrices when averaged over the randomness is 2-dimensional (for example, the spectrum of model B). Many of the authors who have looked at model A believe that its spectrum, as shown in Figs. \[fig2a\] and \[fig2b\], is quasi-0-dimensional: the spectrum seems to consist of many accumulation points. Here we claim that the dimension of the spectrum actually lies between $1$ and $2$ dimensions, in the sense described below.
Distribution of the Characteristic Ratio
========================================
Consider the degree $(N+1)$ characteristic polynomial $\Delta
_{N+1}(z)\equiv \mathrm{\det }(H_{N}+zI_{N+1})$, where $I_{N+1}$ denotes the $(N+1)\times(N+1)$ identity matrix. We easily obtain the recursion relation $$\Delta _{N+1}(z)=z\Delta _{N}(z)-r_{N}\Delta _{N-1}(z)$$ with $\Delta _{0}(z)=1$ and $\Delta _{1}(z)=z.$ Note that this second order recursion relation can be expressed in terms of transfer matrices as
$$\left(
\begin{array}{c}
\Delta _{N+1}(z) \\
\Delta _{N}(z)
\end{array}
\right) =\left(
\begin{array}{cc}
z & -r_{N} \\
1 & 0
\end{array}
\right) \left(
\begin{array}{c}
\Delta _{N}(z) \\
\Delta _{N-1}(z)
\end{array}
\right) \label{TM}$$
very similar to those defined for the wave function $\psi $, with the transfer matrix given by
$$U_{N}=\left(
\begin{array}{cc}
z & -r_{N} \\
1 & 0
\end{array}
\right). \label{TM1}$$
Following Dyson and Schmidt [@Dyson; @Schmidt] we consider the characteristic ratio $$y_{k}(z)\equiv \frac{\Delta _{k}(z)}{\Delta _{k-1}(z)}, \label{y}$$ which satisfies the recursion relation $$y_{k+1}=z-\frac{r_{k}}{y_{k}} \label{recursion}$$ with initial condition $y_1=z$.
The hope is that, while the characteristic polynomials $\Delta _{k}(z)$ obviously changes dramatically as $k$ varies, the characteristic ratio $%
y_{k}(z)$ might converge asymptotically.
From the definition in eq. (\[y\]), it is clear that a point $z$ is in the spectrum of the $N \times N$ matrix iff $y_N(z)=0$ and $y_{N+1}(z)=\infty$. Thus a point $z$ belongs to the spectrum if the corresponding set of variables $\{y_N\}$ is unbounded, that is, if the probability of escape of the variable $y_n$ to $\infty$ is finite. As is shown in [@Bau], this condition is sufficient to determine the spectrum along the real axis ($z \in
R$), but is insufficient in the complex case.
Let $P_{k}(y_{k})$ be the probability distribution of $y_{k}$ (note that $y_{k}$ is complex). Then $$P_{k+1}(y_{k+1})=\left\langle{\int d^{2}y_{k}\,P_{k}(y_{k})\delta^{(2)}(y_{k+1}-z+
\frac{r_{k}}{y_{k}})}\right\rangle$$ where the brackets denote the average over the disorder variables $\{r_{k}\}$. In the thermodynamic limit $k\rightarrow \infty $, under fairly general conditions [@osedelec] it can be shown that the probability distribution has a limit $P(y)$, called the invariant distribution, which is determined by the self-consistent equation $$\begin{aligned}
P(y)&=&\left\langle{\int\,d^{2}t\,P(t)\delta
^{(2)}(y-z+\frac{r}{t})}\right\rangle\\
&=&\left\langle
{\frac{\mid r\mid ^{2}}{\mid z-y\mid ^{4}}P(\frac{r}{z-y})}\right\rangle,\end{aligned}$$ where we have used the fact that $\delta ^{(2)}(f(z))$ near a zero, $z_{0}$, of $f(z)$ is given by $\delta ^{(2)}(z-z_{0})\,|\frac{df}{dz}|^{-2}$.
For model A, we obtain the amusing equation $$P(y)=\frac{1}{2\mid z-y\mid ^{4}}\left( P({\frac{1}{z-y}})+P({\frac{-1}{z-y}
})\right). \label{DS}$$
This type of equation has been studied extensively for the real case in [@aeppli; @orland; @luck]. It can be shown that it can be solved by the Ansatz $$P(y)=\sum_{j}a_{j}\delta ^{(2)}(y-b_{j}(z)), \label{ansatz}$$ where the $b_{j}$’s depend on $z$ whereas the $a_{j}$’s don’t. Note that the index $j$ is not necessarily an integer and can refer to a continuous set. In addition, it can be shown that the $b_{j}(z)$ are the stable fixed points of the product of any sequence of transfer matrices $U$ (see the next section).
Plugging (\[ansatz\]) into (\[DS\]) we have $$\begin{aligned}
\sum_{j}a_{j}\delta ^{(2)}(y-b_{j})=\frac{1}{2}\sum_{j}a_{j}&&\!\!\!\!\!\!\{\delta
^{(2)}(y-z+\frac{1}{b_{j}})\\
&&+\delta ^{(2)}(y-z-\frac{1}{b_{j}})\}.\end{aligned}$$ Since the right hand side has twice as many delta function spikes as the left hand side, for the two sides to match we expect that, in general, the index $j$ would have to run over an infinite set.
For a given complex number $z$, we demand that the two sets of complex numbers $\{b_{j}\}$ and $\{z-1/b_{j},z+1/{b_{j}}\}$ be the same. This very stringent condition should then determine the $%
\{b_{j}(z)\}.$
To see how this works, focus on a specific $b_{1}$ (Since the label $j$ has not been specified this can represent any $b_{j}$). It is equal to either $z-%
{1}/{b_{2}}$ or $z+{1}/{b_{2}}$ for some $b_{2}.$ But $b_{2}$ must in its turn be equal to either $z-{1}/{b_{3}}$ or $z+{1}/{b_{3}}$ for some $b_{3}$. This process of identification must continue until we return to $b_{1}$. Indeed, if the process of return to $b_{1}$ occurs in a finite number $L$ of steps, then it will repeat indefinitely (since the system is back at its starting point $b_{1})$. It is this infinite repetition which gives a finite weight to the $\delta -$function at $b_{1}$. By contrast, if the number of steps needed to return to the initial point $b_{1}$ is infinite, then the weight associated with this point vanishes, and it will not be present in the spectrum. We thus conclude that the support of the distribution of $P(y)$ is the closure of the set of all the stable fixed points of the product of any sequence of transfer matrices $U$. We also conclude that the support of the density of states of the non-Hermitian matrix, in the thermodynamic limit, is given by the zeroes of any stable fixed point: $b_j(z)=0$. What is important to notice is that the $a_{j}$ are independent of $z$ and depend only on the length of the word. Thus, we conclude that the set of complex numbers $\{b_{j}(z)\}$ is determined by the solution of an infinite number of fixed point equations.
The Theory of Words
===================
It is useful here to introduce the theory of words. A word $w$ of length $L$ is defined as the sequence $\{w_1,w_2,\cdots ,w_L\}$ where the letters $w_j=\pm 1$. In other words, we have a binary alphabet. Let us also define the repetition of a given word a specific number of times as a simple sentence. We can then string together simple sentences to form paragraphs.
For a given word $w$ of length $L$, consider a function $f_{L}(b;z,w)$ to be constructed iteratively. For notational simplicity we will suppress the dependence of $f_{L}$ on $b,$ $z,$ and $w,$ indicating only its dependence on the length $L$ of the word $w.$ The iteration begins with $$f_{1}=z-\frac{w_{1}}{b}$$ and continues with $$f_{j+1}=z-\frac{w_{j+1}}{f_{j}}.$$ We define $f(b;z,w)\equiv f_{L}.$
The set of complex numbers $\{b_{j}(z)\}$ are then determined as follows. Consider the set of all possible words. For each word $w$, determine the solution of the fixed point equation $$b=f(b;z,w).$$ By considering small deviations from the solution, we see that the solution is a stable fixed point only if $$\left|\frac{\partial f(b;z,w)}{\partial b}\right|<1. \label{stable}$$ The set of all possible words $w$ generates the set of complex numbers $%
\{b_{j}(z)\}.$ In other words, $b$ is determined by a continued fraction equation, since $$f(b;z,w)=z-\frac{w_{L}}{z-\frac{w_{L-1}}{z-\frac{w_{L-2}}{\frac{\ddots }{z-%
\frac{w_{1}}{b}}}}}.$$ We see that $f_{j}$ has the form $$f_{j}=\frac{\alpha _{j}b+\beta _{j}}{\alpha _{j-1}b+\beta _{j-1}}, \label{fj}$$ with the polynomials $\alpha _{j}$ and $\beta _{j}$ determined by the recursion relations $$\alpha _{j+1}=z\alpha _{j}-w_{j+1}\alpha _{j-1} \label{alpha}$$ and $$\beta _{j+1}=z\beta _{j}-w_{j+1}\beta _{j-1}, \label{beta}$$ with the initial condition $\alpha _{0}=1,$ $\alpha _{1}=z,$ $\beta _{0}=0,$ and $\beta _{1}=-w_{1}.$ We notice that $\alpha _{j}$ and $\beta _{j}$ satisfy the same recursion relation as that satisfied by $\Delta _{j}$ with the correspondence $w_{j+1}\leftrightarrow r_{j}.$ Note also that (\[alpha\]) and (\[beta\]) can be packaged as the matrix equation $$\left(
\begin{array}{ll}
\alpha _{j+1} & \beta _{j+1} \\
\alpha _{j} & \beta _{j}
\end{array}
\right) =\left(
\begin{array}{ll}
z & -w_{j+1} \\
1 & 0
\end{array}
\right) \left(
\begin{array}{ll}
\alpha _{j} & \beta _{j} \\
\alpha _{j-1} & \beta _{j-1}
\end{array}
\right),$$ where the transfer matrix $U_{j}$ defined in the previous section appears. This is closely related to the transfer matrix formalism discussed earlier. Indeed, defining $W_{j}\equiv \left(
\begin{array}{ll}
\alpha _{j} & \beta _{j} \\
\alpha _{j-1} & \beta _{j-1}
\end{array}
\right) $ we have the initial condition $W_{1}=\left(
\begin{array}{ll}
z & -w_{1} \\
1 & 0
\end{array}
\right)$. Hence a given word $w$ of length $L$ can also be characterized by a matrix $$W=\left(
\begin{array}{ll}
\alpha & \beta \\
\gamma & \delta
\end{array}
\right),$$ where for convenience we have written $\alpha =\alpha _{L},$ $%
\beta =\beta _{L},$ $\gamma =\alpha _{L-1}$, and $\delta =\beta _{L-1}.$
For a given word $w,$ the fixed point value $b$ is determined by the quadratic equation $$b=\frac{\alpha b+\beta }{\gamma b+\delta },$$ which is the fixed point equation of the homographic mapping associated with the matrix $M$. The geometric interpretation is clear: The matrix $W$ acts on $2-$component vectors $v$, and we ask for the set of $v$ such that the ratio $b$ of the first component to the second component is left invariant by the transformation. In other words, we look for the projective space left invariant by the transformation $W$: the fixed point value $b$ defines the direction of the invariant ray.
Hence $b$ is given by $$b=\frac{P(z)\pm \sqrt{Q(z)}}{2R(z)}, \label{b}$$ with $P,Q,R$ polynomials of degree $2L$ in $z$, where $L$ denotes the length of the word $w$. Explicitly, $$P_{L}(z)=\alpha _{L}-\beta _{L-1},$$ $$Q_{L}(z)=(\alpha _{L}-\beta _{L-1})^{2}+4\alpha _{L-1}\beta _{L},$$ and $$R_{L}(z)=\alpha _{L-1}.$$
The stability condition (\[stable\]) determines which root of (\[b\]) is to be chosen.
We will see shortly that $b(z)$ determines the spectrum. Anticipating this, we see that if we form a compound word by stringing the word $w$ together twice (for example, the Japanese word “nurunuru”) then we expect the contribution to the spectrum to be the same. But given the preceding discussion, this is obvious, since if a ray is left invariant by $W,$ it is manifestly left invariant by $W^{2}.$
Density of Eigenvalues
======================
Once we have determined $b_j(z)$, how do we extract the density of eigenvalues? The eigenvalues $\{\lambda _{i}\}$ of the matrix $H$ are given by $\Delta
_{N}(z)=\Pi _{i=1}^{N}(z-\lambda _{i})$. From (\[y\]) we have $%
\sum_{k=1}^{N}\log y_{k}\simeq \log \Delta
_{N}(z)=\sum_{i=1}^{N}\log (z-\lambda _{i})$, and thus $$\int \,d^{2}y\,P(y)\log y=\left<\frac{1}{N}\sum_{i=1}^{N}\log
(z-\lambda _{i})\right>. \label{bob}$$ Using the identity (\[Cauchy\]) we can differentiate the right hand side of (\[bob\]) to obtain the density of eigenvalues in the complex plane $$\rho =\frac{1}{\pi }\frac{\partial }{\partial z^{\ast }}\frac{\partial }{%
\partial z}\int\,d^{2}y\,P(y)\log y.$$ Plugging in our solution $$P(y)=\sum_{j}a_{j}\delta ^{(2)}(y-b_{j})$$ we finally deduce that $\rho =\frac{1}{\pi }\frac{\partial }{\partial
z^{\ast }}\frac{\partial }{\partial z}\sum_{j}a_{j}\log b_{j}=\frac{1%
}{\pi }\frac{\partial }{\partial z^{\ast }}\sum_{j}a_{j}\frac{1}{b_{j}}\frac{%
\partial b_{j}}{\partial z}.$
Since the $a_{j}$ do not depend on $z$, the spectrum is determined by the zeroes of the fixed point solutions $b_{j}(z).$
We see from (\[b\]) that the density of eigenvalues is given as a sum over $j$ of terms like $$\frac{\partial }{\partial z^{\ast }}\left\{\frac{1}{P(z)\pm \sqrt{Q(z)}}[%
P^{\prime }(z)\pm \frac{Q^{\prime }(z)}{2\sqrt{Q(z)}}]-\frac{R^{\prime }(z)}{%
R(z)}\right\}.$$ Thus the spectrum consists of isolated poles given by the zeroes of $R$ and $P(z)\pm \sqrt{Q(z)}$, and of the cuts of $\sqrt{Q(z)}$, and is made of isolated points plus curved line segments connecting the zeroes of $Q(z).$
Contrary to what some authors have believed, the spectrum is not $0-$dimensional, but $(0+1+\delta)-$dimensional, with $\delta \leq 1$. Each word gives rise to a line segment, and words which differ slightly from each other gives rise to line segments near each other. Indeed, given a word $w$, it is possible to construct a word $w^{\prime }$ with a spectrum as close to that of $w$ as desired. For that purpose, we may construct $w^{\prime }$ as $
w^{\prime }=w^{l_{1}}vw^{l_{2}}$ where $v$ is any “corrupting” word, and the two lengths $l_{1}$ and $l_{2}$ are sufficiently long. Indeed, in terms of transfer matrices and invariant rays, we see that $w^{l_{2}}$ acting on any initial ray brings it close to the stable invariant ray of $w$. Then the direction of this ray is corrupted by $v$, but it is brought back arbitrarily close to the invariant ray of $w$ by applying the transfer matrix $
w^{l_{1}}$, provided that $l_{1}$ is large enough. Presumably (although this remains to be proven rigorously), the spectrum associated with the corrupted word $w^{\prime }$ can be made as close as we want to that of $w$. We have thus this property that for any word $w$, there is a word $w^{\prime }$ generating a spectrum as close as we want to that of $w$. In Figs. \[fig3a\] and \[fig3b\] we plot the eigenstates of a word $w=\{++-\}$ and the spectrum of the word $
w^{\prime }=w^{16}\{+++\}w^{16}$. We see that the two spectra are very close.
As is clear from this discussion, the spectrum is indeed “incredibly complicated.”
Words and Spectral Curves
=========================
We content ourselves by focusing on the cuts of $\sqrt{Q(z)}.$ Since for a word $w$ of length $L,$ $Q(z)$ is a polynomial of degree $2L$ with $2L$ roots, it gives rise to $L$ curved line segments. The curves are given by the condition $$\mathrm{Im}\,Q(z)=0. \label{imq}$$ (The sign of $\,Q(z)$ depends on the choice for the square root branch cut.) Given a word $w$ of length $L$, the corresponding spectrum must be invariant under a cyclic permutation of the letters $%
w_{1},w_{2},\cdots ,w_{L},$ namely under $w_{1}\rightarrow w_{2},$ $%
w_{2}\rightarrow w_{3},\cdots ,w_{L}\rightarrow w_{1}.$
As an example, for the $3-$letter word $w=\{++-\},$ $R(z)=z^{2}+1$ and $%
Q(z)=z^{6}-2z^{4}+z^{2}+4$, which has roots at $z=\pm i,$ $z=\pm (\sqrt{7}-i)/2$, and $z=\pm (\sqrt{7}+i)/2$.
It is now clear what the words correspond to “physically”: a matrix $H$ with $r_{k}$’s given by an endless repetition of $(+1,+1,-1)$ has a spectrum given by a straight line connecting $z=\pm i,$ and two algebraic curves connecting $z=(\sqrt{7}+i)/2$ to $z=(\sqrt{7}-i)/2$ and $z=-(\sqrt{7}+i)/2$ to $z=-(\sqrt{7}-i)/2$, plus poles at $z=\pm i.$ Notice that the two poles are buried under a cut. In Fig. \[fig4\] we show the spectrum associated with the word $\{++-\}$ together with the spectrum of a random $1000\times1000$ matrix.
We now give a complete study of all $4-$letter words. The polynomial is easily found to be $$Q_{4}(z)=z^{8}-2sz^{6}+(s^{2}+2\kappa )z^{4}-2s\kappa z^{2}+\omega ^{2}$$ with $s\equiv w_{1}+w_{3}+w_{2}+w_{4},$ $\kappa \equiv w_{1}w_{3}+w_{2}w_{4}$, and $\omega \equiv w_{1}w_{3}-w_{2}w_{4}.$ The condition (\[imq\]) for the curves in the spectrum reduces to $$xy\left(y^{2}-x^{2}+\frac{s}{2}\right)\left(x^{4}+y^{4}-6x^{2}y^{2}+s(y^{2}-x^{2})+k\right)=0.$$ There are only three non-trivial $4-$letter words, namely $w=\{+++-\},$ $%
\{++--\},$ and $\{+---\}.$ Their contribution to the spectrum of $H$ together with the spectrum of a random $4000\times4000$ matrix is shown in Fig. \[fig5a\].
In Fig. \[fig5b\] we show the contribution of all one, two, three, and four letter words to the density of states.
Thus, an $N$ by $N$ matrix $H$ with $r_{k}$’s given by repeating the word $w$ has a spectrum determined by the stable fixed point value $b(z)$ corresponding to $w.$ Furthermore, consider an $N$ by $N$ matrix $H$ with $%
r_{k}$’s given by first repeating the word $w_{1}$ (of length $L_{1})$ $%
N_{1} $ times and then by repeating the word $w_{2}$ (of length $L_{2})$ $%
N_{2}$ times. As we would expect, in the limit in which $N_{1}$, $N_{2},$ and $N=N_{1}L_{1}+N_{2}L_{2}$ all tend to infinity, the spectrum of $H$ is given by superposing the spectra of $H_{i}$ $(i=1,2)$, where $H_{i}$ is constructed with $r_{k}$’s given by first repeating the word $w_{i}$ $N_{i}$ times. This clearly generalizes. In Figs. \[fig6a\]–\[fig6c\] we show the spectrum of the word $w_{1}=\{++--\}$, the spectrum of $w_{2}=\{+++-\}$, and the spectrum of the word $w=w_{1}^{20}w_{2}^{20}$.
We see the superposition principle at work.
From this discussion it becomes clear why the spectrum of the matrix $H$ in FZ is so complicated. The sequence $\{r_{1},r_{2},\cdots r_{\infty }\}$ is a book written in the binary alphabet that, in the mathematical limit $%
N\rightarrow \infty $, contains all possible words, sentences, and paragraphs. In fact, $H$ contains everything ever written or that will be written in the history of the universe, including this particular paper. This familiar mind boggling fact accounts for the complicated looking spectrum first observed in FZ.
It also explains why numerical studies of the spectrum suggest that it is $%
0-$dimensional. Even for $N$ as large as $1000$ the sequence contains an infinitesimally small subset of the set of all possible words, sentences, and paragraphs.
Eigenvalues on the Unit Circle
==============================
In the ensemble of all books there are particularly simple books such that $%
\{r_1,r_2,\cdots r_\infty \}$ consists of a word $w$ of length $L$ repeated again and again. In this case, we can determine the spectrum explicitly by two different methods.
Let $W$ be the transfer matrix corresponding to $w.$ In other words, $%
W=\prod\limits_{j=1}^{L}\left(
\begin{array}{ll}
z & -w_{j} \\
1 & 0
\end{array}
\right) ,$ where the matrix product is ordered. After repeating the word $R$ times, we have $$\left(
\begin{array}{l}
\Delta _{RL+1} \\
\Delta _{RL}
\end{array}
\right) =W^{R}\left(
\begin{array}{l}
\Delta _{1} \\
\Delta _{0}
\end{array}
\right).$$ Diagonalizing $W=S^{-1}\left(
\begin{array}{ll}
\lambda _{1}(z) & 0 \\
0 & \lambda _{2}(z)
\end{array}
\right) S,$ we see immediately that $\Delta _{RL}$ is a linear function of $%
\lambda _{1}^{R}(z)$ and $\lambda _{2}^{R}(z):$$$\Delta _{RL}=\alpha \lambda _{1}^{R}+\beta \lambda _{2}^{R}. \label{drl}$$ We remind the reader that all quantities in (\[drl\]) are functions of $z.$
The spectrum of $H$ is determined by the zeroes of $\Delta _{RL}(z)$ as $%
R\rightarrow \infty .$ We note that in this limit the solution of $$\Delta _{RL}(z)=0 \label{zero}$$ does not depend on knowing the detailed form of $\alpha $ and $\beta .$ Indeed, (\[zero\]) implies $$\lambda _{1}=\left(-\frac{\beta }{\alpha }\right)^{\frac{1}{R}}\lambda _{2}$$ or $$\lambda _{1}^{2}=\pm \left(-\frac{\beta }{\alpha }\right)^{\frac{1}{R}},$$ since $\mathrm{\det }\,W=\lambda _{1}\lambda _{2}=\pm 1.$ In the limit $%
R\rightarrow \infty ,$ $(-{\beta }/{\alpha })^{\frac{1}{R}}$ tends towards a ($z-$dependent) complex number of modulus unity. Thus, we conclude that $$|\lambda (z)|=1 \label{unit},$$ namely, that the eigenvalues of $W$ lie on the unit circle. This constraint suffices to determine the eigenvalues of $H.$ Plugging (\[unit\]), that is $\lambda =e^{i\theta },$ into the eigenvalue equation $$\lambda ^{2}-(\mathrm{tr}\,W)\lambda +\mathrm{\det }\,W=0,$$ we obtain $(\mathrm{tr}\,W)=2\cos\theta $ if $\mathrm{\det }\,W=+1$ and $(\mathrm{tr}\,W)=2i\sin \theta $ if $\mathrm{\det }\,W=-1$, which we can combine into the single equation $$(\mathrm{tr}\,W)=2(\mathrm{\det }\,W)^{\frac{1}{2}}\cos\theta
\label{result}$$ after a trivial phase shift.
As $\theta $ ranges from $0$ to $2\pi ,$ this traces out the spectrum in the complex plane. As an example, consider the $4-$letter word $\{+++-\}$, in which case (\[result\]) reduces to $$\label{alg}
z^{4}-2z^{2}=2i\cos\theta.$$ This traces out the algebraic curve shown in Fig. \[fig7\], which is to be compared to Fig. \[fig6b\].
Of course, since $H$ is now translation invariant with period $L$, we can apply Bloch’s theorem to determine the spectrum of $H.$ Imposing $\psi
_{k+L}=e^{i\varphi }\psi _{k}$ we reduce the eigenvalue problem of $H$ to the eigenvalue problem of the $L$ by $L$ matrix $$h_{L}=\left(
\begin{array}{llllll}
0 & 1 & 0 & \cdots & 0 & r_{L}e^{-i\varphi } \\
r_{1} & 0 & 1 & \ddots & 0 & 0 \\
0 & r_{2} & 0 & 1 & \ddots & \vdots \\
\vdots & \ddots & \ddots & \ddots & \ddots & 0 \\
0 & 0 & \ddots & r_{L-2} & 0 & 1 \\
e^{i\varphi } & 0 & \cdots & 0 & r_{L-1} & 0
\end{array}
\right).$$ One can verify that with a suitable relation between $\theta $ and $\varphi $ the eigenvalue equation $\mathrm{Det}\,(zI_{L}-h)=0$ becomes identical to (\[result\]).
Conclusion and Open Questions
=============================
We have seen that the structure of this simple tridiagonal non-Hermitian random matrix possesses an amazing richness. This complexity can be understood if one realizes that the spectrum of the random matrix is the sum of the spectra of all tridiagonal matrices with a periodic subdiagonal obtained by repeating an infinite number of times any finite word of length $L$, weighted by a factor $1/2^{L}$.
The number of lines does not have the cardinal of the continuum. The number of lines is equal to the number of words of any length that can be made with a 2-letter alphabet; this is a countable number.
There are many open questions concerning the fine structure of the spectrum, such as whether the spectrum contains holes in the complex plane (in its domain of definition). We also have not touched upon the question of the nature of the eigenstates. Are they localized or delocalized? Numerical data seems to suggest a localization transition. We hope to address these and other questions in future work.
Acknowledgements
================
One of us (HO) would like to thank M. Bauer, D. Bernard and J.M. Luck for helpful discussions. This work was partially supported by the NSF under grant PHY99-07949 to the ITP.
Appendix: Cyclic Invariants
===========================
As remarked in the text, the coefficients of $Q_{L}(z),$ an even polynomial of degree $2L,$ must be constructed out of the cyclic invariants made of $%
w_{1},w_{2},\cdots ,w_{L}.$ There is presumably a well-developed mathematical theory of cyclic invariants, but what we need we can easily deduce here.
For any $L,$ we have the two obvious cyclic invariants $s=\sum_{j=1}^Lw_j$ and $p=\Pi _{j=1}^Lw_j.$ The number of cyclic invariants grows rapidly with $%
L.$ Apparently different cyclic invariants can be constructed out of other cyclic invariants, for example, $(\sum_{j=1}^Lw_j)^2-(%
\sum_{j=1}^Lw_j^2)=2\sum_{j\neq i}^Lw_jw_i.$
It is easy to work out $Q_{L}(z)$ for low values of $L,$ as follows: $$Q_{2}(z)=z^{4}-2sz^{2}+d^{2}$$ with $d\equiv w_{1}-w_{2}$, $$Q_{3}(z)=(z^{3}-sz)^{2}-4p=z^{6}-2sz^{4}+s^{2}z^{2}-4p,$$ where we have written $Q_{3}(z)$ in a form which shows that its roots can be found explicitly, $$Q_{4}(z)=z^{8}-2sz^{6}+(s^{2}+2\kappa )z^{4}-2s\kappa z^{2}+\omega ^{2}$$ with $\kappa \equiv w_{1}w_{3}+w_{2}w_{4}$ and $\omega \equiv
w_{1}w_{3}-w_{2}w_{4},$ $$Q_{5}(z)=z^{10}-2sz^{8}+(s^{2}+2\kappa )z^{6}-2s\kappa z^{4}+\kappa
^{2}z^{2}-4p,$$ with $\kappa \equiv w_{1}w_{3}+w_{1}w_{4}+w_{2}w_{4}+w_{2}w_{5}+w_{3}w_{5},$ $$\begin{aligned}
Q_{6}(z)&=&z^{12}-2sz^{10}+(s^{2}+2\kappa )z^{8}-2(s\kappa +\rho
)z^{6}\nonumber\\
&&+(\kappa ^{2}+2s\rho )z^{4}-2\kappa \rho z^{2}+\delta ^{2},\end{aligned}$$ where $$\begin{aligned}
\kappa
&=&w_{1}w_{3}+w_{1}w_{4}+w_{2}w_{4}+w_{1}w_{5}+w_{2}w_{5}+w_{3}w_{5}\nonumber\\
&&+w_{2}w_{6}+w_{3}w_{6}+w_{4}w_{6}\nonumber\\
\rho &=&w_{1}w_{3}w_{5}+w_{2}w_{4}w_{6}\nonumber\\
\delta &=&w_{1}w_{3}w_{5}-w_{2}w_{4}w_{6}\nonumber\end{aligned}$$ and $$\begin{aligned}
Q_{7}(z) &=&z^{14}-2sz^{12}+(s^{2}+2\kappa )z^{10}-2(s\kappa +\rho
)z^{8} \nonumber\\
&&+(\kappa ^{2}+2s\rho )z^{6}-2\kappa \rho z^{4}+\rho ^{2}z^{2}-4p,\end{aligned}$$ with $$\begin{aligned}
\kappa &=&w_1w_3+w_1w_4+w_2w_4+w_1w_5+w_2w_5+w_3w_5 \nonumber\\
&&+w_1w_6+w_2w_6+w_3w_6+w_4w_6+w_2w_7+w_3w_7 \nonumber\\
&&+w_4w_7+w_5w_7\nonumber\\
\rho &=&w_1w_3w_5+w_1w_3w_6+w_1w_4w_6+w_2w_4w_6+ \nonumber\\
&&w_2w_4w_7+w_2w_5w_7+w_3w_5w_7\nonumber.\end{aligned}$$ The quantities $d,\kappa ,\omega ,$ $\kappa ,\rho ,\delta $ are manifestly cyclic invariants.
[99]{} J. Feinberg and A. Zee, Phys. Rev. E **59**, 6433 (1999).
N. Hatano and D. R. Nelson, Phys. Rev. Lett. **77**, 570 (1996).
N. Hatano and D. R. Nelson, Phys. Rev. B **56**, 8651 (1997).
J. Feinberg and A. Zee, Nucl. Phys. B **552** 599 (1999).
D. R. Nelson and N. M. Shnerb, Phys. Rev. E **58**, 1383 (1998).
I. Y. Goldsheid and B. A. Khoruzhenko, Phys. Rev. Lett. **80**, 2897 (1998).
P. W. Brouwer, P. G. Silvestrov, and C. W. J. Beenakker, Phys. Rev. B **56**, R4333 (1997).
K. B. Efetov, Phys. Rev. Lett. **79** 491 (1997).
J. Feinberg and A. Zee, Nucl. Phys. B **504** 579 (1997).
E. Brézin and A. Zee, Nucl. Phys. B **509**, 599 (1998).
A. Zee, Physica A **254**, 300 (1998).
N. Hatano, Physica A **254**, 317 (1998).
C. Mudry, P. W. Brouwer, B. I. Halperin, V. Gurarie, and A. Zee, Phys. Rev. B **58**, 13539 (1998).
B. Derrida, J. Lykke Jacobsen, and R. Zeitak, J. Stat. Phys. **98**, 31 (2000).
G.M. Cicuta, M. Contedini, and L. Molinari, J. Stat. Phys. **98**, 685 (2000).
D. Gross and A. Zee, unpublished.
M. Bauer, D. Bernard, and J.M. Luck, J. Phys. A **34**, 2659 (2001).
C. Sire and P.L. Krapivsky, J. Phys. A **34** 9065 (2001).
F. J. Dyson, Phys. Rev. **92**, 1331 (1953).
H. Schmidt, Phys. Rev. **105**, 425 (1957).
V.I. Oseledec, Trans. Moscow Math. Soc. **19**, 197 (1968).
R. Bruinsma and G. Aeppli, Phys. Rev. Lett. **50**, 1494 (1983).
J.M. Normand, M.L. Mehta, and H. Orland, J. Phys. A **18**, 621 (1985).
J.M. Luck, “Systèmes Désordonnés Unidimensionnels (Paris: Collection Aléa-Saclay, 1992).
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
Marijan Ribarič and Luka Šušteršič[^1]\
Jožef Stefan Institute, p.p.3000, 1001 Ljubljana, Slovenia
title: Search for an equation of motion of a classical pointlike charge
---
0.5 in PACS numbers: 03.20.+i; 41.70.+t
Keywords: Electrodynamics; Point charge; Equation of motion 0.5 in In 1892 H.A. Lorentz started the search for a classical equation of motion for pointlike charged particles that takes into account the radiation reaction force. This search culminated in the Lorentz-Abraham-Dirac equation of motion, which is not satisfactory since it exhibits self-acceleration causing runaway solutions. In spite of ongoing efforts for more than a century, there is yet no acceptable classical equation of motion for a pointlike charge, cf. the recent paper by Rohrlich [@Rohrlich1] and the comments about his proposal [@mi001; @Rohrlich2; @Zhidkov]. So, it is still an open question how to augment continuous classical electrodynamics with the physical concept of pointlike charged particles. The pointlike charge is presently only a common and handy computational device, which we generalized by the expansions in terms of co-moving moments of time-dependent, moving charges and currents [@mi003].
This open question suggests that our present understanding of classical electrodynamics is actually not as complete as is usually taken for granted. There are actually many serious conceptual flaws and crack in our understanding of the consequences of electromagnetic forces between electric charges and currents.
To faciliate the search for an equation of motion for a pointlike charge, in our monograph [@miknjiga] we:
- clarified the issues by concisely formulating this open question,
- gave a systematic collection of conditions which prospective equations have to meet, and of physical properties they are desired to possess, and
- gathered and collated various theoretical concepts and results which might be *useful and relevant to anyone interested in discussing and proposing answers to this open question of theoretical physics.*
To this end we considered about 80 references from 1903 to 1989 where some of the ideas, concepts, conclusions or formulae analogous or relevant to those considered in our monograph may be found.
To obtain an equation of motion, generalizing Schott [@Schott] we introduced the acceleration four-momentum ${{\bf B}}$ implicitely defined by the relativistic differential energy-momentum balance equation $${\dot{{\bf \beta}}}- ({\dot{{\bf \beta}}}\cdot{\dot{{\bf \beta}}}){{\bf \beta}}+ \tau_0\gamma d{{\bf B}}/dt = {\bf f} \,,
\label{spenacba}$$ where: (i) ${{\bf \beta}}= (\gamma,\gamma{\bf v}/c)$, with $\gamma = (1 - |{\bf v}/c|)^{-1/2}$, is the four-velocity of the pointlike charge; (ii) ${\dot{{\bf \beta}}}= \tau_0\gamma d{{\bf \beta}}/dt$, is the acceleration four-vector, with $\tau_0 = q^2/6\pi\epsilon_0 mc^3$, $q$ and $m$ being the pointlike charge charge and mass, respectively; (iii) ${\bf f} = (\tau_0/mc)\gamma({\bf v}\cdot {\bf F}/c, {\bf F})$ is the external four-force, with ${\bf F}$ being the external force acting on the pointlike charge.
If the acceleration four-momentum ${{\bf B}}$ is explicitly known and depends solely on the external four-force ${\bf f}$ and four-momentum ${{\bf \beta}}$ but not on its derivatives, relation (\[spenacba\]) could be used as a Newtonian equation of motion; and for ${{\bf B}}= -{\dot{{\bf \beta}}}$, it is just the Lorentz-Abraham-Dirac equation. We pointed out sixteen qualitative properties that a general equation of motion for pointlike charged particles ought to possess [@miknjiga; @mi001]; in particular, *it has to have more than two free parameters, with unforeseeable mathematical implications.*
Furthermore, we proposed a conceptually new mathematical model—differential relations describing the asymptotic behaviour of trajectories of classical pointlike charged particles in response to a small and slowly changing external force $f_{ext} = \omega F(\omega t)$, $\omega > 0$ [@mi002; @miknjiga]. They take account of radiation, and of Dirac’s and Bhabha’s conditions about conservation of energy and linear momentum, and of angular and boost momenta. We can use them for describing, investigating and evaluating the dynamic behaviour of classical pointlike charged particles in response to a small and slowly changing external force to any desired order of $\omega$. But we can not use these differential relations as equations of motion for computing their trajectories; though we may transform each of them into an appropriate Newtonian equation of motion accurate up to the same order of $\omega$. We proposed that the Lorentz-Abraham-Dirac equation is not an equation of motion for computing trajectories, but just a differential relation of order $\omega^2$, the lowest order that still takes account of radiation.
[99]{}
F. Rohrlich, Phys. Lett. A **283 (2001) 276.**
M. Ribarič and L. Šušteršič, Phys. Lett A **295 (2002) 318.**
F. Rohrlich, Phys. Lett. A **295 (2002) 320.**
A. Zhidkov, J. Koga, A. Sasaki and M. Uesaka, Phys. Rev. Lett. **88 (2002) 185002; G. Compagno and F. Persico, J. Phys. A-Math. Gen. **35 (2002) 3629; W.E. Baylis and J. Huschilt, Phys. Lett. A **301 (2002) 7; F. Rohrlich, Phys. Lett. A **303 (2002) 307; R. Rivera and D. Villarroel, Phys. Rev. E **66 (2002) 046618; D. Villarroel, Phys. Rev. E **66 (2002) 046624; R.F. O’Connell, Phys. Lett. A **313 (2003) 491; T. Petrosky, G. Ordonez and I. Progogine, Phys. Rev. A **68 (2003) 022107; D. Vogt and P.S. Letelier, Gen. Relativ. Gravit. **35 (2003) 2261; J.A.E. Roa-Neri and J.L. Jimenez, Found. Phys. **34 (2004) 547; V.I. Berezhiani, R.D. Hazeltine and S.M. Mahajan, Phys. Rev. E **69 (2004) 056406; R.D. Hazeltine and S.M. Mahajan, Phys. Rev. E **70 (2004) 036404; R.D. Hazeltine and S.M. Mahajan, Phys. Rev. E **70 (2004) 046407; J. Koga, Phys. Rev. E **70 (2004) 046502; V.V. Lidsky, Theor. Math. Phys. **143 (2005) 583.******************************
M. Ribarič and L. Šušteršič, SIAM J. Appl. Math. **55 (1995) 593.**
M. Ribarič and L. Šušteršič, *Conservation Laws and Open Questions of Classical Electrodynamics, World Scientific, Singapore 1990, Chapters 9–11.*
G.A. Schott, Ann. Phys. (Leipzig) **25 (1908) 63.**
M. Ribarič and L. Šušteršič, Phys. Lett. A**139 (1989) 5.**
[^1]: Corresponding author. Phone +386 1 477 3258; fax +386 1 423 1569; electronic address: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This work presents an in-depth analysis of the majority of the deep neural networks (DNNs) proposed in the state of the art for image recognition. For each DNN multiple performance indices are observed, such as recognition accuracy, model complexity, computational complexity, memory usage, and inference time. The behavior of such performance indices and some combinations of them are analyzed and discussed. To measure the indices we experiment the use of DNNs on two different computer architectures, a workstation equipped with a NVIDIA Titan X Pascal and an embedded system based on a NVIDIA Jetson TX1 board. This experimentation allows a direct comparison between DNNs running on machines with very different computational capacity. This study is useful for researchers to have a complete view of what solutions have been explored so far and in which research directions are worth exploring in the future; and for practitioners to select the DNN architecture(s) that better fit the resource constraints of practical deployments and applications. To complete this work, all the DNNs, as well as the software used for the analysis, are available online.'
address:
- 'University of Milano-Bicocca, Department of Informatics, Systems and Communication, viale Sarca, 336, 20126 Milano, Italy\'
- 'Sorbonne Université, CNRS, LIP6, F-75005 Paris, France '
author:
- ', , .'
bibliography:
- 'references.bib'
title: |
Benchmark Analysis of Representative\
Deep Neural Network Architectures
---
Deep neural networks, Convolutional neural networks, Image recognition.
=-15pt
Introduction {#sec:introduction}
============
Deep neural networks (DNNs) have achieved remarkable results in many computer vision tasks [@lecun2015deep]. AlexNet [@krizhevsky2012imagenet], that is the first DNN presented in the literature in 2012, drastically increased the recognition accuracy (about 10% higher) with respect to traditional methods on the 1000-class ImageNet Large-Scale Visual Recognition Competition (ImageNet-1k) [@russakovsky2015imagenet]. Since then, literature has worked both in designing more accurate networks as well as in designing more efficient networks from a computational-cost point of view.
Although there is a lot of literature discussing new architectures from the point of view of the layers composition and recognition performance, there are few papers that analyze the aspects related to the computational cost (memory usage, inference time, etc.), and more importantly how computational cost impacts on the recognition accuracy.
Canziani *et al.* [@canziani2016analysis] in the first half of 2016 proposed a comprehensive analysis of some DNN architectures by performing experiments on an embedded system based on a NVIDIA Jetson TX1 board. They measured accuracy, power consumption, memory footprint, number of parameters and operations count, and more importantly they analyzed the relationship between these performance indices. It is a valuable work, but it has been focused on a limited number (i.e. 14) of DNNs and more importantly the experimentation has been carried out only on the NVIDIA Jetson TX1 board. In [@huang2017speed], speed/accuracy trade-off of modern DNN-based detection systems has been explored by re-implementing a set of meta-architectures inspired by well-known detection networks in the state of the art. Results include performance comparisons between an Intel Xeon CPU and a NVIDIA Titan X GPU.
The aim of this work is to provide a more comprehensive and complete analysis of existing DNNs for image recognition and most importantly to provide an analysis on two hardware platforms with a very different computational capacity: a workstation equipped with a NVIDIA Titan X Pascal (often referred to as Titan Xp) and an embedded system based on a NVIDIA Jetson TX1. To this aim we analyze and compare more than 40 state-of-the-art DNN architectures in terms of computational cost and accuracy. In particular we experiment the selected DNN architectures on the ImageNet-1k challenge and we measure: accuracy rate, model complexity, memory usage, computational complexity, and inference time. Further, we analyze relationships between these performance indices that provide insights for: 1) understanding what solutions have been explored so far and in what direction it would be appropriate to go in the future; 2) selecting the DNN architecture that better fits the resource constraints of practical deployments and applications.
The most important findings are that: i) the recognition accuracy does not increase as the number of operations increases; ii) there is not a linear relationship between model complexity and accuracy; iii) the desired throughput places an upper bound to the achievable accuracy; iv) not all the DNN models use their parameters with the same level of efficiency; v) almost all models are capable of super real-time performance on a high-end GPU, while just some of them can guarantee it on an embedded system; vi) even DNNs with a very low level model complexity have a minimum GPU memory footprint of about 0.6GB.
The rest of paper is organized as follows: in Section \[sec:benchmarking\] hardware and software used for experiments are detailed; in Section \[sec:architectures\] the considered DNN architectures are briefly introduced; in Section \[sec:metrics\] the measured performance indices are described; finally, in Section \[sec:results\] obtained results are reported and analyzed, and Section \[sec:discussion\] presents our final considerations.
Benchmarking {#sec:benchmarking}
============
We implement the benchmark framework for DNNs comparison in Python. The PyTorch package [@paszke2017automatic] is used for neural networks processing with cuDNN-v5.1 and CUDA-v9.0 as back-end. All the code for the estimation of the adopted performance indices, as well as all the considered DNN models are made publicly available [@github_celona]. We run all the experiments on a workstation and on an embedded system:
1. The workstation is equipped with an Intel Core I7-7700 CPU @ 3.60GHZ, 16GB DDR4 RAM 2400 MHz, NVIDIA Titan X Pascal GPU with 3840 CUDA cores (top-of-the-line consumer GPU). The operating system is Ubuntu 16.04.
2. The embedded system is a NVIDIA Jetson TX1 board with 64-bit ARM^^A57 CPU @ 2GHz, 4GB LPDDR4 1600MHz, NVIDIA Maxwell GPU with 256 CUDA cores. The board includes the JetPack-2.3 SDK.
The use of these two different systems allows to highlight how critical the computational resources can be depending on the DNN model adopted especially in terms of memory usage and inference time.
Architectures {#sec:architectures}
=============
In this section we briefly describe the analyzed architectures. We select different architectures, some of which have been designed to be more performing in terms of effectiveness, while others have been designed to be more efficient and therefore more suitable for embedded vision applications. In some cases there is a number following the name of the architecture. Such a number depicts the number of layers that contains parameters to be learned (i.e. convolutional or fully connected layers).
We consider the following architectures: AlexNet [@krizhevsky2012imagenet]; the family of VGG architectures [@simonyan2014very] (VGG-11, -13, -16, and -19) without and with the use of Batch Normalization (BN) layers [@ioffe2015batch]; BN-Inception [@ioffe2015batch]; GoogLeNet [@Szegedy_2015_CVPR]; SqueezeNet-v1.0 and -v1.1 [@iandola2016squeezenet]; ResNet-18, -34, -50, -101, and -152 [@he2016deep]; Inception-v3 [@szegedy2016rethinking]; Inception-v4 and Inception-ResNet-v2 [@szegedy2016inc]; DenseNet-121, -169, and -201 with growth rate corresponding to 32, and DenseNet-161 with growth rate equal to 48 [@huang2017densely]; ResNeXt-101 (32x4d), and ResNeXt-101 (64x4d), where the numbers inside the brackets denote respectively the number of groups per convolutional layer and the bottleneck width [@xie2017aggregated]; Xception [@Chollet_2017_CVPR]; DualPathNet-68, -98, and -131, [@chen2017dual]; SE-ResNet-50, SENet-154, SE-ResNet-101, SE-ResNet-152, SE-ResNeXt-50 (32x4d), SE-ResNeXt-101 (32x4d) [@Hu_2018_CVPR]; NASNet-A-Large, and NASNet-A-Mobile, whose architecture is directly learned [@Zoph_2018_CVPR].
Furthermore, we also consider the following efficientcy-oriented models: MobileNet-v1 [@howard2017mobilenets], MobileNet-v2 [@sandler2018mobilenetv2], and ShuffleNet [@Zhang_2018_CVPR].
Performance indices {#sec:metrics}
===================
In order to perform a direct and fair comparison, we exactly reproduce the same sampling policies: we directly collect models trained using the PyTorch framework [@paszke2017automatic], or we collect models trained with other deep learning frameworks and then we convert them in PyTorch.
All the pre-trained models expect input images normalized in the same way, i.e. mini-batches of RGB images with shape $3 \times H \times W$, where $H$ and $W$ are expected to be:
- 331 pixels for the NASNet-A-Large model;
- 229 pixels for InceptionResNet-v2, Inception-v3, Inception-v4, and Xception models;
- 224 pixels for all the other models considered.
We consider multiple performance indices useful for a comprehensive benchmark of DNN models. Specifically, we measure: accuracy rate, model complexity, memory usage, computational complexity, and inference time.
Accuracy rate {#sec:accuracy}
-------------
We estimate Top-1 and Top-5 accuracy on the ImageNet-1k validation set for image classification task. The predictions are computed by evaluating the central crop only. Slightly better performances can be achieved by considering the average prediction coming from multiple crops (four corners plus central crop and their horizontal flips).
Model complexity {#sec:model_complexity}
----------------
We analyze model complexity by counting the total amount of learnable parameters. Specifically, we collect the size of the parameter file in terms of MB for the considered models. This information is very useful for understanding the minimum amount of GPU memory required for each model.
Memory usage {#sec:memory}
------------
We evaluate the total memory consumption, which includes all the memory that is allocated, i.e. the memory allocated for the network model and the memory required while processing the batch. We measure memory usage for different batch sizes: 1, 2, 4, 8, 16, 32, and 64.
Computational complexity {#sec:computational_complexity}
------------------------
We measure the computational cost of each DNN model considered using the floating-point operations (FLOPs) in the number of multiply-adds as in [@xie2017aggregated]. More in detail, the multiply-adds are counted as two FLOPs because, in many recent models, convolutions are bias-free and it makes sense to count multiply and add as separate FLOPs.
Inference time {#sec:inference-time}
--------------
We report inference time per image for each DNN model for both the NVIDIA Titan X Pascal GPU and the NVIDIA Jetson TX1. We measure inference time in terms of milliseconds and by considering the same batch sizes described in Section \[sec:memory\]. For statistical validation the reported time corresponds to the average over 10 runs.
Results {#sec:results}
=======
Accuracy-rate vs Computational Complexity vs Model Complexity
-------------------------------------------------------------
The ball charts reported in Figures \[fig:accuracyvsflops\] (a) and (b) show Top-1 and Top-5 accuracy on the ImageNet-1k validation set with respect to the computational complexity of the considered architectures for a single forward pass measured for both the workstation and the embedded board, The ball size corresponds to the model complexity. From the plots it can be seen that the DNN model reaching the highest Top-1 and Top-5 accuracy is the NASNet-A-Large that is also the one having the highest computational complexity. Among the models having the lowest computational complexity instead (i.e. lower than 5 G-FLOPs), SE-ResNeXt-50 (32x4d) is the one reaching the highest Top-1 and Top-5 accuracy showing at the same time a low level of model complexity, with approximately 2.76 M-params. Overall, it seems that there is no relationship between computational complexity and recognition accuracy, for instance SENet-154 needs about $3\times$ the number of operations that are needed by SE-ResNeXt-101(32x4d) while having almost the same accuracy. Moreover, it seems that there is no relationship also between model complexity and recognition accuracy: for instance VGG-13 has a much higher level of model complexity (size of the ball) than ResNet-18 while having almost the same accuracy.
Accuracy-rate vs Learning Power
-------------------------------
It is known that DNNs are inefficient in the use of their full learning power (measured as the number of parameters with respect to the degrees of freedom). Although many papers exist that exploit this feature to produce compressed DNN models with the same accuracy of the original models [@han2015deep] we want here to measure how efficiently each model uses its parameters. We follow [@canziani2016analysis] and measure it as Top-1 accuracy density, i.e. Top-1 accuracy divided by the number of parameters. The higher is this value and the higher is the efficiency. The plot is reported in Figure \[fig:accuracydensity\](a), where it can be seen that the models that use their parameters most efficiently are the SqueezeNets, ShuffleNet, the MobileNets and NASNet-A-Mobile. To focus to the density information, we plot the Top-1 accuracy with respect to the Top-1 accuracy density (see Figure \[fig:accuracydensity\](b)), that permits to find more easily the desired trade-off. In this way it is possible to easily see that among the most efficient models, NASNet-A-Mobile and MobileNet-v2 are the two providing a much higher Top-1 accuracy. Among the models having the highest Top-1 accuracy (i.e. higher than 80%) we can observe how the models using their parameters more efficiently are Inception-v4 and SE-ResNeXt-101 (32x4d).
Inference time {#inference-time}
--------------
Average per image inference time over 10 runs for all the DNN models considered are reported in Tables \[table:inferencetime\](a) and (b) for batch size equal to 1, 2, 4, 8, 16, 32, and 64 on both the Titan Xp and the Jetson. Inference time is measured in milliseconds and the entries in Tables \[table:inferencetime\](a) and (b) are color coded to easily convert them in frames per second (FPS). From the table it is possible to see that all the DNN models considered are able to achieve super real-time performances on the Titan Xp with the only exception of SENet-154, when a batch size of 1 is considered. On the Jetson instead, only a few models are able to achieve super real-time performances when a batch size of 1 is considered, namely: the SqueezeNets, the MobileNets, ResNet-18, GoogLeNet, and AlexNet. Missing measurements are due to the lack of enough system memory required to process the larger batches.
Accuracy-rate vs Inference time
-------------------------------
In Figure \[fig:accuracyvsfps\](a) and (b) we report the plots of the top-1 accuracy with respect to the number of images processed per second (i.e. the number of inferences per second) with a batch size of 1 on both the Titan Xp and the Jetson TX1. On each plot the linear upper bound is also reported; the two have almost the same intercept ($\approx83.3$ for Titan Xp and $\approx83.0$ for the Jetson TX1), but the first has a slope that is almost $8.3\times$ smaller than the second one ($-0.0244$ *vs.* $-0.2025$); these bounds show that the Titan Xp guarantees a lower decay of the maximum accuracy achievable when a larger throughput is needed. Note that this bound appears a curve instead of line in the plots because of the logarithmic scale of the images per second axis. From the Titan Xp plot it is possible to see that if one targets a throughput of more than 250 FPS, the model giving the highest accuracy is ResNet-34, with 73.27% Top-1 accuracy; with a target of more than 125 FPS the model giving the highest accuracy is Xception, with 78,79% Top-1 accuracy; with a target of more than 62.5 FPS the model giving the highest accuracy is SE-ResNeXt-50 (32x4d), with 79,11% Top-1 accuracy; with a target of more than 30 FPS the model giving the highest accuracy is NASNet-A-Large, with 82,50% Top-1 accuracy. This analysis shows how even the most accurate model in the state of the art, i.e. NASNet-A-Large, is able to provide super real-time performance (30.96 FPS) on the Titan Xp. Considering the Jetson TX1 plot it is possible to see that if one targets super real-time performance, the model giving the highest accuracy is MobileNet-v2, with a Top-1 accuracy of 71.81%; if one targets a Top-1 accuracy larger than 75%, the maximum throughput is achieved by ResNet-50 (18,83 FPS); targeting a Top-1 accuracy larger than 80%, the maximum throughput is achieved by SE-ResNeXt-101 (32x4d) (7,16 FPS); targeting the highest Top-1 accuracy in the state of the art the throughput achievable is 2,29 FPS.
[cc]{} &\
&\
(a)&(b)\
\[table:inferencetime\]
Memory usage {#memory-usage}
------------
In Table \[tab:memory\] we analyze the memory consumption for all the DNN models considered for different batch sizes on the Titan Xp. From the memory footprints reported it can be seen that when a batch size of 1 is considered, most models require less than 1GB of memory, with the exception of NASNet-A-Large, the SE-ResNets, the SE-ResNeXTs, SENet-154, the VGGs and Xception. However none of them requires more than 1.5GB for a batch size of 1.
Memory usage vs Model Complexity
--------------------------------
In Figure \[fig:linee\] we analyze the relationship between the initial static allocation of the model parameters (i.e. the model complexity) and the total memory utilization for a batch size of 1 on the Titan Xp. We can see that the relationship is linear, and follows two different lines with approximately the same slope (i.e. 1.10 and 1.15) and different intercepts (i.e. 910 and 639 respectively). This means that the model complexity can be used to reliably estimate the total memory utilization. From the plots we can observe that families of models belong to the same line, as for example the VGGs, the SE-ResNets ans SqueezeNets lie on the line with higher intercept, while the ResNets, DualPathNets, DenseNets, Inception nets and MobileNets line on the line with the lower intercept. In particular we can observe how models having the smallest complexity (i.e. SqueezeNet-v1.0 and SqueezeNet-v1.1 both with 5MB) have a 943MB and 921MB memory footprint, while models having slightly higher complexity (i.e. MobileNet-v1 and MobileNet-v2 with respectively 17MB and 14MB) have a much smaller memory footprint, equal to 650MB and 648MB respectively.
Best DNN at given constraints
-----------------------------
Table \[tab:constraints\] shows the best DNN architectures in terms of recognition accuracy when specific hardware resources are given as computational constraints. This analysis is done for both the Titan Xp and Jetson TX1. We define the following constraints:
- Memory usage: high ($\leq$1.4GB), medium ($\leq$1.0GB) and low ($\leq$0.7GB);
- Computational time: half real-time (@15FPS), real-time (@30FPS), super real-time (@60FPS);
A Titan Xp, with a low memory usage as constraint, achieves a recognition accuracy of at most 75.95% by using the DPN-68 network independently of the computational time. Having more resources, for instance medium and high memory usage, Titan Xp achieves a recognition accuracy of at most 79.11% by using the SE-ResNeXt-50 (32x4d) with a super real-time throughput. Having no requirements in terms of memory usage, the Jetson TX1 achieves a recognition accuracy of at most 69.52% by using the MobileNet-v1, which guarantees a super real-time throughput. To have a DNN running on a Jetson that is comparable in terms of recognition accuracy to the best DNNs running on the Titan Xp, a memory size of at least 1GB is needed. In this case the most performing is the ResNet-50, able to guarantee an half real-time throughput, with a recognition accuracy of 76.01%.
[c]{}\
\
Conclusion {#sec:discussion}
==========
The design of Deep neural networks (DNNs) with increasing complexity able to improve the performance of the ImageNet-1k competition plays a central rule in advancing the state-of-the-art also on other vision tasks. In this article we present a comparison between different DNNs in order to provide an immediate and comprehensive tool for guiding in the selection of the appropriate architecture responding to resource constraints in practical deployments and applications. Specifically, we analyze more than 40 state-of-the-art DNN architectures trained on ImageNet-1k in terms of accuracy, number of parameters, memory usage, computational complexity, and inference time.
The key findings of this paper are the following:
- the recognition accuracy does not increase as the number of operations increases: in fact, there are some architectures that with a relatively low number of operations, such as the SE-ResNeXt-50 (32x4d), achieve very high accuracy (see Figures \[fig:accuracyvsflops\]a and b). This finding is independent on the computer architecture experimented;
- there is not a linear relationship between model complexity and accuracy (see Figures \[fig:accuracyvsflops\]a and b);
- not all the DNN models use their parameters with the same level of efficiency (see Figures \[fig:accuracydensity\]a and b);
- the desired throughput (expressed for example as the number of inferences per second) places an upper bound to the achievable accuracy (see Figures \[fig:accuracyvsfps\]a and b);
- model complexity can be used to reliably estimate the total memory utilization (see Figure \[fig:linee\]);
- almost all models are capable of real-time or super real-time performance on a high-end GPU, while just a few of them can guarantee them on an embedded system (see Tables \[tab:inference\]a and b);
- even DNNs with a very low level model complexity have a minimum GPU memory footprint of about 0.6GB (see Table \[tab:memory\]).
All the DNNs considered, as well as the software used for the analysis, are available online [@github_celona]. We plan to add to the repository interactive plots that allow other researchers to better explore the results of this study, and more importantly to effortlessly add new entries.
Acknowledgments {#acknowledgments .unnumbered}
===============
We gratefully acknowledge the support of NVIDIA Corporation with the donation of the Titan X Pascal GPU used for this research.
[Simone Bianco]{} is assistant professor of Computer Science at Department of Informatics, Systems and Communication of the University of Milano-Bicocca, Italy. He obtained the PhD in Computer Science from the University of Milano-Bicocca, in 2010. He obtained the BSc and the MSc degree in Mathematics from the University of Milano-Bicocca, respectively in 2003 and 2006. His current research interests include computer vision, machine learning, optimization algorithms, and color imaging.
[Remi Cadene]{} is a PhD student and teaching assistant at LIP6 (Computer Science laboratory) of Sorbonne University, France. In 2016, he received a MSc degree in Computer Science at Sorbonne University. His primary research interests are in the fields of Machine Learning, Computer Vision and Natural Language Processing. He is currently focusing on Neural Networks, Multimodal Learning and Weakly Supervised Learning.
[Luigi Celona]{} is currently a postdoctoral fellow at DISCo (Department of Informatics, Systems and Communication) of the University of Milano-Bicocca, Italy. In 2018 and 2014, he obtained respectively the PhD and the MSc degree in Computer Science at DISCo. In 2011, he obtained the BSc degree in Computer Science from the University of Messina. His current research interests focus on image analysis and classification, machine learning and face analysis.
[Paolo Napoletano]{} is assistant professor of Computer Science at Department of Informatics, Systems and Communication of the University of Milano-Bicocca. In 2007, he received a PhD in Information Engineering from the University of Salerno. In 2003, he received a Master’s degree in Telecommunications Engineering from the University of Naples Federico II. His current research interests include machine learning for multi-modal data classification and understanding.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This work evaluates the model dependence of the electric and Coulomb quadrupole amplitudes (E2, C2) in the predominantly M1 (magnetic dipole-quark spin flip) $\gamma^* N \rightarrow \Delta$ transition. Both the model-to-model dependence and the intrinsic model uncertainties are evaluated and found to be comparable to each other and no larger than the experimental errors. It is confirmed that the quadrupole amplitudes have been accurately measured indicating significant non-zero angular momentum components in the proton and $\Delta$.'
author:
- 'S. Stave'
- 'A. M. Bernstein'
- 'I. Nakagawa'
title: 'Accuracy of Extracted Multipoles from $\gamma^*N\rightarrow\Delta$ Data'
---
[ address=[[Department of Physics, Laboratory for Nuclear Science, Massachusetts Institute of Technology, Cambridge, Massachusetts 02139, USA]{}]{},altaddress=[[Current address: Triangle Universities Nuclear Laboratory/Duke University, Durham, North Carolina 27708, USA]{}]{} ]{}
[ address=[[Department of Physics, Laboratory for Nuclear Science, Massachusetts Institute of Technology, Cambridge, Massachusetts 02139, USA]{}]{} ]{}
[ address=[[Department of Physics, Laboratory for Nuclear Science, Massachusetts Institute of Technology, Cambridge, Massachusetts 02139, USA]{}]{}, address=[[Department of Physics and Astronomy, University of Kentucky, Lexington, Kentucky 40206 USA]{}]{}, ,altaddress=[[Current address: Radiation Laboratory, RIKEN, 2-1 Hirosawa, Wako, Saitama 351-0198, Japan]{}]{} ]{}
Physics Motivation
==================
Experimental confirmation of the presence of non-spherical hadron amplitudes (i.e. d states in quark models or p wave $\pi$-N states) is fundamental and has been the subject of intense experimental and theoretical interest (for reviews see [@nstar2001; @cnp; @amb]). This effort has focused on the measurement of the electric and Coulomb quadrupole amplitudes (E2, C2) in the predominantly M1 (magnetic dipole-quark spin flip) $\gamma^* N \rightarrow \Delta$ transition. Since the proton has spin 1/2, no quadrupole moment can be measured. However, the $\Delta$ has spin 3/2 so the $\gamma^*N\rightarrow
\Delta$ reaction can be studied for quadrupole amplitudes in the nucleon and $\Delta$. Due to spin and parity conservation in the $\gamma^*N(J^\pi=1/2^+) \rightarrow \Delta(J^\pi=3/2^+)$ reaction, only three multipoles can contribute to the transition: the magnetic dipole ($M1$), the electric quadrupole ($E2$), and the Coulomb quadrupole ($C2$) photon absorption multipoles. The corresponding resonant pion production multipoles are $M_{1+}^{3/2}$, $E_{1+}^{3/2}$, and $S_{1+}^{3/2}$. The relative quadrupole to dipole ratios are EMR=Re($E_{1+}^{3/2}/M_{1+}^{3/2}$) and CMR=Re($S_{1+}^{3/2}/M_{1+}^{3/2}$). In the quark model, the non-spherical amplitudes in the nucleon and $\Delta$ are caused by the non-central, tensor interaction between quarks [@glashow]. However, the magnitudes of this effect for the predicted E2 and C2 amplitudes[@capstick_karl] are at least an order of magnitude too small to explain the experimental results and even the dominant M1 matrix element is $\simeq$ 30% low [@amb; @capstick_karl]. A likely cause of these dynamical shortcomings is that the quark model does not respect chiral symmetry, whose spontaneous breaking leads to strong emission of virtual pions (Nambu-Goldstone Bosons)[@amb]. These couple to nucleons as $\vec{\sigma}\cdot \vec{p}$ where $\vec{\sigma}$ is the nucleon spin, and $\vec{p}$ is the pion momentum. The coupling is strong in the p wave and mixes in non-zero angular momentum components.
However, the multipoles are not observables and must be extracted from the measured cross sections. The five-fold differential cross section for the $p(\vec{e},e'p)\pi^{0}$ reaction is written as five two-fold differential cross sections with an explicit $\phi^*$ dependence as [@drechsel_tiator]
$$\frac{d^5\sigma}{d\Omega_f dE_f d\Omega} = \Gamma (\sigma_T + \epsilon
\sigma_L + v_{LT}\sigma_{LT} \cos\phi^*
+ \epsilon \sigma_{TT} \cos 2\phi^*
+ h p_{e} v_{LT'}\sigma_{LT'} \sin \phi^*)$$
where $\epsilon$ is the transverse polarization of the virtual photon, $v_{LT}=\sqrt{2\epsilon(1+\epsilon)}$, $v_{LT'}=\sqrt{2\epsilon(1-\epsilon)}$, $\Gamma$ is the virtual photon flux, $\phi^*$ is the pion center of mass azimuthal angle with respect to the electron scattering plane, $h$ is the electron helicity, and $p_{e}$ is the magnitude of the electron longitudinal polarization. The virtual photon differential cross sections ($\sigma_{T},\sigma_{L},\sigma_{LT},\sigma_{TT},\sigma_{LT'}$) are all functions of the center of mass energy $W$, the four momentum transfer squared $Q^2$, and the pion center of mass polar angle $\theta_{\pi q}^{*}$ (measured from the momentum transfer direction). They are bilinear combinations of the multipoles [@drechsel_tiator].
Resonant Multipole Fitting
==========================
The current experiments [@beck; @blanpied; @warren; @mertz; @kunz; @sparveris; @joo; @frolov; @pospischil; @bartsch; @elsner; @stave] do not have sufficient polarization data to perform a model independent multipole analysis and must rely upon models for the non-resonant (background) amplitudes. The standard procedure to extract the multipoles is to use the models to fit the data. Their background terms are unaltered and the three isospin =3/2 resonance multipoles $$R_{i}^{3/2} = M_{1+}^{3/2}, E_{1+}^{3/2}, S_{1+}^{3/2}$$ are fit to the data. Specifically we introduced multiplicative factors, $\lambda(R_{i})$ for the $I=3/2$ multipoles so that the phase, and hence unitarity, is preserved. For the charge channels with a proton target and an outgoing neutral pion (e.g. $\gamma +p \rightarrow \pi^{0} p$): $$R_{i}(\pi^0 p) = R_{i}^{1/2} +\frac{2}{3} \lambda(R_{i}) R_{i}^{3/2}$$ where $R_{i}$ represents any of the three photo-pion multipoles $M_{1+}, E_{1+}, S_{1+}$ for the final charge state or in the isospin 1/2 and 3/2 states.
Three parameter, resonant multipole fits were performed on data taken at $Q^2=0.060$ [@stave] and $Q^2=0.126$ (GeV/c)$^2$ [@warren; @mertz; @kunz; @sparveris; @stave_thesis], one $Q^2$ value at a time, using four representative calculations: the phenomenological MAID 2003 [@maid1] and SAID[@said]models, and the dynamical models of Sato-Lee [@sato_lee] and DMT [@dmt]. The fits are presented in terms of $M_{1+}^{3/2}$, EMR = E2/M1 = Re($E_{1+}/M_{1+})$, and CMR = C2/M1 = Re($S_{1+}/M_{1+})$. At least one multipole is expressed in absolute terms rather than as a ratio because some models can give accurate predictions for ratios but not for absolute sizes. Figure \[fig:before\_after\_fit\] shows our new, low $Q^2$ data along with several model predictions before and after the three resonant parameter fitting. The convergence is rather significant. Only the data points in the top three plots were included in the fits and yet the parallel cross section as a function of the center of mass energy $W$ converged nicely. Note that since the Sato-Lee model does not include higher resonances, it was not expected to fit the data well at higher $W$ explaining the deviation observed in Fig. \[fig:before\_after\_fit\]. Also, as expected, the $\sigma_{LT'}$ curves did not converge since this time reversal odd observable [@rask_don] is primarily sensitive to background amplitudes and the fit is only for resonant amplitudes.
The bottom half of Figure \[fig:before\_after\_fit\] shows the “spherical” calculated curves when the resonant quadrupole amplitudes ($E_{1+}^{3/2}$ in $\sigma_{0}$ and $S_{1+}^{3/2}$ in $\sigma_{TL}$, see Appendix for multipole expansions of the observables) are set equal to zero. The difference between the spherical and full curves shows the sensitivity of these cross sections to the quadrupole amplitudes and demonstrates the basis of the measurement of the $E_{1+}$ and $S_{1+}$ multipoles. The small spread in the spherical curves indicates their sensitivity to the model dependence of the background amplitudes. Figure \[fig:EMR\_CMR\] shows the values for the EMR and CMR for the four models before and after fitting. It is seen that there is a very strong convergence of these values after fitting. We have quoted the average value of these parameters as the measured value and are using the RMS deviation to estimate the model-to-model error[@sparveris; @stave] since these four models are sufficiently different to have a reasonable estimate of the present state of model dependence of the multipoles. At the present time the model-to-model and experimental errors are approximately equal.
![\[fig:before\_after\_fit\]$Q^2=0.060$ (GeV/c)$^2$ data with model predictions before fitting (top panels) and after the three resonant parameter fit (bottom panels) along with the EFT predictions from Pascalutsa and Venderhaeghen (PV) [@pasc]. Note the convergence of the models except for the background sensitive $\sigma_{LT'}$ points. Data from [@stave] and include the statistical and systematic errors. The lines with dots on them are the fitted models with the $E_{1+}$ and $S_{1+}$ quadrupole terms set to zero. The models are MAID 2003 [@maid1], SAID [@said], DMT [@dmt] and Sato-Lee (SL) [@sato_lee]. ](Q06_XS_SoH_all.eps "fig:"){height=".4\textheight"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[fig:before\_after\_fit\]$Q^2=0.060$ (GeV/c)$^2$ data with model predictions before fitting (top panels) and after the three resonant parameter fit (bottom panels) along with the EFT predictions from Pascalutsa and Venderhaeghen (PV) [@pasc]. Note the convergence of the models except for the background sensitive $\sigma_{LT'}$ points. Data from [@stave] and include the statistical and systematic errors. The lines with dots on them are the fitted models with the $E_{1+}$ and $S_{1+}$ quadrupole terms set to zero. The models are MAID 2003 [@maid1], SAID [@said], DMT [@dmt] and Sato-Lee (SL) [@sato_lee]. ](Q06_XS_SoH_allfit.eps "fig:"){height=".4\textheight"}
![\[fig:EMR\_CMR\]$Q^2=0.060$ (GeV/c)$^2$ extracted EMR and CMR before and after fitting. The light error band is the average of the fitting errors and the darker band is the RMS deviation of the models added in quadrature. The models are MAID 2003 [@maid1], SAID [@said], DMT [@dmt] and Sato-Lee [@sato_lee] and the effective field theory calculations of Pascalutsa and Vanderhaeghen (PV) [@pasc] and Gail and Hemmert (GH) [@gail_hemmert]. The convergence of the models after fitting is clear.](Mainz_fits_EMR_CMR.eps){height=".4\textheight"}
In a way what we are observing is the fact that the electro-pion production process shows us two separate faces, depending on the observable and on the center of mass energy $W$ that we choose. The best way to extract the three resonant amplitudes is to measure the time reversal even observables ($\sigma_{0}, \sigma_{TT},
\sigma_{LT}$) [@rask_don] at or near the resonance energy $W = 1232$ MeV. On the other hand, the best way to test the model calculations is to examine time reversal odd observables such as $\sigma_{LT'}$ [@rask_don] right on resonance. In addition, we also have off resonance data. These are sensitive to both the shape and phase of the $M_{1+}$ multipole and also the background amplitudes. The $\gamma p \rightarrow \pi^{+}n$ charge channel is also more sensitive to the background amplitudes particularly the $I = 1/2$ amplitudes. Such background sensitive data in combination with model studies are essential if the field is to progress to the stage where the model errors are significantly smaller than the experimental ones.
Intrinsic Model Errors in Determination of the Resonant Multipoles
==================================================================
Beyond Three Parameter Fits: Including Background Multipoles
-------------------------------------------------------------
This work expands the three resonant parameter fits to include the influence of the background multipoles on the resonant amplitudes derived from fitting the experimental data. In this way we will be able to make reasonable estimates of the intrinsic model errors due to uncertainties in the background multipoles and to see if this leads to any suggestions to reduce them. First, we include the remaining s and p wave multipoles: $E_{0+}, L_{0+}$, $M_{1-},L_{1-}$. Next, we estimate the influence of the higher partial waves using the CGLN invariant amplitudes $F_{i}, i=1,\dots,6$ [@drechsel_tiator]. We introduce a new combination of higher order multipoles we call $\overline{F_{i}}$. These combinations show that many small multipoles can have a cumulative effect as will be seen shortly. The ${\overline{F_{i}}}$s are varied using a scaling factor $\lambda_i$ as $$\begin{aligned}
F_i & = & F_i^{S\&P} + \lambda_i{\overline{F_{i}}} \nonumber \\
\hline
F_1 & = & {E_{0+}}+ 3{{M_{1+}}}\cos\theta + 3{{E_{1+}}}\cos\theta + \lambda_1{\overline{F_{1}}} \nonumber \\
F_2 & = & {M_{1-}}+ 2{{M_{1+}}} + \lambda_2{\overline{F_{2}}} \nonumber \\
F_3 & = & 3({{E_{1+}}} - {{M_{1+}}}) + \lambda_3{\overline{F_{3}}} \nonumber \\
F_4 & = & \lambda_4{\overline{F_{4}}} \nonumber \\
F_5 & = & {L_{0+}}+ 6{{L_{1+}}}\cos\theta + \lambda_5{\overline{F_{5}}}\nonumber \\
F_6 & = & {L_{1-}}- 2 {{L_{1+}}} + \lambda_6{\overline{F_{6}}}.\end{aligned}$$
Next, we allow the $I=1/2$ part of the charge channel multipole to vary in a way similar to the $I=3/2$ part. $$R_i(\pi^0p) = \lambda(R_i^{1/2}) R_i^{1/2} + \lambda(R_i) \frac{2}{3}R_i^{3/2}.$$
This new fitting procedure introduces thirteen background amplitudes, too many to determine with the available data. So, we set out to determine how much the resonant parameters are affected by the uncertainties in the background amplitudes. We try to quantify this effect and to determine which parameters have a strong effect and which are correlated.
First, we look for parameters which are correlated with the resonant parameters. Figure \[fig:corr\] shows some examples of negative, positive and no correlation. Fitting parameters are plotted on each axis and the ellipse indicates the region where 68% (1 $\sigma$) of the fits are expected to fall if many similar data sets are fit. The ellipse with axes close to the x and y axes shows no correlation. However, the other ellipses are rotated indicating that as one parameter tends one direction, the other parameter tends to go with it or away from it. This indicates a correlation between the parameters. While error ellipse plots are useful in a qualitative way, they are difficult to use in a quantitative manner. Changes in the scale of the axes will change the angle of the ellipse hiding or exaggerating correlations. Therefore, we use the correlation coefficient, $r$, to indicate the level of correlation between two parameters: $$r=\frac{\sigma_{12}}{\sigma_{11}\sigma_{22}}$$ where the error and curvature matrices $$\left(\begin{array}{cc}
\epsilon_{11} & \epsilon_{12} \\
\epsilon_{21} & \epsilon_{22}
\end{array}\right)^{-1}
=
\left(\begin{array}{cc}
\sigma_{11}^2 & \sigma_{12} \\
\sigma_{21} & \sigma_{22}^2
\end{array}\right)$$ are used [@PDH]. $r$ varies from -1 to 1 and is insensitive to the parameter scales since the scale factor for each parameter cancels in the ratio. Figure \[fig:corr\_coeff\] shows the various correlation coefficients for each background amplitude with $E_{1+}^{3/2},
M_{1+}^{3/2}$, and $L_{1+}^{3/2}$ using MAID 2003 and combined Bates and Mainz data at $Q^2=0.126$ [(GeV/c)$^2$]{}. The square of $r$ indicates how much of the variance is explained by a linear relation between the two variables. A rule of thumb is that two variables are correlated if $r^2 \ge
0.5$ and uncorrelated if $r^2 \le 0.1$. This then leads to the ranges in $r$: $0.7 \le |r| \le 1.0 =$ large correlation, $0.3 \le |r| < 0.7 =$ medium correlation, $0.0 \le |r| < 0.3 =$ small correlation.
![\[fig:corr\]Examples of negative correlation (left), no correlation (center), and positive correlation (right). Note: F2b=${\overline{F_{2}}}$. ](Bates_M2003_E1p_v_F2b.eps "fig:") ![\[fig:corr\]Examples of negative correlation (left), no correlation (center), and positive correlation (right). Note: F2b=${\overline{F_{2}}}$. ](Bates_M2003_L1p_v_E0p.eps "fig:") ![\[fig:corr\]Examples of negative correlation (left), no correlation (center), and positive correlation (right). Note: F2b=${\overline{F_{2}}}$. ](Bates_M2003_E1p_v_M1m.eps "fig:")\
![\[fig:corr\_coeff\]Correlation coefficients for background amplitudes with the resonant amplitudes $E_{1+}^{3/2},M_{1+}^{3/2},L_{1+}^{3/2}$ using MAID 2003 for Mainz and Bates $Q^2=0.126$ [(GeV/c)$^2$ ]{}data. The central region of each plot is for small correlation with the next region being medium correlation and the next being large correlation. Note: F1b=${\overline{F_{1}}}$, F2b=${\overline{F_{2}}}$, etc. ](Bates93_M2003_corr_coeff.eps)
The next check is for sensitivity. If a parameter is large, zeroing it out should affect the $\chi^2$ by a large amount. For example, when $L_{0+}$ is turned off, the model predictions change noticeably (see Fig. \[fig:sens\_L0\]). Other background terms can have significant effects as well like the ${\overline{F_{i}}}$s and the remaining s and p do in Fig. \[fig:sens\_other\]. Table \[table:sens\] shows the $\chi^2$/d.o.f. that results from turning off the various background amplitudes in the MAID 2003 model. It also indicates how strongly the amplitude was correlated with any of the resonant multipoles.
In Figs. \[fig:sens\_L0\] and \[fig:sens\_other\], a combined cross section[@cnp; @mertz; @sparveris] $\sigma_{E2}=\sigma_0(\theta_{\pi
q}^*) + \sigma_{TT}(\theta_{\pi q}^*) -
\sigma_0(\theta_{\pi q}^*=\pi)$ is shown. In this linear combination the dominant $M_{1+}$ multipole contribution cancels out and shows the effect of the smaller $E_{1+}$ quadrupole contribution. (See Appendix for the expansion of the observables in terms of multipoles.)
![\[fig:sens\_L0\]Sensitivity to $L_{0+}$ in MAID 2003 using data at $Q^2=0.126$ [(GeV/c)$^2$]{}. The solid curve is the full MAID 2003 model and the dashed line is with the $L_{0+}$ multipole set to zero. The effect is particularly evident for $\sigma_{LT}$. Data are from [@mertz; @kunz; @sparveris] and include statistical, systematic and model errors. ](Bates_M2003_L0p0.eps)
![\[fig:sens\_other\]Sensitivity to ${\overline{F_{i}}}$ and s and p in MAID 2003 using data at $Q^2=0.126$ [(GeV/c)$^2$]{}. As in Fig. \[fig:sens\_L0\], the solid curve is the full MAID 2003 while the dashed line is now MAID 2003 with all the ${\overline{F_{i}}}$s turned off and the dotted line shows the results using only the resonant multipoles. This shows the significant effect the background has on some observables and that the ${\overline{F_{i}}}$ amplitudes can also have effects of similar size. Data are from [@mertz; @kunz; @sparveris] and include statistical, systematic and model errors. ](Q127_M2003_scan.eps)
Extra Par. $\chi^2$/d.o.f
---------------------- ----------------
$L_{0+}$ [**7.42**]{}
$E_{0+}$ 6.09
${\overline{F_{1}}}$ [**5.60**]{}
$L_{1+}^{1/2}$ [**5.59**]{}
$M_{1-}$ [**4.10**]{}
$E_{1+}^{1/2}$ [**3.31**]{}
${\overline{F_{5}}}$ [**1.85**]{}
$M_{1+}^{1/2}$ 1.85
${\overline{F_{2}}}$ [**1.80**]{}
${\overline{F_{6}}}$ 1.72
$L_{1-}$ 1.55
${\overline{F_{3}}}$ 1.24
— 1.21
${\overline{F_{4}}}$ 1.17
: \[table:sens\]Gives $\chi^2/$d.o.f. resulting from turning off the corresponding background parameter in MAID 2003 arranged from most to least sensitive using the combined Bates and Mainz $Q^2=0.126$ [(GeV/c)$^2$ ]{}data. The type face indicates the level of correlation with any of the three resonant multipoles: [**Large Correlation**]{} Medium or Small Correlation.
Using the criteria of correlation ($|r|>0.7$) and sensitivity ($\chi^2$ increases by 50% upon removal), several parameters were identified as significant using the $Q^2=0.126$ (GeV/c)$^2$ data set. Those multipoles are shown in Table \[table:sens\_filter\] and include two of the s and p multipoles, two of the isospin = 1/2 multipoles and three ${\overline{F_{i}}}$ terms. Looking at Fig. \[fig:res\_pars\] all seven terms shown in Table \[table:sens\_filter\], when varied, lead to shifted central values or larger error bars for the resonant multipoles. Some, like the $E_{1+}^{1/2}$ with $M_{1+}^{3/2}$ are shifted but not outside the error bars and with no increase in the error bar size and so are not considered significant.
Effort was made to search for a set of criteria using the correlation coefficient and the change in $\chi^2$/d.o.f. that would identify all of the parameters which Fig. \[fig:res\_pars\] identifies as significant. The criteria for significance were an increased error and/or a shift in central value. Either indicates a significant effect on the resonant multipole determination. In order to make the criteria robust, other models were put through the selection process as well. In addition to the MAID 2003 model, Sato-Lee, SAID, and DMT were all used. The best identifier of significant parameters turned out to be the single test of $|r|>0.7$. In almost every case, this alone identified all the significant parameters. The $\chi^2$ sensitivity would identify some of the sensitive parameters but not others.
----- ----- ----------------------------------------------------------------------
E1+ vs. $E_{1+}^{1/2}$, $M_{1-}$, ${\overline{F_{1}}}$, ${\overline{F_{2}}}$
L1+ vs. $L_{1+}^{1/2}$, $L_{0+}$, ${\overline{F_{5}}}$
----- ----- ----------------------------------------------------------------------
: \[table:sens\_filter\] Background amplitudes for which the quadrupole amplitudes show sensitivity for MAID 2003 using the criteria listed in the text and the combined Bates and Mainz $Q^2=0.126$ [(GeV/c)$^2$ ]{}data.
Effect of Background on Resonant Amplitudes
===========================================
In order to see the effect the sensitive background amplitudes have on the extracted results, the resonant parameters resulting from each four parameter fit were plotted in Fig. \[fig:res\_pars\]. The horizontal bar indicates the position and error of the three parameter fit. The background amplitudes identified as significant do have an effect on the extracted multipoles relative to the three parameter fit. For each sensitive background parameter, the error increases and in most cases the central values shift. What is also interesting is that the ${\overline{F_{i}}}$s have a significant effect. This indicates that many small amplitudes can combine to have a large effect.
To try to quantify the effect of the various background amplitudes on the resonant amplitudes, the RMS deviation of the various four parameter fits was taken for each model and identified as the intrinsic model error. For some fits, the RMS deviation was small and so the average of the four parameter fitting errors was used instead. In both cases, an estimate of the intrinsic error in each model was obtained. The results are shown in Fig. \[fig:res\_par\_all\] and Table \[table:res\_par\_all\] along with the average and RMS deviation of the three parameter fits (model-to-model error). The figure and table indicate that the model-to-model variation is about the same size as the intrinsic model error (specifically for $M_{1+}$ and the CMR the model-to-model error is larger while for the EMR it is somewhat smaller). However, the new error determination procedure is able to use one model alone instead of comparing it with other models. Each models’s error can be assessed independently of the other models.
![\[fig:res\_pars\]Effect of sensitive parameters on extracted resonance values using MAID 2003 for the combined Mainz and Bates data at $Q^2=0.126$ (GeV/c)$^2$. Note the larger error bars and shifted central positions of certain combinations of multipoles. The horizontal band indicates the value and error resulting from the three resonant parameter fit. Similar plots were made for the remaining models. Their results are shown in Fig. \[fig:res\_par\_all\] and Table \[table:res\_par\_all\]. ](Bates93_M2003_res_fits.eps)
![\[fig:res\_par\_all\]Comparison of the estimated intrinsic and model-to-model errors for all four models at $Q^2=0.126$ (GeV/c)$^2$, $W=1232$ MeV. The intrinsic error was estimated using the RMS deviation across all the four parameter fits. If the deviation was small, then the average of the four parameter fit errors was used. Cases where the RMS deviation was used are indicated by the dashed vertical error lines. The horizontal band indicates the RMS deviation of the three resonant parameter fits using the four different models.](Bates93_res_par_all2.eps)
---------- ------------------------- ------------- ---------------- ------------------ ------ ---------- ------
3 par. avg. Model-to-model Intrinsic errors
error MAID DMT Sato-Lee SAID
$M_{1+}$ \[$10^{-3}/m_{\pi^+}$\] 40.94 0.50 0.20 0.29 0.23 0.22
EMR \[%\] -1.65 0.45 0.47 0.78 0.43 0.27
CMR \[%\] -6.43 0.75 0.12 0.47 0.39 0.15
---------- ------------------------- ------------- ---------------- ------------------ ------ ---------- ------
: \[table:res\_par\_all\]Preliminary fitting results and comparison of the estimated intrinsic and model-to-model errors for all four models for the combined Mainz and Bates data at $Q^2=0.126$ (GeV/c)$^2$, $W=1232$ MeV. See Fig. \[fig:res\_par\_all\] for calculation details.
While looking to improve the fits, an exhausive search was performed of all combinations of the three resonant parameters and any combination of the 13 remaining parameters. No significant improvement was found for either $Q^2=0.060$ or $0.126$ (GeV/c)$^2$.
It is time, then, to look beyond fitting the multipoles. It is possible to modify internal model parameters (form factors, coupling constants) which affect many multipoles simultaneously but in different ways. This may allow the models to fit the data better. However, this fitting most likely needs to be performed by the model authors.
The understanding of the $\Delta$ will also be improved with experiments that are closer to complete. With target and recoil polarization, more observables are accessable and these have different combinations of multipoles. These new combinations will further constrain the models allowing better fits and smaller uncertainties in the backgrounds. Until new data are available, though, fitting the data and improving the models remain the best options.
Summary and Conclusion
======================
Experimental results using the $\gamma^* p\rightarrow p \pi^0$ reaction have advanced the understanding of the shape of the proton and the $\Delta$. However, the analysis process begins with extracting multipoles (which are not observables) from cross sections (which are). Without complete experiments including target and recoil polarization, the extraction must rely upon models for the background amplitudes. Performing standard three resonant parameter fits has allowed a good deal of progress to be made. Near resonance, fits using various models converge at $Q^2=0.060$ and 0.126 (GeV/c)$^2$ despite the differences in the model backgrounds. However, what has not been fully understood is the effect these differing backgrounds can have on the resonant parameters.
To answer that question, we have added thirteen more background amplitudes to our three parameter fits and systematically examined the effect of each one on all three resonant multipoles. Those additional background amplitudes are the four remaining s and p wave amplitudes, three isospin 1/2 amplitudes and six amplitudes we have constructed, the ${\overline{F_{i}}}$s. The large effect of some of the ${\overline{F_{i}}}$ terms shows how small multipoles which may have been ignored separately, can combine to have a sizable effect on the resonant amplitudes.
As part of the systematic examination of the additional background amplitudes, correlations were found between them and the resonant amplitudes which led to larger errors and/or shifts in the values of the extracted resonant multipoles. We also found that while some amplitudes exhibit a large sensitivity in $\chi^2$, no universal criteria could be found which would predict a sensitivity in the resonant amplitudes. Some background amplitudes which were sensitive did not affect the fits while others which were not sensitive did.
However, varying the background amplitudes which were highly correlated with the resonant multipoles did affect the extracted resonant multipoles. Previous works have shown that the experimental and model-to-model errors are similar in size [@stave; @sparveris]. What this exercise has shown is that the intrinsic model error is also similar in size to the model-to-model error. The current data really are challenging the existing models. So, without improvement in the models or more complete experiments, this is as far as the current data can take us.
In general, the models agree with the data in a qualitative way but a good quantitative decription will require further refinement of those models. It is possible that some of the models may be made to fit the data much better with adjustment of the proper parameters. Adjusting a form factor or coupling constant within the model will change many multipoles in ways that are different from how they were varied in this study. Once the models are improved, they can be further tested with experiments that utilize target and recoil polarization. These introduce new combinations of multipoles which will further constrain the models. However, until new data are available, improvement of the models is the only option which will allow a better understanding of the $\Delta$.
Finally we return to the question that has primarily motivated this field: Do we have definitive evidence that the nucleon and $\Delta$ have non-spherical components and if so how large? Based on this study we follow reference [@cnp] and present Figure \[fig:M2003\_bands\] which indicates the final sensitivity to the quadrupole amplitudes. On the right is $\sigma_{LT}$ which is sensitive to the $S_{1+}$ quadrupole term. On the left is a special construction, $\sigma_{E2}=\sigma_0(\theta_{\pi q}^*) +
\sigma_{TT}(\theta_{\pi q}^*) -
\sigma_0(\theta_{\pi q}^*=\pi)$[@cnp; @mertz; @sparveris] which cancels out the dominant $M_{1+}$ multipole contribution and shows the effect of the smaller $E_{1+}$ quadrupole contribution. In Fig. \[fig:M2003\_bands\], the range of predictions using all the four parameter fits was found by cycling through all the fits and storing the maximum and minimum. In this way, a high probability region was identified where the physical multipole would be expected to be. For comparison, the same procedure was repeated but with the quadrupole amplitudes set to zero. On the basis of this study of the uncertainties in the resonance and background amplitudes we agree with the previous conclusion[@cnp] that a significant contribution of quadrupole amplitudes has been observed.
![\[fig:M2003\_bands\]$\sigma_{E2}$ (left) and $\sigma_{LT}$ (right) with error bands showing the range of all the MAID 2003 four parameter fits with (blue/dark region) and without (green/gray region) the quadrupole terms. The data are clearly in the region indicating quadrupole strength similar in size to that found in the models. These plots represent fits using the MAID 2003 model. The light region surrounding the MAID 2003 fit was found using the errors in the three parameter fit with fixed background terms. Data are from [@mertz; @kunz; @sparveris] and include statistical, systematic and model errors. ](Bates93_M2003_SigE2_bands.eps "fig:") ![\[fig:M2003\_bands\]$\sigma_{E2}$ (left) and $\sigma_{LT}$ (right) with error bands showing the range of all the MAID 2003 four parameter fits with (blue/dark region) and without (green/gray region) the quadrupole terms. The data are clearly in the region indicating quadrupole strength similar in size to that found in the models. These plots represent fits using the MAID 2003 model. The light region surrounding the MAID 2003 fit was found using the errors in the three parameter fit with fixed background terms. Data are from [@mertz; @kunz; @sparveris] and include statistical, systematic and model errors. ](Bates93_M2003_SigLT_bands.eps "fig:")
The authors would like to thank L. Tiator, D. Drechsel, T.-S. H. Lee, V. Pascalutsa, M. Vanderhaeghen, T. Gail and T. Hemmert for their assistance with valuable discussions and for sharing their unpublished work. This work is supported at MIT by the U.S. DOE under Grant No. DE-FG02-94ER40818.
[27]{} natexlab\#1[\#1]{}\[1\][“\#1”]{} url \#1[`#1`]{}urlprefix\[2\]\[\][[\#2](#2)]{}
D. Drechsel, and L. Tiator, editors, *Proceedings of the Workshop on the Physics of Excited Nucleons*, World Scientific, 2001.
C. N. Papanicolas, *Eur. Phys. J.* **A18**, 141–145 (2003).
A. M. Bernstein, *Eur. Phys. J.* **A17**, 349–355 (2003).
S. L. Glashow, *Physica* **A96**, 27–30 (1979).
S. Capstick, and G. Karl, *Phys. Rev.* **D41**, 2767 (1990).
D. Drechsel, and L. Tiator, *J. Phys.* **G18**, 449–497 (1992).
R. Beck, et al., *Phys. Rev.* **C61**, 035204 (2000).
G. Blanpied, et al., *Phys. Rev.* **C64**, 025203 (2001).
G. A. Warren, et al., *Phys. Rev.* **C58**, 3722 (1998).
C. Mertz, et al., *Phys. Rev. Lett.* **86**, 2963–2966 (2001).
C. Kunz, et al., *Phys. Lett.* **B564**, 21–26 (2003).
N. F. Sparveris, et al., *Phys. Rev. Lett.* **94**, 022003 (2005).
K. Joo, et al., *Phys. Rev. Lett.* **88**, 122001 (2002).
V. V. Frolov, et al., *Phys. Rev. Lett.* **82**, 45–48 (1999).
T. Pospischil, et al., *Phys. Rev. Lett.* **86**, 2959–2962 (2001).
P. Bartsch, et al., *Phys. Rev. Lett.* **88**, 142001 (2002).
D. Elsner, et al., *Eur. Phys. J.* **A27**, 91–97 (2006).
S. Stave, et al. (2006), .
S. Stave, [PhD]{} dissertation (unpublished), MIT (2006).
D. Drechsel, O. Hanstein, S. S. Kamalov, and L. Tiator, *Nucl. Phys.* **A645**, 145–174 (1999).
R. A. Arndt, et al., *Phys. Rev.* **C66**, 055213 (2002), .
T. Sato, and T. S. H. Lee, *Phys. Rev.* **C63**, 055201 (2001).
S. S. Kamalov, et al., *Phys. Lett.* **B522**, 27–36 (2001).
A. S. Raskin, and T. W. Donnelly, *Ann. Phys.* **191**, 78 (1989).
V. Pascalutsa, and M. Vanderhaeghen, *Phys. Rev. Lett.* **95**, 232001 (2005).
T. A. Gail, and T. R. Hemmert (2005), .
S. Eidelman, et al., *Phys. Lett.* **B592**, 1 (2004).
Appendix:Contribution of Higher Partial Waves in the Leading Multipole Approximation
====================================================================================
The response functions can be expanded keeping only the terms which interfere with the dominant $M_{1+}$ multipole. The multipoles for $L\ge2$ have been combined into the ${\overline{F_{i}}}$s in the following expansions which are called the Leading Multipole Approximation (LMA):
$$\begin{aligned}
R_{T}^{LMA} & = & \left(\frac{5}{2}-\frac{3}{2}\cos^2\theta\right){\vert{M_{1+}}\vert^2} \nonumber \\
& + & 6\cos^2\theta{\rm Re}\left[{{E_{1+}}^*}{M_{1+}}\right] \nonumber \\
& - & \sin^2\theta\left(3{\rm Re}\left[{{E_{1+}}^*}{M_{1+}}\right] +
{\rm Re}\left[{{M_{1+}}^*}{\overline{F_{3}}}\right]\right) \nonumber\\
& + & (1-3\cos^2\theta)\left({\rm Re}\left[{{M_{1+}}^*}{M_{1-}}\right] +
{\rm Re}\left[{{M_{1+}}^*}{\overline{F_{2}}}\right]\right)\nonumber\\
& + & 2\cos\theta\left({\rm Re}\left[{{E_{0+}}^*}{M_{1+}}\right]+
{\rm Re}\left[{{M_{1+}}^*}{\overline{F_{1}}}\right]\right)\end{aligned}$$
$$\begin{aligned}
R_{TT}^{LMA} & = &
-{\rm Re}\left[{{M_{1+}}^*}(3{E_{1+}}+3{\overline{F_{2}}}+{\overline{F_{3}}}+3{M_{1-}})\right]\sin^2\theta \nonumber \\
& - & \frac{3}{2}{\vert{M_{1+}}\vert^2}\sin^2\theta\end{aligned}$$
$$\begin{aligned}
R_{LT}^{LMA} & = &
\sin\theta{\rm Re}\left[{{M_{1+}}^*}({\overline{F_{5}}}+{L_{0+}}+6{L_{1+}}\cos\theta)\right]\end{aligned}$$
$$\begin{aligned}
R_{LT'}^{LMA} & = &
\sin\theta{\rm Im}\left[{{M_{1+}}^*}({\overline{F_{5}}}+{L_{0+}}+6{L_{1+}}\cos\theta)\right]\end{aligned}$$
$$\begin{aligned}
\sigma_{E2}^{LMA} & = &
-12\sin^2\theta {\rm Re}\left[{{E_{1+}}^*}{M_{1+}}\right]\nonumber \\
& - & 2{\rm Re}\left[{{M_{1+}}^*}{\overline{F_{3}}}\right]\sin^2\theta \nonumber \\
& + & \left(2\cos\theta
+2\right)\left({\rm Re}\left[{{E_{0+}}^*}{M_{1+}}\right] +
{\rm Re}\left[{{M_{1+}}^*}{\overline{F_{1}}}\right]\right)\end{aligned}$$
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Fusion-fission dynamics is investigated with a special emphasis on fusion reactions at low energy for which shell effects and pairing correlations can play a crucial role leading in particular to multi-modal fission. To follow the dynamical evolution of an excited and rotating nucleus we solve a 2-dimensional Langevin equation taking explicitly light-particle evaporation into account. The confrontation theory-experiment is demonstrated to give interesting information on the model presented, its qualities as well as its shortcomings.'
author:
- |
Ch. Schmitt$^1$, J. Bartel$^1$, K. Pomorski$^{1,2}$, A. Surowiec$^2$\
[*$^1$IReS – IN2P3/CNRS and Université Louis Pasteur, Strasbourg, France*]{}\
[*$^2$Katedra Fizyki Teoretycznej, Uniwersytet M. C. Skłodowskiej, Lublin, Poland*]{}
title: 'Fission-Fragment Mass Distribution and Particle Evaporation at low Energies [^1]'
---
Introduction
============
The description of the dynamical evolution of a compound nucleus along its way to fission, i.e. from its rather compact ground-state shape to its scission configuration, represents an intricate problem. Many ingredients enter into the description of such a process, starting from a sufficiently precise account of the formation of the compound system, to the determination of the multi-dimensional energy landscape, to the coupling between the collective dynamics and the intrinsic degrees of freedom of the nucleus, to the concept used to describe light-particle evaporation which can occur all along the fission path. As a general microscopic treatment is completely out of scope, different theoretical approaches based on a more or less classical picture [@GPW79]-[@Fro98] have been proposed.
We have developed such a model describing the time evolution of a highly excited rotating nucleus and its subsequent decay through symmetric fission with pre-fission light-particle emission [@PBR96]. The aim of the present paper is to extend our theory to lower energy. Through a comparison with the available experimental data, in particular fission-fragment mass distributions and neutron pre-scission multiplicities, we hope to get some valuable information on the behavior of transport coefficients at low energy.\
Evolution of an excited rotating nucleus towards fission
========================================================
To study the time evolution of an excited rotating nucleus, the system is assumed to follow a stochastic Langevin equation of motion [@PoStru90] of one or several collective variables that describe in an appropriate and sufficiently flexible way the deformation of the nucleus along its path to fission .\
Description of nuclear shapes
-----------------------------
To describe the large variety of deformed shapes that can appear in the fission process, the Trentalange–Koonin–Sierk (TKS) nuclear shape parametrization [@TKS80] is used. In the case of an axially symmetric system the nuclear surface is given by
$$\rho_s^2\,(z) = R_0^2 \, \sum_{\ell=0}^\Lambda \alpha_{\ell} \,
P_{\ell}\,\left (\frac{z-{\bar z}}{z_0} \right )
= R_0^2 \, \sum_{\ell=0}^\Lambda \alpha_{\ell} \, P_{\ell}\,(u) \; ,
\hspace{ 0.3cm} z_0 = \frac{2 R_0}{3 \alpha_0} \; ,
\hspace{0.3cm} u = \frac{z-{\bar z}}{z_0}
\label{eqon1}$$
with $2 z_0$ the elongation of the shape in z direction, ${\bar z}$ its geometrical center and $R_0$ the radius of the corresponding spherical nucleus. The deformation parameters $\alpha_\ell$ define the shape.
This parametrization is strongly related to the well known Funny Hills $\{c,h, \alpha\}$ parametrization [@BDJ72] recalled below :
$$\rho_s^2\,(z) = c^2 \, R_0^2 \left\{ \begin{array}{ll}
(1 - u^2) \, (A + \alpha \, u + B \, u^2) \; , \; & B \geq 0 \\
(1 - u^2) \, (A + \alpha \, u) \; {\rm exp}(B \, c^3 \, u^2) \; ,\; & B < 0
\end{array}\right.
\label{eqon2}$$
with $z_0 = c \, R_0$ and where $A$ and $B$ are related to $c$ and $h$ through $$A = \frac{1}{c^3} - \frac{1}{5} B \; , \;\;\;\;\;\;\;\;
B = 2 \, h + \frac{1}{2} (c-1) \;\; .$$
We have tested the convergence of these parametrizations for the description of symmetric fission-barrier heights and compared it to the results obtained using the expansion of the nuclear surface in spherical harmonics. The agreement with experiment was better with the TKS parametrization using 3 parameters $\alpha_2$, $\alpha_4$, $\alpha_6$ than with the later including deformation parameters up to $\beta_{14}$, thus showing the fast convergence of the TKS parametrization.
Fission dynamics and Langevin equation
--------------------------------------
Fission dynamics is investigated through the resolution of the Langevin equation which for the generalized coordinates $q_i$ is given by
$$\begin{aligned}
&& \!\!\!\!\!\!\!\!
{dq_i\over dt} = \sum_{j} \; [M^{-1}(\vec q)]_{i\, j} \; p_j
\nonumber\\[ -1.0ex]
&& \label{eqn:rho}\\[ -1.0ex]
&& \!\!\!\!\!\!\!\!
{dp_i\over dt} = - {1\over 2} \sum_{j,k} \,
{d[M^{-1}(\vec q)]_{jk} \over dq_i} \; p_j \; p_k
- {dV(\vec q)\over dq_i}
- \sum_{j,k} \gamma_{ij}(\vec q) \; [M^{-1}(\vec q)]_{jk} \; p_k
% - \sum_{j,k} [\gamma(\vec q)]_{ij} \; [M^{-1}(\vec q)]_{jk} \; p_k
+ F_i(t) \nonumber\end{aligned}$$
${}$\
where $p_i$ are the canonical momenta associated with the coordinates $q_i$. $[M(\vec q)]$ represents the tensor of inertia determined in our approach in the irrotational incompressible fluid approximation of Werner-Wheeler as developed by Davies, Sierk and Nix [@Da76] and $[\gamma(\vec q)]$ corresponds to the friction tensor calculated in the framework of the so-called [*wall and window friction*]{} model [@BRS77; @Fe87]. The collective potential $V(\vec q)$ is defined in our approach as the Helmholtz free energy at given deformation [@PBR96; @BMR96]. The term $F_i(t)$ stands for the random Langevin force which couples collective dynamics to the intrinsic degrees of freedom. We have $F_i(t) \!\!=\!\! \sum_{j} g_{ij}(\vec q) \; G_j(t)$ where the strength tensor $[g(\vec q)]$ is given by the diffusion tensor $[D(\vec q)]$ through $D_{ij} \!\!=\!\! \sum_{k} g_{ik} \; g_{jk}$ and $\vec G(t)$ is a stochastic function. In our model it is assumed that diffusion is related to friction through the Einstein relation $[D(\vec q)] \!=\! [\gamma(\vec q)] \; T$ where T corresponds to the nuclear temperature [@PBR96]. The explicit expressions of these quantities in the TKS parametrization have been presented and discussed in details in [@BMR96].
The friction model we are using is based on a classical concept valid at high energy. When going to lower temperatures this picture can only be considered as an upper limit since nucleon-nucleon collisions become less and less frequent thus reducing friction [@PH91]. We also know that the Einstein relation is in principle only valid at high energy [@HoIv99]. We shall come back to these approximations in section 4.5. and show that one has to modify this simplified description at low temperature to correctly describe the experimental data.
Another quantity entering the Langevin equation and whose temperature dependence requires special attention is the potential $V(\vec q\,)$ namely because of the vanishing of quantal effects at high excitation energy. In our approach valid up to now for symmetric fission, it consisted of a temperature dependent Liquid Drop Model (LDM) term only. At lower energy we have to add to this macroscopic contribution the shell corrections which are evaluated at each deformation using the Strutinsky’s approach [@S67] and the pairing correlations which we calculate in the framework of the BCS model [@BCS57] with a constant pairing strength ([*seniority scheme*]{}) [@NTS69].
The generalized coordinates $q_i$ which enter the Langevin equation are either chosen as the deformation parameters generating the nuclear shape (e.g. coefficients $\alpha_{\ell}$) or as more physically relevant quantities (elongation, mass asymmetry, etc) which are determined through these parameters [@BMR96].
Up to now [@PBR96; @PNS00] we have investigated the case of highly excited compound nuclei giving rise to symmetric fission. Such a process can be described approximately by a single collective coordinate characterizing the nuclear elongation as explained in ref. [@BMR96]. This approach has been proven quite successful reproducing experimental pre-scission neutron multiplicities with an accuracy of $10 - 20 \%$ for nuclei ranging from $^{126}$Ba to the region of superheavy elements [@PNS00]. As our aim in the present paper is to investigate systems at lower energy, one has to be able to describe [*multi-modal*]{} fission caused by the competition between symmetric and asymmetric splitting generated by the quantal effects present at low temperature. Dealing with asymmetric shapes, we need to take at least two collective variables (e.g. elongation and asymmetry) into account describing the compound nucleus along its deformation process. For this purpose we choose to use the Funny-Hills parametrization and to restrict ourselves to the 2-dimensional ($c, \alpha$) deformation space imposing $h = 0$. Indeed, one can show that the influence of the [*neck parameter*]{} $h$ can be considered as rather small, at least in the semi-classical limit [@CSLi00; @CS02].\
Entrance channel effects
------------------------
In order to solve the Langevin equation of motion one needs to specify the initial conditions of the trajectory (for reasonable statistics we need to consider $10^4$ to $10^6$ trajectories) from which the compound system starts and evolves either through the fission channel or ending up as an evaporation residue. The initial conditions for $\vec q_0$ and $\vec p_0$ are fixed to the ground-state deformation and drawn from a normalized gaussian distribution respectively [@PBR96]. The nuclear systems we have investigated so far were generated through heavy-ion collisions which can lead to a large variety of the angular momentum of the synthesized nucleus. The initial spin distribution of the former is determined in our model by solving a Langevin equation [@Prz94] describing the evolution of the two colliding ions from an infinite distance up to fusion. The Langevin equation (3) is then solved in order to describe the dynamical evolution of the synthesized nucleus taking particle emission into account by coupling the Langevin equation to the Master equations governing this evaporation process. For each trajectory we start with a given compound system characterized by its excitation energy and angular momentum. The final prediction, which can be compared to experiment, is then determined by weighting the calculations made at given angular momentum by the fusion-fission cross section [@PNS00]. ${}$\
FIG. 1. Differential fusion (solid line) and fission (histogram) cross-section for the reaction $^{28}$Si$ + ^{98}$Mo$ \rightarrow ^{126}$Ba at $E^*_{tot} = 118.5$ MeV.
FIG. 2. Fission barriers for the nucleus $^{126}$Ba as function of angular momentum at fixed total excitation energy (left) and of temperature at fixed angular momentum (right).\
Fig. 1 shows the fusion and fission cross-sections obtained for the system $^{28}$Si$ + ^{98}$Mo$ \rightarrow ^{126}$Ba at a total excitation energy of $E^*_{tot} = 118.5$ MeV. One notices that fission yields are rather small and located in the tail of the spin distribution at high values of the angular momentum where fission barriers are low. A study of the fission-barrier height as function of angular momentum and thermal excitation energy is given in Fig. 2 from which we conclude that a careful description of the fusion cross-section through its initial spin distribution is necessary if one wants to describe the competition between the decay by fission and light-particle evaporation.\
Light-particle emission
=======================
${}$\
Fission dynamics of an excited rotating nucleus usually goes along with the emission of light particles (we will consider neutrons, protons and $\alpha$ particles). This evaporation process is governed by the emission width $\Gamma_{\nu}^{\mu \kappa}(E^*,L)$ for emitting a particle of type $\nu$, energy $\varepsilon_{\mu}$ and angular momentum $\ell_{\kappa}$ from a nucleus characterized by its thermal excitation energy $E^*$ and its rotational angular momentum $L$. In order to determine $\Gamma_{\nu}^{\mu \kappa}(E^*,L)$ we use two different prescriptions.
FIG. 3. Emission widths for neutrons, protons and $\alpha$ particles emitted from the system $^{160}$Yb (E$^*$=50 MeV, L=40$\hbar$) as a function of elongation.
In Weisskopf’s evaporation theory [@Wei37] the decay rates are essentially evaluated through the level densities of the mother and the daughter nuclei and the transmission coefficient for emitting the particle from a given point of the nuclear surface into a given direction as explained in ref. [@PBR96]. In practice it is not possible to discuss the values of the emission width for each energy, angular momentum and position of the emission point on the nuclear surface. We therefore use them to determine the probability $\Gamma_{\nu} (E^*,L)$ of emitting a given particle from a given nucleus at given deformation. This simplified procedure calculates a transmission coefficient obtained by an averaging over the different emission directions and over the whole surface of the deformed nucleus. A detailed description of this procedure can be found in ref. [@PBR96]. Also other groups [@SDP91; @AAR96; @Fro98] have dealed with particle emission in connection with fission dynamics but, to our knowledge, none of them has taken nuclear deformation explicitly into account as we have, even if it is in an approximate way. In Fig. 3 the evaporation rate $\Gamma_{\nu}$ is displayed for a hot rotating nucleus $^{160}$Yb. It becomes obvious that the deformation dependence of $\Gamma_{\nu}$ is essential and that assuming a deformation independent emission width could probably lead to wrong predictions.
The second approach we used so far to describe particle emission calculates the transition rates $\Gamma^{\mu\kappa}_\nu$ through the probability that a particle which impinges on the nuclear surface at a given point $\vec r_0\,'$ and with a given velocity $\vec v_0\,'$ is actually transmitted [@DPR95]. In this framework the quantity $\Gamma^{\mu\kappa}_\nu$ is determined as $$\Gamma^{\mu\kappa}_\nu = {d^2 n_\nu \over d\varepsilon_\mu d\ell_\kappa}
\Delta\varepsilon \,\Delta \ell \; .
\label{eqtw10}$$
The number $n_\nu$ of particles of type $\nu$ which are emitted per time unit through the surface $S$ of the fissioning nucleus is given by $$n_\nu = \int\limits_S \! d\sigma \int \! d^3 p' \, f_\nu(\vec r_0\,',
\vec p\,') \, v'_\perp(\vec r_0\,') \, w_\nu( v'_\perp (\vec r_0\,'))$$ where the quantity $f_\nu(\vec r\,' , \vec p\,')$ corresponds to the quasi-classical phase-space distribution function [@DPR95].\
${}$\
${}$\
FIG. 4. Emission rates $\Gamma_n$ and $\Gamma_p$ for neutrons and protons obtained in the Weisskopf and the distribution function (called Thomas-Fermi here) approaches at deformations close to the spherical shape ($q=0.73$) and to the scission configuration ($q=2.01$). Fig. 4 gives a comparison of the neutron and proton emission rates in the two evaporation models obtained for different values of a collective coordinate $q$ related to nuclear elongation [@BMR96] for the system $^{188}$Pt ($E^* \!\!=\!\! 100$ MeV, $L \!\!=\!\! 0 \hbar$) [@SPS01]. Both models yield emission rates that are reasonably close for both types of particles for all elongations except for an increase of the distribution function approach relative to the Weisskopf prediction in the case of protons for very large deformations. One should also notice the deformation dependence of the proton emission width $\Gamma_p$ that can be easily understood if one keeps in mind that the Coulomb barrier which charged particles have to overcome depends on the direction of the emission (an emission along the tips is favored compared to an emission perpendicular to the symmetry axis).
The determination of the phase-space distribution function is quite intricate in the case of $\alpha$-particles which are composite particles. We are presently working on a model which determines the $\alpha$-particle distribution function $f_\alpha$ through those of two correlated protons and neutrons respectively [@SPS01].\
Theoretical results of fission dynamics
=======================================
Influence of quantal effects
----------------------------
### One- versus 2-dimensional Langevin equation
In the framework of the 2-dimensional Langevin equation solved in the ($c, \alpha$) deformation space, the LDM energy landscape is displayed on Fig. 5 together with a typical fission trajectory for the compound nucleus $^{227}$Pa at a total excitation energy of $E^*_{tot} \!=\! 26$ MeV and an angular momentum of $L = 60 \hbar$. We choose this specific nuclear system because it was the object of a recent experimental campaign [@CS02]. As no shell effects are taken into account here, only the symmetric fission valley is present. Consequently the compound nucleus starting from its ground-state deformation ($c \!=\! 1.11 , \alpha \!=\! 0$), naturally ends up in the symmetric fission channel. In this calculation we have not coupled particle emission to the Langevin equation and therefore cannot make any statement on particle multiplicities. The fission time, defined as the average time which a trajectory takes to reach the scission point, is in the present 2-dimensional treatment reduced by about 7% (5.96 $10^{-17}$ sec versus 6.36 $10^{-17}$ sec) as compared to its 1-dimensional value [@CSLi00].
Let us try to understand this result since it might seem astonishing that resolving the 2-dimensional Langevin equation, where trajectories can fill out more effectively the deformation space (as it is demonstrated with the typical trajectory drawn on Fig. 5), would lead to shorter fission times than when the compound nucleus follows the deepest symmetric fission valley of the 1-dimensional picture. In fact we have to think of the Langevin equation as an approximation to the Fokker-Planck one which deals with probability distributions. In an 1-dimensional space the system is constrained whereas, the more the dimensionality is increased, the less constraints one has.\
FIG. 5. Temperature dependent LDM energy landscape and typical fission trajectory for the compound system $^{227}$Pa ($E^*_{tot} = 26$ MeV, $L = 60 \hbar$).
This result on fission times seems also to indicate that in the case of highly excited nuclei our previous 1-dimensional description was already fairly accurate. The small change in times should indeed imply a rather small change in pre-fission particle multiplicities.
### Influence of shell effects and pairing correlations
Let us now go one step further by including in our potential energy calculation quantal effects and their dependence on temperature. It is generally admitted that shell corrections have disappeared for temperatures above $2.5$ to $3$ MeV whereas pairing correlations have already vanished at $T \!\approx\! 1.5$ MeV or even before. In order to take care of the $T$-dependence of quantal corrections we multiply their values obtained at $T \!=\! 0$ MeV with a temperature smoothing function which goes to zero at $T \!\!=\! 3$ MeV for shell corrections and at $T \!=\! 1.5$ MeV in the case of pairing [@I78] - [@EgRo00]. The energy landscape then obtained for $E^*_{tot} \!=\! 26\,$MeV and $L \!=\! 60\,\hbar$ is drawn on Fig. 6. Comparing the landscapes in Figs. 5 and 6, one notices the appearance, due to the presence of microscopic corrections, of asymmetric fission channels beyond $c \approx 1.7$ ending up in well pronounced valleys around $\alpha \approx \pm 0.035$.
The resolution of the Langevin equation in the landscapes of Figs. 5 and 6 gives rise to the distributions for the asymmetry parameter and fission-fragment masses presented in Figs. 7 and 8 respectively. Whereas symmetric distributions were obviously expected for the pure LDM landscape, the distributions obtained in the case where quantal effects are present are a little surprising, because of their very strong asymmetry in spite of the rather flat energy landscape of Fig. 6 in the asymmetry direction $\alpha$ for large elongations ($c \approx 2.0$). However one should not forget that the fragment mass distribution is decided all along the fission path and not only in the immediate neighborhood of the exit point [@MN70] - [@NPPW79]. As the asymmetric valley is around $1$ MeV deeper than the symmetric fission path in the vicinity of $c \approx 1.8 - 1.9$ where $\alpha \approx \pm 0.035$, the predominant part of the trajectories finally ends up in this asymmetric channel.\
FIG. 6. Same as Fig. 5 but with inclusion of quantal effects.\
FIG. 7. Fission-fragment distribution as function of the asymmetry parameter $\alpha$ (left) and of the fragment mass (right) when quantal effects are omitted.\
FIG. 8. Same as Fig. 7 but with inclusion of quantal effects.\
It is interesting to notice that the average fission time is increased from 8.0 $10^{-17}$ sec to 16.2 $10^{-17}$ sec when going from the LDM picture to the one with shell and pairing corrections.\
FIG. 9. Same as Fig. 6 with a typical asymmetric fission trajectory. In the LDM landscape the symmetric fission barrier at $L = 60$ MeV is of the order of $0.7$ MeV. On the other hand when quantal effects are included, the asymmetric fragment partition related to $\alpha \approx \pm 0.035$ corresponds to a barrier of $0.1$ MeV. One notices that in spite of a lower barrier height when including microscopic corrections the corresponding fission time is larger than in the semi-classical picture. One understands this result if we compare the typical symmetric trajectory drawn on Fig. 5 to the typical asymmetric trajectory drawn on Fig. 9. In the former case the system cannot overcome the rather high [*mountain top*]{} at $c \approx 1.55$ for $\alpha \in [-0.06 \,,\, 0.06]$ and consequently has to bypass it along $\alpha \approx \pm 0.1$ before reaching the asymmetric valley for $\alpha \approx \pm 0.035$. One could say that the path is [*longer*]{}. Let us note that a small energy difference (of the order of a few hundreds keV) between valleys can lead to very different fragment distributions what suggests the strong dependence of the dynamics on the [*details*]{} of the energy landscape. This drastic sensitivity to the structure of the landscape requires to be careful when one performs energy calculations in the deformation space.
Dynamics including light-particle evaporation
---------------------------------------------
Even if we know that particle evaporation is strongly reduced at low energy we have to admit that we do not have yet a complete description evaluating the emission widths $\Gamma_{\nu}$ at low temperature since one can seriously question the validity of the Weisskopf’s theory at such energies and since our development of the more microscopic phase-space distribution function approach is not complete. Nevertheless in order to investigate particle evaporation, we consider in this section the compound nucleus $^{227}$Pa at a higher total excitation energy of $56$ MeV for which we believe that the Weisskopf’s approach should be approximately valid.\
[||p[3.5cm]{}||\*[2]{}[c|]{}|]{} & $\;\;\; V^{LDM} \;\;\;$ & $\;\; V^{LDM} + \delta E \;\;$\
$\;\; \sigma_{fis} / \sigma_{tot}$ ($\%$) & $99.8$ & $98.5$\
$\;\; {\bar t}_{fis} (\times 10^{-17} sec)$ & $2.335$ & $3.275$\
$\;\; {M}_{n}$ & $1.806$ & $2.153$\
$\;\; {M}_{p}$ & $0.010$ & $0.006$\
$\;\; {M}_{\alpha}$ & $0.017$ & $0.011$\
Table 1 : Influence of quantal corrections on fission probability, average fission time and light-particle multiplicities obtained for the system $^{227}$Pa ($E^*_{tot} \!=\! 56$ MeV, $L \!=\! 60 \hbar$).\
In Table 1 we compare the fission cross section, average fission time and light-particle multiplicities obtained for the pure LDM description to the ones related to the potential energy surface including shell and pairing corrections. As in the case without particle evaporation, one observes an increase of the fission time when quantal effects are taken into account. Whereas the neutron pre-scission multiplicity is larger in the calculations with microscopic corrections, charged particle multiplicities are smaller. With Fig. 3 we have seen that neutrons can be emitted whatever the nuclear elongation, i.e. all along the fission path, and that their emission probability increases with increasing deformation. A longer fission time should therefore lead to a larger neutron multiplicity. Charged particles are preferentially emitted at large deformations (see again Fig. 3). One has, however, to remember that when charged particle emission is favored a substantial amount of the available excitation energy of the emitting nucleus can already have been carried away through neutron emission. In addition one finds that the gradient of the potential energy for the asymmetric fission path including quantal effects is larger between saddle and scission points than the one of the symmetric valley of the LDM landscape. This suggests that the corresponding time scale for the descent from saddle to scission is smaller in the case when shell and pairing effects are present what again favors a reduction of charged particle multiplicities.
Influence of excitation energy and angular momentum
---------------------------------------------------
As shown in Fig. 10 an increase of the total excitation energy of the system from $26$ to $56$ MeV (which for a given angular momentum $L = 60 \hbar$ implies an increase of the thermal excitation energy) leads to a larger contribution to the symmetric fission mode. This result is obviously due to the vanishing of quantal effects when the nuclear temperature increases. However it can also be partly explained by a larger diffusion generated by the larger temperature (see Einstein relation). The corresponding larger oscillations thus allow the nucleus to [*explore more easily*]{} the energy landscape being able to overcome higher barriers and consequently to pass from one valley to another instead of being trapped preferentially in the deepest valley (which is asymmetric for the system presently considered). In ref [@CS02] we also investigated the impact of the angular momentum on fission dynamics and obtained an relative increase of the symmetric fission cross section for increasing angular momentum due to a decrease of the fission barrier height.
\
FIG. 10. Fission-fragment mass distributions for two different values of the total excitation energy of the compound nucleus $^{227}$Pa at an angular momentum of $L = 60 \hbar$.\
Evaluation of shell corrections close to the scission point
-----------------------------------------------------------
In the framework of the Strutinsky method, shell correction calculations need to determine nuclear single particle levels which in our approach are the eigenvectors of a deformed Saxon Woods potential of standard parametrization [@BDJ72]. In practice these states are obtained by an expansion in the basis of a deformed harmonic oscillator. This oscillator basis is an [*one-center*]{} basis which is probably not so well adapted if one is interested in describing shapes near the scission point. Indeed, for such strongly elongated and possibly [*necked-in*]{} surfaces, a [*two-center*]{} basis seems to be more adapted taking the structure of the nascent fragments better into account. We stress this technical detail in order to focus on the importance of a careful determination of shell effects at very large deformations. To illustrate this point we compare on Fig. 11 the fragment mass distribution obtained when the dynamical calculation is artificially stopped at an elongation $c_{scis} = 1.8$ to the one obtained when this calculation is carried through up to the geometrical scission point $c_{scis} \!=\! c_{geo}$ where the splitting into two fragments takes place. The broad distribution related to $c_{scis} \!=\! 1.8$ can be easily understood with Fig. $6$ where the quite flat potential landscape in the $\alpha$ direction around $c \approx 1.8$ can give rise to a large variety of mass partitions. In spite of this, the final distribution at $c_{geo}$ is rather strongly asymmetric. Moreover the value $c \!=\! 1.8$ corresponds to a quite important elongation, i.e. an elongation for which one can already have a reasonable idea of the asymmetry of the nascent fission fragments [@NPPW79]. The present investigation points out the importance of quantal effects for $c \!>\! 1.8$ and with it the necessity of their accurate determination for these largest deformations. To avoid problems related to the choice of this one-center basis we perform the diagonalisation taking a very large number of basis states into account.
\
FIG. 11. Fission-fragment mass distributions obtained for the system $^{227}$Pa ($E^*_{tot} = 56$ MeV, $L = 60 \hbar$) for $c_{scis} \!=\! 1.8$ and $c_{scis} \!=\! c_{geo}$ (see text).\
Temperature dependence of transport coefficients
------------------------------------------------
As mentioned in section 2.2. we probably overestimate friction at low temperature. As demonstrated on Fig. 12 the reduction of friction by a factor of two ($0.5 \; w \& w$) results in a striking difference as compared to the full wall-and-window friction ($w \& w$). A larger friction causes a decrease of the kinetic energy of the system which is therefore more sensitive to the [*fine structure*]{} of the landscape and consequently is more easily trapped in the deepest valleys. A smaller friction, on the contrary, allows the system, with larger kinetic energy, to move more freely through the landscape, to overcome more easily eventual barriers, resulting in a broader distribution. Reducing friction by a constant factor is obviously an extremely crude approximation to a real temperature dependent viscosity. We use this picture here simply to investigate the influence of friction on fragment distributions and light-particle multiplicities.
The procedure used in order to simulate in an approximate way the vanishing of quantal effects with temperature (see section 4.1.2) is still nowadays subject of controversies, in particular what pairing is concerned [@I78] - [@EgRo00]. Our investigations dealing with this point (for details see ref [@CS02]) showed that the T-dependence of shell and pairing corrections cannot be neglected, even if our system is already in the beginning of its decay at quite low excitation energy, which can still decrease along the fission path (namely due to particle evaporation).
\
FIG. 12. Fission-fragment mass distributions obtained for the system $^{227}$Pa ($E^*_{tot} \!=\! 56$ MeV, $L \!=\! 60 \hbar$) with the full ($w \& w$) and a reduced ($0.5 \; w \& w$) friction (see text).\
Confrontation with experimental data
====================================
As the agreement theory-experiment at high excitation energy is quite promising [@PNS00], we would like to compare in the present section our predictions to the available experimental data concerning the fission process of the nucleus $^{227}$Pa synthesized at a total excitation energy of $E^*_{tot} \!=\! 26$ MeV [@CS02]. In the calculations we should obviously take particle evaporation into account. Since we do not have for the moment a complete reliable evaporation theory at our disposal at low temperature we first performed dynamical calculations at higher energy for which we believe that the Weisskopf’s approach is about reasonable. This study showed us that the influence of particle evaporation on the fission fragment mass distribution can be neglected [@CS02]. As the probability of emitting particles decreases with excitation energy [@PBR96], we also expect a really small impact of evaporation on the mass distribution at $26$ MeV. Consequently we compare in Fig. 13 mass distributions obtained for $E^*_{tot} \!=\! 26\,$MeV without taking evaporation into account with the experimental mass distribution. We have considered in the theoretical calculations three different frictions : $25 \%$ of the wall and window value, $20 \%$ and $15 \%$.
The experimental analysis has exhibited a multi-modal fission-fragment mass distribution [@CS02] composed of three modes : the symmetric one and two asymmetric modes centered around mass $A \!=\! 132$ corresponding to the double magic $^{132}$Sn nucleus and around mass $A \!=\! 140$ related to the deformed $^{140}$Ba nucleus, explained [@BG90; @IVP00] by the closure of the deformed neutron shell $N \!=\! 84$. The comparison with our predictions shows that in the case of a friction corresponding to $15 \%$ of the wall-and-window value the model reproduces quite well the symmetric fission mode. We would like to mention here that microscopic calculations performed by Hofmann and Ivanyuk [@HIR01] indicate that such a reduced viscosity is about what is to be expected at such low excitation energy. However our calculation gives only rise to the asymmetric $A \!=\! 132$ channel, the $A = 140$ mode being absent.\
${}$ $\;\; 0.25 \; w \& w
\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; 0.2 \; w \& w
\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; 0.15 \; w \& w$\
FIG. 13. Experimental (solid line) and theoretical (histograms) fission fragment mass distributions for the system $^{227}$Pa ($E^*_{tot} = 26$ MeV) for different values of the friction.\
In order to understand the disagreement between our model and the experimental data for asymmetric fission we have to remember that we have chosen to describe nuclear shapes in the 2-dimensional deformation space ($c, \alpha$) imposing $h \!=\! 0$. Taking $h$ different from zero will allow us to consider a larger variety of nuclear configurations. We thus believe that with the restricted 2-dimensional parametrization we are not able to give a description of the deformed shape of $^{140}$Ba but that when taking $h \!\neq\! 0$ into account we will describe that shape and the corresponding asymmetric fission valley so that a part of the trajectories which, for $h \!=\! 0$, end up in the $A \!=\! 132$ channel will reach, in the case of $h \neq 0$, the previously missing $A \!=\! 140$ valley. The contribution to the fission mode $A \!=\! 132$ will then decrease while the one of the $A = 140$ channel will increase, thus reaching a better agreement between theory and experiment when we will have extended the present 2-dimensional treatment to a 3-dimensional one. Investigations along this direction are under way.\
Discussion and conclusions
==========================
With the purpose to study multi-modal fission, we have developed a model describing the dynamics of the fission process by the resolution of a 2-dimensional Langevin equation coupled to the Master equations governing particle emission. Starting from a more or less classical description proven as rather successful for describing symmetric fission at high excitation energy, we extended our theory to multi-modal fission by increasing the dimensionality of the deformation space in which the Langevin equation is solved in order to be able to deal with asymmetric shapes and by including quantal effects (shell and pairing corrections) in the potential-energy calculations. Our investigations show the strong sensitivity of the dynamics on the structure of the potential-energy landscape what implies the necessity for a careful description of the later, in particular in the determination of shell and pairing corrections at large deformation.
Comparing theoretical and experimental fission-fragment mass distributions one observes a rather promising agreement which, as we believe, could still be considerably improved if the 2-dimensional treatment is extended to a 3-dimensional one. We also point out the importance of taking into account the temperature dependence of nuclear friction which as we have seen should be significantly reduced at low energy. Another crucial aspect of the problem lies in the necessity of a reliable evaporation theory at low excitation energy.
Up to now the general analysis was that pre-scission light-particle multiplicities were the quantities to investigate [@Fro98; @PNS00] for a better understanding of fission dynamics. Our present study shows, on the contrary, that at low excitation energies where the number of emitted particles is small and, in the frequent case where the competition between symmetric and asymmetric channels exhibits multi-modal fission, the fragment mass distribution is probably more relevant, in particular for investigating transport coefficients like nuclear friction.\
[**Acknowledgments**]{}\
Two of us (C.S. and J.B.) are very grateful for the hospitality extended to them on many occasions by the Theory Department of the Marie-Curie-Skłodowska University in Lublin. A.S. and K.P. are in turn very thankful for many fruitful visits they were able to do to the Strasbourg Nuclear Research Center IReS and its Nuclear Theory group. K.P. especially acknowledges a PAST position granted to him by the University Louis Pasteur and the French Ministry of National Education and Research. This work was partially sponsored by the Polish Committee of Scientific Research KBN No. 2P 03B 115 19 and the POLONIUM fellowship No. 007/IN2P3/4788/2002.
[99]{} P. Grangé, H.C. Pauli and H.A. Weidenmüller, Phys. Lett. [**B88**]{} (1979) 9;\
Z. Phys. [**A296**]{} (1980) 107. E. Strumberger, W. Dietrich and K. Pomorski, Nucl. Phys. [**A529**]{} (1991) 522. W. Przystupa, K. Pomorski, Nucl. Phys. [**A572**]{} (1994) 153. Y. Abe, S. Ayik, P.-G. Reinhard and E. Suraud, Phys. Rep. [**275**]{} (1996) 49. K. Pomorski, J. Bartel, J. Richert and K. Dietrich, Nucl. Phys. [**A605**]{} (1996) 87. P. Fröbrich, I.I. Gontchar, Phys. Rep. [**292**]{} (1998) 131. K. Pomorski, E. Strumberger, Ann.Univ.MCS, Poland, sec.AAA XLV (1990) 113. S. Trentalange, S.E. Koonin, A. Sierk, Phys. Rev. [**C22**]{} (1980) 1159. M. Brack, J. Damgaard, A.S. Jensen, H.C. Pauli, V.M. Strutinsky, C.Y.Wong, Rev. Mod. Phys. [**44**]{} (1972) 320. K.T.R. Davies, A.J. Sierk, J.R. Nix, Phys. Rev. [**13C**]{} (1976) 2385. J. Blocki, J. Randrup, W. J. Swiatecki and C.W. Tsang, Ann. Phys. [**105**]{} (1977) 427. H. Feldmeier, Rep. Prog. Phys. [**50**]{} (1987) 915. J. Bartel, K. Mahboub, J. Richert and K. Pomorski, Z. Phys. [**A354**]{} (1996) 59. K. Pomorski, H. Hofmann, Phys. Lett. [**B263**]{} (1991) 164. H. Hofmann, F.A. Ivanyuk, Phys. Rev. Lett. [**82**]{} (1999) 4603. V.M. Strutinsky, Nucl. Phys. [**A95**]{} (1967) 420; [**A122**]{} (1968) 1. J. Bardeen, L.N. Cooper and J.R. Schrieffer, Phys. Rev. [**108**]{} (1957) 1175. S.G. Nilsson, C.F. Tsang, A. Sobiczewski, Z. Szymanski, S. Wycech, C. Gustafson, I.-L. Lamm, P. Moller, B. Nilsson, Nucl. Phys. [**A131**]{} (1969) 1. K. Pomorski, B. Nerlo-Pomorska, A. Surowiec, M. Kowal, J. Bartel, K. Dietrich, J. Richert, C. Schmitt, B. Benoit, E. de Goes Brennand, L. Donadille, C. Badimon, Nucl. Phys. [**A679**]{} (2000) 25. C. Schmitt, A. Surowiec, J. Bartel, K. Pomorski, Proccedings of the Int. Conf. Nuclear Physics at Border Lines, Lipari, Italy, May 21-24, 2001, Edited by G.Fazio, G.Giardina, F.Hanappe, G.Immè and N.Rowley, World Scientific. C. Schmitt, PhD Thesis, Université Louis Pasteur Strasbourg ($2002$), IReS $02-04$. V. Weisskopf, Phys. Rev. [**52**]{} (1937) 295. K. Dietrich, K. Pomorski and J. Richert, Z. Phys. [**A351**]{} (1995) 397. A. Surowiec, K. Pomorski, C. Schmitt, Acta Physica Polonica [**B33**]{} (2002). in Nuclear Fission, Contributed Papers, W. Berlin ($1989$), 63. A.V. Ignatyuk, Phys. Lett. [**B76**]{} (1978) 543. A.V. Ignatyuk, Nucl. Phys. [**A346**]{} (1980) 191. A.V. Ignatyuk, G.N. Smirenkin, A.S. Tishin, Yad. Fiz. [**21**]{} (1975) 485. J.L. Egido, L.M. Robledo, V. Martin, Phys. Rev. Lett. [**85**]{} (2000) 054308. P. Moller, S.F. Nilsson, Phys. Lett. [**B31**]{} (1970) 283. J.M. Maruhn, W. Greiner, Phys. Rev. Lett. [**32**]{} (1974) 548. B. Nerlo-Pomorska, K. Pomorski, E. Werner, International Conference Fifty Years Research U. Brosa, S. Grossman, A. Muller, Physics Reports, [**197**]{} (1990) 167. I.V. Pokrovsky et al., Phys. Rev. [**C62**]{} (2000) 014615. H. Hofmann, F.A. Ivanyuk, C. Rummel, S. Yanji, Phys. Rev. [**C64**]{} (2001) 054316.
[^1]: This paper is devoted to Professor Adam Sobiczewski on the occasion of his 70th birthday
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'M. L. Ahnen'
- 'S. Ansoldi'
- 'L. A. Antonelli'
- 'P. Antoranz'
- 'C. Arcaro'
- 'A. Babic'
- 'B. Banerjee'
- 'P. Bangale'
- 'U. Barres de Almeida'
- 'J. A. Barrio'
- 'J. Becerra González'
- 'W. Bednarek'
- 'E. Bernardini'
- 'A. Berti'
- 'B. Biasuzzi'
- 'A. Biland'
- 'O. Blanch'
- 'S. Bonnefoy'
- 'G. Bonnoli'
- 'F. Borracci'
- 'T. Bretz'
- 'R. Carosi'
- 'A. Carosi'
- 'A. Chatterjee'
- 'P. Colin'
- 'E. Colombo'
- 'J. L. Contreras'
- 'J. Cortina'
- 'S. Covino'
- 'P. Cumani'
- 'P. Da Vela'
- 'F. Dazzi'
- 'A. De Angelis'
- 'B. De Lotto'
- 'E. de Oña Wilhelmi'
- 'F. Di Pierro'
- 'M. Doert'
- 'A. Domínguez'
- 'D. Dominis Prester'
- 'D. Dorner'
- 'M. Doro'
- 'S. Einecke'
- 'D. Eisenacher Glawion'
- 'D. Elsaesser'
- 'M. Engelkemeier'
- 'V. Fallah Ramazani'
- 'A. Fernández-Barral [^1]'
- 'D. Fidalgo'
- 'M. V. Fonseca'
- 'L. Font'
- 'K. Frantzen'
- 'C. Fruck $^\star$'
- 'D. Galindo'
- 'R. J. García López'
- 'M. Garczarczyk'
- 'D. Garrido Terrats'
- 'M. Gaug'
- 'P. Giammaria'
- 'N. Godinović'
- 'D. Gora'
- 'D. Guberman'
- 'D. Hadasch'
- 'A. Hahn'
- 'M. Hayashida'
- 'J. Herrera'
- 'J. Hose'
- 'D. Hrupec'
- 'G. Hughes'
- 'W. Idec'
- 'K. Kodani'
- 'Y. Konno'
- 'H. Kubo'
- 'J. Kushida'
- 'A. La Barbera'
- 'D. Lelas'
- 'E. Lindfors'
- 'S. Lombardi'
- 'F. Longo'
- 'M. López'
- 'R. López-Coto'
- 'P. Majumdar'
- 'M. Makariev'
- 'K. Mallot'
- 'G. Maneva'
- 'M. Manganaro'
- 'K. Mannheim'
- 'L. Maraschi'
- 'B. Marcote'
- 'M. Mariotti'
- 'M. Martínez'
- 'D. Mazin'
- 'U. Menzel'
- 'J. M. Miranda'
- 'R. Mirzoyan'
- 'A. Moralejo'
- 'E. Moretti'
- 'D. Nakajima'
- 'V. Neustroev'
- 'A. Niedzwiecki'
- 'M. Nievas Rosillo'
- 'K. Nilsson'
- 'K. Nishijima'
- 'K. Noda'
- 'L. Nogués'
- 'S. Paiano'
- 'J. Palacio'
- 'M. Palatiello'
- 'D. Paneque'
- 'R. Paoletti'
- 'J. M. Paredes'
- 'X. Paredes-Fortuny'
- 'G. Pedaletti'
- 'M. Peresano'
- 'L. Perri'
- 'M. Persic'
- 'J. Poutanen'
- 'P. G. Prada Moroni'
- 'E. Prandini'
- 'I. Puljak'
- 'J. R. Garcia'
- 'I. Reichardt'
- 'W. Rhode'
- 'M. Ribó'
- 'J. Rico'
- 'T. Saito'
- 'K. Satalecka'
- 'S. Schroeder'
- 'T. Schweizer'
- 'A. Sillanpää'
- 'J. Sitarek'
- 'I. Snidaric'
- 'D. Sobczynska'
- 'A. Stamerra'
- 'M. Strzys'
- 'T. Surić'
- 'L. Takalo'
- 'F. Tavecchio'
- 'P. Temnikov'
- 'T. Terzić'
- 'D. Tescaro'
- 'M. Teshima'
- 'D. F. Torres'
- 'T. Toyama'
- 'A. Treves'
- 'G. Vanzo'
- 'M. Vazquez Acosta'
- 'I. Vovk'
- 'J. E. Ward'
- 'M. Will'
- 'M. H. Wu'
- 'R. Zanin'
bibliography:
- 'bibliography.bib'
date: 'Received ... ; accepted ...'
title: 'Very-high-energy gamma-ray observations of the Type Ia Supernova SN 2014J with the MAGIC telescopes'
---
[In this work we present data from observations with the MAGIC telescopes of SN 2014J detected in January 21 2014, the closest Type Ia supernova since Imaging Air Cherenkov Telescopes started to operate.]{} [We probe the possibility of very-high-energy (VHE; $E\geq100$ GeV) gamma rays produced in the early stages of Type Ia supernova explosions.]{} [We performed follow-up observations after this supernova explosion for 5 days, between January 27 and February 2 in 2014. We search for gamma-ray signal in the energy range between 100 GeV and several TeV from the location of SN 2014J using data from a total of $\sim5.5$ hours of observations. Prospects for observing gamma-rays of hadronic origin from SN 2014J in the near future are also being addressed.]{} [No significant excess was detected from the direction of SN 2014J. Upper limits at 95$\%$ confidence level on the integral flux, assuming a power-law spectrum, d$F/$d$E\propto E^{-\Gamma}$, with a spectral index of $\Gamma=2.6$, for energies higher than 300 GeV and 700 GeV, are established at $1.3\times10^{-12}$ and $4.1\times10^{-13}$ photons cm$^{-2}$s$^{-1}$, respectively.]{} [For the first time, upper limits on the VHE emission of a Type Ia supernova are established. The energy fraction isotropically emitted into TeV gamma rays during the first $\sim10$ days after the supernova explosion for energies greater than 300 GeV is limited to $10^{-6}$ of the total available energy budget ($\sim 10^{51}$ erg). Within the assumed theoretical scenario, the MAGIC upper limits on the VHE emission suggest that SN 2014J will not be detectable in the future by any current or planned generation of Imaging Atmospheric Cherenkov Telescopes.]{}
Introduction
============
Type Ia supernovae (SNe) are extremely luminous stellar explosions, which are believed to originate from primary carbon-oxygen white dwarfs (WD) in binary systems reaching the Chandrasekhar mass limit of 1.4 M$_{\odot}$ [@Chandrasekhar1931]. When this happens, the electron-degenerate core can no longer support the gravitational pressure, leading to an implosion of the progenitor WD. Thereby the temperature grows up to the carbon fusion point, giving rise to a thermonuclear explosion releasing so much nuclear energy ($\sim10^{51}$ erg; @1993ApJ...412..192B) that no compact remnant is expected. The nature of the companion star is still unclear, although two classical progenitor scenarios have been promoted: single-degenerate model, in which the WD accretes material from a red giant star [@1973ApJ...186.1007W], and the double-degenerate model, in which the explosion is produced by the merging of two WDs [@1984ApJS...54..335I]. Type Ia SNe have been used to provide information on the Galactic chemical evolution [@1995ApJS...98..617T] and to measure cosmological parameters [e.g. @Perlmutter:1998np] since they can be used as standard candles thanks to their consistent luminosity [@Branch1992]. Still, the evolutionary path that leads to a carbon-oxygen WD which exceeds the Chandrasekhar limit is not well understood yet.\
SN 2014J was detected on January 21 2014 (MJD 56678) by the UCL Observatory [@2014CBET.3792....1F] and classified as a Type Ia SN with the Dual Imaging Spectrograph on the ARC 3.5 m telescope (January 22; @Goobar2014). It is located in the starburst galaxy M82 at a distance of 3.6 Mpc [@karachentsev_masses_2006]. Its proximity has granted it the title of the nearest Type Ia SN in the past 42 years and motivated large multiwavelength follow-up observations from radio to very-high-energy (VHE; $E\geq100$ GeV) gamma rays.\
Deep studies of color excess and reddening estimation were carried out on SN 2014J. These studies helped to understand the properties of the dust that affects the brightness of the SN as well as to provide new clues on the progenitors, which are important parameters in the cosmology investigation. [@2014ApJ...788L..21A] presented, for the first time, a characterization of the reddening of a Type Ia SN in a full range from $0.2~\mu$m to $2~\mu$m. Their results, with reddening values of $E(B-V)\sim1.3$ and $R_{V}\sim1.4$, are compatible with a power-law extinction, expected in the case of multiple scattering scenarios. In the same wavelength band, from UV to near infrared (NIR), [@2014MNRAS.443.2887F] found reddening parameter values of $E(B-V)\sim1.2$ and $R_{V}\sim1.4$. In this model, the extinction is explained to be caused by a combination of the galaxy dust and a dusty circumstellar medium. However, although compatible with an extinction law with a low value of $R_{V}\sim1.4$ and consistent as well with previous mentioned results, [@2015ApJ...805...74B], making use of *Swift*-UVOT data, suggested that most of the reddening is caused by the interstellar dust. Optical and NIR linear polarimetric observations of the source presented in [@2014ApJ...795L...4K] supports the scenario where the extinction is mostly produced by the interstellar dust. These evidences favor the double-degenerate scenario for SN 2014J, where less circumstellar dust is expected than in cases with a giant companion star. This type of companion are indeed ruled out by several authors as possible progenitors in SN 2014J, e.g. [@pereztorres2014], with the most sensitive study in the radio band of a Type Ia SN, or [@Margutti:2014uha] in the X-ray band. The former reported non-detection from the observations performed with eMERLIN and EVN. These results, compared with a detailed modeling of the radio emission from the source, allowed to exclude the single-degenerate scenario in favour of the double-degenerate one with constant density medium of $n \lesssim1.3$ cm$^{-3}$.
Several authors have speculated about the possibility of SN explosions being able to produce gamma-ray emission at detectable level by current and/or future telescopes. However, these models generally consider Type II SNe due to the strong wind of the progenitors (e.g. and ). Nevertheless, given the proximity of SN 2014J, this event provides a good exploratory opportunity to probe the eventual production of VHE gamma rays during the first days after such an explosion.
In this work, we present the analysis results of SN 2014J observations performed with the MAGIC telescopes.
Observations $\&$ Results
=========================
---------------- ----------- ------- ------------------ -------------------------------- --------------------------------
Eff. Time Zd UL ($E>300$ GeV) UL ($E>700$ GeV)
\[yyyy-mm-dd\] \[MJD\] \[h\] \[degrees\] \[photons cm$^{-2}$ s$^{-1}$\] \[photons cm$^{-2}$ s$^{-1}$\]
2014-01-27 56684.23 0.43 47-50 $1.03\times10^{-11}$ $-$
2014-01-28 56685.06 1.41 40-43 $2.19\times10^{-12}$ $1.55\times10^{-12}$
2014-01-29 56686.09 1.30 40-42 $4.55\times10^{-12}$ $5.97\times10^{-13}$
2014-02-01 56689.07 0.98 40-42 $3.14\times10^{-12}$ $9.98\times10^{-13}$
2014-02-02 56690.08 1.30 40-42 $3.35\times10^{-12}$ $1.76\times10^{-12}$
Total $-$ 5.41 40-50 $1.30\times10^{-12}$ $4.10\times10^{-13}$
---------------- ----------- ------- ------------------ -------------------------------- --------------------------------
![Distribution of the squared angular distance, $\theta^{2}$, after 5.41 hours of observation between the reconstructed arrival direction of the gamma-ray candidate events and the position of the source in the camera (red empty circles). The $\theta^{2}$ distribution of the background events (black points) is also displayed. The vertical dashed line at $\theta^{2}$=0.02 deg$^{2}$ defines the expected signal region.[]{data-label="fig:theta2LE"}](./theta2plot.pdf){width="\hsize"}
![MAGIC daily integral ULs from the direction of SN 2014J for energies above 300 GeV (red circles) and 700 GeV (blue squares). The integral UL for energies above 700 GeV was not computed for the first night (MJD 56684) due to low statistics (see also Table \[tab:dailyuls\]). The horizontal black dashed line indicates zero flux level and the vertical green line indicates the day of the SN explosion (MJD 56678), just six days before the beginning of the MAGIC observations.[]{data-label="fig:dailyuls"}](./LC.pdf){width="\hsize"}
The MAGIC stereo system at the Observatorio del Roque de los Muchachos on the Canary island of La Palma, Spain (28.8${^\circ}$N, 17.8${^\circ}$ W, 2200 m a.s.l.), consists of two 17 m Imaging Air Cherenkov Telescopes (IACTs). The MAGIC telescopes reach one of the lowest trigger energy thresholds among current IACTs (50 GeV). The observations were carried out in stereoscopic mode, which means that only shower images seen in both telescopes are recorded and analyzed. This mode provides a sensitivity of 0.66 $\pm$ 0.03 % of the Crab Nebula flux in 50 hours of observation for energies above 220 GeV [@2016APh....72...76A].\
SN 2014J was observed under moderate moonlight conditions from January 27 to 29 and on February 1 and 2 under dark-night conditions at medium zenith angles (from 40${^\circ}$ to 52${^\circ}$). The MAGIC observations started six days after the first detection by the UCL Observatory because of the adverse weather conditions. The complete data set up to 50${^\circ}$ ($\sim5.5$ hours) was used for the analysis given the overall good quality of the data (concerning weather, light conditions and performance of the system).\
Figure \[fig:theta2LE\] shows the distribution of squared angular distance ($\theta^2$) between the reconstructed gamma-ray direction and the position of either SN 2014J (on-source histogram) or the center of the background control region (off-source histogram). The resulting excess of the on-source histogram over the background from the region, where gamma-ray events from SN 2014J are expected is compatible with zero excess. The significance computed using Eq. 17 of [@LiMa1983] is $0.90 \sigma$.
Upper limits (ULs) on the flux were computed following the @Rolke2005 method for 95% confidence level (CL), assuming a Gaussian background and a systematic uncertainty of 30% on the effective area of the instrument [@Albert2008]. The spectrum that we assumed was a power-law function, d$F/$d$E \propto E^{-\Gamma}$, with a spectral index of 2.6. Variations of $\sim$20% in the spectral index produced changes in the integral ULs of less than 5%, and hence small deviations from the used spectral index do not critically affect the reported ULs. The ULs above 300 GeV and 700 GeV for the single-night observations are reported in Table \[tab:dailyuls\] and depicted in Figure \[fig:dailyuls\].
After $\sim5.5$ hours of observations with the MAGIC telescopes, we establish an integral UL on the gamma-ray flux for energies above 300 GeV of $1.3\times10^{-12}$ photons cm$^{-2}$s$^{-1}$ at 95% CL, which corresponds to $\sim 1.0 \%$ in units of the Crab Nebula flux (CU) in the same energy range. For energies above 700 GeV, the integral UL is $4.1\times10^{-13}$ photons cm$^{-2}$s$^{-1}$, corresponding to $\sim 1.1 \%$ CU at the same CL. Our ULs for $E>700$ GeV are already close to the flux from the host galaxy M82 measured by VERITAS in the same energy range, $(3.7\pm0.8_{stat}\pm0.7_{syst})\times10^{-13}$ photons cm$^{-2}$ s$^{-1}$ [@2009Natur.462..770V], which constitutes an irreducible background for our measurement. Under the hypothesis that M82 has a gamma-ray spectrum of d$F/$d$E = 3\times 10^{-16} (E/1000$ GeV$)^{-2.5}$ photons cm$^{-2}$ s$^{-1}$ GeV$^{-1}$, as measured by VERITAS, the expected number of excess events in our observations would be 9.4, with a 95% CL lower limit at -6.1. The observed number of excess events by MAGIC is -4.2, with an associated p-value of $8.4\times10^{-2}$, hence consistent with the VHE flux of M82 measured by VERITAS.
Discussion
==========
VHE gamma rays are typically secondary products of particle acceleration to $\gtrsim$ TeV energies, either in hadronic or leptonic processes [@2013APh....43...71A]. The primary particles of such processes can either be protons or electrons. In the first case, the gamma-ray emission is produced by neutral pions from the inelastic collisions between the protons accelerated in the SN and the ambient atomic nuclei. In the latter case, the gamma rays result from inverse Compton (IC) process of the accelerated electrons on the ambient photons. In both cases, the environment plays an important role for the production of VHE gamma-ray radiation. A near and young supernova ($\sim1$ week old) emitting in this energy regime could shed light on the progenitors of these thermonuclear stellar explosions.\
Although we did not detect VHE gamma rays right after the explosion, using the known distance of M82 [$d_\mathrm{M82} = 3.6$ Mpc @karachentsev_masses_2006] and assuming, as before, a spectral index of 2.6, one can convert the measured flux UL into an UL on the power emitted into VHE gamma rays. Therefore, given the integral UL for energies greater than 300 GeV, $1.3\times10^{-12}$photons cm$^{-2}$s$^{-1}$, the resulting UL on the power emitted is of the order of $10^{39}$ erg s$^{-1}$. If one now assumes an emission period of the order of 10 days, the total energy emitted in VHE gamma rays during this period is smaller than $10^{45}$ erg, which is about $10^{-6}$ of the total available energy budget of the SN explosion.\
Models of the evolution of young supernova remnants (SNRs) can be used to estimate the expected emission from the region in the future. One of the most important parameters to be assumed is the density profile of the SN ejecta. In this work, we considered a simple power-law density profile, which allows us to use the [@Dwarkadas:2013nwa] model to obtain an analytic solution for the estimated flux. Other density profiles have been used in the literature: Models like W7 or WDD1, applied by [@Nomoto] and [@1999ApJS..125..439I], are usually utilized in Type Ia SN studies, but they are based on the single-degenerate scenario. [@1998ApJ...497..807D] discussed a possible exponential density profile, which could represent better the SN ejecta structure than the power-law one. However, the exponential profile cannot provide an analytic result as the one assumed in this work, which can give a correct solution within the order of magnitude, as explained in [@Dwarkadas:2013nwa].\
Thus, making use of Eq. 10 in [@Dwarkadas:2013nwa], one can obtain the time-dependent emission assuming a hadronic origin. Considering only this hadronic origin, we can establish a lower limit on the total gamma-ray radiation. As discussed above, gamma-ray emission can be also expected from IC processes. Nevertheless, purely leptonic scenarios have been studied and discarded by several authors, e.g. .\
The expected flux depends strongly not only on the assumed density structure of the SNR but also on the density profile of the surrounding interstellar medium (ISM). As shown by different authors, we can consider double-degenerate scenario in the case of SN 2014J, i.e. two WDs progenitors. WDs do not suffer wind-driven mass-loss and therefore, they are not expected to modify the surrounding medium. Nevertheless, different assumptions, from the lack of certainty on the progenitors, have also been studied (see e.g. @2000ApJ...541..418D). We can then assume that the Type Ia SN explosion took place in a uniform density medium. In this work, we used a density of $n=2.2\times10^{-24}$ g/cm$^{3}$, based on [@pereztorres2014], assuming that all the content in the host galaxy of our source, M82, stems from the neutral hydrogen, H$_{\text{I}}$. This homogeneous medium assumption leads to an increasing flux emission, above a certain gamma-ray energy, with time in the free-expansion SNR stage, as shown below in the expression given by [@Dwarkadas:2013nwa]:
$$\label{eq:Dwarkadasflux}
F_{\gamma}(>1 \text{TeV},t)=\frac{3q_{\gamma}\xi(\kappa C_{1})^{5}m^{3}}{6(5m-2)\beta\mu m_{p}d^{2}}n^{2}t^{5m-2}$$
where the assumed parameters in this work are
- $q_{\gamma} = 1\times10^{-19}$ cm$^{3}$s$^{-1}$erg$^{-1}$H-atom$^{-1}$ (for energies greater than 1 TeV) is the emissivity of gamma rays normalised to the cosmic ray energy density tabulated in [@Drury1994]. This value corresponds to a spectral index of 4.6 of the parent cosmic ray distribution, which was selected according to the assumed spectral index in this work, $\Gamma=2.6$
- $\xi = 0.1$ is the fraction of the total SN explosion energy converted to cosmic ray energy, so an efficient cosmic ray acceleration is assumed
- $\kappa = 1.2$ is the ratio between the radius of the forward shock and the contact discontinuity (which separates ejecta and reverse shock)
- $C_{1} = 1.25\times10^{13}$cm/s$^{m}$ is referred to as a constant related to the kinematics of the SN. This value is calculated from the relation given by [@Dwarkadas:2013nwa], $R_{shock}=\kappa C_{1}t^{m}$. In turn, $R_{shock}$ is obtained from Eq. 2 in [@Gabici2016], by assuming an explosion energy of $10^{51}$ erg, a mass of the ejecta of $1.4$ M$_{\odot}$ and a ISM density of $1.3$ cm$^{-3}$, whose value is constrained by [@pereztorres2014]
- $\beta = 0.5$ represents the volume fraction of the already shocked region from which the emission arises
- $\mu = 1.4$ is the mean molecular weight
- $m_{p} = 1.6\times10^{-24}$g is the proton mass
- $d = 3.6$ Mpc is the distance to our source
- $t$ is the elapsed time since the explosion, and
- $m$ is the expansion parameter.
The expansion parameter varies along the free-expansion phase in different ways according to the assumed model for the density structure of the SN ejecta after the explosion. In this work, we make use of the power-law profile with a density profile proportional to $R^{-7}$ [@1982ApJ...258..790C], where $R$ is the outer radius of the ejecta. The initial value of the expansion parameter is very unalike depending on the density profile assumed, but in all cases evolve to $m=0.40$ [@1998ApJ...497..807D]. This limit at 0.40 is constrained by the beginning of the Sedov-Taylor phase.\
The expansion parameter for the power-law profile keeps constant at 0.57 in the first years of the free-expansion stage. Given this value, the expected flux above 1 TeV (constrained by the emissivity of gamma rays, $q_{\gamma}$, tabulated in @Drury1994) at the time of the MAGIC observations ($t=6$ days) from Eq. \[eq:Dwarkadasflux\] is $\approx 10^{-24}$ photons cm$^{-2}$s$^{-1}$. This flux is consistent with the UL at 95% CL derived from MAGIC data in the same energy range, $2.8\times10^{-13}$ photons cm$^{-2}$s$^{-1}$ and hence, the power-law density profile could be considered as possible model to describe the density structure of SN 2014J, considering all the assumptions and parameters selection discussed above. On the other hand, this model predicts a constant parameter of $m=0.57$ during the first $\sim300$ years, after which it starts dropping gradually [@1998ApJ...497..807D]. Although the flux keeps increasing with time according to Eq. \[eq:Dwarkadasflux\], with this low expansion parameter it will still be about $10^{-21}$ photons cm$^{-2}$s$^{-1}$ 100 years after the SN occurred, which is well below the sensitivity of the current and planned VHE observatories (several orders of magnitude below the sensitivity that the future Cherenkov Telescope Array, CTA[^2], will reach).
Conclusions
===========
The MAGIC telescopes performed observations of the nearest Type Ia SN in the last decades, SN2014J. No VHE gamma-ray emission was detected. Integral ULs for energies above 300 GeV and 700 GeV were established at $1.3\times10^{-12}$ photons cm$^{-2}$s$^{-1}$ and $4.1\times10^{-13}$ photons cm$^{-2}$s$^{-1}$, respectively, for a 95% CL and assuming a power-law spectrum. The flux UL at $E>\SI{300}{\GeV}$ corresponds to an emission power of $< 10^{39}$ erg s$^{-1}$ or a total maximal emitted VHE gamma-ray energy during the observational period – approximately ten days – of $< 10^{45}$ erg, which is about $10^{-6}$ times the total energy budget of a Type Ia SN explosion ($\sim 10^{51}$ erg). Following [@Dwarkadas:2013nwa] model for hadronic gamma-ray flux, a power-law density profile proportional to $R^{-7}$ is consistent with our ULs, although, due to the uncertainties in several parameters, this cannot exclude other, more sophisticated, theoretical scenarios. Assuming this SN density profile and a constant density medium, we can estimate an expected emission from the region of the source of $\approx 10^{-24}$ photons cm$^{-2}$s$^{-1}$. Following these assumptions, this flux would not increase enough in a near future to be detectable by any current or future generation of IACTs, as CTA.
We would like to thank the Instituto de Astrofísica de Canarias for the excellent working conditions at the Observatorio del Roque de los Muchachos in La Palma. The financial support of the German BMBF and MPG, the Italian INFN and INAF, the Swiss National Fund SNF, the ERDF under the Spanish MINECO (FPA2015-69818-P, FPA2012-36668, FPA2015-68278-P, FPA2015-69210-C6-2-R, FPA2015-69210-C6-4-R, FPA2015-69210-C6-6-R, AYA2013-47447-C3-1-P, AYA2015-71042-P, ESP2015-71662-C2-2-P, CSD2009-00064), and the Japanese JSPS and MEXT is gratefully acknowledged. This work was also supported by the Spanish Centro de Excelencia “Severo Ochoa” SEV-2012-0234 and SEV-2015-0548, and Unidad de Excelencia “María de Maeztu” MDM-2014-0369, by grant 268740 of the Academy of Finland, by the Croatian Science Foundation (HrZZ) Project 09/176 and the University of Rijeka Project 13.12.1.3.02, by the DFG Collaborative Research Centers SFB823/C4 and SFB876/C3, and by the Polish MNiSzW grant 745/N-HESS-MAGIC/2010/0.
[^1]: Corresponding authors: A. Fernández-Barral, , C. Fruck,
[^2]: as shown in https://www.cta-observatory.org/science/cta-performance/
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Conjecturally, almost all graphs are determined by their spectra. This problem has also been studied for variants such as the spectra of the Laplacian and signless Laplacian. Here we consider the problem of determining graphs with Ihara and Bartholdi zeta functions, which are also computable in polynomial time. These zeta functions are geometrically motivated, but can be viewed as certain generalizations of characteristic polynomials. After discussing some graph properties determined by zeta functions, we show that large classes of cospectral graphs can be distinguished with zeta functions and enumerate graphs distinguished by zeta functions on $\le 11$ vertices. This leads us to conjecture that almost all graphs which are not determined by their spectrum are determined by zeta functions.
Along the way, we make some observations about the usual types of spectra and disprove a conjecture of Setyadi and Storm about Ihara zeta functions determining degree sequences.
address:
- 'Department of Mathematics, University of Oklahoma, Norman, OK 73019'
- 'Department of Mathematics, University of Oklahoma, Norman, OK 73019'
author:
- Christina Durfee
- 'Kimball Martin$^\ast$'
title: Distinguishing graphs with zeta functions and generalized spectra
---
[^1]
Introduction
============
A fundamental problem in spectral graph theory is: when can we distinguish (unlabeled simple) graphs by their spectra? If the answer were always, as was once conjectured, that would mean we could solve the graph isomorphism problem in polynomial time. However, many pairs of non-isomorphic cospectral graphs have since been found, and various constructions of cospectral pairs are known. Nevertheless, Haemers conjectured that almost all graphs are determined by their spectra (DS), i.e., the fraction of graphs of order $n$ which are DS goes to 1 as $n \to \infty$. In fact, from the numerical data for $n \le 11$ in Haemers–Spence [@HS], it appears that more graphs are determined by their Laplacian spectra ($L$-DS), and even more by their signless Laplacian spectra ($|L|$-DS). We refer to surveys by van Dam and Haemers for more details [@vDH1], [@vDH2].
This question of which graphs are DS can be thought of geometrically. Knowing the adjacency spectrum of $G$ is equivalent to knowing what one might call the [*walk length spectrum*]{} of $G$—the set $\{ (\ell, w_G(\ell)) : \ell \ge 0 \}$, where $w_G(\ell)$ is the number of closed walks of length $\ell$ on $G$—together with the order $n$ of $G$. Hence the above question can be stated: when do the order and walk length spectrum determine $G$?
From the point of view of Riemannian geometry and covering space theory, it is more natural to look at [*geodesics*]{} on $G$ rather than arbitrary walks. Roughly, geodesics are paths with no backtracking and they correspond to lines in the universal cover of $G$ (see Section \[sec:ihara\]). The [*geodesic length spectrum*]{} of $G$ is then the set $\{ (\ell, a_G(\ell)) : \ell \ge 0 \}$ of numbers of (primitive) closed geodesics of length $\ell$ for all $\ell$. For graphs, the Ihara zeta function $Z_G(t)$ of $G$ encodes the geodesic length spectrum, and knowing one is equivalent to knowing the other.
This is entirely analogous to the Selberg zeta function encoding the (geodesic) length spectrum for Riemann surfaces. For compact Riemann surfaces, Huber’s theorem says that knowing the Selberg zeta function, i.e., the length spectrum, is equivalent to knowing the spectrum of the Laplacian (see, e.g., [@buser]). Similarly, for connected regular graphs, knowing the Ihara zeta function is equivalent to knowing the (adjacency or Laplacian) spectrum (see Section \[sec:ihara-prop\]). However, this is not the case for irregular graphs (cf. Table \[tab:md2\]).
Here, we suggest that the Ihara zeta function—which is computable in polynomial time—provides a more effective way to differentiate (irregular) graphs than the usual spectra studied (Conjecture \[conj:main\]). One heuristic for why this should be the case is that geodesics capture much of the geometry of the graph better than arbitrary paths. Another is that the Ihara zeta function typically encodes more information—at least for md2 graphs (graphs with no vertices of degree $\le 1$), the reciprocal of the Ihara zeta function is a polynomial whose degree is twice the number of edges, and thus typically has more coefficients than the characteristic polynomial.
The obvious drawback of the Ihara zeta function is that it cannot detect leaves or isolated nodes. For this reason (at least in part), most studies of the Ihara zeta function restrict to md2 graphs. Most studies also restrict to connected graphs, which is convenient to discuss covering space theory, but this is less crucial for the problem of distinguishing graphs. (As with characteristic polynomials, the zeta function of a disconnected graph is the product of the zeta functions of its components.) Instead of making such restrictions here, we consider four methods of addressing this defect to distinguish graphs (undefined notation explained subsequently):
1. Use the order and the Ihara zeta $Z_{G^*}(t)$ function of the cone $G^*$ of $G$. Equivalently use the order and the characteristic polynomial ${\varphi}_{T^*}(\lambda) = \det(\lambda I - T^*)$.
2. Use the order, size and the Ihara zeta functions $Z_G(t)$ and $Z_{\bar G}(t)$ of $G$ and its complement $\bar G$. Equivalently, use the order and the characteristic polynomials ${\varphi}_{T}(\lambda) = \det(\lambda I - T)$ and ${\varphi}_{\bar T}(\lambda) = \det(\lambda I - \bar T)$.
3. Use the order and the Bartholdi zeta function ${\mathcal{Z}}_G(t,u)$ of $G$. Equivalently, use the generalized characteristic polynomial ${\varphi}_{AD}(\lambda, x) = \det(\lambda I - A + xD)$.
4. Use the order and the Bartholdi zeta functions ${\mathcal{Z}}_G(t,u)$ and ${\mathcal{Z}}_{\bar G}(t,u)$ of $G$ and $\bar G$. Equivalently, use the generalized characteristic polynomial ${\varphi}_{ADJ}(\lambda, x, y) = \det(\lambda I - A + xD + yJ)$.
Here $I$ and $J$ denote the identity and all-ones matrices of the appropriate sizes, and $A$ and $D$ denote the adjacency and degree matrices for a fixed ordering of vertices on $G$. Further, $T$ (resp. $T^*$, $\bar T$) denotes Hashimoto’s oriented edge matrix (see Section \[sec:ihara\]) of $G$ (resp. $G^*$, $\bar G$)—whose spectrum contains the same information as $Z_G$ together with the size of $G$ (cf. ). The two-variable Bartholdi zeta function ${\mathcal{Z}}_G(t,u)$ is a generalization of the one-variable Ihara function $Z_G(t): {\mathbb{C}}\to {\mathbb{C}}$ which counts “closed geodesics with $r$ backtracks”—see Section \[sec:bartholdi\], in particular Equation for the relation with ${\varphi}_{AD}$.
Note that knowing ${\varphi}_{AD}$ implies knowing the spectra of $A$, the Laplacian $L$, and the signless Laplacian $|L|$. Similarly knowing ${\varphi}_{ADJ}$ implies knowing all of these spectra for both $G$ and $\bar G$.
We will explain the precise motivation behind these choices as we consider each in turn in the body of the paper. The basic ideas are that the Ihara zeta function cannot detect “dangling edges,” i.e., edges to a degree 1 node, but (i) taking cones or complements turns degree 1 nodes into higher degree nodes and (ii) the Bartholdi zeta function does detect dangling edges. In analogy with the DS terminology, we say that a graph $G$ is DZ\* (resp. [DZ$\bar{\textrm{Z}}$]{}, [D${\mathcal{Z}}$]{}, [D${\mathcal{Z}}\bar {\mathcal{Z}}$]{}) if it is uniquely determined by method (M1) (resp. (M2), (M3), (M4)).
We will see in Section \[sec:M4\] that (M4) is provably stronger than each of (M1), (M2) and (M3), i.e., being DZ\* (resp. [DZ$\bar{\textrm{Z}}$]{}, [D${\mathcal{Z}}$]{}) implies [D${\mathcal{Z}}\bar {\mathcal{Z}}$]{}. However there are no obvious implications among (M1), (M2) and (M3) (but there are some indications that (M2) may be stronger than (M1)—see the end of Section \[sec5\]). Nevertheless, a surprising outcome of our calculations in Table \[tab:main\] is that for graphs on $n \le 11$ vertices, methods (M1), (M2) and (M4) are always equivalent (whereas (M3) has slightly less discriminating power). This suggests one gains little by using the Bartholdi zeta function, and one is in practice justified in just considering the Ihara zeta function, which is both conceptually and computationally simpler. However, all of these methods run in polynomial time.
Now we outline the contents of the paper.
In Section \[sec:ihara\], we first recall basic facts about the Ihara zeta function, and discuss some graph properties it determines. Then we consider some basic properties determined by methods (M1) and (M2). In particular, these zeta invariants force strong restrictions on the degree sequence of a graph. Setyadi–Storm [@SS], in their enumeration of pairs of connected md2 graphs with the same Ihara zeta function on $n \le 11$ vertices, found that the Ihara zeta function of a connected md2 graph determines the degree sequence for $n \le 11$, and conjectured this holds for all $n$, but we give a counterexample to this conjecture on 12 vertices (Example \[ex:crabsquid\]). Nevertheless, we show that knowing the zeta functions of sufficiently many cones of $G$ algorithmically determines the degree sequence (Lemma \[lem:ds\]).
In Section \[sec:bartholdi\], we recall basic facts about the Bartholdi zeta function and discuss the relative strengths of methods (M1)–(M4). To show that (M4) is stronger than (M1), we show that two graphs have the same spectra with respect to $A+xD$ and $\bar A+ x\bar D$ for fixed $x$, if and only if the same is true of their joins with another graph (Theorem \[thm:cone-comp\]). In particular, two graphs have the same $A$- and $\bar A$- (or $|L|$- and $\overline{|L|}$-) spectra if and only if the same is true for their cones (Corollary \[cor:cone-comp\]; cf. Table \[tab:cone-comp\] for related data).
In Section \[sec:constructions\], we show that methods (M1)–(M4) distinguish a large class of cospectral and Laplacian cospectral pairs coming from well known constructions: GM switching, coalescence and join. On the other hand, these methods will not distinguish graphs arising from the more restrictive GM\* switching. Section \[sec:new-con\] presents a new construction of pairs of graphs which cannot be distinguished by these methods. This construction is interesting because it constructs graphs $G_1$, $G_2$ such that the generalized adjacency matrices $A_1 + xD_1$ and $A_2 + xD_2$ are miraculously conjugate for all $x$ but not “uniformly conjugate” (as happens in GM\* switching) i.e., there is no invertible matrix $P$ such that $P(A_2+xD_2)P^{-1} = A_1 + xD_1$ for all $x$.
In Section \[sec5\], we enumerate all graphs on $\le 11$ vertices which are not DZ\* (resp. [DZ$\bar{\textrm{Z}}$]{}, [D${\mathcal{Z}}$]{}, [D${\mathcal{Z}}\bar {\mathcal{Z}}$]{}) in Table \[tab:main\]. We state Conjecture \[conj:main\], which asserts almost all graphs are DZ\* (resp. [DZ$\bar{\textrm{Z}}$]{}, [D${\mathcal{Z}}$]{}, [D${\mathcal{Z}}\bar {\mathcal{Z}}$]{}), and further that almost all non-DS graphs are DZ\* (resp. [DZ$\bar{\textrm{Z}}$]{}, [D${\mathcal{Z}}$]{}, [D${\mathcal{Z}}\bar {\mathcal{Z}}$]{}). Table \[tab:md2\] indicates that there is no essential difference in just using the Ihara zeta function $Z_G$ compared to methods (M1)–(M4) when restricting to md2 graphs. This suggests Conjecture \[conj:md2\], which is the analogue of Conjecture \[conj:main\] for determining md2 graphs $G$ using only $Z_G$. We also compare the effectiveness of combining different kinds of spectra for $n \le 11$ in Table \[tab:combine\]. In particular, this suggests that using two of the usual spectra, such as $A$ and $L$ or $A$ and $|L|$, is much more effective at distinguishing graphs than any single one.
Our calculations were done using Sage [@sage], including nauty [@nauty], and standard Unix tools. In practice, we found that to check if two graphs have the same zeta function, it almost always sufficed to compute the two numbers $\det |L|$ and $\det(4D+2A-3I)$, which are essentially (the residue of) $Z_G(1)$ and $Z_G(-2)$. For convenience of the interested reader, when mentioning particular examples of graphs, we include (non-canonical) [`graph6`]{} strings to specify the graphs, which can be used to easily reconstruct the graphs in Sage.
As a final remark, we note that there are more general notions of zeta functions of graphs, such as path and edge zeta functions (see [@ST] or [@terras]). We do not consider these here.
We thank C. Storm for discussions about [@SS], as well as the referee for helpful comments.
The Ihara zeta function {#sec:ihara}
=======================
We begin by defining closed geodesics and the Ihara zeta function. This notion of a geodesic on a graph corresponds to a bi-infinite simple path in the universal cover, but we do not explain this here—see, e.g., [@ST] or [@terras]. Ihara zeta functions are a special case of more general zeta functions of multigraphs considered by Hashimoto [@hashimoto], generalizing the zeta functions originally defined by Ihara [@ihara].
Fix a finite (simple) graph $G = (V, E)$. Let $n = |V|$ and $m = |E|$. We will denote walks by sequences of adjacent oriented (or directed) edges $e_i$. For an oriented edge $e_i = (u, v)$, let $e_i^{-1} = (v,u)$ be the edge with reversed orientation. Suppose $\gamma = (e_1, e_2, \ldots, e_\ell)$ denotes a closed walk of length $\ell$ in $G$. We say $\gamma$ is a [*closed geodesic*]{} if $e_i \ne e_{i+1}^{-1}$ for $1 \le i < \ell$ and $e_1 \ne e_{\ell}^{-1}$. The former condition is often expressed saying $\gamma$ has no backtracking and the latter that $\gamma$ has no tails. Write $k\gamma$ for the concatenation of $k$ copies of $\gamma$. We say $\gamma$ is [*primitive*]{} if $\gamma$ is not of the form $k \delta$ for a closed geodesic $\delta$ and some $k \ge 2$.
Let $\sigma(\gamma) = (e_2, \ldots, e_\ell, e_1)$. Then $\sigma(\gamma)$ is also a closed geodesic of length $\ell$, and it is primitive if $\gamma$ is. Thus the cyclic group $\langle \sigma \rangle$ of order $\ell$ generated by $\sigma$ acts on all (primitive) closed geodesics of length $\ell$. The $\langle \sigma \rangle$-orbits thus partition the set of (primitive) closed geodesics of length $\ell$ into equivalence classes. Let $a(\ell)=a_G(\ell)$ denote the number of primitive closed geodesics of length $\ell$ up to equivalence.
Note that, as $G$ is simple, there are no closed geodesics of length $< 3$. The closed geodesics of length 3, 4 or 5 are just the cycles of length 3, 4 or 5. The closed geodesics of length 6 are just the cycles of length 6 together with the concatenations of 2 simple cycles of length 3 starting at a fixed base point. If there are two distinct cycles based at a vertex $v_0$, going around the first cycle $k$ times and going around the second cycle once or more is a primitive closed geodesic. Hence, for connected md2 graphs, the lengths of primitive closed geodesics are unbounded unless $G$ is a circuit.
The (Ihara) zeta function of $G$ is $$\label{eq:zeta-def}
Z_G(t) = \prod_\gamma (1-t^{\ell(\gamma)})^{-1}
= \prod_{\ell > 2} (1-t^{\ell})^{-a(\ell)} = \exp \left( \sum_{\ell > 2} \sum_{k \ge 1} a(\ell) \frac{t^{\ell k}}k \right),$$ where, in the first product, $\gamma$ runs over a set of representatives for the equivalences classes of primitive closed geodesics, and $\ell(\gamma)$ denotes the length of $\gamma$. Since no closed geodesics will involve degree 0 or degree 1 nodes, $Z_G(t) = Z_{G^{\dagger}}(t)$, where $G^{\dagger}$ is the “pruned graph” obtained by successively deleting degree 0 and degree 1 nodes until one is either left with an md2 graph or the null graph (the graph on 0 vertices).
Note if $G$ is a disjoint union of two subgraphs, $G = G_1 \sqcup G_2$, then $a_G(\ell) = a_{G_1}(\ell) + a_{G_2}(\ell)$. Hence the zeta function of a graph is the product of the zeta functions of its connected components.
The zeta function is [*a priori*]{} an infinite product, but turns out to be a rational function and thus is meromorphic on ${\mathbb{C}}$. Namely, Hashimoto [@hashimoto], [@hashimoto3] and Bass [@bass] gave two determinant formulas for $Z_G$, which have been subsequently retreated many times (e.g., [@ST]). The Bass determinant formula [@bass] is $$\label{eq:bass-det}
Z_G(t) = (1-t^2)^{n-m} \det(I - tA + t^2(D-I))^{-1}.$$ Note the right hand side is invariant under adding nodes of degree 0 or 1 to $G$.
Let $\{ e_1, \ldots, e_{2m} \}$ denote the set of oriented edges of $G$. The [*oriented edge matrix*]{} (with respect to this ordering of oriented edges) $T$ is the $2m \times 2m$ matrix whose $(i,j)$-entry is 1 if $e_i = (u, v)$, $e_j = (v, w)$ and $u \ne w$; or 0 otherwise. The Hashimoto determinant formula [@hashimoto], [@hashimoto3] (cf. [@ST Thm 3]) is $$\label{eq:hashimoto-det}
Z_G(t) = \det(I - tT)^{-1} = (t^{2m} {\varphi}_T(t^{-1}))^{-1},$$ where ${\varphi}_T(\lambda) = \det(\lambda I - T)$ is the characteristic polynomial of $T$. Again, one can check that $\det(I - tT)$ is invariant under adding degree 0 or 1 nodes to $G$.
Since $G^{\dagger}\ne G$ in general, $Z_G$ does not determine $m$ (or $n$). This means knowing $Z_G$ is not exactly the same as knowing the spectrum of $T$, but it almost is: ${\varphi}_T$ determines both $Z_G$ and $m$, and conversely. Indeed, the degree of ${\varphi}_T$ is $2m$, so one can recover $Z_G$ from ${\varphi}_T$. The converse is obvious. This observation relates to method (M2).
Properties determined by the Ihara zeta function {#sec:ihara-prop}
------------------------------------------------
Here we summarize some elementary graph properties determined by the Ihara zeta function. This question has previously been considered mainly for md2 and connected md2 multigraphs (e.g., [@czarneski] and [@cooper]).
First, examining the coefficients of $t^k$ in the logarithm of shows $Z_G$ determines $a(\ell)$ for all $\ell$, hence knowing $Z_G$ is the same as knowing the primitive geodesic length spectrum. Note that the number of non-primitive geodesics of a given length $l$ is the sum of the numbers of primitive geodesics of proper divisors of $l$. Hence the primitive geodesic length spectrum determines the full geodesic length spectrum and vice versa. Consequently $Z_G$ determines the number of cycles of length 3, length 4 and length 5 in $G$. It is not true that $Z_G$ determines the number of cycles of length 6.
The pair of graphs [`HheadXZ`]{} and [`Hhf@eS|`]{} pictured below, each with 9 vertices and 18 edges, have the same zeta functions but a different number of cycles of length 6 (46 and 50).
(0,0) rectangle (2.0cm,2.0cm); (v0)(v1) (v0)(v4) (v0)(v5) (v0)(v7) (v1)(v2) (v1)(v6) (v1)(v8) (v2)(v3) (v2)(v5) (v2)(v7) (v3)(v4) (v3)(v6) (v3)(v8) (v4)(v7) (v4)(v8) (v5)(v7) (v6)(v8) (v7)(v8)
(0,0) rectangle (2.0cm,2.0cm); (v0)(v1) (v0)(v4) (v0)(v5) (v0)(v7) (v1)(v2) (v1)(v5) (v1)(v7) (v2)(v3) (v2)(v6) (v2)(v8) (v3)(v4) (v3)(v6) (v3)(v8) (v4)(v7) (v4)(v8) (v5)(v8) (v6)(v7) (v7)(v8)
\[ex:9verts\]
We remark $Z_G(t) = 1$ if and only if $G$ has no cycles, so $Z_G$ can determine whether a graph is a forest or not. The girth $g$ of $G$ will be the smallest $\ell$ such that $a(\ell) \ne 0$. So $Z_G$ determines $g$ and, if $G$ is connected, Scott and Storm [@scott-storm] showed $Z_G$ determines the number of cycles of length $\ell$ for any $\ell < 2g$.
From it is clear that $Z_{G}(t)^{-1}$ is a polynomial in $t$ of degree $2m^{\dagger}$, where $m^{\dagger}$ is the size of $G^{\dagger}$. In general, though $Z_G$ does not determine $m$ or $n$, it does determine $m = m^{\dagger}$ if $G$ is md2. If we assume that $G$ is connected (but not necessarily md2), then we can say $Z_G$ determines $m-n$. Namely, if $Z_G(t)=1$ then $G$ is a tree and $m-n=-1$. Otherwise $G^{\dagger}$ is not the null graph so $m-n=m^{\dagger}-n^{\dagger}$, and $m^{\dagger}$ and $n^{\dagger}$ are determined by $Z_G$, as we will see below. Note that $Z_G$ does not determine $m-n$ if we drop the assumption that $G$ is connected, since all forests have trivial zeta function.
To get clean statements, for the rest of this section assume $G$ is md2 and connected. For general connected $G$, the results here can be viewed as results about $G^{\dagger}$ when $G^{\dagger}$ is not the null graph. Note that if $G$ is connected, then $G^{\dagger}$ is also connected.
In this case $m \ge n$, and Hashimoto [@hashimoto] showed (i) $Z_G(t)$ has a pole at $t=1$, which is of order $2= m-n+2$ if $G$ is a circuit and $m-n+1$ otherwise, and (ii) $Z_G(t)$ has a pole at $t=-1$ of order $m-n$, $m-n+1$ or $m-n+2$ according to whether $G$ is non-bipartite, a bipartite non-circuit or a bipartite circuit. Hashimoto [@hashimoto2] later showed the residue at $t=1$ for non-circuits is $2^{m-n+1}(n-m) \kappa(G)$, where $\kappa$ denotes complexity of $G$, i.e., the number of spanning trees of $G$. Note that $G$ is bipartite if and only if its geodesics all have even length, and recall $G$ is a circuit if and only if $a(\ell) = 0$ for $\ell$ sufficiently large.
Consequently, $Z_G$ determines $n$, $m$, $\kappa(G)$, whether $G$ is bipartite, and whether $G$ is a circuit. In fact, since $\det(I-tA + t^2(D-I))$ is simply $\det(D+A)=\det |L| $ at $t=-1$, tells us $\lim_{t \to -1} (1-t^2)^{n-m} Z_G(t)^{-1} = \det |L|$, which is 0 if and only if $G$ has a bipartite component. So $Z_G(t)$ also determines the product of eigenvalues of $|L|$.
Cooper [@cooper] also showed $Z_G$ determines whether $G$ is regular, and if so, the degree of regularity as well as the spectrum. In fact if $G$ is $(q+1)$-regular, then the spectrum conversely determines $Z_G$ by $$Z_G(t) = (1-t^2)^{n-m} \prod_i (1- \lambda_i t + q t^2)^{-1}$$ where $\{ \lambda_1, \ldots, \lambda_n \}$ are the eigenvalues of $A$ (attributed to A. Mellein [@czarneski]).
Write $V = \{ v_1, \ldots, v_n \}$, $d_i = \deg(v_i)$ and $q_i = d_i - 1$. We see the leading term of $Z_G(t)^{-1}$ is $(\prod q_i) t^{2m}$. Consequently, for $G$ md2, $Z_G(t)$ determines $\prod q_i$. One also knows $2m = \sum d_i$, so knowing $Z_G(t)$ places rather strong restrictions on the degree sequence for md2 graphs.
In fact, Setyadi–Storm [@SS], based on their enumeration of zeta functions of connected md2 graphs on $\le 11$ vertices, conjectured that connected md2 graphs with the same Ihara zeta function have identical degree sequences. However, we give a counterexample.
The graphs [`K??CA?_FEcdk`]{} and [`K??CA?_ccWNk`]{} on 12 vertices and 16 edges drawn below (“the crab and the squid”) have the same zeta function, but their degree sequences are (5, 5, 4, 2, 2, 2, 2, 2, 2, 2, 2, 2) and (7, 3, 3, 3, 2, 2, 2, 2, 2, 2, 2, 2).
(0,0) rectangle (2.0cm,2.0cm); (v0)(v6) (v0)(v10) (v1)(v7) (v1)(v10) (v2)(v8) (v2)(v11) (v3)(v9) (v3)(v10) (v4)(v9) (v4)(v11) (v5)(v9) (v5)(v11) (v6)(v10) (v7)(v11) (v8)(v11) (v9)(v10)
(0,0) rectangle (2.0cm,2.0cm); (v0)(v6) (v0)(v9) (v0)(v10) (v1)(v7) (v1)(v11) (v2)(v8) (v2)(v11) (v3)(v9) (v3)(v11) (v4)(v10) (v4)(v11) (v5)(v10) (v5)(v11) (v6)(v9) (v7)(v11) (v8)(v11)
\[ex:crabsquid\]
Zeta functions of cones—(M1)
----------------------------
The most obvious issue of using zeta functions to distinguish arbitrary graphs is that no closed geodesics will pass through “dangling links,” i.e., paths not contained in cycles. To resolve this issue, an obvious thing to try is connecting all the degree 1 nodes to a new vertex. However, doing this to just the degree 1 nodes is not a nice operation on graphs (it is not injective), so it makes more sense to look at the [*cone*]{} $G^*$ of $G$. This is just the join of $G$ with a point: $G^* = G \vee K_1$, where $\vee$ denotes the operation of graph join. The new vertex in $G^*$ is denoted $v_{n+1}$.
Clearly $G^*$ has order $n+1$ and increases the degree of each vertex in $G$ by 1. That is, $G^*$ has degree sequence $(d_i^*)$ where $d_i^* = d_i+1$ for $1 \le i \le n$ and $d_{n+1}^* = n$. Hence if $G$ is an md1 graph (a graph with no degree 0 vertices), then $G^*$ is an md2 graph. A graph $H$ on $n+1$ vertices will be the cone of some graph $G$ if and only if there is a vertex of degree $n$, and $G$ can be recovered by deleting any vertex of degree $n$.
We propose to use $Z_{G^*}$ to study $G$. Now any edge in $G$ will appear in some closed geodesic in $G^*$, so it is reasonable to expect that $Z_{G^*}$ encodes much of the structure of $G$. Of course, degree 0 nodes in $G$ become degree 1 nodes in $G^*$ and still are not detected by $Z_{G^*}$. If one wanted to, say, count the degree 0 nodes only using zeta functions, one could also look at the zeta function of the double cone $G^{**} = G \vee K_2$. Instead, we will simply fix $n$, and see how to determine the number of degree 0 nodes from $n$ and $Z_{G^*}$.
Hence, for the rest of this section, we will assume we know the order $n$ of $G$ (except where stated otherwise) and see what can be deduced about $G$ from $Z_{G^*}$.
We first show that we can determine the number of degree 0 nodes from $n$ and $Z_{G^*}$. Let $G'$ be the graph obtained from $G$ by removing all degree 0 nodes. Note $G$ is the empty graph on $n$ vertices if and only if $Z_{G^*} = 1$. Otherwise, $(G^*)^{\dagger}= (G')^*$ and $(G')^*$ is connected md2, so $Z_{G^*}$ determines the number of vertices and edges for $(G')^*$ and hence also for $G'$. Combining this with knowing $n$, we can determine $G$ from $G'$. Hence in considerations below we may and will assume $G=G'$, i.e., $G$ has no degree 0 nodes, thus $G^*$ is connected md2.
From the previous section, we immediately see $Z_{G^*}$ determines the following: $n$, $m$, $\prod d_i$, and the number of triangles in $G$, which equals $\frac 12 a_{G^*}(3) - m$. The latter follows as a 3-cycle in $G^*$ corresponds to either a 3-cycle in $G$ or a triangle formed from an edge in $G$ with $v_{n+1}$. Similarly, $a_{G^*}(4) = a_{G}(4) + 2\sum \binom{d_i}{2}$, since the latter term is the number of directed paths of length 2 in $G$. In particular, $Z_{G^*}$ determines $a_G(4) + \sum d_i^2$. Also, one knows $a_{G^*}(5)$, which is $a_G(5)$ plus the number of directed paths of length 3 in $G$.
An elementary consequence of knowing $n$, $2m = \sum d_i$ and $\prod d_i$ is the following: if we fix $a, b$ and $c$ and know that each vertex has degree $a$, $b$ or $c$, then $Z_{G^*}$ determines the degree sequence. For general $G$, we remark that one can determine the degree sequence by looking at sufficiently many cones. Namely, let $G^{*(r)} = G \vee K_r$.
Let $G$ be a graph of possibly unknown order $n \ge 1$. There exists a finite number $r$ such that if $H_1, \ldots, H_{r}$ is any known sequence of graphs with distinct orders, then $Z_{G \vee H_1}, \ldots, Z_{G \vee H_{r}}$ determines the degree sequence of $G$. One may take $r$ to be at most the order of $G$ (or 3 for $G$ with order less than 3). In particular, $Z_{G^*}, \ldots, Z_{G^{*(r)}}$ determines the degree sequence of $G$. \[lem:ds\]
We will take $r$ to be at least 3, in order to guarantee that one of the $H_j$ has order at least 2 (we allow one of the $H_j$’s to be the null graph). We will assume that we know the degree sequences of the $H_j$ and describe how the degree sequence of $G$ can be computed from $Z_{G \vee H_1}, \ldots, Z_{G \vee H_{r}}$.
Let $H$ be one of the $H_j$’s of order $h\geq 2$. Then $G\vee H$ is connected and if $n\geq 2$ it is also md2. It follows that the order of $(G\vee H)^{\dagger}$ is less than $h+2$ if and only if $n=1$. Otherwise it has order $h+n$ and, in either case, we can find $n$ from $Z_{G\vee H}$. Assume from now on that $n\geq 2$. Then $Z_{G\vee H}$ determines the number of edges of $G\vee H$ from which we can compute the number of edges of $G$, which we will call $m$.
Let $x_i$ be the number of vertices of $G$ of degree $i$ for $0\leq i \leq n-1$. From $n$ and $m$ we have the two linear equations $\sum_{i=0}^{n-1} x_i = n$ and $\sum_{i=0}^{n-1} i x_i = 2m$. Let $h_j$ be the order of $H_j$ for $1\leq j \leq r$. First assume that each $h_j$ is at least 2 so $G\vee H_j$ are all connected md2 graphs. Then $Z_{G\vee H_j}$ determines the products of the degrees of $G\vee H_j$ minus 1. The degrees coming from vertices of $H_j$ are known and hence $Z_{G\vee H_j}$ will determine $\prod_{i=0}^{n-1} (i +h_j - 1)^{x_i}$, or equivalently, $\sum_{i=0}^{n-1} \ln(i+h_j-1)x_i$. Thus we have a linear system with $r+2$ equations and $n$ unknowns. The coefficient matrix $A$ of this linear system has $1, i, \ln(i+h_1 -1), \ln(i+h_2 - 1), ..., \ln(i+h_r - 1)$ as the entries of its $i$-th column, for $0\leq i \leq n-1$. If $\mathbf{c}= {\begin{bmatrix}}c_1 & c_2 & ... &c_{r+2}{\end{bmatrix}}$ is a vector with $\mathbf{c}A = \mathbf{0}$, then the function $f(t) = c_1 + c_2t+ c_3\ln(t + h_1-1) + c_4\ln(t+h_2-1)+...+ c_{r+2}\ln(t+h_{r}-1)$ has zeros at $0, 1, 2,...,n-1$. As the $h_j$’s are distinct, this function changes direction at most $r$ times and thus has at most $r+1$ zeros unless all $c_i$ are zero. If we take $r=n-2$ then the $c_i$ must all be 0 so $A$ is an invertible $n\times n$ matrix and the linear system has exactly one solution. If we drop the assumption that the $h_j$’s are all at least 2, then at worst we can ignore $Z_G$ and $Z_{G^*}$ and take $r=n$.
If we allow $n \ge 0$ in the above proposition, the order of $G$ can be determined using the $Z_{G\vee H_j}$’s, with one exception: if all the $H_j$’s are empty graphs, then the $Z_{G\vee H_j}$ will not distinguish between the null graph and the graph on one vertex. Also, using fewer than $n$ joins often suffices to determine the degree sequence of $G$. For instance, the proof shows $n-2$ works if all $h_j$’s are $\ge 2$. It also ignores the degree information gotten from $Z_{G^*}$ if one of the $H_j$ has order 1. If we also take into account that the $x_i$ must be nonnegative integers, this could significantly decrease the number of joins needed in many cases.
We also comment that for regular graphs, knowing $Z_{G^*}$ is equivalent to knowing the spectrum of $G^*$ [@BS].
Zeta functions of graphs and their complements—(M2)
---------------------------------------------------
We begin by thinking of another way to encode the dangling nodes or links in the zeta function. By a dangling node (link), we mean a vertex (edge) not contained in any cycle, i.e., any node (edge) not in $G^{\dagger}$.
First note that if we fix $n$, we know the number $n - n^{\dagger}$ of dangling nodes, since $Z_G$ determines the order $n^{\dagger}$ of $G^{\dagger}$. In particular, $n$ and $Z_G$ will tell us if $G$ is connected md2.
Recall that knowing the spectrum of $T$ gives slightly more information than just $Z_G$—it also tells us $m$. Since $Z_G$ tells us $m^{\dagger}$, knowing ${\varphi}_T$ tell us $m-m^{\dagger}$ which is the number dangling links. However, this tells us nothing about the structure of dangling links—e.g., ${\varphi}_T$ cannot distinguish among a cycle $C_n$ with leaves added, $C_n$ with a path attached and $C_n$ disjoint union a forest, provided the number of edges match.
If we want to somehow account for dangling nodes with zeta functions, we can also try looking at both $G$ and $\bar G$. Degree 0 and degree 1 vertices in $G$ now have degree $n-1$ and $n-2$ in $\bar G$, so for $n > 3$ they will now have degree at least 2. However, it is still possible that some of these vertices are dangling nodes in $\bar G$. Nevertheless, we can show the following.
\[prop:M2\] Let $G$ be a graph of order $n$. Then at least one of the following holds:
\(i) $G$ is determined by ${\varphi}_T$ and ${\varphi}_{\bar T}$ (or equivalently, by $Z_G$, $Z_{\bar G}$ and $m$); or
\(ii) for any vertex $v \in V$, we have $v \in G^{\dagger}$ or $v \in (\bar G)^{\dagger}$.
We remark this is not true if (i) is just replaced with “$G$ is determined by $Z_G$ and $Z_{\bar G}$.” For instance, let $G_1 = K_3 \sqcup \{ v \}$ and $G_2$ be the graph obtained from $G_1$ by adding an edge from $v$ to one vertex in $K_3$. Then $\bar G_1$ is a tree and $\bar G_2$ is a forest, so $Z_{G_1} = Z_{G_2} = Z_{K_3}$ and $Z_{\bar G_1} = Z_{\bar G_2} = 1$. Thus we cannot distinguish $G_1$ and $G_2$ by looking at their zeta functions and their complements’ zeta functions. Moreover, $v$ does not lie in $G_i^{\dagger}$ or $(\bar G_i)^{\dagger}$ for $i = 1, 2$.
Before the proof, we give some numerical evidence that (M2)—using ${\varphi}_T$ and ${\varphi}_{\bar T}$ (or $Z_G$, $Z_{\bar G}$ and $m$) to distinguish graphs of order $n$—is much better than just using $Z_G$ and $Z_{\bar G}$, or just $Z_G$ or ${\varphi}_T$. This data is presented in Table \[table1\]. The second column is the total number of graphs on $n$ nodes. Subsequent columns contain the number of graphs $G$ on $n$ nodes which are not determined respectively by $Z_G$, by $Z_G$ and $Z_{\bar G}$, by ${\varphi}_T$, and by both ${\varphi}_T$ and ${\varphi}_{\bar T}$.
$n$ \# graphs $Z$ $Z\bar Z$ $T$ $T \bar T$
----- ----------- -------- ----------- -------- ------------
2 2 2 2 0 0
3 4 3 2 0 0
4 11 8 4 4 0
5 34 23 8 15 0
6 156 94 22 75 0
7 1,044 534 68 449 0
8 12,346 4,889 312 4,297 0
9 274,668 76,807 350 68,708 2
: Number of small graphs not distinguished by zeta invariants
\[table1\]
The pair of graphs $(G_1, G_2)$ on 9 vertices with the same ${\varphi}_T$ and ${\varphi}_{\bar T}$ are the graphs with 18 edges pictured in Example \[ex:crabsquid\]. We remark that $G_1 \simeq \bar G_2$ in this example.
One explanation for why (M2) is so effective is that in most cases it restricts the problem of distinguishing arbitrary graphs with zeta functions to looking just at connected md2 graphs, where zeta functions give us a lot of information (cf. Section \[sec:ihara-prop\]). Similarly we can determine a lot about $G$ if $\bar G$ is connected md2. And if neither $G$ nor $\bar G$ is connected md2, this places strong constraints on $G$, as we will see in the proof.
We first treat a special case, where we can say something stronger.
Suppose $G$ is a forest of order $n$. Then at least one of the following holds:
\(i) $\bar G$ is connected md2, or
\(ii) ${\varphi}_T$ and ${\varphi}_{\bar T}$ determine $G$ among all graphs of order $n$.
By Table \[table1\], we see ${\varphi}_T$ and ${\varphi}_{\bar T}$ determine $G$ for at least $n \le 8$. Assume $n \ge 5$. Then $G$ has at most 1 vertex of degree $\ge n-2$. If it has none, $\bar G$ is connected md2, so assume it does. Then $G$ must be one of the following three graphs: a star graph on $n$ nodes, a star-like graph on $n$ nodes with exactly 1 node distance 2 from the hub (so the other $n-2$ non-hub nodes are adjacent to the hub), or a star graph on $n-1$ nodes union a point. The latter case can be distinguished from the previous two by counting edges. The former two can be distinguished by looking at the number of edges in $(\bar G)^{\dagger}$. Since $Z_G$ detects forests, ${\varphi}_T$ and ${\varphi}_{\bar T}$ determine these 3 graphs among all graphs of order $n$.
If $G$ or $\bar G$ is md2, then clearly $(ii)$ holds, so we may assume neither is md2. Switching $G$ and $\bar G$ if necessary, we may also assume $G$ is connected. In light of the lemma, we may further assume $G^{\dagger}= (V^{\dagger}, E^{\dagger})$ is not the null graph, and hence $G^{\dagger}$ has order $n^{\dagger}\geq 3$.
Let $H = (W, F)$ be the subgraph (which is a forest) of $G$ induced from $W=V-V^{\dagger}$. Note that, in $G$, no $w \in W$ can be adjacent to more than 1 vertex in $V^{\dagger}$. Also, in any connected component of $H$, there is exactly one vertex which is adjacent to a vertex in $V^{\dagger}$. Suppose that $H$ contains at least two vertices $w_1, w_2$. If $w_1$ and $w_2$ are not adjacent in $G$ then there exists $v\in V^{\dagger}$ such that neither $w_1$ nor $w_2$ is adjacent to $v$ in $G$ and hence $\bar G$ contains the triangle $w_1, w_2, v$. If $w_1$ and $w_2$ are adjacent in $G$, then at most one of them is adjacent to a vertex in $V^{\dagger}$ so there exist $v_1, v_2 \in V^{\dagger}$ which are not adjacent to either $w_1$ or $w_2$. Then $\bar G$ contains the 4-cycle $\{w_1,w_2\}\vee \{v_1,v_2\}$. In either case, both $w_1$ and $w_2$ appear in $(\bar G)^{\dagger}$.
It remains to consider the case where $|W|=1$. Let $w$ be the unique vertex in $W$ and let $v$ be the unique vertex in $V^{\dagger}$ which is adjacent to $w$ in $G$. If any two vertices $v_1, v_2$ in $V^{\dagger}-\{ v\}$ are not adjacent, then $w, v_1, v_2$ forms a triangle in $\bar G$. Also, if there exists $v_1,v_2\in V^{\dagger}$ which are not adjacent to $v$ then $\{w,v\}\vee \{v_1,v_2\}$ is a 4-cycle in $\bar G$. In either case, $w$ appears in $(\bar G)^{\dagger}$. We may therefore assume the graph induced by $V^{\dagger}-\{ v\}$ is a complete graph and there is a most 1 vertex in $V^{\dagger}- \{v\}$ which does not connect to $v$. This leaves us with 2 possibilities for $G$, but we see that $\bar G$ for these two possibilities are a star-like graph on $n$ nodes with exactly 1 node distance 2 from the hub or a star graph on $n-1$ nodes union a point. In the proof of the last lemma, we showed that these two graphs are distinguished by ${\varphi}_T$ and ${\varphi}_{\bar T}$ among all graphs of order $n$.
Note that the proofs of the lemma and proposition show that for $n\geq 5$, there are exactly three pairs of graphs $(G, \bar G)$ of order $n$ for which it is not true that every vertex $v$ of $G$ is in $G^{\dagger}$ or ${\bar G}^{\dagger}$. These are the following three graphs and their complements: the star graph on $n$ nodes, the star-like graph on $n$ nodes with exactly 1 node distance 2 from the hub, and the star graph on $n-1$ nodes union a point. Hence in all but these three cases, every vertex of $G$ appears in a geodesic of $G$ or $\bar G$ (or both).
The Bartholdi zeta function {#sec:bartholdi}
===========================
An alternative to using zeta functions of graphs related to $G$ in order to study $G$ is to consider a more general notion of zeta function which involves dangling nodes and links. If we think about the adjacency spectrum, or equivalently the closed walk spectrum, it can distinguish things like path graphs from star graphs because backtracking is allowed in closed walks. On the other hand, the closed walk spectrum loses a lot of information contained in the geodesic length spectrum. Bartholdi [@bartholdi] introduced a more general zeta function which encodes both the closed walk spectrum and the geodesic length spectrum.
Let $\gamma = (e_1, \ldots, e_\ell)$ be a closed walk of length $\ell = \ell(\gamma)$. The number of backtracks in $\gamma$ is the number of $1 \le i < \ell$ such that $e_{i+1} = e_i^{-1}$. We say $\gamma$ has a tail if $e_\ell = e_1^{-1}$. The cyclic bump count ${\mathrm{cbc}}(\gamma)$ is the number of backtracks in $\gamma$ plus 1 or 0, according to whether $\gamma$ has a tail or not. The cyclic permutation group $\langle \sigma \rangle$ defined in Section \[sec:ihara\] acts on closed walks of length $\ell$ and preserves the cyclic bump count. A closed walk is primitive if it is not of the form $k\gamma$ for $k > 1$. Let $a(\ell; c) = a_G(\ell;c)$ denote the number of $\langle \sigma \rangle$ orbits of primitive closed walks in $G$ with $\ell(\gamma) = \ell$ and ${\mathrm{cbc}}(\gamma) = c$. Note $a(\ell; 0) = a(\ell)$ since ${\mathrm{cbc}}(\gamma) = 0$ means $\gamma$ is a geodesic. The [*Bartholdi zeta function*]{} is $$\label{eq:barth-def}
{\mathcal{Z}}_G(t, u) = \prod_\gamma (1-u^{{\mathrm{cbc}}(\gamma)} t^{\ell(\gamma)} )^{-1} = \prod_{c, \ell} (1 - u^c t^\ell)^{-a(\ell; c)}
= \exp \left( \sum_{c, \ell} a(\ell; c) \sum_{k\geq 1} \frac{u^{ck} t^{\ell k}}k \right),$$ where $\gamma$ runs over $\langle \sigma \rangle$ equivalence classes of primitive closed walks in $G$. Note this gives the Ihara zeta function when $u=0$: ${\mathcal{Z}}_G(t,0) = Z_G(t)$.
Bartholdi [@bartholdi] proved an analogue of the Bass determinant formula: $$\label{eq:barth-bass}
{\mathcal{Z}}_G(t,u) = (1 - (1-u)^2t^2)^{n-m} \det( I - tA + (1-u)(D - (1-u)I)t^2)^{-1}.$$
Properties determined by the Bartholdi zeta function—(M3)
---------------------------------------------------------
First observe tells us that knowing ${\mathcal{Z}}_G$ is equivalent to knowing all of the numbers $a(\ell; c)$. Since $a(2; 1) = 2m$, ${\mathcal{Z}}_G$ determines $m$. However, ${\mathcal{Z}}_G$ does not determine $n$ as adding isolated vertices does not change ${\mathcal{Z}}_G$. We see $\mathcal Z_G$ determines the number of 3-, 4-, and 5-cycles in $G$ since $Z_G$ does. However it does not determine the number of 6-cycles, as the pair of graphs in Example \[ex:crabsquid\] have the same Bartholdi zeta function.
[@KL] observed that one can rewrite in terms of the generalized characteristic polynomial ${\varphi}_{AD}^G(\lambda, x) = \det(\lambda I - A + xD)$ by $$\label{eq:phiAD}
{\mathcal{Z}}_G(t,u) = (1 - (1-u)^2t^2)^{n-m} t^{-n} {\varphi}_{AD}^G( t^{-1} - (1-u)^2t, (1-u)t )^{-1},$$ We will write ${\varphi}_{AD} = {\varphi}_{AD}^G$ if the graph $G$ is clear from context. It is stated in [@WLLX] that ${\mathcal{Z}}_G(t,u)$ determines ${\varphi}_{AD}$ and vice versa, but this is not true without further qualification in the same way that $Z_G$ does not determine ${\varphi}_T$. Namely, ${\varphi}_{AD}$ determines $m$ and $n$, so also determines ${\mathcal{Z}}_G$. On the other hand, ${\mathcal{Z}}_G$ does not determine $n$. However, since ${\mathcal{Z}}_G$ determines $m$ we see that ${\mathcal{Z}}_G$ and $n$ determine ${\varphi}_{AD}$ and vice versa.
From now on, we now consider (M3): what can be determined from $n$ and ${\mathcal{Z}}_G$, or equivalently, ${\varphi}_{AD}$?
By specializing ${\varphi}_{AD}(\lambda, x)$ to $x = 0, \pm 1$, we see ${\varphi}_{AD}$ determines the spectra of $A$, $L$ and $|L|$. There is much literature about what these spectra individually determine about $G$, and various families of graphs that are determined by such spectra (e.g., see the books [@BH], [@CRS] and the survey articles [@vDH1], [@vDH2]). We just recall a few things determined by knowing all of these spectra: the number of edges, regularity, the number of components, the number of bipartite components, the complexity and the closed walk spectrum.
From $Z_G$, we also know whether $G$ is connected md2, and all of the things discussed in Section \[sec:ihara-prop\]. In addition, [@WLLX] proves that ${\varphi}_{AD}$ determines the degree sequence of $G$.
Bartholdi zeta functions of complements—(M4) {#sec:M4}
--------------------------------------------
Finally, consider our last method (M4): what can be determined from $n$, ${\mathcal{Z}}_G$ and ${\mathcal{Z}}_{\bar G}$.
Equation tells us these quantities determine the spectra of $A+xD$ and $A+xD-J$ for all $x$. When $x=0$, [@JN] showed this determines the spectra of $A + yJ$ for all $y$, and the proof (see [@vDH1] for a simpler proof) in fact works for $x \ne 0$, i.e., $n$, ${\mathcal{Z}}_G$ and ${\mathcal{Z}}_{\bar G}$ determine the generalized characteristic polynomial ${\varphi}_{ADJ}(\lambda, x, y) = {\varphi}_{ADJ}^G(\lambda,x,y) = \det(\lambda I - A + xD +y J)$. The converse, that ${\varphi}_{ADJ}$ determines $n$, ${\mathcal{Z}}_G$ and ${\mathcal{Z}}_{\bar G}$, is straightforward.
Consequently, (M4) determines everything (M2) and (M3) do. In fact, we show below that (M4) determines everything (M1) does.
Let $X$ and $Y$ be $n\times n$ and $m\times m$ matrices respectively and let $J_{nm}$ denote the $n\times m$ all ones matrix. Let $M$ be the block matrix $M={\begin{bmatrix}}X & J_{nm} \\ J_{nm} & Y {\end{bmatrix}}$. We will write $J$ for the square all ones matrix when the order is clear from context. The following are true:
1. The spectra of $M$ and $J-M$ are determined by the spectra of $X, J-X, Y$, and $J-Y$.
2. The spectra of $X$ and $J-X$ are determined by the spectra of $M$, $J-M$, $Y$, and $J-Y$.
Note that, for an $r \times r$ matrix, one can determine the spectrum of the matrix from the traces of the first $r$ powers of the matrix and, conversely, one can determine the traces of all powers from the spectrum.
We can prove using induction that $M^k$ has the form ${\begin{bmatrix}}X^k + A_k & B_k\\ C_k & Y^k+D_k{\end{bmatrix}}$ where the matrices $A_k,B_k,C_k,D_k$ are sums of matrices which are the product of $k$ matrices coming from $\{X,Y, J_{nm}, J_{mn}\}$. For $k\geq 2$, each product of $k$ matrices appearing in $A_k$ and $D_k$ is such that at most $k-2$ of the $k$ matrices are $X$’s or $Y$’s, and for $B_k$ and $C_k$ at most $k-1$ are $X$’s or $Y$’s.
From the equation for $M^k$, we get that ${\mathrm{tr}}(M^k) = {\mathrm{tr}}(X^k)+{\mathrm{tr}}(Y^k) +{\mathrm{tr}}(A_k)+{\mathrm{tr}}(D_k)$. We also use the following properties: ${\mathrm{tr}}(XJ_{nm}YJ_{mn})= (\sum_{i,j} x_{ij}) (\sum_{i,j} y_{ij}) = {\mathrm{tr}}(XJ){\mathrm{tr}}(YJ)$, trace is invariant under cyclic permutations of products, and the product of any two all ones matrices is a scalar multiple of an all ones matrix. From these properties, it follows that ${\mathrm{tr}}(A_k)$ and ${\mathrm{tr}}(D_k)$ can be determined from ${\mathrm{tr}}(X^iJ)$, and ${\mathrm{tr}}(Y^iJ)$ for $i=0,1,2,..,k-2$. This leads to a relation of the form ${\mathrm{tr}}((J-X)^k) = (-1)^k({\mathrm{tr}}(X^k) -k{\mathrm{tr}}(X^{k-1}J) + \cdots)$, where the omitted terms are determined by ${\mathrm{tr}}(X^iJ)$ for $i=0,1,2,...,k-2$, as well as a similar relation for $Y$.
We now prove the two statements. Note that in both cases we can determine $n$ and $m$. Also note that $J-M$ is the block diagonal matrix with $J-X$ and $J-Y$ on the diagonals, so any two of the spectra of $J-M, J-X,$ and $J-Y$ determine the third.
First suppose we know the spectra of the matrices $X, J -X, Y$, and $J-Y$. Then we know ${\mathrm{tr}}((J-X)^i)$ and ${\mathrm{tr}}(X^i)$ for all $i$. From this, we can recursively determine ${\mathrm{tr}}(X^iJ)$ for all $i$, and similarly for $Y$. We can therefore compute ${\mathrm{tr}}(M^k) = {\mathrm{tr}}(X^k)+{\mathrm{tr}}(Y^k) +{\mathrm{tr}}(A_k)+{\mathrm{tr}}(D_k)$ for all $k$ and thus determine the spectrum of $M$.
Suppose now that we know the spectra of $Y, J-Y, M, J-M$. As mentioned above, this tells us the spectrum of $J-X$. It remains to show that we can compute ${\mathrm{tr}}(X^k)$ for $k=1,2,...,n$. For $k=1,2$ we can find ${\mathrm{tr}}(X)$ and ${\mathrm{tr}}(X^2)$ from ${\mathrm{tr}}(M)={\mathrm{tr}}(X)+{\mathrm{tr}}(Y)$ and ${\mathrm{tr}}(M^2)= {\mathrm{tr}}(X^2)+{\mathrm{tr}}(Y^2)+{\mathrm{tr}}(J_{mn}J_{nm})+{\mathrm{tr}}(J_{nm}J_{mn})$. If $k>2$ and we know ${\mathrm{tr}}(X^i)$ for $i<k$, then from ${\mathrm{tr}}((J-X)^i)$ and ${\mathrm{tr}}(X^i)$ for $i=1,2,..,k-1$ we can find ${\mathrm{tr}}(X^iJ)$ for $i=1,2,...,k-2$, and therefore also ${\mathrm{tr}}(A_k)$ and ${\mathrm{tr}}(D_k)$. From these, we can compute ${\mathrm{tr}}(X^k)$ from ${\mathrm{tr}}(M^k) = {\mathrm{tr}}(X^k)+{\mathrm{tr}}(Y^k) +{\mathrm{tr}}(A_k)+{\mathrm{tr}}(D_k)$.
If $D$ is the degree matrix of $G$, we write $\bar D$ for the degree matrix of $\bar G$.
Fix $x \in {\mathbb{C}}$. Let $G_1$ and $G_2$ be two graphs of the same order and let $H$ be any graph. The graphs $G_1\vee H$ and $G_2\vee H$ are cospectral with respect to $A+xD$ and $\bar A + x\bar D$ if and only if $G_1$ and $G_2$ are cospectral with respect to $A+xD$ and $\bar A+ x\bar D$.
\[thm:cone-comp\]
Let $M=A+xD$ and note that, for an order $r$ graph, $\bar M := \bar A + x \bar D = J-M + (x(r-1)-1)I$. Thus if we know $x$ and $r$, knowing the spectrum of $\bar M$ is equivalent to knowing the spectrum of $J-M$. Let $n$ be the order of $G_1$ and $G_2$, let $m$ be the order of $H$, and denote their respective matrices $M=A+xD$ by $M_{G_i}$ and $M_H$. The join $G_i\vee H$ has matrix $M_{G_i\vee H} = {\begin{bmatrix}}M_{G_i} + xm I_n & J_{nm}\\ J_{mn} & M_H + xnI_m {\end{bmatrix}}$. We can thus apply the previous lemma to show that if $G_1$ and $G_2$ are cospectral with respect to $M$ and $\bar M$, then so are $G_1\vee H$ and $G_2\vee H$ and vice versa.
This is already well known in the case that $x=-1$, which corresponds to the Laplacian—in fact less is needed in this case since the Laplacian spectrum of $G$ determines that of $\bar G$.
\[cor:cone-comp\] The graphs $G_1$ and $G_2$ have the same spectra with respect to $A$ and $\bar A$ (or $|L|$ and $\overline{|L|}$), if and only if the same is true for $G_1^*$ and $G_2^*$. Similarly, $G_1$ and $G_2$ have the same generalized characteristic polynomial ${\varphi}_{ADJ}$ if and only if the same is true for $G_1^*$ and $G_2^*$, i.e., ${\varphi}_{ADJ}^{G_1} = {\varphi}_{ADJ}^{G_2}$ if and only if ${\varphi}_{ADJ}^{G_1^*} = {\varphi}_{ADJ}^{G_2^*}$.
Constructions {#sec:constructions}
=============
Here we discuss three well-known constructions of cospectral graphs, and show that in many cases, such graphs are distinguished by methods (M1)–(M4) using constraints on degree distributions. This provides some evidence for our conjectures in the next section that most cospectral graphs can be distinguished by any of (M1)–(M4).
In the last part of this section, we give a new construction for graphs which cannot be distinguished by (M1)–(M4), which generalizes the pair of graphs from Example \[ex:9verts\].
GM switching
------------
Godsil and McKay present a method for constructing cospectral pairs of graphs [@GM], which is now referred to as GM switching. We say $G$ satisfies the $(k+1)$-GM (or just GM) condition if there is an ordering of the vertices such that the adjacency matrix can be written in the form $$A = {\begin{bmatrix}}B_1 & B_{12} & \cdots & B_{1k} & N_1 \\ {}^t B_{12} & B_2 & \cdots & B_{2k} & N_2 \\
\vdots & &\ddots & & \vdots \\ {}^t B_{1k} & \cdots & {}^t B_{k-1,k} & B_k & N_k \\
{}^t N_1 & \cdots & {}^t N_{k-1} & {}^t N_k & C {\end{bmatrix}}$$ where (i) each $G_{B_i}$ is regular, (ii) each $B_{i,j}$ has constant row and column sums, and (iii) each column of each $N_i$ has exactly 0, $b_i/2$ or $b_i$ 1’s. Here $b_i$ is the order of each square matrix $B_i$, and we assume at least one of the $b_i$’s is even. Note $k+1$ is the number of diagonal blocks, so the ordering of the vertices and the size of the $B_i$’s determines a partition of the vertex set into $k+1$ subsets $V_{B_1}, \ldots, V_{B_k}, V_C$.
Let ${\widetilde}N_i$ be the matrix formed from $N_i$ by replacing each column $\mathbf{v}$ of $N_i$ which consists of $b_i/2$ ones by the column $J_{b_i,1}-\mathbf{v}$. The GM switch of $G$ is the graph ${\widetilde}G$ with adjacency matrix ${\widetilde}A$ which is gotten from $A$ by replacing each $N_i$ with ${\widetilde}N_i$. Then $G$ and ${\widetilde}G$ are cospectral with respect to the adjacency matrix. The proof is to exhibit a matrix $Q$ which conjugates $A$ to ${\widetilde}A$ and has constant row and column sums of 1, and hence commutes with the all ones matrix $J$. It follows that the complements of $G$ and ${\widetilde}G$ are also cospectral with respect to the adjacency matrix. Also, by Corollary \[cor:cone-comp\], the cones of $G$ and ${\widetilde}G$ are cospectral with respect to the adjacency matrix.
Haemers and Spence introduced a special case of GM switching [@HS], which they called GM\* switching, which gives cospectral pairs with respect to any matrix of the form $A+xD$. They only define GM\* switching in the case where $k=1$, but it works in the more general setup as well. We say $G$ satisfies the $(k+1)$-GM\* condition if $G$ satisfies the $(k+1)$-GM condition with the additional requirement that each vertex in $G_{B_i}$ has the same degree in $G$ (i.e., each $N_i$ has constant row sums). This condition guarantees that $G$ and ${\widetilde}G$ have the same degree matrix $D$. Also $D$ and $J$ commute with $Q$, the matrix which conjugates $A$ to ${\widetilde}A$, and hence $A+xD+yJ$ and ${\widetilde}A+xD+yJ$ are cospectral for any $x,y$ so $G$ and ${\widetilde}G$ have the same generalized characteristic polynomial ${\varphi}_{ADJ}$. Also, by Corollary \[cor:cone-comp\], the cones of $G$ and ${\widetilde}G$ have the same ${\varphi}_{ADJ}$. We call this $(k+1)$-GM\* switching and say that $(G, {\widetilde}G)$ is a $(k+1)$-GM\* pair provided $G \not \simeq {\widetilde}G$.
In the next section, we will find that GM\* switching accounts for a significant percentage of graphs up to 11 vertices which have the same ${\varphi}_{ADJ}$ (see Table \[tab:main\]). For the graphs on up to 10 vertices, 3-GM\* switching does not produce any such examples that 2-GM\* does not. However for 11 vertices, there are 108 3-GM\* pairs which cannot be obtained by 2-GM\* switching, even if one allows successive 2-GM\* switching. (It happens that $(G_1, G_2)$ can be a 3-GM\* pair but not a 2-GM\* pair, but that $G_1$ and $G_2$ are both obtained as 2-GM\* switches from a third graph $G_3$.)
\[ex:3GMst\] The graphs [`J?BD?oX[F[?`]{} and [``` J?`CP``LE{? ``` ]{} on 11 vertices and 18 edges drawn below are a 3-GM\* pair, but not a 2-GM\* pair. Further, neither of these form a 2-GM\* pair with any other graph. The subgraphs $G_{B_1}, G_{B_2}$, and $G_C$ are the subgraphs induced by vertices 1–6, 7–9, and 10–11 respectively.
(0,0) rectangle (3.0cm,3.75cm); (v0)(v2) (v0)(v8) (v0)(v10) (v1)(v5) (v1)(v8) (v1)(v10) (v2)(v6) (v2)(v10) (v3)(v4) (v3)(v7) (v3)(v9) (v4)(v7) (v4)(v9) (v5)(v6) (v5)(v9) (v6)(v10) (v7)(v10) (v8)(v10)
(0,0) rectangle (3.0cm,3.5cm); (v0)(v2) (v0)(v8) (v0)(v9) (v1)(v5) (v1)(v8) (v1)(v9) (v2)(v6) (v2)(v9) (v3)(v4) (v3)(v7) (v3)(v10) (v4)(v7) (v4)(v10) (v5)(v6) (v5)(v10) (v6)(v10) (v7)(v10) (v8)(v10)
We now focus on the case of 2-GM switching. For simplicity, we omit the subscripts and write the adjacency matrix as $$A = {\begin{bmatrix}}B & N\\ {}^t N & C {\end{bmatrix}}.$$ For $G \not \simeq {\widetilde}G$, we need $B$ to have even size $\ge 4$. Further, the larger $B$ is, the less likely it is that the GM condition will be satisfied. So Haemers and Spence [@HS Thm 3] use GM switching with B of size 4 to get a lower bound on the number of non-DS graphs. The following shows that most cospectral pairs thus constructed are distinguished by zeta functions.
\[thm:GM-zeta\] Suppose $G_B$ is regular on $4$ vertices and $G_C$ is an md1 (resp. md2) graph on $n$ vertices. Then the proportion of GM-admissible choices of $N$ such that the pair $(G, {\widetilde}G)$ of [*labeled*]{} graphs formed by GM switching on $(B,C,N)$ which are distinguished by $Z^*$ (resp. $Z$) goes to $1$ as $n \to \infty$.
For each column of $N$, there are $2 + \binom{4}{2} = 8$ possible choices, which we count with equal probability. The probability that at least 2 columns are all ones is $$1 - \left[\binom{n}{0} \left(\frac 18\right)^0 \left(\frac 78\right)^n + \binom{n}{1} \left(\frac 18\right)^1 \left(\frac 78\right)^{n-1}\right] = 1 - \left(\frac{7+n}8 \right)\left( \frac 7 8 \right)^{n-1}.$$ When this is satisfied, $G$ and ${\widetilde}G$ are md1 if $G_C$ is md1 and md2 if $G_C$ is md2. Since this probability tends to 1 as $n \to \infty$, from now on, we will assume $G$ and ${\widetilde}G$ are md1 in the case of $Z^*$ or md2 in the case of $Z$.
If $G$ and ${\widetilde}G$ are not distinguished by $Z^*$ (resp. $Z$), then the products of the vertex degrees (resp. degrees minus 1) must be the same. Since all vertices coming from $G_C$ have the same degree, this is equivalent to knowing that the products of the degrees of vertices (resp. degrees minus 1) from $B$ are the same. Let $k$ be the degree of a vertex in $G_B$ plus the number of all ones columns in $N$ in the case of $Z^*$, or this number minus 1 in the case of $Z$. Let $N_0$ be the submatrix of $N$ formed by removing the columns consisting of all zeroes or all ones. Let $x_i$ denote the number of 1’s in the $i$-th row of $N_0$. Then $G$ and ${\widetilde}G$ being distinguished by $Z^*$ or $Z$ implies $$\label{GM-degcond}
\prod_{i=1}^4 (k+x_i) = \prod_{i=1}^4 (k+n_0 - x_i),$$ where $n_0$ is the number of columns in $N_0$.
View $n_0$ as fixed for now. Since $x_4 = 2n_0 - x_1 - x_2 - x_3$, for fixed $x_1, x_2$, the solutions to in $x_3$ are the solutions to a degree 2 polynomial in $x_3$, of which there are at most 2. Now view the top 2 rows of $N_0$ as fixed. Let $r$ be the number of columns in $N_0$ which have exactly one 1 in the first 2 entries. Then there are $r+1$ choices for $x_3$. There are $2^r$ choices for row 3 of $N_0$, hence for an integer $0\leq y \leq r$, the probability that $x_3=y$ is $\binom{r}{y} \frac 1{2^r}\leq \binom{r}{\lceil r/2 \rceil}\frac 1{2^r}$. Given $r \ge 3$, the probability that $x_3$ avoids solving is at least $$1 - 2 \binom{r}{\lceil r/2 \rceil} \frac 1{2^r} \ge 1 - \frac{2}{\sqrt{3 {\lceil r/2 \rceil} + 1}} \ge 1 - \sqrt{\frac 8{3r}}.$$ (Here we use the inequality $\binom{2n}{n} \le \frac{2^{2n}}{\sqrt{3n+1}}$, which implies $\binom{r}{\lceil r/2 \rceil} \le \frac{2^r}{\sqrt{3 {\lceil r/2 \rceil} + 1}}$.) Also note $$P( width(N_0) = n_0) = \binom{n}{n_0} \left(\frac 34\right)^{n_0} \left(\frac 14\right)^{n-n_0}$$ and for a fixed $n_0$ the probability that the number of columns in $N_0$ which have exactly one 1 in the first 2 entries is $r$ is $$P(r|n_0) = \binom{n_0}{r} \left(\frac 23\right)^{r} \left(\frac 13\right)^{n_0-r}.$$
Fix $0 < \delta < 1$. By the law of large numbers, for any $\epsilon>0$, the probability that $r\ge n_0 ( \frac 23 -\epsilon)$ goes to 1 as $n_0 \to \infty$. Hence the probability that $r \ge n_0^\delta$ goes to 1 as $n_0 \to \infty$. Similarly, the probability that $n_0 \ge n^\delta$ goes to 1 as $n_0 \to \infty$. So the probability that $x_3$ avoids the solutions of is at least $$P(n_0 \ge n^\delta) P(r \ge n_0^\delta) \left(1 - \sqrt{\frac 8{3n_0^\delta}}\right) \ge P(r \ge n^{\delta^2} ) \left(1 - \sqrt{\frac 8{3n^{\delta^2}}} \right),$$ which goes to 1 as $n \to \infty$.
Let $g_n$ denote the number of simple graphs (up to isomorphism) of order $n$.
The number of graphs $G$ on $n$ vertices for which there exists a cospectral nonisomorphic graph ${\widetilde}G$ but $Z_{\vphantom{{\widetilde}G}G} \ne Z_{{\widetilde}G}$ is at least $n^3 g_{n-1} ( \frac 1{24} - o(1) )$. The same statement is true with $Z_{\vphantom{{\widetilde}G}G} \ne Z_{{\widetilde}G}$ replaced by $Z_{\vphantom{{\widetilde}G^*}G^*} \ne Z_{{\widetilde}G^*}$.
Haemers and Spence [@HS Thm 3] show there are at least $n^3 g_{n-1} ( \frac 1{24} - o(1) )$ non-isomorphic pairs $(G, {\widetilde}G)$ obtained from GM switching with $B$ of size 4. Further, for almost all of these pairs, both graphs are md2. Since, in the argument above, we can replace the condition that $C$ is md2 or md1 with knowing $G$ and ${\widetilde}G$ are md2 or md1, we see almost all such pairs are distinguished by $Z$ or $Z^*$.
Coalescence construction
------------------------
Suppose $G_1$ and $G_2$ are two graphs of order $n$ with the same adjacency spectra but different degree sequences. Assume there are vertices $x_1$ of $G_1$ and $x_2$ of $G_2$ such that $G_1 - \{ x_1 \}$ and $G_2 - \{ x_2 \}$ are cospectral. Let $U_i$ be the vertex set of $G_i - \{ x_i \}$. Let $\Gamma$ be any graph with a fixed vertex $y$. Let $G_i'$ be the coalescence of $(G_i, x_i)$ with $(\Gamma, y)$, i.e., the union of $G_i$ and $\Gamma$ after identification of $x_i$ with $y$.
With notation as above, ${\mathrm{Spec}}(G_1') = {\mathrm{Spec}}(G_2')$ for any $(\Gamma, y)$. If $G_1$ and $G_2$ are md2 and $\prod_{v \in U_2} (\deg(v) -1) \ne \prod_{v \in U_2} (\deg(v) - 1)$, then $Z_{G_1'} \ne Z_{G_2'}$. Similarly, if $G_1$ and $G_2$ are md1 and $\prod_{v \in U_1} \deg(v) \ne \prod_{v \in U_2} \deg(v)$, then $Z_{G_1'^*} \ne Z_{G_2'^*}$.
The cospectrality is due to Schwenk [@schwenk].
For the distinction by zeta functions, note that the fact that $G_1 - \{ x_1 \}$ and $G_2 - \{ x_2\}$ are cospectral implies that $\deg_{G_1}(x_1) = \deg_{G_2}(x_2)$ as the spectrum determines the number of edges. Note $\deg_{G_i}(v) = \deg_{G_i'}(v)$ for any $v \in U_i$ and $\deg_{G_1'}(v) = \deg_{G_2'}(v)$ for any vertex $v$ of $\Gamma$. The assertions follow as $Z_G$ (resp. $Z_{G^*}$) determines the product of the degrees minus 1 (degrees) of the pruned graph $G^{\dagger}$.
Let $G_1$ and $G_2$ be the graphs [`F?zPw`]{} and [`F@Rfo`]{} on 7 vertices and 10 edges pictured below, where $x_1$ and $x_2$ are the white vertices.
(0,0) rectangle (2.0cm,2.25cm); (v0)(v4) (v0)(v5) (v1)(v4) (v1)(v5) (v2)(v4) (v2)(v6) (v3)(v5) (v3)(v6) (v4)(v6) (v5)(v6)
(0,0) rectangle (2.0cm,2.0cm); (v0)(v5) (v0)(v6) (v1)(v4) (v1)(v5) (v1)(v6) (v2)(v3) (v2)(v5) (v2)(v6) (v3)(v6) (v4)(v6)
Then $G_1$ and $G_2$ are cospectral but not isomorphic, whereas $G_1 - {x_1} $ and $G_2 -{x_2}$ are cospectral because they are isomorphic. Here $x_1$ and $x_2$ have degree 2. The other vertex degrees are $(4,4,4,2,2,2)$ for $G_1$ and $(5,3,3,3,2,2)$ for $G_2$. It is clear that the products of the degrees and the products of the degrees minus 1 are different. Thus, for any coalescences $G_1'$ and $G_2'$ of $(G_1, x_1)$ and $(G_2, x_2)$ with any $(\Gamma,y)$, we have $Z_{G_1'} \ne Z_{G_2'}$ and $Z_{G_1'^*} \ne Z_{G_2'^*}$.
We assumed that $G_1$ and $G_2$ are md2 or md1 in this proposition for simplicity, but this is not necessary. What one really needs is a condition on pruned subgraphs of $G_1'$ and $G_2'$.
Note the coalescence construction includes the case of a disjoint union. If $H_1$ and $H_2$ are cospectral, we can take $G_i$ to be $H_i$ disjoint union a single vertex $x_i$. Then $G_1$ and $G_2$ are cospectral, as are $G_1 - \{ x_1 \} = H_1$ and $G_2 - \{ x_2 \} = H_2$ by assumption. Let $\Gamma$ be any graph and $y$ any vertex in $\Gamma$. Then the coalescence of $(G_i, x_i)$ with $(\Gamma, y)$ is simply the disjoint union $H_i \sqcup \Gamma$ of $H_i$ with $\Gamma$. However in the case of disjoint unions, we already know the stronger statement that $Z_{H_1} \ne Z_{H_2}$ implies $Z_{G_1'} \ne Z_{G_2'}$ since zeta functions factor into products over their connected components (though this factorization is not true for the zeta of the cones $Z_{G_i'^*}$). E.g., if $H_1$ and $H_2$ are the unique pair of cospectral graphs on 5 vertices, then $Z_{H_1 \sqcup \Gamma} \ne Z_{H_2 \sqcup \Gamma}$ for any $\Gamma$, though $G_1$ and $G_2$ are not md2.
Join construction
-----------------
Suppose $G_1$ and $G_2$ are two graphs of order $n$ which have the same Laplacian spectra but different degree sequences. Let $\Gamma$ be an arbitrary graph. If $G_1$ or $G_2$ has an isolated vertex, then assume $\Gamma$ has at least 2 vertices. Then the joins $G_1 \vee \Gamma$ and $G_2 \vee \Gamma$ are connected and md2.
With notation as above, ${\mathrm{Spec}}_L(G_1 \vee \Gamma) = {\mathrm{Spec}}_L(G_2 \vee \Gamma)$ for any $\Gamma$. However $Z_{G_1 \vee \Gamma} \ne Z_{G_2 \vee \Gamma}$ for all but finitely many $\Gamma$. Specifically, there is a finite set $S$ consisting of at most $n-1$ integers such that $Z_{G_1 \vee \Gamma} \ne Z_{G_2 \vee \Gamma}$ for any $\Gamma$ whose order $r$ does not lie in $S$.
The first part is true for any $\Gamma$.
Let $d_1, \ldots, d_n$ be the degree sequence for $G_1$ and $d_1', \ldots, d_n'$ be the degree sequence for $G_2$. Let $\delta_1, \ldots, \delta_r$ be the degree sequence for $\Gamma$. Then the degree sequence for $G_1 \vee \Gamma$ is $d_1+r, \ldots, d_n + r, \delta_1 + n, \ldots, \delta_r + n$ and similarly for $G_2 \vee \Gamma$. Assume $r \ge 2$ so $G_1 \vee \Gamma$ and $G_2 \vee \Gamma$ are md2. Consequently $Z_{G_1 \vee \Gamma}$ determines $$\prod (d_i + r - 1) \cdot \prod (\delta_j + n - 1)$$ and similarly for $G_2 \vee \Gamma$. Hence $Z_{G_1 \vee \Gamma} \ne Z_{G_2 \vee \Gamma}$ if $$\prod (d_i + r - 1) \ne \prod (d_i' + r - 1).$$ Consider the polynomial $$f(x) = \prod (d_i + x) - \prod (d_i' + x),$$ which has degree $< n$. The previous equation holds if and only if $r-1$ is not a root of $f(x)$, which can only happen for at most $n-1$ values of $r$.
Note, one can replace the bound on the size of $S$ by the number of differing vertex degrees (counting multiplicity) with $G_1$ and $G_2$. Also, this proposition gives a non-algorithmic proof of Lemma \[lem:ds\] by showing that any two order $n$ graphs with different degree sequences can by distinguished by the zeta functions of $n$ joins with graphs of distinct orders.
Let $G_1$ and $G_2$ be [`ECZo`]{} and [`EEr_`]{}. These are graphs on 6 vertices with the same Laplacian spectra, but $G_1$ has degree sequence $(4,2,2,2,2,2)$ and $G_2$ has degree sequence $(3,3,3,2,2,1)$.
(0,0) rectangle (2.0cm,2.25cm); (v0)(v3) (v0)(v5) (v1)(v4) (v1)(v5) (v2)(v4) (v2)(v5) (v3)(v5)
(0,0) rectangle (2.0cm,2.0cm); (v0)(v3) (v0)(v4) (v0)(v5) (v1)(v3) (v1)(v4) (v1)(v5) (v2)(v5)
For any (non-null) graph $\Gamma$, the graphs $G_1 \vee \Gamma$ and $G_2 \vee \Gamma$ are automatically md2. The polynomial $f(x)$ in the previous proof is just $f(x) = (4+x)(2+x)^5 - (3+x)^3(2+x)^2(1+x) =(2+x)^2 (2x+5)$. This has no positive roots, so $Z_{G_1 \vee \Gamma} \ne Z_{G_2 \vee \Gamma}$ for any $\Gamma$.
The same conclusion is true if we replace $G_1$ and $G_2$ by their complements.
There are at least $2g_{n-6}$ pairs of graphs $(G_1, G_2)$ of order $n$ which have the same Laplacian spectra but different Ihara zeta functions.
A new construction {#sec:new-con}
------------------
There is one pair of non-isomorphic md2 graphs on 9 vertices which have the same generalized characteristic polynomial ${\varphi}_{ADJ}$, the pair from Example \[ex:9verts\]. This is the smallest such example. This pair does not occur as the result of GM\* switching for the following reason: the matrices $A+xD$ are similar for all $x$, but there is no matrix $P$ such that $P^{-1}(A_1+xD_1)P = (A_2+xD_2)$ for all $x$, as must be the case for GM\* pairs. (In [@SS], it was mistakenly written that all pairs of connected md2 graphs on $n \le 11$ vertices with the same zeta function, adjacency spectrum, and Laplacian spectra are obtained by GM\* switching, but this is false as this example shows. The second author of that paper informed us that the conclusion of that sentence, “are obtained by GM\* switching,” should be “have the same ${\varphi}_{ADJ}$.”)
There do however exist matrices $P$ and $R$ such that $(A_1+xD_1)(P+xR) = (P+xR)(A_2+xD_2)$ and $P+xR$ commutes with $J$ for all $x$ and is invertible for all real valued $x$. Here we prove this example is part of a more general construction of pairs of graphs with the same ${\varphi}_{ADJ}$ which cannot be explained by GM\* switching. The fact that the conjugating matrix $P+xR$ needs to depend on $x$ makes this construction more delicate (and complicated) than GM\* switching (and as far as we know, there are no other such constructions). So we first describe it in a concrete way, and then remark afterwards what are the necessary abstract conditions for this construction to work.
\[our-construction\] Consider adjacency matrices $A_1, A_2$ of the following form: $$A_1 = {\begin{bmatrix}}B_1 & B_{12} & \cdots & B_{1k} & N_1 \\ {}^t B_{12} & B_2 & \cdots & B_{2k} & N_2 \\
\vdots & &\ddots & & \vdots \\ {}^t B_{1k} & \cdots & {}^t B_{k-1,k} & B_k & N_k \\
{}^t N_1 & \cdots & {}^t N_{k-1} & {}^t N_k & C {\end{bmatrix}},
A_2 = {\begin{bmatrix}}B_1 & {}^t B_{12} & \cdots & {}^t B_{1k} & N_1 \\B_{12} & B_2 & \cdots & {}^t B_{2k} & N_2 \\
\vdots & &\ddots & & \vdots \\ B_{1k} & \cdots & B_{k-1,k} & B_k & N_k \\
{}^t N_1 & \cdots & {}^t N_{k-1} & {}^t N_k & C {\end{bmatrix}}$$ Here we allow the $B_i$’s to be chosen from among the following $4 \times 4$ adjacency matrices given in $2 \times 2$ block form by: $${\begin{bmatrix}}0 & 0\\ 0 & 0 {\end{bmatrix}}, {\begin{bmatrix}}0 & I \\ I & w {\end{bmatrix}},{\begin{bmatrix}}0 & J \\ J & 0 {\end{bmatrix}}, {\begin{bmatrix}}w & 0 \\ 0 & w {\end{bmatrix}}, {\begin{bmatrix}}w & J \\ J & w {\end{bmatrix}},$$ where $w = J - I$. The $B_{ij}$’s ($i<j$) are allowed to be arbitrarily chosen from among the following 0-1 matrices with constant row and column sums: $${\begin{bmatrix}}0 & 0\\ 0 & 0 {\end{bmatrix}}, {\begin{bmatrix}}J & J \\ J & J {\end{bmatrix}}, {\begin{bmatrix}}0 & J \\ J & 0 {\end{bmatrix}}, {\begin{bmatrix}}J & 0 \\ 0 & J {\end{bmatrix}}, {\begin{bmatrix}}I & 0 \\ 0 & I{\end{bmatrix}}, {\begin{bmatrix}}w & 0\\ 0 & w {\end{bmatrix}}, {\begin{bmatrix}}0 & I \\ w & 0 {\end{bmatrix}}, {\begin{bmatrix}}0 & w \\ I & 0{\end{bmatrix}},$$ $${\begin{bmatrix}}I & J \\ J & I{\end{bmatrix}}, {\begin{bmatrix}}w & J \\ J & w {\end{bmatrix}}, {\begin{bmatrix}}J & w \\ I & J {\end{bmatrix}}, {\begin{bmatrix}}J & I \\ w & J {\end{bmatrix}}, {\begin{bmatrix}}I & I \\ w & I {\end{bmatrix}}, {\begin{bmatrix}}I & w \\ I & I {\end{bmatrix}}, {\begin{bmatrix}}w & I \\ w & w {\end{bmatrix}}, {\begin{bmatrix}}w & w \\ I & w {\end{bmatrix}}.$$ We allow $C$ to be any $m\times m$ adjacency matrix, and the $N_i$’s are $4\times m$ matrices such that each column of $N_i$ consists of all zeros or all ones.
Then ${\varphi}_{ADJ}^{G_1} = {\varphi}_{ADJ}^{G_2}$, where $G_i$ is the graph with adjacency matrix $A_i$.
Note that as each $B_{ij}$ has constant row and column sums and each $N_i$ has constant row sums, $G_1$ and $G_2$ have the same degree matrix, call this $D$. Consider the following $4\times 4$ matrices: $$Q = {\begin{bmatrix}}w & w-I\\ w-I & I {\end{bmatrix}}, S = \frac{1}{2}{\begin{bmatrix}}I-w & 0 \\ 0 & w-I{\end{bmatrix}}$$ Let $P$ be the block diagonal matrix whose first $k$ blocks are the matrix $Q$ and last block is the $m\times m$ identity matrix. Let $R$ be the block diagonal matrix whose first $k$ blocks are the matrix $S$ and last block is the $m\times m$ zero matrix.
One can check that $(A_1+xD)(P+xR) = (P+xR)(A_2+xD)$ for all $x$. Note that $P$ has constant row and column sums of 1 and $R$ has constant row and column sums of $0$ so $P+xR$ commutes with $J$ for all $x$. Therefore $(A_1+xD + yJ)(P+xR) = (P+xR)(A_2+xD + yJ)$ for all $x,y$. Also, $\det(P+xR) = \det(Q+xS)^k = (-x^2+2x-5)^k$ which is nonzero for all real $x$. It follows that $A_1+xD + yJ$ and $A_2+xD + yJ$ have the same eigenvalues for all $x,y$ so $G_1$ and $G_2$ have the same ${\varphi}_{ADJ}$.
\[ex:9verts2\] The pair of graphs $G_1$ and $G_2$ from Example \[ex:9verts\] can be obtained from the above construction as follows. Take $k=2$ and $B_1 = B_2 = {\begin{bmatrix}}0 & I \\ I & w {\end{bmatrix}}$ and $B_{12}= {\begin{bmatrix}}I & I \\ w & I {\end{bmatrix}}$. Take $C$ to be the $1\times 1$ zero matrix, $N_1$ to be the $4\times 1$ zero vector and $N_2$ to be the $4\times 1$ all ones vector. The resulting graphs are the graphs from Example \[ex:9verts\], which we redraw below with the vertices labeled in the order that they appear in this description.
(1)(3) (3)(4) (4)(2) (5)(7) (7)(8) (8)(6) (1)(5) (1)(7) (3)(7) (3)(6) (4)(5) (4)(8) (2)(8) (2)(6) (5)(9) (6)(9) (7)(9) (8)(9)
(1)(3) (3)(4) (4)(2) (5)(7) (7)(8) (8)(6) (1)(5) (1)(8) (3)(7) (3)(5) (4)(6) (4)(8) (2)(7) (2)(6) (5)(9) (6)(9) (7)(9) (8)(9)
If in the previous example, we take $B_1, B_2, B_{12}$ as before but let $C$ be any adjacency matrix and $N_1 = 0$ and $N_2=J$ (appropriately sized all zero and all ones matrices respectively), then this construction always results in a non-isomorphic pair of graphs. Hence this construction produces at least $g_{n-8}$ non-isomorphic graphs with the same ${\varphi}_{ADJ}$ on $n$ vertices. On 10 vertices, there are 6 different non-isomorphic pairs of graphs that can be built using this construction.
This construction can be made more general. Given a set of $n\times n$ adjacency matrices $B_i$ and a set of $n\times n$ matrices $B_{ij}$ with 0 and 1 entries and constant row and column sums, then we can do an analogous construction if there exist $n\times n$ matrices $Q$ and $S$ with the following properties. First, $Q$ and $S$ should have constant row and column sums of 1 and 0 respectively and have $\det(Q+xS)$ not equal to the zero polynomial. If $D_i$ is the degree matrix of $M_i$, then we need $QB_i= B_iQ$, $D_iS=SD_i$, and $D_iQ-QD_i = SB_i-B_iS$ for all $i$. Also, we need $B_{ij}Q = Q{}^tB_{ij}$, ${}^tB_{ij} Q = QB_{ij}$, $B_{ij}S = S{}^tB_{ij}$, and ${}^tB_{ij}S = SB_{ij}$. If all of these properties are satisfied, then any $A_1$ and $A_2$ built in the same manner as in the construction will result in graphs with the same ${\varphi}_{ADJ}$. Note that at least one of the $B_{ij}$ must be non-symmetric for $A_1$ and $A_2$ to be different, but this does not guarantee that the resulting graphs will be non-isomorphic.
Calculations and Conjectures {#sec5}
============================
In this section, we present our computational study of comparing graphs via the usual spectral methods ($A$, $L$ and $|L|$) with methods (M1)–(M4). Our findings, together with the above results, lead us to several formal and informal conjectures.
In [@HS], Haemers and Spence enumerated the number of graphs on $n \le 11$ vertices which are not DS, not $L$-DS, and not $|L|$-DS (see also [@BS2] and [@spence:web] for some errata and additional data for $n=12$). This data is summarized in the $A$, $L$ and $|L|$ columns of Table \[tab:main\]. They also enumerated the number of non-DS graphs explained by 2-GM switching and 2-GM\* switching. This data is recalled in the 2-GM and 2-GM\* columns of Table \[tab:main\]. The 2-GM column is a lower bound for the $A$ column, and the 2-GM\* will be a lower bound for all other columns in the table.
Our main numerical results are a similar enumeration of the number of graphs on $n \le 11$ vertices which are not determined by methods (M1)–(M4) (considering these methods separately—considering them in tandem is equivalent to just using (M4)). The ${\mathcal{Z}}$ column of Table \[tab:main\] gives the number of order $n$ graphs which are not [D${\mathcal{Z}}$]{}, i.e., not determined by (M3). (Recall that non-[D${\mathcal{Z}}$]{} implies non-DS, non-$L$-DS, and non-$|L|$-DS.) It—surprisingly to us—turned out that for $n \le 11$ being DZ\*, [DZ$\bar{\textrm{Z}}$]{} and [D${\mathcal{Z}}\bar {\mathcal{Z}}$]{} (i.e., not determined by (M1), (M2) or (M4)) are all equivalent. The number of graphs which do not satisfy these conditions is given in the ${\mathcal{Z}}\bar{{\mathcal{Z}}}$ column.
$n$ \# graphs $A$ $L$ $|L|$ ${\mathcal{Z}}$ ${\mathcal{Z}}\bar {\mathcal{Z}}$ 2-GM 2-GM\*
--------- --------------- ------------- ------------ ------------ ----------------- ----------------------------------- ------------- -----------
$\le 5$ 51 2 0 6 0 0 0 0
6 156 10 4 16 0 0 0 0
7 1,044 110 130 102 0 0 40 0
8 12,346 1,722 1,767 1,201 0 0 1,054 0
9 274,668 51,038 42,595 19,001 2 2 38,258 0
10 12,005,168 2,560,606 1,412,438 636,607 10,146 10,140 2,047,008 9,480
11 1,018,997,864 215,331,676 91,274,836 38,966,935 1,353,426 1,353,402 176,895,408 1,297,220
: Counting graphs not determined by spectral invariants (M3) is in the ${\mathcal{Z}}$ column; (M1), (M2) and (M4) are in the ${\mathcal{Z}}\bar {\mathcal{Z}}$ column
\[tab:main\]
This data suggests several things: (i) there is not much difference among (M1)–(M4); (ii) any of (M1)–(M4) seems to do much better than $A$, $L$ or $|L|$; (iii) most GM-pairs are distinguished by (M1)–(M4); (iv) most graphs not distinguished by (M1)–(M4) are explained by 2-GM\* switching. (We note that for $n=11$, 216 additional non-[D${\mathcal{Z}}\bar {\mathcal{Z}}$]{} graphs are explained by 3-GM\* switching, but none for $n \le 10$.[^2] Several more non-[D${\mathcal{Z}}\bar {\mathcal{Z}}$]{} graphs for $9 \le n \le 11$ may also be explained by Construction \[our-construction\].) These suggestions are congruous with our work in previous sections. We just state one of them as a formal conjecture.
\[conj:main\] We have: (i) almost all graphs are DZ\*; and (ii) almost all non-DS graphs are DZ\*. Further, (iii) if $\mathcal H_n$ (resp. $\mathcal H'_n$) is the set of non-DS (resp. non-DZ\*) graphs of order $n$, then $\# \mathcal H'_n / \# \mathcal H_n \to 0$ as $n \to \infty$.
The above statements are also true if we replace DZ\* with any of [DZ$\bar{\textrm{Z}}$]{}, [D${\mathcal{Z}}$]{}, and [D${\mathcal{Z}}\bar {\mathcal{Z}}$]{}.
In light of Haemers’ conjecture that almost all graphs are DS, (ii) seems to be much stronger than (i). Note (iii) is strictly stronger than (ii) as not all DS graphs are DZ\*. One reason to believe (ii) is that it appears a significant fraction of non-DS graphs are explained by GM switching. Theorem \[thm:GM-zeta\] suggests that most of these are distinguished by (M1)–(M4). Further, most of the graphs not distinguished by (M1)–(M4) are explained by GM\* switching, which appears to account for just a trivial fraction of graphs obtained by GM switching (see [@HS]). Note this reasoning does not apply to distinguishing graphs which are not $L$-DS or $|L|$-DS.
In fact, Theorem \[thm:GM-zeta\] suggests that just the Ihara zeta function $Z_G$ by itself might differentiate most pairs of graphs with the same adjacency spectrum. However, if we are just looking at $Z_G$, it is more reasonable to restrict to md2 graphs. Let us say an md2 graph is DZ if there is no other md2 graph of the same order with the same Ihara zeta function.
Setyadi–Storm [@SS] enumerated all pairs of connected md2 graphs with the same Ihara zeta function for $n \le 11$. This does not exactly tell us the number of non-DZ md2 graphs (even among connected graphs), so we did a similar enumeration to Table \[tab:main\] for md2 graphs with $n \le 10$. The results are in Table \[tab:md2\], with the column headings meaning the same things as in Table \[tab:main\], except restricted to md2 graphs. The only new column is the $Z$ column, which is the number of non-DZ md2 graphs on $n$ vertices.
$n$ \# graphs $A$ $L$ $|L|$ $Z$ ${\mathcal{Z}}$ ${\mathcal{Z}}\bar {\mathcal{Z}}$
--------- ----------- ----------- ----------- --------- -------- ----------------- -----------------------------------
$\le 6$ 77 0 0 4 0 0 0
7 510 26 64 37 0 0 0
8 7,459 744 1,156 725 2 0 0
9 197,867 32,713 31,353 13,878 6 2 2
10 9,808,968 1,727,629 1,184,460 535,080 10,130 10,094 10,088
: \[tab:md2\] Counting md2 graphs not determined by spectral invariants
We remark that while computing this data, we discovered some small errors in the tables in [@SS] for $n=11, 12$. Namely, in the cases where more than 2 connected md2 graphs have the same Ihara zeta function, [@SS] undercounts the number of pairs. For instance, for $n=10$, the entries in the [@SS] tables should be augmented by 1 for $m=20, 21, 24, 25$. In each of these cases, there is one triple of connected md2 graphs all with the same zeta function.
This data, together with our work above, suggests that, when restricting to md2 graphs, there is little difference between methods (M1)–(M4) and just using the Ihara zeta function. Hence we are led to the following conjecture.
\[conj:md2\] Almost all md2 graphs are DZ, and almost all non-DS md2 graphs are DZ. In fact, if $\mathcal H_n^{(2)}$ (resp. $\mathcal H_n^{\prime (2)}$) is the set of non-DS (resp. non-DZ) md2 graphs of order $n$, then $\# \mathcal H_n^{\prime (2)} / \# \mathcal H_n^{(2)} \to 0$ as $n \to \infty$.
Since almost all graphs are md2 for $n$ large, this first statement would imply that almost all graphs of order $n$ are determined by their Ihara zeta function.
We performed the above calculations by generating the graphs with fixed $n$ and $m$ using nauty [@nauty] in Sage 6.1.1 [@sage]. Then we used Sage to compute $\det|L| = \lim_{t \to -1} (1-t^2)^{n-m} Z_G(-1)^{-1}$ and $\det(4D+2A-3I) = (-3)^{n-m} Z_G(-2)^{-1}$ for each $G$, and similarly for $G^*$ and $\bar G$. By sorting the graphs according to these values for $G^*$ or $G$ and $\bar G$, we made a list of candidate non-DZ\* and non-[DZ$\bar{\textrm{Z}}$]{} graphs. For each of these pairs, we check to see if the corresponding Ihara zeta functions match up. We compute the Ihara zeta functions by first constructing $G^{\dagger}$, then using the Hashimoto determinant formula. This is much faster than using Sage 6.1.1’s built-in function to compute $Z_G(t)^{-1}$. To conserve memory, for $n = 10, 11$, we wrote out the data of the special values of the zeta function with the graph6 string to a file for fixed $n$ and $m$, and sorted these files using the Unix [`sort`]{} tool. Then for each pair of graphs with the same $Z_G$ and $Z_{\bar G}$, we compared their generalized characteristic polynomials ${\varphi}_{ADJ}$. This takes care of (M1), (M2) and (M4). We will return to (M3) momentarily.
This method of making a first pass by checking just 2 values of the zeta function $Z_G(-1)$ and $Z_G(-2)$ is based on the heuristic idea that it is unlikely that 2 graphs $G_1$ and $G_2$ will have the same spectra for several different random linear combinations of $A$ and $D$, unless they have the same spectra for all linear combinations of $A$ and $D$. (This is part of the reason we believe the Ihara zeta function has essentially the same distinguishing power as the Bartholdi zeta function.) From some computational experimentation, we discovered that just knowing $n$, $m$ and the 2 values $Z_G(-1)$ and $Z_G(-2)$ almost always determines $Z_G(t)$ (for $n \le 11$).
This heuristic also suggests that using 2 independent spectra to distinguish graphs is much better than a single one. To further test this idea, we enumerated graphs which are not determined by the following sets of spectra: $A$; $A$ and $L$; $A$, $L$ and $|L|$; all three of these plus the Ihara zeta function, and finally ${\varphi}_{AD}$. These numbers are respectively given in the 3rd through 7th columns of Table \[tab:combine\].
$n$ \# graphs $A$ $AL$ $AL|L|$ $AL|L|Z$ ${\mathcal{Z}}$
--------- --------------- ------------- ----------- ----------- ----------- -----------------
$\le 5$ 51 2 0 0 0 0
6 156 10 0 0 0 0
7 1,044 110 0 0 0 0
8 12,346 1,722 0 0 0 0
9 274,668 51,038 82 2 2 2
10 12,005,168 2,560,606 13,948 10,718 10,150 10,146
11 1,018,997,864 215,331,676 1,468,790 1,361,246 1,353,498 1,353,426
: \[tab:combine\] Counting graphs not determined by combining spectral invariants
Note that even though the numbers of graphs not distinguished by $A$ or $L$ or $|L|$ individually are quite large (cf. Table \[tab:main\]), the number of graphs not distinguished by combining 2 or more of these shrinks drastically. (We did not compute the numbers for combining $A$ and $|L|$ or $L$ and $|L|$ for all $n \le 11$, but we expect a similar phenomenon to hold in these cases also. For instance when $n=9$, there are only 8 graphs not distinguished by $A$ and $|L|$, and 4,405 graphs not distinguished by $L$ and $|L|$.) This suggests that, say, using $A$ and either $L$ or $|L|$, to distinguish graphs is closer in effectiveness to using (M1)–(M4) than to using just $A$ or just $L$ or just $|L|$.
The calculations for Table \[tab:combine\] were obtained by comparing spectra via spectral moments, and successively sieving out graphs. E.g., once we know all pairs or graphs with the same $A$- and $L$-spectra, we search through these to see which also have the same $|L|$-spectra. To compute the last column (i.e., (M3)), we looked at all pairs with the same $A$-, $L$-, $|L|$-spectra and Ihara zeta functions, and checked their generalized characteristic polynomials ${\varphi}_{AD}$.
We conclude with a couple of final remarks about related calculations.
First, Brouwer and Spence [@BS2] find that quite large families can have the same adjacency spectrum (46 graphs can have the same spectrum on 11 vertices, and this goes up to 128 on 12 vertices). It seems much rarer for a large family of graphs to have the same zeta functions. We found that on 10 vertices there are 4 triples of graphs with the same ${\varphi}_{ADJ}$, and no larger families. On 11 vertices, there are 1,442 triples of graphs with the same ${\varphi}_{ADJ}$, and 192 quadruples, but no larger families.
Second, one might wonder about using cones or complements with the usual spectra. Knowing the Laplacian spectrum of $G$ is equivalent to knowing it for $\bar G$ or $G^*$, but what about $A$ or $|L|$? From Corollary \[cor:cone-comp\], we know that knowing the $A$- or $|L|$-spectrum of $G$ and $\bar G$ implies the same for $G^*$ (and conversely knowing the $A$- or $|L|$-spectrum of $G^*$ and $\overline{G^*}$ implies the same for $G$). In Table \[tab:cone-comp\], we enumerate the graphs $G$ on $n \le 10$ vertices which are not determined by the following: spectrum of $G^*$, spectra of $G$ and $\bar G$, $|L|$-spectrum of $G^*$ and $|L|$-spectrum of $G$ and $\bar G$, listed respectively in columns 3 through 6. The data in the $A \bar A$ columns is already in [@HS]. We note there has been recent work towards showing almost all graphs are distinguished by the $A$- and $\bar A$-spectra—see, e.g., [@wang] or [@vDH2].
$n$ \# graphs $A^*$ $A \bar A$ $|L|^*$ $ |L| \overline{|L|}$
--------- ------------ ----------- ------------ --------- -----------------------
$\le 4$ 17 0 0 2 2
5 34 0 0 4 4
6 156 0 0 16 16
7 1,044 44 40 102 102
8 12,346 1,194 1,166 1,139 1,139
9 274,668 44,120 43,811 18,748 18,748
10 12,005,168 2,423,121 2,418,152 633,232 633,226
: \[tab:cone-comp\] Counting graphs not determined by cones and complements for $A$ and $|L|$
We remark that for $A$- or $|L|$-spectra, there appears to be little difference between using $G$ and $\bar G$ versus $G^*$ versus just $G$ (cf. Table \[tab:main\]; see [@HS], [@spence:web] for the $A \bar A$ data when $n=11, 12$). However, unlike the case of $L$-spectra, there is at least some difference. In particular, the $A$- or $|L|$-spectrum of $G^*$ does not always determine that of $G$, but Table \[tab:cone-comp\] suggests it usually does (particularly for $|L|$).
The coincidence of methods (M1) and (M2) on $n \le 11$ vertices means that for graphs with $n \le 11$ vertices of size $m$, $Z_{G^*}$ always determines $Z_G$ and $Z_{\bar G}$, and vice versa. Consequently, for $n \le 11$ vertices, knowing $Z_{G^*}$ is equivalent to knowing $Z_{(\bar G)^*}$. By analogy with the $A$-, $L$-, and $|L|$-spectra, Corollary \[cor:cone-comp\] suggests that $Z_G$ and $Z_{\bar G}$ (at least almost) always determine $Z_{G^*}$, and Table \[tab:cone-comp\] suggests that $Z_{G^*}$ almost always determines $Z_G$ and $Z_{\bar G}$. That is, we expect (M1) and (M2) to be almost always equivalent, but there seems to be no reason to expect this is always the case. Similarly, we have no reason to expect (M2) and (M4) always give the same results.
[^1]: ${}^\ast$Partially supported by a Simons Collaboration Grant.
[^2]: To enumerate graphs obtained by 3-GM\* switching, we observe that to get a non-isomorphic switch ${\widetilde}G$ from $G$, one needs at least one of the $B_i$’s to be even of size $\ge 4$, say $B_1$. One may also assume $B_2$ has size $> 1$, and the size of $B_2$ is not relatively prime to the size of $B_1$. Then for each $G$ which is not distinguished by ${\varphi}_{ADJ}$, one can check if it satisfies the 3-GM\* condition by iterating first through all “admissible” choices for $V_{B_1}$, then through choices for $V_{B_2}$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This short example shows a contrived example on how to format the authors’ information for [*IJCAI–19 Proceedings*]{} using LaTeX.'
author:
- 'First Author$^1$[^1]'
- Second Author$^2$
- |
Third Author$^{2,3}$Fourth Author$^4$\
$^1$First Affiliation\
$^2$Second Affiliation\
$^3$Third Affiliation\
$^4$Fourth Affiliation\
{first, second}@example.com, [email protected], [email protected]
title: 'IJCAI–19 Example on typesetting multiple authors'
---
Introduction
============
This short example shows a contrived example on how to format the authors’ information for [*IJCAI–19 Proceedings*]{}.
Author names
============
Each author name must be followed by:
- A newline [\\\\]{} command for the last author.
- An [\\And]{} command for the second to last author.
- An [\\and]{} command for the other authors.
Affiliations
============
After all authors, start the affiliations section by using the [\\affiliations]{} command. Each affiliation must be terminated by a newline [\\\\]{} command. Make sure that you include the newline on the last affiliation too.
Mapping authors to affiliations
===============================
If some scenarios, the affiliation of each author is clear without any further indication (*e.g.*, all authors share the same affiliation, all authors have a single and different affiliation). In these situations you don’t need to do anything special.
In more complex scenarios you will have to clearly indicate the affiliation(s) for each author. This is done by using numeric math superscripts [\${\^$i,j, \ldots$}\$]{}. You must use numbers, not symbols, because those are reserved for footnotes in this section (should you need them). Check the authors definition in this example for reference.
Emails
======
This section is optional, and can be omitted entirely if you prefer. If you want to include e-mails, you should either include all authors’ e-mails or just the contact author(s)’ ones.
Start the e-mails section with the [\\emails]{} command. After that, write all emails you want to include separated by a comma and a space, following the same order used for the authors (*i.e.*, the first e-mail should correspond to the first author, the second e-mail to the second author and so on).
You may “contract" consecutive e-mails on the same domain as shown in this example (write the users’ part within curly brackets, followed by the domain name). Only e-mails of the exact same domain may be contracted. For instance, contracting “[email protected]" and “[email protected]" is not allowed because the domains are different.
[^1]: Contact Author
|
{
"pile_set_name": "ArXiv"
}
|
**Reduction Algorithm for the NPMLE for the Distribution Function\
of Bivariate Interval Censored Data**\
**Marloes H. Maathuis[^1]\
Department of Statistics, University of Washington, Seattle, WA 98195**\
**Key words**: Computational Geometry; Maximal Clique; Maximal Intersection; Parameter Reduction.\
INTRODUCTION
============
We consider the nonparametric maximum likelihood estimator (NPMLE) for the distribution function of bivariate interval censored data. Let $(X,Y)$ be the variables of interest and let $F_0$ be their joint distribution function. Suppose that there is a censoring mechanism, independent of $(X,Y)$, so that $(X,Y)$ cannot be observed directly. Thus, instead of a realization $(x,y)$, we observe a rectangular region $R \subset
{\mathbb{R}}^2$ that is known to contain $(x,y)$. We call $R$ an *observation rectangle*. Our data consists of $n$ i.i.d.observation rectangles $R_1,\dots,R_n$, and our goal is to compute the NPMLE $\hat F_n$ of $F_0$.
Let $\mathcal F$ denote the class of all bivariate distribution functions. Furthermore, for each $F\in \mathcal F$, let $P_F(R_i)$ denote the probability that the pair of random variables $(X,Y)$ is in observation rectangle $R_i$ when $(X,Y) \sim F$. Then, omitting the part that does not depend on $F$, we can write the log likelihood as $$\begin{aligned}
\label{eq:l_n}
l_n(F)=\sum_{i=1}^n \log(P_F(R_i)),\end{aligned}$$ and an NPMLE $\hat F_n \in \mathcal F$ is defined by $$\begin{aligned}
l_n(\hat F_n) = \max_{F\in \mathcal{F}} \hspace{.1cm} l_n(F).\end{aligned}$$ As stated here, this is an infinite dimensional optimization problem. However, the number of parameters can be reduced by generalizing the reasoning of for univariate censored data. By doing this (see e.g. , , ) one can easily derive that:
- The NPMLE can only assign mass to a finite number of disjoint rectangles. We denote these rectangles by $A_1,\dots,A_m$ and call them *maximal intersections*, following .
- The NPMLE is indifferent to the distribution of mass within the maximal intersections.
The second property implies that the NPMLE is non-unique, in the sense that we cannot determine the distribution of mass within the maximal intersections. call this *representational non-uniqueness*. Hence, we can at best hope to determine the amount of mass assigned to each maximal intersection. Let $\alpha_j$ be the mass assigned to maximal intersection $A_j$, and let $\alpha=(\alpha_1,\dots,\alpha_m)$. Then $P_F(R_i)$ in equation (\[eq:l\_n\]) is simply the sum of the probability masses of the maximal intersections that are subsets of $R_i$: $$\begin{aligned}
P_F(R_i) = P_{\alpha}(R_i) = \sum_{j=1}^m \alpha_j {1_{ \{ {A_j \subset R_i} \} }}.\end{aligned}$$ We can then express the log likelihood in terms of $\alpha$: $$\begin{aligned}
\label{eq: log likelihood alpha}
l_n(\alpha) & = \sum_{i=1}^n \log (P_{\alpha}(R_i)) =
\sum_{i=1}^n \log\Bigl( \sum_{j=1}^m \alpha_j {1_{ \{ {A_j \subset
R_i} \} }}\Bigr).\end{aligned}$$ Let $\mathcal{K} = \left\{ \alpha \in {\mathbb{R}}^m: \alpha_j \ge 0,
j=1,\dots,m \right\}$ and $\mathcal{A} =\{\alpha\in {\mathbb{R}}^m \st
\mathbf{1}^T \alpha=1\}$, where $\mathbf{1}$ is the all-one vector in ${\mathbb{R}}^m$. Then an NPMLE $\hat \alpha \in \mathcal K \cap \mathcal
A$ is defined by $$\begin{aligned}
\label{eq:red opt ind}
l_n(\hat \alpha) = \max_{\alpha \in \mathcal K \cap \mathcal A}
l_n(\alpha).\end{aligned}$$ This is an $m$-dimensional convex constrained optimization problem that does not need to have a unique solution in $\alpha$. This forms a second source of non-uniqueness in the NPMLE. call this *mixture non-uniqueness*.
Asymptotic properties of the NPMLE for univariate interval censored data have been studied by . In contrast to the consistency problems of the NPMLE for bivariate right censored data, the NPMLE for bivariate interval censored data has been shown to be consistent (). This implies that the NPMLE can be used in practical applications to estimate the distribution function of bivariate interval censored data, for example to analyze data from AIDS clinical trials (see e.g. ).
From the discussion above, it follows that the computation of the NPMLE consists of two steps. First, in the *reduction step*, we need to find the maximal intersections $A_1,\dots,A_m$. This reduces the dimensionality of the problem. Then, in the *optimization step*, we need to solve the optimization problem defined in .\
In this paper we focus on the reduction step. We distinguish between two types of reduction algorithms that reflect a trade-off between computation time and space:
- type 1: The reduction algorithm computes the maximal intersections $A_1,\dots,A_m$.
- type 2: The reduction algorithm computes the *clique matrix*, an $m\times n$ matrix $C$ with elements $C_{ji}={1_{ \{ {A_j\subset
R_i} \} }}$.
For $n$ observation rectangles, the number of maximal intersections is $O(n^2)$. Hence, given the observation rectangles, one can compute the clique matrix from the maximal intersections and vice versa in $O(n^3)$ time.
We need $O(n^2)$ space to store the maximal intersections, while we need $O(n^3)$ space to store the clique matrix. Thus, type 1 algorithms require an order of magnitude less space to store the output. On the other hand, the choice of reduction algorithm determines the amount of computational overhead in the optimization step, where the values of the indicator functions ${1_{ \{ {A_j \subset R_i} \} }}$ are needed repeatedly. Namely, using a type 1 algorithm requires repeated computation of these indicator functions, while such computations are avoided with a type 2 algorithm. Thus, if we use a type 1 reduction algorithm, the computational overhead in the optimization step is increased by a constant factor.
Finally, it should be noted that the clique matrix provides useful information about mixture uniqueness of the NPMLE. For example, properties of the clique matrix are used to derive sufficient conditions for mixture uniqueness by and . We can also use the clique matrix to describe the equivalence class of solutions to . Let $\hat \alpha$ be a solution, and consider $\left(C^T
\hat \alpha\right)_i = P_{\hat \alpha}(R_i)$, $i=1,\dots,n$. Since the log likelihood is strictly concave in $P_{\alpha}(R_i)$, the vector $C^T \hat \alpha$ is unique. Thus, the equivalence class of NPMLEs is exactly the set $\{\alpha \in \mathcal K \cap \mathcal A \st C^T \alpha = C^T \hat
\alpha\}$, since all $\alpha$’s in this set yield the same likelihood value.
We now give a brief overview of existing reduction algorithms. provide a simple, but not very efficient, type 1 algorithm. introduce a type 2 algorithm of time complexity $O(n^5)$. proposes a type 1 algorithm that is of comparable speed. The algorithm with the best time complexity so far is the $O(n^3)$ type 1 algorithm of . Finally, gives an $O(n \log n)$ algorithm for a somewhat different but related problem, namely that of finding the largest number of rectangles having a non-empty intersection.
In this paper, we introduce two new reduction algorithms. The algorithm we initially developed, the *Tree* algorithm, is only mentioned briefly. It is based on the algorithm of , and is a fast but complex type 2 algorithm. Later, we realized the reduction problem could be solved in a much simpler way if one is only interested in finding the maximal intersections. This resulted in the *HeightMap* algorithm, a very fast and simple type 1 algorithm of time complexity $O(n^2)$. We discuss this algorithm in detail and also give it in pseudo code. Finally, we compare the performance of our algorithms with the algorithms of , and , using simulated data. We show that our algorithms, and especially the HeightMap algorithm, are significantly faster.
HEIGHT MAP ALGORITHM
====================
Recall that we want to find the maximal intersections $A_1,\dots,A_m$ of a set of observation rectangles $R_1,\dots,R_n$. There exist several equivalent definitions for the concept of maximal intersection in the literature. use the following: $A_j\ne \emptyset$ is a maximal intersection if and only if it is a finite intersection of the $R_i$’s such that for each $i$ $A_j \cap R_i = \emptyset$ or $A_j \cap R_i = A_j$. use a graph theoretic perspective: maximal intersections are the real representations of maximal cliques in the intersection graph of the observation rectangles.
We view the maximal intersections in yet another way. We define a *height map* of the observation rectangles. This height map is a function $h: \mathbb{R}^2 \rightarrow \mathbb{N}$, where $h(x,y)$ is defined to be the number of observation rectangles that contain the point $(x,y)$. The concept of the height map is illustrated in Figure \[fig:canonrectheightmap\]. It is easily seen that the maximal intersections are exactly the local maxima of the height map. This is true whenever there are no ties between the observation rectangles, and this observation forms the basis of our algorithm.\
![An example of six observation rectangles and their height map. The grey rectangles are the maximal intersections. Note that they correspond exactly to the local maxima of the height map.[]{data-label="fig:canonrectheightmap"}](canonheightmap.eps){width=".6\textwidth"}
Canonical rectangles
--------------------
We represent each rectangle $R_i$ as $(x_{1,i},x_{2,i},y_{1,i},y_{2,i})$. The point $(x_{1,i},y_{1,i})$ is the lower left corner and $(x_{2,i},y_{2,i})$ is the upper right corner of the rectangle. We call $(x_{1,i},x_{2,i}]$ the $x$-interval, and $(y_{1,i},y_{2,i}]$ the $y$-interval of $R_i$. Furthermore, we use boolean variables $(c^x_{1,i},c^x_{2,i},c^y_{1,i},c^y_{2,i})$ to indicate whether an endpoint is closed. As default we assume that left endpoints are open and right endpoints are closed, so that $(c^x_{1,i},c^x_{2,i},c^y_{1,i},c^y_{2,i}) = (0,1,0,1).$
We now transform the observation rectangles $R_1,\dots,R_n$ into *canonical rectangles* with the same intersection structure. We call a set of $n$ rectangles canonical if all $x$-coordinates are different and all $y$-coordinates are different, and if they take on values in the set $\{1,2,\dots,2n\}$. An example of a set of canonical rectangles is given in Figure \[fig:canonrectheightmap\].
We perform this transformation as follows. We consider the $x$-coordinates and $y$-coordinates separately and replace them by their order statistics. The only complication lies in the fact that there might be ties in the data. Hence, we need to define how to break ties. We explain the basic idea using the examples given in Figure \[fig: examples canonical\]. In (a) we have an open left endpoint $x_{1,i}$ and a closed right endpoint $x_{2,j}$ with $x_{1,i}=x_{2,j}$ and $i\ne j$. Then the $x$-intervals of $R_i$ and $R_j$ have no overlap. Therefore, we sort the endpoints so that the corresponding canonical intervals have no overlap, i.e. we let $x_{2,j}$ be smaller. In (b) we have a closed left endpoint $x_{1,i}$ and a closed right endpoint $x_{2,j}$ with $x_{1,i}=x_{2,j}$ and $i\ne j$. Now the $x$-intervals of $R_i$ and $R_j$ do have overlap. Therefore, we sort the endpoints so that the corresponding canonical intervals overlap, i.e. we let $x_{1,i}$ be smaller. In this way, we can consider all possible combinations of endpoints. By listing the results in a table, we found a compact way to code an algorithm for comparing endpoints. It is given in pseudo code (Algorithm 1).
The reason for transforming the observation rectangles into canonical rectangles is twofold. First, it forces us in the very beginning to deal with ties and with the fact whether endpoints are open or closed. As a consequence, we do not have to account for ties and open or closed endpoints in the actual algorithm. Second, it simplifies the reduction algorithm, since the column and row numbers in the height map directly correspond to the $x$- and $y$-coordinates of the canonical rectangles.
Building the height map
-----------------------
After transforming the rectangles, we build up the height map. To this end, we use a sweeping technique commonly used in the field of computational geometry (). By using this technique, we do not need to store the entire height map. Instead, we only store one column at a time, in an array $h_1,\dots,h_{2n}$. To build up the height map, we start with $h_1,\dots,h_{2n}=0$. This is column 1 of the height map. We then sweep through the plane, column by column, from left to right. Every time we move to a new column, we either enter or leave one observation rectangle. Thus, to compute the values of the height map in the next column, we respectively increment or decrement the values in the corresponding cells by 1. For example, when we move from the first to the second column of the height map in Figure \[fig:canonrectheightmap\], we enter rectangle $R_1$. $R_1$ has $y$-interval $(7,12]$ which corresponds to rows 8 to 12 in the height map. Hence, we increment $h_8,\dots,h_{12}$ by 1.\
Finding local maxima
--------------------
During the process of building up the height map, we can find its local maxima, or equivalently, the maximal intersections. We denote the maximal intersections in the same way as the observation rectangles: $A_j=(x_{1,j},x_{2,j},y_{1,j},y_{2,j})$. Suppose we apply the sweeping technique to the height map in Figure \[fig:canonrectheightmap\], and suppose we are in column 5. We then are about to leave rectangle $R_2$. The $y$-interval of $R_2$ is $(5,11]$, which corresponds to rows 6 to 11 in the height map. Hence, the values of the height map will decrease by 1 in rows 6 to 11, and will not change in the remaining rows. Since the values of the height map are going to decrease, we may leave areas of local maxima. Therefore, we need to look for local maxima in rows 6 to 11 of column 5. We find two local maxima: the cell in row 6, and the cells in rows 9 and 10. These local maxima in column 5 correspond to local maxima in the height map, say $A_1$ and $A_2$ respectively. For $A_1$, we know that $(y_{1,1},y_{2,1})=(5,6)$ and for $A_2$ we know that $(y_{1,2},y_{2,2})=(8,10)$. Furthermore, from the fact that we currently are in column 5, we know that $x_{2,1}=x_{2,2}=5$. Finally, we obtain the values of $x_{1,1}$ and $x_{1,2}$ from the left boundaries of the rectangles that were last entered. For the cell in row 6 this is $R_4$ with left boundary 4. Hence, $A_1 =
(4,5,5,6)$. For the cells in rows 9 and 10, we last entered rectangle $R_3$ with left boundary 3. Hence, $A_2=(3,5,8,10)$. From this example we see that we need an additional array, $e_1,\dots, e_{2n}$, where $e_k$ contains the index of the rectangle that was last entered in row $k$ of the height map.
After finding the first local maxima we can continue the above procedure. However, not every local maximum in the array $h$ corresponds to a local maximum in the complete height map. To illustrate this problem, suppose that we are in column 6 of the height map in Figure \[fig:canonrectheightmap\]. We then are about to leave rectangle $R_1$ with $y$-interval $(7,12]$. Applying the above procedure, we look for local maxima in rows 8 to 12 of column 6, and we find a maximum in rows 9 and 10. However, this does not correspond to a local maximum in the height map. It merely is a remainder from the maximal intersection $A_2$ that we found earlier. Namely, the local maximum in column 6 is formed by the set $\{R_1,R_3\}$ which is a subset of the set $\{R_1,R_2,R_3\}$ that forms $A_2$. We can prevent the output of such pseudo local maxima as follows. After we output a maximal intersection $A_j$, we set $e_k := 0$ for one of the rows in $A_j$. Then, a local maximum in the array $h$ corresponds to a maximal intersection if and only if $e_k >0$ for all of its cells. In the example in Figure \[fig:canonrectheightmap\], this means that after we output $A_1$ and $A_2$ we need to set $e_k := 0$ for one of their rows. $A_1$ only consists of row 6, and therefore we set $e_6 := 0$. $A_2$ consists of rows 9 and 10, and we choose to set $e_9 :=0$. Then, when we find the local maximum in rows 9 and 10 of column 6, we know it does not correspond to a maximal intersection since $e_9 = 0$.
Summarizing, we sweep through the plane from left to right, column by column. At each step in the sweeping process we either enter or leave a canonical rectangle. When we enter a rectangle $R_i$ with $y$ interval $(y_{1,i},y_{2,i}]$, we increment $h_k$ by 1 and set $e_k:=i$ for $k= y_{1,i}+1,\dots,y_{2,i}$. When we leave a rectangle $R_i$, we first look for local maxima in $h_k$ for $k=y_{1,i}+1,\dots,y_{2,i}$. For each local maximum that we find in $h$, we check whether $e_k > 0$ for all of its cells. If this is the case, we output the corresponding maximal intersection and set $e_k := 0$ for one of the cells in the local maximum. Finally, we decrement $h_k$ by 1 for $k= y_{1,i}+1,\dots,y_{2,i}$. The complete algorithm is given in pseudo code (Algorithm 2). An R-package of the algorithm is available at http://www.stat.washington.edu/marloes.
Time and space complexity
-------------------------
We can easily determine the time and space complexity of the algorithm. In order to transform a set of rectangles into canonical rectangles, we need to sort the endpoints of their $x$-intervals and $y$-intervals. This takes $O(n \log n)$ time and $O(n)$ space. At each step in the sweeping process, we need to update at most $2n$ cells of the arrays $h$ and $e$. Furthermore, we may need to find local maxima in at most $2n$ cells, and we may need to check whether $e_k >0$ for at most $2n$ cells. Thus, the time complexity of one sweeping step is $O(n)$. Combining this with the fact that the number of sweeping steps is $O(n)$ gives a total time complexity of $O(n^2)$. With respect to the space complexity, we need to store the arrays $h$ and $e$. Hence, the space complexity for computing the maximal intersections is $O(n)$. However, storing the maximal intersections takes $O(n^2)$ space.
EVALUATION OF THE ALGORITHMS
============================
We compared our algorithms with the algorithms of , , and , using simulated data. We generated bivariate current status data according to a very simple exponential model: $$\begin{aligned}
\label{model:expbcs}
X,Y,U,V \sim \exp(1),\end{aligned}$$ where $X$ and $Y$ are the variables of interest, $U$ is the observation time for $X$, $V$ is the observation time for $Y$, and $X$, $Y$, $U$ and $V$ are mutually independent. Thus, the observation rectangles were generated as follows: $$\begin{aligned}
\begin{array}{llll}
X_i \le U_i, & Y_i \le V_i & \Rightarrow & R_i=(0,U_i,0,V_i) \\
X_i \le U_i, & Y_i > V_i & \Rightarrow & R_i =
(0,U_i,V_i,\infty) \\
X_i > U_i, & Y_i \le V_i & \Rightarrow & R_i =
(U_i,\infty,0,V_i) \\
X_i > U_i,& Y_i > V_i & \Rightarrow & R_i =
(U_i,\infty,V_i,\infty)
\end{array}\end{aligned}$$ We used sample sizes $50$, $100$, $250$, $500$, $1,\!000$, $2,\!500$, $5,\!000$ and $10,\!000$. For each sample size, we ran 50 simulations on a Pentium IV 2.4GHz computer with 512 MB of RAM and we recorded the user times of the algorithms. For each algorithm, we omitted sample sizes that took over $1,\!000$ seconds to run. All algorithms were implemented in C.
The results of the simulation are shown in Table \[table: results simulation\]. We see that the Tree algorithm, and especially the HeightMap algorithm are significantly faster than the other algorithms. The HeightMap algorithm runs sample sizes of $10,\!000$ in less than two seconds.
-------- ---- ------ ---- ------ ----- ----- ----- ----- --- ------ --- ------ --- ------ --- ------ --- ------ --- ------
$n$
50 0 0004 0 0028 0 029 0 011 0 0010 0 0042 0 0012 0 0044 0 0006 0 0031
100 0 001 0 0036 0 52 0 14 0 0052 0 0079 0 0036 0 0072 0 0008 0 0040
250 0 061 0 015 26 0 47 0 0 083 0 014 0 016 0 0053 0 0018 0 0056
500 1 3 0 48 540 0 100 0 0 91 0 11 0 058 0 0087 0 0060 0 0083
1,000 46 0 63 0 13 0 1 0 0 29 0 032 0 019 0 0082
2,500 470 0 30 0 3 1 0 10 0 10 0 011
5,000 25 0 0 37 0 38 0 014
10,000 180 0 2 7 1 4 0 029
-------- ---- ------ ---- ------ ----- ----- ----- ----- --- ------ --- ------ --- ------ --- ------ --- ------ --- ------
: Mean and standard deviation of the user time in seconds, over 50 simulations per sample size from model . Cells with NA indicate that simulations took over 1,000 seconds to run and were therefore omitted.[]{data-label="table: results simulation"}
![ Log-log plot of the mean user time in seconds versus the sample size, over 50 simulations per sample size from model . For each algorithm, the estimated slope of its graph is given. These slopes can be used as empirical estimates of the time complexity of the algorithms. []{data-label="fig:reduc algorithms"}](ComparisonReductionAlgorithms.ps){width=".6\textwidth"}
To get an empirical idea of the time complexity of the algorithms, Figure \[fig:reduc algorithms\] shows a log-log plot of the mean user time versus the sample size. We fitted least squares lines through the last 4 points of each algorithm. The slopes of these lines can be used as empirical estimates of the time complexity of the algorithms. We see that the estimated slope of the HeightMap algorithm is 1.9, which agrees with the theoretical time complexity of $O(n^2)$ that we derived earlier. Furthermore, we see that the HeightMap algorithm is about an order faster than the Tree algorithm, which is about an order faster than the algorithm of Bogaerts and Lesaffre. Finally, note that the empirical time complexity of the algorithm of Bogaerts and Lesaffre is greater than the theoretical complexity of $O(n^3)$ that they derived.\
MULTIVARIATE HEIGHTMAP ALGORITHM
================================
The height map algorithm can be easily generalized to higher dimensional data. For example, for 3-dimensional interval censored data the observation sets $R_i$ take the form of 3-dimensional blocks $(x_{1,i},x_{2,i},y_{1,i},y_{2,i},z_{1,i},z_{2,i})$. In this situation the height map is a function $h: {\mathbb{R}}^3 \to \mathbb
N$, where $h(x,y,z)$ is the number of observation sets that contain the point $(x,y,z)$. The maximal intersections again correspond to local maxima of the height map. By first transforming the observation sets into canonical sets, this implies that we need to find the local maxima of a $2n \times 2n
\times 2n$ matrix. We can do this by sweeping through the matrix, slice by slice, say along the $z$-coordinate. We only store one slice of the height map at a time, so that $h$ and $e$ are now $2n
\times 2n$ matrices. At each step in the sweeping process, we either enter or leave an observation set $R_i$. When we enter an observation set, we update the corresponding values of $h$ and $e$, i.e. we set $h_{k,l} := h_{k,l} + 1$ and $e_{k,l} := i$ for all $k=x_{1,i}+1,\dots,x_{2,i}$ and $l=y_{1,i}+1,\dots,y_{2,i}$. When we leave an observation set, we look for local maxima in the cells of the rectangle $(x_{1,i},x_{2,i},y_{1,i},y_{2,i})$, using the height map algorithm for 2-dimensional data. For each local maximum that we find, we check whether $e_{k,l}>0$ for all of its cells. If this is the case, we output the corresponding maximal intersection and set $e_{k,l} := 0$ for one of the cells in the local maximum. Finally, we decrement $h_{k,l}$ by 1 for $k=
x_{1,i}+1,\dots,x_{2,i}$ and $l = y_{1,i}+1,\dots,y_{2,i}$.
For $d$-dimensional data, the time complexity of a sweeping step is $O(n^{d-1})$. Since the number of sweeping steps is $O(n)$, this gives a total time complexity of $O(n^d)$. With respect to the space complexity, we need to store the matrices $h$ and $e$. Hence, the space complexity to compute the maximal intersections is $O(n^{d-1})$. However, storing the maximal intersections takes $O(n^d)$ space.\
ACKNOWLEDGEMENTS {#acknowledgements .unnumbered}
================
This research was partly supported by NSF grant DMS-0203320. The author would like to thank Kris Bogaerts, Shuguang Song, and Alain Vandal for providing the code of their algorithms, and a referee for suggesting to consider generalizing the HeightMap algorithm to $d$-dimensional data. Finally, the author would like to thank Piet Groeneboom, Steven Schimmel and Jon Wellner for their contributions, support and encouragement.\
APPENDIX: PSEUDO CODE {#appendix-pseudo-code .unnumbered}
=====================
[CompareEndpoints]{} [$A$,$B$]{} [Two endpoint descriptors $A=(x_{k,i},c^x_{k,i})$ and $B=(x_{l,j},c^x_{l,j})$]{} [A boolean value indicating $A < B$]{}
[HeightMapAlgorithm2D]{}[$R_1,\dots,R_n$]{}[A set of $n$ 2-dimensional observation rectangles $R_1,\dots,R_n$]{} [The corresponding maximal intersections $A_1,\dots,A_m$]{}
Betensky, R. A. and Finkelstein, D. M. (1999). “A Nonparametric Maximum Likelihood Estimator for Bivariate Censored Data,” , 18, 3089–3100.
Bogaerts, K. and Lesaffre, E. (2004). “A New Fast Algorithm to Find the Regions of Possible Support for Bivariate Interval Censored Data,” , 13, 330–340.
Gentleman, R. and Geyer, C. J. (1994). “Maximum Likelihood for Interval Censored Data: Consistency and Computation,” , 81, 618–623.
Gentleman, R. and Vandal, A. C. (2001). “Computational Algorithms for Censored-Data Problems using Intersection Graphs,” , 10, 403–421.
———— (2002). “Nonparametric Estimation of the Bivariate CDF for Arbitrarily Censored Data,” , 30, 557–571.
Groeneboom, P. and Wellner, J. A. (1992). Birkhäuser, Boston.
Lee, D. T. (1983). “Maximum Clique Problem of Rectangle Graphs,” , 1, 91–107.
Song, S. (2001). Ph.D. thesis, University of Washington.
Turnbull, B. W. (1976). “The Empirical Distribution Function with Arbitrarily Grouped, Censored, and Truncated Data,” , 38, 290–295.
Van der Vaart, A. W. and Wellner, J. A. (2000). “Preservation Theorems for [G]{}livenko-[C]{}antelli and Uniform [G]{}livenko-[C]{}antelli Classes,” , 115–133, Birkhäuser, Boston.
Wong, G. Y. and Yu, Q. (1999). “Generalized MLE of a Joint Distribution Function with Multivariate Interval-Censored Data,” , 69, 155–166.
[^1]: Marloes H. Maathuis is a Ph.D. student in the Department of Statistics, University of Washington, Seattle, WA 98195 (email: [email protected]).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We model a sonic black hole analog in a quasi one-dimensional Bose-Einstein condensate, using a Gross-Pitaevskii equation matching the configuration of a recent experiment by Steinhauer \[Nat. Phys. [**10**]{}, 864 (2014)\]. The model agrees well with important features of the experimental observations, demonstrating their hydrodynamic nature. We find that a zero-frequency bow wave is generated at the inner (white hole) horizon, which grows in proportion to the square of the background condensate density. The relative motion of the black and white hole horizons produces a Doppler shift of the bow wave at the black hole, where it stimulates the emission of monochromatic Hawking radiation. The mechanism is confirmed using temporal and spatial windowed Fourier spectra of the condensate. Mean field behavior similar to that in the experiment can thus be fully explained without the presence of self-amplifying Hawking radiation.'
author:
- 'Yi-Hsieh Wang$^{1,2}$'
- 'Ted Jacobson$^{3}$'
- 'Mark Edwards$^{1,4}$'
- 'Charles W. Clark$^{1,2}$'
bibliography:
- 'HawkingCondensateFinal.bib'
title: 'Mechanism of stimulated Hawking radiation in a laboratory Bose-Einstein condensate'
---
Introduction and Summary {#sec:intro}
========================
Hawking radiation [@hawking1; @hawking2] is a pair creation process, resulting from a vacuum instability of quantum fields at a black hole horizon. The radiation is thermal, with temperature, $T = 62 \, \mathrm{ nK }\, M_{\odot} / M$, for a spherical black hole of mass $M$, where $M_{\odot}$ is the mass of the Sun. Such low temperature Hawking radiation (HR) from solar mass or larger black holes will likely never be observed. However, sonic analogs of HR can exist in hydrodynamic systems possessing a “sonic horizon," where the flow transitions from subsonic to supersonic [@unruh1981]. In such systems, the Hawking temperature is proportional to Planck’s constant times the velocity gradient at the horizon. To produce observable [*quantum*]{} HR, the system must not be much warmer than this temperature, and for this reason Bose-Einstein condensates (BECs) are a natural candidate [@robertson2012; @PhysRevD.85.024021; @PhysRevA.80.043601; @Gardiner2007; @PhysRevLett.113.090405; @PhysRevLett.85.4643]. In a sonic analog, the experimenter has access to the regions both inside and outside of the horizon, thus enabling measurements of correlations that could exhibit quantum entanglement of Hawking quanta and their interior partners [@Balbinot:2007de; @PhysRevA.80.043603; @PhysRevA.85.013621; @Boiron2015; @Steinhauer:2015saa; @Steinhauer:2015ava].
{width="7.2in"}
In 2014, Steinhauer [@nphys3104] reported on a remarkable experiment that implements a sonic analog black hole in a needle–shaped BEC of $^{87}$Rb. The BEC is swept by a negative step potential with the energy equivalent of a few nK. The sweep generates a low-density “cavity" of supersonic flow in the interior of the BEC, bounded by black-hole (BH) and white-hole (WH) horizons \[Figs. \[fig:exp\](a)-(c)\]. Steinhauer observed exponential growth of a standing wave between the horizons, and measured the density-density, two-point, connected correlation function, which displayed correlations between points within the cavity, as well as between a point in the cavity and a point outside the BH. He interpreted the latter as a signal of Hawking radiation phonons correlated with their partners behind the horizon, and he interpreted the growing standing wave and internal correlation function as evidence for self-amplifying Hawking radiation.
In a flow with a supersonic cavity bounded by BH and WH horizons, Hawking radiation can be self-amplifying if the phonon group velocity becomes supersonic at high wavenumbers[@jacobson1999]. The negative energy partner of a Hawking phonon is trapped in the cavity, so that the parent state for subsequent Hawking radiation is no longer the vacuum, but instead is an excited state. This produces stimulated emission of Hawking radiation, which amplifies the trapped negative energy mode. The repetition of this process leads to exponential growth of the negative energy mode and the associated Hawking emission. This is called the “black hole laser" mechanism[@jacobson1999]. The behavior of this lasing mechanism, and its potential role in laboratory realizations of Hawking radiation has been extensively investigated[@Coutant:2009cu; @parentani2010]. It could enhance the signal of Hawking radiation but, even if the initial trigger for the lasing were spontaneous emission of Hawking radiation, the amplified signal, once it had grown significantly, would be a coherent state of phonons which would be difficult to distinguish from a classical wave. Moreover, a lasing mode could be excited by a classical seed.
In the work reported here, we have investigated the dynamics of the condensate of the experiment of Ref. [@nphys3104], using primarily the one dimensional, time-dependent Gross-Pitaevskii (GP) equation. (We used the three-dimensional GP equation only to check that the qualitative features of the dynamics are the same as for the one-dimensional case.) The GP equation is a nonlinear Schrodinger equation, approximating the behavior of the expectation value of the many body field operator, which captures the classical, hydrodynamic aspects of the BEC, as well as interference phenomena. Effects of quantum fluctuations can be treated approximately by adding to the GP wavefunction an initial distribution of random fluctuations with the Gaussian statistics of the zero point fluctuations, and averaging over an ensemble of realizations. That is called the truncated Wigner approximation (TWA). In this paper we do not include the fluctuations, because our purpose here was to understand first the mean field behavior entailed by the experimental conditions.
Our study has revealed the importance, in configurations similar to those in the experiment, of two features that are not present in previous, idealized studies of sonic analog black holes: (i) the condensate density increases towards the center of the atom trap and, (ii) as a result of this inhomogeneity, and the shape of the trap potential, there is no single inertial frame in which the condensate is stationary. A key consequence of the latter feature is that the white hole horizon, defined in the locally relevant sense explained below, is not at rest with respect to the black hole horizon.
As a result of these features, growth of a standing wave and emission of Hawking radiation both occur purely hydrodynamically, in a way that appears to be similar to what was observed in the experiment. As described below, we have identified the mechanisms producing these simulated phenomena, and we find that the black hole laser mechanism plays no role. This raises doubts concerning whether lasing actually plays any role in the experiment. Even though our simulations do not take quantum fluctuations into account, the mean field behavior they reveal would persist in the presence of quantum fluctuations, and it already seems to account for the observed mean field behavior. On the other hand, even if the conditions for lasing existed in the system, we may not have seen that instability because of the absence of the necessary seed fluctuations.
There is an important feature of the experimental measurements that our simulations in this paper do not address, and that is the connected density-density correlation function. This is because our simulations are deterministic. In order to understand the relation between this correlation function and the mean field behavior, as well as to check for lasing instability in the presence of quantum fluctuations, in another paper [@correlations] we have studied the system in the presence of both quantum fluctuations, and fluctuations in atom number from one run to the next. We find there that the correlation function is engendered by fluctuation-induced modulation of the deterministic standing wave. In fact, it results primarily from the varying number of atoms in the condensate. When averaging over GP simulations with a 10-20% variation in the number of atoms, and no quantum fluctuations, a correlation function similar to the measured one is produced. The addition of quantum fluctuations improves the agreement, but their effect is sub-dominant. That is, we find that the observed correlation arises by modulation of the standing wave, and is unrelated to the intrinsic quantum correlations that would be present without the standing wave. In particular, we find that the quantum fluctuations do not trigger any instability.
Returning now to the results of the simulations reported here, let us summarize our findings. As the potential step is swept through the condensate, a growing hydrodynamic standing wave (Fig. \[fig:GP1\]) arises between the horizons, which appears similar to the one observed in the experiment. This wave is Bogoliubov-Čerenkov radiation (BCR) [@Carusotto], generated at the WH [@Coutant:2012zh; @Busch:2014hla; @Mayoral2011] \[see Figs. \[fig:exp\](d) and (e)\]. It is also known as a zero-frequency undulation, and is reminiscent of a ship’s bow wave.
Three independent lines of evidence all indicate the BCR nature of this standing wave. First, the growth rate of the standing wave (Fig. \[fig:GP\_growth\]) matches very closely the square of the background density, which changes as the step sweeps into denser parts of the BEC, as would be expected from the BCR mechanism. (The time dependence of this growth also appears roughly compatible with that observed in the experiment.) Second, a spacetime plot of the magnitude of the deviation of the GP wave function from the background flow \[Fig. \[fig:stimulated\_HR\](b)\] shows that the standing wave first arises at the WH, and then propagates to the BH \[as illustrated in Fig. \[fig:stimulated\_HR\](a)\]. And third, the standing wave has zero frequency in the WH frame. This is evident by inspection of Fig. \[fig:stimulated\_HR\](b), which shows that the lines of constant phase are parallel to the WH horizon worldline.
Figures \[fig:stimulated\_HR\](a) and (b) also reveal that the WH horizon has a smaller velocity than the BH horizon. The BCR is therefore Doppler shifted to a nonzero frequency in the BH frame (which is the rest frame of the potential step), as can also be seen in the figure. Although the relative velocity of the horizons is small, this Doppler shift is larger than might be expected, because the BCR has a very large wave vector. With its nonzero frequency in the BH frame, the BCR classically stimulates emission of Hawking radiation at the BH horizon, i.e., “pair production" of an outgoing wave and partner radiation inside the horizon, as seen in Fig. \[fig:stimulated\_HR\](b). We verified that the partner has the same frequency as the Doppler-shifted BCR, using a temporal windowed Fourier transform, Figs. \[fig:stimulated\_HR\](c) and (d). This shows that only two frequencies are present: that of the background flow, and that of the BCR. It may be possible to perform a similar analysis on the experimental data.
We also explored trap parameters near those that roughly matched the experiment, seeking a regime that could yield a more distinct signal both for our analysis and in future experiments, We identified a slightly modified regime, in which all features are qualitatively the same as those in the experimental regime, only much sharper \[see Fig. \[fig:M20\](b)\]. In the modified regime we made a thorough spectral analysis of the condensate, using both temporal and spatial windowed Fourier transforms \[Figs. \[fig:STFT\_wt\](b) and (c)\]. This enabled us to establish that, despite the significant inhomogeneity in the system, the Bogoliubov–de Gennes dispersion relation gives a remarkably accurate prediction for the temporal and spatial spectral content of the BEC, with the only inputs being (1) the assumption that the BCR has zero frequency in the WH frame, and (2) the local sound speed and flow velocity. \[We used spatial windowed Fourier transform of the density and of the GP wave function to identify the background flow and determine its density (for the sound speed) and wavevector (for the velocity).\] This detailed spectral analysis gives us confidence that there is nothing going on beyond the mechanisms we have identified. It also allows us to establish that the Hawking temperature prediction is consistent with the relative amplitude of the Hawking radiation and partner waves, insofar as would be expected. More generally, it reveals the utility of windowed Fourier transform in characterizing the local structure of an inhomogeneous BEC flow.
Finally, we conclude this introduction by mentioning that related studies having some overlap with ours have been reported in Refs. [@Tettamanti; @SR]. We comment on the relation between that work and our conclusions in Sec. \[sec:comparison\].
Methods {#sec:method}
=======
Reference [@nphys3104] reported a step-sweeping experiment on a quasi-one-dimensional (quasi-1D) condensate. A detailed discussion of our approach to simulating this experiment is given in Appendix \[app:sim\_details\]. We describe a three-dimensional (3D) model of the experimental potential and condensate in Appendix \[app:1dGPE\], while in Appendix \[app:3dGPE\] we discuss and evaluate the criteria for applicability of the reduction to a 1D model, and compare with a simulation using the 3D GPE. We find that the qualitative features in the 1D simulation are consistent with the results of 3D simulation. This indicates that the results of our detailed analysis of the 1D simulation should apply as well to the 3D system.
The GP wavefunction, $\Psi(x,t)$, in our simulation involves a condensate component and the excitation modes generated during the sweep: $$\begin{aligned}
\label{eq:psi0}
\Psi(x,t)=\Psi_{\rm bf}(x,t)+\sum_{j}\psi_{j}(x,t),\end{aligned}$$ where $\Psi_{\rm bf}$ indicates the background condensate flow, and $\psi_{j}$ denotes its excitation mode satisfying the Bogoliubov–de Gennes (BdG) equation [@pethick]. In our simulation, there are three modes that we observed: the BCR mode, the HR mode, and the HR partner (labeled by $j=$ BCR, HR, and p, respectively). The role of each mode will be explained in the later sections. In regions where the flow is slowly varying, each component in Eq. \[eq:psi0\] behaves locally as a Wentzel-Kramers-Brillouin (WKB) plane wave with a characteristic frequency $\omega$ and wavevector $k$. Here we introduce two techniques to resolve individual components and their spectral properties.
Windowed Fourier transform {#sec:WFT0}
--------------------------
A windowed Fourier transform (WFT) [@gomes99fourier] is a method that brings out the “local" spectral elements of a function in the neighborhood of a given position or time. It differs from the normal Fourier transform by including a Gaussian function centered at the position ($x$) or time ($t$) of interest. The spatial WFT $F(k,x)$ of a function $f(x)$ is defined as: $$\begin{aligned}
\label{eq:FTkx0}
F(k,x)= \int^{\infty}_{-\infty} dy\, f(y)w(y-x;D)e^{-iky},\end{aligned}$$ where $w(y-x;D)=\exp (-(y-x)^2/D^2)/\left(\sqrt{\pi}D\right)$ is a Gaussian window function of width $D$. With the filtering of the window, the transformed function $F(k,x)$ constitutes a local Fourier transform of $f(x)$, capturing features that vary on length scales much smaller than $D$. For instance, given a function, $f(x)=f_{q}(x) \exp (iqx)$, with wavevector $q$ and slowly varying amplitude $f_{q}(x)$, the transformed function is $F(k,x)\approx f_{q}(x) \exp (-(k-q)^2(D/2)^2)$ : a Gaussian in $k$-space, centered at $k=q$ with width $2/D$, and the peak height is the local amplitude, $f_{q}(x)$.
The WFT is able to resolve locally (at a given $x$ or $t$) the Fourier components in Eq. \[eq:psi0\] as peaks in the resulting wavevector (or frequency) spectrum, in which peak position and height indicate the wavevector and amplitude of each component. Specifically, for a background condensate flow, $\Psi_{\rm bf}\sim |\Psi_{\rm bf}(x)|e^{ik_{\rm bf}(x)x}$, its spatial WFT exhibits the local wavevector $k_{\rm bf}(x)$ for each $x$, which determines the local flow velocity in the laboratory frame, $\hbar k_{\rm bf}(x)/m$. This is shown as the main streak in Fig. \[fig:exp\](b). Similarly, the spatial WFT of the density, $n(x)=|\Psi(x)|^2$, separates the background condensate density $n_{\rm bf}(x)$ ($k=0$) from the superimposed spatial oscillations (with nonzero $k$). An example is shown in Fig. \[fig:GP\_growth\](b), which is the spatial WFT of density profile in Fig. \[fig:GP1\](g), evaluated at the center of the oscillatory region. The spectrum has a central peak $n_{\rm bf}$ as the background density, and two side peaks $n_k$, indicating the oscillatory component.
Moving average of the GP wavefunction {#sec:smoothing0}
-------------------------------------
To separate fast-oscillating components in Eq. \[eq:psi0\] from the slowly-varying parts, we implement a smoothing procedure on the GP wavefunction. The procedure is equivalent to calculating the moving average of a discrete data set, which smooths out short-range fluctuations. Here the moving average of wavefunction $\Psi(x)$ is defined as $$\begin{aligned}
\label{eq:smoothing0}
\bar{\Psi}(x)=\frac{1}{2D_s} \int^{x+D_s}_{x-D_s} dy\, \Psi(y),\end{aligned}$$ where the integral serves as a square window of width $2D_s$ centered at $x$, over which $\Psi(x)$ is being averaged. For components in $\Psi(x)$ with wavelength shorter than $D_s$ (i.e., $D_s>\pi/k$), the integral would give rise to an average of zero, leaving those that are slowly varying in space (i.e., $\pi/k > D_s$) in $\bar{\Psi}(x)$, and the difference $\delta\Psi \equiv \Psi -\bar\Psi$ characterizes the part of $\Psi$ composed roughly of wavevectors $k\gtrsim \pi/D_s$. Later in Sec. \[sec:BCRStim\], we use the above procedure at each time and exhibit a spacetime diagram of $|\delta\Psi(x,t)|$, in which the slowly-varying part of the background flow ($\Psi_{\rm bf}$ with $k\sim 0$) is removed to bring out $\psi_{\rm HR}$.
Analysis of the simulated experiment {#sec:experiments}
====================================
Formation of the BH-WH cavity
------------------------------
In the experiment of [@nphys3104] a BH-WH cavity is established in a quasi-one-dimensional, laboratory BEC held by a confining potential, $U(x)$, as shown in Fig.\[fig:exp\](a). By sweeping a potential step of depth $U_{s}$ at uniform speed $v_{s}$ across the BEC, BH and WH are established ($U_s$ is on the order of $10^{-9}$ K, and $v_s$ is 0.21 mm/s). Atoms are accelerated in the direction opposite to the step motion due to the precipitous drop in the potential. This creates a supersonic flow behind the step and forms a BH at the step edge, $x_{\mathrm{BH}}$. The accelerated atoms gradually slow as they recede from the step, due to the rising potential. This causes the flow to become subsonic at a critical distance $L$ behind the step, forming a WH, $x_{\mathrm{WH}}$. Not far beyond $x_{\rm WH}$ the flow velocity in the laboratory frame drops to zero, roughly where $U(x)-U_{s}=U(x_{\mathrm{BH}})$ (this implies that $L$ increases slightly as $x_{\mathrm{BH}}$ moves toward the center of the trap). This procedure produces the flow structure shown in Fig. \[fig:exp\](c).
To determine the flow structure, we implement the spatial WFT described in Sec. \[sec:WFT0\]. Figure \[fig:exp\](b) is a local wavevector spectrum $|\Psi(k,x)|^2$ with $D=5$ $\mu$m, defined in the laboratory frame at a moment during the sweep \[Fig. \[fig:GP1\](e)\]. There is a dominant streak, indicating the background condensate flow, $\Psi_{\rm bf}$, for which the peak position at each $x$ defines the background wavevector, $k_{\rm bf}(x)$. The regions with zero wavevector, $k_{\rm bf}\sim 0$, correspond to the non-accelerated, subsonic BEC; the region behind the step with $k_{\rm bf}\sim -1.4$ $\mu$m$^{-1}$ corresponds to the accelerated, supersonic flow. The blue (dark gray) curve in Fig.\[fig:exp\](c) is minus the flow velocity in the rest frame of the step, $v(x)=v_{\rm bf}(x)-v_{\rm s}$, where $v_{\rm bf}(x)=\hbar k_{\rm bf}(x)/m$ is the background flow velocity in the laboratory frame. The green (light gray) curve is the local speed of sound $c(x)=\sqrt{g{n_{\rm bf}}(x)/m}$, where $g$ is a coupling constant defined in Appendix \[app:1dGPE\], and ${n_{\rm bf}}(x)$ is the local density of the background flow, which we identify here using a WFT of the density $n(x)$ (see Appendix \[app:FT\_vc\] for details).
### Locating the black and white hole horizons {#sec:Locating}
The black hole horizon is defined as the location where a right moving phonon is at rest in the step frame. This corresponds to the right intersection of $c(x)$ and $-v(x)$ in Fig. \[fig:exp\](c). The step frame is distinguished as the one in which the system is closest to being stationary near the step. In particular, $c(x)$ and $v(x)$ are nearly steady where the trapped BEC spills over the moving step.
The definition of the WH horizon is not as simple, because the step frame is not a global stationary frame of the system, due to the spatial variation in the trap potential and background condensate density. Instead, what is dynamically important is the location of the transition from supersonic to subsonic in the frame in which conditions are [*locally*]{} stationary. In particular, this is the locus of Čerenkov radiation, which arises from the accessibility of negative energy modes in a frame in which energy is conserved, i.e., in which conditions are stationary.
At early times in the sweep of the step, the WH horizon so defined is located where the BEC density, and therefore the sound speed, is significantly smaller. It therefore starts out moving much more slowly than the step. As the sweep progresses it accelerates smoothly, until it reaches a uniform velocity slightly less than that of the step. It is then approximately located at the left intersection of $c(x,t)$ and $-v(x,t)$, and the distance $L$ between the two horizons grows slowly and uniformly in time. This behavior can be seen in Figs. \[fig:stimulated\_HR\](b) and \[fig:M20\](b).
Cavity standing wave
--------------------
Figure \[fig:GP1\] shows comparisons of the simulated density profile with experiment. Figures (a)-(g) show the BEC density for $U_s = k \times 6$ nK after the launch of a sweep at 20 ms intervals, where $k$ is the Boltzmann constant. Figure (h) corresponds to the density profile at $t=$ 120 ms for $U_s = k \times 3$ nK. The coordinate origin in each panel has been displaced to coincide with $x_{\rm BH}$.
![\[fig:GP1\] (a-g) Density vs. time of a swept BEC at 20 ms intervals with step $U_s/k=\text{ 6 nK}$, scaled by a common factor to match experiment, and viewed in the moving frame where $x=0$ defines the step edge; panel (h): $U_s/k =\text{ 3 nK}$ at 120 ms. Blue (dark gray): experiment [@nphys3104]; red (gray): present simulation. ](steinhauer_evolution_density_072417){width="3.5in"}
The density exhibits a standing-wave pattern behind the step, with amplitude growing in time. Considering that the experimental observations involve an average over any quantities that fluctuate from one run to another, the GP simulation qualitatively matches the overall evolution seen in the experiment. In particular, the growth, wavelength, and phase of the wave pattern are similar to each other.
Čerenkov mechanism {#sec:mechanism}
------------------
In the following, we present multiple lines of evidence showing that the standing wave results from the Bogoliubov-Čerenkov radiation (BCR) effect, in a process closely analogous to the flow past an obstacle studied in [@Carusotto]. This evidence is based on the wavevector and frequency spectra of the standing wave, and the growth rate of the standing wave, which we will show is due to the increasing BEC density. We also establish that the Hawking radiation is emitted by this system and that the partner mode slightly modulates the standing wave.
As illustrated in Fig.\[fig:exp\](d), an obstacle in a stationary supersonic flow produces an upstream, Bogoliubov-Čerenkov standing wave [@Leboeuf2001; @Carusotto], analogous to a bow wave on water [@Carusotto:2012fy]. It was observed in Ref. [@Mayoral2011] that such a standing wave is generated at a WH, triggered by an incident wavepacket on the stationary flow, and saturating at an amplitude determined by nonlinear effects. A similar standing wave, generated by inhomogeneity at a WH horizon, can be seen in Fig. 4 of Ref. [@PhysRevA.91.053603]. In our case, the subsonic component to the left of the WH serves as an obstacle in the supersonic flow, generating a Bogoliubov-Čerenkov wave.
### Wavevector spectrum {#sec:kspectrum}
In the WFT spectrum in Fig. \[fig:exp\](b), we observe in the cavity region an excitation mode at $k\sim1.4$ $\mu$m$^{-1}$ coming from the WH, which is roughly the reflection of the supersonic flow $\Psi_{\rm bf}$ with $k\sim-1.4$ $\mu$m$^{-1}$. The interference of the two results in the standing wave in the density profile shown in Fig. \[fig:GP1\](e), which has a wavevector with twice the above value, $k\sim3$ $\mu$m$^{-1}$.
The relation $k_{\rm BCR} \sim-k_{\rm bf}$ is expected from energy conservation: the flow structure is approximately time-independent in the rest frame of the WH horizon, so we expect that the BCR production process should conserve energy. One way to view it is that incoming atoms reflect from the flow transition at the horizon. The velocity of the WH horizon is quite low compared to the flow velocity upstream, so to a good approximation energy conservation in the WH frame implies that the laboratory frame wavevector should simply reverse sign. A more precise account of this given in Sec. \[sec:M2\] below, using a linearized mode analysis.
Another consequence of the approximate local time independence in the WH frame is that the frequency of the standing wave generated there should have zero frequency in that frame. Indeed it does, but we postpone the demonstration of that to later in this section.
### Growth of the standing wave {#comparison}
![\[fig:GP\_growth\] (a) Simulated growth of the standing-wave pattern in the supersonic region for $U_{s}/k =$ 6 nK. Solid green: normalized standing-wave amplitude $\bar{n}_k(t)$, $\bar{n}_{k}(t)=n_{k}(t)/n_{k}(0)$, for which $\ln[\bar{n}_{k}(120)]\sim 4.4$. Dashed black: the square of background density, $\bar{n}_{\rm bf}(t)$, scaled to match the final standing-wave amplitude, $\bar{n}_{\rm bf}^2(t)=n_{\rm bf}^2(t)[\bar{n}_{k}(120)/n_{\rm bf}^2(120)]$. The growth of $n_{\rm bf}$ and $n_{k}$ is determined from a spatial WFT of $n(x)$ at $x=-12$ $\mu$m with window width $D=6.5$ $\mu$m. Inset (b) shows the windowed wavevector spectrum at $\text{t=120 ms}$.](steinhauer_evolution_growth_072417){width="3.in"}
Reference [@nphys3104] reported exponential growth of the oscillatory density pattern in the BH-WH cavity, and suggested that it results from the black hole laser effect. Our simulations exhibit similar growth, but lead us to attribute it to a different mechanism. Figure \[fig:GP\_growth\] displays the growth of the background flow density $n_{\rm bf}$ and of a standing wave, $n_k$, defined by the peaks of the WFT of the density at $x_{\rm BH}-12 \mu$m, as shown in the inset. Note that this spectrum is different from the one in Fig. \[fig:exp\](b), which is the squared modulus of the spatial WFT of the wavefunction. Over 120 ms the standing wave density grows by $\sim \exp(4.4)$. Figure \[fig:GP\_growth\] shows that $n_k$ grows in proportion to $n_{\rm bf}^2$. The two oscillation features superimposed on the growth curve coincide with the variations seen in Fig. \[fig:stimulated\_HR\](b), and will be discussed in Sec. \[sec:comparison\].
To understand this quadratic relationship between the standing wave amplitude and the background flow density, we begin by noting that the step moves toward the region of higher BEC density \[see Fig. \[fig:exp\](a)\], so the background density $n_{\rm bf}$ also grows in the cavity. Now the saturated amplitude ($n_k$) of a BCR standing wave should be proportional to both the strength of the obstacle ($V_{\rm ob}$), and the density of the background flow ($n_{\rm bf}$) [@Carusotto; @pethick] \[see Fig.\[fig:exp\](e)\]. The “obstacle" in the present case has a strength proportional to the BEC density to the left of the WH, which grows similarly to that on the right, so it follows that the saturated wave amplitude $n_k$ should grow as $n_{\rm bf}^2$.
The very close agreement with this scaling relation displayed in Fig.\[fig:GP\_growth\](a), gives further compelling evidence that the standing wave observed inside the supersonic cavity is in fact BCR, and indicates that its growth results from the increase of background density, rather than from a black hole laser instability. Moreover, effects due to classical or quantum fluctuations, not included in our simulation, could not remove this robust, large BCR wave, but rather would have to appear in addition to it.
The absence of fluctuations in our simulation here implies that we are unable to capture the behavior of the density-density correlation function. That correlation function was measured in the experiment [@nphys3104], and displays a checkerboard pattern with periodicity very close to that of the standing wave. The growth of the checkerboard pattern was quantified in [@nphys3104] via the Fourier power spectrum of the correlation, and found to grow by a factor $\sim\exp(3.3)$. In [@correlations] we have shown, by introducing quantum and atom-number fluctuations into our simulations, that this checkerboard pattern results directly from the presence of the underlying BCR standing wave, modulated by the fluctuations.
### Spacetime portrait {#sec:BCRStim}
In this subsection we present a spacetime portrait for the evolving BEC. This portrait illustrates by visual inspection that the standing wave is generated from the WH, and has zero frequency in the WH reference frame. Its frequency in the step frame is nonzero, due to a Doppler shift arising because the WH recedes from the BH as the system evolves. The spacetime portrait also reveals a signal of Hawking radiation, which is stimulated by the BCR at the BH. We further verify this mechanism quantitatively through a windowed frequency spectrum evaluated inside the cavity, which reveals that the only frequencies present are those of the background condensate and the BCR.
{width="7in"}
The spacetime portrait for the simulated experiment, Fig. \[fig:stimulated\_HR\](b), displays $|\delta\Psi|=|\Psi-\bar\Psi|$. To resolve the HR outside the cavity, we subtract the moving average $\bar\Psi$. using the procedure described in Sec. \[sec:smoothing0\]. $\bar\Psi$ approximates the dominant, slowly-varying background \[with $k\sim0$ as in Fig. \[fig:exp\](b)\]. We choose the smoothing window $D_s=5.4\ \mu$m, such that it is large enough so that $\bar\Psi\approx0$ between the horizons, yet small enough to capture the slow variations of the background outside the horizon.
The portrait displays an interference pattern between the background supersonic flow and excited modes of $\delta\Psi$. The evolution of the BH is indicated by the diagonal orange (right) line. To clearly display HR upstream of the horizon, we have multiplied $|\delta\Psi|$ there by a factor of 10. At the beginning of the evolution, as the condensate spills over the step, a left-moving flow develops, indicated by the growing light gray area. When this flow reaches the WH, at $t\approx 10$ ms, a standing wave (BCR) is generated. In Fig. \[fig:stimulated\_HR\](b), it is clear by inspection of the dashed red line and solid red (left) line \[which is parallel to diagonal orange (right) line\] that the standing wave has zero frequency in the WH rest frame, but nonzero frequency in the BH rest frame. Since its frequency is nonzero in the BH frame, the BCR can stimulate production of Hawking pairs at the BH horizon. (Zero frequency Hawking pairs do not exist.)
The stimulated HR is seen in the spacetime portrait Fig. \[fig:stimulated\_HR\](b). The BCR first reaches the BH at $t\approx 20$ ms, stimulating emission of HR. Hawking radiation first appears at around 25 ms, but is not visible on the grayscale plot until around 40 ms. The left–moving partner radiation (p-mode) resulting from the “pair creation" forms a “V"-shape with the HR, and makes an interference pattern with the BCR that can first be seen around $t\approx 40$ ms. (Were there no mode present to stimulate the pair creation, it would nevertheless occur spontaneously, as in the Hawking effect for an astrophysical black hole.)
Figure \[fig:stimulated\_HR\](a) is a schematic illustration of the mechanism just described. This is viewed in the laboratory frame, where the BH moves at velocity $v_s$ and the WH with a slightly smaller velocity, $v_s-\Delta v$, as indicated by the dashed red line. As seen by inspection of Fig. \[fig:stimulated\_HR\](b), the BCR (i.e., the standing wave) has zero phase velocity with respect to the WH, corresponding to zero frequency in the WH frame. Since the WH velocity is less than the BH velocity \[as shown in Figs.\[fig:stimulated\_HR\](a) and (b)\], this gives rise to a nonzero frequency in the BH frame. Note that, although the relative velocity of the BH and WH is rather small, the BCR wavelength is rather short, so that the BCR frequency in the BH frame is not small. As the BCR mode ($\psi_{\rm BCR}$) propagates to the BH, stimulates the emission of HR ($\psi_{\rm HR}$) and its partner ($\psi_{\rm p}$) at the latter frequency, with the associated wavevectors determined by the Bogoliubov-de Gennes (BdG) spectrum.
### Frequency spectrum
In Fig. \[fig:stimulated\_HR\](c) and (d), we show the windowed frequency spectrum in the experimental regime in the supersonic region. The frequency is computed in the rest frame of the moving step, along the solid red (left) line in Fig. \[fig:stimulated\_HR\](b). The long streak, which starts from the beginning of the evolution, corresponds to the background flow, $\Psi_{\rm bf}$. The short streak, which is separated from the long streak by $\Delta\omega\sim 0.11(3)$ rad/ms, corresponds to the BCR and the p-mode. $\Delta\omega$ is nonzero because of the Doppler shift between the WH and BH frames.
The Doppler effect due to the recession of the WH can be estimated using the velocity difference between the two horizons, $\Delta v\sim0.03$ mm/s. The shifted frequency is the product of $\Delta v$ and the BCR wavevector, $\Delta k_{\rm BCR}$, $$\label{eq:shift}
\Delta \omega=-\Delta k_{\rm BCR}\Delta v=- 0.09 \mbox{ rad/ms}.$$ We find that the WFT frequency agrees with the prediction in Eq. \[eq:shift\] to within the uncertainty.
The quantitative agreement with zero WH frequency shifted to the BH frame establishes that the mechanism illustrated in Fig. \[fig:stimulated\_HR\](a) is operative. In particular, the Doppler shifting of frequency between the two horizons plays an essential role in the process, and the partner mode of the Hawking radiation has the same frequency as that of the BCR. The fact that the partner waves match this frequency shows that they are stimulated by the BCR, rather than being self amplifying.
Enhanced parameter regime {#sec:M2}
=========================
In addition to simulating the system using parameters close to those of the experiment, we have explored a different parameter regime, in which the phenomena observed in the experiment, in particular the Hawking radiation, are more sharply displayed. This was helpful in developing an understanding of the behavior of the system, and it may prove useful for optimizing the Hawking radiation signal in future experiments.
In the experimental regime, the signal of HR is too weak to be directly seen in the density profile in Fig. \[fig:GP1\]. With the help of the spacetime portrait in Fig. \[fig:stimulated\_HR\](b), one observes emission from the BH that resembles HR, but with irregularity. This irregularity may be due to the long wavelength of the p-mode, $\lambda_{\rm p}$, relative to the cavity size, $L$. Since $\lambda_{\rm p}\sim L$, the p-mode does not behave as a WKB mode on a slowly-varying background. This, in concert with the time dependence of the cavity size, may lead to the irregular emission of the HR mode.
In the enhanced parameter regime, we lower the $\lambda_{\rm p}/L$ ratio by modifying the parameters of the trapping potential (axial trap frequency $\omega_x$) and the step potential ($U_s$ , $v_s$). Figure \[fig:M20\](a) shows the density profile in one such modified regime (case M2), from which sharper signals of HR and p-mode have been observed, with suppressed $\lambda_{\rm p}/L$ ratio. In this case, the BEC is twice as long as in the experiment of [@nphys3104], the step size is halved relative to the 6 nK step, and the step speed is about the same. The details of the investigation of parameter regimes are summarized in Appendix \[app:parameter\].
The spacetime portrait of the modified regime is shown in Fig. \[fig:M20\](b). The BCR-stimulated pair production mechanism illustrated in Fig. \[fig:stimulated\_HR\](a) can be seen very clearly, with more distinct features than in the experimental regime \[Fig. \[fig:stimulated\_HR\](b)\]: (i) the BCR, with phase parallel to the WH, which grows substantially prior to the pair creation, (ii) the “V"-shaped HR pair, stimulated by the BCR. Note that the frequency of $|\delta\Psi|$ appears doubled outside the BH compared to that inside. This is because $\delta\Psi$ contains very little background flow component with which to interfere outside the BH, so the visible interference is between the positive and negative relative frequency parts of the HR.
Furthermore, since the HR and the p-mode have enhanced signals and regular wavelengths here, their spectral properties can be captured by WFTs. In the following, we analyze the properties of the modes based on the Bogoliubov-de Gennes theory.
![\[fig:M20\] Stimulated pair production in the enhanced regime, M2. Panel (a): density $n(x)$ at $t = 650$ ms, along the horizontal green line in (b); panel (b): spacetime portrait. The diagonal red (left) and blue (right) lines indicate the paths on which the windowed frequency spectra of Fig. \[fig:STFT\_wt\](b) are calculated. The wavevector spectrum along the horizontal green line is shown in Fig. \[fig:STFT\_wt\](c). The magenta (left) dot and cyan (right) dot correspond to a correlated Hawking pair, for which the thermal prediction is being tested.](bhbec_evolution_M2_072517){width="3.1in"}
{width="7.3in"}
BdG mode analysis {#sec:BdG}
-----------------
The BdG theory of linearized modes [@pethick] (Appendix \[app:BdG\]) can be used to predict the temporal and spatial WFT spectra of the BEC, starting from only one input assumption: that the standing wave has zero frequency in the WH frame. This will further verify the mechanism we have proposed for the excitations of the BEC. In addition, it will demonstrate the remarkable accuracy of BdG analysis when combined with WFT in an inhomogeneous setting.
To make contact with the notion of linearized, BdG modes and their dispersion relation, we locally factor the full GP wavefunction $\Psi(x,t)$ into a homogeneous background $\Psi_{\rm bf}$ and the deviation $\psi$, so that the deviation is locally a superposition of harmonic modes of the form $$\begin{aligned}
\label{eq:BdG1}
\psi_j &=& \left( u_{j} e^{-i\D\omega_{j} t+i \D k_{j}x} +v_{j}^{*} e^{+i\D\omega_{j} t-i \D k_{j}x}\right)\nonumber\\
&& ~~~\times e^{-i\omega_{\rm bf} t+ik_{\rm bf}x},\end{aligned}$$ where $j=$p, HR, BCR. Each BdG mode is composed of two components, with opposite frequency and wavevector, $\pm(\D\omega_j,\D k_j)$, relative to those of the background flow, $(\omega_{\rm bf},k_{\rm bf})$. The BdG dispersion relation is given by $$\begin{aligned}
\label{eq:dispersion}
\Delta\omega &=&\sqrt{c^2\Delta k^2+ (\hbar \Delta k^2/2m)^2}+v_{\rm bf, o}\Delta k,\end{aligned}$$ where $v_{\rm bf,o}$ is the velocity of the condensate with respect to the “observer" frame in which the frequency is defined. The square root term gives the frequency in the comoving frame of the condensate, $\Delta\omega_{\rm cm}$. The first term in the square root describes long wavelength sound modes, while the second term corresponds to the kinetic energy of the atoms, and dominates at large wavevectors. The amplitudes of two components of $\psi_j$ are given by $$\begin{aligned}
\label{eq:uv}
(u_j,v_j)=\frac{1}{2\pi}\sqrt{\left|\frac{d\Delta k}{d\Delta\omega}\right|} \left(\frac{1}{\sqrt{1-D^2}},\frac{D}{\sqrt{1-D^2}} \right),\end{aligned}$$ where $D=[\hbar\Delta\omega_{\rm cm}- \hbar^2 \Delta k^2/2m - mc^2]/mc^2$ [@parentani2010]. Note that $D$ goes to zero for $\Delta k\gg mc/\hbar=1/\sqrt{2}\xi$, where $\xi$ is the healing length.
In the enhanced regime, the dispersion relations evaluated inside and outside the BH at $t=650$ ms are shown in Fig. \[fig:STFT\_wt\](a). The red (gray) and blue (dark gray) solid curves indicate the dispersion relation in the BH frame ($v_{\rm bf,o}=v_{\rm bf,BH}$), at $x_{\rm I}=x_{\rm BH}-$26 $\mu$m and $x_{\rm O}=x_{\rm BH}+$26 $\mu$m, respectively. These points correspond to the intersections of the diagonal red (left) and blue (right) lines with the horizontal green line in Fig. \[fig:M20\](b). The dashed red curve also indicates the dispersion relation at $x_{\rm I}$, but referred to the WH frame ($v_{\rm bf,o}=v_{\rm bf,WH}$). We use the numerically measured values of the local flow velocity and sound speed, determined from the background flow $\Psi_{\rm bf}$, which can be identified by a spatial WFT (despite the appearance of additional excitations). The WH velocity is approximated by the speed of the left edge of $|\delta\Psi(x,t)|$ \[see Fig. \[fig:M20\](b) and Appendix \[app:dispersion\]\], while the BH velocity is that of the step.
The BCR has zero frequency in the WH frame, so the BCR wavevector should satisfy $\Delta\omega(\Delta k_{\rm BCR})=0$ in that frame. This is indicated graphically by the intersection of the dashed red dispersion curve in Fig. \[fig:STFT\_wt\](a) with the $\Delta k$ axis, which yields $\Delta k_{\rm BCR}\sim 2.5\, \mu{\rm m}^{-1}$.
Due to the recession of WH relative to the BH, the frequency of BCR in the BH frame corresponds to $\Delta \omega=- \Delta k_{\rm BCR}\Delta v=- 0.2 \mbox{rad/ms}$, where $\Delta v$ is the BH velocity relative to the WH. This frequency is indicated by the lower dashed horizontal line, which intersects the solid red (gray) curve at the vertical line, $ \Delta k_{\rm BCR}$. (Note that similar reasoning can be applied for the upper dashed horizon line, which intersects the component, $v^*_{\rm BCR}$, at the opposite frequency and wavevector.) If the HR and partner modes are indeed stimulated by the BCR, they should share the same frequency with the BCR in the BH frame, so their wavevectors should lie at the intersections of the shifted BCR frequency (dashed black) lines with the solid blue (dark gray) and solid red (gray) dispersion curves, respectively.
As the BCR mode propagates toward the BH horizon, the dispersion curve lifts upwards due to the change of flow velocity and sound speed, and the wavevector “redshifts", until the mode coincides with the local minimum of the dispersion relation. At that stage the WKB description breaks down, and the mode converts to a superposition of other modes that share the same frequency. These are the Hawking radiation and partner modes. The modes are labeled by “$u$" or “$v^*$", according to the corresponding component of the BdG mode . Modes whose $u$-component has negative (positive) relative frequency in the step frame have negative (positive) energy relative to the condensate [@PhysRevA.80.043601]. The BCR and partner modes thus have negative energy, while the Hawking mode has positive energy.
Spectral comparison with BdG prediction {#sec:spectralHR}
---------------------------------------
To capture the spectral properties of the modes observed in Fig. \[fig:M20\](b), and compare with the prediction in Fig. \[fig:STFT\_wt\](a), we apply the spatial and temporal WFTs on $\Psi(x,t)$ and $\delta\Psi(x,t)$. On the left-hand side of the BH ($x<x_{\rm BH}$), we calculate the WFTs of $\Psi(x,t)$; on the right-hand side of the BH ($x>x_{\rm BH}$), we take $\delta\Psi(x,t)$ and multiply it by 10 to subtract the background and bring out the HR. The left panel of Fig. \[fig:STFT\_wt\](b) shows the windowed frequency spectra of $\Psi(x_{\rm I}(t),t)$ ($\omega=\text{0--0.5 rad/ms}$) and $\delta\Psi(x_{\rm O}(t),t)$ ($\omega=\text{0.5--0.7 rad/ms}$), in the BH frame, along the diagonal red (left) and diagonal blue (right) lines in Fig. \[fig:M20\](b) with Gaussian width $T=\text{55 ms}$. The streak in the center corresponds to the background flow $\Psi_{\rm bf}$, and indicates the frequency $\omega_{\rm bf}\sim \text{0.36 rad/ms}$. The two other streaks located symmetrically about the center correspond to HR ($\omega\sim \text{0.56 rad/ms}$), and the BCR and the p-mode ($\omega\sim \text{0.15 rad/ms}$). The full frequency spectra at $t=$ 650 ms for $x_{\rm I}$(diagonal red line, left) and $x_{\rm O}$(blue, right) are shown on the right panel.
The left panel of Fig. \[fig:STFT\_wt\](c) shows the windowed wavevector spectrum as a function of position, in the laboratory frame. It is defined by WFTs of $\Psi(x,t_0)$ ($x<x_{\rm BH}$) and $\delta\Psi(x,t_0)$ ($x>x_{\rm BH}$) at $t_0=650$ ms, along the horizontal green line in Fig. \[fig:M20\](b), with width $D=$ 21 $\mu\text{m}$ for $\Psi$, and 12 $\mu\text{m}$ for $\delta\Psi$. The background flow spectrum between the horizons is centered on a large negative wavevector at each $x$, and extends from the BH to the WH. As in the experimental regime \[Fig. \[fig:exp\](b)\], the BCR spectrum is roughly the reflection of the background flow, $\Psi_{\rm bf}$. This feature was explained qualitatively in Sec. \[sec:kspectrum\]. Here we can explain it quantitatively, using the dispersion relation . As can be seen in Fig. \[fig:STFT\_wt\](a), the point at $\Delta\omega=0$ in the WH frame is close to the single-particle regime (i.e., $\Delta\omega_{\rm cm}\sim \hbar \Delta k^2/2m$), so that $\Delta k_{\rm BCR}\sim -2mv_{\rm bf, WH}/\hbar$. The flow velocity $v_{\rm bf, WH}$ relative to the WH is approximately the same as the velocity in the laboratory frame, which is $\hbar k_{\rm bf}/m$. Therefore $\Delta k_{\rm BCR}\sim -2k_{\rm bf}$, hence the wavevector of $\psi_{\rm BCR}$ in Eq. \[eq:BdG1\] becomes $k_{\rm bf}+\Delta k_{\rm BCR}\sim-k_{\rm bf}$.
The HR and p-mode spectra extend outward and inward from the BH, with positive and negative wavevectors, respectively. The wavevector spectra at $x_{\rm I}$(diagonal red line, left) and $x_{\rm O}$(diagonal blue, right) are shown on the right panel, with the modes labeled (except for $v^{*}_{\rm BCR}$) in the figure.
We compare the WFT spectra \[Figs. \[fig:STFT\_wt\](b) and (c)\] with the BdG dispersion relations \[Fig. \[fig:STFT\_wt\](a)\] at $x_{\rm I}$ and $x_{\rm O}$, corresponding to the intersections of the diagonal red (left) and diagonal blue (right) lines with the horizontal green line in Fig. \[fig:M20\](b). The numerical values of $\Delta \omega$ and $\Delta k$ obtained from the WFT spectra of the GP solution are displayed in Table \[table:FT\], along with those predicted from the BdG dispersion relations. The inputs to the BdG prediction are just (i) the assumption of zero frequency in the WH frame, and (ii) the velocity of the BH frame relative to the WH frame. The GP spectra and BdG predictions agree to within 5%. Note that the flow is not perfectly stationary, so that the zero frequency of the initial BCR is not perfectly conserved. Also, the speed of the WH changes slightly over time, which gives rise to the uncertainty in $\Delta \omega_{\rm BdG}$ and $\Delta k_{\rm BdG}$.
Modes $\Delta\omega_{\rm FT}$ $\Delta\omega_{\rm BdG}$ $\Delta k_{\rm FT}$ $\Delta k_{\rm BdG}$
--------------- ------------------------- -------------------------- --------------------- ----------------------
$u_{\rm BCR}$ -0.21(3) -0.20$\pm0.01$ 2.42(8) 2.45$\pm0.02$
$u_{\rm p}$ -0.21(3) -0.20$\pm0.01$ 0.26(8) 0.26$\pm0.01$
$u_{\rm HR}$ 0.20(3) 0.20$\pm0.01$ 0.65(12) 0.65$\pm0.02$
: Numerical values of relative mode frequency $\Delta\omega$(rad/ms) and wavevector $\Delta k$($\mu$m$^{-1}$) from the GP Fourier spectra (FT) and from the WH-zero-frequency BdG dispersion relation (BdG). The uncertainty for the former is estimated by the widths of the Gaussians fitting the spectral peaks in Figs. \[fig:STFT\_wt\](b) and (c), and the uncertainty for the latter is due to the variation of the speed of WH.[]{data-label="table:FT"}
Hawking temperature {#sec:temp}
-------------------
The spontaneous emission from a black-hole horizon is thermal, with temperature $T_{\rm H}=\hbar\kappa/(2\pi k)$, where $\kappa$ is the surface gravity [@hawking1; @hawking2]. In the sonic analog, the surface gravity becomes $\kappa = d(v+c)/dx$, evaluated at the horizon [@unruh1981]. The Hawking mode of the stimulated radiation is excited with a coefficient $\beta$, and the partner mode with a coefficient $\alpha$, corresponding, in effect, to transmission and reflection coefficients. The ratio $|\beta/\alpha|$ carries the signature of the thermal prediction [@Unruh:1994je; @Unruh:2014hua], $$\label{ratio}
\frac{|\beta|}{|\alpha|}=
\frac{|V_{\rm HR}/v_{\rm HR}|}{|U_{\rm p}/u_{\rm p}|}
=\exp(-\pi\Delta\omega/\kappa).$$ Here $(U_{\rm p},V_{\rm HR})$ are the full mode amplitudes, which can be captured from the WFT spectra, and $(u_{\rm p},v_{\rm HR})$ are the normalized BdG amplitudes defined in Eq. \[eq:uv\].
To test the thermal prediction we evaluate the mode amplitudes at a pair of points $x_{\rm p}$ and $x_{\rm HR}$ with a common retarded time, defined by phase velocity, at the BH. These points are denoted by the magenta (left) and cyan (right) dots on the horizontal green line in Fig. \[fig:M20\](b). The common retarded time on the horizon is $t=588$ ms, for which we find the surface gravity $\kappa\sim350$ s$^{-1}$ (using $v$ and $c$ computed directly from the GP wavefunction, see Appendix \[app:temperature\]. The thermal prediction for $\omega=200$ rad/s is $|\beta/\alpha|=0.17^{+0.05}_{-0.04}$, allowing for a 5% uncertainty in $\omega$ and a 10% uncertainty in $\kappa$. This agrees reasonably well with the ratio 0.21 computed directly from the amplitudes according to the thermal prediction. The Hawking temperature for the case M2 depicted in Fig. \[fig:M20\](b) is $T_{\rm H}=0.43\,$ nK. The temperature equivalent of the chemical potential, $\mu$, in that case is $\mu/k=2.5\,$ nK.
Several factors could play a role in causing the GP ratio $|\beta/\alpha|$ to differ from the thermal prediction. First, the latter is exponentially sensitive to the value of $\kappa$, so the time dependence of $\kappa$ can introduce a significant effect. Second, phonon dispersion can produce deviations that depend on how large are $\kappa$ and $\omega$ compared to the sound speed over the healing length, $c/\xi\sim 320{\rm s}^{-1}$, and on how wide is the linear regime of the function $c+v$ around the BH ($\sim$ 6 $\mu{\rm m}$), compared to $c/\kappa\sim$ 1.2 $\mu{\rm m}$ [@two_regimes]. And third, nonlinearity of the GP modes could lead to deviations from the linear prediction.
Discussion
==========
Comments on the lasing mechanism
--------------------------------
The results in this paper, together with those of [@correlations] which includes fluctuations, establish that the BH laser phenomenon is not present, or at least not significant, in our simulations of the experiment of Ref. [@nphys3104]. Instead, we have traced the standing wave and its growth to the BCR mechanism. But the question remains as to why lasing does not occur, given that the system exhibits a flow structure of the type that can lead to the laser instability.
We can suggest two factors that may lie behind this: one is that the growth rate of the instability may be too slow to show any significant growth during the timespan of the step sweep. The other is that the time dependence of the flow structure and cavity size may lead to a detuning of the instability.
Regarding the time available for the laser instability, what matters is both the number of “cycles” that occurs, and the amplification factor in each cycle. We estimate using the dispersion relation a “round trip" time for modes propagating between the BH and WH of $\sim 50$ ms. This is consistent with Fig. \[fig:stimulated\_HR\](b), from which one can see that the p-mode that is created at the BH at $t\sim$ 60 ms \[Fig. \[fig:stimulated\_HR\](b)\] takes $\sim 30$ ms to reach the WH, and that the BCR takes $\sim 20$ ms to propagate from the WH to the BH. An initial partner mode excitation that reaches the left end of the cavity just after the WH has formed at $t\sim 20$ ms, i.e., at the earliest possible time, would thus have time for no more than two return trips to the BH before $t=120$ ms.
Regarding the time-dependent flow, the black hole laser scenario was originally introduced, and has been studied, in the setting of stationary flow structure. In the experiment, however, the background density changes significantly as the step moves from the edge of the condensate to the center, changing the sound speed. In addition, the size of the supersonic cavity is time dependent due to the receding WH. This causes a repetitive Doppler shift ($\propto \Delta k \Delta v$) on the modes inside the cavity. The Doppler shift is greater for modes possessing a larger wavevector, for which the frequency shift in one cycle is comparable to the initial frequency. This is explained in more detail in Appendix \[app:FT\_regime\_exp\], where the detuning effect is illustrated on the dispersion relation graph in Fig. \[fig:FT\_steinhauer\](c). The upshot is that the effect of moving WH horizon cannot simply be treated as an adiabatic evolution of the static case.
An unstable mode responsible for the laser effect contains right-moving components with large wavevector, arising from mode conversion of the left-moving p mode at the WH horizon, which suffer a significant Doppler shift relative to the frequency of the incoming p mode. When the cavity is small, the lasing can be dominated by a single, fastest growing unstable mode [@Coutant:2009cu]. In that setting, the above time-dependent effects might “detune" the laser, inhibiting the self-amplification mechanism.
Comparison with other simulated results {#sec:comparison}
---------------------------------------
A study having some overlap with ours was reported by Tettamanti [*et al.*]{} [@Tettamanti]. They established the hydrodynamic character of the experimental observations, and identified the Bogoliubov-Čerenkov (BCR) mode as responsible for initiating the instability, both of which are consistent with our findings. Our accounts differ, however, regarding the subsequent evolution. They report that the resulting Hawking radiation is self-amplifying, and that the growing wave pattern between the horizons results from this amplification and the interference between counter-propagating waves. However, it is hard to glean from the paper on what basis that conclusion was drawn. The local wavevector spectrum in the standing wave region is shown in \[Fig. \[fig:exp\](b)\], and discussed in Sec. \[sec:mechanism\]. We find that the standing wave pattern is simply the result of interference between the BCR mode and the background flow, and grows due to the growing condensate density. An additional, long wavelength, left-propagating partner mode of Hawking radiation is evident in the spacetime portrait \[Fig. \[fig:stimulated\_HR\](b)\]. However, this appears well after the standing wave has formed, and has much smaller amplitude.
Another numerical investigation of this system, by Steinhauer and de Nova [@SR], appeared while we were preparing revisions of our manuscript. They report findings indicating that the BCR component to the standing wave can not be generated at the WH horizon because it appears before the WH horizon forms. We suspect that this discrepancy with our findings may be traced to their use of a definition of the WH horizon that does not coincide with the location of the stationary, super-to-subsonic transition, as explained more fully in Sec. \[sec:Locating\].
In addition, Ref. [@SR] argues that self-amplifying Hawking radiation can be distinguished from what they call the “background ripple" by the presence of time dependence, since the power in a monochromatic traveling wave such as the BCR would not oscillate in time. To exhibit the time dependence, they evaluate a temporal Fourier transform of the spatial Fourier transform of the density, normalized by the square of the spatial average of the density in the cavity. This has a peak at a characteristic frequency which, they assert, is a signature of the self-amplifying Hawking radiation.
We also find time dependence associated with the standing wave in the cavity, which can be seen in the oscillatory features in Fig. \[fig:GP\_growth\](a). These oscillations result from interference with the Hawking partner mode, as can be seen in Fig. \[fig:stimulated\_HR\](b). This might be the source of the time dependence found in Ref. [@SR]. The dominant period for this time dependence, which can be read from Fig. \[fig:GP\_growth\](a), is of order 40 ms. This corresponds to an angular frequency of order 0.16 rad/ms, i.e., $\omega/\omega_{\rm max}\sim 0.2$ (where $\omega_{\rm max}\sim 0.75$ rad/ms is the maximum allowed frequency in the cavity region), which is not far from the values found in Ref. [@SR].
Conclusion {#sec:conclusion}
==========
To conclude, we summarize the evidence that the black hole laser effect plays no role in our GP simulations. First, we find only one unstable mode, the BCR mode, which grows in proportion to the square of the background flow density. Lasing action, by contrast, would have no reason to satisfy this relation. Second, this growing mode has zero frequency in the WH frame, which differs from the BH frame due to a nonzero relative motion of the two horizons. And, third, the windowed Fourier spectrum in the cavity reveals no other frequency components. The Hawking partner radiation has the same frequency as the BCR mode, as would be expected if it results from stimulation by the monochromatic BCR.
Our simulations did not include quantum fluctuations capable of spontaneously producing Hawking radiation. However, in a related paper [@correlations] we have studied the effect of quantum fluctuations using the truncated Wigner approximation, and we found no evidence there of any mode growth beyond that found here. Moreover, the results of Ref. [@correlations] reveal that the key features of the observed density correlation function can all be produced by modulation of the BCR standing wave caused by atom number variations and quantum fluctuations.
Finally, we investigated various regimes of potential experimental parameters, and found a regime where a sharper signal of HR is obtained, and in which a BdG mode description is valid. This enabled us to carry out a detailed quantitative check of our proposed mechanism, stimulation of the Hawking radiation by a Doppler shifted, zero frequency BCR standing wave. This enhanced parameter regime could provide a useful guide for future experimental investigations of stimulated Hawking radiation in this setting.
Acknowledgements
================
We thank R. Parentani for numerous helpful discussions and suggestions, and J. Steinhauer for stimulating correspondence and criticism. This material is based upon work supported by the U.S. National Science Foundation Physics Frontier Center at JQI and grants PHY–1407744, PHY–1004975 and PHY–0758111, and by the Army Research Office Atomtronics MURI.
Characterization of the experimental condensates and description of simulation procedures {#app:sim_details}
=========================================================================================
The condensate the in the experiment of Ref. [@nphys3104] is tightly confined in two transverse dimensions, and elongated in the third dimension, with a scale ratio $\sim 1:20$. It is thus approximately one-dimensional, and for all the simulations in this paper we have employed the one-dimensional description, except for some comparative simulations described in this appendix. Here we begin by describing a three-dimensional model of the system, and then proceed to explain the reduction to an effective, one-dimensional model. We also compare this 1D model to a 3D one, and find that the 1D model accurately captures the important features of the 3D dynamics.
The Gross–Pitaevskii (GP) equation, giving the mean field description of a BEC in three dimensions, takes the form $$i\hbar\frac{\partial\Psi({\bf r},t)}{\partial t} =
\left(-\frac{\hbar^{2}}{2m}\nabla^{2} + V({\bf r},t) + g_{\rm 3D}N\left|\Psi\right|^{2}\right)
\Psi({\bf r},t),
\label{tdgp3d}$$ where $N$ is the number of condensate atoms, $m$ is the mass of a condensate atom, $g_{\rm 3D}=4\pi\hbar^{2}a/m$ where $a$ is the $s$–wave scattering length, and $V({\bf r},t)$ is the full external potential. In the case of the experiment, the potential is given by $$V({\bf r},t) = U({\bf r}) + U_{\rm step}({\bf r},t).$$ The potential in which the initial condensate was formed in the experiment of Ref. [@nphys3104] is denoted by $U({\bf r})$, and $U_{\rm step}({\bf r},t)$ is the potential for the step that was swept along the length of the condensate, as shown schematically in Fig. \[fig:exp\] of our paper. The wave function for the initial condensate, $\Psi_{0}({\bf r})$, satisfies the time–independent GP equation: $$\left(-\frac{\hbar^{2}}{2m}\nabla^{2} + U({\bf r}) + g_{\rm 3D}N\left|\Psi_{0}({\bf r})\right|^{2}\right)
\Psi_{0}({\bf r}) = \mu_{0}\Psi_{0}({\bf r}),
\label{tigp3d}$$ where $\mu_{0}$ is the chemical potential of the ground-state condensate.
We begin by describing the most accurate 3D GP model, given the information in Ref. [@nphys3104] about the experiment. We first found the potential, $U({\bf r})$, that is produced by the red–detuned trapping laser specified in Ref. [@nphys3104]. There, the laser beam characteristics are stated in terms of its wavelength, $\lambda = 812 $ nm, and the beam waist, $w_{0} = 5 $ mm. We used these data to model the trapping laser light as a focused ideal Gaussian laser beam. Thus the trapping potential is proportional to the beam intensity: $$U({\bf r}) = U_{0}\left[1 -
\left(\frac{w_0}{w\left(x\right)}\right)^2
\exp\left(\frac{-2\rho^2}{w^2\left(x\right)}\right)\right].
\label{Intensity}$$ where $x$ is the axis of light propagation, $\rho = \sqrt{y^2 + z^2}$ is the transverse (axial) radial coordinate, $w_0$ is the beam waist, $U_{0}$ is proportional to the peak laser intensity and $$w(x) = w_0 \sqrt{1 + \left(\frac{x}{x_0}\right)^2} \, ,
\quad
{\rm where}
\quad
x_0 = \frac{\pi w_0^2}{\lambda}.
\label{wofz}$$ We have chosen the origin of energy so that $U({\bf r})$ vanishes at the center of the trap, $U(0) = 0$. Reference [@nphys3104] also states that, since the long axis of the needle–shaped condensate lies in a horizontal plane, the effect of gravity is mostly (all but 9%) compensated for by an external magnetic field with a vertical gradient. For simplicity, in our model we take the gravitational and compensating magnetic forces to cancel exactly.
Specification of the wavelength and beam waist fixes all the parameters in $U({\bf r})$ except for $U_{0}$. Reference [@nphys3104] gives the axial trap frequency as $\omega_{\rho}/2\pi = 123$ Hz. We used this frequency to determine $U_{0}$ by expanding $U({\bf r})$ to second order about ${\bf r}= 0 $ : $$\label{U(r)}
U({\bf r}) \approx
\left(\frac{2U_{0}}{w_{0}^{2}}\right)\rho^{2} +
\left(\frac{U_{0}}{x_0^{2}}\right)x^{2} \equiv
\frac{1}{2}m\omega_{\rho}^{2}\rho^{2} +
\frac{1}{2}m\omega_{x}^{2}x^{2}.$$ Thus, $U_{0} = (1/4)m\omega_{\rho}^{2}w_{0}^{2}\approx 39 \, k \, \mathrm{nK}$, where $k$ is the Boltzmann constant. This completes our determination of $U({\bf r})$ from the experimental parameters. The full list of experimental parameters is given in Table \[Parameters\].
To determine the number of atoms in the condensate (which was not explicitly stated in Ref. [@nphys3104]) we simulate the initial condensate for different atom numbers $N$, as shown in Fig. \[initial\_condensates\]. We find that $N = 6000$ gives a best match of the axial length to that determined experimentally. Note that the chemical potential for $N = 6000$ is $\mu_{0}/k =10.8\,$ nK, which includes the radial kinetic and potential energies, and is greater than the reported experimental value, $\mu_{0}/k =8\,$ nK. It is not reported in [@nphys3104] how this value was determined; however, for other quasi-low-dimensional BECs, a chemical potential usually refers to the maximal interaction energy determined by the maximal integrated density in the loosely-confined direction [@1DBEC2001; @pethick]. In the simulated condensate with $N = 6000$, the maximal interaction energy in the axial direction is about 7.3 nK, which is comparable to the reported experimental value, 8 nK.
![Optical densities of the condensate ground states from 3D GP simulations for atom numbers $2000 < N < 6000$. Each plot is also labeled with its associated chemical potential. Full horizontal and vertical scales are 100 and 12 micrometers, respectively, and the color box scale denoting optical density is graduated in arbitrary units.[]{data-label="initial_condensates"}](steinhauer_becs_multiplot_5x1){width="3.3in"}
{width="7in"}
{width="7in"}
Parameter Value Units
------------------------------------------------ --------------- --------
atom $^{87}$Rb
atomic state $F=2, M_F =2$
trapping laser wavelength $\lambda$ 812 nm
beam waist $w_0$ 5 micron
radial trap frequency $\nu$ 123 Hz
transverse energy level spacing $E$ 6 nK $k$
healing length $\xi$ 2 micron
nominal chemical potential $\mu$ 8 nK $k$
\*actual chemical potential $\mu$ 10.4 nK $k$
\*axial length scale $x_0$ from eq. (\[wofz\]) 97 micron
\*number of condensate atoms $N$ 6000 atoms
: Parameters of the trapped BEC as reported in, or inferred from (\*), Ref. [@nphys3104]. Uncertainties are not stated in Ref. [@nphys3104], and we do not attempt to estimate them in this work.[]{data-label="Parameters"}
Reduction to a one-dimensional system {#app:1dGPE}
-------------------------------------
When a condensate is tightly confined in the radial direction, and the integrated density $n$ in the axial direction satisfies $na\ll1$, it can be viewed as quasi-one-dimensional [@1DGPE; @stringari]. For the experimental configuration, the axial to radial frequency ratio is given by $\omega_x/\omega_\rho= w_0 / (\sqrt{2}x_0) =4.5 \, \mathrm{Hz}/123 \, \mathrm{Hz}$, which is much smaller than unity. This shows that the system is tightly-confined in the radial direction. Using the maximum of the integrated density from the 3D simulation, $n_{\rm max}\sim120$ $\mu{\rm m}^{-1}$, we estimate that the integrated 1D density satisfies $an_{\rm max}\approx 0.65$, which is less than, but not much smaller than, unity. It is therefore not guaranteed that the 1D system is truly quasi-one dimensional. However, as discussed in Sec. \[app:3dGPE\], we compared the 3D and 1D simulations and found the essential features to be quite similar.
To implement the simplest sort of 1D reduction, we approximate the wavefunction in the radial direction by the solution of a harmonic oscillator, so that $$\Psi(\mathbf{r},t) = \frac{\exp\left[-\rho^2 /\left(2 d^2\right)\right]}{d\sqrt{\pi}}\Psi_{\rm 1D}(x,t),$$ where $d = \sqrt{\hbar/\left(m \omega_{\rho}\right)}$. Integrating the 3D GP equation over the Cartesian coordinates $y$ and $z$, we obtain a 1D GP equation with an effective interaction coefficient $g_{\rm 1D}=g_{\rm 3D}m \omega_{\rho}/h$: $$\begin{aligned}
i\hbar\frac{\partial\Psi_{\rm 1D}(x,t)}{\partial t} &=&
\left(-\frac{\hbar^{2}}{2m}\frac{\partial^2 }{\partial x^2} + V(x,t)\right)\Psi_{\rm 1D}(x,t)\nonumber\\
&+& g_{\rm 1D}N\left|\Psi_{\rm 1D}(x,t)\right|^{2} \Psi_{\rm 1D}(x,t),
\label{tdgp1d}\end{aligned}$$ where $h$ is the Planck constant and $V(x,t)$ is the full external potential $V(\bf{r},t)$ evaluated at $y=z=0$. We define the coefficient of the nonlinear term as, $g=g_{\rm 1D}N$, which is used extensively in the paper.
We take the number of atoms determined by the 3D GP equation, $N=6000$, and use the 1DGPE \[Eq. \[tdgp1d\]\] to simulate the step-sweeping experiment. Note that Ref. [@Tettamanti] simulated the dynamics by using a 1D nonpolynomial nonlinear Schrodinger equation (NPSE) [@1DNPSE], which incorporates the effect of a variable axial density on the transverse shape of the GP wave function (under the condition that the axial derivative of the transverse wave function is much smaller than the transverse derivative). Here we use a simpler 1D GP equation, which assumes a fixed transverse shape of the wave function. In the paper, and in other sections of the Appendix, we drop the subscript “1D” in $\Psi_{\rm 1D}(x,t)$ when referring to the 1D GP wavefunction.
Solution of the time–dependent 1D GP equation {#app:1dTDGPE}
---------------------------------------------
The time-dependent 1D GP equation is solved by using the split-step Crank-Nicholson algorithm [@adhikari] on a 1D spatial grid of 320 $\mu$m with 4800 points, first propagating in imaginary time to obtain the initial stationary condensate, then propagating in real time with the given initial state to simulate the dynamics. To simulate the step-sweeping experiment [@nphys3104], we use a step potential $U_{\rm step}(x,t)$, which takes the form $$\label{eq:step}
U_{\rm step}(x,t)=-U_{\rm s}\Theta(x_{\rm s}(t)-x),$$ where $\Theta$ is the Heaviside step function, $U_{\rm s}$ is the step strength, which takes the values of $U_{\rm s}/k=$ 3 nK and 6 nK, and $x_{\rm s}(t)$ represents the step location, moving at a constant speed, $v_s=$ 0.21 mm/s.
Growing standing wave, spacetime portrait, and frequency spectrum from a 3D simulation {#app:3dGPE}
--------------------------------------------------------------------------------------
To test the accuracy of the 1D simulation, we also simulated the step-sweeping experiment using the 3D GP equation, with the potential , assuming the condensate shares the axial symmetry of the potential. Figures \[fig:GP2\](a)-(g) show the integrated density profiles with a potential step, $U_{\rm s}/k=$ 5 nK, which is adjusted slightly to match the cavity size with the experiment. The growth of the standing-wave amplitude, $\bar{n}_k$, and that of the background density, $\bar{n}_{\rm bf}$, are shown in Fig. \[fig:GP2\](h). The standing wave grows by $\sim\exp(4.8)$, which is greater than in the 1D simulation ($\sim\exp(4.4)$), but the growth relation, $\bar{n}_{k}\propto\bar{n}^2_{\rm bf}$, is preserved in the 3D simulation.
Similarly, we calculate the spacetime portrait and the local frequency spectrum using the GP wavefunction at the center of the radial trap, $\rho=0$. The spacetime portrait in Fig. \[fig:FT\_3d\](a) shows very similar features as those in the 1D simulation, including the standing wave parallel to the WH, and the stimulated Hawking pair. The WH recession can also be seen in the portrait, which gives rise to a Doppler-shifted BCR frequency in the WFT spectrum in Fig. \[fig:FT\_3d\](b) and (c), $\Delta\omega\sim 0.23$ rad/ms.
Although there are some quantitative differences with the 1D simulation, all the qualitative features found in the 1D GPE are preserved here: (i) the growth relation between the standing wave and the background density, (ii) the stimulated HR pair by the BCR, and (iii) the Doppler shift due to the WH recession.
Windowed Fourier transform {#app:FT}
==========================
Here we summarize the basic ideas of the windowed Fourier transform (WFT), and explain our use of it. In \[app:FT\_definition\], we give the definition of WFT used here, and provide a few basic examples to show how it can resolve spectral information on non-stationary phenomena. In \[app:FT\_vc\], we describe the application of the WFT to the determination of flow and sound speeds, $v(x)$ and $c(x)$, in inhomogeneous media. In \[app:FT\_spectral\_analysis\], we discuss calculations of the wavevector and frequency spectra displayed in Figs. \[fig:STFT\_wt\](b) and (c), and the additional spectra that distinguish the partner and BCR modes. In \[app:FT\_regime\_exp\], we show the windowed frequency spectrum for the experimental regime, and a comparison with the dispersion relation.
Definition and examples {#app:FT_definition}
-----------------------
A windowed Fourier transform [@gomes99fourier] $f(k,x)$ of a function $f(x)$ is defined as: $$\begin{aligned}
\label{eq:FTkx1}
f(k,x)= \int^{\infty}_{-\infty} dy\, f(y)w(y-x;D)e^{-iky},\end{aligned}$$ where $w(y-x;D)=\exp (-(y-x)^2/D^2)/\left(\sqrt{\pi}D\right)$ is a Gaussian window function of width $D$. With the filtering of the window, the transformed function $f(k,x)$ constitutes a local Fourier transform of $f(x)$, capturing features that vary on length scales much smaller than $D$. For a plane wave with wavevector $q$ and amplitude $f_{q}$ , $f(x)=f_{q} \exp (iqx)$, the transformed function is $f(k,x)=f_{q} \exp (-(k-q)^2(D/2)^2)$ : a Gaussian in $k$-space, centered at $k=q$ with width $2/D$ and peak amplitude $f_{q}$.
Suppose now that $f(x) = f_{q}(x) \exp(iqx)$, where $f_q(x)$ has weak dependence on $x$, and can be adequately approximated near a point $x_0$ by $$\begin{aligned}
\label{eq:FTkx2}
f_{q}(x)= f_q(x_0)+f_q^{\prime}(x_0)(x-x_0).\end{aligned}$$ Then for sufficiently small values of $D$, the WFT of $f(x)$ near $x=x_0$ is approximately $$\begin{aligned}
\label{eq:FTkx3}
f(k,x_0)&\approx & f_q(x_0)e^{-(k-q)^2(D/2)^2} \nonumber\\
&+&f_q^{\prime}(x_0) \,i\frac{k-q}{2}D^2\, e^{-(k-q)^2(D/2)^2}.\end{aligned}$$ Note that the second term vanishes at the peak position $k=q$, so that $f(q,x_0)\approx f_q(x_0)$.
Finally, let $f(x)$ be composed of a number of such slowly–varying modes, $$\begin{aligned}
\label{eq:FTkx4}
f(x)= \sum\limits_{n} f_{q_n}(x)e^{iq_{n}x}, \end{aligned}$$ so that Eq. \[eq:FTkx3\] becomes $$\begin{aligned}
\label{eq:FTkx5}
f(k,x_0)& \approx & \sum\limits_{n} \left[f_{q_n}(x_0)+f_{q_n}^{\prime}(x_0)\left(i\frac{k-q_n}{2}D^2\right)\right]\nonumber\\
&&~~~\times e^{-(k-q_n)^2(D/2)^2}.\end{aligned}$$ In $k$–space, each mode presents a Gaussian distribution centered on its respective $q_n$, whose peak value of $f(q_n,x_0)$ defines the local mode amplitude. This is how we make quantitative determinations of the mode amplitudes that are discussed in our paper.
Determination of the profiles of flow speed and the speed of sound {#app:FT_vc}
------------------------------------------------------------------
{width="7in"}
As shown in the paper, during the sweep of the step, the time-dependent GP wavefunction $\Psi(x,t)$ exhibits excitation modes on top of the background condensate. To calculate the speed of sound $c(x)$ and flow speed $v(x)$ associated with the background condensate, we extract the condensate from the full GP wavefunction with the help of a WFT.
First, the amplitude of the background flow at a given time $t_0$ can be calculated by applying a spatial WFT on the GP density $|\Psi(x,t_0)|^2 =n(x)$ , where $t_0$ is suppressed for brevity. $$\begin{aligned}
n(k,x)= \int^{\infty}_{-\infty}dy \, n(y) w(y-x;D) e^{-iky}.\end{aligned}$$ Figure \[fig:vc\_supplement\](a) shows the result of a spatial WFT of density $n(x)$ (which corresponds to Fig. \[fig:GP1\](e) in the paper) with width $D=5$ $\mu$m. The central streak at $k\sim 0$ corresponds to the background flow, whose peak value gives rise to the background density $n_{\rm bf}(x)=|n(k\sim0,x)|$, as shown in Fig. \[fig:vc\_supplement\](b). The local speed of sound can then be expressed as by $c(x)=\sqrt{gn_{\rm bf}(x)/m}$. In Fig. \[fig:vc\_supplement\](b), we see that WFT method is appropriate in the slowly varying regions away from the two horizons: in the exterior region, $n_{\rm bf}(x)$ matches the GP density $n(x)$; in the interior region, $n_{\rm bf}(x)$ is at about the average value of the density oscillations. Near the event horizons, on the other hand, the background density changes rather quickly, so that WFT introduces an unwanted averaging. In these regions, it is more appropriate to use the local GP wavefunction directly to define $v(x)$ and $c(x)$, since there are no significant excitations on the background flow, and the definition is strictly local.
Second, the flow velocity can be calculated by a WFT of the GP wavefunction $$\begin{aligned}
\label{eq:FTkx_n}
\Psi(k,x)= \int^{\infty}_{-\infty} dy\, \Psi(y)w(y-x;D)e^{-iky},\end{aligned}$$ where again $t_0$ is suppressed for brevity. Figure \[fig:vc\_supplement\](c) shows the windowed wavevector spectrum $|\Psi(k,x)|^2$ with width $D=5$ $\mu$m. The dominant streak is the background flow, whose peak location $k_{\rm bf}(x)$ gives rise to the flow velocity in the rest frame of the step, $-v(x)=\hbar k_{\rm bf}(x)/m-v_s$, as shown in Fig. \[fig:vc\_supplement\](d); the peak value of the streak also corresponds to the background density $n_{\rm bf}(x)=|\Psi(k_{\rm bf},x)|^2$. In addition, we calculate the velocity profile by using the full GP wavefunction, $v_{\rm GP}(x)=\hbar/(m n(x)) \text{ Im} \left[ \Psi^{*}(x)d\Psi(x)/dx\right]-v_s$. We can see that WFT works well in regions apart from the event horizons, and effectively projects out the spatial oscillation present in $v_{\rm GP}(x)$.
Figures \[fig:vc\_supplement\](e) and (f) compare $v(x)$ and $c(x)$ from the windowed spectra \[Fig. \[fig:vc\_supplement\](e)\] with those obtained from the full GP wavefuntion \[Fig. \[fig:vc\_supplement\](f)\]. In short, near the event horizon, the approach of directly adopting the GP wavefunction gives more accurate speed profiles, with the correct horizon locations and the respective surface gravity; yet away from the horizons, the WFT effectively removes excitations from the background flow, and hence gives a more suitable definition for $v(x)$ and $c(x)$.
Spectral analysis with windowed Fourier transform {#app:FT_spectral_analysis}
-------------------------------------------------
The spectral properties of excitation modes can be obtained by performing spatial and temporal WFTs on the condensate wavefunction. Given a GP wavefunction, $\Psi(x,t)$, we calculate its local wavevector spectrum and frequency spectrum by applying the WFTs. To obtain a local wavevector spectrum, we perform a spatial WFT on the wavefunction $\Psi(x,t_0)$ using Eq. \[eq:FTkx\_n\] at a time $t_0$ in which excitation modes are present. The result is presented in Fig. \[fig:STFT\_wt\](c) in the paper, in which the excitation modes are resolved in addition to the background flow. Note that for the region on the right-hand side of the step, we perform the WFTs on the variation function $\delta\Psi(x,t)$ rather than $\Psi(x,t)$ (see Appendix \[app:separation\]), in order to subtract the background component and bring out the excitation mode in that region.
For a local frequency spectrum, we apply a temporal WFT at position $x_0(t)$ moving at constant speed $v_{\rm s}$ with the potential step: $$\begin{aligned}
\label{eq:FTwt_n}
\Psi(\omega,t)= \int^{\infty}_{-\infty} d\tau\, \Psi(x_0(\tau),\tau)w(\tau-t;T)e^{i\omega \tau},\end{aligned}$$ where $w(\tau-t;T)$ represents a Gaussian window function of width $T$, $w(\tau-t;T)=e^{-(\tau-t)^2/T^2}/\sqrt{\pi}T$; $x_0$ is selected to be both inside ($x_{\rm I}$) and outside the BH cavity ($x_{\rm O}$), which is indicated by the red (left diagonal) and blue (right diagonal) lines in Fig. \[fig:M20\](b). The result is presented in Fig. \[fig:STFT\_wt\](b). In the figure, there are two modes ($\psi_{\rm p}$ and $\psi_{\rm BCR}$) overlapped in the frequency spectrum ($\omega\sim0.15$ $\mu$m) evaluated at position $x_{\rm I}(t)$. To resolve the two modes, we perform a spatial WFT evaluated at $x_{\rm I}(t)$ for various times $$\begin{aligned}
\label{eq:FTkt_n}
\Psi(k,t)= \int^{\infty}_{-\infty} dy\, \Psi(y,t)w(y-x_{\rm I}(t);D)e^{-iky}.\end{aligned}$$ The result is presented in Fig. \[fig:bcr\_p\_spectra\](a), from which $\psi_{\rm BCR}$ and $\psi_{\rm p}$ are separated at different $k$ values, $k_{\rm BCR}$ (solid red line) and $k_{\rm p}$ (dashed red line). Furthermore, by performing a temporal WFT on $\Psi(k,t)$ at the two wavevectors, we resolve the overlapped streaks in the initial frequency spectrum at $\omega\sim0.15$ rad/ms, as shown in Figs. \[fig:bcr\_p\_spectra\](b) and (c).
{width="7in"}
Windowed frequency spectrum and dispersion relation for the experimental regime {#app:FT_regime_exp}
-------------------------------------------------------------------------------
Here we present the frequency spectrum for the experimental regime, in comparison with the prediction from the BdG dispersion relation as in Sec. \[sec:BdG\]. This shows that no black hole laser effect is apparent in our simulation of the experiment of Ref. [@nphys3104].
We apply the temporal WFT on $\Psi(x,t)$ at a position about the center of the cavity, $x_{\rm I}=x_{BH}-12$ $\mu$m, indicated by the diagonal red (left) line in Fig. \[fig:stimulated\_HR\](b). The resulting frequency spectrum is given in Figs. \[fig:FT\_steinhauer\](a) and (b). The streak that appears from early times shows the frequency of the background flow wavefunction $\Psi_{\rm bf}$. The lower streak corresponds to the superposition of the BCR and the partner mode. The cut-through at $t=$ 100 ms is shown in Fig. \[fig:FT\_steinhauer\](b), from which can be seen the relative frequency (of the $u$-components), $\Delta\omega\sim $ -0.11(3) rad/ms.
We also predict this relative frequency using the dispersion relation, as in Fig. \[fig:STFT\_wt\](a) for the enhanced regime. The assumption that the BCR is the zero-frequency mode in the WH frame determines $\Delta k_{\rm BCR}=2.9$ $\mu$m$^{-1}$. Taking into account the velocity difference between the WH and the BH, $\Delta v\sim 0.03$ mm/s, the relative frequency of BCR (and p-mode) in the BH frame is given by $\Delta\omega=-\Delta k_{\rm BCR} \Delta v\sim - 0.09$ rad/ms (for the components $u_{\rm BCR}$ and $u_{\rm p}$), which is indicated by the lower dashed black line in Fig. \[fig:FT\_steinhauer\](c). This predicted frequency is within the uncertainty of the measured WFT value.
{width="7in"}
The wavelength of the partner mode predicted using the dispersion relation is $\lambda_{\rm p}\sim 57$ $\mu$m, which is greater than the width of the supersonic cavity $L\sim 25$ $\mu$m. Therefore the partner cannot be treated in the WKB approximation, and the discrete spectrum of cavity modes modifies the emission, unlike in the M2 regime where the ratio $\lambda_{\rm p}/L$ is smaller. This may explain the irregular wavelength of the HR in the experimental regime seen in Fig. \[fig:stimulated\_HR\](b).
When the p-mode scatters at the WH, it creates a pair of positive-norm ($\psi_{+}$) and negative-norm ($\psi_{-}$) modes [@nphys3104] \[here we only show the former, $u_{+}$ in Fig. \[fig:stimulated\_HR\](c)\], whose frequency is the same as that of the partner in the WH frame. Due to the relative velocity between BH and WH, $u_{+}$ has a shifted frequency in the BH frame \[the horizontal blue (bottom) line in Fig. \[fig:FT\_steinhauer\](c)\], lower than the frequency of the first p-mode, and it stimulates the second p-mode ($u_{p}^{(2)}$) at that shifted frequency. The repetitive scatterings at the horizons therefore do not occur at a single frequency, as they would in the static case, i.e.,with zero WH horizon velocity. For the BCR, it can be seen in Fig. \[fig:FT\_steinhauer\](c) that the frequency shift per cycle is comparable to the frequency itself, $\Delta \omega/\omega \sim 1$. The motion of the WH is therefore not well within the adiabatic regime. Hence, for large wavevectors, like those of the BCR, the static analysis of the black-hole lasing phenomenon is not reliable for predicting what happens with the moving WH horizon.
Separation of fast and slow oscillation of condensate wavefunction {#app:separation}
==================================================================
To separate the HR from the subsonic background flow, we apply a smoothing procedure to separate fast oscillatory modes from the slowly-varying components in the GP wavefunction. The procedure is equivalent to calculating the moving average of a discrete data set, which smooths out short-term fluctuations. Here, the moving average of wavefunction $\Psi(x)$ is defined as $$\begin{aligned}
\label{eq:smoothing}
\bar{\Psi}(x)=\frac{1}{2D} \int^{x+D}_{x-D} dy\, \Psi(y),\end{aligned}$$ where the integral serves as a square window of width $2D$ centered at $x$, over which $\Psi(x)$ is being averaged. For components in $\Psi(x)$ with wavelength much shorter than $D$ (i.e., $D \gg\pi/k$), the integral would give rise to an average of zero, leaving those that are slowly varying in space (i.e., $\pi/k \gg D$) in $\bar{\Psi}(x)$.
{width="7in"}
{width="7in"}
According to Fig. \[fig:vc\_supplement\](c), the background flow in the subsonic region ($\Psi_{\rm bf}^{\rm sub}$) has $k\sim0$ and can be separated from the GP wavefunction through Eq. \[eq:smoothing\], such that $\bar{\Psi}(x) \approx \Psi_{\rm bf}^{\rm sub}(x)$. Furthermore, highly oscillatory components in the wavefunction, including all the excitation modes ($\psi_j$) and the supersonic background flow ($\Psi_{\rm bf}^{\rm sup}$), can be obtained by subtracting the GP wavefunction with the non-oscillatory component, $\delta\Psi(x) = \Psi(x)-\bar{\Psi}(x)$. Thus, the variation $\delta\Psi(x)$ can be expressed as $$\begin{aligned}
\label{eq:delta_psi}
\delta\Psi \approx \psi_{\rm p}+\psi_{\rm HR}+\psi_{\rm BCR}+\Psi_{\rm bf}^{\rm sup}.\end{aligned}$$ Figs. \[fig:psi0\_supplement\](a) and (c) show the application of Eq. \[eq:smoothing\] to obtain a smoothed profile of $|\bar{\Psi}(x)|$ in the experiment of Ref. [@nphys3104], and for a simulation in the M2 enhanced regime. A separate calculation using the spatial WFT is shown in Figs. \[fig:psi0\_supplement\](b) and (d), in which $|\Psi_{\rm bf}^{\rm sub}|$ is evaluated by taking the peak amplitude $|\Psi(k,x)|$ at $k\sim0$. Both approaches agree with the GP wavefunction at regions away from the event horizons, capturing the background component outside the supersonic cavity. This gives rise the variation function $\delta\Psi(x)$, which nicely approximates the components in Eq. \[eq:delta\_psi\]. Figure \[fig:delta\_psi\_supplement\] shows the variation $\delta\Psi(x)$ for the M2 regime, which agrees with the GP wavefunction inside the BH cavity. Note that in Fig. \[fig:delta\_psi\_supplement\](b) we have multiplied $\delta\Psi(x)$ by a factor of 10 for $x>x_{\rm BH}$ to bring out the HR mode in the exterior region.
BEC parameter regimes in which Hawking radiation has greater visibility {#app:parameter}
=======================================================================
To find a more distinctive signature of HR, we study the GP evolution in different parameter regimes where the frequency of the trapping potential, $\omega_{x}$, and the depth, $U_{\rm s}$, and speed, $v_{\rm s}$, of the potential step are varied away from the values ($\omega^{0}_x$, $U_{\rm s}^{0}$, $v_{\rm s}^{0}$) reported in Ref. [@nphys3104], which are given in Appendix \[app:sim\_details\]. We find that by choosing an appropriate set of experimental parameters, the HR can be observed with well-resolved wavelengths and frequencies.
Figure \[fig:evolution\_modified\] shows four representative cases for our investigation: E1, E2, M1, and M2. Regimes E1 and E2 use the same trapping frequency as the experimental value $\omega^{0}_x$, but adopt a greater step speed $v_s=$ 1.5$v_s^0$; case E1 uses the same step strength as Fig. \[fig:stimulated\_HR\](b) in the paper, $U_{\rm s}/k=6$ nK; case E2 has a greater step strength, $U_{\rm s}/k=9$ nK. Note that $\omega^0_x$ and $v_{s}^0$ are the reference values taken from [@nphys3104], $\omega^0_x= (2\pi) \times 4.5$ Hz, and $v_{\rm s}^0=$ 0.21 mm/s.
{width="6.7in"}
Regimes M1 and M2 are the cases equivalent to E1 and E2 with a modified trapping frequency, $\omega_x=(1/4)\omega^0_x$. Here, we use some scaling relations to determine the step speed $v_{\rm s}$ and depth $U_{\rm s}$ that give rise to an equivalent flow structure with the modified trapping frequency. We know that modifying $\omega_x$ changes the speed of sound $c$ \[due to the change of $n(x)$\] and the chemical potential $\mu$, and subsequently changes the flow structure shown in Fig. \[fig:vc\_supplement\]. Using the Thomas-Fermi approximation [@pethick] for a 1D condensate in a harmonic trap, we find that $\mu \propto \omega_x^{2/3}$, and the maximal density $n_{\rm max}\propto \omega_x^{2/3}$ (i.e., $c_{\rm max}\propto \omega_x^{1/3}$). By keeping ratios $U_{\rm s}/\mu$ and $v_{\rm s}/c_{\rm max}$ fixed, we can construct an equivalent flow structure under a different trapping frequency. We define the scaling factor $\gamma=\omega_x/\omega_x^0$, and incorporate $\gamma$ into the ratios. This gives rise to the scaling relations, $U_{\rm s}=\gamma^{2/3}U_{\rm s}^0$ and $v_{\rm s}=\gamma^{1/3}v_{\rm s}^0$. Regime M1 is the modified case for E1, such that $U_{s}/k=\gamma^{2/3}\times 6$ nK, $v_s=$ 1.5$\gamma^{1/3}v_s^0$; likewise, M2 is the modified case for E2, so $U_{\rm s}/k=\gamma^{2/3}\times 9$ nK, $v_s=$ 1.5$\gamma^{1/3}v_s^0$.
Our investigation shows that a clear mode structure occurs in regimes where the background flow is sufficiently homogeneous. Then, the BdG modes can be described as WKB modes with well-characterized frequency and wavevector, as in Ref. [@parentani2010]. In the experimental regime \[Fig. \[fig:stimulated\_HR\](b)\], $\psi_{\rm p}$ has the longest wavelength, and is comparable to the width of the BH cavity, $L$ (the distance between the BH and WH). We find that the mode structure is improved when reducing the wavelength of the partner mode $\psi_{\rm p}$, relative to $L$.
To control the wavelength of $\psi_{\rm p}$, one can refer to the BCR mechanism and the stimulated Hawking effect, and use the dispersion relation shown in Fig. \[fig:STFT\_wt\](a) in the paper. Overall, the wavelength of the p-mode decreases with increasing step speed, $v_{\rm s}$. According to the dispersion relation, the BCR is the zero frequency mode in the WH frame, $\Delta\omega(\Delta k_{\rm BCR})=0$. Increasing $v_s$ increases the flow speed inside the supersonic cavity, which lowers (raises) the positive-$k$ (negative-$k$) branch of the dispersion curve $\Delta\omega(\Delta k)$, and displaces the intersection $\Delta\omega(\Delta k)=0$ to a larger $\Delta k$ value. This further increases the frequency $|\Delta\omega|$ of the Hawking pair, which is proportional to $k_{\rm BCR}$, and displaces the root of the dispersion curves for $\psi_{\rm HR}$ and $\psi_{\rm p}$ to greater $|\Delta k|$ \[see Fig. \[fig:STFT\_wt\](a)\]. In regime E1, we increase $v_{\rm s}$ by 50% over the experimental value. This decreases the p-mode wavelength relative to the cavity length, $L$, and the corresponding HR appears more periodic. We further extend $L$, by increasing step depth, $U_s$. Regime E2 in Fig. \[fig:evolution\_modified\] corresponds to the case with a greater step depth, in which the number of oscillations of $\psi_{\rm p}$ doubles.
Regimes M1 and M2 adopt a smaller trapping frequency, $\omega_x=(1/4)\omega^0_x$. Reducing $\omega_{x}$ increases the size of a BEC, and extends the flow structure, by which excitation modes can be more easily observed and resolved in the laboratory. We can see that cases M1 and M2 have clear mode structures as in E1 and E2, with approximately twice the cavity length. Note that regime M2 is reported in the paper, along with a mode analysis using the spatial and temporal WFTs.
Growth of the BCR mode in the M2 regime {#app:growth}
---------------------------------------
In the paper, we found that in the experimental regime the standing-wave amplitude $n_k$ inside the cavity (which later proved to be the BCR) grows in proportion to the square of the background density $n_{\rm bf}$, $n_k \propto n_{\rm bf}^2$. We use the BCR mechanism to interpret this relationship. If it is indeed the underlying mechanism that occurs in the step-sweeping experiment, the same growth relationship should be found in other parameter regimes.
![\[fig:M2\_growth\] Growth of the standing wave in the M2 regime. Panel (a) simulated growth of the standing-wave pattern in the supersonic region. Solid green: normalized standing-wave amplitude $\bar{n}_k(t)$, $\bar{n}_{k}(t)=n_{k}(t)/n_{k}(0)$. Dashed black: the square of background density, $\bar{n}_{\rm bf}(t)$, scaled to match the final standing-wave amplitude, $\bar{n}_{\rm bf}^2(t)=n_{\rm bf}^2(t)[\bar{n}_{k}(t_f)/n_{\rm bf}^2(t_f)]$. Panel (b): time evolution of $|\delta \Psi(x,t)|$, from which we select a position nearby the WH, denoted by the diagonal red (left) line, to monitor the mode growth. The growth of $n_{\rm bf}$ and $n_{k}$ is determined from a spatial WFT of $n(x)$ with window width $D= 20$ $\mu$m at the position indicated by the diagonal red (left) line in (b). Inset shows the windowed spectrum at $\text{t=650 ms}$.](M2_growth_030217){width="3.3in"}
In Fig. \[fig:M2\_growth\], we monitor the growth of the standing wave at a position nearby the WH \[indicated by the diagonal red (left) line in Fig. \[fig:M2\_growth\](b)\] . We find that the growth of the standing wave $n_k$ \[solid green curve in (a)\] matches that of the background flow $n_{\rm bf}^2$ \[dashed black curve in Fig. \[fig:M2\_growth\](a)\], which is consistent with the observation in the experimental regime. For both regimes, the relationship $n_k \propto n_{\rm bf}^2$ implies that the BCR mechanism along with the increasing background density gives rise to the mode growth inside the cavity, rather than the black-hole lasing effect. Note that the p-mode propagates to the position indicated the diagonal red (left) line at $t\sim 500$ ms, which causes some small oscillations on the growth plot $n_k(t)$.
Bogoliubov-de Gennes mode analysis: asymptotic modes, local dispersion relations, and the thermal prediction {#app:BdG}
============================================================================================================
Here, we present the Bogoliubov–de Gennes (BdG) equations and the relevant calculations discussed in the paper. In Appendix \[app:WKB\], we introduce the standard BdG formalism, and an asymptotic method (WKB) to describe modes on a slowly varying background. In Appendix \[app:dispersion\], we use the dispersion relation to determine spectral properties of the modes. In Appendix \[app:temperature\], we compare the mode amplitudes of the Hawking pair with the thermal prediction using the flow profile at the BH.
BdG equations and asymptotic BdG modes {#app:WKB}
--------------------------------------
Here, we summarize the BdG formulation presented in [@pethick; @parentani2010]. The BdG equations can be obtained by the linearization of the condensate wavefunction:
$$\begin{aligned}
\label{eq:psi}
\Psi(x,t)=\Psi_0(x,t)+\psi(x,t)\end{aligned}$$
where $\Psi_0(x,t)$ corresponds to a stationary condensate, $\Psi_0(x,t)=\sqrt{n(x)}e^{-i \mu t}$, and $\psi(x,t)$ corresponds to a deviation to the background condensate, which can be expressed as $$\begin{aligned}
\label{eq:BdG1_app}
\psi(x,t) &=& e^{-i \mu t}\left( u(x) e^{-i\omega t} +v^{*}(x) e^{i\omega t}\right),\end{aligned}$$ where $u(x)$ and $v^{*}(x)$ satisfy the BdG equations: $$\begin{aligned}
\label{eq:BdG2}
\left [\hbar\omega+\frac{\hbar^2}{2m}\frac{d^2 }{d x^2}-V(x)-2gn(x)+\mu \right ]u(x)&=&gn(x)v(x),\nonumber\\
\left [-\hbar\omega+\frac{\hbar^2}{2m}\frac{d^2 }{d x^2}-V(x)-2gn(x)+\mu \right ]v(x)&=&gn(x)u(x).\nonumber\\\end{aligned}$$ For a homogeneous system, BdG modes can be expressed as plane waves $$\begin{aligned}
\label{eq:uv_plane}
u({x})=u_k \frac{e^{ikx}}{\sqrt{2\pi}} ,\,\,\, v({x})=v_k\frac{e^{ikx}}{\sqrt{2\pi}},\end{aligned}$$ where the normalized mode amplitudes $u_k$ and $v_k$ are $$\begin{aligned}
\label{eq:uv_D}
u_k=\frac{1}{\sqrt{1-D_k^2}} ,\,\,\, v_k=\frac{D_k}{\sqrt{1-D_k^2}},\end{aligned}$$ where $D_k$ gives the ratio between $v_k$ and $u_k$, and is determined by the speed of sound $c=\sqrt{gn/m}$, $$\begin{aligned}
\label{eq:D}
D_k=\frac{1}{mc^2}\left[\sqrt{\hbar^2c^2k^2+\left(\frac{\hbar^2 k^2}{2m}\right)^2}-\frac{\hbar^2 k^2}{2m}-mc^2\right].\nonumber\\\end{aligned}$$ This leads to the dispersion relation $$\begin{aligned}
\label{eq:dispersion_supplement}
\omega(k)^2 &=& c^2k^2+\frac{\hbar^2 k^4}{4m^2}.\end{aligned}$$ Note that here we use $\omega$ and $k$ to indicate the relative frequency $\Delta \omega$ and wavevector $\Delta k$ adopted in the paper. Using Eq. \[eq:dispersion\_supplement\], the BdG modes in the wavevector ($k$) representation can be converted to the frequency ($\omega$) representation, such that $u_{\omega}=u_k/\sqrt{d\omega/dk}$ and $v_{\omega}=v_k/\sqrt{d\omega/dk}$.
Suppose the background condensate is inhomogeneous but varies smoothly in space, the BdG modes can be approximated by the WKB method as described in [@parentani2010]. The WKB-BdG modes in the $\omega$-representation are $$\begin{aligned}
\label{eq:uv_WKB}
u_{\omega}({x})=\sqrt{\frac{\partial k_{\omega}(x)}{\partial\omega}}\frac{1}{\sqrt{1-D_{k_{\omega}(x)}^2}} \frac{e^{i\int^x k_{\omega}(x')dx'}}{\sqrt{2\pi}}, \nonumber\\
v_{\omega}({x})=\sqrt{\frac{\partial k_{\omega}(x)}{\partial\omega}}\frac{D_{k_{\omega}(x)}}{\sqrt{1-D_{k_{\omega}(x)}^2}} \frac{e^{i\int^x k_{\omega}(x')dx'}}{\sqrt{2\pi}}, \end{aligned}$$ where $k_{\omega}(x)$ is determined by the local dispersion relation using the local sound speed $c(x)$ $$\begin{aligned}
\label{eq:dispersion_WKB}
\omega^2=c(x)^2k^2+\frac{\hbar^2 k^4}{4m^2}.\end{aligned}$$
{width="7in"}
Dispersion relations for the BCR mechanism and stimulated pair production {#app:dispersion}
-------------------------------------------------------------------------
{width="6.8in"}
Here, we use the local dispersion relations to determine the wavevectors and frequencies of the three BdG modes: $\psi_{\rm BCR}$, $\psi_{\rm HR}$, and $\psi_{\rm p}$. First, we transform the local dispersion relation from the comoving frame of the condensate \[Eq. \[eq:dispersion\_WKB\], denoted by $\omega_{\rm cm}(k)$\] to an observer frame in which the condensate has nonzero flow velocity $$\begin{aligned}
\label{eq:dispersion_WKB_v}
\omega &=&\sqrt{c(x)^2k^2+\frac{\hbar^2 k^4}{4m^2}}+v_{\rm bf, o}(x)k,\end{aligned}$$ where $v_{\rm bf,o}(x)$ is the local velocity of the condensate with respect to the “observer" frame in which the frequency is defined. Then, we select two points of observation: $x_{\rm I}$ inside the BH cavity, and $x_{\rm O}$ outside. The local speed of sound and the flow velocity can be evaluated by the spatial WFT of $\Psi(x)$, in which the dominant peak location gives the wavevector of background flow in the laboratory frame $k_{\rm bf}(x)$, and its peak value gives the local density $n_{\rm bf}(x)=|\Psi(k_{\rm bf},x)|^2$. The flow velocity in the BH frame is $v_{\rm bf,BH}(x)=\hbar k_{\rm bf}(x)/m-v_{\rm s}$. The WH is defined by the point where $v(x)+c(x)=0$. It is formed in the small transition region that connects the accelerated flow and the $k\sim0$ region on the left \[see Fig. \[fig:exp\](b)\], which can be identified by the left edge of $|\delta\Psi(x,t)|$ in Fig. \[fig:v\_WH\], $x_{\rm L}(t)$. Thus, we approximate the speed of WH by that of the left edge, $v_{\rm L}$, as shown in Fig. \[fig:v\_WH\]. The flow velocity in the WH frame is given by $v_{\rm bf,WH}(x)=\hbar k_{\rm bf}(x)/m-v_{\rm L}$. The change of flow velocity from the WH frame to the BH frame is equal to the velocity difference $\Delta v$ between the two horizons, $v_{\rm bf,BH}-v_{\rm bf,WH}=-\Delta v=-(v_{\rm s}-v_{\rm L})$.
According to the BCR mechanism, the BCR mode has zero frequency in the WH frame, $\omega(k_{\rm BCR})=0$. We can predict the value of $k_{\rm BCR}$ by the local dispersion relation at $x_{\rm I}$ using $c(x_{\rm I})$ and $v_{\rm bf,WH}(x_{\rm I})$, as shown in Fig. \[fig:dispersion\](a). This further determines the frequency of the stimulated pair production \[see Figs. \[fig:dispersion\](b) and (c)\]; the HR and p modes have the same frequency as that of the BCR mode in the BH frame, given by $\omega=-\Delta v k_{\rm BCR}$. This frequency intersects the dispersion relation evaluated at $x_{\rm I}$ to determine $k_{\rm p}$, and the one evaluated at $x_{\rm O}$ to determine $k_{\rm HR}$, as indicated in Figs. \[fig:dispersion\](b) and (c).
Stimulated pair creation with the thermal prediction {#app:temperature}
----------------------------------------------------
The Hawking temperature can be estimated by measuring the amplitudes of the correlated HR and partner mode. The mode mixing process at the BH is expressed by $$\begin{aligned}
\label{eq:alpha_beta}
u_{-\omega}^{\rm BCR}=\alpha_{-\omega}^{\rm p}u_{-\omega}^{\rm p}+\beta_{\omega}^{\rm HR}{\left(v_{\omega}^{\rm HR}\right)}^*,\end{aligned}$$ where $\alpha_{-\omega}^{\rm p}$ and $\beta_{\omega}^{\rm HR}$ are the positive-norm and negative-norm amplitudes of the Hawking pair, and $u_{-\omega}^{\rm BCR}$ and $u_{-\omega}^{\rm p}$ are the $u$ component of the BCR and the partner mode, $(v_{\omega}^{\rm HR})^*$ the $v$ component of the HR mode. The ratio of the amplitudes can be calculated using the thermal prediction [@robertson2012; @PhysRevA.80.043601] $$\begin{aligned}
\label{eq:Hawking}
\left|\frac{\beta_{\omega}^{\rm HR}}{\alpha_{-\omega}^{\rm p}}\right|=e^{-\frac{\pi\omega}{\kappa}},\end{aligned}$$ where $\kappa$ is the surface gravity at the BH determining the Hawking temperature $T_{\rm H}=\hbar\kappa/(2\pi k)$, $$\begin{aligned}
\label{eq:kappa}
\kappa=\left.\frac{d\left(v+c\right)}{dx}\right|_{x_{\rm BH}}.\end{aligned}$$ In Fig. \[fig:M20\](b) in the paper, we trace a correlated Hawking pair \[indicated by the magenta (left) and cyan (right) dots\] generated at $t=$ 588 ms. Using the spatial WFTs, we obtain the mode amplitudes of the pair, ${v_{\omega, {\rm FT}}^{\rm HR}}^*$ and ${u_{-\omega,{\rm FT}}^{\rm p}}$. Using Eqs. \[eq:uv\_WKB\] and \[eq:dispersion\_WKB\], they can be expressed in relation to $\left|\beta_{\omega}^{\rm HR}/\alpha_{-\omega}^{\rm p}\right|$ as $$\begin{aligned}
\label{eq:Hawking_FT}
\left|\frac{{v_{\omega,{\rm FT}}^{\rm HR}}^*}{u_{-\omega,{\rm FT}}^{\rm p}}\right| & = & \left|\frac{\beta_{\omega}^{\rm HR}}{\alpha_{-\omega}^{\rm p}}\right| \sqrt{\left(1-D_{k_{\rm p}}^2\right)\left( \frac{D_{k_{\rm HR}}^2}{1-D_{k_{\rm HR}}^2}\right)} \nonumber\\
&&~\times \left| \frac{\partial \omega/\partial k|_{k_{\rm p}}}{\partial \omega/\partial k|_{k_{\rm HR}}}\right|^{1/2},\end{aligned}$$ where $D_{k_{\rm HR}}$ (and $D_{k_{\rm p}}$) can be evaluated using Eq. \[eq:D\], and $\partial \omega/\partial k|_{k=k_{\rm HR}}$ estimated using the dispersion relation.
![Spatial WFT for a correlated Hawking pair, and surface gravity near the event horizon. (a) Windowed wavevector spectrum of the correlated HR (solid cyan curve) and partner modes (dashed magenta) at $t=650$ ms. (b) Flow velocity $v_{\rm GP}(x)$ and speed of sound $c_{\rm GP}(x)$ at the time at which the pair is created ($t=588$ ms). The BH is indicated by the blue circle, $x_{\rm BH}$. The flow velocity (dashed blue) and the speed of sound (dotted green) calculated from the spatial WFT are plotted for comparison. The surface gravity $\kappa$ is calculated from the speed slopes at the BH. []{data-label="fig:kappa"}](pair_kappa_072117){width="3.1in"}
Quantity Value
--------------------------------------------------------------------- ---------------- -----------
$|{v_{\omega,{\rm FT}}^{\rm HR}}^*/{u_{-\omega,{\rm FT}}^{\rm p}}|$ 0.11
$|D_{\rm p}|$ 0.47 (BdG) 0.40 (FT)
$|D_{\rm HR}|$ 0.39 (BdG) 0.44 (FT)
$\partial \omega/\partial k|_{k_{\rm p}} $ -744 $\mu$m/ms
$\partial \omega/\partial k|_{k_{\rm HR}} $ 409 $\mu$m/ms
$\left|\beta_{\omega}^{\rm HR}/\alpha_{-\omega}^{\rm p}\right|$ 0.21
: Numerical values of the relative mode amplitude and the relevant quantities in Eq. \[eq:Hawking\_FT\].[]{data-label="table:temperature"}
Figure \[fig:kappa\](a) shows the windowed wavevector spectra of a correlated Hawking pair at $t=650$ ms. The numerical values of the relevant quantities in Eq. \[eq:Hawking\_FT\] are given in Table \[table:temperature\]. The ratio $\left|\beta_{\omega}^{\rm HR}/\alpha_{-\omega}^{\rm p}\right|\sim 0.21 \, $. To estimate how well the linear (BdG) approximation works, we measure the quantities $|D_{\rm p}|$ and $|D_{\rm HR}|$ from the wavevector spectrum (denoted by “FT” in Table \[table:temperature\]), which correspond to the ratio between the $u$ and $v$ amplitudes of each mode, as indicated in Eq. \[eq:uv\_WKB\]. They differ from the calculated (BdG) values by 13 % for the HR mode, and 15 % for the partner mode.
We also calculate $\left|\beta_{\omega}^{\rm HR}/\alpha_{-\omega}^{\rm p}\right|$ using the thermal prediction (Eq. \[eq:Hawking\]) and the surface gravity $\kappa$ from (Eq. \[eq:kappa\]). From the data in Fig. \[fig:kappa\](b), we calculate $\kappa = 350 \,$ s$^{-1}$. This corresponds to a Hawking temperature of $T_{\rm H}=$ 0.43 nK. The ratio $\left|\beta_{\omega}^{\rm HR}/\alpha_{-\omega}^{\rm p}\right|\sim 0.17$, as determined by the thermal prediction, differs from the WFT value ($\sim0.21$) by 24%.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: '???.'
address:
- '???, '
- '???, '
author:
-
-
-
title: '???'
---
,
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Federated learning provides a framework to address the challenges of distributed computing, data ownership and privacy over a large number of distributed clients with low computational and communication capabilities. In this paper, we study the problem of learning the exact support of sparse linear regression in the federated learning setup. We provide a simple communication efficient algorithm which only needs one-shot communication with the centralized server to compute the exact support. Our method does not require the clients to solve any optimization problem and thus, can be run on devices with low computational capabilities. Our method is naturally robust to the problems of client failure, model poisoning and straggling clients. We formally prove that our method requires a number of samples per client that is polynomial with respect to the support size, but independent of the dimension of the problem. We require the number of distributed clients to be logarithmic in the dimension of the problem. If the predictor variables are mutually independent then the overall sample complexity matches the optimal sample complexity of the non-federated centralized setting. Furthermore, our method is easy to implement and has an overall polynomial time complexity.'
author:
- |
Adarsh Barik\
Department of Computer Science\
Purdue University\
West Lafayette, Indiana, USA\
`[email protected]`\
- |
Jean Honorio\
Department of Computer Science\
Purdue University\
West Lafayette, Indiana, USA\
`[email protected]`\
title: 'Exact Support Recovery in Federated Regression with One-shot Communication'
---
Introduction {#sec:introduction}
============
Modern day edge devices, with their data acquisition and storage ability, have pushed the need of distributed computing beyond the realms of data centers. Devices such as mobile phones, sensor systems in vehicles, wearable technology and smart homes, within their limited storage and processing capabilities, can constantly collect data and perform simple computations. However, due to data privacy concerns and limitations on network bandwidth and power, it becomes impractical to transmit all the collected data to a centralized server and conduct centralized training.
The nascent field of federated learning [@konevcny2015federated; @konevcny2016federated; @brendan2017aguera; @mohri2019agnostic; @li2020federated] tries to address these concerns. As described in [@konevcny2016federated], federated learning is a machine learning setting where the goal is to train a high-quality centralized model with training data distributed over a large number of clients. Unlike the data centers, the clients collect data samples independently but in a non-i.i.d. fashion. The clients may be highly unbalanced, i.e., the number of samples per client may vary significantly. The clients may also have hardware related constraints. Although the number of clients could be quite large, each client is typically a simple device which has access to a very small number of data samples and can only conduct very basic computations due to limitations on its processing and power capabilities. Furthermore, since battery power is at a premium, the communication between the client and the centralized server acts as a major bottleneck. Due to these constraints, it is common to encounter straggling and faulty clients in the federated learning setting.
In this work, we study the problem of exact support recovery of sparse linear regression in federated learning. [@wainwright2009info] provided an information theoretic lower bound for sparse linear regression. They showed that, in a completely centralized setting where all the data resides in a single server, ${\mathcal{O}}(s \log d)$ samples are necessary for exact support recovery of a $d$-dimensional parameter vector with $s$ non-zero entries. In our setting, none of the clients has the access to necessary number of data samples required for exact support recovery or possess computational capabilities to run complex algorithms. Furthermore, we only allow for one-shot communication between the clients and the centralized server, i.e., clients can send information to the centralized server only once. We propose a novel yet simple algorithm for this setting and show that local clients can collaboratively recover the exact support of the sparse linear regression model with provable theoretical guarantees.
#### Related work.
Despite being a new research area, there has been lot of interest in the field of federated learning. On the experimental side, [@konevcny2015federated] were the first to formally define federated learning and proposed an algorithm with encouraging experimental results. [@konevcny2016federated] came up with strategies to improve the communication efficiency for federated learning. [@brendan2017aguera] proposed a communication efficient algorithm for deep networks. Similarly, [@yurochkin2019bayesian] developed a novel framework for federated learning with neural networks and [@wang2020federated] proposed a federated learning algorithm using matched averaging for neural networks. [@bhagoji2019analyzing] empirically analyzed adversarial attacks on federated learning settings. They specifically studied the threat of model poisoning where the adversary controls a small number of malicious clients (usually 1) with the aim of causing the global model to misclassify. [@li2020fair] studied fair resource allocation in federated learning. On the theoretical side, [@he2018cola] proposed a new decentralized training algorithm with guarantees for convergence rates for linear classification and regression models. [@smith2017cocoa] presented a communication efficient decentralized framework which covers general non-strongly-convex regularizers, including problems like lasso with convergence rate guarantees. They also describe a possible extension of their method to one-shot communication schemes. [@smith2017federated] proposed a multi-task learning based approach for federated learning with convergence rate guarantees which was tolerant to client failure and could handle clients which lag in sending information to the centralized server (also known as straggling clients). [@mohri2019agnostic] proposed a client distribution agnostic framework for federated learning. They also provided Rademacher-based generalization bounds for their proposed approach.
#### Our Contribution.
All the work mentioned above are interesting in their own domain however our contribution is mostly theoretical. The existing theoretical work provide guarantees for convergence rates (which guarantees a small mean squared error in the training set provided enough iterations) or generalization bounds (which guarantees a small mean squared error in the testing set provided enough samples). However, the final solution may not match exactly with the true parameter vector. In this work, we provide provable theoretical guarantees for exact recovery of the support of the true sparse paramater vector of linear regression in federated learning. Support recovery, i.e., correctly detecting the zero and nonzero entries of the parameter vector, is arguably a challenging task. In particular, we show that for a $d$-dimensional $s$-sparse parameter vector ${\mathcal{O}}(\log d)$ clients and ${\mathcal{O}}(s^2 \log s)$ data samples per client are sufficient to recover the exact support. If the predictor variables are mutually independent then we can do exact support recovery with only ${\mathcal{O}}(s)$ data samples per client. Notice that in this case the aggregate sample complexity is ${\mathcal{O}}(s\log d)$ which matches the optimal sample complexity of the centralized setting[@wainwright2009info; @wainwright2009sharp]. We propose a simple yet effective method for exact support recovery and prove that the method is *correct* and efficient in terms of *time* and *sample complexity*. Our method has the following key properties:
- **Simplicity:** We do not solve any optimization problem at the client level. All the computations are simple and let us use our method in devices with low computational power.
- **One shot communication and privacy:** Our method is communication efficient. We only need one round communication of at most $d-$bits from the client to the centralized server. As the communication is kept to a minimum, very little information about the client is passed to the centralized server.
- **Fault tolerance and aversion to model poisoning and straggling:** Our method is naturally robust to client node failure and averse to rogue and straggling clients.
Preliminaries {#sec:preliminaries}
=============
In this section, we collect the notation which we use throughout this paper. We also formally define the support recovery problem for sparse linear regression in federated learning.
Notation and Problem Setup {#subsec:notation and problem setup}
--------------------------
Let $w^* \in {\mathbb{R}}^d$ be a $d-$dimensional parameter with sparsity $s$, i.e., only $s$ out of $d$ entries of $w^*$ are non-zero. We use ${[r]}$ as a shorthand notation to denote the set $\{1,2,\cdots,r\}$. Let $S^*$ be the true support set, i.e., $S^* = \{ r | w^*_r \ne 0, r \in {[d]} \}$. We denote the corresponding complementary non-support set as $S^*_c = \{ r | w^*_r = 0, r \in {[d]} \}$. Assume there are $g$ clients, each with $n_i$ independent samples, for $i \in [g]$. Note that the data distribution across $g$ clients need not be identical. Each client $i \in {[g]}$ contains each data sample in the format $(X_i, y_i)$ where $X_i \in {\mathbb{R}}^d$ are the predictor variables and $y_i \in {\mathbb{R}}$ is the response variable. The data generation process for each client $i \in {[g]}$ is as follows: $$\begin{aligned}
\label{eq:generative model}
\begin{split}
y_i = X_i^{\intercal}w^* + e_i \; ,
\end{split}
\end{aligned}$$ where $e_i$ is a zero mean sub-Gaussian additive noise with variance proxy $\eta_i^2$, where $\eta_i > 0$. Note that all the clients share the same parameter vector $w^*$. The $j$-th entry of $X_i$ is denoted by $X_{ij}, \forall i\in {[g]}, j \in {[d]}$. Each entry $X_{ij}$ of $X_i$ is a zero mean sub-Gaussian random variable with variance proxy $\rho_i^2$, where $\rho_i > 0$. We denote covariance matrix for $X_i$ as $\Sigma^i \in {\mathbb{R}}^{d \times d}$ with diagonal entries $\Sigma^i_{jj} \equiv {\sigma^i_{jj}}^2, \forall j \in {[d]}$ and non-diagonal entries $\Sigma^i_{jk} \equiv \sigma^i_{jk}, \forall j,k \in {[d]}, j \ne k$. If predictor variables are mutually independent then $\sigma^i_{jk} = 0, \forall i \in {[g]}, j,k \in {[d]}, j \ne k$. The $t$-th sample of the $i$-th client is denoted by $(X_i^t, y_i^t), \forall i \in {[g]}, t \in {[n_i]}$. We note that $X_i^t \in {\mathbb{R}}^d$ and $y_i^t \in {\mathbb{R}}$ and denote $j$-th entry of $X_i^t$ as $X_{ij}^t$. Notice that the data distributions for $(X_i, y_i)$ can vary a lot across the clients by varying $\rho_i$ and $\eta_i$, as well as the specific sub-Gaussian probability distribution. The class of sub- Gaussian variates includes for instance Gaussian variables, any bounded random variable (e.g., Bernoulli, multinomial, uniform), any random variable with strictly log-concave density, and any finite mixture of sub-Gaussian variables. Similarly, data samples can be distributed unevenly across the clients by varying $n_i$. In subsequent sections, we use ${\mathbb{P}}(A)$ to denote probability of the event $A$ and ${\mathbb{E}}(A)$ to denote the expectation of the random variable $A$.
Problem Statement {#subsec: problem statement}
-----------------
For our problem, we assume that $n_i < {\mathcal{O}}(s \log d), \forall i \in {[g]}$. Otherwise, support can be trivially recovered by using compressed sensing methods in the client with $n_i = {\mathcal{O}}(s \log d)$ which is the order of necessary and sufficient number of samples required for exact support recovery in linear regression setup [@wainwright2009info; @wainwright2009sharp]. Furthermore, we assume that each of our clients can only do very simple computations and can only do one-shot communication with the centralized server, i.e., each client can only send at most $d$-bits to the centralized server. Considering the above requirements, we are interested in answering the following question:
\[prob:exact support recovery\] Given that each client contains $n_i < {\mathcal{O}}(s\log d)$ data samples generated through the process described in equation , is it possible to efficiently recover the true support of the $s$-sparse shared parameter vector $w^* \in {\mathbb{R}}^d$ by collecting $d$-bits of information from every client only once with provable theoretical guarantees.
The efficiency in exact recovery means that the sample complexity per client should be strictly less than ${\mathcal{O}}(s\log d)$ and that our algorithm should have polynomial time complexity and should also be easy to implement.
Our Method {#sec:methodology}
==========
In this section, we present a simple algorithm to solve problem \[prob:exact support recovery\]. Our main idea is that estimation at the client level can be incorrect for every client but this information can still be aggregated in a careful manner to compute the true support.
Client Level Computations {#subsec:client level computations}
-------------------------
Each client tries to estimate the support of $w^*$ using $n_i$ independent samples available to it. As mentioned previously, $n_i, \forall i \in {[g]}$ is not sufficient to compute correct support of $w^*$ using any method possible [@wainwright2009info]. Let ${\hat{w}}_i \in {\mathbb{R}}^d$ be the estimate of $w^*$ computed by each client $i$. Let $S_i$ be the support of ${\hat{w}}_i$. Each server communicates the computed support (at most $d$ bits) to a centralized server which then computes the final support for $w^*$. The centralized server receives $S_i$ from each client and computes the final support $S = f(S_1, S_2,\cdots,S_g)$. Each client $i, \forall i \in {[g]}$ computes ${\hat{w}}_i$ in the following way: $$\begin{aligned}
\label{eq:what}
\begin{split}
\forall i \in {[g]}, j \in {[d]},\quad {\hat{w}}_{ij} = \frac{1}{{\hat{\sigma}}_{ij}} {\text{sign}}({\hat{\alpha}}_{ij}) \max (0, |{\hat{\alpha}}_{ij}| - \lambda) \; ,
\end{split}
\end{aligned}$$ where ${\hat{w}}_{ij}$ is $j$-th entry of ${\hat{w}}_i$ and $\lambda > 0$ is a regularization parameter. We present the exact procedure to compute a feasible $\lambda$ in later sections. We also define ${\hat{\sigma}}_{ij}$ and ${\hat{\alpha}}_{ij}$ as follows: $$\begin{aligned}
\label{eq:sigma_alpha}
\begin{split}
{\hat{\sigma}}_{ij} \triangleq \frac{1}{n_i} \sum_{t=1}^{n_i} (X_{ij}^t)^2,\quad {\hat{\alpha}}_{ij} \triangleq \frac{1}{n_i} \sum_{t=1}^{n_i} y_i^t X_{ij}^t
\end{split}
\end{aligned}$$ Note that these are simple calculations and can be done in ${\mathcal{O}}(d n_i)$ run time at each client. If $n_i$ can be kept small (which we will show later), this can be done even by a device with low computational ability. The choice of this exact form of ${\hat{w}}_{ij}$ in equation is not arbitrary. To get the intuition behind our choice, consider the following $\ell_1$-regularized (sparse) linear regression problem at each client. $$\begin{aligned}
\label{eq:serverlasso}
\begin{split}
(\forall i\in{[g]}), \quad {\hat{w}}_i = \arg\min_w \frac{1}{n_i}\sum_{t=1}^{n_i} (w^{\intercal}X_i^t - y_i^t)^2 + \lambda \| w \|_1 \; ,
\end{split}
\end{aligned}$$ where $\| \cdot \|_1$ denotes the $\ell_1$ norm of a vector. The construction of ${\hat{w}}_i$ in equation is the exact solution to optimization problem if predictor variables, i.e., the entries in $X_{i}$, are assumed to be uncorrelated. Notice how the solution provided in avoids any computation (or estimation) of the covariance matrix which, in any case, would be incorrect if each client has access to only a few samples. Each client $i$ sends the support $S_i = \{ j | {\hat{w}}_{ij} \ne 0, j \in {[d]} \}$ of ${\hat{w}}_i$ to the centralized server. Note that even in the worst case scenario, each client only sends $d$ bits to the centralized server.
Information Aggregation and Constructing the Final Support {#subsec:constructing final support S}
----------------------------------------------------------
We aggregate supports $S_i, \forall i \in {[g]}$ from all the clients and construct the final support. Before we get to the construction of the final support, we define a random variable $R_{ij}, \forall i \in {[g]}, j \in {[d]}$ which takes value $1$ if $j \in S_i$ and $0$ otherwise.
Thus, random variable $R_{ij}$ indicates whether entry $j$ is in the support $S_i$ of client $i$. Using the random variables $R_{ij}$, we construct the final support $S$ by computing the median of $R_{ij}$ across $i \in {[g]}$. If the median is $1$ then we conclude that $j$ is in the support otherwise we conclude that $j$ is not in the support. More formally, we define a random variable $R_j \triangleq \frac{1}{g}\sum_{i=1}^g R_{ij}$ and if $R_j \geq \frac{1}{2}$, then we conclude that $j \in S$. Otherwise, if $R_j < \frac{1}{2}$, then we conclude that $j \notin S$. The above procedure can be compactly written as the following algorithms running in clients and centralized server:
\[algo:getExactSupport\]
$R_i \leftarrow \{0\}^d$ Send $R_i$ to centralized server
$S \leftarrow \{\}$
Main Results and Analysis {#sec:analysis}
=========================
In this section, we describe and analyze our theoretical results. We present our results in two different settings. In the first setting, we assume that predictor variables are mutually independent. We tackle the more general case of correlated predictors in the second setting.
Mutually Independent Predictors {#subsec:mutually independent predictors}
-------------------------------
In this setting, predictor variables are mutually independent of each other in all the clients, i.e., $\forall i \in {[g]}$, ${\mathbb{E}}(X_{ij} X_{ik}) = 0, \forall j, k \in {[d]}, j \ne k$. In this setting, we state the following result:
\[thm:mutually independent predictors\] For the federated support learning for linear regression as described in Section \[sec:preliminaries\] where predictor variables are mutually independent of each other, if for some $0 < \delta < 1$, each of the $g = {\mathcal{O}}(\log d)$ client has $n_i = {\mathcal{O}}(\frac{1}{\delta^2})$ data samples and if for each $i \in {[g]}$ and $j \in S^*$, $$\begin{aligned}
\begin{split}
8 \delta \rho_i^2 \sqrt{\sum_{k\in S^*} w_k^2} + 8 |\eta_i\rho_i|\delta < \lambda < |w_j^* {\sigma_{jj}^i}^2| - 8 |w_j^*| \rho_i^2 \delta - 8 \rho_i^2 \sqrt{\sum_{k\in S^*, k \ne j} {w_k^*}^2} \delta - 8 |\eta_i\rho_i| \delta \;,
\end{split}
\end{aligned}$$ then Algorithm \[algo:getExactSupport\] recovers the exact support for the shared parameter vector $w^*$.
By taking $\delta = {\mathcal{O}}(\frac{1}{\sqrt{s}})$, we get the following corollary:
\[cor:mutually independent predictors\] For the federated support learning for linear regression as described in Section \[sec:preliminaries\] where predictor variables are mutually independent of each other, if for some constant $0 < K < \sqrt{s}$, each of the $g = {\mathcal{O}}(\log d)$ client has $n_i = {\mathcal{O}}(s)$ data samples and if for each $i \in {[g]}$ and $j \in S^*$, $$\begin{aligned}
\begin{split}
&8 K \rho_i^2 \sqrt{\frac{\sum_{k\in S^*} {w_k^*}^2}{s}} + 8 K \frac{|\eta_i\rho_i|}{\sqrt{s}} < \lambda < |w_j^* {\sigma_{jj}^i}^2| - 8 K \frac{|w_j^*| \rho_i^2}{\sqrt{s}} - 8 K \rho_i^2 \sqrt{\frac{\sum_{k\in S^*, k \ne j} {w_k^*}^2}{s}} - \\
& 8 K \frac{|\eta_i\rho_i|}{\sqrt{s}} \;,
\end{split}
\end{aligned}$$ then Algorithm \[algo:getExactSupport\] recovers the exact support for the shared parameter vector $w^*$.
The choice of such value of $\delta$ is to subdue the growth of the $\sqrt{\sum_{k\in S^*} {w_k^*}^2}$ term which approximately grows as ${\mathcal{O}}(\sqrt{s})$. Later on, we will empirically show that such a choice leads to a feasible range for $\lambda$. Also observe that, the overall sample complexity for our algorithm is ${\mathcal{O}}(s \log d)$ which matches the optimal sample complexity for sparse linear regression[@wainwright2009info; @wainwright2009sharp], i.e., even if we have access to all the samples in a centralized server, we can not have a better sample complexity guarantee for support recovery.
### Proof of Theorem \[thm:mutually independent predictors\] {#subsubsec: proof of theorem mutually independent predictors}
\[proof:theorem mutually independent predictors\] Recall that $R_j = \frac{1}{g} \sum_{i=1}^g R_{ij}$ where $R_{ij}$ is defined in Section \[subsec:constructing final support S\]. We prove that, with high probability, $R_j \geq \frac{1}{2}, \forall j \in S^*$ and $R_j < \frac{1}{2}, \forall j \in S^*_c$. We will provide the proof in two parts. First, we deal with entries $j$ which are in the support of $w^*$, i.e., $j \in S^*$ and then we will deal with $j \in S^*_c$.
#### For entries $j$ in support $S^*$.
We begin our proof by first stating the following lemma.
\[lem:support mcdiarmid\] For $j \in S^*$, let ${\mathbb{E}}(R_j) > \frac{1}{2}$, then $R_j \geq \frac{1}{2}$ with probability at least $1 - 2 \exp( - 2g (-\frac{1}{2} + {\mathbb{E}}(R_j))^2)$.
Next we show that for $j \in S^*$, ${\mathbb{E}}(R_j)$ is indeed greater than $\frac{1}{2}$. To that end, we provide the result of the following lemma.
\[lem:support bound on E(Rj)\] For $j \in S^*$ and some $0 < \delta \leq 1$, if predictors are mutually independent of each other and $ 0 < \lambda < |w_j^* {\sigma_{jj}^i}^2| - 8 |w_j^*| \rho_i^2 \delta - 8 \rho_i^2 \sqrt{\sum_{k\in S^*, k \ne j} {w_k^*}^2} \delta - 8 |\eta_i\rho_i| \delta $ then $ {\mathbb{E}}(R_j) \geq 1 - \frac{6}{g} \sum_{i=1}^g \exp(-n_i\delta^2) $. Furthermore, for $n_i = {\mathcal{O}}(\frac{1}{\delta^2})$, we have ${\mathbb{E}}(R_j) > \frac{1}{2}$.
#### For entries $j$ in non-support $S^*_c$.
Similar to the entries in the support, we begin this part by stating the following result for entries in the non-support.
\[lem:nonsupport mcdiarmid\] For $j \in S^*_c$, let ${\mathbb{E}}(R_j) < \frac{1}{2}$, then $R_j \leq \frac{1}{2}$ with probability at least $1 - 2 \exp( - 2g (\frac{1}{2} - {\mathbb{E}}(R_j))^2)$.
It remains to show that for $j \in S^*_c$, ${\mathbb{E}}(R_j)$ is smaller than $\frac{1}{2}$. In particular, we use the result from the following lemma.
\[lem:nonsupport bound E(Rj)\] For $j \in S^*_c$ and $0 < \delta \leq 1$, if predictors are mutually independent of each other and if $ \lambda > 8 \delta \rho_i^2 \sqrt{\sum_{k\in S^*} w_k^2} + 8 |\eta_i\rho_i|\delta $ then ${\mathbb{E}}(R_j) \leq \frac{4}{g} \sum_{i=1}^g \exp(-n_i \delta^2)$. Furthermore, for $n_i = {\mathcal{O}}(\frac{1}{\delta^2})$, we have ${\mathbb{E}}(R_j) < \frac{1}{2}$.
Results from Lemma \[lem:support bound on E(Rj)\] and \[lem:nonsupport bound E(Rj)\] make sure that Lemma \[lem:support mcdiarmid\] and \[lem:nonsupport mcdiarmid\] hold. We would like these results to hold across all $j \in {[d]}$. This implies that we need a union bound across all the $d$ predictors. Thus, having $g = {\mathcal{O}}(\log d)$ ensures that our results hold for all entries in the support and non-support with high probability.
Correlated predictors {#subsec:correlated predictors}
---------------------
Now that we have dealt with mutually independent predictors, we focus on correlated predictors in this section. As described previously, the covariance matrix for $X_i$ is denoted by $\Sigma^i \in {\mathbb{R}}^{d \times d}$ with diagonal entries $\Sigma^i_{jj} \equiv {\sigma^i_{jj}}^2, \forall j \in {[d]}$ and non-diagonal entries $\Sigma^i_{jk} \equiv \sigma^i_{jk}, \forall j,k \in {[d]}, j \ne k$. Some of the results from the previous subsection can be used for this setting as well. However, correlation between predictors affects some results. Below, we state the main results for this setting before proving them formally.
\[thm:correlated predictors\] For the federated support learning for linear regression as described in Section \[sec:preliminaries\], if for some $0 < \delta < \frac{1}{\sqrt{2}}$, each of the $g = {\mathcal{O}}(\log d)$ client has $n_i = {\mathcal{O}}(\frac{1}{\delta^2} \log s)$ data samples and if for each $i \in {[g]}$, $$\begin{aligned}
\begin{split}
&(\forall j \in S^*_c) |\sum_{k\in S^*} w^*_k \sigma^i_{jk}| + \sum_{k\in S^*} 8 \sqrt{2} |w_k^*| (1 + 4 \max_j \frac{\rho_i^2}{{\sigma^i_{jj}}^2}) \max_j {\sigma^i_{jj}}^2 \delta + 8 |\eta_i \rho_i| \delta < \lambda \\
&< (\forall j \in S^*) |(w^*_j{\sigma^i_{jj}}^2 + \sum_{k\in S^*, k\ne j} w^*_k \sigma^i_{jk})| - 8 |w_j^*| \rho_i^2 \delta - \sum_{k\in S^*, k\ne j} 8 \sqrt{2} |w_k^*| (1 + 4 \max_j \frac{\rho_i^2}{{\sigma^i_{jj}}^2}) \\
&\max_j {\sigma^i_{jj}}^2 \delta - 8 |\eta_i \rho_i| \delta \;,
\end{split}
\end{aligned}$$ then Algorithm \[algo:getExactSupport\] recovers the exact support for the shared parameter vector $w^*$.
By taking $\delta = {\mathcal{O}}(\frac{1}{s})$, we get the following corollary:
\[cor:correlated predictors\] For the federated support learning for linear regression as described in Section \[sec:preliminaries\], if for some constant $0 < K < \frac{s}{\sqrt{2}}$, each of the $g = {\mathcal{O}}(\log d)$ client has $n_i = {\mathcal{O}}(s^2 \log s)$ data samples and if for each $i \in {[g]}$, $$\begin{aligned}
\begin{split}
&(\forall j \in S^*_c) |\sum_{k\in S^*} w^*_k \sigma^i_{jk}| + \sum_{k\in S^*} 8 \sqrt{2} |w_k^*| (1 + 4 \max_j \frac{\rho_i^2}{{\sigma^i_{jj}}^2}) \max_j {\sigma^i_{jj}}^2 \frac{K}{s} + 8 |\eta_i \rho_i| \frac{K}{s} < \lambda \\
&< (\forall j \in S^*) |(w^*_j{\sigma^i_{jj}}^2 + \sum_{k\in S^*, k\ne j} w^*_k \sigma^i_{jk})| - 8 |w_j^*| \rho_i^2 \frac{K}{s} - \sum_{k\in S^*, k\ne j} 8 \sqrt{2} |w_k^*| (1 + 4 \max_j \frac{\rho_i^2}{{\sigma^i_{jj}}^2}) \\
&\max_j {\sigma^i_{jj}}^2 \delta - 8 |\eta_i \rho_i| \frac{K}{s} \;,
\end{split}
\end{aligned}$$ then Algorithm \[algo:getExactSupport\] recovers the exact support for the shared parameter vector $w^*$.
As with the previous case, the choice of such a value of $\delta$ is to subdue the growth of terms which grow as ${\mathcal{O}}(s)$. In our experiments, this leads to a feasible range for $\lambda$. In this case, the overall sample complexity for our algorithm is ${\mathcal{O}}(s^2 \log s \log d)$ which only differs by a factor of $s\log s$ from the optimal sample complexity for support recovery in sparse linear regression in the centralized setting where all the data resides in a single server[@wainwright2009info; @wainwright2009sharp].
### Proof of Theorem \[thm:correlated predictors\] {#subsubsec:proof of theorem correlared predictors}
\[proof:theorem correlared predictors\] Recall that $R_j = \frac{1}{g} \sum_{i=1}^g R_{ij}$ where $R_{ij}$ is defined in Section \[subsec:constructing final support S\]. We will again prove that, with high probability, $R_j \geq \frac{1}{2}, \forall j \in S^*$ and $R_j < \frac{1}{2}, \forall j \in S^*_c$. Some of the results from the previous Section \[proof:theorem mutually independent predictors\] follow without any changes. We provide new results for the remaining parts. Like before first, we deal with entries $j$ which are in the support of $w^*$, i.e., $j \in S^*$ and then we will deal with $j \in S^*_c$.
#### For entries $j$ in support $S^*$.
We observe that Lemma \[lem:support mcdiarmid\] holds even in this case. Thus, we start our proof by stating the following lemma.
\[lem:support bound on E(Rj) correlated\] For $j \in S^*$ and some $0 < \delta \leq \frac{1}{\sqrt{2}}$, if $\forall j \in S^*$, $ 0 < \lambda < |(w^*_j{\sigma^i_{jj}}^2 + \sum_{k\in S^*, k\ne j} w^*_k \sigma^i_{jk})| - 8 |w_j^*| \rho_i^2 \delta - \sum_{k\in S^*, k\ne j} 8 \sqrt{2} |w_k^*| (1 + 4 \max_j \frac{\rho_i^2}{{\sigma^i_{jj}}^2}) \max_j {\sigma^i_{jj}}^2 \delta - 8 |\eta_i \rho_i| \delta $ then $ {\mathbb{E}}(R_j) \geq 1 - \frac{4s}{g} \sum_{i=1}^g \exp(-n_i\delta^2 ) $. Furthermore, for $n_i = {\mathcal{O}}(\frac{1}{\delta^2}\log s)$, we have ${\mathbb{E}}(R_j) > \frac{1}{2}$.
#### For entries $j$ in non-support $S^*_c$.
Again, Lemma \[lem:nonsupport bound E(Rj)\] follows directly. Thus, we present the following lemma to show that for the entries in the non-support ${\mathbb{E}}(R_j) < \frac{1}{2}$.
\[lem:nonsupport bound E(Rj) correlated\] For $j \in S^*_c$ and some $0 < \delta \leq \frac{1}{\sqrt{2}}$, if $ \lambda > |\sum_{k\in S^*} w^*_k \sigma^i_{jk}| + \sum_{k\in S^*} 8 \sqrt{2} |w_k^*| (1 + 4 \max_j \frac{\rho_i^2}{{\sigma^i_{jj}}^2}) \max_j {\sigma^i_{jj}}^2 \delta + 8 |\eta_i \rho_i| \delta $ then ${\mathbb{E}}(R_j) \leq \frac{4s + 2}{g} \sum_{i=1}^g \exp(-n_i \delta^2)$. Furthermore, for $n_i = {\mathcal{O}}(\frac{1}{\delta^2} \log s)$, we have ${\mathbb{E}}(R_j) < \frac{1}{2}$.
Results from Lemmas \[lem:support bound on E(Rj) correlated\] and \[lem:nonsupport bound E(Rj) correlated\] ensure that Lemma \[lem:support mcdiarmid\] and \[lem:nonsupport mcdiarmid\] hold. Since we would like these results to hold across all $j \in {[d]}$, we need a union bound across all the $d$ predictors. Thus, having $g = {\mathcal{O}}(\log d)$ makes sure that our results hold for all entries in the support and non-support with high probability.
Time Complexity {#sub:time complexity}
---------------
Each client does ${\mathcal{O}}(dn_i)$ basic calculations. Thus, from the results of Corollaries \[cor:mutually independent predictors\] and \[cor:correlated predictors\], the time complexity at each client is ${\mathcal{O}}(sd)$ for mutually independent predictors and ${\mathcal{O}}(s^2 d \log s)$ for correlated predictors. The centralized server gathers $d$-bits of information from $g$ clients in ${\mathcal{O}}(dg)$ time.
Discussion on Robustness {#sec:discussion on robustness}
========================
Since our method only relies on the correct calculation of the median, it is naturally robust to failure of few clients. To simulate the effect of model poisoning [@bhagoji2019analyzing] and stragglers, we consider that $0 < \beta < \frac{1}{2}$ portion of clients have gone rogue (are straggling) and transmitting the wrong information to the centralized server. For the worst case scenario, we assume that they report the complement of the support, i.e., they always send a bit “$1$” for entries in the non-support and a bit “$0$” for entries in the support. To accommodate this change in the case of correlated predictors, we slightly change statements of Lemmas \[lem:support bound on E(Rj) correlated\] and \[lem:nonsupport bound E(Rj) correlated\]. Now we have, $ (\forall j \in S^*), \quad {\mathbb{E}}(R_j) \geq (1 - \beta) - \frac{4s}{g} \sum_{i=1}^{(1-\beta)g} \exp(-n_i \delta^2) $ and $ (\forall j \in S^*_c), \quad {\mathbb{E}}(R_j) \leq \frac{4s + 2}{g} \sum_{i=1}^{(1 - \beta)g} \exp(-n_i \delta^2) + \beta $. It is easy to see that, as long as, we have $n_i > \frac{1}{\delta^2} \log(\frac{(8s + 4)(1 - \beta)}{1 - 2 \beta })$ data samples per client, then we still have ${\mathbb{E}}(R_j) > \frac{1}{2}, \forall j \in S^*$ and ${\mathbb{E}}(R_j) < \frac{1}{2}, \forall j \in S^*_c$ and all our results still hold. A similar analysis can be conducted for the case of mutually independent predictors and our results hold as long as we have $n_i > \frac{1}{\delta^2} \log (\frac{12(1 - \beta)}{1 - 2 \beta})$ data samples per client.
Experimental Results {#sec:experimental results}
====================
[.45]{} ![Phase transition curves. Left: Exact support recovery averaged across $30$ runs against varying number of samples per client for $d = 500, 1000$, and $2000$, $s = 3$, $g = {\mathcal{O}}(\log d)$ clients. Right: Exact support recovery averaged across $30$ runs against varying number of clients for $s = 10, 20, 40$, and $50$, $d = 1000$, $n = \max(30, {\mathcal{O}}(s^2 \log s))$ samples per server.[]{data-label="fig:recovery"}](exact_support_rec_num_samples "fig:"){width="\linewidth"}
[.45]{} ![Phase transition curves. Left: Exact support recovery averaged across $30$ runs against varying number of samples per client for $d = 500, 1000$, and $2000$, $s = 3$, $g = {\mathcal{O}}(\log d)$ clients. Right: Exact support recovery averaged across $30$ runs against varying number of clients for $s = 10, 20, 40$, and $50$, $d = 1000$, $n = \max(30, {\mathcal{O}}(s^2 \log s))$ samples per server.[]{data-label="fig:recovery"}](exact_support_rec_num_clients "fig:"){width="\linewidth"}
In this section, we validate our theoretical results by conducting computational experiments. We provide the results for the experiments when predictors are correlated. Data in each client is generated by following generative process described in equation \[eq:generative model\]. Note that predictors and error term in different clients follow different sub-Gaussian distributions. To make it more general, we keep the correlation between entries in the support different than the correlation between one entry in the support and the other entry in the non-support and these further vary across clients. The regularization parameter $\lambda$ is chosen such that condition in corollary \[cor:correlated predictors\] is satisfied for every client and for every entry in support and non-support. All the results reported here are averaged over 30 independent runs. We conduct two separate experiments to verify that $n_i = {\mathcal{O}}(s^2\log s)$ independent samples per client and a total of $g = {\mathcal{O}}(\log d)$ clients are sufficient to recover the true support.
#### Exact support recovery against number of samples per client.
This experiment was conducted for a varying number of predictors ($d = 500, 1000$ and $2000$). For each of them, we fixed the number of clients to be $g = 2 \log d$. The sparsity $s$ is kept fixed at $3$. The number of samples per client $n_i$ is varied with control parameter $C$ as $10^C s^2 \log s$. Performance of our method is measured by assigning value $1$ for exact recovery and $0$ otherwise. We can see in Figure \[fig:recnumsample\], that initially, recovery remains at $0$ and then there is sharp jump after which recovery becomes $1$. Notice how all three curves align perfectly. This validates the result of our theorem and shows that given $g = {\mathcal{O}}(\log d)$ clients, $n_i = {\mathcal{O}}(s^2 \log s)$ samples per client are sufficient to recover the true support.
#### Exact support recovery against number of clients.
The second experiment was conducted for a varying number of non-zero entries ($s=10, 20, 40$ and $50$) in the support of $w^*$. The experiments were run for a setup with $d=1000$ predictors. We fixed the number of samples per client ($n_i$) to be $\max(30, {\mathcal{O}}(s^2 \log s))$. This ensures that a minimum of $30$ samples are available to each client. This is inline with our previous experiment where exact recovery is achieved around $30$ samples per client. The number of clients $g$ is varied with control parameter $C$ as $10^C \log d$. Like previous experiment, performance is measured by assigning value $1$ for exact recovery and $0$ otherwise. We can again see in Figure \[fig:recnumclient\], that initially, recovery is remains at $0$ and then it goes to $1$ as we increase number of clients. We also notice that all four curves align nicely. This validates that given $n_i = {\mathcal{O}}(s^2 \log s)$ independent samples per server, $g = {\mathcal{O}}(\log d)$ clients are sufficient to recover the true support.
Concluding Remarks {#sec:conclusion}
==================
In this paper, we propose a simple and easy to implement method for learning the exact support of parameter vector of linear regression problem in a federated learning setup. We provide theoretical guarantees for the correctness of our method. We also show that our method runs in polynomial sample and time complexity. Furthermore, our method is averse to client failures, model poisoning and straggling clients. As a future direction, it would be interesting to analyze if the bound on the sample complexity in the case of correlated predictors matches corresponding information theoretic lower bounds.
[10]{}
Arjun Nitin Bhagoji, Supriyo Chakraborty, Prateek Mittal, and Seraphin Calo. Analyzing federated learning through an adversarial lens. , 2019.
McMahan H Brendan, Moore Eider, Ramage Daniel, Hampson Seth, and Ag[ü]{}era y Arcas Blaise. Communication-efficient learning of deep networks from decentralized data. In [*Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS)*]{}, 2017.
Lie He, An Bian, and Martin Jaggi. Cola: Decentralized linear learning. In [*Advances in Neural Information Processing Systems*]{}, pages 4536–4546, 2018.
Jakub Kone[č]{}n[y]{}, Brendan McMahan, and Daniel Ramage. Federated optimization: Distributed optimization beyond the datacenter. , 2015.
Jakub Kone[č]{}n[y]{}, H Brendan McMahan, Felix X Yu, Peter Richt[á]{}rik, Ananda Theertha Suresh, and Dave Bacon. Federated learning: Strategies for improving communication efficiency. , 2016.
Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. , 37(3):50–60, 2020.
Tian Li, Maziar Sanjabi, and Virginia Smith. Fair resource allocation in federated learning. , 2020.
Colin McDiarmid. On the method of bounded differences. , 141(1):148–188, 1989.
Mehryar Mohri, Gary Sivek, and Ananda Theertha Suresh. Agnostic federated learning. , 2019.
Pradeep Ravikumar, Martin J Wainwright, Garvesh Raskutti, Bin Yu, et al. High-dimensional covariance estimation by minimizing l1-penalized log-determinant divergence. , 5:935–980, 2011.
Virginia Smith, Chao-Kai Chiang, Maziar Sanjabi, and Ameet S Talwalkar. Federated multi-task learning. In [*Advances in Neural Information Processing Systems*]{}, pages 4424–4434, 2017.
Virginia Smith, Simone Forte, Chenxin Ma, Martin Tak[á]{}[č]{}, Michael I Jordan, and Martin Jaggi. Cocoa: A general framework for communication-efficient distributed optimization. , 18(1):8590–8638, 2017.
M Wainwright. Chapter 2: Basic tail and concentration bounds, 2015.
M. J. Wainwright. Information-theoretic bounds on sparsity recovery in the high-dimensional and noisy setting. , 55:5728–5741, December 2009.
Martin J Wainwright. Sharp thresholds for high-dimensional and noisy sparsity recovery using l1-constrained quadratic programming (lasso). , 55(5):2183–2202, 2009.
Hongyi Wang, Mikhail Yurochkin, Yuekai Sun, Dimitris Papailiopoulos, and Yasaman Khazaeni. Federated learning with matched averaging. , 2020.
Mikhail Yurochkin, Mayank Agarwal, Soumya Ghosh, Kristjan Greenewald, Trong Nghia Hoang, and Yasaman Khazaeni. Bayesian nonparametric federated learning of neural networks. , 2019.
Proof of Lemma \[lem:support mcdiarmid\] {#sec:proof of lemma support mcdiarmid}
========================================
\[proof:support mcdiarmid\] Note that since $w^*$ is fixed, for any $j \in {[d]}$, $R_{ij}$ across $i \in {[g]}$ are independent of each other. $$\begin{aligned}
\begin{split}
{\mathbb{P}}(R_j < \frac{1}{2}) &= {\mathbb{P}}(R_j - {\mathbb{E}}(R_j) < \frac{1}{2} - {\mathbb{E}}(R_j)) \\
&\leq {\mathbb{P}}(|R_j - {\mathbb{E}}(R_j)| > - \frac{1}{2} + {\mathbb{E}}(R_j))
\end{split}
\end{aligned}$$ The last inequality holds as long as ${\mathbb{E}}(R_j) > \frac{1}{2}$. We also note that for a fixed $w^*$, $R_j = f(R_{1j}, R_{2j}, \cdots, R_{gj})$ such that $$\begin{aligned}
\begin{split}
(\forall k \in {[g]}),\quad \sup_{R_{1j}, \cdots, R_{kj}, R_{kj}',\cdots, R_{gj}}|f(R_{1j}, \cdots, R_{kj},\cdots, R_{gj}) - f(R_{1j}, \cdots, R_{kj}',\cdots, R_{gj})| \leq \frac{1}{g}
\end{split}
\end{aligned}$$ Thus using McDiarmid’s inequality [@mcdiarmid1989method], we can write the following: $$\begin{aligned}
\begin{split}
{\mathbb{P}}(|R_j - {\mathbb{E}}(R_j)| > - \frac{1}{2} + {\mathbb{E}}(R_j)) \leq 2 \exp( - 2g (-\frac{1}{2} + {\mathbb{E}}(R_j))^2)
\end{split}
\end{aligned}$$
Proof of Lemma \[lem:support bound on E(Rj)\] {#sec:proof of lemma lem:support bound on E(Rj)}
=============================================
\[proof:support bound on E(Rj)\] By using sum of expectations, we can write ${\mathbb{E}}(R_j)$ in the following form: $$\begin{aligned}
\begin{split}
{\mathbb{E}}(R_j) &= \frac{1}{g} \sum_{i=1}^g {\mathbb{E}}(R_{ij}) \\
&= \frac{1}{g} \sum_{i=1}^g {\mathbb{P}}({\hat{w}}_{ij} \ne 0) \\
&= \frac{1}{g} \sum_{i=1}^g {\mathbb{P}}( \frac{1}{{\hat{\sigma}}_{ij}} {\text{sign}}({\hat{\alpha}}_{ij}) \max (0, |{\hat{\alpha}}_{ij}| - \lambda) \ne 0) \\
&= \frac{1}{g} \sum_{i=1}^g {\mathbb{P}}({\text{sign}}({\hat{\alpha}}_{ij}) \ne 0 \wedge |{\hat{\alpha}}_{ij}| > \lambda) \\
&= \frac{1}{g} \sum_{i=1}^g {\mathbb{P}}( {\hat{\alpha}}_{ij} \ne 0 \mid |{\hat{\alpha}}_{ij}| > \lambda ) {\mathbb{P}}( |{\hat{\alpha}}_{ij}| > \lambda ) \\
&= \frac{1}{g} \sum_{i=1}^g {\mathbb{P}}( |{\hat{\alpha}}_{ij}| > \lambda )
\end{split}
\end{aligned}$$ The last equality follows since for $\lambda > 0$, we have ${\mathbb{P}}( {\hat{\alpha}}_{ij} \ne 0 \mid |{\hat{\alpha}}_{ij}| > \lambda ) = 1$. Now, we put a bound on the term ${\mathbb{P}}( |{\hat{\alpha}}_{ij}| > \lambda )$. $$\begin{aligned}
\begin{split}
{\mathbb{P}}( |{\hat{\alpha}}_{ij}| > \lambda ) &= {\mathbb{P}}( |\frac{1}{n_i} \sum_{t=1}^{n_i} y_i^t X_{ij}^t| > \lambda ) \\
& \text{Expanding $y_i^t$ using equation \eqref{eq:generative model}} \\
&= {\mathbb{P}}( |\frac{1}{n_i} \sum_{t=1}^{n_i} ((X_i^t)^{\intercal}w^* + e_i^t) X_{ij}^t| > \lambda ) \\
&= {\mathbb{P}}( |\frac{1}{n_i} \sum_{t=1}^{n_i} (\sum_{k=1}^d X_{ik}^t w^*_k + e_i^t) X_{ij}^t| > \lambda ) \\
&= {\mathbb{P}}( |\frac{1}{n_i} \sum_{t=1}^{n_i} ((X_{ij}^t)^2 w_j^* + \sum_{k \in S^*, k \ne j} X_{ik}^t X_{ij}^t w^*_k + e_i^t X_{ij}^t)| > \lambda ) \\
&= {\mathbb{P}}( |\frac{1}{n_i} \sum_{t=1}^{n_i} (X_{ij}^t)^2 w_j^* + \sum_{k \in S^*, k \ne j} \frac{1}{n_i} \sum_{t=1}^{n_i} X_{ik}^t X_{ij}^t w^*_k + \frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t| > \lambda )
\end{split}
\end{aligned}$$ Recall that in this setting the covariance matrix for $X_i$ is $\Sigma^i$ with diagonal entries $\Sigma^i_{jj} \equiv {\sigma^i_{jj}}^2, \forall j \in {[d]}$ and non-diagonal entries $\Sigma^i_{jk} \equiv 0, \forall j,k \in {[d]}, j \ne k$. Let $D_{ij} \triangleq w_j^* \sigma_{jj}^i$. Then, $$\begin{aligned}
\begin{split}
{\mathbb{P}}( |{\hat{\alpha}}_{ij}| > \lambda ) &= {\mathbb{P}}( |w_j^* \frac{1}{n_i} \sum_{t=1}^{n_i} ((X_{ij}^t)^2 - {\sigma^i_{jj}}^2) + \sum_{k \in S^*, k \ne j} w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} X_{ik}^t X_{ij}^t + \frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t + D_{ij}| > \lambda )
\end{split}
\end{aligned}$$ Using the reverse triangle inequality $|a + b| \geq |a| - |b|$ recursively, $$\begin{aligned}
\begin{split}
&\geq {\mathbb{P}}( |D_{ij}| - |w_j^* \frac{1}{n_i} \sum_{t=1}^{n_i} ((X_{ij}^t)^2 - {\sigma^i_{jj}}^2)| - \sum_{k \in S^*, k \ne j} |w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} X_{ik}^t X_{ij}^t| - | \frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t| > \lambda ) \\
&\geq {\mathbb{P}}( |D_{ij}| > \lambda + \delta_j + \delta_1 + \delta_e \wedge |w_j^* \frac{1}{n_i} \sum_{t=1}^{n_i} ((X_{ij}^t)^2 - {\sigma^i_{jj}}^2)|< \delta_j \wedge | \sum_{k \in S^*, k \ne j} w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} (X_{ik}^t X_{ij}^t )|\\
&< \delta_1 \wedge | \frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t| < \delta_e ) \\
&\geq 1 - {\mathbb{P}}(|D_{ij}| \leq \lambda + \delta_j + \delta_1 + \delta_e) - {\mathbb{P}}(|w_j^* \frac{1}{n_i} \sum_{t=1}^{n_i} ((X_{ij}^t)^2 - {\sigma^i_{jj}}^2)|\geq \delta_j ) - \\
& {\mathbb{P}}( |\sum_{k \in S^*, k \ne j} w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} X_{ik}^t X_{ij}^t | \geq \delta_1 ) - {\mathbb{P}}(| \frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t| \geq \delta_e )
\end{split}
\end{aligned}$$
The following Lemmas \[lem:bound x\^2 term\], \[lem:uncorrelated wkXikXij\] and \[lem:error term\] provide concentration bounds for separate terms above.
\[lem:bound x\^2 term\] For $0 \leq \delta_j \leq 8 |w_j^*| \rho_i^2, \forall i \in {[g]}, \forall j \in S^*$, $$\begin{aligned}
\begin{split}
{\mathbb{P}}(|w_j^* \frac{1}{n_i} \sum_{t=1}^{n_i} ((X_{ij}^t)^2 - {\sigma^i_{jj}}^2)|\geq \delta_j) \leq 2 \exp( -\frac{n_i(\frac{\delta_j}{|w_j^*|\rho_i^2} )^2}{64} )
\end{split}
\end{aligned}$$
We take $\delta_j = 8 |w_j^*| \rho_i^2 \delta$, then $$\begin{aligned}
\begin{split}
{\mathbb{P}}(|\frac{1}{n_i} \sum_{t=1}^{n_i} ((\frac{X_{ij}^t}{\rho_i})^2 - (\frac{\sigma^i_{jj}}{\rho_i})^2)|\geq \frac{\delta_j}{|w_j^*|\rho_i^2} ) \leq 2\exp(-n_i\delta^2), \quad \forall 0 \leq \delta \leq 1
\end{split}
\end{aligned}$$
\[lem:uncorrelated wkXikXij\] For $0 \leq \delta_1 \leq 8 \rho_i^2 \sqrt{\sum_{k \in S^*, k \ne j} {w_k^*}^2}, \forall i \in {[g]}, \forall j \in S^*$, if predictors are mutually independent of each other then $$\begin{aligned}
\begin{split}
{\mathbb{P}}( |\sum_{k \in S^*, k \ne j} w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} X_{ik}^t X_{ij}^t | \geq \delta_1 ) \leq 2 \exp(- \frac{ n_i (\frac{\delta_1}{\rho_i^2 \sqrt{\sum_{k\in S^*, k \ne j} {w_k^*}^2} })^2 }{64})
\end{split}
\end{aligned}$$
We take $\delta_1 = 8 \delta \rho_i^2 \sqrt{\sum_{k\in S^*, k \ne j} {w_k^*}^2}$, which gives us $$\begin{aligned}
\begin{split}
{\mathbb{P}}( |\sum_{k \in S^*, k \ne j} w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} X_{ik}^t X_{ij}^t | \geq \delta_1 ) \leq 2 \exp(-n_i \delta^2), \quad \forall 0 \leq \delta \leq 1
\end{split}
\end{aligned}$$
\[lem:error term\] For $0 \leq \delta_e \leq 8 |\eta_i \rho_i|, \forall i \in {[g]}, \forall j \in S^*$, $$\begin{aligned}
\begin{split}
{\mathbb{P}}(| \frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t| \geq \delta_e ) \leq 2 \exp(- \frac{n_i (\frac{\delta_e}{|\eta_i\rho_i|})^2}{64})
\end{split}
\end{aligned}$$
We take $\delta_e = 8 |\eta_i \rho_i| \delta$, then $$\begin{aligned}
\begin{split}
{\mathbb{P}}(| \frac{1}{n_i} \sum_{t=1}^{n_i} \frac{e_i^t}{\eta_i} \frac{X_{ij}^t}{\rho_i}| \geq \frac{\delta_e}{|\eta_i\rho_i|} ) \leq 2 \exp(- n_i\delta^2), \quad \forall 0 \leq \delta \leq 1
\end{split}
\end{aligned}$$
By making the appropriate substitutions for $\delta_j, \delta_1$ and $\delta_e$ and noting that $|w_j^* {\sigma_{jj}^i}^2| > \lambda + 8 |w_j^*| \rho_i^2 \delta + 8 \rho_i^2 \sqrt{\sum_{k\in S^*, k \ne j} {w_k^*}^2} \delta + 8 |\eta_i\rho_i| \delta$, we can write: $$\begin{aligned}
\begin{split}
{\mathbb{E}}(R_j) &\geq 1 - \frac{6}{g} \sum_{i=1}^g \exp(-n_i\delta^2)
\end{split}
\end{aligned}$$
It follows that if we take $n_i > \frac{1}{\delta^2} \log 12$ then ${\mathbb{E}}(R_j) > \frac{1}{2}$.
Proof of Lemma \[lem:bound x\^2 term\] {#sec:proof of lemma bound x^2 term}
======================================
Observe that, $$\begin{aligned}
\begin{split}
{\mathbb{P}}(|w_j^* \frac{1}{n_i} \sum_{t=1}^{n_i} ((X_{ij}^t)^2 - {\sigma^i_{jj}}^2)|\geq \delta_j ) &= {\mathbb{P}}(|\frac{1}{n_i} \sum_{t=1}^{n_i} ((X_{ij}^t)^2 - {\sigma^i_{jj}}^2)|\geq \frac{\delta_j}{|w_j^*|} ) \\
&= {\mathbb{P}}(|\frac{1}{n_i} \sum_{t=1}^{n_i} ((\frac{X_{ij}^t}{\rho_i})^2 - (\frac{\sigma^i_{jj}}{\rho_i})^2)|\geq \frac{\delta_j}{|w_j^*|\rho_i^2} )
\end{split}
\end{aligned}$$ Note that $\frac{X_{ij}}{\rho_i}$ is a zero mean sub-Gaussian random variable with a variance proxy $1$. This implies that $(\frac{X_{ij}}{\rho_i})^2$ is a sub-exponential random variable with parameters $(4\sqrt{2}, 4)$. Thus, using concentration bounds for sub-exponential random variables[@wainwright2015chapter], we can write: $$\begin{aligned}
\begin{split}
{\mathbb{P}}(|\frac{1}{n_i} \sum_{t=1}^{n_i} ((\frac{X_{ij}^t}{\rho_i})^2 - (\frac{\sigma^i_{jj}}{\rho_i})^2)|\geq \frac{\delta_j}{|w_j^*|\rho_i^2} ) \leq 2\exp(-\frac{n_i(\frac{\delta_j}{|w_j^*|\rho_i^2} )^2}{64}), \quad \forall 0 \leq \frac{\delta_j}{|w_j^*|\rho_i^2} \leq 8
\end{split}
\end{aligned}$$
Proof of Lemma \[lem:uncorrelated wkXikXij\] {#sec:proof of lem:uncorrelated wkXikXij}
============================================
\[proof:uncorrelated wkXikXij\] Note that, $$\begin{aligned}
\begin{split}
\sum_{k\in S^*, k \ne j} w_k^* \frac{1}{n_i} \sum_{t=1}^{n_i} X_{ik}^t X_{ij}^t = \frac{1}{n_i} \sum_{t=1}^{n_i} (\sum_{k\in S^*, k\ne j} w_k^* X_{ik}^t) X_{ij}^t
\end{split}
\end{aligned}$$ Here $X_{ik}^t$ is a zero mean sub-Gaussian random variable with variance proxy $\rho_i$ and $(\sum_{k\in S^*, k\ne j} w_k^* X_{ik}^t)$ is a zero mean sub-Gaussian random variable with variance proxy $\sqrt{\sum_{k\in S^*, k \ne j} {w_k^*}^2} \rho_i $. Thus, $$\begin{aligned}
\begin{split}
{\mathbb{P}}( |\sum_{k \in S^*, k \ne j} w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} X_{ik}^t X_{ij}^t | \geq \delta_1 ) &= {\mathbb{P}}( |\frac{1}{n_i} \sum_{t=1}^{n_i} (\sum_{k\in S^*, k\ne j} w_k^* X_{ik}^t) X_{ij}^t | \geq \delta_1 ) \\
&= {\mathbb{P}}( |\frac{1}{n_i} \sum_{t=1}^{n_i} \frac{(\sum_{k\in S^*, k\ne j} w_k^* X_{ik}^t)}{\sqrt{\sum_{k\in S^*, k \ne j} {w_k^*}^2} \rho_i} \frac{X_{ij}^t}{\rho_i} | \geq \frac{\delta_1}{\rho_i^2 \sqrt{\sum_{k\in S^*, k \ne j} {w_k^*}^2} } )
\end{split}
\end{aligned}$$ where $\frac{(\sum_{k\in S^*, k\ne j} w_k^* X_{ik}^t)}{\sqrt{\sum_{k\in S^*, k \ne j} {w_k^*}^2} \rho_i}$ and $\frac{X_{ij}^t}{\rho_i}$ are zero mean sub-Gaussian random variables with unit variance proxy. Thus their product is sub-exponential random variable with parameters $(4\sqrt{2}, 4)$. Using concentration bound for sub-exponential random variables [@wainwright2015chapter], we can write: $$\begin{aligned}
\begin{split}
&{\mathbb{P}}( |\frac{1}{n_i} \sum_{t=1}^{n_i} \frac{(\sum_{k\in S^*, k\ne j} w_k^* X_{ik}^t)}{\sqrt{\sum_{k\in S^*, k \ne j} {w_k^*}^2} \rho_i} \frac{X_{ij}^t}{\rho_i} | \geq \frac{\delta_1}{\rho_i^2 \sqrt{\sum_{k\in S^*, k \ne j} {w_k^*}^2} } ) \leq 2 \exp(- \frac{ n_i (\frac{\delta_1}{\rho_i^2 \sqrt{\sum_{k\in S^*, k \ne j} {w_k^*}^2} })^2 }{64}) \\
&\forall 0 \leq \frac{\delta_1}{\rho_i^2 \sqrt{\sum_{k\in S^*, k \ne j} {w_k^*}^2} } \leq 8
\end{split}
\end{aligned}$$
Proof of Lemma \[lem:error term\] {#sec:proof of lemma error term}
=================================
\[proof:error term\] Note that, $$\begin{aligned}
\begin{split}
{\mathbb{P}}(| \frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t| \geq \delta_e ) &= {\mathbb{P}}(| \frac{1}{n_i} \sum_{t=1}^{n_i} \frac{e_i^t}{\eta_i} \frac{X_{ij}^t}{\rho_i}| \geq \frac{\delta_e}{|\eta_i\rho_i|} )
\end{split}
\end{aligned}$$ Again $\frac{X_{ij}}{\rho_i}$ and $\frac{e_i}{\eta_i}$ are mutually independent zero mean sub-Gaussian random variables with a variance proxy $1$. Thus, $\frac{e_i^t}{\eta_i} \frac{X_{ij}^t}{\rho_i}$ is a sub-exponential random variable with parameters $(4\sqrt{2}, 4)$. Thus, using concentration bounds for sub-exponential random variables[@wainwright2015chapter], we can write: $$\begin{aligned}
\begin{split}
{\mathbb{P}}(| \frac{1}{n_i} \sum_{t=1}^{n_i} \frac{e_i^t}{\eta_i} \frac{X_{ij}^t}{\rho_i}| \geq \frac{\delta_e}{|\eta_i\rho_i|} ) \leq 2 \exp(- \frac{n_i (\frac{\delta_e}{|\eta_i\rho_i|})^2}{64}), \quad \forall 0 \leq \frac{\delta_e}{|\eta_i\rho_i|} \leq 8
\end{split}
\end{aligned}$$
Proof of Lemma \[lem:nonsupport mcdiarmid\] {#sec:proof of lemma nonsupport mcdiarmid}
===========================================
\[proof:nonsupport mcdiarmid\] We again note that, $$\begin{aligned}
\begin{split}
{\mathbb{P}}(R_j \geq \frac{1}{2}) &= {\mathbb{P}}(R_j - {\mathbb{E}}(R_j) \geq \frac{1}{2} - {\mathbb{E}}(R_j)) \\
&\leq {\mathbb{P}}(|R_j - {\mathbb{E}}(R_j)| \geq \frac{1}{2} - {\mathbb{E}}(R_j))
\end{split}
\end{aligned}$$ The last inequality holds as long as ${\mathbb{E}}(R_j) < \frac{1}{2}$. Again, by noting that for a fixed $w^*$, $R_j = f(R_{1j}, R_{2j}, \cdots, R_{gj})$ such that $$\begin{aligned}
\begin{split}
(\forall k \in {[g]}),\quad \sup_{R_{1j}, \cdots, R_{kj}, R_{kj}',\cdots, R_{gj}}|f(R_{1j}, \cdots, R_{kj},\cdots, R_{gj}) - f(R_{1j}, \cdots, R_{kj}',\cdots, R_{gj})| \leq \frac{1}{g}
\end{split}
\end{aligned}$$ and using McDiarmid’s inequality [@mcdiarmid1989method], we can write the following: $$\begin{aligned}
\begin{split}
{\mathbb{P}}(|R_j - {\mathbb{E}}(R_j)| \geq \frac{1}{2} - {\mathbb{E}}(R_j)) \leq 2 \exp( - 2g (\frac{1}{2} - {\mathbb{E}}(R_j))^2)
\end{split}
\end{aligned}$$
Proof of Lemma \[lem:nonsupport bound E(Rj)\] {#sec:proof of lemma nonsupport bound E(Rj)}
=============================================
\[proof:nonsupport bound E(Rj)\] Like the proof of Lemma \[lem:support bound on E(Rj)\], we have the same formula for ${\mathbb{E}}(R_j)$ with the difference that $j \in S^*_c$, i.e., $$\begin{aligned}
\begin{split}
{\mathbb{E}}(R_j) = \frac{1}{g} \sum_{i=1}^g {\mathbb{P}}( |{\hat{\alpha}}_{ij}| > \lambda )
\end{split}
\end{aligned}$$ This time, we will put an upper bound on ${\mathbb{P}}( |{\hat{\alpha}}_{ij}| > \lambda )$. $$\begin{aligned}
\begin{split}
{\mathbb{P}}( |{\hat{\alpha}}_{ij}| > \lambda ) &= {\mathbb{P}}( |\frac{1}{n_i} \sum_{t=1}^{n_i} y_i^t X_{ij}^t| > \lambda ) \\
& \text{Expanding $y_i^t$ using equation \eqref{eq:generative model}} \\
&= {\mathbb{P}}( |\frac{1}{n_i} \sum_{t=1}^{n_i} ((X_i^t)^{\intercal}w^* + e_i^t) X_{ij}^t| > \lambda ) \\
&= {\mathbb{P}}( |\frac{1}{n_i} \sum_{t=1}^{n_i} (\sum_{k=1}^d X_{ik}^t w^*_k + e_i^t) X_{ij}^t| > \lambda ) \\
&= {\mathbb{P}}( |\frac{1}{n_i}(\sum_{k \in S^*} X_{ik}^t X_{ij}^t w^*_k + e_i^t X_{ij}^t)| > \lambda ) \\
&= {\mathbb{P}}( | \sum_{k \in S^*} \frac{1}{n_i} \sum_{t=1}^{n_i} X_{ik}^t X_{ij}^t w^*_k + \frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t| > \lambda )
\end{split}
\end{aligned}$$ Using the triangle inequality $|a+b| \leq |a| + |b|$, we can rewrite the above equation as: $$\begin{aligned}
\begin{split}
{\mathbb{P}}( |{\hat{\alpha}}_{ij}| > \lambda ) &\leq {\mathbb{P}}( |\sum_{k \in S^*} w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} X_{ik}^t X_{ij}^t | + |\frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t| > \lambda ) \\
&= 1 - {\mathbb{P}}( | \sum_{k \in S^*} w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} X_{ik}^t X_{ij}^t | + |\frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t| \leq \lambda ) \\
&\leq 1 - {\mathbb{P}}( | \sum_{k \in S^*} w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} X_{ik}^t X_{ij}^t | \leq \delta_1 \wedge |\frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t| \leq \delta_e \wedge 0 \leq \lambda - \delta_1 - \delta_e ) \\
&\leq {\mathbb{P}}(|\sum_{k \in S^*} w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} X_{ik}^t X_{ij}^t | \geq \delta_1 ) + {\mathbb{P}}(|\frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t| \geq \delta_e) + {\mathbb{P}}(0 \geq \lambda - \delta_1 - \delta_e)
\end{split}
\end{aligned}$$
The following Lemmas \[lem:nonsupport uncorrelated wkXikXij\] and \[lem:nonsupport error term\] provide concentration bounds for the separate terms above.
\[lem:nonsupport uncorrelated wkXikXij\] For $0 \leq \delta_1 \leq 8 \rho_i^2 \sqrt{\sum_{k \in S^*} {w_k^*}^2}, \forall i \in {[g]}, \forall j \in S^*_c$, if predictors are mutually independent of each other then $$\begin{aligned}
\begin{split}
{\mathbb{P}}( |\sum_{k \in S^*} w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} X_{ik}^t X_{ij}^t | \geq \delta_1 ) \leq 2 \exp(- \frac{ n_i (\frac{\delta_1}{\rho_i^2 \sqrt{\sum_{k\in S^*} {w_k^*}^2} })^2 }{64})
\end{split}
\end{aligned}$$
We take $\delta_1 = 8 \rho_i^2 \sqrt{\sum_{k \in S^*} {w_k^*}^2} \delta$ for $0 < \delta < 1$. Then, $$\begin{aligned}
\begin{split}
{\mathbb{P}}( |\sum_{k \in S^*} w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} X_{ik}^t X_{ij}^t | \geq \delta_1 ) \leq 2 \exp(-n_i \delta^2)
\end{split}
\end{aligned}$$
\[lem:nonsupport error term\] For $0 \leq \delta_e \leq 8 |\eta_i \rho_i|, \forall i \in {[g]}, \forall j \in S^*_c$, $$\begin{aligned}
\begin{split}
{\mathbb{P}}(| \frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t| \geq \delta_e ) \leq 2 \exp(- \frac{n_i (\frac{\delta_e}{|\eta_i\rho_i|})^2}{64})
\end{split}
\end{aligned}$$
We take $\delta_e = 8 |\eta_i \rho_i| \delta$ for $0 < \delta < 1$, then $$\begin{aligned}
\begin{split}
{\mathbb{P}}(| \frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t| \geq \delta_e ) \leq 2 \exp(-n_i \delta^2)
\end{split}
\end{aligned}$$
Using results of Lemmas \[lem:nonsupport uncorrelated wkXikXij\] and \[lem:nonsupport error term\] and making the appropriate substitutions for $\delta_1$ and $\delta_e$ and noticing that $ \lambda > 8 \delta \rho_i^2 \sqrt{\sum_{k\in S^*} w_k^2} + 8 |\eta_i\rho_i|\delta$, we can write $$\begin{aligned}
\begin{split}
{\mathbb{E}}(R_j) \leq \frac{4}{g} \sum_{i=1}^g \exp(-n_i\delta^2)
\end{split}
\end{aligned}$$ Furthermore, if we have $n_i > \frac{1}{\delta^2} \log 8$ then ${\mathbb{E}}(R_j) < \frac{1}{2}$.
Proof of Lemma \[lem:nonsupport uncorrelated wkXikXij\] {#sec:proof of lemma nonsupport uncorrelated wkXikXij}
=======================================================
We follow the same procedure as the proof of Lemma \[lem:uncorrelated wkXikXij\].
Proof of Lemma \[lem:nonsupport error term\] {#sec:proof of lemma nonsupport error term}
============================================
We follow the same procedure as the proof of Lemma \[lem:error term\].
Proof of Lemma \[lem:support bound on E(Rj) correlated\] {#sec:proof of lemma support bound on E(Rj) correlated}
========================================================
\[proof:support bound on E(Rj) correlated\] We know from the proof of Lemma \[lem:support bound on E(Rj)\] that $$\begin{aligned}
\begin{split}
{\mathbb{E}}(R_j) = \frac{1}{g} \sum_{i=1}^g {\mathbb{P}}(|{\hat{\alpha}}_{ij}| > \lambda)
\end{split}
\end{aligned}$$ and $$\begin{aligned}
\begin{split}
{\mathbb{P}}( |{\hat{\alpha}}_{ij}| > \lambda ) &= {\mathbb{P}}( |\frac{1}{n_i} \sum_{t=1}^{n_i} (X_{ij}^t)^2 w_j^* + \sum_{k \in S^*, k \ne j} \frac{1}{n_i} \sum_{t=1}^{n_i} X_{ik}^t X_{ij}^t w^*_k + \frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t| > \lambda )
\end{split}
\end{aligned}$$ Let $D_{ij} = w^*_j{\mathbb{E}}(X_{ij}^2) + \sum_{k\in S^*, k\ne j} w^*_k {\mathbb{E}}(X_{ik}X_{ij}) = w^*_j{\sigma^i_{jj}}^2 + \sum_{k\in S^*, k\ne j} w^*_k \sigma^i_{jk}$. Then, $$\begin{aligned}
\begin{split}
{\mathbb{P}}( |{\hat{\alpha}}_{ij}| > \lambda ) &= {\mathbb{P}}( |w_j^* \frac{1}{n_i} \sum_{t=1}^{n_i} ((X_{ij}^t)^2 - {\sigma^i_{jj}}^2) + \sum_{k \in S^*, k \ne j} w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} (X_{ik}^t X_{ij}^t - \sigma^i_{jk} )\\
&+ \frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t + D_{ij}| > \lambda )
\end{split}
\end{aligned}$$ Using the reverse triangle inequality $|a + b| \geq |a| - |b|$ recursively, $$\begin{aligned}
\begin{split}
&\geq {\mathbb{P}}( |D_{ij}| - |w_j^* \frac{1}{n_i} \sum_{t=1}^{n_i} ((X_{ij}^t)^2 - {\sigma^i_{jj}}^2)| - \sum_{k \in S^*, k \ne j} |w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} (X_{ik}^t X_{ij}^t - \sigma^i_{jk} )| - | \frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t| > \lambda ) \\
&\geq {\mathbb{P}}( |D_{ij}| > \lambda + \delta_j + \sum_{k \in S^*, k \ne j} \delta_k + \delta_e \wedge |w_j^* \frac{1}{n_i} \sum_{t=1}^{n_i} ((X_{ij}^t)^2 - {\sigma^i_{jj}}^2)|< \delta_j \wedge (\forall k \in S^*, k \ne j)\\
& |w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} (X_{ik}^t X_{ij}^t - \sigma^i_{jk} )|< \delta_k \wedge | \frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t| < \delta_e ) \\
&\geq 1 - {\mathbb{P}}(|D_{ij}| \leq \lambda + \delta_j + \sum_{k \in S^*, k \ne j} \delta_k + \delta_e) - {\mathbb{P}}(|w_j^* \frac{1}{n_i} \sum_{t=1}^{n_i} ((X_{ij}^t)^2 - {\sigma^i_{jj}}^2)|\geq \delta_j ) - \\
& \sum_{k \in S^*, k \ne j} {\mathbb{P}}( |w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} (X_{ik}^t X_{ij}^t - \sigma^i_{jk} )| \geq \delta_k ) - {\mathbb{P}}(| \frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t| \geq \delta_e )
\end{split}
\end{aligned}$$ Now we focus on concentration bounds for individual terms. Note that we already have concentration bounds for ${\mathbb{P}}(|w_j^* \frac{1}{n_i} \sum_{t=1}^{n_i} ((X_{ij}^t)^2 - {\sigma^i_{jj}}^2)|\geq \delta_j ) $ and ${\mathbb{P}}(| \frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t| \geq \delta_e ) $ from Lemmas \[lem:bound x\^2 term\] and \[lem:error term\] respectively. The following lemma provides concentration bound for ${\mathbb{P}}( |w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} (X_{ik}^t X_{ij}^t - \sigma^i_{jk} )| \geq \delta_k )$.
\[lem:support correlated wkXikXij\] For $0 \leq \delta_k \leq 8 \sqrt{2} |w_k^*| (1 + 4 \max_j \frac{\rho_i^2}{{\sigma^i_{jj}}^2}) \max_j {\sigma^i_{jj}}^2, \forall i \in {[g]}, \forall j \in S^*$, then $$\begin{aligned}
\begin{split}
{\mathbb{P}}( |w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} (X_{ik}^t X_{ij}^t - \sigma^i_{jk} )| \geq \delta_k ) \leq 4 \exp( - \frac{n_i (\frac{\delta_k}{|w^*_k|})^2}{128(1 + 4 \max_j \frac{\rho_i^2}{{\sigma^i_{jj}}^2})^2 \max_j{\sigma^i_{jj}}^4})
\end{split}
\end{aligned}$$
We take $\delta_k = 8 \sqrt{2} |w_k^*| (1 + 4 \max_j \frac{\rho_i^2}{{\sigma^i_{jj}}^2}) \max_j {\sigma^i_{jj}}^2 \delta$, then $$\begin{aligned}
\begin{split}
&{\mathbb{P}}( |\frac{1}{n_i} \sum_{t=1}^{n_i} (X_{ik}^t X_{ij}^t - \sigma^i_{jk} )| \geq \frac{\delta_k}{|w^*_k|} ) \leq 4 \exp( - n_i \delta^2), \forall 0 \leq \delta \leq \frac{1}{\sqrt{2}}
\end{split}
\end{aligned}$$ Also taking $\delta_j =8 |w_j^*| \rho_i^2 \delta$ and $\delta_e = 8 |\eta_i \rho_i| \delta $ in Lemmas \[lem:bound x\^2 term\] and \[lem:error term\] respectively and noting that $$\begin{aligned}
\begin{split}
&0 < \lambda < |(w^*_j{\sigma^i_{jj}}^2 + \sum_{k\in S^*, k\ne j} w^*_k \sigma^i_{jk})| - 8 |w_j^*| \rho_i^2 \delta - \sum_{k\in S^*, k\ne j} 8 \sqrt{2} |w_k^*| (1 + 4 \max_j \frac{\rho_i^2}{{\sigma^i_{jj}}^2}) \\
&\max_j {\sigma^i_{jj}}^2 \delta - 8 |\eta_i \rho_i| \delta \;,
\end{split}
\end{aligned}$$ we can write: $$\begin{aligned}
\begin{split}
{\mathbb{E}}(R_j) \geq 1 - \frac{4s}{g} \sum_{i=1}^g \exp(-n_i\delta^2 )
\end{split}
\end{aligned}$$ It follows that if we take $n_i > \frac{1}{\delta^2}\log 8s$, we get ${\mathbb{E}}(R_j) > \frac{1}{2}$.
Proof of Lemma \[lem:support correlated wkXikXij\] {#sec:proof of lemma support correlated wkXikXij}
==================================================
\[proof:support correlated wkXikXij\] Note that, $$\begin{aligned}
\begin{split}
{\mathbb{P}}( |w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} (X_{ik}^t X_{ij}^t - \sigma^i_{jk} )| \geq \delta_k ) &= {\mathbb{P}}( |\frac{1}{n_i} \sum_{t=1}^{n_i} (X_{ik}^t X_{ij}^t - \sigma^i_{jk} )| \geq \frac{\delta_k}{|w^*_k|} )
\end{split}
\end{aligned}$$ Using Lemma 1 from [@ravikumar2011high], we can write $$\begin{aligned}
\begin{split}
&{\mathbb{P}}( |\frac{1}{n_i} \sum_{t=1}^{n_i} (X_{ik}^t X_{ij}^t - \sigma^i_{jk} )| \geq \frac{\delta_k}{|w^*_k|} ) \leq 4 \exp( - \frac{n_i (\frac{\delta_k}{|w^*_k|})^2}{128(1 + 4 \max_j \frac{\rho_i^2}{{\sigma^i_{jj}}^2})^2 \max_j{\sigma^i_{jj}}^4}), \\
& \forall 0 \leq \frac{\delta_k}{|w^*_k|} \leq 8\max_j {\sigma^i_{jj}}^2(1 + 4 \max_j \frac{\rho_i^2}{{\sigma^i_{jj}}^2})
\end{split}
\end{aligned}$$
Proof of Lemma \[lem:nonsupport bound E(Rj) correlated\] {#sec:proof of lemma nonsupport bound E(Rj) correlated}
========================================================
\[proof:nonsupport bound E(Rj) correlated\] Like before, $$\begin{aligned}
\begin{split}
{\mathbb{E}}(R_j) = \frac{1}{g} \sum_{i=1}^g {\mathbb{P}}( |{\hat{\alpha}}_{ij}| > \lambda )
\end{split}
\end{aligned}$$ and $$\begin{aligned}
\begin{split}
{\mathbb{P}}( |{\hat{\alpha}}_{ij}| > \lambda ) = {\mathbb{P}}( | \sum_{k \in S^*} \frac{1}{n_i} \sum_{t=1}^{n_i} X_{ik}^t X_{ij}^t w^*_k + \frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t| > \lambda )
\end{split}
\end{aligned}$$ Let $D_{ij} = \sum_{k\in S^*} w^*_k {\mathbb{E}}(X_{ik}X_{ij}) = \sum_{k\in S^*} w^*_k \sigma^i_{jk}$. Then, $$\begin{aligned}
\begin{split}
{\mathbb{P}}( |{\hat{\alpha}}_{ij}| > \lambda ) &= {\mathbb{P}}( |\sum_{k \in S^*} w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} (X_{ik}^t X_{ij}^t - \sigma^i_{jk} ) + \frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t + D_{ij}| > \lambda )
\end{split}
\end{aligned}$$ Using the triangle inequality $|a+b| \leq |a| + |b|$, $$\begin{aligned}
\begin{split}
{\mathbb{P}}( |{\hat{\alpha}}_{ij}| > \lambda ) &\leq {\mathbb{P}}( \sum_{k \in S^*} |w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} (X_{ik}^t X_{ij}^t - \sigma^i_{jk} )| + |\frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t| + |D_{ij}| > \lambda ) \\
&= 1 - {\mathbb{P}}( \sum_{k \in S^*} |w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} (X_{ik}^t X_{ij}^t - \sigma^i_{jk} )| + |\frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t| + |D_{ij}| \leq \lambda ) \\
&\leq 1 - {\mathbb{P}}( (\forall k \in S^*) |w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} (X_{ik}^t X_{ij}^t - \sigma^i_{jk} )| \leq \delta_k \wedge |\frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t| \leq \delta_e \wedge \\
& |D_{ij}| \leq \lambda - \sum_{k \in S^*} \delta_k - \delta_e ) \\
&\leq \sum_{k \in S^*} {\mathbb{P}}(|w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} (X_{ik}^t X_{ij}^t - \sigma^i_{jk} )| \geq \delta_k ) + {\mathbb{P}}(|\frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t| \geq \delta_e) + \\
& {\mathbb{P}}(|D_{ij}| \geq \lambda - \sum_{k \in S^*} \delta_k - \delta_e)
\end{split}
\end{aligned}$$ Observe that for entries in the non-support, we already have a concentration bound on ${\mathbb{P}}(|\frac{1}{n_i} \sum_{t=1}^{n_i} e_i^t X_{ij}^t| \geq \delta_e)$ from Lemma \[lem:nonsupport error term\]. The following lemma provides concentration bound on ${\mathbb{P}}(|w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} (X_{ik}^t X_{ij}^t - \sigma^i_{jk} )| \geq \delta_k )$.
\[lem:nonsupport correlated wkXikXij\] For $0 \leq \delta_k \leq 8 \sqrt{2} |w_k^*| (1 + 4 \max_j \frac{\rho_i^2}{{\sigma^i_{jj}}^2}) \max_j {\sigma^i_{jj}}^2, \forall i \in {[g]}, \forall j \in S^*_c$, then $$\begin{aligned}
\begin{split}
{\mathbb{P}}( |w^*_k \frac{1}{n_i} \sum_{t=1}^{n_i} (X_{ik}^t X_{ij}^t - \sigma^i_{jk} )| \geq \delta_k ) \leq 4 \exp( - \frac{n_i (\frac{\delta_k}{|w^*_k|})^2}{128(1 + 4 \max_j \frac{\rho_i^2}{{\sigma^i_{jj}}^2})^2 \max_j{\sigma^i_{jj}}^4})
\end{split}
\end{aligned}$$
We take $\delta_k = 8 \sqrt{2} |w_k^*| (1 + 4 \max_j \frac{\rho_i^2}{{\sigma^i_{jj}}^2}) \max_j {\sigma^i_{jj}}^2 \delta$, then $$\begin{aligned}
\begin{split}
&{\mathbb{P}}( |\frac{1}{n_i} \sum_{t=1}^{n_i} (X_{ik}^t X_{ij}^t - \sigma^i_{jk} )| \geq \frac{\delta_k}{|w^*_k|} ) \leq 4 \exp( - n_i\delta^2), \forall 0 \leq \delta \leq \frac{1}{\sqrt{2}}
\end{split}
\end{aligned}$$ Taking $\delta_e = 8 |\eta_i \rho_i| \delta$ and noticing that $$\begin{aligned}
\begin{split}
\lambda > |\sum_{k\in S^*} w^*_k \sigma^i_{jk}| + \sum_{k\in S^*} 8 \sqrt{2} |w_k^*| (1 + 4 \max_j \frac{\rho_i^2}{{\sigma^i_{jj}}^2}) \max_j {\sigma^i_{jj}}^2 \delta + 8 |\eta_i \rho_i| \delta \;,
\end{split}
\end{aligned}$$ we can write $$\begin{aligned}
\begin{split}
{\mathbb{E}}(R_j) \leq &\frac{4s + 2}{g} \sum_{i=1}^g ( \exp(-n_i \delta^2) )
\end{split}
\end{aligned}$$ It also follows that if we have $n_i > \frac{1}{\delta^2}\log(8s + 4) $ then ${\mathbb{E}}(R_j) < \frac{1}{2}$.
Proof of Lemma \[lem:nonsupport correlated wkXikXij\] {#sec:proof of lemma nonsupport correlated wkXikXij}
=====================================================
\[proof:nonsupport correlated wkXikXij\] The proof follows the same procedure as the proof of Lemma \[lem:support correlated wkXikXij\].
Real World Experiment {#sec:real world application}
=====================
In this section, we demonstrate the effectiveness of our method to determine the support of a sparse linear regression setup in a real world data set. We used the BlogFeedback data set[^1] from <https://archive.ics.uci.edu/ml/datasets/BlogFeedback>. This data set contains features extracted from blog posts and the task is to predict how many comments the post will receive using these features. The details about the data set are as follows:
- Number of training samples: $52397$
- Number of features (after removing all zeros columns): $276$
#### Constructing a True Support.
Since the true support for the parameter vector is unknown for real world data, we constructed a “centralized” support for comparison by running LASSO on the complete training data set. This simulates a centralized server with access to all the data. The “centralized” support contains $40$ elements. Our goal is to recover this “centralized” support by running our method in a federated setting.
#### Performance Measures.
We measure the performance of our method by measuring recall, precision and $F_1$-score with respect to our recovered “federated” support. These performance measures are defined as follows: $$\begin{aligned}
\begin{split}
\text{Recall} &= \frac{\text{Number of elements in the ``federated'' support that are in the ``centralized'' support}}{\text{Number of elements in the ``centralized'' support}} \\
\text{Precision} &= \frac{\text{Number of elements in the ``federated'' support that are in the ``centralized'' support}}{\text{Number of elements in the ``federated'' support}} \\
F_1\text{-score} &= \frac{2 \times \text{Recall} \times \text{Precision}}{\text{Recall} + \text{Precision}} \\
\end{split}
\end{aligned}$$
#### Case 1.
For the first experiment, we divided the dataset randomly into $523$ clients with each client containing $100$ samples (except the last one which contains more to account for imbalance). This is a highly distributed setting where each client has access to a small number of samples. We conducted our experiment using $\lambda = 0.08$. Our method recovered a support with 48 elements, with $39$ of them being in the “centralized” support. We report the following results:
- Recall: 97.5%
- Precision: 81.2%
- $F_1$-score: 88.7 %
#### Case 2.
For the second experiment, we divided the dataset randomly into $52$ clients with each client containing $1000$ samples (except the last one which contains more to account for imbalance). This is a setting where we have only a few clients. We again conducted our experiment using $\lambda = 0.08$. Our method recovered a support with 49 elements, which included all of the elements in the “centralized” support. We report the following results:
- Recall: 100%
- Precision: 81.6%
- $F_1$-score: 90 %
We see that in both cases, we recovered almost all of the elements in the “centralized” support while making very few mistakes.
[^1]: Buza, K. (2014). Feedback Prediction for Blogs. In Data Analysis, Machine Learning and Knowledge Discovery (pp. 145-152). Springer International Publishing.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The kinetics of the self-assembly of supramolecular polymers is dictated by how monomers, dimers, trimers etc., attach to and detach from each other. It is for this reasons that researchers have proposed a plethora of pathways to explain the kinetics of various self-assembling supramolecules, including sulfur, linear micelles, living polymers and protein fibrils. Recent observations hint at the importance of a hitherto ignored molecular aggregation pathway that we refer to as “*body evaporation and addition*”. In this pathway, monomers can enter at or dissociate from any point along the backbone of the polymer. In this paper, we compare predictions for the well-established *end evaporation and addition* pathway with those that we obtained for the newly proposed *body evaporation and addition* model. We quantify the lag time, characteristic of nucleated reversible polymerisation, in terms of the time it takes to obtain half of the steady-state polymerised fraction and the apparent growth rate at that point, and obtain power laws for both as a function of the total monomer concentration. We find, perhaps not entirely unexpectedly, that the *body evaporation and addition* pathway speeds up the relaxation of the polymerised monomeric mass relative to that of the *end evaporation and addition*. However, the presence of the *body evaporation and addition* pathway does not affect the dependence of the lag time on the total monomer concentration and it remains the same as that for the case of *end evaporation and addition*. The scaling of the lag time with the forward rate is different for the two models, suggesting that they may be distinguished experimentally.'
author:
- 'Nitin S. Tiwari'
- Paul van der Schoot
title: 'On the Kinetics of Body *versus* End Evaporation and Addition of Supramolecular Polymers'
---
Introduction {#sec1}
============
Supramolecular polymerisation processes are of immense importance in biology and in chemistry. [@tomdegreef] Some classic examples in biology include actin and microtubule self-assembly that play important roles in the context of the mechanics of the cell, and $\beta-$amyloid and prion protein aggregation implicated in neuro-degenerative diseases. [@takalo; @blanchoin; @fletcher; @chiti; @murphy] Similarly, supramolecular polymerisation has major applications in the chemistry of medicine and in molecular electronics. [@aida] In this light, it is not surprising that researchers have long studied the thermodynamic and kinetics of self-assembly. As the time evolution of self-assembly is very much system specific, in particular the early time kinetics, itself the most extensively studied aspect of reversible polymerisation, a whole host of molecular pathways of supramolecular self-assembly have been proposed. [@oosawa; @cates_all; @cohen]
It is generally believed that Oosawa was the first to suggest a model in the context of the polymerisation of actin filaments, where a monomer can be added to or removed from the ends. [@oosawa; @cohen] Oosawa’s model has one important ingredient, known as *nucleation*. This means that a stable critical nucleus of $n_c \ge 1$ monomers must be formed before polymer growth commences. Although Oosawa’s model of self-assembly is in agreement with experimental data in the context of actin polymerisation, it fails to describe many other protein aggregation processes. [@curve_fitting] Indeed, the prevalent molecular pathway for self-assembly is dictated by the complex molecular structure of the monomers involved, as well as by the type of bonding between monomers that form a polymer. This results in molecular pathways that are more complex than the simple pathway proposed by Oosawa, which is sometimes also referred to as *end evaporation and addition* [@dubbeldam; @cates_eea; @semenov_eea].
In the context of the living polymerisation or linear polymers, researchers proposed a plethora of pathways by which self-assembly can occur, e.g., polymer *scission and recombination* [@cates_sr1; @semenov_sr; @cates_sr], *secondary nucleation* [@cohen; @tuomas_review] and *two-stage nucleation*. [@knowles_two_stage] Further work shows that the kinetics of self-assembly is strongly dependent on which of the above mentioned pathways are active in the assembly process. [@curve_fitting] The influence molecular aggregation pathways have on the early time kinetics of linear self-assembly, which is the most studied aspect of the problem in hand, motivates researchers to study and characterize all possible pathways. [@tuomas_science; @cohen; @knowles_two_stage; @Tiwari_2016; @tuomas_review]
With the aim to probe the molecular pathway responsible for linear self-assembly in the context of supramolecular polymerisation, Albertazzi *et al.* recently performed experiments with a self-assembling molecule known as 1,3,5 benzenetricarboxamide or BTA for short. [@albertazzi] By imaging the supramolecular polymers at different assembly times, they were able to investigate monomer mixing on the scale of individual polymers. Because their observations could not be explained by any of the hitherto known molecular pathways, they suggested the need to revisit theoretically and experimentally the dynamic behaviour of supramolecular polymers. From their observations, they conclude that the molecular pathway responsible for self-assembly of BTAs is the one in which the monomer can be removed from and inserted anywhere along the polymer backbone.
We have schematically depicted the novel pathway in Fig. \[fig1\]. At time $t=0$ the solution contains only two types of supramolecular homopolymer, and as time progresses mixing of monomers occurs at the supramolecular level. However, contrary to crowding of differently labeled monomers at the ends, which is to be expected if monomers can only attach on and detach from the ends, the mixing takes place homogeneously along the polymer backbone. We call their proposed pathway the “*body evaporation and addition*” pathway to contrast it with the conventional *end evaporation and addition* pathway, and study theoretically the kinetics of this pathway in the presence of the *scission and recombination* pathway and *primary nucleation*. [@oosawa] The molecular pathway *scission and recombination* allows polymers to break at any point on the backbone resulting into two smaller polymers and recombining two polymers into a longer polymer. *Primary nucleation* is the mechanism by which a critical number of monomers spontaneously self-assemble to form the shortest stable polymer.[@oosawa] Many researchers have concluded that the time evolution of the length distribution of living polymers is typically a mixture of molecular pathways, which are switched on and off dependent on the system of interest. [@tuomas_science; @hong; @hong_moment_closure] It is for this reason that we study the kinetics of the newly proposed pathway in combination with the molecular pathways already referred to. At first glance, *body evaporation and addition* pathway looks similar to *end evaporation and addition*, and naively one would perhaps presume that a simple renormalisation of the rate constants can account for the former. However, a closer look at the problem reveals that for *end evaporation and addition* every polymer has only two ends, resulting into a probability of addition or removal of monomer at the ends that plausibly is independent of the length of the polymer. In contrast, in the case of *body evaporation and addition* the probability of adding or removing a monomer is proportional to the number of bonds in a polymer in which a monomer is being added. Hence, the addition or removal of a monomer along the polymer backbone in the *body evaporation and addition* depends on the size of that particular polymer. As the size of an individual polymer changes as a function of time, it is not possible to simply renormalise the rate constants associated with *body evaporation and addition* pathway and expect it to behave like *end evaporation and addition*.
To study the kinetics of linear self-assembly we start by writing the discrete rate equations. However, the rate equations are highly nonlinear and have so far eluded exact analytical solution except in a few limiting cases. [@cohen; @hong] Hence, we study the kinetics of self-assembly by closing the discrete reaction rate equations by insisting on plausible approximations. This way, we obtain dynamical equations for the first two moments of the polymer length distribution. These are the number of polymers and the polymerised monomeric mass, of which the latter is primarily probed in assembly experiments. [@tuomas_science; @hellstrand; @tuomas_review] We obtain asymptotic analytical solutions of the resulting dynamical equations. From our analytical solutions we quantify the early time kinetics and show that the scaling of the lag time with the total monomer concentration is identical to that of the standard *end evaporation and addition* pathway even in the presence of the proposed *body evaporation and addition*. However, the lag time significantly does decreases with increasing predominance of *body evaporation and addition* pathway. We also show that when only one of the two addition and evaporation pathways is present, the half-time and the apparent growth rate differ in their scaling with the forward rate constants of monomer insertion.
![Schematic showing the effect of *body evaporation and addition* pathway on the mixing and growth of differently dye-labeled but otherwise identical monomers at time $t=0$. Note the homogeneous insertion of black monomers along the green polymer backbone and *vice versa* via the free monomers in the solution.[]{data-label="fig1"}](Fig1.eps){width="6.5in"}
The remainder of this paper is organised as follows: In Section \[sec2\], we introduce the moment equations for the generalized molecular pathway and study the equilibrium properties of the two moments of the length distribution. In Section \[sec3\], the dynamical equations for the moments are solved in the absence of polymer recombination and in the limit of vanishing fragmentation and nucleation rate constants. The analytical solutions are then compared with numerical solutions, and used to calculate the half-time and the apparent growth rate for the polymerisation kinetics. Finally, in Section \[sec4\], we discuss the results and findings of our theoretical analysis.
Master Equations and Moment Equations {#sec2}
=====================================
The molecular pathways that we include in our study are i) *primary nucleation*, ii) *end evaporation and addition*, iii) *body evaporation and addition* and iv) *scission and recombination*. [@cates_all; @knowles_two_stage; @Tiwari_2016] Our moment equations, derived from the generalized rate equations, allow us to probe the quantitative role of the various molecular pathways involved. The molecular pathway of interest can in principle always be made dominant by switching off other pathways completely or asymptotically. In order to study the kinetics of nucleated polymerisation we first transform the molecular pathways into the corresponding reaction representation. $$\begin{aligned}
\label{eq1}
& \text{i) primary nucleation} \hspace{85pt} n_c x \xrightleftharpoons[k_n^-]{k_n^+} y_{n_c}, \\
& \text{ii) end evaporation and addition} \hspace{26pt} y_i + x \xrightleftharpoons[2 k_e^-]{2 k_e^+} y_{i+1}, \\
& \text{iii) body evaporation and addition} \hspace{19pt} y_i + x \xrightleftharpoons[k_b^-(i-1)]{k_b^+(i-1)} y_{i+1}, \\
& \text{iv) scission and recombination} \hspace{40pt} y_i + y_j \xrightleftharpoons[k_f^-]{k_f^+} y_{i+j} \quad \text{for} \quad i,j \le n_c,\end{aligned}$$ where $x$ and $y_i$ are the concentration of monomers and that of polymers of degree of polymerisation $i$, respectively, and $n_c$ is the critical nucleus size, i.e., the degree of polymerisation of smallest stable polymer. Furthermore, $k_n^+$, $k_n^-$, $k_e^+$, $k_e^-$, $k_b^+$, $k_b^-$, $k_f^+$ and $k_f^-$ are the rates of nucleus formation and disintegration (subscript n), monomer addition and removal from the ends (subscript e), the rates of monomer addition and removal from the polymer backbone excluding the ends (subscript b) and polymer recombination and scission (subscript f). The factor of $i-1$ in reaction iii), describing body evaporation and addition, accounts for the fact that a monomer can be added in $i-1$ places on the backbone a polymer of size $i$, and that any one of $i-1$ monomers can be removed from a polymer of size $i+1$ because removal from the ends is forbidden for this pathway.
The reactions are assumed to be reaction limited rather that diffusion limited, implying that the reaction rates are constant in time. The indices $i$ and $j$ for reaction iv) obey $i,j \ge n_c$, where $n_c \ge 2$. We consider the case $n_c \ge 2$ in order to be able to close the sums in the master equation and obtain the dynamical equations for the first two principal moments of the polymer length distribution. In the case $n_c \ge 2$, a monomer is not counted as a polymer, whereas for $n_c=1$, a monomer can be an active monomer, i.e., a polymer of size one, or an inactive monomer. Hence, the master equations for $n_c \ge 2$ are fundamentally different from that for $n_c=1$ polymerisation.
One additional assumption that we employ in order to close the sums in our master equations is that we assume $k_n^-=k_e^-$. Later we will see that this approximation does not alter our results because in order to be able to close our discrete master equation, we neglect disintegration of a nuclei via the *end evaporation and addition* pathway. Our approximation of irreversible nucleus formation has been employed in the past by several researchers and the results have been quantitatively compared with the experimental data on protein polymerisation, justifying our approximation. [@tuomas_science; @hong; @tuomas_review]
Before we delve deeper into our analysis, a few remarks should be made. In principle, we consider four mechanisms that are responsible for the time evolution of the length distribution: the *primary nucleation*, the *end evaporation and addition*, the *body evaporation and addition* and the *scission and recombination* pathway. Our goal in this work is very specific and is to compare the *early time* kinetics of the two pathways of interest, which are the *end evaporation and addition* and the *body evaporation and addition*. However, if we do not include the *scission and recombination* pathway, the resulting dynamical equations will be singular, meaning that we can make a parameter corresponding to this pathway small but never put it to zero. Hence, it is for purely mathematical reasons that we make use of the most general description that formally includes all the aforementioned pathways. This is also the reason why we implement *primary nucleation*. To make the problem mathematically tractable we work in the limit of strongly nucleated systems, where for the time domain of our interest the primary nucleation step is not functional and hence can be ignored for all practical purposes. This said, we will explain in detail all of our approximations and limitations, as and when they come in this paper.
To derive a closed form of the moment equations, we start with the discrete master equation for the reaction schemes defined above. For the polymers, this yields \[eq2\] &=& k\_n\^+ x(t)\^[n\_c]{} \_[i,n\_c]{} + 2 x(t) y\_[i-1]{}(t) - 2 x(t) y\_[i]{}(t) + 2 y\_[i+1]{}(t) - 2 y\_i(t)\
& & + (i-2) x(t) y\_[i-1]{} - (i-1) x(t) y\_[i]{} + (i-1) y\_[i+1]{} - (i-2) y\_i\
& & - (i- 2 n\_c +1) y\_i(t) + 2 \_[j=i+n\_c]{}\^ y\_j(t) + \_[k+l=i]{} y\_k(t) y\_l(t)\
& & - 2 y\_i (t) \_[j=n\_c]{}\^ y\_j, where the first term on the right-hand-side of Eq. \[eq2\] accounts for the formation of the critical nucleus, the next four terms stem from the *end evaporation and addition* pathway, and terms six to nine represent the *body evaporation and addition*. The last four terms result from *scission and recombination*. Here, $\delta_{i,n_c}$ denotes Kronecker delta that acquires value of 1 when $i=n_c$ and is zero otherwise. Notice that Eq. (5) is missing the term for nucleus disintegration. This is due to our approximation of $k_n^- = k_e^-$ that allows nucleus to disintegrate via monomer removal from an end hence the nucleus disintegration term gets absorbed in the end evaporation terms.
The factor of $(i- 2 n_c +1)$ in the tenth term on the right-hand side of Eq. \[eq2\] accounts for the number of bonds allowed to break such that the fragmenting filaments are larger than the nucleus size $n_c$. It should be mentioned that the same term, in principle, should include a factor of $\theta(i-2 n_c)$. However, this factor would prevent us from closing the summations and obtaining the dynamical equations for the first two moments of the full polymer length distribution. Hence, we choose not to include it in our master equation. As a consequence, an inconsistency arises for the dimers and trimers, at least if we focus on the case of $n_c=2$. Indeed, this choice would in principle allow a dimer and a trimer to break via the polymer scission mechanism, yet this is disallowed in our way of implementation of reaction schemes. The reason is that in our final analysis we set the limit of $k_f^- \rightarrow 0$, justifying our approximation. Lastly, in the eleventh term, i.e., polymer scission term, the lower limit $i+n_c$ makes sure that two fragments post-scission are stable polymers of size greater than or equal to $n_c$. The condition of the conservation of mass finally results into a time-dependent equation for the monomers \[eq3\] = -( \_[i=n\_c]{}\^ i y\_i(t) ).
Eq. \[eq2\] is different from previously obtained master equations on account of the additional terms that describe the contribution of the body evaporation $\kmb$ and addition $\kpb$. [@cohen; @hong] We also implement the *scission and recombination* pathway to allow for polymer fragmentation resulting only into fragments greater than or equal to the critical size; the recombining polymers also have to be of the size $i \ge n_c$. The reason for this is that we assume that any fragment of size $l \le n_c$ is highly unstable and quickly disintegrates to $l$ monomers. This prohibits the recombination of fragments smaller than the critical nucleus, as they do not exist in a polymeric state. Additionally, by disallowing the fragmentation that results in a fragment smaller than the critical size we prevent this step from contributing to the free monomeric pool. As a consequence, we completely decouple the *end evaporation and addition* from the *scission and recombination* pathway, i.e., one of the pathways can be switched on or off without affecting the other. [@cohen]
Our master equation as those very much like it discussed at length in the literature are being highly non-linear equations, and have so far eluded exact analytical solution. [@tuomas_review] Hence, a standard practice in the field is not to study the full length distribution, but only the first two moments of it. [@oosawa; @cohen] These are the polymer concentration, $P$, and the polymerised monomeric mass, $M$. The dimensionality of both are in moles per liter if the rate constants are given in molar units. Of these two quantities, the latter quantity is readily measured by means of, e.g., circular dichroism or fluorescence microscopy. [@hellstrand; @kelly_cd] The number concentration of polymers can in principle be quantified by measuring the mean size of the polymers, using techniques such as static and dynamic light scattering, and calculating the ratio of the polymerised mass to the mean degree of polymerisation. [@zhang_rev]
The two principal moments expressed in our variables read \[eq4\] P = \_[i=n\_c]{}\^ y\_i, for the polymer concentration and \[eq5\] M = \_[i=n\_c]{}\^ i y\_i, for the polymerised mass. We obtain the dynamical equations for $P$ and $M$ by extracting the first two principal moments from the full polymer length distribution described by Eq. \[eq2\]. In the process of deducing the dynamical equations for the moments, the only approximation we employ is that we neglect all terms arising from the disintegration of nuclei, which for early times are negligible in number anyway, at least in the limit $k_n^+ \rightarrow 0$. See Appendix A for details.
The dynamical equation for the number of polymers $P(t)$ that we obtain reads \[eq6\] &=& - P(t)\^2 + + k\_n\^+ \^[n\_c]{}, and for the time evolution of the the polymerised monomeric mass $M(t)$ we find \[eq7\] &=& 2 + (-M(t)) (M(t)+P(t))\
& &+ (2 P(t)-M(t))+ n\_c k\_n\^+ (-M(t))\^[n\_c]{}, where $\mtot$ is the total concentration of monomers in the system. A detailed derivation of Eqs. \[eq6\] and \[eq7\] from Eqs. \[eq2\] and \[eq3\] is provided in Appendix A. For this general set of equations the initial polymerised monomeric mass can have any value between and including, 0 and $\mtot$, i.e., $0\le M(0) \le \mtot$. The same holds true for $P(0)$ and $M(t) \ge P(t)$ for all times, where the equality holds only when all the polymers are critical nuclei.
In the absence of the *body evaporation and addition* terms, Eqs. \[eq6\] and \[eq7\] have been compared with kinetic Monte Carlo simulations that do allow for the disintegration of nuclei. [@Tiwari_2016] As expected, the time evolution of the moments obtained from the simulation are in quantitative agreement with Eqs. \[eq6\] and \[eq7\] in the appropriate limit of strongly nucleated polymerisation, justifying our approximation. With regard to the domain of validity of the Eqs. \[eq6\] and \[eq7\], they produce nonzero and positive $M(\infty)$ and $P(\infty)$ only above the so-called critical concentration of monomers. [@paul_review] In the absence of other pathways, except *end evaporation and addition*, the critical concentration obtained from our kinetic equations has a simple analogy to that of the thermodynamic theory of linear self-assembly. [@paul_review] However, in the presence of more complex pathways, the mapping is not so trivial as the kinetic theory demands the introduction of additional energy scales associated with the various pathways. For example, an energy scale for monomer removal or addition along the backbone of a polymer would be needed to characterise the *body evaporation and addition* in addition to an energy scale associated with the monomer addition to or removal from the ends of a polymer.
Notice that Eqs. \[eq6\] and \[eq7\] are highly nonlinear offering little hope of exact analytical solution. This implies that we have not improved the state of affairs significantly in the context of the various simplifications of the equations that we have already implemented. Of course, one may linearise Eqs. \[eq6\] and \[eq7\] and obtain linear solutions, but this is of limited help as a linear solution can never result in sigmoidal kinetics. Hence, with the aim to reduce the nonlinearity of the equations but preserve the most important and generic behaviour of the kinetics of self-assembly, we employ several additional approximations. Our first approximation is to restrict our analysis to the limit $k_n^+ \rightarrow 0$. This way we reduce the degree of the polynomial, which is essential to obtain an analytical solution. In addition, we also neglect the polymer recombination $\kpf P(t)^2$ term in Eq. \[eq6\]. By doing so, we break the reversibility condition and hence do not expect the system to follow the corresponding law of mass action. However, earlier studies on nucleated self-assembly have shown that in the context of the early time kinetics we focus attention on, polymer recombination does not play a significant role and hence can be ignored as in fact we shall also make plausible below. [@knowles_two_stage] Indeed, by identifying the most dominant terms in the dynamical equations for the polymerisation kinetics, the predicted lag phase has been shown to be in a quantitative agreement with theoretical models, at least in the absence of *body evaporation and addition*. [@tuomas_science; @knowles_two_stage; @hong; @hong_moment_closure; @subramaniam] This motivates us to do the same for our reaction pathway, despite it potentially being inaccurate in the long-time limit.
The steady-state solution of Eqs. \[eq6\] and \[eq7\] sheds some light on the role of molecular pathways and how our approximations, i.e., $k_n^+ \rightarrow 0$ and $k_f^+ P^2=0$, impact upon the long-time behaviour. In the steady state, the time derivatives of the two moments are zero and we are left with algebraic equations. If we define $K_f=\kpf/\kmf$ to be the fragmentation equilibrium constant, we can switch off the effect of *scission and recombination*, at least in equilibrium, by taking the asymptotic limit $K_f=0$. If $k_f^+=0$ this will be true for any value of $k_f^- \neq 0$. Within this approximation we can compare the effect of *end evaporation and addition* and *body evaporation and addition* explicitly. In this limit the equilibrium polymerised monomeric mass for a nucleus size of $n_c=2$ becomes \[eq8\] M() = -. In the absence of *body evaporation and addition*, i.e., for $\kpb=\kmb=0$, the polymerised monomeric mass becomes $M(\infty)=\mtot-\kme/\kpe=\mtot-K_e^{-1}$, where $K_e \equiv \kpe/\kme$. In the absence of *body evaporation and addition*, the ratio of rate constants $K_e$ can be mapped onto the equilibrium constant used in the thermodynamic theory of linear self-assembly. [@paul_review] In that case, in the polymerised regime, $y_1(\infty)= K_e^{-1}$, remains equal to the critical concentration. Hence, in this limit the polymerised monomeric mass is in agreement with the thermodynamic theory of self-assembly. [@paul_review] However, in the presence of *body evaporation and addition* pathway the effective elongation constant has to be renormalized to account for the free energy associated with monomer addition to or removal from the ends and that associated with monomer addition or removal along the backbone of the polymer.
In the limit $K_f \rightarrow 0$ the steady-state number of polymers $P$ and average degree of polymerisation $L$ are given by \[eq9\] P() &=& , and \[eq10\] L() &=& = 2 n\_c -1 for all values of $\kpb$ and $\kmb$. Note that the equilibrium average degree of polymerisation $L(\infty)$ only depends on the critical nucleus size, $n_c \ge 2$, not on the concentration. This, of course, disagrees with the thermodynamic theory but, as we shall see below, this does not preclude very large values of $L(t)$ for intermediate times. [@paul_review] Similar results were obtained by Cohen *et al.* in the absence of *body evaporation and addition*. [@knowles_two_stage]
Lag Time Analysis {#sec3}
=================
The steady-state solutions of the moment equations clearly indicate that our generalized reaction schemes do not result into the long polymers in the limit of $t \rightarrow \infty$ to be expected for nucleated polymerisation. However, as shown in Fig. \[fig2\], for intermediate times the mean polymer length overshoots and attains very large values. Also, Fig. \[fig2\] shows that the mean polymer length for very large and very small values of recombination rate constant, $k_f^+=10^{8}, 10^{-8}$, $M s^{-1}$ the term $k_f^+ P^2$ has no effect on the early time kinetics for any reasonable value of $k_f^+$. This justifies our approximation of neglecting the term representing the contribution of polymer recombination as in this paper we are mainly interested in comparing the early time kinetics of *end evaporation and addition* and *body evaporation and addition*. The early time kinetics is characterised by the lag time, i.e., the time intercept of the tangent at the inflection point of the polymerisation curve, i.e., $M(t)$. [@hellstrand] For the sake of simplicity, we limit ourselves to strongly nucleated polymerisation, and set $k_n^+ \rightarrow 0$. This also reduces the degree of polynomial on the right hand side of Eqs. \[eq6\] and \[eq7\], enabling us to obtain analytical solutions. In the limit $k_n^+ \rightarrow 0$, the polymerisation process has to be seeded, i.e., some initial polymerised mass, $M(0)=P(0) \neq 0$, has to be provided, otherwise the system stays in the initial state $M(0)=P(0)=0$.
![Time evolution of the mean length of the polymer $L(t)=M(t)/P(t)$, obtained by numerically solving Eqs. \[eq6\] and \[eq7\] for two values of recombination rate $\kpf=10^{8}$ and $10^{-8} M s^{-1}$. The remainder of the system parameters are $k_n^+=10^{-5} s^{-1}$, $\kpe=5 \times 10^{5} M^{-1} s^{-1}$, $\kme=10^{-2} s^{-1}$, $\kpb=5 \times 10^{5} M^{-1} s^{-1}$, $\kmb=10^{-2} s^{-1}$, $\kmf=10^{-4} s^{-1}$, $\mtot=10^{-5} M$, $M(0)=\mtot \times 10^{-4}$ and $P(0)=\mtot \times 10^{-6}$, where $M$ is moles per liter and $s$ is in seconds. The values of the chosen rate constants do not correspond to any particular experiment. However, the order of magnitude is similar to the parameters found in the literature of protein polymerisation. [@tuomas_review] The initial conditions are chosen to be small but non-zero, because of the necessity to seed the polymerisation process in the limit of $k_n^+ \rightarrow 0$.[]{data-label="fig2"}](Fig2.eps){width="4in"}
The resulting dynamical equations become, after employing the approximations and rearranging terms in Eqs \[eq6\] and \[eq7\], \[eq11\] &=& ( M(t)-(2 n\_c -1) P(t) ), for the polymer number concentration, and \[eq12\] &=& -M(t) (+ M(t)) + P(t) (- M(t)), for the polymerised mass, where we introduce new dynamical constants $\alpha \equiv \kmb -\kpb \mtot$, $\beta \equiv 2(\kmb-\kme)+\kpb\mtot+2\kpe \mtot$ and $\gamma \equiv \kpb+2\kpe$. It should be emphasized that we define the parameters $\alpha$, $\beta$ and $\gamma$ merely for notational simplicity and do not associate any physical meaning to them. Eqs. \[eq11\] and \[eq12\] are nonlinear in nature, hence, the first trivial analysis demands the linearisation of these equations. Researchers have analysed similar equations for variety of pathways and obtained $t^2$ or $t^3$ time dependence for the polymerised mass fraction. [@ferron_early] However, for our particular choice of molecular pathways, the resulting linearised equations always yield simple exponential time dependence that in turn result into a linear time dependence of early times.
![Time evolution of the polymerised mass fraction $M(t)/M(\infty)$ and the polymer concentration $P(t)/P(\infty)$ obtained by numerically solving Eqs. \[eq6\] and \[eq7\] for $\kpf=0$ $M^{-1} s^{-1}$. The remainder of the system parameters are $k_n^+=10^{-5} s^{-1}$, $\kpe=10^{5} M^{-1} s^{-1}$, $\kme=10^{-2} s^{-1}$, $\kpb=10^{5} M^{-1} s^{-1}$, $\kmb=10^{-2} s^{-1}$, $\kmf=10^{-4} s^{-1}$, $\mtot=10^{-5} M$, $M(0)=\mtot \times 10^{-5}$ and $P(0)=\mtot \times 10^{-6}$, where, $M$ is moles per liter and $s$ is in seconds. The inset shows the early time behaviour of the number of polymers $P(t)$, which remains essentially constant during the relaxation of the polymerised monomeric mass, $M(t)$, and only relaxes after that. This figure highlights our claim of separation of time scales between the time evolution of the polymerised mass fraction and that of the polymer concentration. See the main text.[]{data-label="fig3"}](Fig3.eps){width="6.5in"}
Although the dynamical equations Eqs. \[eq11\] and \[eq12\] contain the polymer scission term, we can reduce the effect of polymer scission on the kinetics of the polymerised monomeric mass, $M$, by choosing a very small value for the scission rate constant $\kmf$. The value of $\kmf=10^{-4}$ that we chose is small enough to diminish the effect of scission for the early time kinetics of the polymerised mass, as we find and also show below that this is true for our choice of set of parameters. For this small value of $\kmf$, the typical trajectory of the two moments is shown in Fig. \[fig3\]. We notice that the polymerised mass $M$ evolves much faster in time than the number density of polymers $P$. The inset in Fig. \[fig3\]b shows that in the asymptotic limit of vanishing scission rate $\kmf$, the lag phase for both moments is characterised by an approximately equal time scale. However, after the lag phase, $M$ relaxes to its equilibrium value much faster than $P$ does. Hence, for the early time kinetics of the polymerised monomeric mass, the number concentration of polymers remains effectively constant, i.e., equal to initial value $P(t)=P(0)$.
Making use of this, we can solve Eq. \[eq11\] for $M(t)$ for early times, yielding the explicit solution, \[eq13\] M(t) &=& - , where for the purpose of notational simplicity we define $\sigma \equiv \alpha+\gamma P(0)$ and $\eta \equiv \sqrt{\sigma ^2+4 \beta \kpb P(0)}$. In Fig. \[fig4\] we show that Eq. \[eq13\] agrees with the numerical results obtained in the limit $\kmf \rightarrow 0$. Below we will analyse the effect of molecular pathways by calculating the lag time from Eq. \[eq13\]. For now, the main conclusion is that Eq. \[eq13\] predicts sigmoidal kinetics, i.e., a lag phase followed by exponential growth and subsequent saturation of the solution for $M(t)$.
Next, we solve for $P(t)$ in the long-time limit when $M(t)$ has achieved its equilibrium value $M(\infty)$. For long times, i.e., post-lag phase of the polymerised mass fraction, $P(t)$ is given by \[eq14\] P(t) &=& . Notice that for the analytical solution of the number concentration of polymers, the time scale of the evolution of $P(t)$ depends only on the scission rate constant $\kmf$. This is due to our choice of reaction scheme, where the *scission and recombination* affects only the number of polymers. In contrast, *end evaporation and addition* and *body evaporation and addition* only affect the exchange of monomers between the free monomer pool and the polymer pool.
Having obtained a closed analytical expression for the time evolution of the polymerised mass, $M(t)$, we can now calculate the lag time, $\tlag$. This is achieved by calculating the inflection point or the point of maximum growth rate, i.e., the time at which the second derivative of Eq. \[eq13\] is equal to zero. We then calculate the time intercept of the tangent at the inflection point, resulting in the analytical expression for the lag time. In general, for nucleated polymerisation kinetics the lag time $\tlag$ is a linear combination of two terms. [@hong_moment_closure] These terms represent two important characteristics of the polymerisation kinetics: the half-time $\thalf$ and apparent growth rate $\kapp$. The half-time $\thalf$ is the time at which the polymerised monomeric mass is exactly half of its steady-state (long-time) value and from Eq. \[eq13\] we find that it is equal to the inflection point. Furthermore, the apparent growth rate $\kapp$ is the growth rate of the polymerisation curve at the inflection point (so, the time derivative of $M(t)$ at $t=\thalf$).
Hence, we find \[eq15\] = - , where \[eq16\] =, denotes the half-time and \[eq17\] = - , the apparent growth rate.
To investigate the influence of the overall monomer concentration on the lag time, let us assume that all the other parameters, i.e., the rate constants, are constant and do not depend on the monomer concentration. This also ties in with our assumption that our polymerisation process is reaction-limited and not diffusion-limited. Let us first focus on the case $\kpb=\kpe$ and $\kmb=\kme$, i.e., which is true if *body evaporation and addition* kinetics and *end evaporation and addition* are equally likely. We see that $\alpha = \kmb -\kpb \mtot \approx -\kpb \mtot$ where we have $\kpb \mtot \gg \kmb$ for polymers to exist. The other parameters hence become $\beta = 2(\kmb-\kme)+\kpb\mtot+2\kpe \mtot \approx 3 \kpb \mtot$ and $\gamma = \kpb+2\kpe = 3\kpb \approx $ constant, i.e., independent of total monomer concentration. This further implies that $\sigma =\alpha + \gamma P(0) \approx -\kpb \mtot$, because $\alpha \sim \kpb \mtot \gg \gamma P(0) = 3 \kpb P(0)$ as $\mtot \gg P(0)$. Finally, the denominator in the expression for half-time is $\eta = \sqrt{\sigma^2 + 4 \beta \kpb P(0)} \approx \sqrt{(\kpb \mtot)^2+ 12 \kpb \mtot P(0)}$.
Again, as $P(0) \ll \mtot$, we infer that $\eta \approx \kpb \mtot$. From Eq. \[eq16\] the half-time has a logarithmic numerator resulting in a weak dependence of the numerator on $\mtot$, and effectively we have $\thalf \sim \eta^{-1} \sim \mtot^{-1}$. Similarly, $\kapp=\eta^2/2 \sigma$, where $\eta \approx \kpb \mtot$, and hence the apparent growth rate scales as $\kapp \sim \mtot$. In Fig. \[fig5\], we show, by fitting the concentration dependence of the half-time Eq. \[eq17\] on a double logarithmic scale that our predictions for the power laws are indeed correct. It confirms that the logarithmic correction is indeed negligible. These power laws for the half-time and the apparent growth rate have same exponents as for the case of *end evaporation and addition* in the limit of weak scission obtained before. [@hong]
Let us now examine how the half-time and the apparent growth rate depend on the forward rates of the *body evaporation and addition* and the *end evaporation and addition* pathways, when only one of them is present. In the absence of *end evaporation and addition* pathway, i.e., $\kpe=\kme=0$, we have $\eta \sim \kpb \mtot$, as shown above. This, in combination with the fact that the half-time scales as $\thalf \sim \eta^{-1}$ and that the apparent growth rate scales as $\kapp \sim \sigma/\eta^{-1}$, where $\sigma \sim - \kpb \mtot$, results in scaling of $\thalf \sim (\kpb)^{-1}$ and $\kapp \sim \kpb$. This is confirmed in Fig. \[fig6\].
In the absence of *body evaporation and addition* only primary nucleation and the *end evaporation and addition* pathways are active, in the limit where *scission and recombination* is sufficiently weak. In that limit, our equations are exactly the same as those presented in earlier works. [@hong] From the work of Hong et al., we already know that for weak scission, both the half time and the reciprocal apparent growth rate scale linearly with the total monomer concentration, which is the same as what we found for *body evaporation and addition*. [@hong] In other words, from the monomer concentration dependence of the lag time we cannot distinguish body from end evaporation and addition. However, differences do show up when considering the dependence of the lag time on the forward rate constants. For the *end evaporation and addition*, the half time and the reciprocal apparent growth rate scale as the square root of the forward rate constant of monomer addition, i.e., $\sqrt{\kpe}$, [@hong] which contrasts with what we found for *body evaporation and addition*. For the latter we found a linear scaling with the forward rate constant.
In conclusion, although the scaling of the lag time with the total monomer concentration is the same for both pathways, there is a difference in the scaling with the forward rate constants, or, equivalently, with the equilibrium constants for the two pathways. This in principle provides us with a probe to inspect the existence of *body evaporation and addition* pathway in the polymerisation process, at least if we knew how to control them.
Discussion and Conclusions {#sec4}
==========================
In our model calculations, we compare predictions based on the newly discovered kinetic pathway of *body evaporation and addition* with the well studied pathway of *end evaporation and addition*, in the context of strongly nucleated reversible polymerisation. To be able to do that, we rely on a kinetic scheme that includes a third pathway, known as *scission and recombination*, that in the end we switch off asymptotically. [@cates_sr; @tuomas_review] The presence of this third pathway is needed solely for mathematical reasons, to achieve our purpose. In practice, we cannot exclude the presence of *scission and recombination*, and our set of equations, which focus on the two moments of the full distribution, allow for that.
The way the *scission and recombination* and *end evaporation and addition* pathways are implemented in the literature creates a conundrum, in which a process where a monomer is removed from the end can either be seen as a polymer scission or as an end evaporation event. [@hong; @tuomas_review] To remedy this, we have implemented the *scission and recombination* pathway for strongly nucleated linear reversible polymerisation in such a way that a polymer can break, resulting into the formation of two polymers of size greater than or equal to the *critical nucleus*, which in our model has to be bigger than or equal to a dimer. We modelled polymer recombination, so the merging of two short polymers into a longer one the same way. Also we exclude the case $n_c=1$. This prescription prevents the *scission and recombination* pathway from interfering with the *end evaporation and addition*. So in our prescription *scission and recombination* does not directly affect the monomer pool.
In our view, our alternative implementation of the two pathways is the more sensible one, because detachment of a monomer is inherently different from the breaking of a polymer. Indeed, the latter process should strongly depend on the length of the polymer, whilst the former arguably should not provided the polymer is sufficiently long. [@tom_length_dependence] In addition, our implementation of the *scission and recombination* pathway reproduces the thermodynamically consistent law of mass action for the amount of polymerised material under conditions of equilibrium and in the presence of the *end evaporation and addition* pathway. The amount of polymerised material of course is only one of the moments of the full length distribution. As for the other moments, such as the number of polymers in solution, they suffer from thermodynamically inconsistent long-time behaviour, a drawback that we share with previous studies. [@cohen; @hong; @tuomas_science]
However, if we focus on the short-time behaviour of the reversible polymerisation reactions, we could argue, as in fact is tacitly done in the literature, that the long-time behaviour of the system is inconsequential. [@hong; @tuomas_review] Indeed, many thermodynamically inconsistent pathways have been shown to be in quantitative agreement with experimental findings regarding the early time kinetics, which is the prime focus of almost all experimental works. [@tuomas_science; @knowles_two_stage; @hong; @hong_moment_closure; @subramaniam] For this reason, we also focus on the early time kinetics, when comparing body and end evaporation and addition, in the limit of asymptotically weak scission. Because polymer recombination has been shown theoretically not to substantially alter the early-time polymerisation kinetics, we suppress the recombination step completely for mathematical expedience in order to obtain closed-form solutions, in the final steps of our analysis. [@knowles_two_stage]
In this limit, we find that the typical polymerisation kinetics of *body evaporation and addition* is characterised by a separation of time scales between the polymerised monomeric mass $M$ and the polymer number concentration $P$. This is obvious from the difference in the lag times for these two quantities, i.e., the time scales required to get a substantial growth. Such separation of time scales has also been found for the *end evaporation and addition* pathway in combination with weak polymer scission and no recombination albeit that it is is not as strong. [@hong_moment_closure] To illustrate this, in Fig. \[fig7\], we show the polymerised mass fraction and the polymer concentration as a function of time, where we vary the forward rate of body addition at a fixed value of the forward rate of end addition. We cover the full spectrum from predominant body to predominant end evaporation and addition.
Comparing Fig. 7a and 7b, it is clear that the separation of time scales is many orders of magnitude larger for *body evaporation and addition* than that for *end evaporation and addition*. Indeed, the polymerised mass fraction evolves much faster in time than the number of polymers does, if the *body evaporation and addition* pathway is active. It is this vast separation of time scales that allows us to obtain first order perturbative solutions to the two moment equations that are in quantitative agreement with numerical results for vanishing scission rate constant.
Our perturbative solutions for the polymerised mass fraction and the polymer concentration, are sigmoidal as a function of time and provide us with the closed form solution for the lag time. The lag time associated with the polymerised mass fraction is the one that is experimentally readily accessible, and hence we focus on that. We find, within the limit where it is valid, that the lag time produces only a weak dependence on the initial polymerised monomeric mass. This contrasts with other studies where they do not include *body evaporation and addition* in the reaction scheme, and is a result of the extremely fast kinetics connected with that pathway. [@oosawa; @cohen; @hong; @hong_moment_closure]
In the usual definition of the lag time, it takes the form of the sum of a half-time $\thalf$ and a reciprocal of an apparent growth rate $\kapp$. [@tuomas_review] For *body evaporation and addition* we find the same linear scaling of the the half time and the reciprocal apparent growth rate with the total monomer concentration, $\mtot$, as was found in the literature for the *end evaporation and addition*. [@hong] In fact, for the *bulk evaporation and addition* pathway it is the product $\kpb \mtot$ dictates the scaling of the lag time, implying that both time scales are also linearly dependent on the forward rate for the body addition, $\kpb$. In contrast, for the *end evaporation and addition* pathway both time scale are proportional to the square root of the forward rate constant of the end addition, $\sqrt{\kpe}$. [@hong]
It seems that the newly proposed pathway, although it speeds up the growth of the polymerised mass by providing a larger number of places to insert or remove a monomer from a polymer, does not affect the dependence of the lag time on the total monomer concentration. However, it does affect how the relevant forward rate constant influences that time scale. This means that to distinguish end from body evaporation and addition experimentally by probing the dependence of the lag time on the system parameters, we would need to be able to control this quantity. This is not trivial. It may well be that highly specialised techniques, such as those used in reference [@albertazzi], are required to actually observe it. In fact, this may even be the reason, why it had not been considered before.
Acknowledgements
================
We thank Mariana Oshima Menegon, Shari Finner and Stefan Paquay for a critical reading of the manuscript. This work was supported by the Nederlandse Organisatie voor Wetenschappelijk Onderzoek through Project No. 712.012.007.
Derivation of Moment Equations from Discrete Master Equation
============================================================
We start by writing down the discrete master equation, Eq. \[eq2\], for the reaction schemes described in the main text. Next, we define the principal moments, the number of polymers $P$ and the polymerised monomeric mass $M$, Eqs. (7) and (8).
After substituting Eq. \[eq4\] in Eq. \[eq5\], the equation for the number of polymers $P$ is given by, &=& \_[i=n\_c]{}\^ k\_n\^+ x(t)\^[n\_c]{} \_[i,n\_c]{} + 2 \_[i=n\_c]{}\^ x(t) y\_[i-1]{}(t) - 2 x(t) \_[i=n\_c]{}\^ y\_[i]{}(t) + 2 \_[i=n\_c]{}\^ y\_[i+1]{}(t) - 2 \_[i=n\_c]{}\^ y\_i(t)\
& & + \_[i=n\_c]{}\^ (i-2) y\_[i-1]{}(t) - \_[i=n\_c]{}\^ (i-1) x(t) y\_[i]{}(t) + \_[i=n\_c]{}\^ (i-1) y\_[i+1]{}(t) - \_[i=n\_c]{}\^ (i-2) y\_i(t)\
& & - \_[i=n\_c]{}\^ (i- 2 n\_c +1) y\_i(t) + 2 \_[i=n\_c]{}\^ \_[j=i+n\_c]{}\^ y\_j(t)\
& & + \_[i=n\_c]{}\^ \_[k+l=i]{} y\_k(t) y\_l(t) - 2 \_[i=n\_c]{}\^ y\_i (t) \_[j=n\_c]{}\^ y\_j. To obtain a kinetic equation without any summation signs, we collect and simplify the equation term by term. Let us start with the terms accounting for monomer addition at the end, & & ( \_[i=n\_c]{}\^ y\_[i-1]{} - \_[i=n\_c]{}\^ y\_[i]{} ) = ( \_[i=n\_c-1]{}\^ y\_[i]{} - \_[i=n\_c]{}\^ y\_[i]{} ) = y\_[n\_c-1]{} = 0. Notice that $y_{n_c-1}=0$, because the smallest stable polymer is the critical nucleus of size $n_c$. Similarly for end evaporation, & & ( \_[i=n\_c]{}\^ y\_[i+1]{} - \_[i=n\_c]{}\^ y\_[i]{} ) = ( \_[i=n\_c+1]{}\^ y\_[i]{} - \_[i=n\_c]{}\^ y\_[i]{} ) = - y\_[n\_c]{}. The analysis in the main text assumes the strongly nucleated polymerisation, i.e., $k_n^+ \rightarrow 0$. In this limit the critical nuclei are highly unstable and can be neglected.
For terms six and seven in Eq. (A1) that represent monomer addition on the polymer backbone, we obtain \_[i=n\_c]{}\^ ( (i-2) y\_[i-1]{} - (i-1) y\_i ) = \_[i=n\_c-1]{}\^ (i-1) y\_i - \_[i=n\_c]{}\^ (i-1) y\_i = 0. Similarly, terms eight and nine in Eq. (A1) accounting for monomer removal from the polymer backbone, simplify to \_[i=n\_c]{}\^ (i-1) y\_[i+1]{} - \_[i=n\_c]{}\^ (i-2) y\_i = \_[i=n\_c]{}\^ (i-1) y\_[i+1]{} - \_[i=n\_c-1]{}\^ (i-1) y\_i = - (n\_c-1) y\_[n\_c]{}. Once again under the assumption of strongly nucleated polymerisation, we neglect $y_{n_c}$.
The contribution from polymer scission can be rewritten in terms of a theta function as - \_[i=n\_c]{}\^ (i- 2 n\_c +1) y\_i + 2 \_[i=n\_c]{}\^ \_[j=n\_c]{}\^ y\_j (i-j-n\_c)\
- \_[i=n\_c]{}\^ (i- 2 n\_c +1) y\_i + 2 \_[i=n\_c]{}\^ (1+j-2 n\_c) y\_j = (M-(2 n\_c -1) P), and the contribution of polymeric recombination is \_[k=n\_c]{}\^ \_[l=n\_c]{}\^ y\_k y\_l - 2 \_[i=n\_c]{}\^ y\_i \_[j=n\_c]{}\^ y\_j = ( \_[k=n\_c]{}\^ y\_k ) ( \_[l=n\_c]{}\^ y\_l ) - 2 ( \_[i=n\_c]{}\^ y\_i ) ( \_[j=n\_c]{}\^ y\_j ) = -P\^2 This gives us following equation for the number of polymers $P$, = - P(t)\^2 + ( M(t)-(2 n\_c -1) P(t) ) + k\_n\^+ x(t)\^[n\_c]{} For the polymerised monomeric mass $M(t)$, the evolution equation is, &=& \_[i=n\_c]{}\^ k\_n\^+ i x(t)\^[n\_c]{} \_[i,n\_c]{} + 2 x(t) \_[i=n\_c]{}\^ i y\_[i-1]{}(t) - 2 x(t) \_[i=n\_c]{}\^ i y\_[i]{}(t) + 2 \_[i=n\_c]{}\^ i y\_[i+1]{}(t) - 2 \_[i=n\_c]{}\^ i y\_i(t)\
& & + \_[i=n\_c]{}\^ i (i-2) y\_[i-1]{}(t) - \_[i=n\_c]{}\^ i (i-1) x(t) y\_[i]{}(t) + \_[i=n\_c]{}\^ i (i-1) y\_[i+1]{}(t) - \_[i=n\_c]{}\^ i (i-2) y\_i(t)\
& & - \_[i=n\_c]{}\^ i (i- 2 n\_c +1) y\_i(t) + 2 \_[i=n\_c]{}\^ \_[j=i+n\_c]{}\^ i y\_j(t) + \_[i=n\_c]{}\^ \_[k+l=i]{} i y\_k(t) y\_l(t)\
& & - 2 \_[i=n\_c]{}\^ i y\_i (t) \_[j=n\_c]{}\^ y\_j(t). The terms involving monomer addition at the ends are 2 x(t) \_[i=n\_c]{}\^ i ( y\_[i-1]{}(t)-y\_i(t) ) = 2 ( \_[i=n\_c-1]{}\^ (i+1) y\_i - \_[i=n\_c]{}\^ i y\_i ) = 2 x(t) P(t), and the term arising from monomer removal at the ends are 2 \_[i=n\_c]{}\^ i ( y\_[i+1]{}(t)-y\_i(t) ) = 2 ( \_[i=n\_c+1]{}\^ (i-1) y\_i - \_[i=n\_c]{}\^ i y\_i ) = 2 (-P(t) - n\_c y\_[n\_c]{}). Where we neglect the contribution, $n_c y_{n_c}$.
Next the terms contributed by the monomer addition in the bulk becomes & & \_[i=n\_c]{}\^ ( i (i-2) y\_[i-1]{} - i (i-1) y\_i )\
&=& ( \_[i=n\_c-1]{}\^ (i+1)(i-1) y\_i - \_[i=n\_c]{}\^ i (i-1) y\_i )\
&=& ( \_[i=n\_c]{}\^ (i\^2-1) y\_i + \_[i=n\_c]{}\^ i y\_i - \_[i=n\_c]{}\^ i\^2 y\_i )\
&=& x(t) (M(t) - P(t)), and the terms from monomer removal from the bulk are & & ( \_[i=n\_c]{}\^ i (i-1) y\_[i+1]{} - \_[i=n\_c]{}\^ i (i-2) y\_i )\
&=& ( \_[i=n\_c+1]{}\^ (i-1) (i-2) y\_i - \_[i=n\_c]{}\^ i (i-2) y\_i )\
&=& - \_[i=n\_c+1]{}\^ (i-2) y\_i = (- n\_c y\_[n\_c]{} + 2 P(t)- M(t)). This completes our closure of discrete master equation to obtain moment equations, that are given by, &=& - P(t)\^2 + ( M(t)-(2 n\_c -1) P(t) ) + k\_n\^+ x(t)\^[n\_c]{}, and &=& 2 ( x(t) P(t) - P(t) ) + x(t) (M(t)+P(t)) + (2 P(t)-M(t))+ n\_c k\_n\^+ x(t)\^[n\_c]{}.
T. F. A. De Greef, M. M. J. Smulders, M. Wolffs, A. P. H. J. Schenning, R. P. Sijbesma and E. W. Meijer, Chem. Rev., 109, 5687-5754, (2009). M. Takalo, A. Salminen, H. Soininen, M. Hiltunen and A. Haapasalo, Am. J. Neurodegener. Dis., 2, 1-14 (2013). L. Blanchoin, R. B. Paterski, C. Sykes and J. Plastino, Physiol. Rev. 94, 235-263 (2014). D. A. Fletcher, R. D. Mullins, Nature 463, 485–492 (2010). F. Chiti, C. M. Dobson, Annu. Rev. Biochem. 75, 333–366 (2006). M. M. P. Murphy and H. Levine III, Journal of Alzheimer’s Disease, 19, 311-323 (2010). T. Aida, E. W. Meijer, S. I. Stupp, Science 335, 813–817 (2012). A. Ciferri (Ed.), *Supramolecular polymers*, 2nd Edition (Taylor and Frances, Boca Raton, 2005). S. Katen and A. Zlotnick, Methods Enzymol., 455, 395-417 (2009). C. M. Marques, M. S. Turner and M. E. Cates, Journal of Non-Crystalline Solids, 172-173, 1168-1172 (1994). F. Oosawa and S. Asakura, *Thermodynamics of the Polymerization of Protein* (Academic Press, New York, 1975). S. I. A. Cohen, M. Vendruscolo, C. M. Dobson, E. M. Terentjev, and T. P. J. Knowles, J. Chem. Phys. 135, 065105 (2011). A. M. Morris, M. A. Watzky, R. G. Finke, Biochim Biophys Acta., 1794, 375-397 (2009). C. M. Marques, M. S. Turner and M. E. Cates, J. Chem. Phys., 99, 7260 (1993). J.L.A. Dubbeldam and P.P.A.M. van der Schoot, J. Chem. Phys., 123, 144912 (2005). A. N. Semenov and I. A. Nyrkova, J. Chem. Phys, 134, 114902 (2011). M. E. Cates, Macromolecules, 20, 2289-2296 (1987). M. S. Turner and M. E. Cates, J. Phys. France, 51, 307-316 (1990). I. A. Nyrkova and A. N. Semenov, Eur. Phys. J. E, 24, 167-183 (2007). G. A. Garcia, S. I. A. Cohen, C. M. Dobson and T. P. J. Knowles, Phys. Rev. E, 89, 032712 (2014). N. S. Tiwari and P. van der Schoot, J. Chem. Phys. 144, 235101 (2016). L. Albertazzi, D. van der Zwaag, C. M. A. Leenders, R. Fitzner, R. W. van der Hofstad and E. W. Meijer, Science, 344, 6183, 491-495 (2014). P. Arosio, T. P. J. Knowles and S. Linse, Phys. Chem. Chem. Phys., 17, 7606-7618 (2015). L. Hong, X. Qi and Y. Zhang, J. Phys. Chem. B, 116 (23), 6611–6617 (2012). L. Hong and Wen-An Yong, Biophys. Journal, 140, 533-540 (2013). T. P. J. Knowles, C. A. Waudby, G. L. Devlin, S. I. A. Cohen, A. Aguzzi, M. Vendruscolo, E. M. Terentjev, M. E. Welland and C. M. Dobson, Science, 326, 5959, 1533-1537 (2009). T. P. J. Knowles, M. Vendruscolo and C. M. Dobson, Nat. Rev. Mol. Cell Biology, 15, 384–396 (2014). V. V. Schvadchak, M. M. A. E. Claessens and V. Subramaniam, J. Phys. Chem. B, 119, 1912–1918 (2015). T. C. T. Michaels and T. P. J. Knowles, J. Chem. Phys., 140, 214904 (2014). E. Hellstrand, B. Boland, D. M. Walsh, and S. Linse, ACS Chem. Neurosci., 1, 13-18 (2010). S. M. Kelly, T. J. Jess and N. C. Price, Biochim. Biophys. Acta, 1751, 119-139 (2005). Y. Liu, Z. Wang and X. Zhang, Chem. Soc. Rev., 41, 5922–5932 (2012). M. F. Bishop and F. A. Ferrone, Biophys., 46, 631-644 (1984). I. A. W. Filot,A. R. A. Palmans, P. A. J. Hilbers, R. A. van Santen, E. A. Pidko, and T. F. A. de Greef, J. Phys. Chem. B, 114, 13667–13674 (2010).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present a new compressed representation of free trajectories of moving objects. It combines a partial-sums-based structure that retrieves in constant time the position of the object at any instant, with a hierarchical minimum-bounding-boxes representation that allows determining if the object is seen in a certain rectangular area during a time period. Combined with spatial snapshots at regular intervals, the representation is shown to outperform classical ones by orders of magnitude in space, and also to outperform previous compressed representations in time performance, when using the same amount of space.'
author:
- 'Nieves R. Brisaboa'
- Travis Gagie
- 'Adrián Gómez-Brandón'
- Gonzalo Navarro
- 'José R. Paramá'
bibliography:
- 'bibliografia.bib'
title: 'Efficient Compression and Indexing of Trajectories[^1]'
---
Introduction
============
With the appearance of cheap devices, such as smartphones or GPS trackers, which record the position of moving objects, the need to efficiently store and manage information on trajectories has become commonplace. Although storage, network, and processing capacities are rapidly increasing, the available data grows faster, and demands reduced-size representations [@zheng11]. The first option is to lose precision and discard points of the acquired trajectories, with more or less sophisticated procedures. A second choice is to keep all the points of the trajectories and use differential compression [@TrajStore; @Trajic]. These methods store for each coordinate $(x,y)$ the difference with the previous point. The problem is that, to obtain the coordinates of the $i^{th}$ point, we must add up all the preceding differences. This is a variant of the [*partial sums problem*]{} where the values can be positive or negative.
Our new method, called Constant Time Access Compressed Trajectories (ContaCT), uses an Elias-Fano-based [@Fan71; @Eli74] representation of the differences that allows computing the partial sums in constant time while using space comparable to other differential encoding methods. In addition to constant-time access to the trajectory data, ContaCT provides a hierarchical structure that allows efficiently answering time-interval queries [@PfoserJT00] (i.e., determine if an object is seen inside a rectangular area during a time interval) without the need to follow all the movements of the object in the queried interval. We use ContaCT to represent the trajectories of a large set of objects. At regular time instants, ContaCT includes a spatial snapshot with a structure that supports range queries, which is useful to bound the objects that must be tracked to answer time-interval queries.
Our experiments on a set of real trajectories of ships shows that, while there exist techniques based on grammar-compression that use less space than ContaCT [@BrisaboaGNP16], our index is up to 2.7 times faster when using about the same amount of space. Our index is also much faster than a baseline differentially compressed representation, for about the same space. We also compared ContaCT with a classical MVR-tree, where trajectories are stored as sets of points and time-interval queries reduce to 3D range queries. It turns out that ContaCT required 1,300 times less space, and it was still faster in time-interval queries spanning more than 14 instants.
Background
==========
A trajectory is a sequence of timestamped geographic positions in the two-dimensional space. We assume that the recorded timestamps are regularly placed over time, possibly with periods of time without values. We also assume that the recorded timestamps are exactly the same for all the objects.
Apart from the basic functionality of returning the whole trajectory of an object or its position at some time instant, we deal with the following, more elaborate queries [@PfoserJT00]: *time-slice* returns all the objects in a given query region at a given timestamp, and [*time-interval*]{} returns all the objects that overlap the query region at any time instant of an interval.
**Bitmaps.** A *bitmap* is a binary sequence $B[1,n]$ that supports the following operations: (i) $access(B,i)$ returns the bit $B[i]$, (ii) $rank_b(B,i)$ returns the number of occurrences of bit $b\in{0,1}$ in $B[1,i]$, and (iii) $select_b(B,j)$ returns the position in $B$ of the $j^{th}$ occurrence of bit $b\in{0,1}$. There exist representations using $n+o(n)$ bits that answer all those queries in constant time [@Clark:1996]. When the bitmap has $m \ll n$ 1s, it is possible to use compressed representations that use $m\log(n/m)+O(m)$ bits [@Eli74; @Fan71]. This representation still performs $select_1$ queries in constant time, whereas $access$ and $rank$ require time $O(\log(n/m))$ [@Okanohara:2007:PER:2791188.2791194].
**Partial sums.** Given values $0 < x_1 < x_2 < \ldots < x_m \le n$, we can define the differences $d_i =
x_i-x_{i-1}$ and $d_1=x_1$, so that $x_i = \sum_{j=1}^i d_i$. An Elias-Fano representation of the partial sums is a bitmap $B[1..n]$ with all $B[x_i] = 1$ and all the rest zero, or which is the same, the concatenation of the $d_i$ values written in unary. Therefore, we can retrieve $x_i =
select_1(B,i)$ in constant time, and the space of the representation is $\log(n/m)+O(m)$ bits, close to a differential representation of the $d_i$ values.
Related work
============
**Reducing the size of trajectories.** A lossy way to reduce size is to generate a new trajectory that approximates the original one, by keeping the most representative points. The best known method of this type is the Douglas-Peucker algorithm [@Douglas:1973:ARN]. Other strategies record speed and direction, discarding points that can be reasonably predicted with this data [@TrajcevskiCSWV06]. A lossless way to reduce space is to use differential encodings of the consecutive values $x$, $y$, and time [@TrajStore; @Trajic; @Wang:2014].
**Spatio-temporal indexes.** Spatio-temporal indexes can be classified into three types. The first is a classic multidimensional spatial index, usually the R-tree, augmented with a temporal dimension. For example, the 3DR-tree [@Vazirgiannis1998] uses three-dimensional Minimum Bounding Boxes (MBBs), where the third dimension is the time, to index segments of trajectories. A second approach is the multiversion R-trees, which creates an R-tree for each timestamp and a B-tree to select the relevant R-trees. The best known index of this family is the MV3R-tree [@PapadiasT01]. The third type of index partitions the space statically, and then a temporal index is built for each of the spatial partitions [@ChakkaEP03].
GraCT {#Grct}
-----
The closest predecessor of our work, GraCT [@BrisaboaGNP16], assumes regular timestamps and stores trajectories using two components. At regular time instants, it represents the position of all the objects in a structure called *snapshot*. The positions of objects between snapshots are represented in a structure called *log*.
Let us denote $Sp_{k}$ the snapshot representing the position of all the objects at timestamp $k$. Between two consecutive snapshots $Sp_{k}$ and $Sp_{k+d}$, there is a log for each object, which is denoted ${\cal L}_{k,k+d}(id)$, being $id$ the identifier of the object. The log stores the differences of positions compressed with [*RePair*]{} [@larsson2000off], a grammar-based compressor. In order to speed up the queries over the resulting sequence, the nonterminals are enriched with additional information, mainly the MBB of the trajectory segment encoded by the nonterminal.
Each snapshot is a binary matrix where a cell set to 1 indicates that one or more objects are placed in that position of the space. To store such a (generally sparse) matrix, it uses a $k^2$-tree [@ktree]. The $k^2$-tree is a space- and time- efficient version of a region quadtree [@Sam2006], and is used to filter the objects that may be relevant for a time-instant or time-interval query.
ScdcCT
------
ScdcCT was implemented as a classical compressed baseline to compare against GraCT [@BrisaboaGNP16]. It uses the same components, snapshots and logs, but the logs are compressed with differences and not with grammars. The differences are compressed using $(s,c)$-Dense Codes [@RodrguezBrisaboa07], a fast-to-decode variable-length code that has low redundancy over the zero-order empirical entropy of the sequence. This exploits the fact that short movements to contiguous cells are more frequent than movements to distant cells.
ContaCT
=======
ContaCT uses snapshots and logs, just like GraCT. The main differences are in the log. As explained, in GraCT the log stores the differences of the consecutive positions. In order to know the position of an object at a given timestamp $i$, we access the closest previous snapshot and add up the differences until reaching the desired timestamp. GraCT speeds up this traversal by storing the total differences represented by nonterminals, so that they can be traversed in constant time. This makes GraCT faster than a differential representation that needs to add up all the individual differences, but still it has to traverse a number of symbols that grows proportionally to the distance $d$ between consecutive snapshots. ContaCT completely avoids that sequential traversal of the log.
The log
-------
ContaCT represents each ${\cal L}_{k,k+d}(id)$ with components $time(id)$, $\Delta_X(id)$, $\Delta_Y(id)$.
Time $(id)$ tells the timestamps for which object $id$ has $(x,y)$ coordinates. It stores the $first$ and $last$ positions with data in ${\cal L}_{k,k+d}(id)$, and a bitmap $T(id)$ of $last-first+1$ bits indicating with a 0 that there is data at that time instant.
$\Delta_{X}(id)$ stores the differences of the $x$ coordinate using three bitmaps: $X(id)_t$ indicates, for each position having a 0 in $T(id)$, whether the difference is positive or negative; and $X(id)_{p}$ and $X(id)_n$ store the positive and negative differences, respectively, using Elias-Fano. $\Delta_{Y}(id)$ is analogous.
Given the log ${\cal L}_{k,k+d}(id)$ and a local timestamp $i\in [1,d-1]$, we compute the $x$ coordinate of the object $id$ at that timestamp as follows (analogous for $y$):
1. $dis=rank_1(T(id),i-first+1)$ returns the number of timestamps for which we have no data (the object was missing) until position $i$, counting from the first timestamp with data.
2. $pos=rank_1(X(id)_t,i-dis-first+1)$ and $neg=i-dis-first-pos+1$, are the number of positive and negative differences until position $i$, respectively.
3. $select_1(X(id)_p,pos)-pos-(select_1(X(id)_n,neg)-neg)$ returns the $x$ coordinate at timestamp $i$.
We use the sparse bitmap representation for $X(id)_{p}$ and $X(id)_{n}$, and the plain version for $X(id)_t$ and $T(id)$. The size of the complete structure is $n \log N/n + O(d)$ bits, where $N$ is the sum of the differences in $x$, and $n \le d$ is the number of positions where the object has coordinate information.
![The log of ContaCT for a given object *id*.[]{data-label="estructura"}](images/estructura2.eps)
**Example.** The top of Figure \[estructura\] shows the coordinates of a trajectory. There is no data about the position of the object at timestamps 1, 6, 7, 11, and 12. Timestamps $0$ and $13$ are represented with snapshots. Arrays $X$ and $Y$ contain the absolute coordinates of the trajectory, and $\Delta X$ and $\Delta Y$ the corresponding differences (the arrays are not stored in this form, they are included for clarity).
Below those arrays, we have the data structure $time(id)$: [*First*]{} and [*last*]{} store the first and last timestamps of ${\cal L}_{0,13}(id)$ that have data, and bitmap $T(id)$ has a bit for each timestamp in between. A bit 1 means no data for its timestamp.
The bottom of the figure shows the three bitmaps that represent $\Delta X(id)$. $X(id)_t$ has a bit for each bit set to 0 in $T(id)$, that is, for each position of $\Delta X(id)$ with a value. Each bit of $X(id)_t$ indicates whether the corresponding difference is positive or negative. For each bit of $X(id)_t$ set to 1, $X(id)_p$ stores that value in unary. $X(id)_n$ stores, in the same way, the negative differences.
Let us extract the $x$ coordinate at timestamp $9$. First, we obtain the number of disappearances until timestamp $9$: $dis=rank_1(T(id),i-first+1)=2$. Next, we obtain the number of positive and negative differences until timestamp $9$: $pos=rank_1(X(id)_t,i-dis-first+1)=4$ and $neg=i-dis-first-pos+1=2$. Finally, the $x$ coordinate is $select_1(X(id)_p,pos)-pos-(select_1(X_n,neg)-neg)= select_1(X(id)_p,4)-4-(select_1(X(id)_n,2)-2)$=$16-4-(5-2)=12-3=9$.
Indexing the logs
-----------------
Our representation yields constant-time extraction of whole trajectories and direct access to any point. To solve time-slice and time-interval queries, we may just compute the position or consecutive positions of the object and see if they fall within the query area. Although we can rapidly know the position of an object in a given timestamp, if we have to inspect all the timestamps of a given queried interval, we may spend much time obtaining positions that are outside the region of interest. In order to accelerate these queries over the logs, ContaCT stores an index for each ${\cal L}_{k,k+d}(id)$.
The index is a perfect binary tree that indexes the timestamps of the interval $[k+1,k+d-1]$ containing data (i.e., after being mapped with $T(id)$). Let $C$ indicate the number of timestamps covered by a leaf. Internal nodes cover the ranges covered by all the leaves in their subtree. Each node stores the MBR of the positions of the object during their covered interval of timestamps.
To check the positions of the object in the interval $[b,e]$, where $1 \leq b\leq e< d$, we first compute $b'=rank_0(T(id),b-first)$ and $e'=rank_0(T(id),e-first)$, and then check the timestamps of the tree in the range $[b',e']$. The way to use this tree is described in the next subsection.
![The index of a log.[]{data-label="index"}](images/index.eps)
**Example.** Figure \[index\] shows the index for the trajectory of Figure \[estructura\]. $C$ is 2, so the leaves cover at most 2 timestamps. In ${\cal L}_{0,13}(id)$, there are 7 time instants with values, at timestamps 2, 3, 4, 5, 8, 9, and 10. Therefore the leftmost leaf of the tree covers the positions at timestamps 2 and 3, the next leaf covers the timestamps 4 and 5, and so on.
The root contains the MBR that encloses all the positions in the interval covered by ${\cal L}_{0,13}(id)$. Since there are 7 timestamps with values, we labeled it $R_{1-7}$. At the top right, that MBR is drawn as a rectangle with a solid line. The left child, $R_{1-4}$, covers the positions of the first 4 timestamps. The right child, $R_{5-7}$, covers the timestamps from the $5^{th}$ to the $7^{th}$, and so on. The second-level MBRs are shown at the top right as rectangles with densely dotted lines, whereas the third level MBRs are drawn with scattered dotted lines.
Observe that each log stores the movements of one object between two snapshots, therefore there will be a considerable number of trees. To save space, we store the perfect trees in heap order, avoiding pointers. Each tree is then stored as two arrays, [*nodesX*]{} and [*nodesY*]{}, storing the extremes of the MBRs in each dimension. The children of a node at position $p$ are at $2p$ and $2p+1$.
Further, the arrays [*nodesX*]{} and [*nodesY*]{} are compressed by storing the values of the nodes below the root as differences with respect to their parent. For example, the values at position 2 (corresponding to $R_{1-4}$) of [*nodeX*]{} are stored as the values of the parent (2,10) minus the values at position 2 (2,5), that is, (0,5). As a result, the numbers are smaller, and we use $\lfloor \log m \rfloor+1$ bits for each number, being $m$ the largest difference (the root MBRs are stored separately).
Queries
-------
To answer a time-slice or a time-interval query, we use the closest previous snapshot to filter the objects that cannot possibly make it to the query region within the given time frame, by exploiting the maximum speed at which objects can move. Let $r = [x_1,x_2] \times [y_1,y_2]$ be a rectangular region in the two-dimensional space, and $b<e$ be two timestamps. Let $s$ be the maximum speed, in our dataset, of any object. We denote $ER(r,q)$, the *expanded region* of $r$ at timestamp $q$, the area that contains the points that must be considered from the preceding snapshot. If the timestamp of the preceding snapshot is $k$, then $ER(r,q) = [x_1-s\cdot(q-k),x_2+s\cdot(q-k)] \times [y_1-s\cdot(q-k),y_2+s\cdot(q-k)]$.
**Time-slice.** A time-slice query specifies a region $r$ and a timestamp $q$. Assume $q$ is between snapshots $Sp_k$ and $Sp_{k+d}$. We perform a range query on $Sp_k$ to retrieve all the objects $id$ in $ER(r,q)$. If $q=k$, we simply return all those objects $id$. Otherwise, we access the log $\mathcal{L}_{k,k+d}$ of each such object $id$ to find, in $O(1)$ time, its position at (local) time $q-k$, and report $id$ if the position is within $r$.
**Time-interval.** A time-interval query specifies a region $r$ and an interval $[b,e]$. It can be solved as a sequence $e'-b'+1$ time-slice queries (where $b'$ and $e'$ are described previously), but we exploit the tree of MBRs to speed up the query.
Each object that is within $ER(r,q)$ must be tracked along the timestamps $b$ to $e$, to determine if it has a position inside $r$. We compute $b'$ and $e'$ as described previously and use the MBR tree to quickly filter out the elements that do not qualify. We start at the tree root, and check if (1) the timestamps of the node intersect $[b',e']$ and (2) the root MBR intersects $r$. If not, we abandon the search at that node. Otherwise, we recursively enter its left and right children. When we reach a leaf, we extract all the positions one by one, looking for the first that falls within $r$. We develop specialized procedures to extract the next point faster than a random access in our Elias-Fano representation.
We further prune the search by continuously considering the maximum speed of the objects. Assume $[b',e']$ is within the right child of a node since the left one covers only $[b_1',e_1']$. If the minimum distance between the MBR of the left and $r$, along any coordinate, is $p > s \cdot(b-e_1)$, then there is no need to examine the right child. Here $e_1$ is the original timestamp corresponding to $e_1'$, which is obtained with $select_0(T(id),e_1')+first-1$. The same argument holds symmetrically with the left child. Finally, as we traverse the positions in a leaf, we verify this condition continuously to preempt the scan as soon as possible (we use a special “select-next” method on $T(id)$ to speed up consecutive $select$ queries).
**Example.** Let us run the time-interval query for the area $r=[4,5] \times [4,10]$ and (mapped) time interval $[b',e'] = [2,4]$ in the log of Figure \[index\]. We start at the root, which covers the time range $[1,7]$ and has MBR $[2,10]\times[4,10]$. Since both intersect the query, we continue. Since the tree is perfect, we know that the left subtree covers the timestamps $[1,4]$ and the right one covers $[5,7]$. Since the right child does not intersect the query time interval, we only descend by the left one, $R_{1,4}$. Its MBR is $[2,5] \times [4,7]$, which intersects $r$, so we continue. Its left child, $R_{1,2}$, covers the time interval $[1,2]$, which intersects $[b',e']$, so we enter it. However, its MBR is $[2,3] \times [4,6]$, which does not intersect $r$ and thus we abandon it. The right child of $R_{1,4}$, $R_{3,4}$, also intersects the time interval of the query. Its MBR is $[3,5] \times [5,7]$, which intersects $r$. Finally, since $R_{3,4}$ is a leaf, we access the $3^{rd}$ and $4^{th}$ positions in the log, finding that the object was in $r$ at time instant $4$.
Experimental Results
====================
ContaCT was coded in C++ and uses several data strucures of the SDSL library[@gbmp2014sea]. As baselines, we include GraCT and ScdcCT [@BrisaboaGNP16], also C++ programs, and the MVR-tree from the spatialindex library (`libspatialindex.github.io`). We used a real dataset storing the movements of $3{,}654$ ships on a grid of size $2{,}723 \times 367{,}775$ and $44{,}642$ time instants, whose plain representation requires 395.07 MB; we measure our compression ratios against that size. Appendix \[appendix:dataset\] gives more details on the dataset.
The experiments ran on an Intel^^ Core^TM^ i7-3820 CPU @ 3.60GHz (4 cores) with 10MB of cache and 64GB of RAM, over Ubuntu 12.04.5 LTS with kernel 3.2.0-115 (64 bits), using gcc 4.6.4 with `-O9`. We tested six types of queries:
- [*Object*]{} searches for the position of a specific object at a given timestamp.
- [*Trajectory*]{} returns the positions of an object between two timestamps.
- [*Slice S*]{} and [*Slice L*]{} are time-slice queries for small regions ($272 \times 367$ cells) and large regions ($2723 \times 3677$ cells), respectively.
- [*Interval S*]{} are time-interval queries specifying a small region on small intervall (36 timestamps), and [*Interval L*]{} are time-interval queries specifying large regions on large intervals (90 timestamps).
We measure elapsed times. Each data point averages 20[,]{}000 Object queries, 10[,]{}000 Trajectory queries, or 1[,]{}000 of Slice/Interval queries.
**Compressed representations**. We built ContaCT, ScdcCT and GraCT with different snapshot distances, namely every 120, 240, 360, and 720 timestamps. ContaCT was also built with different values of $C$ (the number of timestamps covered by the leaves of the MBR trees), specifically 20, 40, 80, 160, 320, and 640. We used Elias-Fano on the bitmaps $T(id)$, which were sparse, but turned to plain bitmaps to represent $X(id)$ and $Y(id)$, as they were not sufficiently sparse after mapping from $T(id)$.
Figure \[fig:size\] shows the size with the different settings. All the structure sizes decrease as the distances between snapshots increase, and ContaCT also decreases as $C$ increases. Thanks to its grammar-compression, the densest snapshot sampling of GraCT still uses 11% less space than the sparsest sampling of ContaCT. In turn, ContaCT is smaller than the other differentially compressed representation, ScdcCT, for example by 14% in their sparsest configurations.
Figure \[fig:time-ot-trajectory\] shows the average answer times for Object and Trajectory queries. ContaCT is especially fast on Object queries, thanks to its constant-time extraction mechanism. This makes it mostly independent of the snapshot sampling, and twice as fast as GraCT and three times faster than ScdcCT, even with their fastest configurations. GraCT is faster than ScdcCT, because it can traverse nonterminals of the grammar in constant time. For Trajectory queries, ContaCT is still faster by 20%. The difference decreases because sequential access to trajectories is not comparatively that slow with the other methods. The reason why some curves actually improve with a sparser snapshot sampling is that some extra work is needed when the query goes through various snapshots.
Figure \[fig:time-slice\] shows time-slice queries. The snapshot sampling is now crucial, since it affects the number of candidates that must be considered from the preceding snapshot (the computation of $ER(r,q)$). Since ContaCT can access the desired time instant in constant time, it is considerably faster than the others for a given snapshot sampling. However, GraCT matches ContaCT (and outperforms it for more selective queries) for a similar space usage, because GraCT can use a denser sampling thanks to its better compression of the log. ContaCT, on the other hand, outperforms ScdcCT by far.
Figure \[fig:time-interval\] shows time-interval queries, with various values of $C$ for ContaCT. Even with the nearly smallest-space configuration (snapshot interval 360, $C=160$), ContaCT outpeforms the largest GraCT configuration by a factor of 2, thanks to the MBR trees that index the logs. Using smaller $C$ values does not significantly improve the time, on the other hand, thanks to our optimized leaf traversal procedure. Once again, the baseline ScdcCT is much slower.
**Comparison with a spatio-temporal index.** We compare ContaCT with MVR-tree, a classic spatio-temporal index. We configured MVR-tree to run in main memory. To avoid space problems, we had to build the MVR-tree over a quarter of the input dataset. The size of the MVR-tree on this reduced input was 15.41 GB (including the data), while the maximum-space configuration of ContaCT uses 11.61 MB, three orders of magnitude less.
The MVR-tree can only solve time-slice and time-interval queries. We built ContaCT with different snapshot samplings and $C=80$. Figure \[fig:mvrtree1\] shows that our structure is faster on time time-interval queries, but slower on our time-slice queries. Figure \[fig:mvrtree2\] studies the turning point, by increasing the time span of time-interval queries, using the smallest-space configuration of ContaCT (snapshot period of 720). Note that MVR-tree times increase linearly whereas ContaCT stays essentially constant. ContaCT outperforms MVR-tree on interval lengths over 8 on large-region queries and over 14 in small-region ones.
Conclusions
===========
We have presented ContaCT, a structure to index trajectories of sets of moving objects in compressed form. ContaCT can efficiently retrieve points or segments of individual trajectories, and answer spatio-temporal range queries on the set of objects. ContaCT combines sampled two-dimensional snapshots compressed with $k^2$-trees, with logs differentially compressed and represented with Elias-Fano, which gives constant-time access to trajectory points. It also includes a hierarchical MBR mechanism that, combined with a pruning done on the snapshots, efficiently answers spatio-temporal queries.
Our experiments show that ContaCT compresses the data by a factor of almost 10 and outperforms by far, in space and time, a baseline alternative based on compressing small consecutive differences. ContaCT is also more than 1[,]{}000 times smaller than a classical spatio-temporal index, while being faster on all but very time-narrow queries. Compared with GraCT, the smallest existing representation based on grammar-compressing the trajectories, ContaCT uses more space. However, when both indexes are set to use the same amount of space, ContaCT generally makes better use of it, outperforming GraCT in most queries, by a factor of up to 3.
Future work involves extending ContaCT to more sophisticated queries, such as nearest-neighbor spatio-temporal queries.
Dataset details {#appendix:dataset}
===============
The dataset used in our experimental evaluation corresponds to a real dataset storing the movements of 3,654 boats sailing in the UTM Zone 10 during one month of 2014. It was obtained from MarineCadastre.[^2] Every position emitted by a ship is discretized into a matrix where the cell size is $50 \times 50$ meters. With this data normalization, we obtain a matrix with 1,001,451,325 cells, 2,723 in the $x$-axis and 367,775 in the $y$-axis. As our structure needs the position of the objects at regular timestamps, we preprocessed the signals every minute, sampling the time into 44,642 minutes in one month.
To filter out some obvious GPS errors, we set the maximum speed of our dataset to 55 cells per minute (over 234 km/h) and deleted every movement faster than this speed. In addition, we observe that most of the boats sent their positions frequently when they were moving, but not when they were stopped or moving slowly. This produced logs of boats with many small periods without signals (absence period). Taking into account that an object cannot move too far away during a small interval of time, we interpolated the signals when the absence period was smaller than 15 minutes, filling the periods of absence with these interpolated positions.
With these settings the original dataset occupies 974.43 MB in a plain text file with four columns: *object identifier*, *time instant*, *coordinate x* and *coordinate y*. Every value of these columns are stored as a string. However, to obtain a more precise compression measure, we represent this information in a binary file using two bytes to represent object identifiers (max value 3,653), two bytes for the instant column (max value 44,641), two bytes for the x-axis (max value 2,723) and three bytes for the y-axis (max value 367,775). Therefore, the binary representation of our dataset occupies 395.07MB.
[^1]:
[^2]: `http://marinecadastre.gov/ais/`
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Compressive sensing aims to recover a high-dimensional sparse signal from a relatively small number of measurements. In this paper, a novel design of the measurement matrix is proposed. The design is inspired by the construction of generalized low-density parity-check codes, where the capacity-achieving point-to-point codes serve as subcodes to robustly estimate the signal support. In the case that each entry of the $n$-dimensional $k$-sparse signal lies in a known discrete alphabet, the proposed scheme requires only $O(k \log n)$ measurements and arithmetic operations. In the case of arbitrary, possibly continuous alphabet, an error propagation graph is proposed to characterize the residual estimation error. With $O(k \log^2 n)$ measurements and computational complexity, the reconstruction error can be made arbitrarily small with high probability.'
author:
- |
Xu Chen and Dongning Guo\
Department of Electrical Engineering and Computer Science\
Northwestern University, Evanston, IL, 60208, USA [^1]
bibliography:
- 'IEEEabrv.bib'
- 'all\_bib.bib'
- 'xu\_bib.bib'
title: A GENERALIZED LDPC FRAMEWORK FOR ROBUST AND SUBLINEAR COMPRESSIVE SENSING
---
Introduction {#sec:intro}
============
Compressive sensing aims to recover a high-dimensional sparse signal from a relatively small number of measurements [@donoho2006compressed; @candes2006near]. There are two different designs of the measurement matrices: random construction and deterministic construction. Convex optimization approaches have been first proposed to recover the noiseless signal with $O(k \log (n/k))$ random measurements [@candes2005decoding]. Greedy algorithms that involve lower complexity have been proposed [@tropp2007signal; @needell2009cosamp; @donoho2012sparse]. However, most of the algorithms that are based on random measurement matrix design inevitably involve a complexity of ${\rm poly}(n)$.
Inspired by the error control code designs, deterministic structured measurement matrices have been proposed to reduce the computational complexity to (near) linear time $O(n)$ [@xu2007efficient; @indyk2008near]. In practice, when the signal dimension is many thousands or millions, even linear time complexity often becomes prohibitive. In response, sublinear compressive sensing based on second order Reed-Muller codes has been proposed, but the reconstruction error was not characterized [@applebaum2009chirp; @calderbank2010construction].
Recently, compressive sensing schemes with a novel design of measurement matrix and sublinear recovery algorithms have been developed, requiring $O(k)$ measurements and arithmetic operations under the noiseless setting [@pawar2012hybrid; @bakshi2012sho]. In those schemes, the measurements are split into multiple groups and each group is a sub-vector, which are linear combinations of the [*same*]{} set of signal components. Treating the measurement groups as bins, the design matrix basically hashes the signals to different measurement bins, which is similar to the bipartite graph induced by low-density parity-check (LDPC) code structure. In [@pawar2012hybrid] and [@bakshi2012sho], the measurement vector in each bin is designed to carry the signal support information by leveraging the discrete Fourier transform (DFT) matrix. The design has been extended to the noisy case, involving $O \left( k \log^{1.3} n \right)$ measurements and computational complexity, with the limitation that the signal entries must lie in a known discrete alphabet. In this paper, we propose a generalized LDPC code inspired compressive sensing scheme to further reduce the the number of measurements required and computation complexity. The scheme adopts the sublinear recovery algorithm framework in [@pawar2012hybrid]. For the measurement matrix design, the scheme also adopts the LDPC structure to disperse the signal into measurement bins. The main difference is that each measurement bin is a subcode, where some recently developed capacity-achieving codes are utilized to encode the signal support.
Specifically, this paper makes the following contributions. First, our scheme is the first to achieve nearly order optimal $O(k \log n)$ noisy measurements and computational complexity for the case of known discrete alphabet. Second, the previous design based on DFT matrix is susceptible to quantization errors, while the proposed measurement matrix consists of only $\{0,\pm 1\}$ entries, which are easier and more robust in practice. Third, we propose an error propagation graph with error message passing rules to capture the error propagation for the case of arbitrary signals with unknown alphabet. Analysis shows that with $O \left( k \log^2 n \right)$ measurements and complexity the signal estimation error can be made arbitrarily small as $k$ increase. The proposed design and error propagation graph have potential applications in sparse Fourier transform [@pawar2013computing] and Walsh-Hadamard transform with arbitrary signal alphabet [@chen2015robust].
System Model
============
Suppose $\bx \in {\mathbb{R}}^n$ is a $k$-sparse vector. The problem is to recover $\bx$ from the $m$-dimensional ($m \ll n$) measurement vector $$\begin{aligned}
\label{eq:system}
\by = \bA \bx + \bz\end{aligned}$$ where $\bA \in \mathbb{R}^{m \times n}$ is the measurement matrix and $\bz$ is the noise vector with each entry being independently and identically distributed (i.i.d.) Gaussian variables with zero mean and variance $\sigma^2$.
Throughout the paper, we use bold capital letter and bold normal letter to denote a matrix and a vector, respectively. Given a matrix $\bA$, $A_{ij}$ denotes the entry located at the $i$-th row and $j$-th column, and $\ba_i$ denotes the $i$-th column. Given $i \in \{0, \cdots n-1 \}$, $(i)_2$ is the $\log n$-bit binary representation of $i$ with $0$ and $1$ mapped to $1$ and $-1$, respectively. For example, $n=3$, $(2)_2 = [1, -1, 1]$. Let $sgn(x) =1$ if $x \geq 0$ and $sgn(x)=-1$ otherwise.
Measurement Matrix Design
=========================
The LDPC inspired design of the measurement matrix is proposed in [@pawar2012hybrid; @bakshi2012sho]. In particular, the measurement matrix is constructed as $$\begin{aligned}
\label{eq:meas_matrix}
\bA = \bH \odot \bG\end{aligned}$$ where $\bH \in \{0,1\}^{b \times n}$, $\bG \in \mathbb{R}^{c \times n}$ and the $\odot$ operator is defined as $$\begin{aligned}
\bH \odot \bG = \left[
\begin{array}{ccc}
H_{0,0} \bg_0 & \cdots & H_{0,(n-1)} \bg_{n-1} \\
\vdots & \cdots & \vdots \\
H_{b-1,0} \bg_0 & \cdots & H_{b-1,(n-1)} \bg_{n-1}
\end{array}
\right].\end{aligned}$$ The number of measurements is thus $m = b \times c$. For example,
$$\begin{aligned}
\begin{bmatrix}
1 & 0 & 1 \\
0 & 1 & 1
\end{bmatrix}
\odot
\begin{bmatrix}
\bg_0 & \bg_1 & \bg_2
\end{bmatrix}
=
\begin{bmatrix}
\bg_0 & 0 & \bg_2 \\
0 & \bg_1 & \bg_2
\end{bmatrix}.\end{aligned}$$
In fact, $\bH$ is inspired by the parity-check matrix of LDPC codes. The relationship between the signal entries and the measurements can be represented by a bipartite graph. In the bipartite graph, there are $n$ left nodes with $x_i$ corresponding to the $i$-th left nodes, and $b$ right nodes, which are also referred to as bins. The $i$-th left node is connected with the $j$-th bin if $H_{ij} = 1$. The measurement vector is thus grouped into $b$ sub-vectors as $\by = \left[ \by_0^{\dagger}, \cdots, \by_{b-1}^{\dagger} \right]^{\dagger}$, where $\by_j \in \mathbb{R}^c$ is the $j$-th bin value given by $$\begin{aligned}
\by_j = \sum_{i=0}^{n-1} H_{ij} x_i \bg_i + \bz_j.\end{aligned}$$ Fig. \[fig:bipartite\_graph\] illustrates the bipartite representation between signals and measurements. In this paper, we construct $\bH$ from the ensemble of left $d$-regular bipartite graph $\mathcal{G}_d(k,b)$, where every signal is connected to $d$ measurement bins uniformly at random.
![Example of the bipartite graph. Left nodes correspond to signals and right nodes correspond to measurement bins. The left nodes marked in red are nonzero signal components.[]{data-label="fig:bipartite_graph"}](fig_bipartite_graph.eps "fig:"){width="2in"}\
The recovery algorithms adopts the framework proposed in [@li2014sub]. The recovery algorithm calls for a robust bin detection, which can 1) identify if a measurement bin is connected to no nonzero signal component (zeroton), to a single nonzero component (singleton) or to multiple nonzero components (multiton); 2) robustly estimate the signal index and value from singleton bins. It can be proved that by some proper $b = O(k)$, the recovery algorithm can correctly estimate $\bx$ with high probability if we have a robust bin detection. The key challenge is how to design $\bG$ to achieve robust bin detection.
In previous works [@bakshi2012sho; @pawar2012hybrid; @li2014sub], $\bG$ is constructed based on the DFT matrix. The signal index information $i$ is embedded in the phase difference between the entries of $\bg_i$. In this paper, we propose a new design of $\bG$, which only consists of $\{\pm 1\}$ entries and is more robust to noise and quantization errors.
We motivate the design using a simplified setting. Assume 1) a measurement bin $j$ is known to be a singleton, 2) there is no noise, and 3) the sign of the signal $x_i$ that is hashed to bin $j$ is known. The question is how can we design $\bG$ to detect the signal index $i$ and its value? Let $(i)_2$ be the $\log n$-bit binary representation of $i$. If $\bg_i = (i)_2$, then the signal index can be easily recovered based on the signs of each entry in $sgn(x_i) \by_j = | x_i | \bg_i$. A robust design of $\bG$ is to overcome the challenges posed by the three assumptions. First, we let $\bar{\bg}_i$ to be an all-one vector such that the signs of $x_i$ can be estimated. Second, $\tilde{\bg}$ is designed to be coded bits of $(i)_2$ for robust estimation of $(i)_2$ under the noisy setting. The sub-vector length is $\lceil \log n \rceil / R$, where $R$ is the code rate of the applied low-complexity error-control code [@barg2004error]. Third, we let $\dot{\bg}_i$ be a binary vector with each entry generated according to i.i.d. Rademacher distribution for singleton verification.
In all, the $i$-th column of $\bG$ consists of three sub-vectors: $$\begin{aligned}
\bg_i = \left[\tilde{\bg}_i^{\dagger}, \bar{\bg}_i^{\dagger}, \dot{\bg}_i^{\dagger} \right]^{\dagger}\end{aligned}$$ where $\tilde{\bg}_i \in \{\pm 1 \}^{c_0}$, $\bar{\bg}_i \in \{\pm 1 \}^{c_1}$, and $\dot{\bg}_i \in \{\pm 1\}^{c_2}$. Accordingly, the measurement vector for bin $j$ can be split into three sub-vectors: $$\begin{aligned}
\by_j &=
\left[
\begin{array}{cc}
\tilde{\by}_j \\
\bar{\by}_j \\
\dot{\by}_j
\end{array}
\right]
=
\left[
\begin{array}{cc}
\sum_{i=0}^{n-1} H_{ij} x_i \tilde{\bg}_i \\
\sum_{i=0}^{n-1} H_{ij} x_i \bar{\bg}_i \\
\sum_{i=0}^{n-1} H_{ij} x_i \dot{\bg}_i
\end{array}
\right] +
\left[
\begin{array}{cc}
\tilde{\bz}_j \\
\bar{\bz}_j\\
\dot{\bz}_j
\end{array}
\right] .\end{aligned}$$ In our design, we choose $b = O(k)$, $c = O(\log n)$ and $c= O(\log^2 n)$ for signals with known discrete alphabet and arbitrary alphabet, respectively.
In the bipartite graph, each measurement bin can be regarded as a super check node where a subcode is further used to encode the index information of the signals. The structure is similar to that of generalized LDPC codes [@miladinovic2008generalized]. The well-established low-complexity capacity-approaching point-to-point codes can serve as subcodes to enhance the robust design.
Recovery Algorithm Design {#sec:algo}
=========================
We adopt the recovery algorithm framework proposed in [@pawar2012hybrid]. The algorithm is implemented in an iterative “peeling" process. In every iteration, a singleton bin is identified. The index and value of the signal that is hashed to the singleton bin are estimated. The contribution of the estimated signal to the other connected bins are cancelled out (peeled off).
The main difference of our work lies in the signal support estimation from a singleton bin, referred to as the [*singleton test*]{}, which is described in Algorithm \[algo:bin\_detect\]. In particular, for some $x_i$ that is hashed to a singleton bin $j$, $\tilde{\bg}_i \in \{ \pm 1\}^{c_0}$ encodes the support information $(i)_2$. Suppose the signal sign estimation is correct, i.e., $s = sgn(x_i)$. Without noise, $s \tilde{\by}_j = |x_i| \tilde{\bg}_i$ and thus $sgn(s \tilde{\by}_i)$ is exactly $\tilde{\bg}_i$. Under the noisy setting, some of the signs are flipped due to noise, which can be regarded as transmission over the binary symmetric channel (BSC). With low-complexity codes used as subcodes, $(i)_2$ can be recovered by inputting $sgn(s \tilde{\by}_i)$ to the corresponding decoder with complexity $O( c_0)$ [@barg2004error].
The overall recovery algorithm is described as follows.
First, run the singleton test on every bin using Algorithm \[algo:bin\_detect\]. Let $L$ denote the set of estimated signal indices. Remove the declared singleton bins.
Then, repeat the following until $L = \emptyset$:
Select arbitrary $i \in L$ and remove $i$ from set $L$.
For every remaining bin $j$ with $H_{ij} = 1$, perform the following:
Subtract the signal node $i$ value from bin $j$: $\by_j \leftarrow \by_j - \hat{x}_i \bg_i$.
Run the singleton test on the bin using Algorithm \[algo:bin\_detect\]. If it is a singleton, add the output index to $L$ and remove bin $j$.
Algorithm \[algo:bin\_detect\] has a computational complexity of $O(c)$, where $c = c_0 + c_1 + c_2$. Performing the singleton test on all $b$ bin takes complexity $O(bc)$. In each subsequent iteration, we perform Algorithm \[algo:bin\_detect\] only on every (remaining) connected bin of a recovered signal component. Since the left-node degree is constant, each iteration involves computational complexity of $O(c)$. It will be proved that the algorithm terminates after $O(k)$ iterations with high probability. The computational complexity of all the iterations involved is thus $O(k c)$. With the choice of $b$ and $c$, the total complexity is $O( k \log n)$ and $O \left( k \log^2 n \right)$ for signals with discrete alphabet and arbitrary alphabet, respectively.
**Input**: Bin measurements, $\by = [\tilde{\by}^{\dagger} \enskip \bar{\by}^{\dagger} \enskip \dot{\by}^{\dagger}]^{\dagger}$. **Output**: index and estimate, $(i,z_i)$. Claim bin is a zeroton and return $(i,z_i) \leftarrow (\emptyset, 0)$. *Signal sign estimation:* $
s \leftarrow sgn (\bar{\bg}^{\dagger} \bar{\by}).
$ *Signal index estimation:* $
i \leftarrow \text{BSC-Decoder} ( sgn( s \cdot \tilde{\by} ) ).
$ *Singleton verification:* $z' \leftarrow \frac{1}{c_2} \dot{\bg}_{i}^{\dagger} \dot{\by}$. $$\begin{aligned}
z_i & \leftarrow \left\{
\begin{array}{cc}
\frac{1}{c} \bg_i^{\dagger} \by & \text{ arbitrary alphabet}\\
\arg\min_{z' \in \mathcal{X}} ||\dot{\by} - z' \dot{g_{i}}||^2 & \text{discrete alphabet}
\end{array}
\right.\end{aligned}$$
Main Results and Proof {#sec:proof}
======================
\[thm:cs\_known\] Given any $\epsilon >0$, there exists $k_0>0$ such that for every $k > k_0$ and every $n$-dimensional $k$-sparse signal $\bx$ whose entries take their values in a known discrete alphabet, the proposed scheme achieves $\mathsf{P} \{\hat{\bx} \neq \bx \} < \epsilon$. The number of measurements required is $O(k\log n)$. The computational complexity is $O(k\log n)$ arithmetic operations.
\[thm:cs\_general\] Given any $\delta, \epsilon >0$, there exists $k_0>0$ such that for every $k > k_0$ and every $n$-dimensional $k$-sparse signal $\bx$ with $|x_i| \geq
\delta$ for every $i \in {\rm supp} (\bx)$, the proposed scheme achieves $\mathsf{P}
\{ {\rm supp}(\hat{\bx}) \neq {\rm supp}(\bx) \} < \epsilon$ and $\mathsf{P} \{ |\hat{x}_i - x_i|^2 \geq \epsilon \} < \epsilon$ for every $i \in {\rm supp}(\bx) $. The number of measurements required is $O(k\log^2 n)$. The computational complexity is $O(k\log^2 n)$ arithmetic operations.
We focus on the proof of Theorem \[thm:cs\_general\] due to space limitations. Theorem \[thm:cs\_known\] follows as a special case. Unlike signals from discrete alphabet, the signal estimates have residual errors, which propagate to later iterations due to the peeling process. In this paper, we propose an *error propagation graph* to keep track of the accumulated errors.
An error propagation graph for $x_i$ is a subgraph induced by the recovery algorithm, which contains the signal nodes that are estimated in the previous iterations and have paths to $x_i$. Fig. \[fig:err\_prop\] illustrates the the error propagation graph for $x_2$.
![Error propagation graph for signal $x_2$.[]{data-label="fig:err_prop"}](fig_error_prop "fig:"){width="6cm"}\
Define the estimation error of $x_i$ as $$\begin{aligned}
p_i &= x_i - \hat{x}_i.\end{aligned}$$ Let $m(i)$ be the measurement bin used to recover the signal index $i$. Define the point error of $x_i$ as $$\begin{aligned}
\label{eq:res_err}
e_i = - c^{-1} {\bg}_i^{\dagger} {\bz}_{m(i)} .\end{aligned}$$ Then $e_i$ is a Gaussian variable with zero mean and variance $\sigma^2/ c$. We will keep track of $p_i$ using the error propagation graph.
Let $S(t)$ denote the signal indices that are recovered in the $t$-th iteration. Consider the estimation of $x_i$, $i \in S(1)$. The measurement vector of bin $m(i)$ and the residual estimation error are given by $$\begin{aligned}
\by_{m(i)} &= x_i \bg_i + \bz_{m(i)} \\
p_i &= e_i.\end{aligned}$$
Consider the estimation for $x_i$, $i \in S(2)$. With the peeling of $\hat{x}_\ell$, $\ell \in S(1)$, the updated measurement vector of $m(i)$ and the estimation error become $$\begin{aligned}
\by_{m(i)} &= x_i \bg_i + \bz_{m(i)} + \sum_{\ell \in S(1): H_{\ell, {m(i)}} = 1} e_{\ell} \bg_{\ell} \\
p_i &= e_i + \sum_{\ell \in S(1): H_{\ell, {m(i)}} = 1} e_{\ell} \left(- c^{-1} \bg_i^{\dagger} \bg_{\ell} \right),\end{aligned}$$ where $|\bg_i^{\dagger} \bg_{\ell} /c| \leq 1$ for every realization of $\bG$.
The estimation error can be calculated recursively according to some message passing rules over the graph. In particular, let $p_i$ be the estimation error propagated from signal node $i$ and $\bq_j \in \mathbb{R}^{c}$ be the error vector propagated from the measurement bin $j$. The errors can be calculated according to the following rules: $$\begin{aligned}
\label{eq:err_prop_rule1} p_i &= e_i + \sum_{j \in \text{in}(i) } \left( - c^{-1} \bg_i^{\dagger} \bq_j \right) \\
\label{eq:err_prop_rule2} \bq_j &= \sum_{i \in \text{in}(j)} p_i \bg_i,\end{aligned}$$ where $\text{in}(i)$ denotes the indices of the measurement bins (signal nodes) incoming to signal node (measurement bin) $i$.
By induction and the error message passing rules and , the error propagation effect is characterized by the following lemma.
The estimation error of $x_i$, $i \in S(t)$, is calculated as $$\begin{aligned}
p_i & = e_i + \sum_{\ell \in \cup_{j =1}^{t-1} S(j) \cap D(i)} \sum_{p = 1}^{P(\ell, i) } e_{\ell} d_{\ell,p},\end{aligned}$$ where $D(i)$ be the connected subgraph of the bipartite graph containing $i$, $\mathcal{P} (\ell, i)$ is the number of paths from $\ell$ to $i$ in $D(i)$, and $d_{\ell, p}$ is some coefficient depending on both $\bG$ and the path satisfying $|d_{\ell,p}| \leq 1$.
Fig. \[fig:err\_prop\] illustrates an example. The number of paths from $x_0$ to $x_2$ is $\mathcal{P}(0,2) = 2$, with the corresponding coefficients being $d_{0,1} = - c^{-1} \bg_2^{\dagger} \bg_0 $ and $d_{0,2} = c^{-2} \bg_1^{\dagger} \bg_0 \bg_2^{\dagger} \bg_1 $. The number of paths from $x_1$ to $x_2$ is $\mathcal{P}(1,2) = 1$, with the coefficients being $d_{1,1} = - c^{-1} \bg_2^{\dagger} \bg_1$.
We further bound the errors by leveraging results on random hypergraph. The bipartite graph induced by $\bH$ corresponds to a hypergraph where the left nodes and right nodes represent hyperedges and vertices. The hyperedge $i$ is incident on vertex $j$ if $H_{ij} = 1$. Then the random bipartite graph $\mathcal{G}_d (k,b)$ induces a $d$-uniform random hypergraph.
[@karonski2002phase]\[lemma:hypergraph\_component\] Suppose $b/k$ is some constant large enough, then with probability $1- O(1/k)$, $\mathcal{G}_d (k,b)$ contains only trees or unicyclic components, and the largest component contains $O( \log k)$ signal nodes.
Let $\mathcal{E}_H$ denote the event that the bipartite graph satisfies the condition as described in Lemma \[lemma:hypergraph\_component\] with $P \{\mathcal{E}_H\} =1- O(1/k)$. We first bound the detection error probability conditioned on $\mathcal{E}_H$. Suppose $\mathcal{E}_H$ holds, then $\mathcal{P}(i,j) \leq 2$, otherwise the component is not unicyclic. Moreover the largest component $D(j)$ contains $O(\log k)$ signal nodes. Therefore, conditioned on $\bG$, $p_i$ is Gaussian distributed with zero mean and the variance of can be upper bounded as $$\begin{aligned}
\label{eq:var_ub} {\rm var} (p_i) &\leq \left( 1+ \sum_{\ell \in D(i)} P^2(\ell,j) \right) \frac{\sigma^2}{c} = O\left( \log k \frac{\sigma^2}{c} \right).\end{aligned}$$
\[lemma:singleton\_estimate\] Conditioned on that $\mathcal{E}_H$ holds, given any $\delta >0$ and $|x_i| \geq \delta$, $\forall i \in {\rm supp} (\bx)$, the recovery algorithm can correctly identity the signal support with probability $1 - O(1 / n)$ with some $c = O(\log^2 n)$.
The support detection may be subject to zeroton, multiton and singleton detection errors. The error probability of detecting zerotons and multitons can be upper bounded by $O(1/n)$ following similar steps in [@li2014sub] and is omitted due to space limitation. We focus on the singleton detection.
Suppose the measurement vector of a singleton is given by $$\begin{aligned}
\by_j = x_i \bg_i + \bw_j,\end{aligned}$$ where the entries in $\bw_j$ are i.i.d. Gaussian variable with zero mean and variance $\tilde{\sigma}^2$. Then the error probability of sign estimation is calculated as $$\begin{aligned}
~\label{eq:err_sign}
\mathsf{P} \{sgn (\mathbf{1}^{\dagger} \bar{\by}) \neq sgn(x_i) \} &= \mathsf{P} \{ \mathbf{1}^{\dagger} \bar{\bw} / c_1 \geq |x_i| \} \\
&= Q( \sqrt{c_1} |x_i| / \tilde{\sigma}),\end{aligned}$$ where $Q(x)$ is the $Q$-function for standard normal distribution.
Suppose the signs of $x_i$ is correctly detected and we want to detect $(i)_2$ by recovering the signs of $\tilde{\bg}_i$. By compensating the signs of $x_i$ as $ sgn(x_i) \by_j$, the random transformation $$\begin{aligned}
sgn (|x_i| \tilde{\bg}_i ) \to sgn( |x_i| \tilde{\bg}_i + \tilde{\bw}_j )\end{aligned}$$ is equivalent to transmission over a BSC with crossover probability less than $Q(|x_i| / \tilde{\sigma})$ [@chen2015robust].
From the recovery process, $\bw_j$ is the noise plus residual estimate errors given by $\bw_j = \bz_j + \sum_{\ell \in \text{in}(j)} p_{\ell} \bg_{\ell}$. According to , for some $c = O \left( \log^2 n \right)$ and a large enough $n$, the variance of $\bw_j$ is dominated by that of $\bz_j$. The entries of $\bw_j$ have a variance $\tilde{\sigma}^2$ bounded by some constant. Therefore, given that $|x_i| \geq \delta$ for some $\delta$, the worst-case SNR for every singleton estimation is lower bounded by some constant. The error probability of sign estimation is $O(1/n^3)$ with some $c_1 = O(\log n)$. Applying an error control code of length $c_0 = \log n /R$ with a low enough code rate $R$, $(i)_2$ can be decoded correctly with probability $1- O(1/n^3)$.
Note that if $\mathcal{E}_H$ holds and the singleton, multiton and zeroton bins are correctly estimated, the peeling decoder terminates by recovering every nonzero signal entry [@price2011efficient]. Since there are at most $O(k)$ iterations and every iteration involves at most $O(k)$ singletons, the error probability can be upper bounded by $O(k^2 / n^3) = O(1/n)$ using the union bound. Moreover, conditioned on that $i$ is correctly estimated, the probability that the singleton verification is not passed is equivalent to a zeroton detection error, which can be upper bounded by $O(1/n)$. The lemma is hence established.
Support recovery fails only if either $\mathcal{E}_H$ does not hold or that a bin detection error occurs conditioned on $\mathcal{E}_H$ holds. By Lemma \[lemma:hypergraph\_component\] and Lemma \[lemma:singleton\_estimate\], the overall error probability of support recovery is $O \left( \frac{1}{n} + \frac{1}{k} \right)$, vanishing as $k$ increases.
The error probability can be upper bounded as $$\begin{aligned}
\mathsf{P} \left\{ |x_i - \hat{x_i}|^2 \geq \epsilon \right\} & \leq \mathsf{P} \left\{ \mathcal{E}_H^c \right\} + \mathsf{P} \left\{ | p_i | \geq \sqrt{\epsilon} | \mathcal{E}_H \right\} \\
\label{eq:mse} & \leq \mathsf{P} \left\{ \mathcal{E}_H^c \right\} + 2
Q \left( \sqrt{ \frac{\epsilon c}{\log k \sigma^2} }
\right) $$ where follows because $p_i$ is Gaussian variable with zero mean and variance upper bounded by conditioned on $\mathcal{E}_H$ and every realization of $\bG$. By Lemma \[lemma:hypergraph\_component\], the error probability is smaller than any $\epsilon$ with a large enough $k$ and some $c = O \left( \log^2 n \right)$. Hence, Theorem \[thm:cs\_general\] is established.
Simulation
==========
Throughout the simulation, we assume that the nonzero signal amplitude is taken uniformly at random from $[1, 10]$ and define SNR = $1/\sigma^2$, which is the worst-case SNR. The signal dimension is $n = 10^{10}$. The number of measurement bins is chosen to be $b = 3 k$. We adopt a regular random LDPC code with rate $1/2$ as subcode to encode the signal support information, and thus $c_0 = 2 \log n$. We let $c_1 = \log n$ and $c_2 = 2 \log n$. Fig. \[fig:errprob\] and Fig. \[fig:rmse\] plot the error probability of support recovery and relative mean square error, respectively. The relative mean square error is only calculated and averaged over the signals with their support correctly estimated. We run 200 simulations for each SNR. In the simulation, for every sparsity level $k$, the error-control code and nonzero signal entries are generated once and fixed.
Although analysis shows that $c = O \left( \log^2 n \right)$ is sufficient to guarantee vanishing error probability, choosing $c = O(\log n)$ also gives a good performance. The error probability of support recovery and relative mean square error decreases as SNR increases. In order to achieve more reliable signal recovery, we can adopt a more sophisticated error-control code or a code with lower code rate.
![Error probability of support recovery.[]{data-label="fig:errprob"}](fig_errprob "fig:"){width="6cm"}\
![Relative mean square error.[]{data-label="fig:rmse"}](fig_rmse "fig:"){width="6cm"}\
[^1]: This work was supported in part by the National Science Foundation under Grant No. CCF-1423040.
|
{
"pile_set_name": "ArXiv"
}
|
1=1
**Structure of Tate-Shafarevich groups of**
**elliptic curves over global function fields**
0.2in
**M.L. Brown**
0.2in
Institut Fourier, B.P. 74, 38402 Saint Martin d’Hères, France
0.4in
**Shortened Title**
Tate-Shafarevich Groups
0.4in
**Abstract**
The structure of the Tate-Shafarevich groups of a class of elliptic curves over global function fields is determined. These are known to be finite abelian groups from the monograph \[1\] and hence they are direct sums of finite cyclic groups where the orders of these cyclic components are invariants of the Tate-Shafarevich group. This decomposition of the Tate-Shafarevich groups into direct sums of finite cyclic groups depends on the behaviour of Drinfeld-Heegner points on these elliptic curves. These are points analogous to Heegner points on elliptic curves over the rational numbers.
0.2in 0.4in
**AMS 2012 Mathematics Subject Classification Numbers**
11G05, 11G09, 11G20, 11G40, 14G10, 14G17, 14G25, 14H52
0.4in
**Keywords: Elliptic curves, Tate-Shafarevich groups, Function fields**
0.4in
**Table of Contents**
0.2in
Chapter 1. Preliminaries
1.1. Introduction
1.2. Global fields of positive characteristic
1.3. Orders in imaginary quadratic fields
1.4. Ring class fields
1.5. Elliptic curves over global fields of positive characteristic
1.6. The Drinfeld modular curve $X_0^{\rm Drin}(I)$
1.7. Analogue for $F$ of the Shimura-Taniyama-Weil conjecture
1.8. Drinfeld-Heegner points
1.9. Groups and cohomology
1.10. Torsion on elliptic curves $E/F$
1.11. Igusa’s theorem
1.12 Consequences of Igusa’s theorem
Chapter 2. Local duality, Cassels pairings, Tate-Shafarevich groups
2.1. Local duality of elliptic curves
2.2. Selmer groups and Tate-Shafarevich groups
2.3. The Cassels pairing
Chapter 3. The cohomology classes $\gamma_n(c),\delta_n(c)$
3.1. The set $\cal P$ of prime numbers
3.2. Frobenius elements and the set $\Lambda(n)$ of divisors
3.3. A refined Hasse principle for finite group schemes
3.4. Drinfeld-Heegner points and the cohomology classes $\gamma_n(c)$, $\delta_n(c)$
Chapter 4. Structure of the Tate-Shafarevich group and the Selmer group
4.1. Statement of the main theorems
4.2. Cochains for the the cohomology classes $\gamma_n(c), \delta_n(c)$
4.3. Points $P_c$ defined over local fields
4.4. The map $\chi_z$
4.5. Localizations of the classes $\gamma_n(c)$ and $\delta_n(c)$
4.6. The Cassels pairing with a class $\delta_n(c)$
Chapter 5. Construction of cohomology classes and proofs of the main theorems
5.1. $M_r$ is finite for some $r$
5.2. A class $\gamma_n(c)$ in the Selmer group
5.3. Proof of Theorem 4.1.15
5.4. Proof of Theorems 4.1.9 and 4.1.13
5.5. Proof of Theorems 1.1.1 and 4.1.14
5.6. Generators of Tate-Shafarevich groups
0.4in [**Chapter 1. Preliminaries**]{}
0.2in
[**1.1. Introduction**]{} 1 by 1
0.2in Let $F$ be a global field of positive characteristic $p>0$. Let $E/F$ be an elliptic curve with an origin, that is to say a 1-dimensional abelian variety.
In \[1\] it is shown that for a class of these elliptic curves $E/F$, the Tate-Shafarevich group ${\textstyle\coprod\!\!\!\coprod}(E/F)$ is finite and for prime numbers $l$ belonging to a set $\cal S$ of prime numbers given by arithmetic conditions then the $l$-primary component ${\textstyle\coprod\!\!\!\coprod}(E/F)_{l^\infty}$ has order which is explicitly bounded.
In this paper, we determine the structure of the finite abelian group ${\textstyle\coprod\!\!\!\coprod}(E/F)_{l^\infty}$ for the same class of elliptic curves and for all $l$ in the same set of prime numbers $\cal S$ (in the notation of Theorem 4.1.9 below $\cal S$ is the set $\cal P$ with the exclusion of the prime divisors of the order of the Picard group Pic$(A)$). We also determine the structure of the Selmer groups of the elliptic curves in question.
Let $E/F$ be an elliptic curve and let $K$ be an imaginary quadratic extension of $F$ with respect to the place $\infty$ of $F$ (see §1.2). Let Spec $A$ be the non-singular affine curve with function field $F$ and whose point at infinity is $\infty$ (see §1.2). Assume that $E,K,\infty$ satisfy the hypotheses (a), (b), (c) of (4.1.1) in chapter 4. This provides an infinite set of prime numbers $\cal P$ of positive Dirichlet density and defined by arithmetic conditions (see §3.1). Indeed, $\cal P$ contains all except finitely many prime numbers $l\in {\mathds Z}$ of the form $2^sn+1$ where $s\geq 1$ and $n$ is odd such that $q$ is a $2^s$th power non-residue modulo $l$ where $q$ is the order of the exact finite field of constants of $F$.
Fix a prime number $l\in \cal P$. There are sets of divisors $\Lambda^r(n)$, relative to $l$, on $F$ for all integers $n\geq 1$ such that each divisor in $\Lambda^r(n)$ is a sum of $r$ distinct prime divisors and there is a decreasing filtration on $\Lambda^r(1)$. $$\Lambda^r=\Lambda^r(1)\supseteq\Lambda^r(2)\supseteq\ldots$$ For any divisor $c\in\Lambda^r(n)$ there is a corresponding Drinfeld-Heegner point $P_c$ of $E(K[c])$, the group of $K[c]$-rational points of $E$, where $K[c]$ is the ring class field of $K$ with conductor $c$ (see §1.4 and (3.4.9)).
On $E(K[c])$ there is the decreasing $l$-adic filtration $$E(K[c])\supseteq lE(K[c])\supseteq l^2E(K[c])\supseteq\ldots$$ Define $$M_r=\min_{c\in \Lambda^r}(\max (n \in {\mathds N}\ \vert P_c\in l^nE(K[c])))
{\rm \ \ for \ all \ integers \ }r\in {\mathds N}.$$
If the point $P_0\in E(K)$ has infinite order in the group of $K$-rational points $E(K)$ then it can be shown that $M_0,M_1,\ldots$ is a decreasing sequence of non-negative integers (see Lemma 5.1.4). One of the main results of this paper is the following. 0.2in [**1.1.1. Theorem.**]{} [*Suppose that $P_0$ has infinite order in $E(K)$, the group of $K$-rational points of $E$. Let $l$ be a prime number in $\cal P$ and which is coprime to the order of the Picard group of the affine curve ${\rm Spec } \ A$. Let $\epsilon=\pm1$ be the sign in the functional equation of the $L$-function of $E/F$. Then the Tate-Shafarevich group ${\coprod\!\!\!\coprod}(E/F)$ of $E/F$ is finite and its $l$-primary component is given by $${\coprod\!\!\!\!\coprod}(E/F)_{l^\infty}\cong \prod_{ (-1)^i=\epsilon\atop i\geq 0}
({\mathds Z}/l^{M_i-M_{i+1}}{\mathds Z})^2$$ where the product runs over integers $i\in {\mathds N}$ such that $(-1)^i=\epsilon$.*]{} 0.2in A similar statement holds for the Tate-Shafarevich group of the elliptic curve $E\times_FK$ over $K$ (see Theorem 4.1.9) as well as the Selmer groups of these curves (corollary 4.1.19). The main results of this paper are stated in §4.1.
It may be conjectured that for every global field $F$ of characteristic $>0$ there are infinitely many non-isomorphic elliptic curves $E/F$ and infinitely many imaginary quadratic field extensions $K/F$ such that $E,K,\infty$ satisfy the hypotheses of this Theorem 1.1.1 and those of §4.1. If this conjecture holds, then the above theorem and those of §4.1 give infinitely many non-isomorphic elliptic curves over a given global field of positive characteristic whose $l$-primary components of the Tate-Shafarevich group are structurally known for infinitely many prime numbers $l$ satisfying arithmetic conditions.
The method of this paper is related to that of Kolyvagin’s determination of the structure of Tate-Shafarevich groups of a class of elliptic curves over the rational numbers (see \[6\] and \[8\]).
The proofs of the main theorems of this paper stated in §4.1 and Theorem 1.1.1 above require many preliminary results which are explained in chapters 1-5.
Chapter 2 contain basics on Tate local duality, Selmer groups, and the Cassels pairing on Tate-Shafarevich groups. In section 3.1, the set $\cal P$ of prime numbers is defined by arithmetic conditions. In section §3.2, the sets $\Lambda^r(n)$ of divisors on the global field $F$ are defined. Sections §§3.3,3.4 constructs the cohomology classes $\gamma_n(c),\delta_n(c)$ in the cohomology of the elliptic curve $E/F$.
In section §4.1., the main results of this paper are stated which are then proved in §§5.3-5.5 after some further properties of $\gamma_n(c),\delta_n(c)$ are proved in Chapters 4 and 5. We show in particular that the cohomology classes $\delta_n(c)$ define characters, via the Cassels pairing on ${\textstyle\coprod\!\!\!\coprod}(E/F)_{l^\infty}$ which determine the structure of this group. The method of proof of the main results in §4.1. is by the construction of many independent elements of the Tate-Shafarevich group ${\textstyle\coprod\!\!\!\coprod}(E/F)_{l^\infty}$. Finally, §5.6 contains complements to the main results.
While care has been taken to minimize the number of hypotheses required for the main theorems of this paper, these hypotheses are still numerous (see for example §3.1, Definition 3.1.3). The assiduous reader will have an abundance of interesting problems in their elimination.
0.4in [**1.1. Global fields of positive characteristic**]{} 1 by 1
0.2in The notation of this paper is mainly that of the monograph \[1\] and is detailed in the rest of this chapter 1. A few differences arise, notably the sets of divisors $\Lambda(n)$, which are required for the more refined results of this paper.
Let
0.2in $k$ be a finite field of characteristic $p$ with $q=p^m$ elements;
${\overline k}$ be an algebraic closure of $k$;
$C/k$ be a smooth projective irreducible curve over $k$;
$F$ be the function field of $C$; these hypotheses imply that the finite field $k$ is
the exact field of constants of the global field $F$;
$\Sigma_L$, for any global field $L$, be the set of all places of the field $L$;
$\infty\in \Sigma_F$ be a closed point of $C/k$;
$\kappa(z)$ be the residue field at a place $z\in \Sigma_F$ of $F$;
$F_v$ be the completion of $F$ at the place $v\in \Sigma_F$;
$F^{\rm sep}$ be the separable closure of $F$;
$A$ be the coordinate ring $\Gamma(C\setminus\{\infty\}, {\cal O}_C)$ of the affine curve $C\setminus\{\infty\}$;
Div$_+(A)$ be the semi-group of effective $k$-rational divisors on Spec $A$; that
is to say Div$_+(A)$ is the semi-group of effective $k$-rational divisors on $C/k$
which are coprime to the place $\infty$; an element of Div$_+(A)$ may be written
as a finite linear combination $\sum_i n_i z_i$ where $n_i\in \mathds N$ and $z_i$ are prime
divisors on Spec $A$ for all $i$;
Supp$(c)$, for a divisor $c\in $ Div$_+(A)$, be the support of the divisor $c$ which is
the set of prime divisors with non-zero coefficient in $c$;
Pic$(A)$ be the Picard group of $A$, the group of projective $A$-modules of
rank $1$;
$K$ be a separable imaginary quadratic extension field of $F$ with respect to $\infty$,
that is to say $K$ is a quadratic extension field of $F$ in which the place $\infty$
remains inert;
$B$ be the integral closure of $A$ in $K$;
$\tau$ be the non-trivial element of the galois group Gal$(K/F)$.
0.4in[**1.1. Orders in imaginary quadratic field extensions**]{}1 by 1 0.2in Let $K/F$ be the imaginary quadratic field extension with respect to $\infty$ of §1.2.
An [*order*]{} $O$ in $K$ with respect to $A$ is an $A$-subalgebra of $B$ whose fraction field is equal to $K$.
There is a bijection between orders $O_c$ of $K$ with respect to $A$ and effective $k$-rational divisors $c$ in Div$_+(A)$ and it is given by $$c\mapsto A+BI(c)$$ where $I(c)$ is the ideal of $A$ cutting out the divisor $c$. The divisor $c$ is the [*conductor*]{} of the order $O_c$.
\[For more details on orders in imaginary quadratic extensions, see \[1, Chapter 2, §2.2\]
0.4in[**1.1. Ring class fields**]{}1 by 1 0.2in Let
0.2in
$O_c$ be the order of $K$ with respect to $A$ and with conductor $c$
where $c\in$ Div$_+(A)$ (see §1.3);
$A_v$, for each place $v$ of $A$, be the localisation of $A$ at $v$;
${\widehat O}_{c,v}$ be the completion of the semi-local ring $O_c\otimes_AA_v$;
$G_c=K_\infty^*\prod_v {\widehat O}_{c,v}^*$ be the subgroup of the idèle group of the global field $K$
whose components are the units of ${\widehat O}_{c,v}$ for all places $v\not=\infty$ of $F$
and $K_\infty^*$ for the place $v=\infty$ and where in the product $v$ runs over all
places of $F$;
$K[c]$, for any divisor $c\in$ Div$_+(A)$, be the ring class field with conductor $c$
with respect to $\infty$; this is the finite abelian extension field of $K$ defined
by the subgroup $G_c$ of the idèle group of $K$ via the reciprocity map;
$G(c/c')$ be the Galois group of the field extension $K[c]/K[c']$ for divisors $c\geq c'$
of Div$_+(A)$.
0.2in We have these properties of the decomposition of primes in ring class fields (for the proofs, see \[1, Chapter 2, §2.3.13\]): 0.2in (a) The primes ramified in $K[c]/K$ are precisely the primes in the support
of $c$.
\(b) The extension $K[c]/K$ is split completely at the place of $K$ lying
above $\infty$.
\(c) If $z\not\in $ Supp$(c)$ then for any positive integer $n$, the galois extension
$K[c+nz]/K[c]$ is totally ramified at all places of $K[c]$ above $z$. 0.2in \[See \[1, Chapter 2, §2.3\] for more details on ring class fields.\]
0.4in [**1.1. Elliptic curves over global fields of positive characteristic**]{} 1 by 1 0.2in Let
0.2in
$C/k$ be a smooth projective irredicible curve over $k$ (as in §1.2);
$X/k$ be an elliptic surface over $C$, that is to say $X/k$ is a smooth
projective irreducible surface, equipped with a morphism $f:X\to C$
which has a section, such that all fibres of $f$, except a finite number, are
elliptic curves;
$E/F$ be the generic fibre of $f:X\to C$, which is an elliptic curve $E$ over $F$
equipped with an origin where $F$ is the function field of $C$.
0.2in
The conductor of an elliptic curve $E/F$ is an effective $k$-rational divisor on $F$ supported only at the places of bad reduction of $E$ and whose multiplicities are defined in terms of the Galois representation of Gal$(F^{\rm sep}/F)$ given by $E$.
\[See \[1, §§B.11.1-B.11.4 for the definition of the conductor of $E/F$.\]
0.4in[**1.1. The Drinfeld modular curve $X_0^{\rm Drin}(I)$** ]{}1 by 1 0.2in Let 0.2in
$I$ be a non-zero ideal of $A$.
$X_0^{\rm Drin}(I)$ be the curve which is the coarse moduli scheme of Drinfeld modules
of rank $2$ for $A$ equipped with an $I$-cyclic structure (see \[1, Definition
2.4.2, p. 23\]); this curve is compactified by a finite number of cusps which
correspond to “degenerate" Drinfeld modules. 0.2in This curve $X_0^{\rm Drin}(I)$ is an analogue for the global field $F$ of the classical modular curve $X_0(N)$ which is the coarse moduli scheme of elliptic curves equipped with a cyclic subgroup of order $N$, where $N\in \mathds N$.
\[For more details, see \[1, §2.4\].\]
0.4in[**1.1. Analogue for $F$ of the Shimura-Taniyama-Weil conjecture**]{} 1 by 1
0.2in Let $E/F$ be an elliptic curve with split (Tate) multiplicative reduction at $\infty$. Let $I$ be the non-zero ideal of the ring $A$ which is the conductor, without the place at $\infty$, of the elliptic curve $E/F$.
According to the work of Drinfeld on the Langlands conjecture, there is a finite surjective morphism of curves over $F$ $$X_0^{\rm Drin}(I)\to E.$$ This result is an analogue for the global field $F$ of the Shimura-Taniyama-Weil conjecture proved by Wiles for semi-stable elliptic curves over the rational numbers.
\[For more details see \[1, §4.7\] and \[1, Appendix B\].\]
0.4in[**1.1. Drinfeld-Heegner points**]{} 1 by 1
0.2in
Let $K$ be an imaginary quadratic extension field of $F$ with respect to $\infty$ (as in §1.2) and let $I$ be a non-zero ideal of $A$.
Let $D$ be a rank $2$ Drinfeld module for $A$ with complex multiplication by an order $\cal O$ of the field $K$ with respect to $A$, that is to say $\cal O$ is a subring of $K$ which is integral over $A$. Let $Z$ be an $I$-cyclic subgroup of $D$. Then the pair $(D,Z)$ represents a non-cuspidal point of the modular curve $X_0^{\rm Drin}(I)$. Such points $(D,Z)$ exist if the prime divisors in the support of $I$ split completely in the field extension $K/F$.
If the quotient Drinfeld module $D/Z$ has the same ring of endomorphisms $\cal O$ as $D$ then the point $(D,Z)$ on $X_0^{\rm Drin}(I)$ is called a [*Drinfeld-Heegner point*]{}.
If $f:X_0^{\rm Drin}(I)\to E$ is a finite morphism of curves where $E/F$ is an elliptic curve (see §§1.5,1.6,1.7), then the point $f(D,Z)$ of the elliptic curve $E$ is also called a Drinfeld-Heegner point.
The Drinfeld-Heegner points $(D,Z)$ and $f(D,Z)$ are rational over the ring class field $K[c]$ where $c$ is the conductor of the order $\cal O$ of $K$ relative to $A$.
\[See \[1, §§2.2,2.3\] or §3.4 below for more details.\]
0.4in[**1.1. Groups and cohomology**]{}1 by 1
0.2in
If $G$ is a discrete abelian group denote by 0.2in $G_m$ the kernel of multiplication by the integer $m \in \mathds N$ on $G$;
$_mG$ the cokernel $G/mG$ of multiplication by the integer $m\in \mathds N$;
$\vert G\vert$ the order of the group $G$ which is either a positive integer or $+\infty$;
ord$(g)$ the order of an element $g\in G$ which is the cardinality of the subgroup
generated by $g$;
exp$(G)$ the exponent of $G$ which is maximum order of an element of $G$;
${\widehat {G }}$ the Pontrjagin dual of $G$ namely the topological group Hom$(G, {\mathds Q}/{\mathds Z})$;
0.2in
If $G$ is a finite abelian group then ${\widehat G}$ may be identified with the group of $1$-dimensional complex characters of $G$, that is to say the group of homomorphisms ${\rm Hom}(G, {\mathds C}^*).$ A character of a finite abelian group is always assumed to be irreducible.
If $F'/F$ is a galois extension of a global field $F$ and if $z\in
\Sigma_F$ is a prime divisor of $F$ unramified in $F'$ we denote by Frob$(z)$, or Frob$_{F'/F}(z)$, the conjugacy class in Gal$(F'/F)$ of Frobenius substitutions associated to $z$.
If $L$ is a field, we shall write $H^i(L,M)$ for the Galois cohomology group $H^i({\rm Gal}(L^{\rm sep}/L),M)$, where $L^{\rm sep}$ is the separable closure of $L$. If $L'/L$ is a finite galois field extension, we write $H^i(L'/L,M)$ for $H^i({\rm Gal}(L'/L),M)$.
If $F$ is a global field and $z$ is a place of $F$ then the restriction, or localization, of a class $c\in H^i(L,M)$ is written $c_z\in H^i(F_z,M)$ where $F_z$ is the completion of $F$ at $z$.
0.4in[**1.1. Torsion on elliptic curves $E/F$**]{}
0.2in Let $E$ be an elliptic curve over a global field $F$ of positive characteristic $p>0$ as in §1.5. Let $K$ be an imaginary quadratic extension field of $F$ with respect to the place $\infty$ of $F$. As in §1.4, let $K[c]$ be the ring class field over $K$ with conductor $c\in $ Div$_+(A)$. Put $$K[A]=\bigcup_{c\in {\rm Div}_+(A)} K[c].$$ That is to say $K[A]$ is a field which is the union of all the ring class fields $K[c]$ in some algebraic closure of $K$.
Let $S$ be a subset of ${\mathds N}^*$ such that if $n_1\in S$ and $n_0$ is any divisor of $n_1$ then $n_0\in S$. A [*quasi-group*]{} $\{ G_n\}_{n\in S}$ relative to $S$ is a family of abelian groups $G_n$ indexed by the elements $n$ of $S$ such that $nG_n=0$ and if $n_0,n_1\in S$ are elements where $n_0$ divides $n_1$ there is a group homomorphism $f_{n_1n_0}:G_{n_1}\to
G_{n_0}$ satisfying the compatibility condition that if $n_0,n_1,n_2\in S$ and $n_0$ divides $n_1$ and $n_1$ divides $n_2$ then $f_{n_2n_0}= f_{n_1n_0}\circ f_{n_2n_1}$. \[See \[1, Chapter 7, §7.1, p. 330\].\]
0.2in[**1.1.1 Proposition.**]{} (\[1, Proposition 7.3.8.\]) [*The quasi-group $$\{ E(K[A])_n\}_{n\in \mathds N}$$ is trivial, that is to say the order of the torsion group $E(K[A])_n$ is bounded independently of $n$ and there is a finite set $\cal E$ of prime numbers such that for all integers $n$ prime to all elements of $\cal E$ we have $$E(K[A])_n\cong 0.\ \ \ {\sqcap \!\!\!\!\sqcup}$$*]{}
0.2in[**1.1.2. Proposition.**]{} (\[1, Proposition 7.14.2\]) [*Let $\cal E$ be the finite set of prime numbers of Proposition 1.1.1. For any divisor $c$ of [Div]{}$_+(A)$, the restriction homomorphism $$H^1(K,E_n)\to H^1(K[c],E_n)^{{\cal G}_c}$$ is an isomorphism for all integers $n$ prime to $\cal E$ where ${\cal G}_c={\rm Gal}(K[c]/K)$.*]{} 0.2in This follows from Proposition 1.1.1 and the Hochschild-Serre spectral sequence $$H^i({\cal G}_c, H^j(K[c], E_n(K^{\rm sep})))\Rightarrow H^{i+j}
(K, E_n(K^{\rm sep}))$$(more details are given in \[1, Proposition 7.14.2\]).${\sqcap \!\!\!\!\sqcup}$
1 by 1 0.4in [ **1.1. Igusa’s theorem**]{} 0.2in This section is a summary of the results of Igusa for the Galois action on torsion points of elliptic curves over global fields of positive characteristic.
0.2in (1.1.1) As in §1.2, let $F$ be a global field of positive characteristic $p$ where $k$ is the exact field of constants of $F$ and let $E/F$ be an elliptic curve. Let
0.2in
$G=$ Gal$(F^{\rm sep}/F)$, where $F^{\rm sep}$ is the separable closure of $F$;
$n$ be an integer prime to $p$;
$E_n$ be the finite $F$-group scheme of $n$-torsion points of $E$;
$E_\infty$ be the torsion subgroup of $E(F^{\rm sep})$ of order prime to $p$. 0.2in
The elliptic curve $E/F$ is said to be \[isotrivial\] [*isotrivial*]{} if there is a finite galois extension field $F'$ of $F$ such that the curve $E\times_FF'$ is definable over a finite subfield of $F'$.
0.2in (1.1.2) The action of the galois group $G$ on $E_n$ provides a homomophism $$\rho_n:G\to {\rm Aut}(E_n)\cong {\rm GL}_2({\mathds Z}/n{\mathds Z}).$$ The determinant $${\rm det}:{\rm Aut}(E_n)\to ({\mathds Z}/n{\mathds Z})^*$$ induces a homomorphism $$G\to ({\mathds Z}/n{\mathds Z})^*.$$ Let $H_n$ be the subgroup of $({\mathds Z}/n{\mathds Z})^*$ generated by the powers of $q=\vert k\vert$ modulo $n$. Then $H_n$ is naturally isomorphic to the Galois group of the field of $n$th roots of unity over $k$. Let $\Gamma_n$ be the subgroup of GL$_2({\mathds Z}/n{\mathds Z})$ defined by the exact sequence of finite groups, where det is the restriction to $\Gamma_n$ of the determinant homomorphism on GL$_2({\mathds Z}/n{\mathds Z})$, $$\matrix{&&&&&{\rm det}&&&\cr
0&\to& {\rm SL}_2({\mathds Z}/n{\mathds Z})&\to &\Gamma_n&\to& H_n&\to
&0.\cr}\leqno{(1.\the\count1.3)}$$ 0.2in (1.1.4) Passing to the projective limit of the previous exact sequence over all integers $n$ prime to $p$ we obtain the exact sequence of profinite groups $$0\quad \longrightarrow\quad
{\rm SL}_2({\widehat{\mathds Z}}^{(p)})\quad \longrightarrow\quad {\widehat\Gamma}\quad \longrightarrow\quad{\widehat H}\quad \longrightarrow\quad 0$$ where $\widehat H$ is the subgroup of ${\widehat{\mathds Z}}^{(p)*}$ topologically generated by $q$, where $${\widehat{\mathds Z}}^{(p)} =\prod_{l\not=p}{\mathds Z}_l$$ is the profinite prime-to-$p$ completion of $\mathds Z$, and $\widehat{\Gamma}$ is a closed subgroup of ${\rm GL}_2({\widehat{\mathds
Z}}^{(p)})$.
0.2in (1.1.5) Passing to the projective limit of the exact sequence $(1.\the\count1.3)$ where $n$ runs over all powers of a prime number $l$ where $l\not=p$, we obtain the exact sequence $$0\quad \longrightarrow\quad
{\rm SL}_2({{\mathds Z}}_l)\quad \longrightarrow\quad {\widehat\Gamma}_l\quad \longrightarrow\quad{\widehat H}_l\quad \longrightarrow\quad 0.$$
0.2in [**1.1.6. Theorem.**]{} (Igusa \[I\]). [*Suppose that $E/F$ is not isotrivial. Then the profinite group ${\rm Gal}(F(E_\infty)/F)$ is an open subgroup of $\widehat \Gamma$.${\sqcap \!\!\!\!\sqcup}$*]{}
0.2in This result has the following consequence. 0.2in [**1.1.7. Theorem.**]{} [*Suppose that $E/F$ is not isotrivial. Then for all prime numbers $l\not=p$ the profinite group ${\rm
Gal}(F(E_{l^\infty})/F)$ is an open subgroup of ${\widehat \Gamma}_l$ and is equal to ${\widehat \Gamma}_l$ for all but finitely many $l$.${\sqcap \!\!\!\!\sqcup}$*]{} 0.2in [*1.1.8. Remarks.*]{} (1) Suppose that the curve $E/F$ is isotrivial. Then it is easy to show that the group ${\rm Gal}(F(E_\infty)/F)$ is an extension of a finite group by the abelian profinite group ${\widehat{\mathds Z}}^{(p)}$.
\(2) Let $E$ be an elliptic curve defined over a number field $L$. The galois action on the torsion points of $E/L$ is known and depends principally on whether or not $E$ has complex multiplication.
\[See \[S5\] and \[S5, §4.5\] for both the cases of complex multiplication and without complex multiplication. See also \[1, Chapter 7, Remarks 7.2.8\] for more details.\]
1 by 1
0.4in
[**1.1. Consequences of Igusa’s theorem**]{} 0.2in For a finite group $G$ and a ${\mathds
Z}[G]$-module $M$, let ${ H}^i(G,M)$ denote the standard cohomology groups of $G$ acting on $M$ (see \[1, §5.6\], see also \[9, Chap. 1\] for the Tate modified cohomology groups).0.2in
[**1.1.1. Proposition.**]{} [*Let $E/F$ be an elliptic curve and let ${\mathds N}^{(p)}$ be the set of positive integers coprime to $p$, where $p$ is the characteristic of $F$. Write $G_n$ for the group ${\rm Gal}(F(E_n)/F)$.*]{}
\(i) [*Let $i=0$ or $1$. Then* ]{} $$\{ { H}^i(G_n,E_{n})\}_{n\in {\mathds N}^{(p)}}$$ [*is a trivial quasi-group.* ]{}
\(ii) [*There is a finite set $\cal N$ of prime numbers including $p$ such that for all prime numbers $l\not\in {\cal N}$ we have* ]{} $$H^i(G_{l^n},E_{l^n})=0{\textsl{ \ \ for \ all }} \ n\geq 1{\textsl{
\ and \ all
\ }} i\geq 0.$$
\[Part (i) may be restated as: for $i=0$ or $1$ and for all $n$ coprime to $p$, the order of the group ${ H}^i(G_n,E_{n}) $ is bounded independently of $n$ and there is a finite set of prime numbers such that for all integers $n$ coprime to this set of prime numbers we have ${ H}^i(G_n,E_{n}) \cong 0$. For the proof of this proposition, see \[1, Chapter 7, Proposition 7.3.1\] ${\sqcap \!\!\!\!\sqcup}$0.2in
0.2in(1.12.2) For each prime number $l$ different from $p$, select once and for all a basis of the Tate module $T_l(E)$ over ${\mathds Z}_l$, the $l$-adic completion of $\mathds Z$; this fixes for the rest of this paper, for every prime number $l\not=p$, an isomorphism of ${\rm
Gal}(F(E_{l^\infty})/F)$ with a subgroup of ${\rm GL}_2({\mathds Z}_l)$ and the two groups may then be identified with each other.
0.2in [**1.1.3. Proposition.**]{}
*Let $E/F$ be an elliptic curve which is not isotrivial. Let $L$ be a finite extension field of $F$ in which $k$ is algebraically closed. Then there is an infinite set $S$ of prime numbers of positive Dirichlet density such that for all $l\in S$ we have*
[(a)]{} the fields $F(E_{l^\infty})$ and $L$ are linearly disjoint over $F$;
[(b)]{} $E(L)_{l^\infty}=0$;
[(c)]{} $\pmatrix{1&0\cr 0&-1\cr}\in {\rm Gal}(F(E_{l^\infty})/F)$.
0.2in \[For the proof, see \[1, Proposition 7.3.10\].\] ${\sqcap \!\!\!\!\sqcup}$ 0.4in
[**Chapter 2. Local duality, Cassels pairings, Tate-Shafarevich groups**]{} 0.4in [**2.1. Local duality of elliptic curves**]{} 0.2in This section is a brief summary of Tate-Poitou local duality for elliptic curves over a local field.
\[For more details on local duality of abelian varieties, see \[9, Chapter I and Chapter III, §7\].\]
0.2in(2.1.1) Let 0.2in $L$ be a non-archimedian complete local field;
$L^{\rm sep}$ be the separable closure of $L$;
$E/L$ be an elliptic curve over $L$;
$n\geq 1$ be an integer coprime to the characteristic of $L$;
$G$ be the Galois group Gal$(L^{\rm sep}/L)$.
0.2in(2.1.2) Let $\mu_{n}$ be the multiplicative subgroup of $L^{{\rm sep}}$ of $n$th roots of unity. Then $\mu_{n}$ is a finite $G$-module. Let $E_{n}$ denote the $G$-module of $n$-torsion points of $E(L^{\rm sep})$. We have an abelian group isomorphism $$E_{n}\cong {{\mathds Z}\over n{\mathds Z}}\oplus {{\mathds Z}\over
n{\mathds Z}}.$$ Denote by $\{,\}$ the Weil pairing $$\{,\}:E_{n}\times E_{n}\to \mu_{n}.$$ This is a perfect pairing of $G$-modules. In particular, we have an isomorphism of $G$-modules $$E_{n}\cong {\rm Hom}_G( E_{n}, \mu_{n}).$$
0.2in(2.1.3) \[WP1\] The Weil pairing induces a cup-product pairing in Galois cohomology $$H^1(L,E_{n})\times H^1(L,E_{n})\to H^2(L,\mu_{n}).$$ This is a non-degenerate anti-symmetric pairing of abelian groups. By local class field theory, we have a canonical isomorphism of groups, where Br$(L)$ is the Brauer group of $L$, $${\rm Br}(L)\cong {{\mathds Q}\over {\mathds Z}}.$$ This induces an isomorphism $$H^2(L,\mu_{n})={\rm Br}(L)_{n}\cong {{\mathds Z}\over n{\mathds Z}}.$$ 0.2in[**2.1.4. Theorem.**]{} (Tate-Poitou local duality). [ *The cup product pairing $$<,>_v:H^1(L,E_{n})\times H^1(L,E_{n})\to {{\mathds Z}\over n{\mathds Z}}\leqno{(2.1.5)}\label{aa}$$ obtained from the Weil pairing is an alternating and non-degenerate pairing of ${{\mathds Z}\over n{\mathds Z}}$-modules.*]{} 0.2in \[For the proof, see \[9, Chapter I, Cor. 2.3\].\]${\sqcap \!\!\!\!\sqcup}$
0.2in[**2.1.6. Theorem.**]{} [*Assume that $n$ is prime to the residue field characteristic of $L$.*]{}
\(i) [*The subgroup ${}_n\!E(L)$ of $ H^1(L,E_{n})$ is isotropic for the alternating pairing $<,>_v$.*]{}
\(ii) [*The cup product pairing $<,>_v$ on $ H^1(L,E_{n})$ induces a non-degenerate pairing of abelian groups, where ${}_nE(L)=E(L)/nE(L)$, $$[,]_v:{}_n\!E(L)\times H^1(L,E)_{n}\to {{\mathds Z}\over n{\mathds Z}}.$$* ]{} 0.2in
\[For the proof, see \[1, Chapter 7, Theorem 7.15.6, p.403\] for part (i) and \[9, Chap. I, Cor. 3.4 and Remark 3.6, Chapter III, Theorem 7.8\] for part (ii). Note that part (ii) holds without the hypothesis that $n$ be coprime to the characteristic of $L$, see \[9, Chap. III, §7\].\] ${\sqcap \!\!\!\!\sqcup}$
0.2in(2.1.7) Suppose now that $K$ is a global field of positive characteristic, $\Sigma_K$ is the set of all places of $K$, and that the integer $n$ is coprime to the characteristic of $K$. Let $E/K$ be an elliptic curve.
0.2in [**2.1.8. Proposition.**]{} [*Let $c$ and $c'$ be two elements of $H^1(K,E_n)$. Denote by $c_v$ and $c'_v$ the induced elements of $H^1(K_v,E_n)$ for all places $v\in \Sigma_K$ of $K$ where $K_v$ is the completion of $K$ at $v$. Then we have $$\sum_{v\in \Sigma_K}<c_v, c_v'>_v=0.$$*]{}
0.2in [*Proof.*]{} The sum of the local invariants of a global class in $H^2(K,{\mathds G}_m)$ is zero.${\sqcap \!\!\!\!\sqcup}$ 0.4in [**2.2. Selmer groups and Tate-Shafarevich groups**]{} 0.2in
(2.2.1) Let 0.2in $E/F$ be an elliptic curve as in §1.5;
$n\in {\mathds N}$ be an integer coprime to the characteristic $p$ of the global field $F$.
0.2in
(2.2.2) For a place $v$ of the field $F$, we write $F_v$ for the completion of $F$ at the place $v$ (as in §1.2). The exact sequence of commutative group schemes over $F$, obtained from the morphism of multiplication by $n$, $$0\longrightarrow E_n\longrightarrow E\ {\buildrel n\over \longrightarrow }\ E\longrightarrow 0$$ gives rise to a commutative diagram, where the maps res$_v$ are the restriction homomorphisms at $v$ and the rows are exact sequences of abelian groups, $$\matrix{0&\to & _{n}E(F)&\to & H^1(F,E(F^{\rm sep})_{n})&\to & H^1(F,E(F^{\rm
sep}))_{n}&\to & 0\cr &&\downarrow{\rm res}_v&&\downarrow{\rm res}_v&&\downarrow{\rm
res}_v&&\cr 0&\to & _{n}E(F_v)&\to & H^1(F_v,E(F^{\rm sep}_v)_{n})&\to &
H^1(F_v,E(F^{\rm sep}_v))_{n}&\to & 0\cr}$$ As in §1.9, $_nE(F)$ denotes the cokernel $E(F)/nE(F)$ and $E(F^{\rm sep})_n$ denotes the $n$-torsion subgroup of $E(F^{\rm sep})$.
0.2in
(2.2.3) The [*Tate-Shafarevich group*]{} ${\textstyle\coprod\!\!\!\coprod}(E/F)$ of $E/F$ is defined as $${\textstyle\coprod\!\!\!\coprod}(E/F)={\rm ker}\{ H^1(F,E)\to
\prod_{v\in \Sigma_F}H^1(F_v,E)\}.$$ This group ${\textstyle\coprod\!\!\!\coprod}(E/F)$ is known to be a torsion of cofinite type (see \[10\]).
The [*$n$-Selmer group*]{} is defined as, in terms of the commutative diagram of (2.2.2) and where ${\rm res}_v$ denotes the middle vertical homomorphism of the diagram, $${\rm Sel}_{n}(E/F)=\bigcap_{v\in \Sigma_F}{\rm
res}_v^{-1}(_{n}E(F_v)).$$ Therefore ${\rm Sel}_{n}(E/F)$ is a subgroup of $H^1(F,E(F^{\rm sep})_n)$ and is a finite abelian group. We then have the exact sequence of torsion abelian groups from the commutative diagram of (2.2.2), where ${\textstyle\coprod\!\!\!\coprod}(E/F)_{n}$ is the $n$-torsion subgroup of ${\textstyle\coprod\!\!\!\coprod}(E/F)$, $$0\to \ _{n}E(F)\to {\rm Sel}_{n}(E/F)\to
{\textstyle\coprod\!\!\!\coprod}(E/F)_{n}\to 0.$$
0.2in
(2.2.4) Let $F'/F$ be a finite separable galois field extension. We write ${\textstyle\coprod\!\!\!\coprod}(E/F')$ in place of ${\textstyle\coprod\!\!\!\coprod}(E\times_FF'/F')$ for the Tate-Shafarevich group of $E\times_FF'$ over $F'$ obtained by ground field extension from $F$ to $F'$; similarly, for the Selmer quasi-group, we write ${\rm Sel}_{n}(E/F')$ in place of ${\rm Sel}_{n}(E\times_FF'/F')$.
0.4in [**2.2.5. Proposition.**]{} [*Let $r$ be the degree of the finite separable field extension $F'/F$. For any torsion abelian group $\cal A$ write $\cal A_{({\rm non}\ r )} $ for the torsion subgroup of $\cal A$ of order coprime to $r$. The restriction homomorphism provides isomorphisms for all integers $n$ coprime to $r$*]{} $${\textstyle\coprod\!\!\!\coprod} (E/F)_{
({\rm non}\ r )}\ \ {\buildrel {\buildrel{\rm
res}\over\cong }\over \longrightarrow} \ \
{\textstyle\coprod\!\!\!\coprod} (E/F')_{({\rm non}\ r )}{}^{{\rm Gal}(F'/F)}$$ $${\rm Sel}_n(E/F) \ \
{\buildrel {\buildrel{\rm res}\over\cong }\over \longrightarrow}
\ \
{\rm Sel}_n(E/F'){}^{{\rm Gal}(F'/F)}.$$ [*Proof.*]{} The definition of the Tate-Shafarevich groups provides a commutative diagram with exact rows $$\matrix{ 0&\!\!\to\!\! &\!\!
{\textstyle\coprod\!\!\!\coprod}(E/F')^{{\rm
Gal}(F'/F)}\!\! &\!\!\to\!\! & \!\!H^1(F',E)^{{\rm Gal}(F'/F)}\!\!
&\!\!\to\!\! &\!\!\displaystyle{\bigg(\prod_{v\in
\Sigma_{F'}}H^1(F'_v,E)}\bigg)^{{\rm Gal}(F'/F)}\cr &&\uparrow&&\uparrow&&\uparrow\cr
0&\!\!\to\!\! &\!\!{\textstyle\coprod\!\!\!\coprod}(E/F)\!\!
&\!\!\to\!\! & \!\!H^1(F,E)\!\!&\!\!\to\!\!
&\!\!\displaystyle{\prod_{v\in \Sigma_{F}}H^1(F_v,E)}\cr}$$ For any place $v$ of $F$, the $F_v$-algebra $F_v\otimes_FF'$ is étale and is the product of the completions of $F'$ at the places lying over $v$. The inflation restriction sequence provides isomorphisms for any integer $s$ coprime to the order of Gal$(F'/F)$ $$\begin{aligned}
H^1(F,E)_{s}&\cong H^1(F',E)^{{\rm
Gal}(F'/F)}_{s}\phantom{{{\rm Gal}(F'/F)}
{\rm \ \ all\ places \ } v{\rm \ of \ } F}\\
H^1(F_v,E)_{s}&\cong H^1(F_v\otimes_FF',E)_{s}^{{\rm Gal}(F'/F)}
{\rm \ for \ all\ places \ } v{\rm \ of \ } F.\\\end{aligned}$$ The isomorphism of the proposition for the Tate-Shafarevich groups now follows by a diagram chase. The corresponding isomorphism for the $n$-Selmer groups follows from the commutative diagram with exact rows $$\matrix{0&\to& \ _{n}E(F)&\to& {\rm Sel}_{n}(E/F)&\to&
{\textstyle\coprod\!\!\!\coprod}(E/F)_{n}&\to& 0\cr
&&\downarrow\cong &&\downarrow&&\downarrow\cong&&\cr
0&\to& (\ _{n}E(F'))^{{\rm Gal}(F'/F)}&\to& {\rm Sel}_{n}(E/F')^{{\rm Gal}(F'/F)}&\to&
{\textstyle\coprod\!\!\!\coprod}(E/F')_{n}^{{\rm Gal}(F'/F)}&\to& 0\cr
}$$ as required. ${\sqcap \!\!\!\!\sqcup}$
0.2in [*2.2.6. Remark.*]{} This section is a generalised form of \[1, Chapter 7, §7.9\].
0.4in [**2.3. The Cassels pairing**]{}
0.2in(2.3.1) Let $E/F$ be an elliptic curve over the global field $F$ of characteristic $p>0$ as in §1.5. The Tate-Shafarevich group $\coprod\!\!\!\coprod (E/F)$ of $E/F$ is equipped with the anti-symmetric Cassels pairing $$<,>_{\rm Cassels}\ :\ \coprod\!\!\!\!\coprod(E/F)
\times
\coprod\!\!\!\!\coprod(E/F)\to {\mathds Q}/{\mathds Z}$$ which is non-degenerate if $\coprod\!\!\!\coprod$$(E/F)$ is finite. The Tate-Shafarevich group $\coprod\!\!\!\coprod$$(E/F)$ is a torsion group of cofinite type (see \[10\]).
In this section, the Cassels pairing is defined for $E/F$ for the non-$p$ part $\coprod\!\!\!\coprod (E/F)_{ ({\rm non}-p) }$ of $\coprod\!\!\!\coprod (E/F)$, that is to say the subgroup of the Tate-Shafarevich group of order coprime to the characteristic $p$.
0.2in(2.3.2) For any place $v\in \Sigma_F$ of $F$, we have the commutative diagram with exact rows obtained from restriction from $F$ to $ F_v$ for any integer $m$ where $\partial_m$ is the connecting homomorphism induced from the morphism of multiplication by $m$ on $E$ $$\matrix{
0&\to& E(F)_m&\to &E(F)&\to&E(F)&{\buildrel \partial_m\over \to} & H^1(F,E_m)&\to & H^1(F,E)_m\cr
&&\downarrow &&\downarrow
&&\downarrow
&&\downarrow
&&\downarrow
\cr
0&\to& E(F_v)_m&\to &E(F_v)&\to&E(F_v)&{\buildrel \partial_m\over \to} & H^1(F_v,E_m)&\to & H^1(F_v,E)_m\cr}$$
0.2in(2.3.3) Let $a,b\in$ $\coprod\!\!\!\coprod (E/F)_{ ({\rm non}-p) }$. Let $m\geq 1$ be the order of $a$ and $n\geq 1$ be the order of $b$ where $m,n$ are coprime to the characteristic $p$ of $F$. Then we have 0.2in $$a\in \coprod\!\!\!\!\coprod(E/F)_m{\rm \ \ and \ \ } b\in \coprod\!\!\!\!\coprod(E/F)_{n}.$$
0.2in Select elements $$a^{(1)}\in H^1(F, E_m) {\rm \ \ and \ \ } b^{(1)}\in H^1(F, E_{n})$$ mapping to $a$ and $b$ respectively in the commutative diagram of (2.3.2).
For any element $h\in H^1(F,E)$ denote by $h_v$ the restriction of $h$ to $H^1(F_v,E)$ for any place $v$ of $F$ and similarly for cochains.
0.2in(2.3.4) Suppose first that $a$ is divisible by $n$ in $H^1(F,E)$, say $a=na_1$ where $a_1\in H^1(F,E)_{mn}$. We may select local points $y_v\in {}_nE(F_v)$ such that $$\partial_n(y_v)=b^{(1)}_v,{\rm \ \ for \ all \ places\ }v\in \Sigma_F,$$ as $b^{(1)}_v$ maps to zero in $H^1(F_v,E)$. Let $a_{1,v}$ denote the localization in $H^1(F_v,E)_{mn}$ of $a_1$. Note that since $a\in $ $\coprod\!\!\!\coprod$$(E/F)_m$ we have $a_{1,v}\in H^1(F_v,E)_{n}$ for all $v$. For any $v\in \Sigma_F$, denote by $$[,]_v: H^1(F_v,E)_n \times {}_n E(F_v)\to {{\mathds Z}\over n{\mathds Z}}.$$ the local pairing as in Theorem 2.1.6. Then the Cassels pairing is given in terms of the local pairing by the formula $$<a,b>_{\rm Cassels}=\sum_{v\in \Sigma_F}[a_{1,v},y_v]_v \leqno{(2.3.5)}$$ where the sum runs over all places of $F$.
0.2in(2.3.6) We have $$a_{1,v}=i_*(c_{1,v})$$ for some $c_{1,v}\in H^1(F_v,E_{n})$ for all $v$ where $i$ is the inclusion of group schemes $E_n\hookrightarrow E$. We then have that the Cassels pairing is also given in terms of the cup product pairing $<,>_v$ of Theorem 2.1.4 by $$<a,b>_{\rm Cassels}=\sum_v
<c_{1,v},b^{(1)}_v>_v \leqno{(2.3.7)}$$ as we have for all places $v$ $$[a_{1,v},y_v]_v =<c_{1,v},\partial_n(y_v)>_v .$$
0.2in(2.3.8) In the previous two paragraphs, the global element $a_1\in H^1(F,E)_{mn}$ such that $na_1=a$ may not exist. Nevertheless, in a suitable sense it always exists locally and in general the Cassels pairing is defined as follows.
Select elements $a^{(1)}$ and $b^{(1)}$ of $H^1(F, E_m)$ and $H^1(F, E_{n})$ mapping to $a$ and $b$ respectively. For each valuation $v$ of $F$, let $a^{(1)}_v$ be the localisation in $H^1(F_v,E_m)$ of $a^{(1)}\in H^1(F,E_m)$. For each valuation $v$ of $F$, $a$ maps to zero in $H^1(F_v,E)$ and hence $a^{(1)}_v$ lies in the image of $E(F_v)$ under $\partial_m$. Then we can lift, by the diagram where id is the identity map, $$\matrix{ E(F_v)&{\buildrel \partial_m\over \to} & H^1(F_v,E_m) \cr
{\rm id}\uparrow&&\phantom{n}\uparrow n\cr
E(F_v)&{\buildrel \partial_{mn}\over \to} & H^1(F_v,E_{mn})\cr}$$ $a^{(1)}_v\in H^1(F_v,E_m)$ to an element $a^{(1)}_{v,1}\in H^1(F_v,E_{mn})$ so that $na^{(1)}_{v,1}=a^{(1)}_v$ and $a^{(1)}_{v,1}$ is in the image of $E(F_v)$ under $\partial_{mn}$.
Let $\beta$ be a cocycle in ${\rm Cocy}^1(F,E_m)$ representing $a^{(1)}\in H^1(F,E_m)$ and lift it to a cochain $\beta_1\in {\rm Coch}^1(F,E_{mn})$. Select a cocycle $\beta_{v,1}\in {\rm Cocy}^1(F_v,E_{mn})$ representing the element $a^{(1)}_{v,1}\in H^1(F_v, E_{mn})$ and a cocycle $\beta'\in {\rm Cocy}^1(F,E_{n})$ representing $b^{(1)}\in H^1(F,E_n)$.
The coboundary $d\beta_1$ of $\beta_1$ takes values in $E_{n}$ as $\beta_1$ is a cochain lifting the cocycle $\beta$ with values in $E_m$. The cup product $d\beta_1\cup \beta'$ represents an element of $H^3({\rm Gal}(F^{\rm sep}/F),
{\mathds G}_m)$ where ${\mathds G}_m$ is the multiplicative group scheme over $F$ and where this cup product is induced by the Weil pairing $E_n\times E_n\to {\mathds G}_m$ (see (2.1.3)). But this last cohomology group $H^3({\rm Gal}(F^{\rm sep}/F),
{\mathds G}_m)$ is zero (see \[9, Chap. 1, 4.18 or 4.21\]). Hence we have $$d\beta_1\cup \beta'=d\epsilon$$ for some 2-cochain $\epsilon\in {\rm Coch}^2(F,{\mathds G}_{\bf m})$.
The cochain localized at $v$ $$((\beta_{v,1}-\beta_{1,v})\cup \beta_v')+\epsilon_v$$ is a 2-cocycle in ${\rm Cocy}^2(F_v, {\mathds G}_{ m})$. Denote by $${\rm inv}_v: {\rm Br}( F_v)=H^2(F_v, {\mathds G}_m) \to {\mathds Q}/{\mathds Z}$$ the canonical isomorphism (the invariant map) given by local class field theory. Define $$<a,b>_{\rm Cassels}=\sum_{v\in \Sigma_F}
{\rm inv}_v(((\beta_{v,1}-\beta_{1,v})\cup \beta_v')+\epsilon_v)\in {\mathds Q}/{\mathds Z}.$$ This can be checked to be independent of the selections made and defines the pairing on $\coprod\!\!\!\coprod (E/F)_{ ({\rm non}-p) }$.
0.2in
[*2.3.9. Remarks.*]{} (1) For this paper, only that part of the construction of the Cassels pairing in paragraphs (2.3.2)-(2.3.6) is required. This is because, under the hypotheses of the main theorems of this paper stated in §4.1, the subgroup $\coprod\!\!\!\coprod (E/K)$ of $H^1(K,E)$ has the following divisibility property.
Under the hypotheses and notations of the main theorems of this paper stated in §4.1, for all prime numbers $l\in \cal P$ where $l$ is coprime to Pic$(A)$ and for any integer $n\geq 0$ there is a subgroup $H$ of $ H^1(K,E)$ such that $l^nH=\coprod\!\!\!\coprod (E/K)_{l^\infty}$.
This divisibility holds because for any sufficiently large positive integer $a$ the group $\coprod\!\!\!\coprod (E/K)_{l^\infty}$ is generated by the cohomology classes $\delta_{M_0}(c)$ where $c$ ranges of the elements of $\Lambda(a)$ (Theorem 5.6.2) and that $l^n\delta_{M_0+n}(c)=\delta_{M_0}(c)$ if $c\in \Lambda(M_0+n)$ (lemma 4.2.1(iii)).
It would be interesting to have examples of elliptic curves $E/K$ and prime numbers $l$ for which $\coprod\!\!\!\coprod (E/K)_{l^\infty}$ is non-zero but does not have this divisibility property as a subgroup of $H^1(K,E)$.
\(2) The Cassels pairing for abelian varieties over global fields can be defined in several ways. For a more geometric construction of the pairing than that given above, see \[9, Chap. I, Remark 6.11, p. 98\]. For the special case of Jacobians of curves, see \[9, Chap. I, Remark 6.12, p. 100\]. For the construction of the pairing including the $p$-torsion part of the Tate-Shafarevich group, where $p$ is the characteristic of the base field, see \[9, Chap. II, Theorem 5.6, pp. 247-248\].
0.2in
[**Chapter 3. The cohomology classes $\gamma_n(c),\delta_n(c)$**]{} 0.4in [**3.1. The set $\cal P$ of prime numbers**]{}
0.2in We define a set $\cal P$ of prime numbers by arithmetic conditions. For prime numbers $l$ of $\cal P$ we shall consider in Chapters 4 and 5 the structure of the $l$-primary component of the Tate-Shafarevich group ${\coprod\!\!\!\coprod}(E/F)$ of elliptic curves $E/F$. 0.2in (3.1.1) For a place $v\in \Sigma_F$ of the global field $F$, let
0.1in
$F_v$ denote the completion of $F$ at $v$;
$F_v^{\rm nr}$ denote the maximal unramified extension of the local field $F_v$
(that is, $F_v^{\rm nr}$ is the field of fractions of the strict henselisation of the
valuation ring of $F_v$);
$O_v$ denote the discrete valuation ring of the local field $F_v$;
$\infty\in \Sigma_F$ be a place of $F$;
$K$ be a separable imaginary quadratic extension field of $F$ with respect to the
place $\infty$ as in §1.2;
$E/F$ be an elliptic curve (as in §1.5);
$\cal
E$ denote the Néron model over $O_v$ of the elliptic curve $E\times_FF_v/F_v$;
${\cal E}_0$ denote the closed fibre of ${\cal E}/O_v$;
$\pi_0({\cal E}_0)$ be the group of connected components of ${\cal E}_0$
as a Gal$(F_v^{\rm nr}/F_v)$-module.
0.2in
Define similarly $K_w$, $K_w^{\rm nr}$ for a place $w\in \Sigma_K$ of the imaginary quadratic extension field $K$ of $F$.
0.2in[**3.1.2. Theorem.**]{} (\[9, Chap. I, Prop. 3.8, p.57\]). [*Write $G={\rm Gal}(F_v^{\rm nr}/F_v)$. There is an isomorphism $$H^1(G, E(F_v^{\rm nr}))\cong H^1(
G, \pi_0({\cal E}_0)).$$ In particular, $H^1(G, E(F_v^{\rm nr}))$ is a finite group for all $v$ and if the elliptic curve $E$ has good reduction at $v$ then $H^1(G,
E(F_v^{\rm nr}))=0$.*]{}${\sqcap \!\!\!\!\sqcup}$
0.2in [**3.1.3. Definition.**]{} Let $\cal P$ be the set of all prime numbers such that for all $l\in \cal P$ we have 0.2in (a) $p$, $2$, and the prime factors of $\vert B^*\vert/\vert A^*\vert$ are not in $\cal P$;
\(b) $H^i(K(E_{l^n})/K,E_{l^n})=0$ for all integers $n\geq
1$ and for all $i\geq 0$
(see Proposition 1.12.1);
\(c) the natural injection Gal$(F(E_{l^\infty})/F)\to
{\widehat \Gamma}_l$ is an isomorphism
(see §1.11 and Igusa’s Theorem 1.11.7);
\(d) $H^1(K^{\rm nr}_z/K_z,E)_{l^\infty}=0$ for all places $z$ of $K$ (see Theorem 3.1.2 above);
\(e) $K$ and $F(E_{l^\infty})$ are linearly disjoint over $F$ (see Proposition 1.12.3 or \[1, Chapter 7, 7.3.10\]);
\(f) $\cal P$ excludes the prime numbers of the finite set $\cal E$ of Proposition 1.10.1, that is
to say, we have $E(K[A])_{l^m}=0$ for all $l\in {\cal P}$, for all $m\geq 1$, where $K[A]$
is defined in §1.10;
\(g) $\pmatrix{1&0\cr0&-1\cr}\in {\rm Gal}(F(E_{l^\infty})/F)$ (see Proposition 1.12.3, which is a
consequence of Igusa’s Theorem 1.11.6);
0.2in [*3.1.4. Remarks.*]{} (1) The first 6 conditions (a),(b),(c),(d),(e),(f) of this definition hold for all except finitely many prime numbers $l$. Only the last condition (g) fails to hold in general for all but finitely many prime numbers. The set $\cal P$ is infinite and has positive Dirichlet density.
The set $\cal P$ can therefore be obtained from the set $S$ of prime numbers provided by Proposition 1.12.3 by deleting a finite number of elements. More precisely, (see \[1, Chapter 7, Remarks 7.3.14(1)\] and also \[1, Chapter 7, Remarks 7.7.6(1)\]) the set $\cal
P$ consists of all but finitely many prime numbers $l$ of the form $2^sn+1$ where $s\geq 1$ and $n$ is odd such that $q=\vert k\vert$ is a $2^s$th power non-residue modulo $l$.
0.1in
\(2) The set $\cal P$ of prime numbers of Definition 3.1.3 above coincides with the set of prime numbers written ${\cal P}\setminus {\cal F}$ of \[1, Lemma 7.14.11\]. The only difference between the Definition 3.1.3 above of the set of prime numbers $\cal P$ and the similar definition \[1, Definition 7.10.3\] is the extra hypothesis (f) of 3.1.3 above which excludes from $\cal P$ the finitely many prime numbers of the exceptional set $\cal E$ of Proposition 1.10.1.
0.2in[**3.1.5. Lemma.**]{} [*For any integers $0\leq m\leq n $ and for all prime numbers $l\in \cal P$ the inclusion of group schemes $E_{l^m}\to E_{l^n}$ induces an injection of cohomology groups $$H^1(K, E_{l^m}){\longrightarrow} H^1(K, E_{l^{n}}).$$* ]{} 0.2in [*Proof of Lemma 3.1.5.*]{} This follows from the long exact sequence induced by the isogeny on the finite group scheme $E_n$ of multiplication by $l^{m}$ $$0\to E_{l^m}(K)\to E_{l^{n}}(K)\to E_{l^{n-m}}(K)\to$$ $$H^1(K, E_{l^m})\to H^1(K, E_{l^{n}})\to H^1(K, E_{l^{n-m}})\to\ldots$$ together with the non-existence of $K$-rational $l$-torsion on $E$ (by Proposition 1.10.1 and the definition of $\cal P$, Definition 3.1.3(f)). ${\sqcap \!\!\!\!\sqcup}$
0.2in (3.1.6) We write, where the limits are set-theoretic unions by the previous Lemma 3.1.5, $$H^1(K, E_{l^\infty})=\lim_{\longrightarrow\atop n}H^1(K, E_{l^n})$$ and $${\rm Sel}_{l^\infty}(E/K)=\lim_{\longrightarrow\atop n} {\rm Sel}_{l^n}(E/K).$$ 0.2in (3.1.7) Let ${\mathds Q}(l)$, for any prime number $l$, be the additive group of rational numbers with denominators a power of $l$; the quotient group ${\mathds Q}(l)/{\mathds Z}$ is a divisible abelian group where every element is annihilated by a power of $l$.
We have the exact sequence of abelian groups for all prime numbers $l\in \cal P$ where $E(K)_{{\rm tors}}$ denotes the torsion subgroup of $E(K)$ $$0\longrightarrow {E(K) \over E(K)_{\rm tors}}\otimes_{\mathds Z} {{\mathds Q}(l)
\over {\mathds Z}}
\longrightarrow {\rm Sel}_{l^\infty}(E/K)\longrightarrow
{\textstyle \coprod\!\!\!\coprod}(E/K)_{l^\infty} \longrightarrow 0\leqno{(3.1.8)}$$ This exact sequence (3.1.8) is obtained from Lemma 3.1.5, the exact sequence of paragraph (2.2.3), and because $E(K)$ has no $l$-torsion (Definition 3.1.3(f)).
The exact sequence (3.1.8) splits and gives the isomorphism $${\rm Sel}_{l^\infty}(E/K)\cong {E(K) \over E(K)_{\rm tors}}\otimes_{\mathds Z} {{\mathds Q}(l)
\over {\mathds Z}}\ \ \oplus \ \
{\textstyle \coprod\!\!\!\coprod}(E/K)_{l^\infty}.\leqno{(3.1.9)}$$ This holds as the abelian group ${E(K) \over E(K)_{\rm tors}}\otimes_{\mathds Z} {{\mathds Q}(l)
\over {\mathds Z}}$, where $E(K)$ is a finitely generated group, is injective in the category of abelian groups.
It follows from the isomorphism (3.1.9) that Sel$_{l^m}(E/K)$ is precisely the subgroup of $ {\rm Sel}_{l^\infty}(E/K)$ annihilated by $l^m$ that is to say we have for all $m\geq 0$ and all $l\in {\cal P}$ $${\rm Sel}_{l^m}(E/K)\cong ( {\rm Sel}_{l^\infty}(E/K))_{l^m}.\leqno{(3.1.10)}$$
[**3.2. Frobenius elements and the set ${\Lambda}(n)$ of divisors**]{}
0.2in(3.2.1) Let
0.2in $F$ be the function field of the curve $C/k$ as in §1.2;
$\infty\in \Sigma_F$ be a closed point of $C/k$;
$K$ be a separable imaginary quadratic extension field of $F$ with respect to $\infty$
as in §1.2;
$\tau\in {\rm Gal}(K/F)$ be the non-trivial element of the Galois group of $K/F$:
$E/F$ be an elliptic curve with conductor $I$;
$\cal P$ be the set of prime numbers associated to $E,F,K$ as in §3.1;
$l\in \cal P$ be a prime number in $\cal P$.
0.2in As in paragraph (1.12.2), for every prime number $l\not=p$, an isomorphism is fixed between ${\rm
Gal}(F(E_{l^\infty})/F)$ and a subgroup of ${\rm GL}_2({\mathds Z}_l)$ by fixing a basis of the corresponding Tate module.
0.2in(3.2.2) For each integer $n\geq 1$ and the chosen $l\in {\cal P}$, let $\tau_\infty\in {\rm Gal}(K(E_{l^n})/F)$ be the unique element of this Galois group satisfying the two conditions: 0.1in ${\displaystyle{\ \tau_\infty\vert_{F(E_{l^n})}=\pmatrix{1&0\cr0&-1\cr} }}$ that is to say the restriction of $\tau_\infty$ to the field
extension $F(E_{l^n})/F$ is ${\displaystyle{\pmatrix{1&0\cr0&-1\cr} }};$ 0.1in ${\displaystyle{\tau_\infty\vert_{K}=\tau }}$ is the non-trivial element of Gal$(K/F).$ 0.1in
The elements $\tau$ and $\tau_\infty$ have exact order 2.
0.2in(3.2.3) For any ${\mathds Z}[{\rm
Gal}(K/F)]$-module $M$ on which multiplication by $2$ is an isomorphism, we have a decomposition of $M$ as a sum of eigenspaces under the action of $\tau$, the non-trivial element of Gal$(K/F)$, $$M\cong M^+\oplus M^-$$ where $M^+$ is the submodule of $M$ on which $\tau $ acts as $1$ and where $M^-$ is similarly the submodule of $M$ on which $\tau $ acts as $-1$.
0.2in[**3.2.4. Definition.**]{} (i) For a prime divisor $z\in \Sigma_F$, unramified in the field extension $K(E_{l^n})/F$, let $${\rm Frob}(z)$$ denote the conjugacy class of Gal$(K(E_{l^n})/F)$ containing the Frobenius substitutions of the prime divisors above $z$.
\(ii) For $l\in \cal P$, let $\Lambda^1(n)$ be the set of prime divisors $z$ in $\Sigma_F$, of support coprime to $\infty$ and Supp$(I)$ and the discriminant of $K/F$, which satisfy $${\rm Frob}(z)=[\tau_\infty]$$where $[\tau_\infty]$ denotes the conjugacy class in Gal$(K(E_{l^n})/F)$ of $\tau_\infty$.
\(iii) For $r\geq 0$, let $\Lambda^r(n)$ be the set of effective divisors $z_1+\ldots+z_r$ on the affine curve Spec $A$ which have $r$ prime components $z_i$ all of which have multiplicity 1 and belong to $\Lambda^1(n)$.
We conventionally put for all $n\geq 1$ $$\Lambda^0(n)=\{0\}$$ which is the set consisting of the zero divisor on $C/k$.
\(iv) Put $$\Lambda^r=\Lambda^r(1)$$ $$\Lambda(n)=\bigcup_{r\geq 0} \Lambda^r( n)$$ $$\Lambda=\bigcup_{n\geq 1} \Lambda( n).$$ We have $$\Lambda^r(n)=\left\{\matrix{ z_1+\ldots +z_r\in {\rm Div }_+(A)\ &\vert& \ z_1,\ldots,z_r {\rm \ are \ distinct\ prime \ divisors\ such \ }\cr
&\vert&
{\rm that \ for \ all \ }i, \ z_i{\rm \ is \ prime \ to \ } \infty, \ {\rm Supp}( I)
\cr
&\vert& {\rm and \ the \ discriminant\ of \ }K/F{\rm \ and \ }\cr
&\vert&{\rm Frob}
(z_i)=[\tau_\infty]{\rm \ on \ }
{ K}(E_{l^n})\cr}\right\} .$$ The set $\Lambda^r$ has a decreasing filtration $$\Lambda^r=\Lambda^r(1)\supseteq \Lambda^r(2)\supseteq\Lambda^r(3)\supseteq\ldots.$$ 0.2in [*3.2.5. Remarks.*]{} (1) The prime divisors in $\Lambda(n)$ are infinite in number, by the Chebotarev density theorem, remain prime in the field extension $K/F$, and their liftings to $K$ split completely in $K(E_{l^n})/K$. Furthermore, $E$ has good reduction at all prime divisors of $\Lambda(n)$.
Note that the prime number $l\in \cal P$ is considered to be fixed and the sets $\Lambda^r(n)$ depend on $l$.
0.1in(2) The set $\Lambda(n)$ is defined for any prime number $l$ in $\cal P$ and contains only effective divisors on Spec $A$ consisting of sums of distinct prime divisors whose corresponding Frobenius conjugacy classes in ${\rm Gal}(K(E_{l^n})/F)$ are all the same and equal to $[\tau_\infty]$. This unique Frobenius conjugacy class is of a special kind, in particular its elements have order $2$.
0.2in(3.2.6) For any prime number $\lambda$ distinct from the characteristic of $F$, let $$\rho:{\rm Gal}(F^{\rm sep}/F)\to {\rm Aut}_{{\mathds
Q}_\lambda}(T_\lambda(E)\otimes_{{\mathds Z}_\lambda}{\mathds Q}_\lambda)$$ denote the galois representation on the $\lambda$-adic Tate module $T_\lambda(E)$ of the elliptic curve $E$; that is to say $$T_\lambda(E)\otimes_{{\mathds
Z}_\lambda}{\mathds Q}_\lambda=H^1_{\rm \acute et}(E\otimes_FF^{\rm sep},{\mathds
Q}_\lambda)^{*}$$ where $*$ denotes the dual ${\mathds Q}_\lambda$-vector space.
If $z$ is a prime of $F$ let $I_z$ be an inertia subgroup of ${\rm Gal}(F^{\rm
sep}/F)$ at $z$; let $$a_z={\rm Tr}(\rho({\rm Frob}(z))\
\vert \ (T_\lambda(E)\otimes_{{\mathds Z}_\lambda}{\mathds Q}_\lambda)^{I_z}).$$ That is to say, $a_z$ is the trace of the Frobenius at $z$ on the part of the Tate module invariant under $I_z$. Then we have $a_z\in {\mathds Z}$ (see \[1, Chapter 5, Examples 5.3.18(1)\]).
0.2in [**3.2.7. Lemma.**]{} [*Suppose that $z\in \Sigma_F$ is a prime divisor of $F$ and $l\in {\cal P}$; write $\kappa(z)$ for the residue field at $z$. .* ]{}
\(i) [*If $z\in \Lambda^1(n)$ then we have $a_z\equiv \vert \kappa(z)\vert +1\equiv 0$ [(mod $l^n$)]{}.*]{}
\(ii) [*If $z\in \Lambda^1(n)$ and ${\cal E}_{0,z}$ denotes the closed fibre over $z$ of the Néron model of $E/F$ and $y$ is the prime divisor of $K$ lying over $z$ then we have group isomorphisms, for $\delta=+1$ or $-1$,* ]{} $${\cal
E}_{0,z}(\kappa(y))_{l^n}^\delta\cong {\cal
E}_{0,z}(\kappa(z))_{l^n}\cong {\mathds Z}/l^n{\mathds Z}.$$
0.2in [*Proof.*]{} (i) As $a_z$ is the trace of a Frobenius above $z$ on the Tate module of $E/F$, we have by the Grothendieck-Lefschetz trace formula, where $\kappa(z)$ is the residue field of $F$ at $z$ and ${\cal E}_{0,z}$ is the closed fibre over $z$ of the Néron model of $E/F$, $$a_z=\vert\kappa(z)\vert+1-\vert{\cal E}_{0,z}(\kappa(z))\vert.$$
On the one hand, let $K'$ be the field $K(E_{l^n})$. The characteristic polynomial of the Frobenius Frob$_{K'/F}(z)$ above $z$ acting on the $l$-adic Tate module of $E/F$ is equal to $$X^2-a_zX+\vert\kappa(z)\vert$$
On the other hand as $z\in \Lambda^1(n)$, the prime $z$ is a place of good reduction of $E/F$ (see remark 3.2.5(1)). Therefore the characteristic polynomial of the Frobenius Frob$_{K'/F}(z)=[\tau_\infty]$ above $z$ acting on $E_{l^n}$, the group scheme of $l^n$-torsion points of $E$, is equal to by (3.2.2)(a) $$X^2-1
\ {\rm (mod\ \ } l^n {\rm )}.$$
Comparing these two quadratic polynomials modulo $l^n$ proves the congruences in part (i) of the lemma.
\(ii) Let $y $ be the unique place of $K$ lying over the place $z$ of $F$ where $z\in \Lambda^1(n)$. Then $\kappa(y)$ is a quadratic extension of $\kappa(z)$. Furthermore, as Frob$_{K'/F}(z)=[\tau_\infty]$ where $\tau_\infty$ has order 2 (see (3.2.2), remark 3.2.5(2)), the prime $y$ splits completely in the extension $K'/K$. The map of reduction modulo a prime of $K'$ over $z$ $$E(K')_{l^n}\to {\cal E}_{0,z}(\kappa(y))_{l^n}$$ is an isomorphism. Hence we have $${\cal E}_{0,z}(\kappa(y))_{l^n}\cong
\bigg({{\mathds Z}\over l^n{\mathds Z}}\bigg)^2.$$ As $l$ is an odd prime number and the roots $\pm1$ of the characteristic polynomial $X^2-1$ of $\tau_\infty$ on $E_{l^n}$ are rational over the prime field ${\mathds Z}/l{\mathds Z}$, the action of $\tau_\infty $ on ${\cal E}_{0,z}(\kappa(y))_{l^n}$ decomposes into a sum over the eigenspaces of $\tau_\infty$. Hence we have for $\delta=\pm1$ $${\cal E}_{0,z}(\kappa(y))_{l^n}^{\delta}\cong
{\mathds Z}/l^n{\mathds Z}.$$ Furthermore, we have $${\cal
E}_{0,z}(\kappa(y))^{+}={\cal E}_{0,z}(\kappa(z)).$$ The result follows from this. ${\sqcap \!\!\!\!\sqcup}$
0.2in[*3.2.8. Remarks.*]{} (1) Let $z\in \Sigma_F$ be a prime divisor which is coprime to $\infty$, Supp$(I)$ and the discriminant of $K/F$ and which is inert in the field extension $K/F$. Then it can be shown, in a similar way to the proof of Lemma 3.2.7, that $z\in \Lambda^1(n)$ if and only if $$a_z\equiv \vert\kappa(z)\vert+1\equiv 0 \ ({\rm mod}\ l^n)$$ and the Frobenius Frob$(z)$ does [*not*]{} act as a homothety on $T_l(E)\otimes_{{\mathds Z}_l} {\mathds Z}/l^n{\mathds Z}$ where $a_z$ is the trace of the Frobenius Frob$(z)$ on the Tate module as in §3.2.6.
0.1in (2) Fix a prime $l\in \cal P$. For two integers $a,b\in
\mathds Z$ denote by $v_l(a,b)$ the valuation at $l$ of the greatest common divisor of $a$ and $b$. For a prime divisor $z$ of Spec $A$ define $\alpha(z)$ by the equation $$\alpha(z)=v_l(\vert \kappa(z)\vert +1, a_z)$$ where as in §3.2.6 $$a_z=\vert \kappa(z)\vert +1-\vert {\cal E}_0(\kappa(z))\vert$$ and where ${\cal E}_0$ is the closed fibre of the Néron model of $E$ at $z$.
The prime divisors $z$ of Spec $A$ in the set $\Lambda^1$ have the property that they are coprime to $I$, remain prime in $K/F$, and satisfy $$\alpha(z)=v_l(\vert \kappa(z)\vert +1, a_z)\geq 1.$$
0.1in (3) If $r>0$ and $c\in \Lambda^r$ we write $$\alpha(c)=\min_{z\in {\rm Supp}( c)} \alpha(z), \ \ \alpha(0)=+\infty.$$ We have the following equation $$\Lambda^r(n)=\{c\in \Lambda^r\ \vert \ \alpha(c)\geq n\}$$ where $n$ is any integer $\geq 1$; $\alpha(c)$ is the greatest integer $n$ such that $c\in \Lambda^r(n)$.
The set $\Lambda^r$ is then equipped with a decreasing filtration $$\Lambda^r=\Lambda^r(1)\supseteq \Lambda^r(2)\supseteq \Lambda^r(3)\supseteq \ldots$$
0.1in (4) In the notation of the monograph \[1, Chapter 7, §7.11\], we have $\Lambda(n)={\cal D}_{l^n}$
0.4in [**3.3. A refined Hasse principle for finite group schemes**]{}
0.2in The finite group schemes in question are $E_{l^n}/F$ and the Hasse principle concerns the localization at places of $F$ of finite groups of their principal homogenous spaces. 0.2in(3.3.1) Suppose that
0.2in
$E$ is an elliptic curve defined over $F$;
$K$ is an imaginary quadratic field over $F$ with respect to $\infty$:
$l\in {\cal P}$ is a, necessarily odd, prime number in the set $\cal P$ (see §3.1);
$n \geq 1$ is an integer;
$L_n$ is the field $K(E_{l^n})$ which is galois over $F$.
0.2in [**3.3.2 Lemma.**]{} [*The restriction map from $K$ to $L_n$ induces an isomorphism $$H^1(K,E_{l^n}(L_n))\
{\buildrel \cong \over \longrightarrow }\ {\rm Hom}_{{\rm Gal}(L_n/K)}
({\rm Gal}(L_n^{\rm ab}/L_n), E_{l^n}(L_n))$$ where $L_n^{\rm ab}$ is the maximal separable abelian extension of $L_n$.*]{} 0.2in [*Proof.*]{} The Hochschild-Serre spectral sequence, where as in (3.3.1) $L_n=K(E_{l^n})$, $$H^p(L_n/K,H^q(L_n,E_{l^n}(L_n)))\Rightarrow H^{p+q}(K,E_{l^n}(L_n))$$ gives rise to a short exact sequence of low degree terms $$0\to H^1(L_n/K,E_{l^n}(L_n))\to H^1(K,E_{l^n}(L_n))\to
H^1(L_n,E_{l^n}(L_n))^{{\rm Gal}(L_n/K)}\to$$ $$\to H^2(L_n/K,E_{l^n}(L_n))\to H^2(K,E_{l^n}(L_n)).$$ The two cohomology groups $H^1(L_n/K,E_{l^n}(L_n))$ and $H^2(L_n/K,E_{l^n}(L_n))$ are zero by the definition of $\cal P$ (Definition 3.1.3(b)); hence this exact sequence shows that restriction map of the lemma is an isomorphism. $$H^1(K,E_{l^n}(L_n))\to H^1(L_n,
E_{l^n}(L_n))^{{\rm
Gal}(L_n/K)}$$ is an isomorphism of [Gal]{}$(L_n/F)$-modules. The isomorphism of the lemma follows immediately. ${\sqcap \!\!\!\!\sqcup}$
0.2in[**3.3.3. Proposition.**]{} [*If $S$ is a finite subgroup of $H^1(K,E_{l^n})$ then there is a finite abelian extension $L_{S,n}$ of $L_n$ and an isomorphism of Gal$(L_n/{K})$-modules $${\rm Gal}(L_{S,n}/L_n)
\cong {\rm Hom}(S,E_{l^n}(L_n)) \leqno{(3.3.4)}$$ $$\!\!\!\!\!\!\!\sigma\mapsto \phi_\sigma$$ and an isomorphism of abelian groups (if $S$ is a ${\mathds Z}[{\rm Gal}(K/F)]$-module then an isomorphism of ${\mathds Z}[{\rm Gal}(K/F)]$-modules) $$S\cong {\rm Hom}_{{\rm Gal}(L_n/K)}({\rm Gal}(L_{S,n}/L_n), E_{l^n}(L_n)).$$* ]{}
0.2in For the proof of Proposition 3.3.3, see \[1, Chapter 7, Corollary 7.18.10\]. Note that the set of prime numbers ${\cal P}\setminus F$ of \[1, Chapter 7, (7.18.1) and corollary 7.18.10\] coincides with the set $\cal P$ of prime numbers defined in §3.1 above. ${\sqcap \!\!\!\!\sqcup}$
0.2in(3.3.5) For a prime divisor $v$ of $K$, denote by $${\rm res}_{v}: H^1(K,E_{l^n})\longrightarrow H^1(K_{v},E_{l^n})$$ the restriction homomorphism at $v$ where $K_v$ is the completion of $K$ at $v$.
Let $S$ be a finite subgroup of $H^1(K, E_{l^n})$. For $s\in S$ and $v$ a place of $K$, we have in the notation of Proposition 3.3.3 applied to the subgroup $S$ $${\rm res}_v(s)=0\Leftrightarrow
\phi_\sigma(s)=0{\rm \ for \ all
\ } \sigma\in D_{v'}$$ where $v'$ is a prime divisor of $L_n$ above $v$ and $D_{v'}$ is the decomposition group of a prime divisor of $L_{S,n}$ above $v'$ and where $\phi_\sigma\in {\rm Hom}(S,E_{l^n}(L_n))$ as in (3.3.4).
0.2in(3.3.6) Let $\tau_\infty\in $ Gal$(L_n/{F})$ be the element of order $2$ of paragraph (3.2.2). The element $\tau_\infty$ acts as $-1$ on the $l^n$th roots of unity and preserves the Weil pairing on $E_{l^n}$ (referring to the exact sequence (1.11.3), det$(\tau_\infty)=-1$ acts on the $l^n$th roots of unity).
We obtain (see paragraph (3.2.3)) as $l$ is odd the decompositions into eigenspaces under the action of $\tau_\infty$ $$H^1(K, E_{l^n})\cong H^1(K, E_{l^n})^+\oplus H^1(K, E_{l^n})^-$$ $$E_{l^n}(L_n) \cong E_{l^n}(L_n)^+\oplus E_{l^n}(L_n)^-.$$ We obtain that the $+1$ eigenspace $${\rm Hom}(H^1(K, E_{l^n}), E_{l^n}(L_n))^{+}$$ is isomorphic to the profinite group of ${\mathds Z}[\tau_\infty]$-homomorphims from $H^1(K, E_{l^n})$ to $E_{l^n}(L_n)$ and hence we obtain an isomorphism between this $\tau_\infty$-invariant subgroup and the profinite Pontrjagin dual of the discrete torsion abelian group $H^1(K,E_{l^n})$ namely we have an isomorphism, where a basis of $T_l(E)$ is fixed as in (1.12.2), $${\rm Hom}(H^1(K,E_{l^n}), E_{l^n}(L_n))^{+}\ \cong\ H^1(K,E_{l^n})^{*{}}$$ where the Pontrjagin dual is given by $$H^1(K,E_{l^n})^{*{}}={\rm Hom}(H^1(K,E_{l^n}), {\mathds Q}/{\mathds Z})
.$$ For a finite subgroup $S$ of $H^1(K, E_{l^n})$ there is similarly an isomorphism $${\widehat S}\cong {\rm Hom}(S, E_{l^n}(L_n))^+.$$ 0.2in [**3.3.7. Proposition.**]{} [*Let $S$ be a finite subgroup of $H^1(K,E_{l^n})$ and let $\chi\in {\widehat S}$. For any integer $t\geq n$, there is a set of positive Dirichlet density of prime divisors $z\in \Lambda^1(t)$ of $F$ unramified in $L_t=K(E_{l^t})$ such that $$\chi=\phi_{{\rm Frob}(z^\times)}$$ for some prime divisor $z^\times$ of $L_t$ lying above $z$, where $\phi_{{\rm Frob}(z^\times)}
\in {\rm Hom}(S,E_{l^t}(L_t))^+$ is as in (3.3.4).* ]{}
0.2in [*Proof.*]{} By Lemma 3.1.5, there is an injection of cohomology groups for all $t\geq n$ $$H^1(K, E_{l^n})\to H^1(K,E_{l^t})$$ obtained from the inclusion of finite group schemes $E_{l^n}\subseteq E_{l^t}$. The finite subgroup $S$ is then a subgroup of $H^1(K,E_{l^t})$ for all $t\geq n$. Applying Proposition 3.3.3 to $S$ as a subgroup of $H^1(K,E_{l^t})$ we obtain the abelian extension $L_{S,t}/L_t$ where we write $L_t=K(E_{l^t})$.
By Proposition 3.3.3 there is an element $\sigma\in $ Gal$(L_{S,t}/L_t)$ such that $$\chi=\phi_\sigma.$$ We have $${\rm Gal}(L_{S,t}/L_t)^+\cong {\widehat S}$$ from the isomorphism of (3.3.4) and the isomorphism of paragraph (3.3.6) $${\widehat S}\cong {\rm Hom}(S, E_{l^t}(L_t))^+.$$ In particular we have $\phi_\sigma^{\tau_\infty}=\phi_\sigma$. As the order of Gal$(L_{S,t}/L_t)$ is odd, we then have $\sigma=\rho^{\tau_\infty}.\rho$ for some $\rho\in $ Gal$(L_{S,t}/L_t)$. By the Chebotarev density theorem there is a set of positive Dirichlet density of prime divisors $z\in \Sigma_F$ of $F$ such that Frob$(z)$ in Gal$(L_{S,t}/{F})$ contains $\tau_\infty\rho$ and where $z$ unramified in $L_{S,t}/{F}$. Note the the finitely many prime divisors ramified in the field extension $L_{S,t}/{F}$ depend only on $S$ and not on $t$.
Since the restriction of $\tau_\infty\rho$ to $L_t$ is $\tau_\infty$, we have $z\in \Lambda^1(t)$. We then have that $z$ has residue class extension degree $2$ in $L_t/{F}$ for any prime divisor above $z$ in $L_t$. Hence for any $z^\times$, a prime divisor of $L_t$ lying above $z$, we have $${\rm Frob}(z^\times)=(\tau_\infty\rho)^2=\rho^{\tau_\infty}.\rho=\sigma.\ \ \ {\sqcap \!\!\!\!\sqcup}$$ 0.2in
[**3.3.8. Proposition.**]{} [*Let $S$ be a finite subgroup of $H^1(K,E_{l^n})$ and let $\chi\in {\widehat S}$. Then for any integer $t\geq 1$ there is a set of positive Dirichlet density of prime divisors $z\in \Lambda^1(t)$ such that for the prime divisor $y$ of $K$ lying over $z$, there is a commutative diagram of group homomorphisms $$\matrix{ S&{\buildrel {\rm res}_{y}\over \longrightarrow}&
{\rm res}_{y}(S)\cr
&\chi \searrow &\cong \downarrow \psi\cr
&&\chi(S)\cr}$$ where $\psi$ is an isomorphism of finite cyclic groups.*]{}
0.2in [*Proof.*]{} As $\Lambda^1(m)$, $m\geq 1$, is a decreasing filtration of $\Lambda^1(1)$, we may assume that $t\geq n$. By Lemma 3.1.5, there is an injection of cohomology groups for all $t\geq n$ $$H^1(K, E_{l^n})\to H^1(K,E_{l^t})$$ obtained from the inclusion of finite group schemes $E_{l^n}\subseteq E_{l^t}$. The finite subgroup $S$ is then a subgroup of $H^1(K,E_{l^t})$ for all $t\geq n$.
Select the prime divisor $z\in \Sigma_F$ as in Proposition 3.3.7 applied to $S,\chi$ and where the divisors $z$ are different from the finitely many prime divisors of $F$ where the cohomology classes of the finite group $S$ ramify, that is to say the finitely many prime divisors where the field extension $L_{S,t}/{F}$ of (3.3.4) is ramified. For such a $z$, let $y$ be the prime divisor of $K$ above $z$.
The decomposition group of $y$ in Gal$(L_{S,t}/K)$ is generated by the Frobenius element Frob$({y})$ as the field extension $L_{S,t}/K$ is unramified at $y$. By paragraph (3.3.5) we then have $${\rm ker}(S\ {\buildrel {\rm res}_{y} \over \longrightarrow}\
H^1(K_{y},E_{l^t}))={\rm ker}(\chi).$$ That is to say, the kernel in $S$ of the restriction homomorphism at $y$ of the classes of $S$ is equal to ker$(\chi)$. Hence the two finite cyclic groups $ {\rm res}_{y}(S)$, $ \chi(S)$ are isomorphic and we obtain the isomorphism $\psi$ of the commutative diagram of the corollary. ${\sqcap \!\!\!\!\sqcup}$
0.2in [**3.3.9. Corollary.**]{} [*We have $$\matrix{\Lambda^r(n)
\phantom{=\{0\}}&{\sl is
\ infinite \ for \ all \ integers \ \ }r, n\geq 1;\cr
\Lambda^0(n)=\{0\} &{\sl \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! for \ all \ }n\geq 1.\cr}$$*]{} 0.2in [*Proof.*]{} We have from Proposition 3.3.8 that $\Lambda^1(n)$ has infinitely many elements for all integers $n\geq 1$. As $\Lambda^r(n) $ consists of all sums of $r$ distinct prime divisors of $\Lambda^1(n)$ (see Definition 3.2.4) and as $\Lambda^0(n)=\{0\}$, the corollary then follows. ${\sqcap \!\!\!\!\sqcup}$
0.2in [*3.3.10. Remark.*]{} Proposition 3.3.8 implies that, under the hypotheses of the proposition, the kernel of the natural homomorphism $$\psi:H^1(K,E_{l^n})\to \prod_{z\in \Lambda^1(n)} H^1(K_{y(z)}, E_{l^n})$$ is zero, where $y(z)$ is the place of $K$ over $z\in \Lambda^1(n)$. To show this, it is sufficient to apply the proposition to a faithful character $\chi$ of the finite cyclic group $S$ generated by any element of $H^1(K,E_{l^n})$.
The [*Hasse principle*]{} for the group scheme $E_{l^n}$ and the set of prime divisors $\Lambda^1(n)$ is precisely that the kernel of the homomorphism $\psi$ is zero. It says that a prinicpal homogeneous space of $E_{l^n}$ which is locally trivial at all places of $F$ in $\Lambda^1(n)$ must be globally trival.
In this way, the proposition is a refined form of the Hasse principle for the finite group scheme $E_{l^n}$. It would be interesting to know to what other group schemes the localization properties of finite subgroups of cohomology groups given in Proposition 3.3.8 also hold.
\[For more details on the Hasse principle, see \[9, pp.142-150\].\] 0.4in
2=1 [**3.4. Drinfeld-Heegner points and the cohomology classes $\gamma_n(c), \delta_n(c)$**]{}
0.2in (3.4.2) 2 by 1 Let 0.2in $E/F$ be an elliptic curve equipped with an origin, that is to say $E/F$ is a
1-dimensional abelian variety;
$I$ be the ideal of $A$ which is the conductor of $E/F$ without the component
at $\infty$;
$\epsilon=\pm1$ be the sign in the functional equation of the $L$-function \[Lfunction2\] of the elliptic
curve $E/F$. 0.2in
Assume that (see also \[1, Chapter 4, §§4.3, 4.7, or Chapter 7, §7.6\]):
0.2in
($\alpha $) $E/F$ has split Tate multiplicative reduction at $\infty$ (see §1.7
or \[1, Chapter 4, §4.7\]);
($\beta $) $K$ is an (separable) imaginary quadratic field extension of $F$,
with respect to $\infty$, such that all primes dividing the conductor $I$
split completely in $K$.
0.2in
The hypothesis ($\alpha$) implies that $E/F$ is covered by the Drinfeld modular curve $X_0^{\rm Drin}(I)$ (see (3.4.4) below). The hypothesis ($\beta$) ensures the existence of Drinfeld-Heegner points on $X_0^{\rm Drin}(I)$. The two hypotheses together ensure that there are Drinfeld-Heegner points on the elliptic curve $E/F$. Note that from (3.4.7) to the end of this section §3.4, it is assumed that $K\not=F\otimes_{{\mathds F}_q}{\mathds F}_{q^2}$ where ${\mathds F}_q$ is the exact field of constants of $F$.
In this section we detail this construction of Drinfeld-Heegner points as well as defining the cohomology classes $\gamma_n(c), \delta_n(c)$.
0.2in (3.4.2) 2 by 1 Let $\lambda$ be any prime number of $\mathds Z$ distinct from the characteristic of $F$. Let $\rho$ be the 2-dimensional $\lambda$-adic representation of Gal$(F^{\rm sep}/F)$ corresponding to $E$ where $F^{\rm sep
}$ is the separable closure of $F$; that is to say $\rho $ is the continuous homomorphism $$\rho:{\rm Gal}(F^{\rm sep}/F)\to {\rm End}_{{\mathds Q}_\lambda}(H^1_{\rm \acute
et}(E\otimes_FF^{\rm sep},{\mathds Q}_\lambda)).$$ For each place $v$ of $F$ put$$a_v={\rm Tr}(\rho({\rm Frob}(v))\vert \ H^1_{\rm \acute
et}(E\otimes_FF^{\rm sep},{\mathds Q}_\lambda)^{T_v})$$ where $T_v$ is the inertia subgroup of Gal$(F^{\rm sep}/F)$ over $v$. The representation $\rho$ satisfies $$a_v\in {\mathds Z} {\rm \ for \ all \ } v\in \Sigma_F$$ (see \[1, Chapter 5, example 5.3.18\]).
0.2in(3.4.2) 2 by 1 Let
0.2in $B$ be the integral closure of $A$ in $K$;
$O_c$ be the order of $K$, with respect to $A$, and of conductor $c$ for any divisor
$c\in$ Div$_+(A)$, (see \[1, Chapter 2, §2.2\]);
$\tau$ be the non-trivial element of Gal$(K/F)$;
$IB=I_1I_2$ be a factorisation of ideals of $B$ where $I_1$, $I_2$ are ideals of $B$ such
that $I_2^\tau=I_1$ where $I$ is the ideal of $A$ which is the conductor of $E/F$
without the place at $\infty$, as in (3.4.1); such a factorisation exists
because of the hypothesis that the prime ideal components of $I$ split
completely in $K/F$ (by (3.4.1)($\beta$)).
0.2in
As in §1.4, $K[c]$ denotes the ring class field of $K$ of conductor $c\in $ Div$_+(A)$. In particular $K[0]$ is the Hilbert class field of $K$, that is to say $K[0]$ is the maximal unramified abelian extension of $K$ which is split completely at $\infty$.
0.2in
(3.4.2) 2 by 1 Let ${\cal H}(\rho)$ be the Heegner module of $\rho$ and $K/F$ with exceptional set\[hmod7\] of primes those dividing $I$ and the place $\infty$ with coefficients in $\mathds Z$ (see \[1, Chapter 5, §5.3\]), where $I$ is the conductor of $E$, without the component at $\infty$.
As in §1.7 (see also \[1, Chapter 4, §4.7\] and \[1, Appendix B\]), there is a finite surjective morphism of curves \[modsch10\] over $F$ under the hypothesis ($\alpha$) $$\pi:X_0^{\rm Drin}(I)\to E.$$We may translate $\pi$ in the group scheme $E$ so that $\pi^{-1}(0)$ consists of at least one cusp of $X_0^{\rm Drin}(I)$ (as in \[1, Chapter 4, (4.8.1)\]); this rigidifies the map $\pi$. The cusps of the modular curve $X_0^{\rm
Drin}(I)$ generate a torsion subgroup of the jacobian of this curve \[1, Chapter 2, Theorem 2.4.9\].
Let $a$ be a divisor class in the Picard group Pic$(O_c)$ of the order $O_c$. Assume that $c$ and $I$ are coprime. Then there is a Drinfeld-Heegner point $$(a,I_1,c)\in X_0^{\rm Drin}(I)(K[c])$$ which is a non-cuspidal point of $X_0^{\rm Drin}(I)$ and is rational over the ring class field $K[c]$.
This point $(a,I_1,c)$ is constructed as follows. Fix an embedding $K\to {\widehat{\overline{F}}}_\infty$ where ${\widehat{\overline{F}}}_\infty$ is the completion of the algebraic closure of $F_\infty$ which is the completion of $F$ at $\infty$. Let $L$ be a projective $O_c$-module of rank 1 in the class $a$ and contained as a lattice in ${\widehat{\overline{F}}}_\infty$. Then $I_1(O_c)=I_1\cap O_c$ is an invertible ideal of $O_c$ and $L'=I_1(O_c)^{-1}L$ is a projective $O_c$ module of rank 1 contained as a lattice in ${\widehat{\overline{F}}}_\infty$. Let $D$ and $D'$ be the rank 2 Drinfeld modules for $A$ over the field ${\widehat{\overline{F}}}_\infty$ corresponding respectively to the lattices $L$ and $L'$. Then $D$ and $D'$ have general characteristic and complex multiplication by $O_c$. The inclusion of $O_c$-modules $L\subset L'$ corresponds to an $I$-cyclic isogeny $f:D\to D'$, as its kernel is isomorphic as an $A$-module to $O_c/I_1(O_c)\cong A/I$. The pair $(D, {\rm ker}(f))$ defines the point $(a,I_1,c)$ on $X_0^{\rm Drin}(I)( {\widehat{\overline{F}}}_\infty)$. That this point $(a,I_1,c)$ is defined over $K[c]$ results from the main theorem of complex multiplication (see \[1, Chapter 4, §4.3\] for more details).
0.2in (3.4.2) 2 by 1 The image $$\pi(a,I_1,c)\in E(K[c])$$ is a Drinfeld-Heegner point of $E$ rational over the ring class field $K[c]$ and is written in the notation of \[1, Chap. 4, §4.8\] $$(a,I_1,c,\pi)\in E(K[c]).$$
By \[1, Chap. 5, example 5.3.18\] there is a homomorphism of discrete Gal$(K^{\rm
sep}/K)$-modules $$\begin{aligned}
{\cal H}(\pi):\quad {\cal H}(\rho)^{(0)}&\to E(F^{\rm sep})\\
<a,c>&\mapsto (a,I_1,c,\pi)\\\end{aligned}$$ for all $c\in {\rm Div}_+(A)$ coprime to $I$, $a\in {\rm Pic}(O_c)$. The image of this homomorphism ${\cal H}(\pi)$ consists of the $\mathds Z$-linear combinations of Drinfeld-Heegner points of $E$ rational over all the ring class fields $K[c]$ for all $c$.
Let $<0,0>$ be the element of the Heegner module ${\cal
H}(\rho)^{(0)}$ given by the principal class of Pic$(B)$, where $B$ is the integral closure of $A$ in $K$. Let $$(0,I_1,0,\pi)={\cal H}(\pi)(<0,0>)$$ be the corresponding Drinfeld-Heegner point of $E(K[0])$ (see \[1, Chapter 4, (4.8.2)\]). Let $$P_0={\rm
Tr}_{K[0]/K}(0,I_1,0,\pi)\in E(K).\leqno{(3.4.\the\count2)}$$2 by 1 That is to say, $P_0$ is the trace from $K[0]$ to $K$ of the point $(0,I_1,0,\pi)$; the point $P_0$ belongs to $E(K)$ and the point $(0,I_1,0,\pi)$ belongs to $E(K[0])$.
0.2in(3.4.2) 2 by 1 We now impose for the rest of this section the hypothesis that $K\not=F\otimes_{{\mathds F}_q}{\mathds F}_{q^2}$ where ${\mathds F}_q$ is the exact field of constants of $F$, that is to say $K$ is not obtained from $F$ by ground field extension.
Let $c\in \Lambda(1)$ that is to say $c$ is a sum of distinct prime divisors of $\Lambda^1(1)$ with multiplicity $1$. Let $$y_c=\pi(0,I_1,c)\in E(K[c]).$$ so that $y_c=(0,I_1,c,\pi)$. The field $K[0]$ is the Hilbert class field of $K$. Let $$G_c={\rm Gal}(K[c]/K[0]).$$ As $ K\not=F\otimes_{{\mathds F}_q}{\mathds F}_{q^2}$ by hypothesis, we have that $B^*=A^*$, where $A^*,B^*$ are the unit groups of $A,B$, and hence (by \[1, Chap. 2, (2.3.8), p.19\]) there is a group isomorphism $$G_c\cong \prod_z G_{z}$$ where the product runs over the prime divisors $z$ in the support of $c$ and $$G_{z}={\rm Gal}(K[c]/K[c-z])$$ is a cyclic group of order $\vert\kappa(z)\vert+1$ as $z$ is inert in $K/F$ (by \[1, Chapter 2, (2.3.12), p.20\]). Fix a generator $\sigma_z$ of the cyclic group $G_z$ for all prime divisors $z\in \Lambda^1(1)$.
0.2in(3.4.2) 2 by 1 Write for any prime divisor $z\in \Lambda^1(1)$ of $F$, where $D_z\in {\mathds Z}[G_z]$, $$D_z=-\sum_{i=1}^{\vert\kappa(z)\vert} i.\sigma_z^i.$$ Here $D_z$ is the Kolyvagin element of the map $h_z: G_z\to {\mathds Z}$ where $\sigma_z^{-i}\mapsto -i$ for all $0\leq i\leq \vert \kappa(z)\vert$. Note the minus sign here which agrees with the Kolyvagin elements of \[1, Chap. 5, pp. 175-178\].
0.2in
(3.4.2) 2 by 1 For any divisor $c\in \Lambda(1)$, put $$D_c=\prod_{z\in \ {\rm Supp}( c)} D_z$$ where $D_c\in {\mathds Z}[G_c]$. Then $D_c$ is the Kolyvagin element of the map $h: G_c\to {\mathds Z}$ given by $h=\prod_{z\in {\rm Supp}(z)} h_z$. Let $${\cal G}_c={\rm Gal}(K[c]/K)$$ where there is an exact sequence of finite abelian groups $$0\longrightarrow G_c\longrightarrow {\cal G}_c\longrightarrow {\rm Gal}(K[0]/K)\longrightarrow 0.$$ Let $\cal S$ be a set of coset representatives for $G_c$ in ${\cal G}_c$. Define the point $P_c\in E(K[c])$ by, where $y_c=\pi(0,I_1,c)$ is the element of $E(K[c])$ in (3.4.7), $$P_c=\sum_{s\in {\cal S}} sD_cy_c.$$
0.2in
(3.4.2) 2 by 1 Suppose now that $c\in \Lambda(n)$. We write $P_c$ (mod $l^n$) for the image of $P_c$ in the quotient group ${}_{l^n}E(K[c])$. Then we have that $P_c$ (mod $l^n$) belongs to $$P_c \ ({\rm mod\ } l^n)\ \in \Bigg({}_{l^n}E(K[c])\Bigg)^{{\cal G}_c}.$$ This inclusion follows immediately from the formula in ${\mathds Z}[G_c]$ $$(\sigma_z-1)D_z=-\vert G(c/c-z)\vert +\sum_{g\in G(c/c-z)} g$$ for all $z\in $ Supp$(c)$ and that $\vert G(c/c-z)\vert=\vert\kappa(z)\vert+1$ is divisible by $l^n$ for all $z\in $ Supp$(c)$ (for a detailed proof of this inclusion $P_c \ ({\rm mod\ } l^n)\ \in ({}_{l^n}E(K[c]))^{{\cal G}_c}$ see \[1, Chapter 7, Lemma 7.14.9 and Lemma 7.14.11\]). Furthermore, we have where $y_0$ is defined in (3.4.7) and where this notation $P_0$ agrees with that of (3.4.6) $$P_0={\rm Tr}_{K[0]/K}(y_0)\in E(K).$$
\[The point $P_c$ (mod $l^n$) in $E(K[c])/ l^nE(K[c])$ coincides with the point denoted $P_c$ in the book \[1, Chapter 7, Notation 7.14.10(iii) and Lemma 7.14.11\].\]
0.2in(3.4.2) 2 by 1 Let $S$ be the ring ${\mathds Z}/m{\mathds Z}$ where $m$ is any non-zero integer. The morphism of multiplication by $m$ on the elliptic curve $E$ then provides for any divisor $c$ of Div$_+(A)$ prime to $I$ the following commutative diagram with exact rows and an exact right-hand column (see \[1, Chap. 7, §7.14.5\]):0.2in $$\matrix{&&&&&0&&\cr
&&&&&\downarrow&&\cr
&&&&&\hidewidth H^1(K[c]/K,E(K[c]))_{m}\hidewidth&&\cr
&&&&{_j}&{\rm inf}\downarrow&&\cr
0\to&{_{m}}\!E(K)&\to&H^1(K,E_{m})&\to&H^1(K,E)_{m}&\to&0\cr
&\downarrow&&{\rm res}\downarrow{\rm quasi-isom.}&&{\rm res}\downarrow&&\cr
0\to&(_{m}\!E(K[c]))^{{\cal
G}_c}&{\buildrel\partial\over\to}&H^1(K[c],E_{m})^{{\cal
G}_c}&\to&H^1(K[c],E)_{m}^{{\cal G}_c}&&\cr
&f\uparrow&&&&&&\cr
&({\cal H}_{c,S}^{(0)})^{{\cal
G}_c}&&&&&&\cr}\leqno{(3.4.\the\count2) }$$ 2 by 1 Here $({\cal H}_{c,S}^{(0)})$ is the $c$-component Heegner module with coefficients in the ring $S$ (see \[1, Chap. 7, (7.11.3), and Chap. 5, §5.3\]).
Let $\cal E$ be the finite exceptional set of prime numbers of Propositions 1.10.1 and 1.10.2. The middle restriction homomorphism here in (3.4.12) is a quasi-isomorphism of quasi-groups in $[{\bf
N}^{(p)}]_{\mathds Z}$ where the finite exceptional set of prime numbers is $\cal E$ and is independent of $c$ (Propositions 1.10.1 and 1.10.2, or alternatively \[1, Chap. 7, Proposition 7.14.2\]). In particular this middle restriction homomorphism is an isomorphism for all integers $m$ which are powers of prime numbers of $\cal P$ (Definition 3.1.3) as $\cal P$ excludes the finitely many prime numbers of $\cal E$.
The map $f$ is obtained by sending a generator $<a,c>$ of ${\cal H}^{(0)}_{c,S}$ to its image $(a,I_1,c,\pi)\in E(K[c])$ as in (3.4.4) and (3.4.5).
This diagram (3.4.12) then provides the [ Heegner homomorphism]{}, for all integers $m\in \mathds N$ prime to $\cal E$,$$({\cal H}_{c,S}^{(0)})^{{\cal
G}_c}\to H^1(K,E)$$whose image belongs to $ H^1(K[c]/K,E(K[c]))_{m}$.
0.2in(3.4.2) 2 by 1 Take $m=l^n$ where $l\in \cal P$ so that the middle quasi-isomorphism in diagram (3.4.12) is an isomorphism. From the diagram (3.4.12), define $\gamma_n(c)$ and $\delta_n(c)$ by the formulae $$P_c \ ({\rm mod}\ l^n)\ \in \ \Bigg({_{l^n}E(K[c]))}\Bigg)^{{\cal G}_c};$$ $$\gamma_n(c){\rm \ is \ the \ image\ of \ } P_c\ {\rm ( mod\ } l^n{\rm )\
in \ } H^1(K, E_{l^n});$$ $$\delta_n(c){\rm \ is \ the \ image\ of \ } \gamma_n(c)\ \
{\rm in \ } H^1(K, E)_{l^n}.$$ \[This is the same as the construction in \[1, Chap. 7, §7.14.10\], where we write here $\gamma_n(c), \delta_n(c)$ in place of $\gamma(c), \delta(c)$ to indicate their dependence on the integer $n$.\]
0.2in[**3.4.2 2 by 1. Proposition.**]{} (i) [*The order of $\gamma_{n}(c)$ is equal to the order of $P_c$ [(mod $l^n$)]{} in ${}_{l^n}E(K[c])$.* ]{}
\(ii) [*The exponent $t$ of the order $l^t$ of $\delta_{n}(c)$ is the least integer $t$ such that $$l^t(P_c\ ({\rm mod} \ l^n))\in { l^nE(K[c])+E(K)\over l^nE([c])}.$$*]{}
0.2in [*Proof.*]{} Here $P_c$ (mod $l^n$) denotes the image of $P_c\in E(K[c])$ in ${}_{l^n} E(K[c])$. These results on the orders of $\gamma_n(c)$ and $\delta_n(c)$ follow immediately from their definition and the commutative diagram (3.4.12). ${\sqcap \!\!\!\!\sqcup}$ 0.2in
[**Chapter 4. Structure of the Tate-Shafarevich group and the Selmer group**]{} 0.4in [**4.1. Statement of the main theorems**]{} 0.2in This section contains no proofs. The main Theorems 4.1.9, 4.1.13, 4.1.14 and 4.1.15 of this paper are finally proved in §§5.3-5.5. 0.2in(4.1.1) Throughout this chapter 4, we assume that
0.2in
$E/F$ is an elliptic curve where $F$ is a global field of characteristic $p>0$;
$I$ is the ideal of $A$ which is the conductor of $E/F$ without the component
at $\infty$;
$K$ is an imaginary quadratic extension field of $F$ with respect to $\infty$;
$l\in \cal P$ is a, necessarily odd, prime number in the set of prime numbers $\cal P$
(see §3.1);
$\epsilon=\pm1$ is the sign of the functional equation of the $L$-function $L(E/{F},s)$
of $E/F$;
$\tau$ is the element of order $2$ of the galois group Gal$(K/F)$;
$P_c\in E(K[c])$ are the points defined in (3.4.9) over the ring class fields $K[c]$
for all divisors $c\in \Lambda(1)$, and where $P_0$ belongs to the group $E(K)$
of $K$-rational points;
$\alpha(c)$, for a divisor $c\in \Lambda(1)$, is the largest integer $n$ such that $c\in \Lambda(n)$ if $c\not=0$
and such that $\alpha(0)=+\infty$ (see remarks 3.2.8).
0.2in
Assume that $E,K,F$ satisfy the following hypotheses (as in \[1, Chapter 7, §7.6.1\]):
0.2in (a) $\infty $ is a place of $F$ with residue field equal to $k$;
\(b) $E/F$ has split Tate multiplicative reduction at $\infty$ (see §1.7
or \[1, Chapter 4, §4.7\]);
\(c) $K$ is an (separable) imaginary quadratic field extension of $F$,
with respect to $\infty$, such that all primes dividing the conductor $I$
split completely in $K$ and $ K\not=F\otimes_{{\mathds F}_q}{\mathds F}_{q^2}$.
0.2in
These hypotheses (a),(b),(c) are assumed for the rest of this paper.
0.2in (4.1.2) If $M$ is a ${\rm Gal}(K/F)$-module denote by $M^+$ the submodule of $M$ on which $\tau $ acts by $+1$ and $M^-$ is similarly the submodule on which $\tau $ acts by $-1$.
0.2in(4.1.3) Let $G$ be a finite abelian $l$-group. The [*invariants*]{} of $G$ are the integers $${r_1}\geq {r_2}\geq {r_3}\geq \ldots$$such that $G$ decomposes into elementary components $$G\cong {{\mathds Z}\over l^{r_1}{\mathds Z}}\oplus
{{\mathds Z}\over l^{r_2}{\mathds Z}}\oplus {{\mathds Z}\over l^{r_3}{\mathds Z}}\oplus \ldots$$ The integers $r_1,r_2,r_3\ldots$ are uniquely determined by $G$; the integers $l^{r_1}, l^{r_2},l^{r_3}, \ldots$ are also sometimes called the invariants of $G$.
0.2in(4.1.4) As the elliptic curve $E$ is defined over $F$, the $l$-power torsion subgroup $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}$ of the Tate-Shafarevich group $\coprod\!\!\!\coprod$$(E/K)$ of the elliptic curve $E\times_FK$ over $K$ decomposes into eigenspaces $$\coprod\!\!\!\!\coprod(E/K)_{l^\infty} \cong \coprod\!\!\!\!\coprod(E/K)^+_{l^\infty} \oplus \coprod\!\!\!\!\coprod(E/K)^-_{l^\infty}$$ under the action of the element $\tau\in $ Gal$(K/F)$.
The Cassels pairing on $\coprod\!\!\!\coprod$$(E/K)$ is anti-symmetric and respects this decomposition into eigenspaces; furthermore, the Cassels pairing is non-degenerate if this Tate-Shafarevich group $\coprod\!\!\!\coprod$$(E/K)$ is finite; therefore if $\coprod\!\!\!\coprod$$(E/K)$ is finite, the invariants of the finite abelian $l$-group $\coprod\!\!\!\coprod$$(E/K)_{l^\infty} $ have even multiplicity.
0.2in(4.1.5) Under the hypothesis that the Tate-Shafarevich group $\coprod\!\!\!\coprod(E/K)_{l^\infty}$ be finite, let $$N_1\geq N_3\geq N_5\ldots$$ and $$N_2\geq N_4\geq N_6\ldots$$ be integers such that $${N_1},{N_1},{N_3},{N_3}, N_5,N_5\ldots$$ are the invariants of the finite abelian $l$-group $\coprod\!\!\!\coprod(E/K)_{l^\infty}^{\epsilon} $, which is the $\epsilon$-eigenspace, and $${N_2},{N_2},{N_4},{N_4}, N_6,N_6\ldots$$ are the invariants of the finite abelian $l$-group $\coprod\!\!\!\coprod(E/K)_{l^\infty}^{-\epsilon} $, which is the $-\epsilon$-eigenspace.
That is to say, putting $\nu(r)=(-1)^r\epsilon$ then we have isomorphisms for $r=1$ or $r=2$ $${\coprod\!\!\!\!\coprod}(E/K)^{-\nu(r)}_{l^\infty}\cong \bigoplus_{s\geq 0}({{\mathds Z}/l^{N_{2s+r}}{\mathds Z}})^2.$$ 0.2in [**4.1.6. Definition.**]{} Let $m\geq 1$ be an integer and $c\in \Lambda(m)$. Then the point $P_c$, defined in (3.4.9), belongs to $E(K[c])$ and the point $P_0$ belongs to $E(K)$.
\(1) Write $$l^m\vert P_c{\rm \ \ if \ \ } P_c\in l^mE(K[c])$$ and $$l^m\vert\vert P_c{\rm \ \ if \ \
}P_c\in l^mE(K[c]) \setminus l^{m+1}E(K[c]).$$
\(2) For any integer $r\geq 0$, let $E(r)$ be the abelian group $$E(r)=\bigoplus_{c\in \Lambda^r(1)} {}_{A(c)}E(K[c])$$ where the sum runs over all divisors $c$ of $\Lambda^r(1)$ and where $A(c)=l^{\alpha(c)}$. If $r>0$ the group $E(r)$ is then a direct sum of finite abelian $l$-groups of the form $E(K[c])/l^{\alpha(c)}E(K[c])$. If $r=0$ then $E(0)$ is defined conventionally to be $E(K[0])$.
A point $P_c\in E(K[c])$, where $c\in \Lambda^r(1)$, induces an element in ${}_{A(c)}E(K[c])$ and hence an element of $E(r)$ where all its components in $E(r)$ are zero except possibly for that in ${}_{A(c)}E(K[c])$.
\(3) Let $P(r)$ be the subgroup of $E(r)$ generated by the images in $E(r)$ of the points $P_c$ for all $c\in \Lambda^r(1)$.
Define $M_r$ for any $r\geq 0$ to be the largest integer $n\geq 0$ such that $$P(r)\subseteq l^nE(r).$$ If $P(r)=0$ then there is no such largest integer $M_r$ and we then put $$M_r=+\infty.$$
0.2in
[**4.1.7. Conjecture.**]{} For some $r\geq 0$ we have $M_r<+\infty$.
0.2in
\[See \[7, Conjecture A\], and also \[1, Conjecture 7.14.19\] for an explanation of this conjecture.\] 0.2in [*4.1.8. Remarks.*]{} (1) The order of $\gamma_n(c)$ is $l^{n-m}$ for all $n\geq m$ if and only if $l^m\vert\vert P_c$ by Proposition 3.4.14.
If $M_r$ is finite then there is a divisor $c\in \Lambda^r(1)$ such that $\gamma_{{{M_r}+1}}(c)\not=0$ and $\gamma_{{{M_r}}}(c)=0$; furthermore, $\gamma_{M_r}(c)=0$ for any $c\in \Lambda^r(1)$ from Proposition 3.4.14(i).
\(2) We have $M_0<+\infty$ if and only if $P_0$ has infinite order in $E(K)$.
\[This equivalence requires that $E(K)$ has no $l$-torsion; but this requirement holds by the restriction on the prime number $l$ given by Definition 3.1.3(b). See Lemma 5.1.2 below for details.\]
\(3) We have that $\alpha(c)$ is finite if and only if $c\not=0$ (see Remarks 3.2.8(2) and (3)). Write if $c\not=0$ $${\rm ord}_l(P_c)=\cases{{\displaystyle \max\{ N: l^N\vert P_c\} }&
provided that this maximum is $< \alpha(c)$ \cr
&\cr
{\displaystyle +\infty}& if $\max\{ N: l^N\vert P_c\}$ is $\geq \alpha(c)$.\cr
}$$ Write $${\rm ord}_l(P_0)= \max\{ N: l^N\vert P_0\}$$ which is either an integer or $+\infty$. If $c\not=0$, we have ord$_l(P_c)<+\infty$ if and only if $l^{{\rm ord}_l(P_c)}\vert\vert P_c$. and ord$_l(P_c)<\alpha (c)$.
\(4) By definition $M_r$, for all integers $r\geq 0$, is given by $$M_r=\min\{ {\rm ord}_l(P_c)\vert\ \
c\in \Lambda^r(1)\}.$$
Note that $M_r<+\infty$ if and only if there is $c\in \Lambda^r(1)$ such that $${\rm ord}_l(P_c)< \alpha(c)$$with an evident interpretation in the case where $r=0$, $c=0$ and $\alpha(0)=+\infty$. Furthermore, $M_r<+\infty$ implies that for this selection of $c\in \Lambda^r(1)$ we would have $l^{M_r}\vert\vert P_c$ and $c\in \Lambda^r(\alpha(c))$ and $M_r<\alpha(c)$.
Alternatively, if $r>0$ then we have $M_r<+\infty$ if and only if for some $c\in \Lambda^r(1)$ we have $$c\in \Lambda^r(m+1)\setminus \Lambda^r(m+2)$$ and $${\rm ord}_l(P_c)\leq m.$$
0.2in [**4.1.9. Theorem.**]{} [*Suppose that $P_0$ has infinite order in $E(K)$, the group of $K$-rational points of $E$. Let $l$ be a prime number in $\cal P$ coprime to the order of [Pic]{}$(A)$ the Picard group of $A$. Then the Tate-Shafarevich group ${\coprod\!\!\!\coprod}(E/K)$ is finite and the invariants ${N_i}$, with multiplicity $2$, of the subgroup ${\coprod\!\!\!\coprod}(E/K)_{l^\infty}$ are given by $$N_i=M_{i-1}-M_i, {\sl \ \ for \ \ } i\geq 1.$$ That is to say, we have the isomorphisms of eigenspaces $${\coprod\!\!\!\!\coprod}(E/K)^\epsilon_{l^\infty}\cong \prod_{i{\rm \ even}\atop i\geq 0}
({\mathds Z}/l^{M_i-M_{i+1}}{\mathds Z})^2$$ $${\coprod\!\!\!\!\coprod}(E/K)^{-\epsilon}_{l^\infty}\cong \prod_{i{\rm \ odd}\atop
i\geq 0}
({\mathds Z}/l^{M_i-M_{i+1}}{\mathds Z})^2.$$*]{}
0.2in [*4.1.10. Remarks.*]{} (1) The image of the Drinfeld-Heegner point $P_0$ in $E(K)/E(K)_{\rm tors}$ is an eigenvector for $\tau$ and its eigenvalue is equal to $-\epsilon$ (by \[1, Chapter 4, Theorem 4.8.6, p. 98\], see also \[1, Chapter 7, Lemma 7.14.11, p. 388\]) where $E(K)_{\rm tors}$ is the torsion subgroup of $E(K)$.
\(2) If the Drinfeld-Heegner point $P_0$ has infinite order in $E(K)$ then by \[1, Chapter 7, Theorem 7.7.5 and Remarks 7.7.6(3)\] the Tate-Shafarevich group $
\coprod\!\!\!\coprod$$(E/K)$ is finite and therefore the alternating Cassels pairing on $\coprod\!\!\!\coprod$$(E/K)$ is non-degenerate and by §§4.1.4,4.1.5 the invariants of the $\tau$-eigenspaces of $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}$ have even multiplicity.
0.2in The next corollary is an immediate consequence of Theorem 4.1.9.0.2in [**4.1.11. Corollary.**]{} [*Under the hypotheses of Theorem 4.1.9, the integers $M_i$ satisfy $$M_i-M_{i+1}\geq M_{i+2}-M_{i+3}\geq 0, \ {\sl for \ all\ } i\geq 0,$$ and if $j$ is such that $$M_{j}=M_{j+1}=M_{j+2}$$ then the sequence $M_j, M_{j+1}, M_{j+2},M_{j+3}\ldots $ is constant.*]{} ${\sqcap \!\!\!\!\sqcup}$
0.2in [**4.1.12. Definition.**]{} Let $$G=G_1\times\ldots\times G_r$$ be a direct product of finite cyclic groups $G_i$. The characters $\chi_1,\ldots,\chi_r$ of $G$ form a [*triangular basis for the dual ${\widehat G}$*]{}, relative to the product $G=\prod_iG_i$, if they generate ${\widehat G}$ and $$\chi_i(G_j)=0, {\rm \ \ for \ all \ }j>i.$$
0.2in [**4.1.13. Theorem.**]{} [*Assume that $P_0$ has infinite order in $E(K)$ and $l\in \cal P$ is a prime number coprime to the order of the Picard group [Pic]{}$(A)$. Suppose that the direct product $$D=\prod_{i\geq 1} D_i$$ is a maximal isotropic subgroup of $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}$ for the Cassels pairing and $D_i$ is a finite cyclic group of order $l^{N_i}$ for all $i\geq 1$ and we have direct products $$D^{\epsilon }=\prod_{i{\rm \ odd}} D_i$$ $$D^{-\epsilon} =\prod_{i{\rm \ even}} D_i.$$ Then there are effective divisors $c_1\leq c_2\leq \ldots$ on $F$ such that $c_i\in \Lambda^i(M_{i-1})$ for all $i\geq 1$ and the characters $$d\mapsto <d,\delta_{M_{i-1}}(c_i)>_{\rm Cassels}, \ {\sl \ \ for \ all \ } i$$ form a triangular basis of characters of $D$ relative to the product $D=\prod_iD_i$.*]{} 0.2in [**4.1.14. Theorem.**]{} [*Assume that $P_0$ has infinite order in $E(K)$ and $l\in {\cal P}$ is coprime to the order of [Pic]{}$(A)$. Let $$M_\infty=\min_{i\in {\bf N}} M_i.$$ Then the group ${\mathds Z}P_0$ has finite index in $E(K)$ and the highest power of $l$ dividing the index $[E(K):{\bf Z}P_0]$ equals $l^{M_0}$ that is to say $$\vert (E(K)/{\bf Z}P_0)_{l^\infty}\vert =l^{M_0}.\ \ $$Furthermore, we have $$\vert\ { {\coprod\!\!\!\!\coprod}(E/K)_{l^\infty}}\ \vert =l^{2(M_0-M_\infty)}.$$* ]{}
0.2in[**4.1.15. Theorem.**]{} [*Assume that $P_0$ has infinite order in $E(K)$ and $l\in {\cal P}$ is coprime to the order of [Pic]{}$(A)$. Then for all integers $m\geq 0$, the natural surjection from the Selmer group to the Tate-Shafarevich group $$\pi_m: {\rm Sel}_{l^m}(E/K)\longrightarrow
{\coprod\! \!\!\!\coprod}(E/K)
_{l^m}\leqno{(4.1.16)}$$ splits and we have isomorphisms of eigenspaces $${\rm Sel}_{l^m}(E/K)^{\pm} \cong
\big({}_{l^m}E(K) \big)^{\pm}\oplus
{\coprod\! \!\!\!\coprod}(E/K)
_{l^m}^\pm {\rm \ \
for \ all \ } m\geq 0.\leqno{(4.1.17)}$$ Furthermore, we have isomorphisms for all $m\geq 0$ $$\matrix{\big( {}_{l^m}{E(K)}\big)^{-\epsilon}&\cong&
{\mathds Z}/l^m{\mathds Z} \cr
\big( {}_{l^m}{E(K)}\big)^{\epsilon}&\cong&
0.\cr}
\leqno{(4.1.18)}$$* ]{}
0.2in
The next corollary follows immediately from Theorems 4.1.9 and 4.1.15.
0.2in[**4.1.19. Corollary.**]{} [*Under the hypotheses of Theorem 4.1.15, put $${\widehat
N_i}=\min(m,M_{i-1}-M_i)$$ for all $i$ and for all integers $m$. Then for all integers $m\geq 0$, the invariants of the Selmer $-\epsilon$-eigenspace ${\rm Sel}_{l^m}(E/K)^{-\epsilon} $ are $m, {\widehat N_2}, {\widehat N_2},
{\widehat N_4},{\widehat N_4},\ldots$ and the invariants of the Selmer $\epsilon$-eigenspace ${\rm Sel}_{l^m}(E/K)^{\epsilon} $ are ${\widehat N_1},{\widehat N_1},{\widehat N_3},{\widehat N_3},\ldots$*]{} ${\sqcap \!\!\!\!\sqcup}$
0.2in
[*4.1.20. Remark.* ]{} The main result on the finiteness of Tate-Shafarevich groups proved in the monograph \[1, Chapter 7, Theorem 7.6.5 and Theorem 7.7.5\] is required for the proofs of the main results of this paper stated in this section; in particular, this paper does not provide a different proof of finiteness independent of the book \[1\]. This is also the prinicpal reason why the hypotheses (a), (b), (c) of §4.1.1 are required.
0.4in
[**4.2. Cochains for the cohomology classes $\gamma_n(c), \delta_n(c)$** ]{}
0.2in The notation and hypotheses of (4.1.1) and (4.1.2) hold in this section. 0.2in
[**4.2.1. Lemma.**]{} 2 by 1 [*Let $c\in \Lambda(n)$.*]{}
\(i) [*The cohomology class $\gamma_n(c)$ is represented by the cocycle $$\sigma\mapsto -{(\sigma-1)P_c\over l^n}+\sigma{P_c\over l^n}-{P_c\over l^n}, \ \ {\rm Gal}(K^{\rm sep}/K)\to E(K^{\rm sep})_{l^n}$$ where ${(\sigma-1)P_c\over l^n}$ is the unique $l^n$-division point of $(\sigma-1)P_c$ in $E(K[c])$ and $P_c/l^n$ is a fixed $l^n$-division point of $P_c$.*]{}
\(ii) [*The cohomology class $\delta_n(c)$ is represented by the cocycle, where ${(\sigma-1)P_c\over l^n}$ is the unique $l^n$-division point of $(\sigma-1)P_c$ in $E(K[c])$, $$\sigma\mapsto -{(\sigma-1)P_c\over l^n},
\ \ {\rm Gal}(K^{\rm sep}/K)\to E(K^{\rm sep}) .$$* ]{}
\(iii) [*Let $n\geq m$ be positive integers and let $c\in \Lambda(n)$. Then we have $$l^{m} \delta_n(c)=\delta_{n-m}(c)$$ $$l^{m}\gamma_n(c)=\gamma_{n-m}(c).$$*]{}
0.2in [*Proof of Lemma 4.2.1.*]{} (i) and (ii) These two formulae for cocycles representing the cohomology classes $\gamma_n(c)$ and $\delta_n(c)$ can be extracted from Step 2 of the proof of \[1, Chapter 7, Lemma 7.14.14\]. We reprove these formulae here.
From the diagram (3.4.12) the restriction homomorphism, where ${\cal G}_c=$ Gal$(K[c]/K)$, $$H^1(K,E_{l^n}(K^{\rm sep}))\to H^1(K[c], E_{l^n}(K^{\rm
sep}))^{{\cal G}_c}$$ is an isomorphism as the prime number $l$ belongs to $\cal P$. The point ${ P}_c$ belongs to $ E(K[c])$.
Let ${{ P}_c\over l^n}\in E(K^{\rm sep})$ be a fixed $l^n$th division point of ${ P}_c$, that is to say ${{ P}_c\over l^n}$ is any point which satisfies $l^n({{ P}_c\over l^n})={ P}_c$. Then the cocycle $$\phi:g\mapsto g\big({{ P}_c\over l^n}\big)-{{ P}_c\over l^n}, \ {\rm Gal}(K^{\rm sep}/K[c])\to
E_{l^n}(K^{\rm sep}),$$ represents a cohomology class in $H^1(K[c], E_{l^n}(K^{\rm sep}))^{{\cal
G}_c}$ which is the image of $P_c\ ({\rm mod}
\ l^n)\in ({}_{l^n}E(K[c]))^{{\cal G}_c}$ under the coboundary map $$\partial_{l^n} :
({}_{l^n}E(K[c]))^{{\cal G}_c}\to H^1(K[c], E_{l^n}(K^{\rm sep}))^{{\cal G}_c}$$ (see the diagram (3.4.12)). The inflation of $\phi$ to ${\rm Gal}(K^{\rm sep}/K)$ is given by the cocycle $$\phi^\sharp :{\rm Gal}(K^{\rm sep}/K)\to
E(K^{\rm sep}), \ \ g\mapsto g\big({{ P}_c\over l^n}\big)-{{ P}_c\over l^n}$$ which need not necessarily be annihilated by $l^n$.
For any element $g\in {\rm Gal}(K^{\rm sep}/K)$, denote by $${(g-1){
P}_c\over l^n}$$ the unique $l^n$th root of $(g-1){ P}_c$ in $E(K[c])$. This root exists because $P_c$ belongs to $({}_{l^n}E(K[c])^{{\cal G}_c}$; furthermore, it is unique because $E(K[c])_{l^\infty}=0$ (by Definition 3.1.3(f) and Proposition 1.10.1, \[1, Lemma 7.14.11(i)\].
The cochain $$\psi:{\rm Gal}(K^{\rm sep}/K)\to E(K[c]), \ \ g\mapsto -{(g-1){ P}_c\over
l^n},$$ is a cocycle whose restriction to the subgroup ${\rm Gal}(K^{\rm sep}/K[c])$ is the zero cochain. But $\psi$ need not be annihilated by $l^n$. The cochain $$\phi^\sharp+\psi:{\rm Gal}(K^{\rm sep}/K)\to E(K^{\rm sep}),
g\mapsto g\big({{ P}_c\over
l^n}\big)-{{ P}_c\over l^n} -{(g-1){ P}_c\over
l^n},$$ is a cocycle which is annihilated by $l^n$ and whose restriction to ${\rm Gal}(K^{\rm sep}/K[c])$ is the cocycle $$\phi:{\rm Gal}(K^{\rm sep}/K[c])\to E_{l^n}(K^{\rm
sep}).$$ Hence the cochain $\phi^\sharp+\psi$ is a cocycle $$\phi^\sharp+\psi:{\rm Gal}(K^{\rm sep}/K)\to E_{l^n}(K^{\rm sep})$$ and this cochain represents the cohomology class $\gamma_n(c)$ in $H^1(K,E_{l^n})$. Therefore the cohomology class $\delta_n(c)$ of $H^1(K[c]/K,E)_{l^n}$ is represented by the cocycle $$\psi:{\rm Gal}(K[c]/K)\to E(K^{\rm sep}), \ \ g\mapsto -{(g-1){ P}_c\over
l^n}$$ as required.
\(iii) This follows immediately from the explicit cocycle formulae of parts (i) and (ii) of this lemma. ${\sqcap \!\!\!\!\sqcup}$
0.2in [**4.2.2. Lemma.**]{} 2 by 1 [*Denote by a superscript $\pm1$ the eigenspaces under the action of the non-trivial element of ${\rm Gal}(K/F)$. If $c\in \Lambda^r(n)$ then we have $$P_c \ ({\rm mod\ } l^n)\ \in \Bigg(({}_{l^n}E(K[c]))^{{\cal G}_c}\Bigg)^{-\nu(c)}$$ $$\gamma_n(c)\in H^1(K, E_{l^n})^{-\nu(c)}$$ $$\delta_n(c)\in H^1(K[c]/K, E)_{l^n}^{-\nu(c)}$$ where ${\cal G}_c=$ Gal$(K[c]/K)$, $$\nu(c)=(-1)^{r}\epsilon,$$ $r$ is the number of distinct prime divisors in the support of $c$, and $\epsilon$ is the sign in the functional equation of the $L$-function of $E/F$.*]{}
\[For the proof, see \[1, Lemma 7.14.11\]. Note that the set of prime numbers $\cal P$ is contained in the set of prime numbers denoted ${\cal P}\setminus {\cal F}$ of \[1, Lemma 7.14.11\].\] ${\sqcap \!\!\!\!\sqcup}$
0.4in
[**4.3. Points $P_c$ defined over local fields**]{}
0.2in The notation and hypotheses of (4.1.1) hold in this section. The Drinfeld-Heegner points $(a,I_1,c)$, $(a,I_1, c,\pi)$ are defined in (3.4.4) and (3.4.5).
0.2in[**4.3.1. Proposition.**]{} [*Suppose that $c\in \Lambda(1)$ and $z\in \Lambda^1(1)$ is a prime divisor disjoint from the support of the divisor $c$. Let $y$ be the unique prime of $K$ lying over $z$ and $K_{y}$ be the completion of $K$ at $y$. Then the point $(a,I_1,c)
\in X_0^{\rm Drin}( K[c])$ is definable over $K_{y}$ that is to say $$(a,I_1,c)\in X_0^{\rm Drin}( K_{y}).$$ Furthermore, we have that $$(a,I_1,c,\pi)\in E(K_{y})$$ is a point of the elliptic curve $E$ definable over $K_{y}$.* ]{}
0.2in [*Proof of Proposition 4.3.1.*]{} The prime $z$ is inert and unramified in $K/F$ by Definition 3.2.4 and remark 3.2.5(i). Furthermore the elliptic curve $E/F$ has good reduction at $z$ by remark 3.2.5(i). As $I$ is the conductor of $E/F$ without the component at $\infty$, by (4.1.1), we have that $z$ is disjoint from the support of $I$ and hence the curve $X_0^{\rm Drin}(I)/F
$ also has good reduction $z$ where there may be several disoint components in the closed fibre over $z$.
By \[1, Theorem 4.6.19(ii)\] as $z$ is inert and unramified in $K/F$ the reduction $(a,I_1,c)$ mod $z$ is defined over the quadratic extension field $\kappa(y)$ of $\kappa(z)$. That is to say $(a,I_1,c)$ mod $z$ is a point of the reduction at $z$ of $X_0^{\rm Drin}(I)$ which is defined over $\kappa(y)$.
Let $F_z$ be the completion of $F$ at $z$. As $X_0^{\rm Drin}(I)$ has good reduction at $z$, it follows that the point $(a,I_1,c)$ is defined over the field $K_y$ as this is the unique quadratic extension field of the local field $F_z$ which is unramified over $z$. It then immediately follows that $(a,I_1,c,\pi)$, which is the image of $(a,I_1,c)$ under the morphism $\pi : X_0^{\rm Drin}\to E$ of $F$-schemes (see (3.4.4) and (3.4.5)), is a point of the elliptic curve $E$ defined over $K_{y}$. ${\sqcap \!\!\!\!\sqcup}$
0.2in [**4.3.2. Proposition.**]{} [*If $z\in \Lambda^1(1)$ is a prime divisor disjoint from the support of the divisor $c\in \Lambda(1)$ and $y$ is the prime of $K$ lying over $z$, then the image of $P_c$ in ${}_{l^n}E(K_y)$ via Proposition 4.3.1 is uniquely determined by $P_c$.*]{}
0.2in [*Proof of Proposition 4.3.2.*]{} We have the isomorphism $$K[c]\otimes K_y\cong \prod_i K[c]_{x_i}$$ where $x_i$ are the places of $K[c]$ over the place $y$ of $K$ and $K[c]_{x_i}$ is the completion of $K[c]$ at $x_i$. The galois group ${\cal G}_c=$ Gal$(K[c]/K)$ permutes transitively the places $x_i$ and the completions $K[c]_{x_i}$. We then have $${}_{l^n}E(K[c]\otimes_KK_y)\cong \prod_i {}_{l^n}E(K[c]_{x_i}).$$
The point $P_c$ belongs to $E(K[c])$ by construction. Hence the point $P_c$ (mod $l^n$) of ${}_{l^n}E(K[c])$ induces an element $(q_1,q_2,\ldots)$ of $
{}_{l^n}E(K[c]\otimes_KK_y)$ where $$(q_1,q_2,\ldots)\in\prod_i {}_{l^n}E(K[c]_{x_i})$$ and $$q_i\in {}_{l^n}E(K[c]_{x_i})\ \ {\rm \ for \ all \ }i.$$
By Proposition 4.3.1, $P_c$ is definable over $K_y$ that is to say $P_c\in E(K_y)$ and so we have $q_i\in {}_{l^n}E(K_y)$ for all $i$. But $P_c$ (mod $l^n$) $ \in \Bigg( {}_{l^n}E(K[c])\Bigg)^{{\cal G}_c}$ by (3.4.13) (see also Lemma 4.2.2). It follows that $(q_1,q_2,\ldots)$ is invariant under $
{\cal G}_c$. But the elements $q_i\in {}_{l^n}E(K_y)$ are permuted transitively by ${\cal G}_c$. Hence the elements $q_i$ are all equal and $P_c$ (mod $l^n$) is the point $(q_1,q_1,\ldots)\in \prod_i{}_{l^n}E([c]_{x_i})$ which is in the image of the diagonal map ${}_{l^n}E(K_y)\to \prod_i {}_{l^n}E(K[c]_{x_i})$. Hence the components of the point $P_c$ (mod $l^n$) in ${}_{l^n} E(K[c]\otimes_KK_y)$ are independent of the place $x_i$ and depend only on $P_c$ as required. ${\sqcap \!\!\!\!\sqcup}$
0.4in
[**4.4. The map $\chi_z$**]{} 2=1
0.2in The notation and hypotheses of (4.1.1) hold in this section.
0.2in[**4.4.2. Proposition.**]{} 2 by 1 [*Let $z\in \Lambda^1(n)$ be a prime divisor and let ${\cal E}_0/\kappa(z)$ be the closed fibre of the Néron model of $E/F$ at the place $z$. Let $a_z\in {\mathds Z}$ be the trace of the Frobenius at $z$ on the inertia invariant part of the Tate module of $E$ as in (3.2.6). Let $y$ be the unique place of $K$ over $z$. Let $c\in \Lambda(n)$ be a divisor whose support contains $z$ and put $c'=c-z\in \Lambda(n)$ so that $c'$ has support coprime to $z$.*]{}
\(i) [*The endomorphism $\vert G(c/c')\vert {\rm Frob}(z) -a_z
$ of the elliptic curve ${\cal E}_0/\kappa(z)$ annihilates the abelian group ${\cal E}_0(\kappa(y))$.*]{}
\(ii) [*The group homomorphism $$h: {}_{l^n}{\cal E}_0(\kappa(y))\ \to
\ {\cal E}_0(\kappa(y))_{l^n}, \ \
x\mapsto \bigg( {
\vert G(c/c')\vert {\rm Frob}(z) -a_z
\over l^n}\bigg)x$$ is an isomorphism which commutes with $\tau$.*]{}
0.2in [*Proof of Proposition 4.4.1.*]{} (i) As in §1.4, $G(c/c')$ denotes the galois group Gal$(K[c]/K[c'])$. The Frobenius ${\rm Frob}(z)$ acts on the Tate module of the elliptic curve ${\cal E}_0/\kappa(z)$ with characteristic polynomial $$X^2-Xa_z+\vert \kappa(z)\vert .$$ Writing $F$ for ${\rm Frob}({z})$ it follows that $F^2-Fa_z+\vert\kappa(z)\vert$ annihilates the abelian group ${\cal E}_0(\kappa(y))$. But $F^2$ is the identity automorphism on ${\cal E}_0(\kappa(y))$ as $\kappa(y)/\kappa(z)$ is a quadratic extension of finite fields. Hence $F^2-Fa_z+F^2\vert\kappa(z)\vert$ annihilates ${\cal E}_0(\kappa(y))$. That is to say $F(F(1+\vert\kappa(z)\vert)-a_z)$ annihilates ${\cal E}_0(\kappa(y))$. As $F$ is an automorphism of ${\cal E}_0({\overline {\kappa(y)} })$ we obtain that $F(1+\vert\kappa(z)\vert)-a_z$ annihilates ${\cal E}_0(\kappa(y))$.
As $K\not= F\otimes_{{\mathds F}_q}{\mathds F}_{q^2}$ (by the hypothesis (c) of paragraph (4.1.1)) we have that the unit group $B^*/A^*$ is the trivial group. The Galois group $$G(c/c')={\rm Gal}(K[c]/K[c'])$$ therefore has order (see \[1, Chap. 2, (2.3.8) and (2.3.12)\]) $$\begin{aligned}
\vert G(c/c')\vert =\vert \kappa(z)\vert+1.\\\end{aligned}$$ Hence the endomorphism $\vert G(c/c')\vert {\rm Frob}({z})-a_z$ of ${\cal
E}_0$ annihilates the abelian group ${\cal
E}_0(\kappa(y))$ as required.
\(ii) Let $\alpha,\beta\in \mathds C$ be the complex roots of the characteristic polynomial of Frobenius ${\rm Frob}(z)$ acting on the Tate module of ${\cal E}_0/\kappa(z)$ $$X^2-Xa_z+\vert \kappa(z)\vert .$$ Then we have by the trace formula for the Frobenius automorphism, where $G(c/c')$ is cyclic of order $\vert \kappa(z)\vert+1$ as in the proof of part (i), $$\vert {\cal E}_0(\kappa(z))\vert= \vert G(c/c')\vert-\alpha-\beta=
\vert G(c/c')\vert-a_z$$ and $$\vert {\cal E}_0(\kappa(y))\vert= \vert\kappa(z)\vert^2+1-\alpha^2-\beta^2=(\vert G(c/c')\vert-a_z)(\vert G(c/c')\vert+a_z).$$ We then obtain the decomposition into eigenspaces under the action of the involution $\tau$, the non-trivial element of Gal$(K/F)$, $${\cal E}_0(\kappa(y))\cong {\cal E}_0(\kappa(z))\oplus
{\cal E}_0(\kappa(y))^{-}$$ where $$\vert {\cal E}_0(\kappa(y))^{\delta}\vert= \vert
G(c/c')\vert-\delta a_z{\rm \ \ for \ }\delta=\pm1.$$ Hence the order of the $l^\infty$-torsion is given by $$\vert {\cal E}_0(\kappa(y))^{\delta}_{l^\infty}\vert=
l^{s(\delta)}$$ where $l^{s(\delta)}$ is the highest power of $l$ dividing $\delta \vert G(c/c')\vert-a_z$.
By Lemma 3.2.7(ii), we have group isomorphisms for the $l^n$-torsion $${\cal E}_0(\kappa(y))^{\delta}_{l^n}\cong{{\mathds Z}\over l^n{\mathds Z}}
{\rm \ \ for \ } \delta=\pm1$$ that is to say these groups are cyclic of order $l^n$. It follows from this that there are isomorphisms for the $l^\infty$-torsion $${\cal E}_0(\kappa(y))^{\delta}_{l^\infty}\cong{{\mathds Z}\over l^{s(\delta)}{\mathds Z}}
{\rm \ \ for \ } \delta=\pm1.$$ We obtain that the $l^\infty$-torsion subgroups of ${\cal E}_0(\kappa(y))^{-}_{l^\infty}$ and ${\cal
E}_0(\kappa(y))^{+}_{l^\infty}$ are both cyclic and denoting by $\vert.\vert_l$ the normalised $l$-adic absolute value on ${\mathds Q}$ we have $$\vert {\cal
E}_0(\kappa(y))^{\delta}_{l^\infty}\vert = \vert
\ \ \delta \vert G(c/c')\vert-a_z\vert_l^{-1}{\rm \ \ for \ } \delta=\pm1.$$
Let $g$ be the homomorphism $$g= {\vert G(c/c')\vert {\rm Frob}({z})-a_z\over
l^n}: \ {\cal E}_0\to {\cal E}_0$$ where the integers $\vert G(c/c')\vert$, $a_z$ are both divisible by $l^n$ by Lemma 3.2.7(i). The homomorphism induced by $g$ on the $\kappa(y)$-rational points of ${\cal E}_0$ $$g:{\cal E}_0(\kappa(y))\to {\cal E}_0(\kappa(y))$$ is annihilated by $l^n$ by part (i); furthermore, the subgroup $l^n{\cal E}_0(\kappa(y))$ of ${\cal E}_0(\kappa(y))$ belongs to the kernel of $g$ again by part (i). Hence $g$ induces a homomorphism $$h: {}_{l^n}{\cal E}_0(\kappa(y))\to {\cal
E}_0(\kappa(y))_{l^n}.$$
On each eigencomponent $({}_{l^n}{\cal E}_0(\kappa(y)))^\delta$ under the action of $\tau$, the non-trivial element of Gal$(K/F)$ and where $\delta=\pm1$, the map $g$ is multiplication by the integer $$N(\delta)={\delta \vert G(c/c')\vert- a_z\over l^n}
={\delta \vert {\cal E}_0(\kappa(y))^\delta\vert\over l^n}.$$ It follows from this formula for $N(\delta)$ that the restriction of $h$ to each eigencomponent $({}_{l^n}{\cal E}_0(\kappa(y)))^\delta$ is an injection. As $\tau
$ commutes with $h$, the homomorphism $h$ preserves the $\tau$-eigencomponents and therefore $h$ is an injection. As ${}_{l^n}{\cal E}_0(\kappa(y))$ and $ {\cal
E}_0(\kappa(y))_{l^n}$ have the same number $l^{2n}$ of elements it follows that $h$ is an isomorphism. ${\sqcap \!\!\!\!\sqcup}$
0.2in
0.2in [**[**4.4.2. Proposition**]{}.**]{} 2 by 1 [*Let $z$ be a prime divisor of $F$ which belongs to $\Lambda^1(n)$. Let $y\in
\Sigma_K$ be the unique place of $K$ lying over $z$. Assume that the prime number $l\in {\cal P}$ is coprime to the order of ${\rm Pic}(A)$, the Picard group of the ring $A$. Then there is a homomorphism $$\chi_z:{}_{l^n}E(K_{y})\to H^1(K_{y}, E_{l^n})$$ with the following properties.*]{}
\(1) [*Let $x$ be a place of the ring class field $K[z]$ above the place $z$ of $K$. Then the image of $\chi_z$ is contained in the subgroup $H^1(K[z]_x/K_y,E_{l^n}(K[z]_x))$.*]{}
\(2) [*The homomorphism $\chi_z$ is injective.*]{}
\(3) [*The composition of $\chi_z$ with the homomorphism, obtained from the inclusion of group schemes $E_{l^n}\subset E$, $$H^1(K_{y}, E_{l^n})\to H^1(K_{y},E)_{l^n}$$is an isomorphism $${}_{l^n}E(K_{y})\cong H^1(K_{y}, E)_{l^n}.$$*]{}
\(4) [*For all divisors $c\in \Lambda(n)$ such that $z$ belongs to ${\rm Supp}(c)$, the cohomology class $\gamma_n(c) $ satisfies*]{} $$\gamma_n(c)_{y}=\chi_z(P_{c-z}\ ({\rm mod\ } l^n))$$ [*where $P_{c-z}$ ([mod $l^n$]{}) is an element of ${}_{l^n}E(K_y)$ (Proposition 4.3.2) and $c-z$ has support coprime to $z$.*]{}
0.2in [*Proof of Proposition 4.4.2.*]{} For all divisors $d\in \Lambda(n)$ with support coprime to $z$, by Propositions 4.3.1 and 4.3.2 the point $P_{d}\in E(K[c])$ induces an element of ${}_{l^n}E(K_y)$ which is uniquely determined by $P_{d}$ where $y$ is the place of $K$ over $z$ (remark 3.2.5(1)).
Select a divisor $c\in \Lambda(n)$ with support containing $z$ and put $$c'=c-z\in \Lambda(n).$$ where Supp$(c')$ is coprime to $z$.
Let $z'$ be a prime divisor of $K[c']$ over the place $y$, which is the place of $K$ over $z$. The prime divisor $z^\prime$ is totally ramified in the field extension $K[c]/K[c']$; this follows via class field theory from the definition of the ring class field $K[c]$ (see §1.4; more details are given in \[1, Chapter 2, (2.3.13)\]). Let $z^\times$ be the unique prime divisor of $K[c]$ lying over $z^\prime$.
The cohomology class $\delta_n(c)$ belongs to $H^1(K[c]/K, E(K[c]))_{l^n}$ which is contained via the inflation map in $H^1(K, E(K^{\rm sep}))_{l^n}$ (diagram (3.4.12)).
Let ${\cal E}_0/\kappa(z)$ be the closed fibre above $z$ of the Néron model of $E/F$. We then define a composite isomorphism $\Phi$ as follows where the maps $i,h,j$ are explained below $$\Phi:{}_{l^n}E(K_y)\ {\buildrel i\over \longrightarrow}\ {}_{l^n}{\cal E}_0(\kappa(y))
\ {\buildrel
h\over \longrightarrow} \ {\cal E}_0(\kappa(y))_{l^n}\ {\buildrel
j\over \longrightarrow} \ { E}(K_y)_{l^n}.
\leqno{(4.4.\the\count2)}$$ 2 by 1 Here $i: {}_{l^n}E(K_y) \longrightarrow{}_{l^n}{\cal E}_0(\kappa(y))$ is the isomorphism obtained from the surjective homomorphism $E(K_y) \longrightarrow{\cal E}_0(\kappa(y))$ of reduction at $z$ whose kernel is a pro-$p$-group. The map $h:{}_{l^n}{\cal E}_0(\kappa(y))
{\longrightarrow} {\cal E}_0(\kappa(y))_{l^n}$ is the isomorphism of Proposition 4.4.1(ii). The map $j: {\cal E}_0(\kappa(y))_{l^n} { \longrightarrow} { E}(K_y)_{l^n}$ is the isomorphism obtained from reduction modulo $y$; the map $j$ is an isomorphism because $E$ has good reduction at $y$ and the prime number $l$ is distinct from the characteristic of $F$. The map $\Phi$ is an isomorphism as $i,h ,j$ are isomorphisms.
Let $x'$ be a prime of $K[0]$ lying over $y$ and let $x$ be the unique prime of $K[z]$ over $x'$, where $y$ is the prime of $ K$ over $z$. The prime $x$ over $x'$ is uniquely determined by $x'$ because $K[z]/K[0]$ is totally ramified at $x'$ (see §1.4). The restriction of elements of the Galois group $G(c/c')$ to the fields $K[z]$ and $K[z]_x$ induces isomorphisms $${\rm Gal}(K[z]_x/K[0]_{x'})\cong G(z/0)\cong G(c/c-z)\leqno{(4.4.\the\count2)}$$ 2 by 1 where these groups are cyclic of order $\vert \kappa(z)\vert+1$ (by \[1, Chapter 2, (2.3.12), p.20\], and as $B^*/A^*$ is the trivial group by hypothesis (c) of (4.1.1)). Put $$G={\rm Gal}(K[z]_x/K[0]_{x'}).$$ We have the exact sequence of galois groups $$0\to G\to {\rm Gal}(K[z]_x/K_{y})
\to {\rm Gal}(K[0]_{x'}/K_{y})\to 0.$$ As the field extension $K[z]/K[0]$ is totally ramified at the place $x'$ of $K[0]$ over $z$, the restriction homomorphism of $G(c/0)$ to $K[0]_{x^\prime}$ gives the isomorphism $$G\cong G(z/0)$$ of (4.4.4).
As the field extension $K[z]_x/K[0]_{x'}$ is totally ramified, we have that the group $G$ acts trivially on $E_{l^n}(K[z]_x)$ and that $$E_{l^n}(K[z]_x)=E_{l^n}(K[0]_{x'}).$$ Hence we have an isomorphism $$H^1(G, E_{l^n}(E(K[z]_x))\cong {\rm Hom}(G, E_{l^n}(K[0]_{x'})).\leqno{(4.4.\the\count2)}$$ 2 by 1
We have already fixed in (3.4.7) a generator $\sigma_z$ of $G(c/c')$. Let $\sigma\in G$ be the generator induced by $\sigma_z$ under the isomorphism $G(c/c')\cong G$ of (4.4.4).
For any $P\in {}_{l^n} E(K_y)$, define a homomorphism $$f_P: G\to E_{l^n}(K_y)\leqno{(4.4.\the\count2)}$$ 2 by 1 as follows. Put, where $\sigma$ is the chosen generator of the cyclic group $G$, $$f_P(\sigma)= \Phi(P).$$ As $\Phi(P)$ is a point of $E_{l^n}(K_y)$ and as the order of the cyclic group $G$ is equal to $\vert\kappa(z)\vert+1$ which is divisible by $l^n$, the homomorphism $f_P$ is well defined. Hence $f_P$ defines a cohomology class in $H^1(G,E_{l^n}(K_y))$ where $G$ acts trivially on $E_{l^n}(K_y)$.
As $\Phi: {}_{l^n}E(K_y)\to E(K_y)_{l^n}$ is an isomorphism (see (4.4.3)) and $G$ is cyclic of order divisible by $l^n$, this map $P\mapsto f_P$ defines a group isomorphism $$f: {}_{l^n}E(K_y)\ {\buildrel \cong \over \longrightarrow }\ H^1(G,E_{l^n}(K_y)), \ \ P\mapsto f_P.
\leqno{(4.4.\the\count2)}$$ 2 by 1
We have the Hochschild-Serre spectral sequence $$E^{i,j}_2
\Rightarrow H^{i+j}(K[c]_{z^\times}/K_y, E_{l^n}(K[c]_{z^\times}))$$ where we write $$G_{z^\times/z'}={\rm Gal}(K[c]_{z^\times}/K[c']_{z'}).$$ and $$E^{i,j}_2=H^{i}(K[c']_{z'}/K_y, H^j(G_{z^\times/z'}, E_{l^n}(K[c]_{z^\times}))).$$ The short exact sequence of low degree terms attached to this spectral sequence in part takes the form $$0\to E^{1,0}_2\to H^{1}(K[c]_{z^\times}/K_y, E_{l^n}(K[c]_{z^\times}))\to E^{0,1}_2\to
E^{2,0}_2.$$
As the order of Pic$(A)$ is coprime to $l$ we have that the order of the place $y$ above $z$ in Pic$(B)$ is coprime to $l$, as $y$ is the unique place of $K$ over $z$. It follows that the degree of the field extension $K[c']_{z'}/K_y$, which is equal to the degree of the residue field extensions, is coprime to $l$. Hence we have $$H^{i}(K[c']_{z'}/K_y, H^j(G_{z^\times/z'}, E_{l^n}(K[c]_{z^\times})))\cong 0
{\rm \ \ for \ all \ } i\geq 1$$ as the group $H^j(G_{z^\times/z'}, E_{l^n}(K[c]_{z^\times}))$ is $l$-power torsion for all $i$. The above short exact sequence of low degree terms then becomes an isomorphism $$H^{1}(K[c]_{z^\times}/K_y, E_{l^n}(K[c]_{z^\times}))\cong
H^{0}(K[c']_{z'}/K_y, H^1(G_{z^\times/z'}, E_{l^n}(K[c]_{z^\times}))).
\leqno{(4.4.\the\count2)}$$ 2 by 1 As the field extension $K[c]_{z^\times}/K[c']_{z'}$ is totally ramified, we have $ E_{l^n}(K[c]_{z^\times})= E_{l^n}(K[c']_{z'})$ and the group $G_{z^\times/z'}$ acts trivially on $ E_{l^n}(K[c]_{z^\times})$. This last isomorphism of (4.4.8) then becomes the isomorphism $$H^{1}(K[c]_{z^\times}/K_y, E_{l^n}(K[c]_{z^\times}))\cong
H^{0}(K[c']_{z'}/K_y,{\rm Hom}(G_{z^\times/z'}, E_{l^n}(K[c']_{z'}) ))$$ where $G_{z^\times/z'}$ acts trivially on $E_{l^n}(K[c']_{z'}) $. This then provides the isomorphism $$H^{0}(K[c']_{z'}/K_y, H^1(G_{z^\times/z'}, E_{l^n}(K[c]_{z^\times})))
\cong {\rm Hom}(G_{z^\times/z'}, E_{l^n}(K_y)).$$ This isomorphism combined with the isomorphism of (4.4.8) gives the isomorphism, where $G_{z^\times/z'}$ acts trivially on $ E_{l^n}(K_y)$, $$H^{1}(K[c]_{z^\times}/K_y, E_{l^n}(K[c]_{z^\times}))\cong H^1(G_{z^\times/z'}, E_{l^n}(K_y)).
\leqno{(4.4.\the\count2)}$$ 2 by 1
Take now $c=z$ in the isomorphism of (4.4.9) where $x'$ is a place of $K[0]$ over $y$ and $x$ is the unique place of $K[z]$ over $x'$; we have that the composition of this isomorphism (4.4.9) with the isomorphism $f$ of (4.4.7) gives the isomorphism $$\phi:{}_{l^n}E(K_y)\ {\buildrel \cong \over \longrightarrow} \ H^1(K[z]_x/K_y, E_{l^n}(K[z]_x)). \leqno{(4.4.\the\count2)}$$ 2 by 1
The inflation map from Gal$(K[z]_x/K_y)$ to Gal$(K_y^{\rm sep}/K_y)$ is an injective homomorphism $$H^1(K[z]_x/K_y, E_{l^n}(K[z]_x))\
{\buildrel {\rm inf}\over \longrightarrow}\ H^1(K_y,E_{l^n}(K^{\rm sep}_y)).
\leqno{(4.4.\the\count2)}$$ 2 by 1 The composite of this injective inflation map of (4.4.11) with the isomorphism $\phi$ of (4.4.10) is then defined to be the injective homomorphism $\chi_z$ $$\matrix{\chi_z:&{}_{l^n}E(K_{y})&\to &H^1(K_{y}, E_{l^n})\cr
&P&\mapsto& \{\sigma\mapsto \Phi(P)\}.\cr}
\leqno{(4.4.\the\count2)}$$ 2 by 1 Here $\{\sigma\mapsto \Phi(P)\}$ denotes the cohomology class $f_P$ of (4.4.6) and (4.4.7) defined by $\sigma\mapsto \Phi(P)$ where $\sigma\in G$ is the chosen generator of $G$; this $f_P$ then defines a cohomology class in $H^{1}(K[z]_x/K_y, E_{l^n}(K[z]_x))$ by the isomorphism of (4.4.9) for $c=z$. By inflation, this $f_P$ gives a cohomology class of Gal$(K_y^{\rm sep}/K_y)$ and hence an element of the cohomology group $H^1(K_{y}, E_{l^n})$ and this defines the homomorphism $\chi_z$. This map $\chi_z$ of (4.4.12) is injective by construction.
To prove property (1), by construction the homomorphism $$\chi_z:{}_{l^n}E(K_{y})\longrightarrow H^1(K_{y}, E_{l^n})$$ takes a point $P\in {}_{l^n}E(K_{y})$ to an element $f_P$ of $H^1(G, E_{l^n}(K[z]_x))$ which is inflated to an element of $H^1(K_{y}, E_{l^n}(K^{\rm sep}))$ and this proves the property (1) of the map $\chi_z$.
To prove property (2), the map $$\chi_z: {}_{l^n} E(K_{y})\to
H^1(K_{y}, E_{l^n})^{}, P\mapsto \{\sigma\mapsto \Phi(P)\}$$ is injective by construction, as already noted.
For property (3), we have by definition that the map $\chi_z$ of (4.4.12) is the composition of the isomorphism $\phi$ of (4.4.10) with the inflation map of (4.4.11). That is to say, the map $\chi_z$ factors as $${}_{l^n}E(K_y)\ {\buildrel \phi \over \longrightarrow } \ H^1(K[z]_x/K_y, E_{l^n}(K[z]_x)) \
{\buildrel {\rm inf}\over \longrightarrow}\ H^1(K_y,E_{l^n}(K^{\rm sep}_y))
\leqno{(4.4.\the\count2)}$$ 2 by 1 where $\phi$ is an isomorphism and inf is an injection.
We have an exact sequence obtained from the inclusion of group schemes $E_{l^n}\subset E$, $$0\ \longrightarrow\ {}_{l^n}E(K_y)\ {\buildrel \partial_{l^n}\over
\longrightarrow}\ H^1(K_{y}, E_{l^n})\ {\buildrel \psi\over \longrightarrow} \ H^1(K_{y},E)_{l^n}.$$ The morphism of multiplication by $l^n$ on the Néron model of $E$ over the ring of valuation integers of $F$ at the place $z$ is étale; therefore the image of $\partial_{l^n}$ consists of unramified cohomology classes and more precisely the image of $\partial_{l^n}$ belongs to the subgroup $H^1(K_y^{\rm nr}/K_y, E_{l^n})$ of $H^1(K_y, E_{l^n})$ where $K_y^{nr}$ is the maximal separable unramified extension of $K_y$. The intersection of $H^1(K_y^{\rm nr}/K_y, E_{l^n})$ with $
H^1(K[z]_x/K_y, E_{l^n}(K[z]_x))$, which are both subgroups of $H^1(K_y, E_{l^n})$, is therefore contained in $H^1(K[0]_{x'}/K_y,E_{l^n}(K[0]_{x'}))$; this follows by considering representative cocycles and because the field extension $K[z]_x/K[0]_{x'}$ is totally ramified and $K[0]_{x'}/K_y$ is unramified. But the order of the Picard group Pic$(A)$ is coprime to $l$ by hypothesis; therefore the order of the place $y$ over $z$ in Pic$(B)$ is coprime to $l$ as $y$ is the unique place of $K$ over $z$. It follows that the degree of the field extension $K[0]_{x'}/K_y$ is coprime to $l$ and therefore we have $$H^1(K[0]_{x'}/K_y,E_{l^n}(K[0]_{x'}))=0.$$ It follows that the composition of the injective inflation map $$H^1(K[z]_x/K_y, E_{l^n}(K[z]_x)) \
{\buildrel {\rm inf}\over \longrightarrow}\ H^1(K_y,E_{l^n}(K^{\rm sep}_y))$$ with $\psi: H^1(K_{y}, E_{l^n})\to H^1(K_{y},E)_{l^n}$ is an injection $$H^1(K[z]_x/K_y, E_{l^n}(K[z]_x)) \
{\buildrel \over \longrightarrow}\ H^1(K_{y},E)_{l^n}.
\leqno{(4.4.\the\count2)}$$ 2 by 1
By arithmetic flat local duality, there is an isomorphism of discrete groups $$H^1(K_y,E)\cong {\rm Hom}( E(K_y),{\mathds Q}/{\mathds Z})
\leqno{(4.4.\the\count2)}$$ 2 by 1 for the proof of which, using flat cohomology and local class field theory, see \[9, Chapter 3, Theorem 7.8\]; we only require the prime-to-$p$ version of this duality (4.4.15) and whose simpler proof is explained in \[9, Chapter 1, Remark 3.6\]. From (4.4.15) we obtain the isomorphism of discrete groups $$H^1(K_y,E)_{l^n}\cong {\rm Hom}( {}_{l^n}E(K_y),{\mathds Z}/l^n{\mathds Z}).
\leqno{(4.4.\the\count2)}$$ 2 by 1 It follows from (4.4.16) that $H^1(K_y,E)_{l^n}$ and ${}_{l^n}E(K_y)$ have the same number of elements. As $E_{l^n}(K_y)$ is isomorphic to ${}_{l^n}E(K_y)$ by the isomorphism $\Phi$ of (4.4.3), we have that $H^1(K_y,E)_{l^n}$ and $E_{l^n}(K_y)$ have the same number of elements.
From (4.4.9) for $c=z$, we have the isomorphism, where $G$ is cyclic of order divisible by $l^n$ and which acts trivially on $E_{l^n}(K_y)$, $$H^{1}(K[z]_x/K_y, E_{l^n}(K[z]_x))\cong H^1(G, E_{l^n}(K_y))
\cong E_{l^n}(K_y).$$
It now follows that the finite groups $H^{1}(K[z]_x/K_y, E_{l^n}(K[z]_x))$, $E_{l^n}(K_y)$ and $H^1(K_y,E)_{l^n}$ have the same number of elements and hence the injective homomorphism of (4.4.14) $$H^1(K[z]_x/K_y, E_{l^n}(K[z]_x))\longrightarrow H^1(K_{y}, E)_{l^n}$$ is an isomorphism. It follows that the natural homomorphism, obtained from the inclusion of group schemes $E_{l^n}\subset E$, $$H^1(K_{y}, E_{l^n})\ {\buildrel \psi\over \longrightarrow}\ H^1(K_{y},E)_{l^n}$$ composed with $\chi_z$, where $\chi_z$ factors through the group $H^{1}(K[z]_x/K_y, E_{l^n}(K[z]_x))$ as in (4.4.13), $${}_{l^n}E(K_y)\ {\buildrel \chi_z\over \longrightarrow} \
H^1(K_y,E_{l^n})\ {\buildrel \psi\over \longrightarrow }\
H^1(K_y,E)_{l^n}$$ is an isomorphism $${}_{l^n}E(K_{y})\cong H^1(K_{y}, E)_{l^n}.$$ This proves the property (3).
It only remains to prove property (4). For all divisors $d\in \Lambda(n)$, as in (3.4.10) we shall write $P_d$ (mod $l^n$) for the image of $P_d\in E(K[d])$ in ${}_{l^n}E(K[d])$.
From Lemma 4.2.1, we have that the cohomology class $\gamma_n(c)$ in $H^1(K, E_{l^n})$ is represented by the cocycle, where ${(\sigma-1)P_c\over l^n}$ is the unique $l^n$-division point of $(\sigma-1)P_c$ in $E(K[c])$ and $P_c/l^n$ is a fixed $l^n$-division point of $P_c$, $$\Gamma_n(c):\sigma\mapsto -{(\sigma-1)P_c\over l^n}+\sigma{P_c\over l^n}-{P_c\over l^n}, \ \ {\rm Gal}(K^{\rm sep}/K)\to E(K^{\rm sep})_{l^n}
\leqno{(4.4.\the\count2)}$$ 2 by 1 and the cohomology class $\delta_n(c)$ in $H^1(K, E)_{l^n}$ is represented by the cocycle $$\Delta_n(c): \sigma\mapsto -{(\sigma-1)P_c\over l^n}, \ \
{\rm Gal}(K^{\rm sep}/K)\to E(K^{\rm sep}). \leqno{(4.4.\the\count2)}$$ 2 by 1
The cohomology class $\delta_n(c)_{y}\in H^1(K_{y},
E(K_{y}^{\rm sep}))$ is the restriction at $y$ of the class $\delta_n(c)\in H^1(K,E(K^{\rm sep}))$. Let $Q\in E(K[c])$ be the element given by $$Q=-{(\sigma_z-1){ P}_c\over l^n} \leqno{(4.4.\the\count2)}$$ 2 by 1 and where $\sigma_z$ is the fixed generator of the cyclic group $G(c/c')$.
The definition of $P_c$ (as in \[1, Chap. 7, §7.14.10 pp.386-387\] and (3.4.9) above) is the following. Let $c=\sum_{i=1}^r z_i$ be the decomposition of $c$ as a sum of distinct prime divisors, where we write $z=z_1$. Then we have $$P_c=\sum_{s\in {\cal S}} sD_cy_c$$ where $$y_c=(0,I_1,c,\pi)\in E(K[c])\leqno{(4.4.\the\count2)}$$ 2 by 1 as in (3.4.7) and where $\cal S$ is a set of coset representatives for $G(c/0)$ in Gal$(K[c]/K)$ and $D_c$ is the Kolyvagin element of (3.4.9).
We have an exact sequence of abelian groups $$0\to G(c/c')\to G(c/0)\to G(c'/0)\to 0.$$ We obtain a corresponding decomposition in the group algebra ${\mathds Z}[G(c/0)]$ of the Kolyvagin element $D_c$ $$D_c=D_{c-z}D_z$$
We have already selected in (3.4.7) a generator $\sigma_{z}$ of the cyclic group $G(c/c-z)$. We may then define a map of sets $$h_{z}:G(c/c-z)\to \mathds Z$$ by $$\sigma_{z}^{-s}\mapsto -s,{\rm \ \ for \ } s=0,1,\ldots,\vert
G(c/c-z)\vert-1.$$ The Kolyvagin element of $h_{z}$ is then as in (3.4.8) $$D_{z}=-\sum_{r=1}^{\vert G(c/c-z)\vert-1} r\sigma_{z}^r.$$ We have $$(\sigma_{z}-1)D_{z}=-\vert G(c/c-z)\vert+ e_{G(c/c-z)}\leqno{(4.4.
\the\count2)}$$ 2 by 1 where $e_{G(c/c-z)}$ is the element of the group algebra ${\mathds Z}[G(c/c-z)]$ given by $$e_{G(c/c-z)}=\sum_{g\in G(c/c-z)}g.$$
The element ${ P}_c\in E(K[c])$ may then be written as, where we write $y_{c}=(0,I_1,c,\pi)$ as in (3.4.7) and (4.4.20), $${ P}_c=\sum_{s\in {\cal S}} s D_{c-z}D_z y_{c}.\leqno{(4.4.
\the\count2)}$$ 2 by 1
We have (see [\[1, Chap. 4, (4.8.3) and table 4.8.5]{}\]) $${\vert O_{c'}^*\vert\over \vert
A^*\vert}{\rm Tr}_{K[c]/K[c']}y_{c}=a_zy_{c'}$$ where $a_z\in {\mathds Z}$ is as in Proposition 4.4.1 and (3.2.6) and where $y_{c'}=(0,I_1,c',\pi)$. By definition $Q=-((\sigma_z-1){ P}_c)/l^n$ (see (4.4.19)); hence we have from (4.4.21) $$Q=-\sum_{s\in{\cal S}} s D_{c-z} \big(
{(\sigma_z-1)D_z\over l^n}\big)y_{c}$$ $$= \sum_{s\in{\cal S}} s D_{c-z} \big(
{\vert G(c/c')\vert \over l^n}y_{c}-{ \vert
A^*\vert\over \vert O_{c'}^*\vert}{a_z\over
l^n}y_{c'}\big).$$ As $K$ is not obtained from $F$ by ground field extension (hypothesis (c) of paragraph (4.1.1)), we have $
A^*= O_{c'}^*$; hence we obtain $$Q= \sum_{s\in{\cal S}} s D_{c-z} \big(
{\vert G(c/c')\vert \over l^n}y_{c}-{a_z\over
l^n}y_{c'}\big).\leqno{(4.4.\the\count2)}$$ 2 by 1
As at the beginning of this proof, let $y$ be the unique place of $K$ over $z$, $z'$ be a prime of $K[c']$ over the place $y$ of $K$, and let $z^\times$ be the place of $K[c]$ over the prime $z'$ of $K[c']$ where the field extension $K[c]/K[c']$ is totally ramified at $z'$. Also ${\cal E}_0/\kappa(z)$ denotes the closed fibre above $z$ of the Néron model of $E/F$.
We write ${ Q}_0$ for the image of $Q$ modulo $z^\times$ in ${\cal
E}_0(\kappa(z^\times))$ by passage to the residue field $\kappa(z^\times)$. From the isomorphism (4.4.9), where $\delta_n(c)_y$ belongs to $H^{1}(K[c]_{z^\times}/K_y, E_{l^n}(K[c]_{z^\times}))$, we have that the reduction of $\delta_n(c)_y$ at $y$ belongs to ${\rm Hom}(G(c/c'), {\cal
E}_0(\kappa(y)))_{l^n}$ and is given by the cocycle (see (4.4.18)) $$g\mapsto -{(g-1){ P}_c\over l^n}\ \ ({\rm mod} \ z^\times),
\ \ {\rm Gal}(K[c]_{z^\times}/K[c']_{z'})\cong G(c/c')\to {\cal
E}_0(\kappa(y))$$ and we have that $-{(g-1){ P}_c\over l^n}$ modulo $z^\times$ belongs to the $l^n$-torsion subgroup ${\cal
E}_0(\kappa(y))_{l^n}$ rational over $\kappa(y)$. Hence the point $ Q_0$, the reduction of $Q$ modulo $z^\times$, belongs to ${\cal
E}_0(\kappa(y))_{l^n}$. Note that $-{(g-1){ P}_c}$ modulo $z^\times$ reduces to zero, for all $g\in G(c/c')$, as $K[c]/K[c']$ is totally ramified at $z'$.
Denote by Frob$(z)$ the Frobenius automorphism $x\mapsto x^{\vert\kappa(z)\vert}$ of the closed fibre ${\cal E}_0/\kappa(z)$ over $z$ of the Néron model of $E/F$. The theorem [\[1, Chap. 4, Theorem 4.8.9\] ]{} gives that for the prime $z^{\prime}$ of $K[c']$ above $z$ we have, where $y_{c}=(0,I_1,c,\pi)$ as in (4.4.20), $${\rm
Frob}(z)y_{c}\equiv y_{c'}\ \ ({\rm mod} \ \
z^{\prime}).\leqno{(4.4.\the\count2)}$$ 2 by 1
We obtain from Proposition 4.3.1 that $y_{c}$ mod $z^\times$ is defined over the subfield $\kappa(y)$ of $\kappa(z^\times)$ where $\kappa(y)$ is the quadratic extension field of the finite field $\kappa(z)$. Hence we have from (4.4.24) $${\vert
G(c/c')\vert \over l^n}y_{c}-{a_z\over l^n}y_{c'}\equiv{\vert G(c/c')\vert {\rm
Frob}({z})-a_z\over l^n}y_{c'}\ ({\rm mod}\ z^{\prime}).
\leqno{(4.4.\the\count2)}$$ 2 by 1
The point ${ P}_{c'}\in E(K[c'])$ is given by (see (3.4.9)) $${ P}_{c'}=\sum_{s\in {\cal S}}s
D_{c-z} y_{c'}.$$ We then have from (4.4.23) and (4.4.25) $$Q=\sum_{s\in{\cal S}} s D_{c-z} \big(
{\vert G(c/c')\vert \over l^n}y_{c}-{a_z\over
l^n}y_{c'}\big)$$ $$\equiv \ {Q_0}\ \equiv{\vert G(c/c')\vert {\rm Frob}({z})-a_z\over l^n}{
P}_{c'}\ \ ({\rm mod}\ z^{\prime}). \leqno{(4.4.\the\count2)}$$2 by 1
From Lemma 4.2.2 or \[1, Lemma 7.14.11(ii)\]), we have that ${ P}_{c'}$ (mod $l^n$) belongs to the $-\nu({c'})$-eigenspace for $\tau$ on ${}_{l^n}{
E}(K[c'])$, where $\nu(c')=(-1)^r\epsilon$ and $r$ is the number of prime divisors in the support of $c'$ and $\epsilon$ is the sign in the functional equation of the $L$-function of $E/F$ (as in Lemma 4.2.2). As $z$ is inert in $K/F$, it follows that the image of the reduction $P^\flat_{c'} $ of ${ P}_{c'}$ modulo $z^{\prime}$ belongs to the $-\nu({c'})$-eigenspace for $\tau$ on ${}_{l^n}{\cal
E}_0(\kappa(z^{\prime}))$. Let $$h: {}_{l^n}{\cal E}_0(\kappa(y))\ \to
\ {\cal E}_0(\kappa(y))_{l^n}, \ \
x\mapsto \bigg( {
\vert G(c/c')\vert {\rm Frob}(z) -a_z
\over l^n}\bigg)x$$ be the isomorphism which commutes with $\tau$ of Proposition 4.4.1. As ${ Q}_0$ belongs to the subgroup ${\cal
E}_0(\kappa(y))$ as already noted and also as $P^\flat_{c'}$ (mod $l^n$) is an element of ${}_{l^n} {\cal E}_0(\kappa(y))$ by Proposition 4.3.2, we have by (4.4.26) $${
Q}_0=h(P^\flat_{c'} \ ({\rm mod}\ l^n)).$$ With the notation of (4.4.3) we then have $${
Q}_0=h\circ i(P_{c'} \ ({\rm mod}\ l^n))$$ where $i$ is the reduction isomorphism ${}_{l^n}E(K_y)\to {}_{l^n}{\cal E}_0(\kappa(y))$ and where $P_{c'}$ (mod $l^n$) belongs to ${}_{l^n} E(K_y)$ by Proposition 4.3.2. Furthermore, by (4.4.3) and as $\Phi=j\circ h\circ i$ we have $$\Phi(P_{c'} \ ({\rm mod}\ l^n)) =j(Q_0) \leqno{(4.4.\the\count2)}$$2 by 1 is the unique $l^n$-torsion point of $E(K_y)$ whose reduction at $y$ is $Q_0$.
Let ${\overline {z} }$ be any prime of ${K^{\rm sep}}$ above ${z^\times} $. Restrict the cocycle $$\Gamma_n(c):
{\rm Gal}(K^{\rm sep}/K)\to E(K^{\rm sep})_{l^n}$$ of (4.4.17), which represents $\gamma_n(c)$, to the decomposition group of ${\overline {z} }$. As $K[c]/K[c']$ is totally ramified at $z^\prime$ we have that the Kolyvagin element $D_z$ restricted to the the residue field of $z^{\times}$ satisfies $$D_z=-\vert\kappa(z)\vert(\vert\kappa(z)\vert+1)/2.$$ Furthermore, $l^n$ divides $\vert\kappa(z)\vert+1$ and $l$ is different from $2$; hence we have that (from (4.4.22)) the reduction $P_c^\flat$ of $P_c$ at $z^{\times}$ satisfies $P_c^\flat \in l^n{\cal E}_0(\kappa(z^{\times}))$ and hence as $E$ has good reduction at $z^{\times}$ we have that $$P_c\in l^nE(K[c]_{z^{\times}}).$$
We then have that the cocycle $\Gamma_n(c) $ of (4.4.17) satisfies $$\Gamma_n(c)(\rho)=0 {\rm \ \ for \ all \ } \rho\in
{\rm Gal} (K[c]_{z^{\times}}^{\rm sep}/K[c]_{z^{\times}})$$ and this cocycle, restricted to the decomposition group of ${\overline z}$, factors through the subgroup Gal$(K[c]_{z^{\times}}/K_{y} )$. Furthermore, since $\sigma_z$, which is the selected generator of $G(c/c')$, is in the inertia group of ${y} $ we have that $$\sigma_z{P_c\over l^n}-{P_c\over l^n}$$ reduces to zero modulo ${z^{\times}}$. Hence $$\Gamma_n(c)(\sigma_z)= -{(\sigma_z-1)P_c\over l^n}+\sigma_z{P_c\over l^n}-{P_c\over l^n}$$ is the unique $l^n$-torsion point congruent to $Q$ (mod ${z^{\times}}$) where $Q$ is given by (4.4.19) that is to say $$\Gamma_n(c)(\sigma_z)\equiv -{(\sigma_z-1)P_c\over l^n}\ \
({\rm mod}\ \ {z^{\times}}). \leqno{(4.4.\the\count2)}$$2 by 1 But from (4.4.27), we then obtain $$\Gamma_n(c)(\sigma_z)= \Phi(P_{c'} \ ({\rm mod}\ l^n)).
\leqno{(4.4.\the\count2)}$$2 by 1 Then (4.4.29) shows, as $\Gamma_n(c)(\sigma_z)$ is the unique $l^n$-torsion point with reduction at $y$ coinciding with the reduction of $Q=-{(\sigma_z-1)P_c\over l^n}$ at $y$, that we have an equality of cohomology classes in $H^1(K_y,E_{l^n})$ $$\gamma_n(c)_y=\chi_z(P_{c-z}\ ({\rm mod} \ l^n))$$ where $P_{c-z}$ (mod $l^n$) is an element of ${}_{l^n}E(K_y)$. This proves property (4) and completes the proof of Proposition 4.4.2. ${\sqcap \!\!\!\!\sqcup}$
0.2in
0.2in [*4.4.2. Remarks.*]{} (1) 2 by 1 This Proposition 4.4.2 and its consequence stated in Proposition 4.5.1(iv) below are an extension of \[1, Chapter 7, Lemma 7.14.14(ii)\]. It would be of interest to eliminate the hypothesis that the prime number $l$ be coprime to the order of the Picard group Pic$(A)$ from these Propositions 4.4.2 and 4.5.1(iv).
\(2) The homomorphism $\chi_z$ interchanges the $\tau$-eigenspaces in that for $\delta=\pm1$ we have $$\chi_z({}_{l^n} E(K_{y}))^\delta)\subseteq
H^1(K_{y}, E_{l^n})^{-\delta}.$$ This property, which is not required for this paper, follows from the group Gal$(K[c]/F)$ being generalised dihedral.
0.4in
[**4.5. Localizations of the classes $\gamma_n(c)$ and $\delta_n(c)$**]{} 0.2in The notation and hypotheses of (4.1.1) hold in this section. Let $\coprod\!\!\!\coprod(E/K)$ be the Tate-Shafarevich group of the elliptic curve $E\times_FK$ over $K$. 0.2in [**4.5.1. Proposition.**]{} [ *Let $c\in \Lambda(n)$, let $z$ be a prime divisor in $\Lambda^1(n)$ and let $y$ be the place of $K$ over the place $z$ of $F$.*]{}
\(i) [*Let $v$ be a place of $K$ coprime to $c$. Then we have $\gamma_n(c)_v\in \partial_n(E(K_v))$ that is to say we have $\delta_n(c)_v=0$.*]{}
\(ii) [*If $c$ is coprime to $z$ then we have $\gamma_n(c)_{y}=\partial_n((P_{c} \ ({\rm mod } \ l^n))_y).$*]{}
\(iii) [*If $l^n\vert P_c$ then we have $\delta_n(c)=0$. If $l^n\vert
P_{c-w}$ for all prime divisors $w$ in the support of $c$ then we have $\delta_n(c)\in \coprod\!\!\!\coprod(E/K)_{l^\infty}$.*]{}
\(iv) [*Assume that the prime number $l\in {\cal P}$ is coprime to the order of ${\rm Pic}(A)$, the Picard group of the ring $A$. If $z\in {\rm Supp}(c)$ then we have $${\rm ord}\ \delta_n(c)_{y}={\rm ord}\ \gamma_n(c)_{y}={\rm ord}\ \gamma_n(c-z)_{y}=$$ $${\rm ord}\bigg( (P_{c-z} \ ({\rm mod} \ l^n))_y \bigg)$$ where $(P_{c-z}\ ({\rm mod}\ l^n))_y$ is an element of ${}_{l^n}E(K_y)$.*]{}
0.2in [*Proof.*]{} (i) The field $K[c]$ is a subfield of $K_\infty$ as $\infty$ is split completely in $K[c]/K$ (see \[1, Chapter 2, (2.3.13)\]). We have that the localization $\delta_n(c)_\infty$ at $\infty$ of $\delta_n(c)$ satisfies $\delta_n(c)_\infty\in H^1(K_\infty, E)_{l^n}$ and the localization $ ( P_c\ ({\rm mod}\
l^n))_\infty$ at $\infty$ of $ P_c\ ({\rm mod}\
l^n)$ satisfies $( P_c\ ({\rm mod}\
l^n))_\infty\in
{}_{l^n}E(K_\infty)$. It then follows from the diagram (3.4.12) that $\delta_n(c)_\infty=0$.
Suppose now that $v$ is a place of $K$ such that $v\not=\infty$ and $v\not\in $ Supp$(c)$. We have by construction that $$\delta_n(c)\in H^1(K[c]/K, E(K[c]))_{l^n}.$$ The field extension $K[c]/K$ is unramified at $v$ (see \[1, Chapter 2, (2.3.13)\]). Hence we have that the localization $\delta_n(c)_v$ at $v$ satisfies $$\delta_n(c)_v\in H^1(K_v^{nr}/K_v,E(K_v^{nr}))_{l^n}\subseteq
H^1(K_v, E(K_v^{\rm sep}))_{l^n}$$ where $K_v^{nr}$ is the maximal unramified separable extension of the local field $K_v$. But $H^1(K_v^{nr}/K_v,E(K_v^{nr}))_{l^n}=0$ by Definition 3.1.3(d) of the set of prime numbers $\cal P$ to which $l$ belongs. Hence we have $\delta_n(c)_v=0$; this last vanishing is equivalent to $\gamma_n(c)_v\in \partial_n(E(K_v))$, as required.
\(ii) By Lemma 4.2.1, the cohomology class $\gamma_n(c)$ is represented by the cocycle $$\sigma\mapsto -{(\sigma-1)P_c\over l^n}+\sigma{P_c\over l^n}-{P_c\over l^n}, \ \ {\rm Gal}(K^{\rm sep}/K)\to E(K^{\rm sep})_{l^n}$$ where ${(\sigma-1)P_c\over l^n}$ is the unique $l^n$-division point of $(\sigma-1)P_c$ in $E(K[c])$. Furthermore by the same Lemma 4.2.1, the cohomology class $\delta_n(c)\in H^1(K, E(K^{\rm sep}))_{l^n}$ is represented by the cocycle $$\sigma\mapsto -{(\sigma-1)P_c\over l^n}, \ \ {\rm Gal}(K^{\rm sep}/K)\to E(K^{\rm sep}).$$ By part (i) as $z\not\in $ Supp$(c)$, we have $\delta_n(c)_y=0$ hence a cocycle representing $\delta_n(c)$ when localized at $v$ is cohomologous to zero. Therefore there is an element $a\in E(K_y^{\rm sep})$ such that $$-{(\sigma-1)P_c\over l^n} = \sigma a-a, \ \ {\rm for \ all \ } \sigma\in \ {\rm Gal}(K^{\rm sep}_y/K_y).$$ Hence we have $$-(\sigma-1)(P_c+l^n a) =0, \ \ {\rm for \ all \ } \sigma\in \ {\rm Gal}(K^{\rm sep}_y/K_y).$$ This implies that $$P_c+l^n a\in E(K_y).$$
By Lemma 4.3.1, the localization at $y$ of the point $P_c$ lies in $ E(K_y)$ where $y$ is the place of $K$ over $z$. Furthermore, by Lemma 4.3.2 the point $(P_c{\rm \ (mod \ }l^n))_y $ belongs to ${}_{l^n} E(K_y)$ and its image in this group is uniquely determined by $P_c$.
It follows from this and the above cocycle formulae for $\gamma_n(c)$ and $\delta_n(c)$ that the cohomology class $\gamma_n(c)_y$ is represented by the cocyle $$\sigma\mapsto \sigma{P_c\over l^n}-{P_c\over l^n}, \ \ {\rm Gal}(K_y^{\rm sep}/K_y)\to E(K_y^{\rm sep})_{l^n}.$$ That is to say we have $\gamma_n(c)_{y}=\partial_n((P_{c} \ ({\rm mod } \ l^n))_y)$.
\(iii) From Lemma 4.2.1(ii), the cohomology class $\delta_n(c)$ is represented by the cocycle $$\sigma\mapsto -{(\sigma-1)P_c\over l^n},
\ \ {\rm Gal}(K^{\rm sep}/K)\to E(K^{\rm sep}) .$$ We then evidently have that if $l^n\vert P_c$, that is to say $P_c\in l^nE(K[c])$, then $\delta_n(c)=0$.
Suppose that $l^n\vert P_{c-z}$ for all prime divisors $z$ in the support of $c$. Then for any prime divisor $y$ of $K$ lying over the prime divisor $z$ dividing $c$ we have $\delta_n(c)_{y}=0$ by Proposition 4.4.2(4). It then follows from part (i) of the present Proposition 4.5.1 that $\delta_n(c)_{v}=0$ for all places $v$ of $K$ and hence that $\delta_n(c)$ belongs to the Tate-Shafarevich group $\coprod\!\!\!\coprod(E/K)_{l^\infty}$.
\(iv) From the property (4) of the map $\chi_z$ of Proposition 4.4.2, we have $$\gamma_n(c)_y=\chi_z((P_{c-z}\ ({\rm mod} \ l^n))_y)$$ where $(P_{c-z}\ ({\rm mod} \ l^n))_y$ is an element of ${}_{l^n}E(K_y)$ by Proposition 4.3.2. As the homomorphism $\chi_z$ is injective (property (2) of Proposition 4.4.2), we obtain $${\rm ord}(\gamma_n(c)_y)={\rm ord}((P_{c-z}\ ({\rm mod}\ l^n))_y)$$ where again $(P_{c-z} \ ({\rm mod}\ l^n))_y$ denotes an element of ${}_{l^n}E(K_y)$.
From part (ii) of the present Proposition 4.5.1, we have $$\gamma_n(c-z)_{y}=\partial_n((P_{c-z} \ ({\rm mod } \ l^n))_y)$$ where $z$ is coprime to the support of $c-z$. Hence we obtain $${\rm ord}\big( (P_{c-z} \ ({\rm mod } \ l^n))_y\big)
={\rm ord}(\gamma_n(c-z)_{y}).$$
From property (3) of Proposition 4.4.2, we have that the composition $\chi_z$ with the homomorphism, obtained from the inclusion of group schemes $E_{l^n}\subset E$, $$\theta :H^1(K_{y}, E_{l^n})\to H^1(K_{y},E)_{l^n}$$is an isomorphism $${}_{l^n}E(K_{y})\cong H^1(K_{y}, E)_{l^n}.$$ Hence the image in $ H^1(K_{y}, E)_{l^n}$ under $\theta$ of $\chi_z(P_{c-z} \ ({\rm mod}\ l^n))$ is the cohomology class $\delta_n(c)_y$ because this is the image under $\theta$ of $\gamma_n(c)_y\in H^1(K_y,E_{l^n})$ and furthermore, because $\theta\circ\chi_z$ is an isomorphism, $\delta_n(c)_y$ has the same order as ${\rm ord}\big( (P_{c-z} \ ({\rm mod } \ l^n))_y\big)$, which completes the proof of the proposition. ${\sqcap \!\!\!\!\sqcup}$
1.8in
[**4.6. The Cassels pairing with a class $\delta_n(c)$**]{}
0.2in
The notation and hypotheses of (4.1.1) hold in this section. The torsion abelian group ${\mathds Z}/n{\mathds Z}$ for $n\not=0$ is considered to be a subgroup of ${\mathds Q}/{\mathds Z}$ via the map $1\mapsto 1/n$. Let
0.2in $\coprod\!\!\!\coprod(E/K)$ be the Tate-Shafarevich group of the elliptic curve $E\times_FK$ over $K$;
$< -, ->_{\rm Cassels}$ be the Cassels pairing on $\coprod\!\!\!\coprod(E/K)$ (see §2.3);
$[ , ]_{w}: H^1(K_w,E)_n\times {}_nE(K_w)\to {\mathds Z}/n{\mathds Z}$ be the local Tate pairing for any place
$w\in \Sigma_K$ of $K$ and any integer $n$ coprime to the characteristic of $F$.
0.2in
[**4.6.1. Proposition.**]{} 2 by 1 [*Let $m$ and $n$ be integers $\geq 1$ and let $c\in \Lambda(m+n)$. Suppose that $\delta_m(c)$ belongs to the Tate-Shafarevich group $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}$. Suppose that the element $s\in\coprod\!\!\!\coprod$$(E/K)_{l^\infty}$ has order at most $l^{n}$. Lift the element $s\in\coprod\!\!\!\coprod$$(E/K)_{l^\infty}$ to an element $S\in H^1(K,E_{l^{n}})$ and select points $x(w)\in E(K_{w})$ such that $S_{w}=\partial_{l^{n}}(x(w))$ for all $w\in\Sigma_K$. Then we have $$< \delta_m(c), s>_{\rm Cassels}
= \sum_{y\in\Sigma_K {\rm \ divides\ } {\rm Supp}( c)}
[ \delta_{m+n}(c)_{y}, x(y)]_{y}
\leqno{(4.6.2)}$$ where the sum runs over the places of $K$ which divide an element of [Supp]{}$(c)$.*]{} 2 by 1
0.2in [*Proof.*]{} The construction of the Cassel’s pairing is given in §2.3. As in the statement of the proposition, we may lift the element $s\in\coprod\!\!\!\coprod$$(E/K)_{l^\infty}$ of order $l^{n}$ to an element $S\in H^1(K,E_{l^{n}})$. The points $x(w)\in E(K_{w})$ are then selected such that $S_{w}=\partial_{l^{n}}(x(w))$ for all $w\in\Sigma_K$.
By Lemma 4.2.1(iii) we have $$l^{n} \delta_{m+n}(c)=\delta_m(c).$$ From the formula (2.3.5) for the Cassels pairing, we then obtain $$< \delta_m(c), s>_{\rm Cassels}
= \sum_{w\in \Sigma_K}
[ \delta_{m+n}(c)_{w}, x(w)]_{w}.
\leqno{(4.6.3)}$$2 by 1 where the sum runs over all places $w$ of $K$.
By Proposition 4.5.1(i), we have $\delta_{m+n}(c)_w=0$ for all places $w\in \Sigma_K $ which do not divide an element of Supp$(c)$. Hence there is no contribution to the Cassels pairing in the sum (4.6.3) when $w$ does not divide an element of Supp$(c)$. Hence we obtain the formula (4.6.2). ${\sqcap \!\!\!\!\sqcup}$
0.2in [**4.6.4. Proposition.**]{} [*Let $m$ and $n$ be integers $\geq 1$ and let $c\in \Lambda(m+n)$, $d\in \Lambda(n)$. Suppose that $\delta_m(c)$ and $\delta_{n}(d)$ belong to the Tate-Shafarevich group $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}$. Assume that the prime number $l\in {\cal P}$ is coprime to the order of ${\rm Pic}(A)$. Then we have $$< \delta_m(c), \delta_{n}(d)>_{\rm Cassels}
= \sum_{y\in \Sigma_K {\rm \ divides \ }{\rm Supp}( c)
\setminus {\rm Supp}(d)}
[\delta_{m+n}(c)_{y}, (P_{d} \ ({\rm mod}\ l^n))_{{y}}]_{y}\leqno{(4.6.5)}$$ where the sum runs over the places of $K$ which divide an element of ${\rm Supp}(c)\setminus {\rm Supp}(d)$.*]{}
0.2in [*Proof.*]{} The element $\gamma_{n}(d) \in H^1(K,E_{l^{n}})$ is a lifting of $\delta_{n}(d)\in$ $\coprod\!\!\!\coprod$$(E/K)$ which has order at most $l^{n}$. Suppose that $z\in$ Supp$( c)$ and $y$ is the unique place of $K$ lying over the place $z$ of $F$. If $z$ also satisfies $z\in $ Supp$( d)$ then by Proposition 4.5.1(iv), we have $$\gamma_{n}(d)_{y}=0$$ because $\delta_{n}(d)_{y}=0$ as we have $\delta_{n}(d)\in $ $\coprod\!\!\!\coprod$$(E/K)$. If on the other hand $z$ satisfies $z\not\in $ Supp$( d)$ then by Lemma 4.5.1(ii), we have $$\gamma_{n}(d)_{y}=\partial_{l^{n}}((P_{d}\ ({\rm mod}\ l^n))_{y}).$$ where $(P_{d}\ ({\rm mod}\ l^n))_{y}$ denotes the localization at $y$ of $P_d$ (mod $l^n$). The formula (4.6.5) to be proved now follows from the formula (4.6.2) of Proposition 4.6.1. ${\sqcap \!\!\!\!\sqcup}$
[**Chapter 5. Construction of cohomology classes and proofs of the main theorems**]{}
0.4in[**5.1. $M_r$ is finite for some $r$** ]{} 0.2in
0.2in(5.1.1) Throughout this chapter 5, the notation of (4.1.1) of the previous chapter 4 remains valid and the elliptic curve $E/F$ and quadratic field extension $K/F$ satisfy the hypotheses (a), (b), (c) of (4.1.1).
We further let in this chapter
0.2in
$N^+,N^-$ be the eigenspaces under the action of the involution $\tau$ whenever $N$
is a ${\mathds Z}[{\rm Gal}(K/F)]$-module on which multiplication by $2$ is an isomorphism;
$\epsilon=\pm1$ is the sign of the functional equation of the $L$-function $L(E/{F},s)$
of $E/F$;
$\nu(r)=(-1)^r\epsilon$ for any natural number $r\in \mathds N$;
$\nu(c)=(-1)^r\epsilon$ for any divisor $c$ of $F$ with exactly $r$ distinct prime divisors in
its support;
$P_c\in E(K[c])$ be the Drinfeld-Heegner points of paragraph (3.4.9) for
all $c\in \Lambda(1)$, where $P_0\in E(K)$;
$M_r$, for all $r\in{\mathds N}$, be the quantities in ${\mathds N}\cup\infty$ given in Definition 4.1.6;
$\coprod\!\!\!\coprod$$(E/K)$ be the Tate-Shafarevich group of the elliptic curve $E\times_FK$ over
$K$ (see §2.2);
Sel$_n(E/K)$ be the $n$-Selmer group of $E\times_FK$ over $K$ for any integer $n$ coprime
to the characteristic of $F$ (see §2.2));
$
{\displaystyle {\rm Sel}_{a^\infty}(E/K)=\lim_{\longrightarrow\atop n} {\rm Sel}_{a^n}(E/K)}$ for any number $a$ coprime to the characteristic
of $F$ (see also (3.1.6)). 0.2in
In this section §5.1 some consequences are presented of the hypothesis that $M_r$ be finite for some $r$.
0.2in [**5.1.2. Lemma.**]{} [*The Drinfeld-Heegner $P_0$ has infinite order in $E(K)$ if and only if $M_0$ is finite. If $P_0$ has infinite order then we have $$l^{M_0}=\vert (E(K)/{\mathds Z}P_0)_{l^\infty} \vert$$ $${\rm ord}\ \gamma_{M_0+m}(0)= l^m, {\sl \ \ for \ all\ } m\geq 0,$$ $$\gamma_{M_0+m}(0)\in {\rm Sel}_{l^\infty}(E/K)^{-\epsilon}
{\sl \ \ for \ all\ } m\geq 0.$$*]{}
0.2in [*Proof of Lemma 5.1.2.*]{} The definition of $M_0$ is that (Definition 4.1.6) $$M_0={\rm ord}_l( P_0)=\max\{ m\vert\ P_0\in l^m E(K[0])\}.$$ The group $E(K[0])$ has no $l$-torsion (as $l\in {\cal P}$, see Proposition 1.10.1 and Definition 3.1.3(f)) and is a finitely generated group by the Mordell-Weil theorem. Hence $P_0$ has infinite order in $E(K)$ if and only if $M_0$ is finite.
We have $\gamma_{M_0+m}(0)\in {\rm Sel}_{l^\infty}(E/K)^{-\epsilon}
$, for all $ m\geq 1,$ by Lemma 4.2.2 and because $\gamma_{M_0+m}(0)=\partial_{l^{M_0+m}}(P_0)$.
Assume now that $P_0$ has infinite order. By \[1, Chap. 7, Theorem 7.6.5\] the point $P_0$ generates a subgroup of $E(K)$ of finite index (note that in the notation of this theorem \[1, Chap. 7, Theorem 7.6.5\], we have $P_0=x_0$).
By definition we have $$\vert (E(K)/{\mathds Z}P_0)_{l^\infty} \vert ={\max\{ l^m\vert \ P_0\in l^mE(K)\}}.$$ where the group $E(K[0])$ has no $l^\infty$-torsion (as $l\in {\cal P}$, see Proposition 1.10.1 and Definition 3.1.3(f)).
We have the Hochschild-Serre spectral sequence $$H^i({\rm Gal}(K[0]/K),H^j(K[0], E_{l^n}))\Rightarrow H^{i+j}(
K, E_{l^n}).$$ The short exact sequence of low degree terms of this spectral sequence is in part $$0\to H^1({\rm Gal}(K[0]/K), E_{l^n}(K[0]))
\to H^1(K, E_{l^n})\to H^1(K[0], E_{l^n})^{{\rm Gal}
(K[0]/K)}$$ $$\to H^2({\rm Gal}(K[0]/K), E_{l^n}(K[0])).$$ The two extreme terms $H^1({\rm Gal}(K[0]/K), E_{l^n}(K[0]))$ and $H^2({\rm Gal}(K[0]/K), E_{l^n}(K[0]))$ are both zero as $E_{l^n}(K[0])$ is zero as already noted. Hence this short exact sequence provides the isomorphism $$H^1(K, E_{l^n})\cong H^1(K[0], E_{l^n})^{{\rm Gal}
(K[0]/K)}\leqno({5.1.3)}$$ which is induced from the injection $E(K)\to E(K[0])$. The short exact sequence of sheaves for the étale topology on Spec $K$ $$0\longrightarrow E_{l^n}\longrightarrow E{\buildrel l^n\over \longrightarrow} E\longrightarrow 0$$ then provides the commutative diagram of cohomology groups $$\matrix{
0&\longrightarrow &E_{l^n}(K)&\longrightarrow & E(K) &{\buildrel l^n\over \longrightarrow} &E(K)&\longrightarrow & H^1(K, E_{l^n})&\ldots\cr
&&\downarrow&&\downarrow&&\downarrow&&\downarrow&&\cr
0&\longrightarrow& E_{l^n}(K[0])&\longrightarrow & E(K[0]) &{\buildrel l^n\over \longrightarrow}& E(K[0])& \longrightarrow & H^1(K[0], E_{l^n})&\ldots}$$ The isomorphism $H^1(K, E_{l^n})\cong H^1(K[0], E_{l^n})^{{\rm Gal}
(K[0]/K)}$ of (5.1.3) together with this commutative diagram then shows that the homomorphism $$E(K)/l^nE(K)\to E(K[0])/l^nE(K[0])$$ is injective. Hence these two numbers $l^{M_0}$ and $\vert (E(K)/{\mathds Z}P_0)_{l^\infty} \vert$ are equal.
We obtain that the order of $\gamma_{M_0+m}(0)$ is equal to $l^m$ from Proposition 3.4.14(i). ${\sqcap \!\!\!\!\sqcup}$
0.2in
[**5.1.4. Lemma.**]{} [*Assume that the prime number $l\in {\cal P}$ is coprime to the order of ${\rm Pic}(A)$. Suppose that $M_r$ is finite for some integer $r\geq 0$. Then $M_s$ is finite for all $s\geq r$ and $$M_r,M_{r+1},M_{r+2}\ldots$$ is a decreasing sequence of non-negative integers.*]{}
0.2in [*Proof of Lemma 5.1.4.*]{} Suppose that $M_s$ is finite for some $s\geq0$. Then there is a divisor $$c\in \Lambda^s(M_s+1)$$ which satisfies $l^{M_s}\vert\vert P_c$ and $M_s<\alpha(c)$; hence the cohomology class $\gamma_{M_s+1}(c)$ is non-zero (by Proposition 3.4.14(i)). From Proposition 3.3.8, where we take $n=M_s+1$, we obtain a prime divisor $$z\in \Lambda^1(M_s+1)\leqno{(5.1.5)}$$ coprime to $c$ such that the localization $\gamma_{M_s+1}(c)_y$ is non-zero where $y$ is the unique prime of $K$ lying over $z$ (see remark 3.2.5(1)). Then by Proposition 4.5.1(ii), we have $P_c\not\in l^{M_s+1}E(K_{y})$; that is to say $${\rm ord}((P_c\ ({\rm mod}\ l^{M_s+1}))_y)> 1.\leqno{(5.1.6)}$$ It follows from Proposition 4.5.1(iv) that ord$\ \gamma_n (c+z)_y\ > 1$, where $n=M_s+1$, and in particular we have $\gamma_n (c+z)\not=0$. Therefore by Proposition 3.4.14(i)), or the definition of $\gamma_n(c+z)$, we have $P_{c+z}\not\in l^{M_s+1} E(K[c+z])$. It follows that we have $$c+z\in \Lambda^{s+1}(M_s+1)$$ and $M_{s+1}$ is finite and we have $M_{s+1}\leq M_s$. As $M_r$ is finite by hypothesis, we then have that $M_s$ is finite and $M_s\geq M_{s+1}$ for all $s\geq r$ by induction. ${\sqcap \!\!\!\!\sqcup}$
0.2in [**5.1.7. Lemma.**]{} [*Suppose that $M_{r-1}$ is finite where $r\geq 1$. Assume that the prime number $l\in {\cal P}$ is coprime to the order of ${\rm Pic}(A)$. Let $c\in \Lambda^{r}(M_{r-1})$ and put $\nu(c)=(-1)^r\epsilon$. Then we have* ]{}
\(i) $\delta_{M_{r-1}}(c)\in$ $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}^{-\nu(c)}$
\(ii) [*$\gamma_{M_{r-1}}(c)\in{\rm Sel}_{l^{\infty}}(E/K)^{-\nu(c)}$;* ]{}
\(iii) [*the order of $\delta_{M_{r-1}}(c)$ is at most $l^{M_{r-1}-M_{r}}$.*]{}
0.2in [*Proof of Lemma 5.1.7.*]{} By the previous Lemma 5.1.4, we have that $M_r$ is finite and $M_{r-1}\geq M_r$. The set $\Lambda^{r}(M_{r-1})$ is non-empty by corollary 3.3.9; therefore there is an element $c\in \Lambda^{r}(M_{r-1})$. It follows from the definition of the $M_i$ that $l^{M_{r}}\vert P_c$ and $l^{M_{r-1}}\vert P_{c-z}$ for all prime divisors $z$ in the support of $c$. From Proposition 4.5.1(iii) and Lemma 4.2.2 we obtain that $$\delta_{M_{r-1}}(c)\in\ \coprod\!\!\!\!\coprod(E/K)^{-\nu(c)}$$ Hence $\gamma_
{M_{r-1}}(c)$ belongs to the Selmer group ${\rm Sel}_{l^{\infty}}(E/K)^{-\nu(c)}$ (see Lemma 4.2.2). The order of the element $\delta_{M_{r-1}}(c)$ is at most $l^{M_{r-1}-M_{r}}$ by Proposition 3.4.14(ii). ${
\sqcap \!\!\!\!\sqcup}$
0.2in [**5.1.8. Lemma.**]{} [*Let $z\in \Lambda^1(n)$ be a prime divisor. Let $y$ be the unique prime of $K$ lying above $z$.* ]{}
\(i) [*The Tate pairing on $H^1(K_y,E_{l^n})$ of Theorem 2.1.4 induces a non-degenerate pairing, where $\delta=\pm1$, $$({}_{l^n}E(K_{y}))^\delta \times H^1(K_{y},E)_{l^n}^{\delta }\longrightarrow {\mathds Z}/l^n{\mathds Z}$$ of eigenspaces under the action of $\tau$ which are finite cyclic groups of order $l^n$.* ]{}
\(ii) [*The image of the homomorphism $\chi_z: {}_{l^n}E(K_y)\to H^1(K_y, E_{l^n})$ is an isotropic subgroup for the alternating Tate pairing on $H^1(K_{y}, E_{l^n})$ and ${\rm Im}(\chi_z)^\delta\cong
{\mathds Z}/l^n{\mathds Z}$ for $\delta=\pm1$.* ]{}
0.2in
[*Proof of Lemma 5.1.8.*]{} (i) We have isomorphisms of Gal$(K/F)$-modules $${}_{l^n}E(K_y)\cong E(K_y)_{l^n},\ \ H^1(K_y, E)_{l^n}\cong
{\rm Hom}(\mu_{l^n}(K_{l^n}), E(K_y)).$$ The first isomorphism here follows from Proposition 4.4.1. As $\vert\kappa(z)\vert+1\equiv 0$ (mod $l^n$) and $\kappa(y)^*$ has $\vert\kappa(z)\vert^2-1$ elements we have that $\mu_{l^n}(K_y)$ has $l^n$ elements and is contained in the $-1$ eigenspace under $\tau$ of $\kappa(y)^*$. Hence we have for $\delta=\pm1$ $${}_{l^n}E(K_y)^\delta\cong H^1(K_y,E)_{l^n}^{-\delta}$$ and these groups are cyclic of order $l^n$ by Lemma 3.2.7(ii).
\(ii) The map $\chi_z$ is defined in Proposition 4.4.2. We have by property (3) of $\chi_z$ of this Proposition 4.4.2 that the composite of $\chi_z$ with the surjective homomorphism $H^1(K_y, E_{l^n})\to H^1(K_y,E)_{l^n}$ is an isomorphism. This implies that that we have for $\delta=\pm1$, where Im$(\chi_z)$ denotes the image of $\chi_z$ in $H^1(K_y, E_{l^n})$, $${\rm Im}(\chi_z)^\delta \cong H^1(K_y,E)_{l^n}^\delta.$$
Hence part (i) on Tate duality implies that for $\delta=\pm1$ $${\rm Im}(\chi_z)^\delta\cong {\mathds Z}/l^n{\mathds Z}.$$ This evidently shows that ${\rm Im}(\chi_z)^\delta$ is an isotropic subgroup of $H^1(K_{y}, E_{l^n})$ for the anti-symmetric Tate pairing. Since the cup product on $H^1(K_{y}, E_{l^n})$ is Gal$(K/{ F})$-equivariant it follows that Im$(\chi_z)
\cong {\rm Im}(\chi_z)^{+1}\oplus {\rm Im}(\chi_z)^{-1}$ is an isotropic subgroup of $H^1(K_{y}, E_{l^n})$. ${\sqcap \!\!\!\!\sqcup}$
0.2in [**5.1.9. Lemma.**]{} [*Let $z\in \Lambda^1(n)$ be a prime divisor where $n\geq 1$. Let $y$ be the unique prime of $K$ lying above $z$. Let $S$ be a finite set of prime divisors of $\Lambda^1(n)$ not containing $z$. Let $\delta=\pm1$. Then there is a non-zero element $h\in H^1(K, E_{l^n})^\delta $ in the $\delta$-eigenspace satisfying these two conditions:* ]{}
\(a) [*$h_x\in \partial_{l^n}(E(K_x))$ for any place $x$ of $K$ not lying over a place of $S\cup \{ z\}$;* ]{}
(b)
*$h_{{x}}\in {\rm Im}( \chi_w) $ for all $w\in S$ where $x$ is the unique place of $K$ over $w$ and*
where $\chi_w:{}_{l^n}E(K_x)\to H^1(K_x,E_{l^n})$ is the homomorphism
of Proposition 4.4.2.
0.2in [*Proof of Lemma 5.1.9.*]{} The places of $S$ remain inert in the field extension $K/F$; for any $u\in S$ denote by $u^\sharp$ the corresponding place of $K$ over $u$.
For any place $v\in \Sigma_K$ of $K$ we have the exact sequence $$0\ \longrightarrow \ {}_{l^n}E(K_v)\ {\buildrel \partial_{l^n}\over \longrightarrow }\ H^1(K_v, E_{l^n})\ \longrightarrow \ H^1(K_v, E)_{l^n}\ \longrightarrow\ 0\leqno{(5.1.10)}$$ where the extremities of this sequence are in duality by Theorem 2.1.6.
Let $\delta=\pm1$. The elliptic curve $E/F$ has good reduction at all places of $S\cup\{z\}$ (remark 3.2.5). Let $U$ be the finite subset of $\Sigma_F$ of places of $F$ given by $$U=S\cup\{z\}\cup\{{\rm bad \ reduction \ places \ of \ }E/F {\rm
\ in \ } \Sigma_F\}.$$ Let $U_K$ be the finite set of all places of $K$ over the places of $U$. For all $v\in U_K$ dividing an element $u$ of $U\setminus\{ z\}$ put $$H_{v}={\rm Im}(\chi_u)^\delta{\rm \ \ if } \ u\in S$$ $$H_v=\partial_{l^n}(E(K_v))^\delta{\rm \ \ if \ } u\in U\setminus (S\cup\{z\}).$$ Then we have $$\vert H_{v}\vert^2=\vert H^1(K_v, E_{l^n})^\delta \vert {\rm \ \ for \ all \ } v{\rm \ dividing \ a\ place
\ } u \in U\setminus \{z\}.$$ For the case where $u\in S$ this follows from Lemma 5.1.8(ii). For the case where $u\in U\setminus (S\cup\{z\})$, that is to say a place of bad reduction, this follows from the exact sequence (5.1.10) and that the extremities of this sequence are in duality.
From Tate local and global duality, there is a self dual exact sequence (see \[9, Chap. 1, Theorem 4.10, p.70\]) $$H^1(K_U/K,E_{l^n})\to \bigoplus_{v\in U_K} H^1(K_v,E_{l^n})\to H^1(K_U/K, E_{l^n})^*$$where $N^*$, for a ${\mathds Z}/l^n{\mathds Z}$-module $N$, denotes ${\rm Hom}(N, {\mathds Z}/l^n{\mathds Z})$, and where $K_U$ is the maximal extension of $K$ unramified outside $U_K$. Hence the image $I$ of $H^1(K_U/K, E_{l^n})$ in $\bigoplus_{v\in U_K} H^1(K_v,E_{l^n})$ is a maximal isotropic subgroup, for the Tate pairing, of the group $$\bigoplus_{v\in U_K} H^1(K_v,E_{l^n}).$$ Since at the place $y$ lying over $z$ we have $H^1(K_y,E_{l^n})^\delta \not=0 $ by Lemma 5.1.8(i), the subgroup $I^\delta$ is of order strictly larger than that of $$\bigoplus_{u\in S} {H^1(K_{u^\sharp},E_{l^n})^\delta\over H_{u^\sharp}}.$$ Hence the natural homomorphism $$H^1(K_U/K,E_{l^n})^\delta \to \bigoplus_{u\in S} {H^1(K_{u^\sharp},E_{l^n})^\delta \over H_{u^\sharp}}$$ has non-zero kernel. Therefore we may select a non-zero element $h\in H^1(K_U/K, E_{l^n})^\delta$ in this kernel. Then $h$ satisfies the condition (b), that is to say we have $h_{u^\sharp}\in H_{u^\sharp}$ for all $u\in S$. Furthermore, $h$ satisfies the condition (a) by the selection of $H_v$ if $v\in\Sigma_K$ is a bad reduction place of $E$ and because we have $H^1(K_v^{\rm un}/K_v, E_{l^n})=\partial_{l^n}(E(K_v))$ if $v\in \Sigma_K$ is a good reduction place of $E$, where $K_v^{\rm un}$ is the maximal unramified separable extension of $K_v$, ${\sqcap \!\!\!\!\sqcup}$
0.2in
[*5.1.10. Remark.*]{} The Lemma 5.1.9 is a technical result required for the proof of Proposition 5.2.1 in the next section.
0.4in [**5.2. A class $\gamma_n(c)$ in the Selmer group**]{}
0.2in The notation and hypotheses (5.1.1) of section 5.1 hold also for this section.
0.2in [**5.2.1. Proposition.**]{} [*Let $r\geq 1$ be an integer and put $\nu(r)=(-1)^r\epsilon$. Let $G$ be a subgroup of the Selmer eigenspace ${\rm Sel}_{l^\infty}(E/K)^{-\nu(r)}$ such that $${\rm rank}( G)\leq r.$$ Assume that $M_{r-1}$ is finite and that $l\in {\cal P}$ is coprime to the order of [Pic]{}$(A)$. Then there is a cohomology class $\gamma_{M_{r-1}}(c)$ for some divisor $c\in \Lambda^r(M_{r-1})$ such that:*]{}
\(i) [*$\gamma_{M_{r-1}}(c)$ belongs to the Selmer eigenspace $ {\rm Sel}_{l^\infty}(E/K)^{-\nu(r)}$;* ]{}
\(ii) [*$\gamma_{M_{r-1}}(c)$ has order $l^{M_{r-1}-M_r}$;*]{}
\(iii) [ *${\mathds Z}\gamma_{M_{r-1}}(c)\cap G=\{0\}.$* ]{} 0.2in
[*Proof of Proposition 5.2.1.*]{} The group $ {\rm Sel}_{l^\infty}(E/K)$ is an $l^\infty$-torsion group, that is to say every element is annihilated by a power of $l$. Hence the subgroup $G$ of finite rank is therefore finite.
Let exp$(G)$ be the exponent of $G$, that is to say the largest order of an element of the abelian $l$-group $G$. Select an integer $$m\geq 1$$ such that $$l^m\geq \max\{ {\rm exp }( G),l^{M_{r-1}}\}$$ and put $$L=K(E_{l^m}).$$ Put $$t={\rm rank}(G)$$ where $${\rm rank}(G)={\rm dim}_{{\mathds Z}/l{\mathds Z}} \ G/lG$$ and where $t\leq r$ by hypothesis. Note that $G$ is a subgroup of Hom$({\rm Gal}(L^{\rm sep}/L), E_{l^m}(L))$, as $l^m\geq $ exp$(G)$ (see Lemma 3.3.2), and the elements of $\Lambda^1(m)$ are unramified in $L/{ F}$. As in §1.2, for any divisor $c$ on $F$, ${\rm Supp}(c)$ denotes the set of distinct prime divisors in the support of $c$.
For any divisor $c$ in $\Lambda^r(1)$, put $$\Xi(c)={\rm Supp}(c)\cap\Lambda^1(m)$$ that is to say $\Xi(c)$ is the set of prime divisors in the support of $c$ which belong to $\Lambda^1(m)$; in particular, $\Xi(c)$ depends on $m$. For any prime divisor $z$ of $F$ in $\Lambda^1(1)$, select a place $z^\times$ of $L=K(E_{l^m})$ lying over $z$. Let $\Gamma(c)\subseteq {\widehat {\ G\ }}$ be the subgroup of characters of the abelian group $G$ generated by the set of characters $$\big\{\phi_{{\rm Frob}(z^\times)}\ \ \vert\ \ \ z\in \Xi(c)\big\}$$ (as in Proposition 3.3.3 and (3.3.4), applied to the finite group $G$, and Proposition 3.3.7). Put $$s(c) ={\rm dim}_{{\mathds Z}/l{\mathds Z}}\ \ {\Gamma(c)+l
{\widehat {\ G\ }}
\over l{\widehat {\ G\ }}}.$$ That is to say, $s(c)$ is the dimension of the image of $\Gamma(c)$ in the vector space ${\widehat {\ G\ }}/l{\widehat {\ G\ }}$ of dimension $t$; the non-negative integer $s(c)$ is then at most equal to $t$. Put $$n(c)=\vert \Xi(c)\vert$$ that is to say $n(c)$ is the cardinality of $\Xi(c)$ and which is a non-negative integer at most equal to $r$, the number of prime divisors in the support of $c$.
Define the [*defect*]{} $\Delta(c)$ of a divisor $c\in \Lambda^r(1)$ on $F$ to be $$\Delta(c)=\max(t-s(c), r-n(c)).$$ Then we have $$0\leq \Delta(c)\leq r$$ and we have $$\Delta(c)=0{\rm \ \ if \ and \ only \ if \ \ } \Gamma(c)={\widehat {\ G\ }} {\rm \ and \ }
c\in \Lambda^r(m).$$ This equivalence holds because $s(c)=t$ if and only if $\Gamma(c)={\widehat {\ G\ }}$ by Nakayama’a Lemma.
We have $M_{r-1}\geq M_r$ by Lemma 5.1.4 and in particular $M_r$ is finite as $M_{r-1}$ is assumed to be finite. We may then select a divisor $$d\in \Lambda^r(M_r+1)\leqno{(5.2.2)}$$ such that $$l^{M_r}\vert\vert P_d.$$ Then the cohomology class $\gamma_{M_r+1}(d)$ is defined and belongs to $H^1(K, E_{l^{M_r+1}})^{-\nu(r)}$ and has exact order $l$ by Lemma 4.2.2 and Proposition 3.4.14(i).
Suppose that $M_r=M_{r-1}$. Then $\gamma_{M_{r-1}}(d)$ is equal to zero and evidently belongs to the Selmer eigenspace ${\rm Sel}_{l^\infty}(E/K)^{-\nu(r)}$ which proves the lemma in this trivial case where $M_r=M_{r-1}$.
We may suppose for the rest of the proof of this Proposition 5.2.1 that $M_{r-1}>M_r$.
Assume that the defect of the divisor $d$ already selected in (5.2.2) satisfies $$\Delta(d)>0.$$ That is to say either (where $s(d)<t$) the image of $\Gamma(d)$ is a proper subspace of ${\widehat {\ G\ }}/l{\widehat {\ G\ }}$ or (where $n(d)<r$) that $d\not\in \Lambda^r(m)$. These two possibilities $s(d)<t$ and $n(d)<r$ for the divisor $d$ are not mutually exclusive.
We select a character $\psi\in
{\widehat {\ G\ }}$ and a prime divisor $z_0\in$ Supp$(d)$ by the following recipe.
0.2in[*Selection of the character $\psi$*]{}
If $s(d)<t$ select a character $\psi\in
{\widehat {\ G\ }}$ such that $$\psi(\gamma_{M_r+1}(d))\not=0, {\rm \ \ if \ } \gamma_{M_r+1}(d)\in G\leqno{(5.2.3)}$$ and $$\psi\in {\widehat {\ G\ }}\setminus (\Gamma(d)+l{\widehat {\ G\ }})\leqno{(5.2.4)}$$ where the condition (5.2.3) is vacuous if $\gamma_{M_r+1}(d)\not\in G$. A character $\psi\in {\widehat {\ G\ }}$ exists which satisfies the two conditions (5.2.3) and (5.2.4) because a finite group cannot be the union of two proper subgroups.
If $s(d)=t$ and $n(d)<r$ select any character $\psi\in {\widehat {\ G\ }}$ such that $$\psi(\gamma_{M_r+1}(d))\not=0, {\rm \ \ if \ } \gamma_{M_r+1}(d)\in G\leqno{(5.2.5)}$$ where the condition (5.2.5) is vacuous if $\gamma_{M_r+1}(d)\not\in G$. 0.6in[*Selection of the divisor $z_0\in {\rm Supp}(d)$*]{}
If $s(d)<t$ select a prime divisor $z_0\in $ Supp$(d)$ such that $$\big\{\phi_{{\rm Frob}(z^\times)}\ \ \vert\ \ \ z\in \Xi(d-z_0)\big\}$$ is a generating set for $\Gamma(d)$ modulo $l{\widehat {\ G\ }}$. This choice is possible because $t\leq r$ and Supp$(d)$ has $r$ distinct elements.
If $s(d)=t$ and $n(d)<r$ select a prime divisor $z_0\in$ Supp$(d)$ such that $z_0\not\in \Lambda^1(m)$.
0.2in We have now defined the pair $\psi,z_0$ for the divisor $d$ where the defect $\Delta(d)>0$. By Lemma 5.1.9 we may select a cohomology class $h$ in the $\nu(r)$-eigenspace $$h\in H^1(K, E_l)^{\nu(r)}$$ of the group scheme $E_l$ such that 0.2in (5.2.6) $h\not=0$;
(5.2.7) $h_v\in \partial_l(E(K_v)) $ for all places $v$ of $K$ coprime to ${\rm Supp}(d)$;
(5.2.8) $h_{y}\in $ Im$(\chi_z)$ for all $z\in {\rm Supp}(d-z_0)$ where $y$ is the prime of $K$ lying
over $z$ and $\chi_z$ is the homomorphism of §4.4. 0.2in Note that $H^1(K, E_l)^{\nu(r)}$ is a subgroup of $H^1(K, E_{l^m})$ by Lemma 3.1.5.
Since $h$ is in a different eigenspace from $G$ and $ \gamma_{M_r+1}(d)$, which both belong to the $-\nu(r)$-eigenspace, we have $$(G+{\mathds Z}\gamma_{M_r+1}(d))\cap {\mathds Z}h=0$$ where $G+{\mathds Z}\gamma_{M_r+1}(d)$ is the subgroup of $H^1(K, E_{l^m})^{-\nu(r)}$ generated by $G$ and $\gamma_{M_r+1}(d)$ and where ${\mathds Z}h$ is the subgroup of $H^1(K, E_{l^m})^{\nu(r)}$ generated by $h$.
Let $D$ be the subgroup of $H^1(K, E_{l^m})$ generated by $G$, $\gamma_{M_r+1}(d)$ and $h$ $$D=G+{\mathds Z} \gamma_{M_r+1}(d)+{\mathds Z}h.$$ Then $D$ is a finite subgroup of $H^1(K,E_{l^m})$ by Lemma 3.1.5. As $D$ is isomorphic to the direct product of $G+{\mathds Z}\gamma_{M_r+1}(d)$ and ${\mathds Z}h$, we can select a homomorphism $$\chi:
D\to {\mathds Q}/{\mathds Z}$$ such that (by (5.2.3), (5.2.5), and (5.2.6)) 0.2in
(5.2.9) $\chi\vert_G=\psi$;
(5.2.10) $ \chi(\gamma_{M_r+1}(d))\not=0$;
(5.2.11) $\chi(h)\not=0$.
0.2in By Proposition 3.3.7 applied to the finite group $D$, there is a prime divisor $$z_1\in \Lambda^1(m)\leqno{(5.2.12)}$$ distinct from the elements of Supp$(d)$ such that, in the notation of (3.3.4), $$\chi=\phi_{{\rm Frob}(z^\times_1)}\leqno{(5.2.13)}$$ where $z^\times_1$ is a prime divisor of $L=K(E_{l^m})$ above $z_1$. We then obtain the cohomology class $\gamma_{M_r+1}(d+z_1)$ associated to the divisor $d+z_1$.
For all places $v$ of $K$, denote the alternating cup product induced by the Weil pairing of $h$ with $\gamma_{M_r+1}(d+z_1)$ localized at $v$ by (see Theorem 2.1.4) $$<\gamma_{M_r+1}(d+z_1)_v, h_v>_v$$ which is an element of ${\mathds Q}/{\mathds Z}$. The sum of local pairings over all places of $K$ $$\sum_{{\rm all\ places \ } v {\rm \ of \ } K}<\gamma_{M_r+1}(d+z_1)_v, h_v>_v=0\leqno{(5.2.14)}$$ is zero by Proposition 2.1.8.
If $v$ does not divide any element of $ {\rm Supp}(d+z_1)$ then $$\gamma_{M_r+1}(d+z_1)_v\in \partial_{l^{M_r+1}}(E(K_v))$$ by Proposition 4.5.1(i) and $$h_v\in \partial_l(E(K_v))$$ by (5.2.7), that is to say both localized elements $\gamma_{M_r+1}(d+z_1)_v$, $h_v$ are in the image of the map $$\partial_{l^m}: E(K_v)\to H^1(K_v,E_{l^m}).$$But the image of this map $\partial_{l^m}$ in $H^1(K_v,E_{l^m})$ is an isotropic subgroup for the alternating pairing $<,>_v$ (see \[1, p. 403, Theorem 7.15.6\] or Theorem 2.1.6(i)). Therefore we have $$<\gamma_{M_r+1}(d+z_1)_v, h_v>_v=0{\rm \ \ for \ all \ } v{\rm \ coprime \ to \ Supp}(d+z_1).$$
If $y$ is a place of $K$ which divides an element $z$ of $ {\rm Supp}(d-z_0)$, then we have $h_y\in $ Im$(\chi_z)$ by (5.2.8) and $$\gamma_{M_r+1}(d+z_1)_y\in {\rm Im}(\chi_z)$$ by property (4) of Proposition 4.4.2, which is the main property of the map $\chi_z$. As Im$(\chi_z)$ is isotropic for the cup product $<,>_v$ by Lemma 5.1.8(ii), we have $$<\gamma_{M_r+1}(d+z_1)_y, h_y>_y=0 {\rm \ \ for \ all \ } y{\rm
\ dividing\
an\ element \ of \ Supp}(d-z_0).$$
Therefore the only possible non-zero terms in the sum $\sum_v<\gamma_{M_r+1}(d+z_1)_{v}, h_{v}>_{v}$ of (5.2.14) are for the places of $K$ lying over the places $z_0$ and $z_1$. For these places $z_i\in \Lambda^1$ of $F$, for $i=0,1$, denote by $y_i$ the corresponding place of $K$ lying over the place $z_i$ which remains inert in $K/F$.
From (5.2.10) and (3.3.5) and that $\chi=\phi_{{\rm Frob}(z_1^\times)}$ by (5.2.13), we have that this localization $\gamma_{M_r+1}(d)_{y_1}$ at $y_1$ is non-zero; hence by Proposition 4.5.1(iv) the localization $$\delta_{M_r+1}(d+z_1)_{y_1}\in H^1(K_{y_1}, E)_{l^{M_r+1}}^{\nu(r)}$$ is non-zero. Furthermore we have $$h_{y_1}\in\partial_l (E(K_{y_1}))^{\nu(r)}$$ by (5.2.7) and $h_{y_1}$ is non-zero by (5.2.11) and (5.2.13). Hence we have by Lemma 5.1.8(i) $$<\gamma_{M_r+1}(d+z_1)_{y_1}, h_{y_1}>_{y_1}\not=0.$$ Since the sum $\sum_v<\gamma_{M_r+1}(d+z_1)_{v}, h_{v}>_{v}$ of (5.2.14) is zero, this implies that the local term at $y_0$ $$<\gamma_{M_r+1}(d+z_1)_{y_0}, h_{y_0}>_{y_0}$$ is non-zero. Hence we have $\gamma_{M_r+1}(d+z_1)_{y_0}\not=0$ and so by Proposition 4.5.1(iv) $$P_{d+z_1-z_0}\not\in l^{M_r+1}E(K_{y_0})$$ hence we obtain, where $z_1\in \Lambda^1(m)$, $$l^{M_r}\vert\vert P_{d+z_1-z_0}\leqno{(5.2.15)}$$ because we have $l^{M_r}\vert P_{d+z_1-z_0}$ and $d+z_1-z_0\in \Lambda^r(M_r+1)$ as well as $d\in \Lambda^r(M_r+1)$ and $z_1\in \Lambda^1(m)$.
Put $$c_1=d+z_1-z_0 \in \Lambda^r(M_r+1).$$ On the one hand, we have that if $s(d)<t$ then the group $\Gamma(c_1)$ is generated by $\Gamma(d)$ and $\psi$ $$\Gamma(c_1)=\Gamma(d)+{\mathds Z}\psi \leqno{(5.2.16)}$$ because $\chi=\phi_{{\rm Frob}(z^\times_1)}$ by (5.2.13) and $\chi\vert_G=\psi$ by (5.2.9) in this case where $s(d)<t$. The condition that $z_1\in
\Lambda(m)$ is satisfied by the choice of $z_1$ in (5.2.12). We then have from (5.2.16) if $s(d)<t$, as $\psi\not\in \Gamma(d)+l{\widehat {\ G\ }}$ by (5.2.4), $$s(c_1)={\rm dim}_{{\mathds Z}/l{\mathds Z}} { \Gamma(c_1)+l{\widehat {\ G\ }}\over l{\widehat {\ G\ }}}= s(d)+1.$$
On the other hand, if $s(d)=t$ and $n(d)<r$ then $n(c_1)=n(d)+1$ as $z_1\in \Lambda^1(m)$ by (5.2.12) and $z_0\not\in \Lambda^1(m)$ by the selection of $z_0\in $ Supp$(d)$. Hence we have in both cases that the defect of $c_1$ is given by $$\Delta(c_1)=\Delta(d)-1.$$
We may by this method construct by induction a sequence of divisors $c_1,c_2,c_3\ldots$ in $\Lambda^r(M_r+1)$ such that their defects are strictly decreasing $$\Delta(d)>\Delta(c_1) > \Delta(c_2)> \ldots$$ and where, as in (5.2.15), $$l^{M_r}\vert\vert P_{c_i}{\rm \ \ for \ all \ } i.\leqno{(5.2.17)}$$ This sequence must terminate in a divisor $c$ with zero defect $\Delta(c)=0$, that is to say $\Gamma(c)={\widehat {\ G\ }}$, by Nakayama’s Lemma, and $c\in \Lambda^r(m)$ and where $$l^{M_r}\vert\vert P_{c}.\leqno{(5.2.18)}$$
The cohomology class $\gamma_m(c)$ is therefore defined and as $\Gamma(c)={\widehat {\ G\ }}$ we have, where $z^\times$ is a place of $K(E_{l^m})$ over $z$, $$\{ g\in G\vert\ \phi_{{\rm Frob}(z^\times)} (g)=0 {\rm \ for \ all\ } z\in
{\rm Supp}(c)\}=0.\leqno{(5.2.19)}$$ We have from Lemma 5.1.7(ii) that $$\gamma_{M_{r-1}}(c)\in
{\rm Sel}_{l^\infty}(E/K)^{-\nu(r)}.\leqno{(5.2.20)}.$$ We have by the definition of $M_{r-1}$ that $\gamma_{M_{r-1}}(c-z)=0$ for any prime divisor $z$ in the support of $c$ as $c-z$ has $r-1$ elements in its support. But by Proposition 4.5.1(iv), we have for any $z\in {\rm Supp}(c)$ where $y$ is the place of $K$ over $z$, $${\rm ord}\ \gamma_{M_{r-1}}(c)_{y}= {\rm ord}\ \gamma_{M_{r-1}}(c-z)_{y}.$$ Hence we obtain for any $z\in {\rm Supp}(c)$ that $\gamma_{M_{r-1}}(c)_{y}=0$ where $y$ is the prime of $K$ over $z$. We then obtain from (5.2.19), as $g_y= \phi_{{\rm Frob}(z^\times)} (g)$ where $y$ is the place of $K$ above $z$, that $$G\cap {\mathds Z}\gamma_{M_{r-1}}(c)=0.$$ Since $l^{M_r}\vert\vert P_c$, by (5.2.18), we have that $\gamma_m(c)$ has order $l^{m-M_r}$ for all $m\geq M_r$ by Proposition 3.4.14(i). Hence $\gamma_{M_{r-1}}(c)$ has order $l^{M_{r-1}-M_r}$ and belongs to the Selmer group Sel$_{l^\infty}(E/K)^{-\nu(r)}$, by (5.2.20), which proves the proposition. ${\sqcap \!\!\!\!\sqcup}$
1.6in [**5.3. Proof of Theorem 4.1.15**]{}
0.2in(5.3.1) The notation (5.1.1) of section §5.1 holds also for this section.
0.2in[**5.3.2. Lemma.**]{} [*Let $A$ be a finite abelian $p$-group where $p$ is a prime number and with invariants $I_1\geq I_2\geq \ldots \geq I_r$. Let $B$ be a subgroup of $A$ with invariants $I_1\geq I_2\geq \ldots\geq I_s$ where $s\leq r$. Then there is a subgroup $C$ of $A$ such that $$A=B\oplus C$$ that is to say, $A$ is the direct sum of the subgroups $B,C$. The invariants of $C$ are $I_{s+1}\geq I_{s+1}\geq \ldots\geq I_r$.* ]{}
0.2in \[The proof of this result follows from the structure theorem of finite abelian groups and is omitted.\] ${\sqcap \!\!\!\!\sqcup}$
0.2in For the proof of Theorem 4.1.15, as $P_0$ has infinite order in $E(K)$, the Tate-Shafarevich group $\coprod\!\!\!\coprod$$
(E/K)$ is finite and $P_0$ generates a subgroup of finite index in $E(K)$ by \[1, Chapter 7, Theorem 7.6.5\] which is one of the principal results of this monograph \[1\].
The image of $P_0$ in ${}_{l^m}E(K)$ belongs to the $-\epsilon$-eigenspace of ${}_{l^m}E(K)$ for all $m$ (by \[1, Chapter 7, Lemma 7.14.11\] or Lemma 4.2.2 above). It follows from the decomposition into eigenspaces $${}_{l^m}E(K)\cong (
{}_{l^m}E(K) )^\epsilon\oplus ({}_{l^m}E(K))^{-\epsilon}$$ that the $\epsilon$-eigenspace $({}_{l^m}E(K))^\epsilon$ is a finite abelian group of order bounded independently of $m$ and hence $E(K)^\epsilon$ is a finite abelian group. As $E(K)$ has no $l$-torsion (as $l\in {\cal P}$ see Proposition 1.10.1 and Definition 3.1.3(f)) it follows that $$({}_{l^m}E(K))^\epsilon=0{\rm \ \ \ for\ all \ } m\leqno{(5.3.3)}$$ and that $$({}_{l^m}E(K))^{-\epsilon}\cong {\mathds Z}/l^m{\mathds Z} {\rm \ \
for \ all \ }m\leqno{(5.3.4)}$$ which proves the isomorphisms (4.1.18).
The $l^m$-Selmer group ${\rm Sel}_{l^m}(E/K)^\pm$ belongs to an exact sequence of eigenspaces $$0\longrightarrow \big( {}_{l^m}{E(K)}\big)^{\pm}\longrightarrow
{\rm Sel}_{l^m}(E/K)^\pm\longrightarrow
{\coprod\! \!\!\!\coprod}(E/K)
_{l^m}^\pm\longrightarrow 0.\leqno{(5.3.5)}$$ Hence this exact sequence induces an isomorphism of $\epsilon$-eigenspaces $${\rm Sel}_{l^\infty}(E/K)^\epsilon\cong
{\coprod\! \!\!\!\coprod}(E/K)_{l^\infty}^\epsilon.\leqno{(5.3.6)}$$
The largest invariant of the $-\epsilon$-eigenspace ${\rm Sel}_{l^m}(E/K)^{-\epsilon}$ is at most equal to $m$. But this group ${\rm Sel}_{l^m}(E/K)^{-\epsilon}$ contains the subgroup $({}_{l^m}E(K))^{-\epsilon}\cong {\mathds Z}/l^m{\mathds Z}$ with invariant $m$. By Lemma 5.3.2 applied to the eigenspace ${\rm Sel}_{l^m}(E/K)^{-\epsilon}$ and the subgroup $({}_{l^m}E(K))^{-\epsilon}$, we have that the $-\epsilon$-eigencomponent of the exact sequence (5.3.5) splits. Hence we obtain an isomorphism of $-\epsilon$-eigenspaces $${\rm Sel}_{l^m}(E/K)^{-\epsilon} \cong
{}_{l^m}E(K)\oplus
{\coprod\! \!\!\!\coprod}(E/K)
_{l^m}^{-\epsilon} {\rm \ \
for \ all \ } m.\leqno{(5.3.7)}$$ The isomorphisms (5.3.6) and (5.3.7) prove the theorem. ${\sqcap \!\!\!\!\sqcup}$
0.4in
[**5.4. Proof of Theorems 4.1.9 and 4.1.13**]{}
0.2in
(5.4.1) The principle of the proof is to construct inductively divisors $c_{k}\in \Lambda^{k}$, $k=1,2\ldots$, such that the cohomology classes $\delta_{M_{k-1}}(c_{k})$ in the Tate-Shafarevich group $ {\coprod\! \!\!\coprod}(E/K)
_{l^\infty}$ form a basis of a maximal isotropic subgroup with respect to the Cassels pairing and where $\delta_{M_{k-1}}(c_{k})$ has order $l^{N_{k}}$ for all $k$. The inductive step is provided by the next Lemma 5.4.3.
0.2in(5.4.2) The notation and hypotheses (5.1.1) of the previous section holds also for this section. We further denote by, where $l\in \cal P$, 0.2in $\Sigma_K$ the set of all places of the global field $K$;
$[,]_v: {}_{l^m}E(K_v)\times H^1(K_v,E)_{l^m}\to {\mathds Z}/l^m{\mathds Z}$ the non-degenerate local pairing at
the place $v$ of $K$ induced by the cup product, as in Theorem 2.1.6. 0.2in
If $z_i\in \Lambda^1(1)$ is a prime divisor of $F$ indexed by an integer $i$, denote by $y_i$ the prime divisor of $K$ lying above the place $z_i$ of $F$ where this place $z_i$ is inert in the field extension $K/F$.
0.2in[**5.4.3. Lemma.**]{} [*Assume that $l\in {\cal P}$ is coprime to the order of [Pic]{}$(A)$. Let $s\geq 1$ be a positive integer and let $r,t\geq0$ be non-negative integers. Let $S$ be a subgroup of [Sel]{}$_{l^s}(E/K)$. Let $e\in {\rm Sel}_{l^s}(E/K)^{-\nu(r+1)}$, $\gamma_s(c)\in {\rm Sel}_{l^s}(E/K)^{-\nu(r)}$, where $c\in \Lambda^r(s+t)$, be elements of the Selmer group where $$S\cap (e,\gamma_s(c))=0$$ and where $(e,\gamma_s(c))$ is the subgroup of the Selmer group generated by $e,\gamma_s(c)$. Suppose also that $e, \gamma_s(c)$ both have order $l^n$ where $n\leq s$. Then there are infinitely many prime divisors $z\in \Lambda^1(s+t)$ coprime to ${\rm Supp}(c)$ such that if $y$ is the place of $K$ over $z$ then we have*]{}
\(1) $S_y=0$;
\(2) [*the value in ${\mathds Z}/l^n{\mathds Z}$ of the local pairing at $y$ with the class $\delta_s(c+z)$ $$[\delta_s(c+z)_y,w_y ]_y$$ has order $l^n$ where $w_y\in {}_{l^n}E(K_y)$ is a point such that $e_y=\partial_{l^n }(w_y)$.*]{}
0.2in [*Proof of Lemma 5.4.3.*]{} Note that from (3.1.10), we have that Sel$_{l^n}(E/K)$ is the subgroup of Sel$_{l^s}(E/K)$ annihilated by $l^n$ and in particular Sel$_{l^n}(E/K)$ contains $e$ and $\gamma_s(c)$.
Let $T$ be the subgroup of the Selmer group ${\rm Sel}_{l^s}(E/K)$ generated by $S, e$ and $\gamma_s(c)$. Then we have an isomorphism $$T\cong S\oplus (e,\gamma_s(c)).$$
For a fixed non-zero element $x\in T$, the set of characters $$\chi:T\to {\mathds Z}/l^s{\mathds Z}$$ such that $${\rm ord}(\chi(x))<{\rm ord}(x)$$ is a proper subgroup of ${\widehat T}$. This follows as the subgroups of ${\mathds Z}/l^s{\mathds Z}$ are linearly ordered ${\mathds Z}/l^s{\mathds Z}\supseteq l{\mathds Z}/l^s{\mathds Z}\supseteq
l^2{\mathds Z}/l^s{\mathds Z}\ldots$. As a group cannot be the union of two proper subgroups, there is then a character $\chi:T\to {\mathds Z}/l^s{\mathds Z}$ such that $${\rm ord}(\chi(e))={\rm ord}(e)$$ $${\rm ord}(\chi(\gamma_s(c))={\rm ord}(\gamma_s(c))$$ $$\chi(S)=0.$$
By Proposition 3.3.8 applied to the subgroup $T$ and the character $\chi$, we may select a prime divisor $z\in \Lambda^1(s+t)$ satisfying, where $y$ is the place of $K$ lying over $z$, $${\rm ord}(e_y)={\rm ord}(e)$$ $${\rm ord}(\gamma_s(c)_y)={\rm ord}(\gamma_s(c))$$ $$S_y=0$$ where the subscript $y$ denotes the restriction at $y$ of elements of the Selmer group Sel$_{l^s}(E/K)$. In particular, condition (1) of the lemma is satisfied for this $z$.
The class $\delta_s(c+z)$, associated to the divisor $c+z\in \Lambda^{r+1}(s+t)$, then belongs to $H^1(K,E)_{l^s}^{-\nu(r+1)}$ and $e$ belongs to $H^1(K,E_{l^s})^{-\nu(r+1)}$.
We may select a point $w_y\in ({}_{l^n}E(K_y))^{-\nu(r+1)}$ such that $e_y=\partial_{l^n}(w_y)$ where $e\in$ Sel$_{l^n}(E/K)$ as already noted. Then $w_y$ has order $l^n$ in ${}_{l^n}E(K_y)$ as $e_y$ has order $l^n$.
Furthermore because $${\rm ord}(\gamma_s(c)_y)={\rm ord}(\gamma_s(c))$$ and by Proposition 4.5.1(iv) we must have that $${\rm ord}(\delta_s(c+z)_y)={\rm ord}(\gamma_s(c))=l^n.$$
The class $\delta_s(c+z)_y$, of order $l^{n}$, belongs to the subgroup $H^1(K_y,E)_{l^n}^{-\nu(r+1)}$ of $H^1(K_y,E)_{l^s}^{-\nu(r+1)}$ and $w_y$, of order $l^{n}$, belongs to $({}_{l^n}E(K_y))^{-\nu(r+1)}$. In particular, $\delta_s(c+z)$ and $w_y$ both belong to the $-\nu(r+1)$-eigenspaces of their respective spaces. Hence the local term $$[\delta_s(c+z)_y,w_y ]_y$$ has order $l^n$. This follows from the non-degeneracy of the local pairing: by Lemma 5.1.8(i), if $z\in \Lambda^1(n)$ and $y\in \Sigma_K$ is the place of $K$ over $z$ then the two elements $$f\in {}_{l^n}E(K_{y} ), \
d\in H^1(K_{y} ,E)_{l^n}$$ give via the Tate pairing an element $$[f,d]_{y}$$ which is non-zero if they are in the same $\tau$-eigenspace and the product of their orders is $\geq l^n$. Therefore condition (2) is satisfied by $z$ which proves the lemma.${\sqcap \!\!\!\!\sqcup}$
0.2in We now prove simultaneously the two Theorems 4.1.9 and 4.1.13.
As $P_0$ has infinite order in $E(K)$, the Tate-Shafarevich group $\coprod\!\!\!\coprod$$
(E/K)$ is finite and $P_0$ generates a subgroup of finite index in $E(K)$ by \[1, Chapter 7, Theorem 7.6.5\].
By definition of the invariants $N_i$ of the finite abelian group $\coprod\!\!\!\coprod$$
(E/K)_{l^\infty}$ in §4.1.5, there is a maximal isotropic subgroup $D$ of $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}$, with respect to the non-degenerate anti-symmetric Cassels pairing, where $$D=\prod_i D_i$$ and each $D_i$ is a cyclic group of order $l^{N_i}$ and where $$D^{\epsilon }= \prod_{ i{\rm \ odd}} D_i$$ and $$D^{-\epsilon }=\prod_{i{\rm \ even}} D_i
.$$
From Theorem 4.1.15, we have the decomposition of eigenspaces $${\rm Sel}_{l^m}(E/K)^{\pm} \cong
\big({}_{l^m}E(K) \big)^{\pm}\oplus
{\coprod\! \!\!\!\coprod}(E/K)
_{l^m}^\pm {\rm \ \
for \ all \ } m\geq 0\leqno{(5.4.4)}$$ where the projection onto the second factor is given by the natural surjection $$\pi_m: {\rm Sel}_{l^m}(E/K)\longrightarrow {\coprod\! \!\!\!\coprod}(E/K)
_{l^m}.$$
Let $m$ be an integer such that $l^m$ is greater than or equal to the exponent of the finite group ${\coprod\! \!\!\coprod}(E/K)
_{l^\infty}$, that is to say $m\geq \max_iN_i$. For each integer $i$, let $d_i\in
\coprod\!\!\!\coprod$$(E/K)_{l^\infty}^{-\nu(i)}$ be a generator of $D_i$ and let $e_i$ such that $$e_i\in {\rm Sel}_{l^m}(E/K)^{-\nu(i)}, \ {\rm where\ } \pi_m(e_i)=d_i,
\ {\rm ord}(e_i)=l^{N_i},\leqno{(5.4.5)}$$ be the lifting of $d_i$ to the Selmer eigenspace ${\rm Sel}_{l^m}(E/K)^{-\nu(i)}$ via the decomposition (5.4.4) such that $e_i$ has zero component in the subgroup ${}_{l^m} E(K)$; in particular, we take $e_i$ to have order equal to $l^{N_i}$ for all $i$ and to belong to the $-\nu(i)$-eigenspace as $d_i$ has order $l^{N_i}$. For each valuation $v$ of $K$ and each $i$, select $w_{i,v}\in {}_{l^{N_i}}E(K_v)$ such that the localization $e_{i,v}$ of $e_i$ at $v$ satisfies $$e_{i,v}=\partial_{l^{N_i}}(w_{i,v}).\leqno{(5.4.6)}$$ Here $$\partial_{l^{N_i}}: {}_{l^{N_i}}E(K_v)\to H^1(K_v, E_{l^{N_i}})\leqno{(5.4.7)}$$ denotes the connecting homomorphism associated to the morphism $l^{N_i}:E\to E$ of multiplication by $l^{N_i}$.
The cohomology class $\gamma_{M_0+N_1}(0)$ belongs to the Selmer group Sel$_{l^m}(E/K)^{-\epsilon }$ and has order $l^{N_1}$ as $l^{M_0}\vert\vert P_0$, by Lemma 5.1.2 or Proposition 3.4.14(i), $${\rm ord}(\gamma_{M_0+N_1}(0))=l^{N_1}.$$ The element $e_1\in$ Sel$_{l^m}(E/K)^{\epsilon }$ has the same order $${\rm ord}(e_1)= l^{N_1}.$$ Let $S$ be the subgroup of the Selmer group ${\rm Sel}_{l^m}(E/K)$ generated by $e_i$ for all $i\geq 2$. The element $\gamma_{M_0+N_1}(0)$ belongs to the component ${}_{l^m}E(K)$ in the decomposition (5.4.4) of ${\rm Sel}_{l^m}(E/K) $. Hence we have that the subgroup of ${\rm Sel}_{l^m}(E/K)$ generated by $S,e_1, \gamma_{M_0+N_1}(0)$ is direct sum $$S\oplus {\mathds Z}e_1\oplus {\mathds Z}\gamma_{M_0+N_1}(0).$$
We may now apply Lemma 5.4.3 to $S$ and the elements $e_1, \gamma_{M_0+N_1}(0)$ where we take the parameters of the lemma to be $$c=0, r=0, s=M_0+N_1, t=0, n=N_1.\leqno{(5.4.8)}$$ There is according to the lemma a prime divisor $z_1\in \Lambda^1(M_0+N_1)$ which satisfies the following 2 conditions, where $y_1\in \Sigma_K$ is the prime divisor of $K$ lying over $z_1$ and where the subscript $y_1$ denotes localization at $y_1$, and where the point $w_{1,y_1}$ in $(E(K_{y_1})/l^{N_1}E(K_{y_1}))^{\epsilon }$ is such that $\partial_{l^{N_1}} (w_{1,y_1})=e_{1,y_1}$:- $${\rm ord} [\delta_{M_0+N_1}(z_1)_{y_1},w_{1,y_1}]_{y_1} =l^{N_1}\leqno{(5.4.9)}$$ and $$S_{y_1}=0.\leqno{(5.4.10)}$$ Here in (5.4.9), $\delta_{M_0+N_1}(z_1)$ is the cohomology class in $H^1(K,E)_{l^{M_0+N_1}}^\epsilon$ associated to $z_1$.
Let $$\delta_{M_0}(z_1)\in {\coprod\! \!\!\!\coprod}(E/K)^{\epsilon
}_{l^\infty}$$ be the cohomology class associated to this prime divisor $z_1\in \Lambda^1(M_0+N_1)$; that $\delta_{M_0}(z_1)$ belongs to the Tate-Shafarevich group $\coprod\!\!\!\coprod$$(E/K)^{\epsilon
}_{l^\infty}$ follows from Lemma 5.1.7(i).
On the one hand, $l^{N_1}$ is the maximum order of an element of $\coprod\!\!\!\coprod$$(E/K)^{\epsilon }_{l^\infty}$. From Proposition 5.2.1 and the isomorphism of (5.4.4), there is an element in the
Selmer group ${\rm Sel}_{l^\infty}(E/K)^{\epsilon }$, and hence in the Tate-Shafarevich group $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}^{\epsilon }$, of order $l^{M_0-M_1}$. It follows that we have the inequality $$M_0-M_1\leq N_1.\leqno{(5.4.11)}$$
On the other hand, the Cassels pairing gives, as $l^n
\delta_{M_0}(z_1)=\delta_{M_0-n}(z_1)$ if $n\leq M_0$, $$<\delta_{M_0}(z_1), l^nd>_{\rm Cassels}=
<l^n\delta_{M_0}(z_1), d>_{\rm Cassels}\leqno{(5.4.12)}$$ $$=<\delta_{M_0-n}(z_1), d>_{\rm Cassels}$$ and where we have $$l^{N_i}\delta_{M_0+N_i-n}(z_1)=\delta_{M_0-n}(z_1).$$
By the construction of the Cassels pairing as a sum of local pairings, more precisely from Proposition 4.6.1 and equation (4.6.2), we obtain from this last equation (5.4.12) that for all $0\leq n\leq N_i-1$ where $l^{N_i}$ is the order of $d_i$ $$<\delta_{M_0}(z_1), l^nd_i>_{\rm Cassels}=\sum_{v\in \Sigma_K}
\ [\delta_{M_0+N_i-n}(z_1)_v,w_{i,v}]_{v}\leqno{(5.4.13)}$$ $$=[\delta_{M_0+N_i-n}(z_1)_{y_1} ,w_{i,y_1} ]_{y_1}$$ where $e_{i,v} =\partial_{l^{N_i}}(w_{i,v} )$ as in (5.4.6) for all $v\in \Sigma_K$, as we have that $\delta_{M_0+N_i-n}(z_1)_v=0$ for all places $v$ of $K$ not dividing $z_1$ (by Proposition 4.5.1(i)).
The term $$[\delta_{M_0+N_i-n}(z_1)_{y_1} ,w_{i,y_1} ]_{y_1}$$ is zero if $i\geq 2$ by (5.4.10). Hence we have from (5.4.13) $$<\delta_{M_0}(z_1), d_i>_{\rm Cassels}=0{\rm \ \ for \ } i\geq 2.\leqno{(5.4.14)}$$
Let $i=1$. From the sum formula (5.4.13) we have $$<\delta_{M_0}(z_1), l^nd_1>_{\rm Cassels}=[\delta_{M_0+N_1-n}(z_1)_{y_1} ,w_{1,y_1} ]_{y_1}.\leqno{(5.4.15)}$$
By (5.4.9), the term $[\delta_{M_0+N_i}(z_1)_{y_1} ,w_{1,y_1} ]_{y_1}$ has order $l^{N_1}$. Hence by (5.4.15), the element $<\delta_{M_0}(z_1), d_1>_{\rm Cassels}$ of ${\mathds Q}/{\mathds Z}$ has order $l^{N_1}$. Hence the character $$d\mapsto <\delta_{M_0}(z_1), l^nd>_{\rm Cassels},\ \ \
\coprod\!\!\!\!\coprod(E/K)_{l^\infty}^{\epsilon }\to {\mathds Q}/{\mathds Z},$$ is non-zero for all $n$ such that $0\leq n\leq N_1-1$ and more precisely $$<\delta_{M_0}(z_1), l^nd_1>_{\rm Cassels}$$ is non-zero for all $n$ such that $0\leq n\leq N_1-1$.
Therefore the character $$d\mapsto <\delta_{M_0}(z_1),d>_{\rm Cassels},\ \ \
\coprod\!\!\!\!\coprod(E/K)_{l^\infty}^{\epsilon }\to {\mathds Q}/{\mathds Z},$$ vanishes on $\prod_{i\geq 2}D_i$ (by (5.4.14)) and its restriction to $D_1$ generates the dual ${\widehat D}_1$. Hence the element $\delta_{M_0}(z_1)$ of $\coprod\!\!\!\coprod$$(E/K)^{\epsilon
}_{l^\infty}$ has order at least $l^{N_1}$ as this is the order of the cyclic group $D_1$. Since $\delta_{M_0}(z_1)$ has order at most $l^{M_0-M_1}$, by definition of the cohomology class $\delta_{M_0}(z_1)$ (see Lemma 5.1.7(iii)), we obtain that $$N_1\leq M_0-M_1.$$
Hence we must have from the inequality (5.4.11) $$N_1=M_0-M_1.\leqno{(5.4.16)}$$ It follows from this equality that $\delta_{M_0}(z_1)$ has order $l^{M_0-M_1}$ and therefore $l^{M_1+1}$ does not divide $P_{z_1}$ and hence we have $$l^{M_1}\vert\vert P_{z_1}.$$
In summary, we have shown that $\delta_{M_0}(z_1)$ has order $l^{N_1}=l^{M_0-M_1}$, $l^{M_1}\vert\vert P_{z_1}$, $S_{y_1}=0$, and the character $d\mapsto <\delta_{M_0}(z_1),d>_{\rm Cassels}$ vanishes on $\prod_{i\geq 2}D_i$ and its restriction to $D_1$ generates the dual ${\widehat D}_1$.
We now proceed by induction. Suppose there are prime divisors $z_1,\ldots,z_k\in \Lambda^1(M_0+N_1)$ such that $$e_{i,y_j}=0, \ {\rm \ for \ all \ \ }i\not=j,{\rm \ and \ for \ all \ } j{\rm
\ such \ that \ } 1\leq j\leq k\leqno{(5.4.17)}$$ and if $$c_j=z_1+\ldots+ z_j$$ then $$l^{M_j}\vert\vert P_{c_j},\ \delta_{M_{j-1}}(c_j)\in
{\coprod\! \!\!\!\coprod}(E/K)^{-\nu(j)}
_{l^\infty}, \ {\rm \ for \ all \ } 1\leq j\leq k\leqno{(5.4.18)}$$ and the characters $$\chi_j: d\mapsto <\delta_{M_{j-1}}(c_j)
,d>_{\rm Cassels}, \ 1\leq j\leq k$$ vanish on $\prod_{i\geq k+1}D_{i}
$ and form a triangular basis of the dual of $\prod_{i\leq k} D_i$ such that the restriction of $\chi_j$ to the cyclic subgroup $D_j$ is a basis for the dual ${\widehat D}_j$ for all $j=1,\ldots, k$. Suppose further we have shown that $$M_{j-1}-M_j=N_j, {\rm \ for \ } 1\leq j\leq k.\leqno{(5.4.19)}$$ and $${\rm ord}\ \delta_{M_{j-1}}(c_j)= l^{N_j}{\rm \ for \ } 1\leq j\leq k.\leqno{(5.4.20)}$$
We have already proved the existence of the divisor $c_1=z_1$ and these properties of the previous paragraph of $e_{i,y_1}$, $P_{c_1}$, $\delta_{M_0}(c_1)$, $\chi_1$, $N_1=M_0-M_1$, including (5.4.17), (5.4.18), (5.4.19), and (5.4.20) for $k=1$. Let $y_i\in \Sigma_K$ be the place of $K$ above $z_i$ for all $i=1,\ldots,k$.
Let $m$ be the integer already selected such that $l^m\geq $ exp$({\coprod\! \!\!\coprod}(E/K)
_{l^\infty})$. The order of $\delta_{M_{k-1}}(c_k)$ in $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}$ is the same as its order as a character on $D$ via the non-degenerate Cassels pairing. Since $D$ is an isotropic subgroup of ${\coprod\! \!\!\coprod}(E/K)
_{l^\infty}$, it follows that $${\mathds Z}\delta_{M_{k-1}}(c_k)\cap D=0.\leqno{(5.4.21)}$$ We have $$\gamma_{M_k+N_{k+1}}(c_k)=l^{N_k-N_{k+1}}\gamma_{M_{k-1}}(c_k)$$ as $N_k=M_{k-1}-M_k$ from (5.4.19) and where $N_k-N_{k+1}\geq 0$ by the definition of the integers $N_i$ as the invariants of $\coprod\!\!\!\coprod$$(E/K)^{
}_{l^\infty}$ in decreasing order. It follows from the induction hypothesis that the cohomology class $\delta_{M_k+N_{k+1}}(c_k)$ has order $l^{N_{k+1}}$, by (5.4.20), and belongs to $\coprod\!\!\!\coprod$$(E/K)^{-\nu(k)
}_{l^\infty}$.
We have $l^{M_k}\vert\vert P_{c_k}$ by the induction hypothesis (5.4.18). Hence by Lemma 3.4.14(i), the cohomology class $\gamma_{M_k+N_{k+1}}(c_k)$ then has the same order as its homomorphic image $\delta_{M_k+N_{k+1}}(c_k)$ namely $l^{N_{k+1}}$.
Let $S$ be the subgroup of the Selmer group Sel$_{l^m}(E/K)$ generated by the elements $e_i$ for all $i\not=k+1$. Let $T$ be the subgroup of the Selmer group Sel$_{l^m}(E/K)$ generated by $\gamma_{M_k+N_{k+1}}(c_k), e_{k+1}, S$. From the isomorphism (5.4.4) and that ${\mathds Z}\delta_{M_k+N_{k+1}}(c_k)$ has trivial intersection with $D$ by (5.4.21), there is an equality of subgroups of the Selmer group ${\rm Sel}_{l^m}(E/K)$, where the sums on the right hand side are direct, $$T= {\mathds Z} \gamma_{M_k+N_{k+1}}(c_k)\oplus {\mathds Z}e_{k+1}\oplus S.$$
We may now apply Lemma 5.4.3 to $S$ and the elements $e_{k+1}, \gamma_{M_k+N_{k+1}}(c_k)$ where we take $$c=c_k, r=k, s=M_k+N_{k+1}, t=M_0+N_1- s, n=N_{k+1}.\leqno{(5.4.22)}$$ There is then according to the lemma a prime divisor $z_{k+1}\in \Lambda^1(M_0+N_1)$ which satisfies the following 2 conditions $${\rm ord} [\delta_{M_k+N_{k+1}}(c_k+z_{k+1}),w_{{k+1},y_{k+1}}]_{y_{k+1}} =l^{N_{k+1}}\leqno{(5.4.23)}$$ and $$S_{y_{k+1}}=0\leqno{(5.4.24)}$$ where $y_{k+1}\in \Sigma_K$ is the prime divisor of $K$ lying over $z_{k+1}$ and where the subscript $y_{k+1}$ denotes localization at $y_{k+1}$, and where the point $w_{{k+1},y_{k+1}}$ in $(E(K_{y_{k+1}})/l^{N_{k+1}}E(K_{y_{k+1}}))^{\epsilon }$ is such that $\partial_{l^{N_{k+1}}} (w_{{k+1},y_{k+1}})=e_{{k+1},y_{k+1}}$.
For this selection of $z_{k+1}\in \Lambda^1(M_0+N_{1})$, note that $M_k+N_{k+1}\leq M_0+N_1$ and so $t\geq 0$, where $t$ is the parameter of (5.4.22), because $M_r,N_r$ are both decreasing sequences of integers and that $H^1(K, E_{l^{M_k+N_{k+1}}})$ is a subgroup of $H^1(K, E_{l^{M_0+N_{1}}})$ by Lemma 3.1.5.
Let $c_{k+1}$ be the divisor which is the sum of the $z_i$, for $i=1,\ldots, k+1$, $$c_{k+1}=\sum_{j=1}^{k+1} z_{j}.\leqno{(5.4.25)}$$ Let $$\delta_{M_k}(c_{k+1})\in {\coprod\! \!\!\!\coprod}(E/K)^{-\nu(k+1)
}_{l^\infty}$$ be the cohomology class associated to this divisor $c_{k+1}\in \Lambda^{k+1}(M_0+N_1)$; that $\delta_{M_k}(c_{k+1})$ belongs to the Tate-Shafarevich group $\coprod\!\!\!\coprod$$(E/K)^{-\nu(k+1)
}_{l^\infty}$ follows from Lemma 5.1.7(i).
Then for $0\leq n\leq N_{k+1}-1$ by the construction of the Cassels pairing as a sum of local terms, more precisely from Proposition 4.6.1 and equation (4.6.2), we have the following sum formulae as $d_i$ has order $l^{N_i}$ $$<\delta_{M_k}(c_{k+1}), l^nd_i>_{\rm Cassels}=<\delta_{M_k-n}(c_{k+1}), d_i>_{\rm Cassels}\leqno{(5.4.26)}$$ $$=\sum_{v\in \Sigma_K}[\delta_{M_k-n+N_i}(c_{k+1})_{v}, w_{i,v}]_{v}
=\sum_{j=1}^{k+1}
[\delta_{M_k-n+N_i}(c_{k+1})_{y_j}, w_{i,y_j}]_{y_j}.$$ because $\delta_{M_k-n+N_i}(c_{k+1})_{v}=0$ for all $v\in \Sigma_K$ not dividing an element of Supp$(c_{k+1})$ by Proposition 4.5.1(i). Here $ w_{i,y_j}\in {}_{l^{N_i}}E(K_{y_j})$ is an element already chosen (see (5.4.6)) such that $$\partial_{l^{N_i}}( w_{i,y_j})=e_{i,y_j}.$$
We have the following sum for the Cassels pairing, obtained from those of (5.4.26), $$<\delta_{M_k}(c_{k+1}), l^nd_i>_{\rm Cassels}=\sum_{j=1}^{k+1}
[\delta_{M_k-n+N_i}(c_{k+1})_{y_j}, w_{i,y_j}]_{y_j}.\leqno{(5.4.27)}$$ We have $w_{i,y_j}=0$ for all $i\not=j$ and all $j$ such that $1\leq j\leq k+1$ by (5.4.17), for $j\leq k$, and by (5.4.24), for $j=k+1$. It follows that for $i\geq k+1$ all terms $[\delta_{M_k-n+N_i}(c_{k+1})_{y_j}, w_{i,y_j}]_{y_j}$ of this sum (5.4.27) are zero except the last $[\delta_{M_k-n+N_i}(c_{k+1})_{y_{k+1}}, w_{i,y_{k+1}}]_{y_{k+1}}$ and the entire sum is zero for $i>k+1$.
By (5.4.23) the local term $$[\delta_{M_k+N_{k+1}-n}(c_{k+1})_{y_{k+1}}, w_{k+1,y_{k+1}}]_{y_{k+1}}$$ is non-zero for all integers $n$ such that $0\leq n\leq N_{k+1}-1.$
Therefore the character, by (5.4.27), $$\chi_{k+1}: \ d\mapsto <\delta_{M_k}(c_{k+1}),d>_{\rm Cassels}$$ vanishes on $\prod_{i> k+1}D_{i}$ and its restriction to $D_{k+1}$ generates ${\widehat D}_{k+1}$, as $D_{k+1}$ has order $l^{N_{k+1}}$ by definition. Hence $\chi_{k+1}$ extends the triangular basis $\chi_1,\ldots, \chi_k$ to generate $\prod_{i\leq k+1}{\widehat D}_i$ and $\delta_{M_k}(c_{k+1})$ has order at least $l^{N_{k+1}}$. Since it has order at most $l^{M_k-M_{k+1}}$, by definition of the cohomology class $\delta_{M_k}(c_{k+1})$ (see Lemma 5.1.7(iii)), we conclude that $$N_{k+1}\leq M_k-M_{k+1}\leqno{(5.4.28)}$$ and also $$l^{N_{k+1}}\leq \ {\rm ord}\ \delta_{M_k}(c_{k+1})\ \leq l^{M_k-M_{k+1}}.\leqno{(5.4.29)}$$
Let $C_k$ be the subgroup of the Selmer eigencomponent Sel$_{l^m}(E/K)^{-\nu({k+1})}$ given by $$C_k=\bigg(\gamma_{M_0+N_1}(0),e_1,\ldots, e_k, \gamma_{M_0}(c_1),\ldots, \gamma_{M_{k-1}}(c_k)\bigg)^{-\nu({k+1})}.$$ If $k$ is even then $C_k$ is generated by $e_1, \gamma_{M_0}(c_1), e_3, \gamma_{M_2}(c_3), \ldots , e_{k-1}, \gamma_{M_{k-2}}(c_{k-1})$ of which there are $k$ in number. If $k$ is odd then $C_k$ is generated by $\gamma_{M_0+N_1}(0),e_2, \gamma_{M_1}(c_2), e_4, \gamma_{M_3}(c_4), \ldots , e_{k-1}, \gamma_{M_{k-2}}(c_{k-1})$ of which there are $k$ in number. We then have for all integers $k$ $${\rm rank}(C_k)\leq k.\leqno{(5.4.30)}$$
The elements $e_1,\ldots, e_k$ generate a subgroup of Sel$_{l^m}(E/K)$ isomorphic to $\prod_{i\leq k} D_i$ by the decomposition (5.4.4) of the Selmer group. Furthermore, the elements $\gamma_{M_0}(c_1),\ldots, \gamma_{M_{k-1}}(c_k)$ generate a subgroup of Sel$_{l^m}(E/K)$ isomorphic to the dual of $\prod_{i\leq k} D_i$ as $\gamma_{M_{i-1}}(c_i)$ has the same order as $\delta_{M_{i-1}}(c_i)$ for all $i=1,\ldots, k$ (see (5.4.18), (5.4.19), (5.4.20), and Proposition 3.4.14(i)). In the decomposition (5.4.4) of Sel$_{l^m}(E/K)$, take $S$ to be the subgroup ${}_{l^m}E(K) \oplus_{1\leq i\leq k} \prod D_i\oplus \bigoplus_{1\leq i\leq k} {\mathds Z} \gamma_{M_{i-1}}(c_i) $. Then we have $C_k=S^{-\nu({k+1})}$.
If a finite abelian group $G$ is a direct product $G_1\times
G_2$ of 2 subgroups and $g\in G$ is such that ${\mathds Z}g\cap G_1=0$ then the order of the element $g$ is at most the exponent of $G_2$. Then by the previous remark where we take $G={\rm Sel}_{l^m}(E/K)^{-\nu({k+1})} $ and $G_1=C_k$, we have that $l^{N_{k+1}}$ is the maximum order of an element $c\in {\rm Sel}_{l^m}(E/K)^{-\nu({k+1})} $ if $${\mathds Z}c\cap C_k=0$$ by the decomposition (5.4.4) of the Selmer group ${\rm Sel}_{l^m}(E/K)$.
On the other hand, by Proposition 5.2.1 and (5.4.30) applied to subgroup $C_k$ of the Selmer group there is an element in ${\rm Sel}_{l^m}(E/K)^{-\nu({k+1})}$ of order $l^{M_k-M_{k+1}}$ satisfying $
{\mathds Z}c\cap C_k=0$. Hence we have $$M_k-M_{k+1}\leq N_{k+1}$$ and so by (5.4.28) we have $$N_{k+1}=M_k-M_{k+1}.\leqno{(5.4.31)}$$ It follows from this equality and (5.4.29) that $${\rm ord}\ \delta_{M_k}(c_{k+1})= l^{M_k-M_{k+1}}.\leqno{(5.4.32)}$$Therefore $l^{M_{k+1}+1}$ does not divide $P_{c_{k+1}}$ and $l^{M_{k+1}}\vert P_{c_{k+1}}$ and hence we have $$l^{M_{k+1}}\vert\vert P_{c_{k+1}}.\leqno{(5.4.33)}$$ The properties (5.4.31), (5.4.32), and (5.4.33) complete the proof of the induction step and this proves the Theorems 4.1.9 and 4.1.13. ${\sqcap \!\!\!\!\sqcup}$
0.4in
[**5.5. Proof of Theorems 1.1.1. and 4.1.14** ]{}
0.2in
[*Proof of Theorem 1.1.1.*]{} From Proposition 2.2.5, we have as $l\in
{\cal P}$ is an odd prime number that $
{\coprod\!\!\!\coprod}(E/F)_{l^\infty}\cong{\coprod\!\!\!\coprod}(E/K)_{l^\infty}^{+1}
$. The Theorem now follows immediately from Theorem 4.1.9. ${\sqcap \!\!\!\!\sqcup}$
0.2in [*Proof of Theorem 4.1.14.*]{} As $P_0$ has infinite order, by Lemma 5.1.2 and \[1, Chap. 7, Theorem 7.6.5\] we have that $M_0$ is finite, the group ${\mathds Z}P_0$ has finite index in $E(K)$, and the highest power of $l$ dividing the index $[E(K):{\mathds Z}P_0]$ is equal to $$l^{M_0}={\big\vert} (E(K)/{\mathds Z}P_0)_{l^\infty}{\big\vert}.$$
By Theorem 4.1.9, we have that $${\bigg\vert } {\coprod\!\!\!\!\coprod}(E/K)_{l^\infty}{\bigg\vert} =\prod_{i\geq 0}l^{2(M_i-M_{i+1})}= l^{2(M_0-M_\infty)}$$ where $$M_\infty=\min_{i\in {\bf N}} M_i$$ and where this minimum exists as the $M_i$ form a decreasing sequence of non-negative integers by Lemma 5.1.4. ${\sqcap \!\!\!\!\sqcup}$
0.4in [**5.6. Generators of Tate-Shafarevich groups**]{} 0.2in
0.2in(5.6.1) The notation (5.1.1) of the section §5.1 holds also for this section. We further denote by 0.2in $\Sigma_K$ the set of all places of the global field $K$;
$[,]_v: {}_{l^m}E(K_v)\times H^1(K_v,E)_{l^m}\to {\mathds Z}/l^m{\mathds Z}$ the non-degenerate local pairing at
the place $v$ of $K$ induced by the cup product, as in Theorem 2.1.6.
0.2in [**5.6.2. Theorem.**]{} [*Let $l$ be a prime number belonging to $\cal P$; assume that $l$ is coprime to the order of ${\rm Pic}(A)$. Suppose that $P_0$ has infinite order in $E(K)$ and let $a$ be an integer such that $a\geq 2M_0$. Then we have*]{}
\(1) [*the classes $ \delta_{M_0}(c)$, for all $c\in \Lambda^1(a) $, generate $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}^{\epsilon }$;*]{}
\(2) [*the classes $ \delta_{M_1}(c)$, for all $c\in \Lambda^2(a) $, generate $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}^{-\epsilon }$.*]{}
0.2in [**5.6.3. Theorem.**]{} [*Under the hypotheses of Theorem 5.6.2, the classes $ \delta_{M_r }(c)$, for all $c\in \Lambda^r(a) $, generate $\coprod\!\!\!\coprod$$(E/F)_{l^\infty}$ where $r=(3-\epsilon)/2$.*]{}
0.2in
[*Proof of Theorem 5.6.3.*]{} This evidently follows from Theorem 5.6.2 and Proposition 2.2.5. ${\sqcap \!\!\!\!\sqcup}$
0.2in [*Proof of Theorem 5.6.2.*]{} By Theorem 4.1.9 above or \[1, Chap. 7, Theorem 7.6.5\], the group $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}$ is finite. By Lemmas 5.1.2 and 5.1.4, the quantities $M_0,M_1,M_2,\ldots $ are all finite and form a decreasing sequence of non-negative integers. We fix an integer $a\geq 2M_0$. The classes $ \delta_{M_0}(z)$, for $z\in \Lambda^1(a) $, and the classes $ \delta_{M_1}(z_1+z_2)$, for $z_1+z_2\in \Lambda^2(a) $, belong to $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}$ by Lemma 5.1.7(i). We now prove separately the two parts of the theorem.
0.2in (1) Suppose that $d\in $ $\coprod\!\!\!\coprod$$(E/K)^{\epsilon }_{l^\infty}$ has order exactly $l^M$ for some $M>0$ and is in the $\epsilon$-eigenspace of $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}$. By Theorem 4.1.15 there is an isomorphism of $\epsilon$-components $${\rm Sel}_{l^\infty}(E/K)^\epsilon\cong
{\coprod\! \!\!\!\coprod}(E/K)_{l^\infty}^\epsilon\leqno{(5.6.4)}$$ induced from the natural surjection of the Selmer group onto the Tate-Shafarevich group.
By Theorem 4.1.9, we have $M\leq M_0$. By the isomorphism (5.6.4), we may lift $d$ to an element of the Selmer group $e\in {\rm
Sel}_{l^\infty}(E/K)^\epsilon $ of order $l^M$. The cohomology class $\gamma_{M_0+M}(0)$ belongs to the $-\epsilon$-eigenspace $ {\rm
Sel}_{l^\infty}(E/K)^{-\epsilon} $ of the Selmer group and has order $l^M$ by Lemma 5.1.2.
We may apply Lemma 5.4.3 to the elements $e$, $\gamma_{M_0+M}(0)$ and the subgroup $S=0$ where we take the parameters of the lemma to be $$c=0, r=0, s=M_0+M, t= a-(M_0+M), n=M .\leqno{(5.6.5)}$$ Note that $t\geq 0$ as $a\geq 2M_0\geq M_0+M$. There is according to the lemma a prime divisor $z\in \Lambda^1(a)$ which satisfies the following condition $${\rm ord}[ \delta_{M_0+M}(z)_y, w_y]_y=l^M\leqno{(5.6.6)}$$ where $y\in \Sigma_K$ is the prime divisor of $K$ lying over $z$ and where the subscript $y$ denotes localization at $y$, and where the point $w_{y}$ in $(E(K_{y})/l^{M}E(K_{y}))^{\epsilon }$ is such that $\partial_{l^{M}} (w_{y})=e_{y}$. Here in (5.6.6), $\delta_{M_0+M}(z)$ is the cohomology class in $H^1(K,E)_{l^{M_0+M}}$ associated to $z$.
Let $\delta_{M_0}(z)$ be the cohomology class of $\coprod\!\!\!\coprod$$(E/K)^{\epsilon }_{l^\infty}$ associated to $z$, where this class belongs to the Tate-Shafarevich group by Lemma 5.1.7.
The Cassels pairing gives, as $l^u\delta_{M_0}(z)=\delta_{M_0-u}(z)$ if $u\leq M_0$, $$<\delta_{M_0}(z), l^ud>_{\rm Cassels}=<\delta_{M_0-u}(z), d>_{\rm Cassels}. \leqno{(5.6.7)}$$ As $d$ has order $l^{M}$ and that we have $$l^{M}\delta_{M_0+M-u}(z)=\delta_{M_0-u}(z),$$ by the construction of the Cassels pairing in terms of local Tate pairings (Proposition 4.6.1 and equation (4.6.2)) we obtain from this last equation (5.6.7) $$<\delta_{M_0}(z), l^ud>_{\rm Cassels}=[\delta_{M_0+M-u}(z)_{y} ,w_{y} ]_{y}$$ where as above $$e_{y} =\partial_{l^M}(w_{y} ),$$ and where $w_y\in {}_{l^M}E(K_y)$ as we have that $\delta_{M_0+M-u}(z)_v=0$ for all places $v$ of $K$ not dividing $z$ by Proposition 4.5.1(i).
By (5.6.6) the element given by the Tate pairing $$[\delta_{M_0+M-u}(z)_{y} ,w_{y} ]_{y}$$ of ${\mathds Q}/{\mathds Z}$ is non-zero for all integers $u$ such that $1\leq u\leq M-1$ and hence $$<\delta_{M_0}(z), l^{u}d>_{\rm Cassels}$$ is non-zero for all integers $u$ such that $1\leq u\leq M-1$. We obtain that the character, where $ z\in \Lambda^1(a)$, $$f\mapsto < \delta_{M_0}(z), f>_{\rm Cassels},
\ \ \coprod\!\!\!\!\coprod(E/K)_{l^\infty}\to {\mathds Q}/{\mathds Z},$$ of $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}$ when restricted to ${\mathds Z}d$ generates the dual ${\widehat{ {\mathds Z}d}}$ of this subgroup ${ {\mathds Z}d}$.
As $d$ is any element of the abelian group $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}^{\epsilon }$, the non-degeneracy of Cassels pairing implies that the classes $\{ \delta_{M_0}(z),z\in \Lambda^1(a)\} $ generate the $\epsilon$-eigenspace $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}^{\epsilon }$ which proves the part of the theorem for $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}^{\epsilon }$.
0.2in (2) Suppose that $f\in $ $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}^{-\epsilon }$ has order exactly $l^{M'}$ and is in the $-\epsilon$-eigenspace of $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}$. We have $M'\leq M_0$ by Theorem 4.1.9.
We have from Theorem 4.1.15, an isomorphism compatible with the $\tau$-eigenspaces $${\rm Sel}_{l^m}(E/K) \cong
{}_{l^m}E(K)\oplus
{\coprod\! \!\!\!\coprod}(E/K)
_{l^\infty} {\rm \ \
for \ all \ } m\geq N_1$$ from which as stated in this theorem we obtain the isomorphism, taking $m=a\geq 2M_0\geq N_1$, $$\Delta_a :{\rm Sel}_{l^a}(E/K)^{-\epsilon} \cong
{\mathds Z}/l^a{\mathds Z} \oplus
{\coprod\! \!\!\!\coprod}(E/K)
_{l^\infty}^{-\epsilon} .$$ The projection of ${\rm Sel}_{l^a}(E/K)$ onto the second factor ${\coprod\! \!\!\coprod}(E/K)
_{l^\infty}$ of this direct sum is the natural surjection of the Selmer group onto the Tate-Shafarevich group.
Lift $f$ to $
g\in {\rm Sel}_{l^a}(E/K)^{-\epsilon}$ via this isomorphism $\Delta_a$ for the integer $a$ and where $g$ has order $l^{M'}$ and has zero component in the first term ${\mathds Z}/l^a{\mathds Z}$ of this decomposition of the Selmer group.
By Theorem 4.1.9 or alternatively Proposition 5.2.1, there is an element $d\in $ $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}^{\epsilon }$ of order exactly $
l^{M_0-M_1}$. Lift, via the isomorphism (5.6.4), $d$ to an element of the Selmer group $e\in {\rm Sel}_{l^{a}}(E/K)^\epsilon$ which is also of order $l^{M_0-M_1}$.
The cohomology class $\gamma_{2M_0-M_1}(0)$ belongs to the $-\epsilon$-eigenspace $ {\rm
Sel}_{l^a}(E/K)^{-\epsilon} $ of the Selmer group and has order $l^{M_0-M_1}$ by Lemma 5.1.2.
Let $T$ be the subgroup of $ {\rm
Sel}_{l^a}(E/K) $ generated by the three elements $\gamma_{2M_0-M_1}(0),
e$ and $g$. As the two elements $\gamma_{2M_0-M_1}(0), g$ belong to the different components of $ {\rm
Sel}_{l^a}(E/K)^{-\epsilon} $ under the isomorphism $\Delta_a$ and as $e$ belongs to a different eigenspace $ {\rm
Sel}_{l^a}(E/K)^{\epsilon} $, the group $T$ is the direct sum of the subgroups generated by these three elements, that is to say we have $$T\cong {\mathds Z}\gamma_{2M_0-M_1}(0)\oplus {\mathds Z}e\oplus
{\mathds Z}g.$$
We may apply Lemma 5.4.3 to the subgroup $S={\mathds Z}g$ and the elements $\gamma_{2M_0-M_1}(0)$ and $e$ of the Selmer group; we take the parameters of the lemma to be $$c=0,r=0, s=2M_0-M_1, t=a-(2M_0-M_1), n=M_0-M_1.\leqno{(5.6.8)}$$ Note that $t\geq 0$ as $a\geq 2M_0$. There is according to the lemma a prime divisor $z_1\in \Lambda^1(a)$ which satisfies the following 2 conditions $${\rm ord }[\delta_{2M_0-M_1}(z_1)_{y_1} ,w_{y_1} ]_{y_1} =l^{M_0-M_1}\leqno{(5.6.9)}$$ $$({\mathds Z} g)_{y_1}=0\leqno{(5.6.10)}$$ where $y_1\in \Sigma_K$ is the prime divisor of $K$ lying over $z_1$, and where the point $w_{y_1}$ in $(E(K_{y_1})/l^{M_0-M_1}E(K_{y_1}))^{\epsilon }$ is such that $\partial_{l^{M_0-M_1}} (w_{y_1})=e_{y_1}$ . Here in (5.6.9), $\delta_{2M_0-M_1}(z_1)$ is the cohomology class in $H^1(K,E)_{l^{2M_0-M_1}}$ associated to $z_1$.
Let $$\delta_{M_0}(z_1)\in {\coprod\! \!\!\!\coprod}(E/K)^{\epsilon
}_{l^\infty}$$ be the cohomology class associated to this prime divisor $z_1\in \Lambda^1(a)$; that $\delta_{M_0}(z_1)$ belongs to $\coprod\!\!\!\coprod$$(E/K)^{\epsilon
}_{l^\infty}$ follows from Lemma 5.1.7(i).
The Cassels pairing gives, as $l^u
\delta_{M_0}(z_1)=\delta_{M_0-u}(z_1)$ if $u\leq M_0$, $$<\delta_{M_0}(z_1), l^ud>_{\rm Cassels}=
<l^u\delta_{M_0}(z_1), d>_{\rm Cassels}\leqno{(5.6.11)}$$ $$=<\delta_{M_0-u}(z_1), d>_{\rm Cassels}$$ where $$l^{M_0-M_1}\delta_{2M_0-M_1-u}(z_1)=\delta_{M_0-u}(z_1).$$
By the construction of the Cassels pairing as a sum of local terms, more precisely from Proposition 4.6.1 and equation (4.6.2), we obtain from this last equation (5.6.11) that for all $0\leq u\leq M_0-M_1-1$ where $M_0-M_1$ is the order of $d$ $$<\delta_{M_0}(z_1), l^ud>_{\rm Cassels}=\sum_{v\in \Sigma_K}
\ [\delta_{2M_0-M_1-u}(z_1)_v,w_{v}]_{v}\leqno{(5.6.12)}$$ $$=[\delta_{2M_0-M_1-u}(z_1)_{y_1} ,w_{y_1} ]_{y_1}$$ where $e_{v} =\partial_{l^{M_0-M_1}}(w_{v} )$ for all $v\in \Sigma_K$, and where we have that $\delta_{2M_0-M_1-u}(z_1)_v=0$ for all places $v$ of $K$ not dividing $z_1$ (by Proposition 4.5.1(i)).
From the sum formula (5.6.12) we have $$<\delta_{M_0}(z_1), l^ud>_{\rm Cassels}=[\delta_{2M_0-M_1-u}(z_1)_{y_1} ,w_{y_1} ]_{y_1}.$$ By (5.6.9), $[\delta_{2M_0-M_1}(z_1)_{y_1} ,w_{y_1} ]_{y_1} $ has order $l^{M_0-M_1}$.
Hence the map $$s\mapsto <\delta_{M_0}(z_1), l^us>_{\rm Cassels},\ \ \
\coprod\!\!\!\!\coprod(E/K)_{l^\infty}^{\epsilon }\to {\mathds Q}/{\mathds Z},$$ is non-zero for all $u$ such that $0\leq u\leq M_0-M_1-1$ and more precisely $$<\delta_{M_0}(z_1), l^ud>_{\rm Cassels}$$ is non-zero for all $u$ such that $0\leq u\leq M_0-M_1-1$.
Therefore the character $$s\mapsto <\delta_{M_0}(z_1),s>_{\rm Cassels},\ \ \
\coprod\!\!\!\!\coprod(E/K)_{l^\infty}^{\epsilon }\to {\mathds Q}/{\mathds Z},$$ is such that its restriction to ${\mathds Z}d$ generates the dual ${\widehat {{\mathds Z}d}}$. Hence $\delta_{M_0}(z_1)$ has order at least $l^{M_0-M_1}$ as this is the order of the cyclic group ${\mathds Z}d$. Since $\delta_{M_0}(z_1)$ has order at most $l^{M_0-M_1}$ , by definition of the cohomology class $\delta_{M_0}(z_1)$ (see Lemma 5.1.7(iii)), we obtain that $\delta_{M_0}(z_1)$ has order given by $${\rm ord}\ \delta_{M_0}(z_1)= l^{M_0-M_1}.$$
We have $l^{M_1}\vert P_{z_1}$ by definition of $M_1$. It follows that $\gamma_{M_0}(z_1)$ has order $\leq l^{M_0-M_1}$ by Proposition 3.4.14(i). As $\delta_{M_0}(z_1)$ has order $l^{M_0-M_1}$ by the previous paragraph and $\delta_{M_0}(z_1)$ is a homomorphic image of $\gamma_{M_0}(z_1)$ we must have that $\gamma_{M_0}(z_1)$ has order exactly given by $${\rm ord}\ \gamma_{M_0}(z_1)=l^{M_0-M_1}.$$ Therefore we have by Proposition 3.4.14(i) $$l^{M_1}\vert\vert P_{z_1}.$$
We may apply Lemma 5.4.3 to the subgroup $S=0$ and the two elements $\gamma_{M_1+M'}(z_1)$ and $g$ of the Selmer group which both have order $l^{M'}$; we take the parameters of the lemma to be $$c=z_1,r=1, s=M_1+M', t=a-(M_1+M'), n=M'.\leqno{(5.6.13)}$$ Note that $t\geq 0$ as $a\geq 2M_0$ and $M', M_1\leq M_0$. There is according to the lemma a prime divisor $z_2\in \Lambda^1(a)$ which satisfies the following condition $${\rm ord }[\delta_{M_1+M'}(z_1+z_2)_{y_2} ,x_{y_2} ]_{y_2} =l^{M'}\leqno{(5.6.14)}$$ where $y_2\in \Sigma_K$ is the prime divisor of $K$ lying over $z_2$, and where the point $x_{y_2}$ in $(E(K_{y_2})/l^{M'}E(K_{y_2}))^{\epsilon }$ is such that $\partial_{l^{M'}} (x_{y_2})=e_{y_2}$. Here in (5.6.14), $\delta_{M_1+M'}(z_1+z_2)$ is the cohomology class in $H^1(K,E)_{l^{M_1+M'}}$ associated to $z_1+z_2$.
Let $$\delta_{M_1}(z_1+z_2)$$ be the cohomology class of $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}^{-\epsilon}$ associated to the divisor $z_1+z_2\in \Lambda^2(a)$, where this class belongs to the Tate-Shafarevich group by Lemma 5.1.7(i).
Then the Cassels pairing gives, as $l^u\delta_{M_1}(z_1+z_2)=\delta_{M_1-u}(z_1+z_2)$ if $u\leq M_1$, $$<\delta_{M_1}(z_1+z_2), l^uf>_{\rm Cassels}=
<\delta_{M_1-u}(z_1+z_2), f>_{\rm Cassels}. \leqno{(5.6.15)}$$ As $f$ has order $l^{M'}$ and that we have $$l^{M'}\delta_{M_1+M'-u}(z_1+z_2)=\delta_{M_1-u}(z_1+z_2),$$ by the construction of the Cassels pairing (Proposition 4.6.1 and equation (4.6.2)) we obtain from this last equation (5.6.15) $$<\delta_{M_1}(z_1+z_2), l^uf>_{\rm Cassels}=\sum_{z\in {\rm \ Supp}(z_1+z_2)} [\delta_{M_1+M'-u}(z_1+z_2)_{y} ,x_{y} ]_{y} \leqno{(5.6.16)}$$ where $$g_{y} =\partial_{l^{M'}}(x_{y} ),\leqno{(5.6.17)}$$ and where $$x_y\in ({}_{l^{M'}}E(K_y))^{-\epsilon}$$ and because we have that $\delta_{M_1+M-u}(z_1+z_2)_v=0$ for all places $v$ of $K$ not dividing $z_1+z_2$ by Proposition 4.5.1(i). Here in (5.6.16) and (5.6.17), $z$ runs over the places $z_1,z_2$ and $y\in \Sigma_K$ runs over the place of $K$ above $z$, that is to say above $z_1,z_2$.
The first term $[\delta_{M_1+M'-u}(z_1+z_2)_{y_1} ,x_{y_1} ]_{y_1}$ in the sum (5.6.16) is zero by (5.6.10) and (5.6.17).
By (5.6.14), the second term, an element of ${\mathds Q}/{\mathds Z}$, $$[\delta_{M_1+M'-u}(z_1+z_2)_{y_2} ,x_{y_2} ]_{y_2}$$ is non-zero for $0\leq u\leq M'-1$ and hence $\delta_{M_1}(z_1+z_2)$ has order at least $l^{M'}$. We obtain from this and (5.6.16) that the Cassels pairing $$<\delta_{M_1}(z_1+z_2), l^{u}f>_{\rm Cassels}=[\delta_{M_1+M'-u}(z_1+z_2)_{y_1} ,x_{y_1} ]_{y_1}$$ is non-zero for all $u$ such that $0\leq u\leq M'-1$. Hence this character $$s\mapsto <\delta_{M_1}(z_1+z_2), s>_{\rm Cassels},
\ \ \coprod\!\!\!\!\coprod(E/K)_{l^\infty}^{-\epsilon} \to {\mathds
Q}/{\mathds Z}$$ has restriction to ${\mathds Z}f$ which generates the dual of this subgroup ${\mathds Z}f$ of order $l^{M'}$.
As $f$ is any element of the abelian group $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}^{-\epsilon} $, the non-degeneracy of the Cassels pairing on $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}^{-\epsilon} $ implies that the classes $\delta_{M_1}(z_1+z_2)$, for divisors $z_1+z_2\in \Lambda^2(a)$, generate the $-\epsilon$ eigenspace $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}^{-\epsilon} $ which proves the part of the theorem for $\coprod\!\!\!\coprod$$(E/K)_{l^\infty}^{-\epsilon} $. ${\sqcap \!\!\!\!\sqcup}$
0.4in[**References**]{}
\[1\] Brown M.L., Heegner Modules and Elliptic Curves. Springer Lecture Notes in Mathematics No. 1849. Springer Verlag, Berlin-Heidelberg-New York, 2004.
\[2\] Gross B.H., Kolyvagin’s work on modular elliptic curves. In: $L$-functions and Arithmetic (eds. J.H. Coates and M.J. Taylor) Cambridge University Press 1990, pp. 235-256.
\[3\] Gross B.H., Zagier D., Heegner points and derivatives of $L$-series. Invent. Math. [**84**]{}, 225-320 (1986)
\[4\] Kolyvagin V.A., Euler systems. In: The Grothendieck Festschrift (Vol. II). P. Cartier et al., eds., pp. 435-483. Progress in Mathematics 8. Birkhäuser, Boston 1990
\[5\] Kolyvagin V.A., Finiteness of $E({\mathds Q})$ and $\coprod\!\!\!\coprod$$(E/{\mathds Q})$ for a class of Weil curves. Math. USSR Vol. [**32**]{} (1989) No. 3.
\[6\] Kolyvagin V.A., On the structure of Shafarevich-Tate groups, Proceedings of USA-USSR Symposium on Algebraic Geometry, Chicago 1989. Lecture Notes in Mathematics No. 1479, Springer Verlag, Berlin-Heidelberg-New York 1991
\[7\] Kolyvagin V.A., On the structure of Selmer groups, Math. Annalen [**291**]{}, 253-259 (1991)
\[8\] Mccallum W.G., Kolyvagin’s work on Shafarevich-Tate groups. In: $L$-functions and Arithmetic (eds. J.H. Coates and M.J. Taylor) Cambridge University Press 1990, pp. 295-316.
\[9\] Milne J.S., Arithmetic Duality Theorems, Perspectives in Mathematics, Academic Press, Boston 1986
\[10\] Raynaud, M., Caractéristique d’Euler-Poincaré d’un faisceau et cohomologie des variétés abéliennes. In: Dix Exposés sur la Cohomologie des Schémas. Ed. A. Grothendieck, pp. 12-30. North Holland Publ. Comp. Amsterdam 1967.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A topologically-invariant and additive homology class is mostly not a natural transformation as it is. In this paper we discuss turning such a homology class into a natural transformation; i.e., a “categorification" of it. In a general categorical set-up we introduce *a generalized relative Grothendieck group* from a cospan of functors of categories and also consider *a categorification of additive invariants on objects*. As an example, we obtain a general theory of characteristic homology classes of singular varieties.'
address:
- 'Jörg Schürmann: Westf. Wilhelms-Universität, Mathematisches Institut, Einsteinstrasse 62, 48149 Münster, Germany '
- 'Department of Mathematics and Computer Science, Faculty of Science, Kagoshima University, 21-35 Korimoto 1-chome, Kagoshima 890-0065, Japan'
author:
- Jörg Schürmann and Shoji Yokura
title: Grothendieck groups and a categorification of additive invariants
---
\[equation\][Theorem]{} \[equation\][Proposition]{} \[equation\][Problem]{} \[equation\][Corollary]{} \[equation\][Conjecture]{} \[equation\][Lemma]{} \[equation\][Example]{} \[equation\][Definition]{} \[equation\][Remark]{}
Introduction {#intro}
============
Characteristic classes are important invariants in modern geometry and topology to investigate other invariants. Classically characteristic classes with values in cohomology theory were considered for real or complex vector bundles. The main feature of them is that they are formulated as natural transformations from the contravariant functor of vector bundles to the cohomology theory. When it comes to the case of possibly singular varieties, characteristic classes are considered in homology theory, instead of cohomology theory and still formulated as *natural transformations from a covariant functor $\m F$ to a (suitable) homology theory $H_*$, satisfying the normalization condition that for a smooth variety $X$ the value of a distinguished element $\Delta_X$ of $\m F(X)$ is equal to the Poincaré dual of the corresponding characteristic cohomology class of the tangent bundle*: $$\tau_{c\ell}: \m F(-) \to H_*(-) \quad \text{such that for $X$ smooth} \quad \tau_{c\ell}(\Delta_X) = c\ell(TX) \cap[X].$$ Most important and well-studied theories of characteristic homology classes of singular varieties are the following ones, where we consider here for simplicity only the category of complex algebraic varieties and functoriality is required only for proper morphisms. Below, in (1) and (2) the homology theory $H_*(X)$ is either Chow groups $CH_*(X)$ or the even-degree Borel–Moore homology groups $H^{BM}_{2*}(X)$, whereas in (3) $X$ is assumed to be compact and thus the homology theory $H_*(X)$ is the usual even-degree homology group $H_{2*}(X)$:
1. MacPherson’s Chern class [@MacPherson] $c^{Mac}_*: F(-) \to H_*(-)$ is the unique natural transformation from the covariant functor $F$ of constructible functions to the homology $H_*$, satisfying the normalization condition that for a smooth variety $X$ the value of the characteristic function is the Poincaré dual of the total Chern class of the tangent bundle: $$c^{Mac}_*(\jeden_X) = c(TX) \cap [X].$$
2. Baum–Fulton–MacPherson’s Todd class [@BFM] $td^{BFM}_*: G_0(X) \to H_*(X)\otimes \bQ$ is the unique natural transformation from the covariant functor $G_0$ of Grothendieck groups of coherent sheaves to the rationalized homology $H_*\otimes \bQ$, satisfying the normalization condition that for a smooth variety $X$ the value of (the class of) the structure sheaf is the Poincaré dual of the total Todd class of the tangent bundle: $$td^{BFM}_*([\m O_X]) = td(TX) \cap [X].$$
3. Goresky– MacPherson’s homology $L$-class [@GM], which is extended as a natural transformation by Cappell-Shaneson [@CS] (also see [@Yokura-TAMS]), $L^{CS}_*: \Omega(X) \to H_*(X)\otimes \bQ$ is a natural transformation from the covariant functor $\Omega$ of cobordism groups of self-dual constructible sheaf complexes to the rationalized homology theory $H_*\otimes \bQ$, satisfying the normalization condition that for a smooth variety $X$ the value of the (class of the) shifted constant sheaf is the Poincaré dual of the total Hirzebruch–Thom’s $L$-class of the tangent bundle: $$L^{CS}_*([\bQ_X[\op{dim} X]]) = L(TX) \cap [X].$$ Here $X$ is assumed to be compact.\
Recently these three theories $c^{Mac}_*, td^{BFM}_*$ and $L^{CS}_*$ are “unified" by the motivic Hirzeruch class [@BSY] (see also [@SY], [@Schuermann-MSRI] and [@Yokura-MSRI]) ${T_y}_*: K_0(\m V_{\bC}/X) \to H_*(X)\otimes \bQ[y]$, which is the unique natural transformation such that for a smooth variety $X$ the value of the isomorphism class of the identity $X \xrightarrow {id_X} X$ is the Poincaré dual of the generalized Hirzebruch–Todd class $T_y(TX)$ of the tangent bundle: $${T_y}_*([X \xrightarrow {id_X} X]) = T_y(TX) \cap [X].$$
Here $K_0(\m V_{\bC}/X)$ is the relative Grothendieck group of the category $\m V_{\bC}$ of complex algebraic varieties, i.e., the free abelian group generated by the isomorphism classes $[V \xrightarrow {h} X]$ of morphism $h \in hom_{\m V_{\bC}}(V, X)$ modulo the relations
- $[V_1 \xrightarrow {h_1} X] + [V_2 \xrightarrow {h_2} X] = [V_1 \sqcup V_2 \xrightarrow {h_1 + h_2} X],$ with $\sqcup$ the disjoint union, and
- $[V \xrightarrow {h} X] = [V \setminus W \xrightarrow {h_{V \setminus W}} X] + [ W \xrightarrow {h_{W}} X]$ for $W \subset V$ a closed subvariety of $V$.
The generalized Hirzebruch–Todd class $T_y(E)$ of the complex vector bundle $E$ (see [@Hirzebruch; @HBJ]) is defined to be $$T_y(E) := \prod _{i=1}^{\op {rank} E} \left (\frac {\alp _i(1+y)}{1-e^{-\alp _i(1+y)}} - \alp _i y \right ) \in H^*(X) \otimes \bQ[y],$$ where $\alp _i$ is the Chern root of $E$, i.e., $\displaystyle c(E) = \prod_{i=1}^{\op{rank} E} (1 + \alp_i).$ Note that
1. $T_{-1}(E) =c(E)$ is the Chern class,
2. $T_{0}(E) =td(E)$ is the Todd class and
3. $T_{1}(E) =L(E)$ is the Thom–Hirzebruch $L$-class.\
The “unification" means the existence of the following commutative diagrams of natural transformations:
$$\xymatrix{
K_0(\m V_{\bC}/X) \ar[dd]_{{T_{-1}}_*} \ar[dr]^{\, \,const} && K_0(\m V_{\bC}/X) \ar[dd]_{{T_{0}}_*} \ar[dr]^{\, \,coh} && K_0(\m V_{\bC}/X)\ar[dd]_{{T_{1}}_*} \ar[dr]^{\, \,sd} \\
& F(X) \ar[dl] ^{c^{Mac}_*\otimes \bQ} && G_0(X) \ar[dl]^{td^{BFM}_*} && \Omega(X) \ar[dl]^{L^{CS}_*} \\
H_*(X)\otimes \bQ, && H_*(X)\otimes \bQ, && H_*(X)\otimes \bQ. }$$
Here $const:K_0(\m V_{\bC}/X) \to F(X)$ is defined by $const([V \xrightarrow {h} X]):= h_* \jeden _V$. The other two comparison transformations are characterized by $coh([V \xrightarrow {h} X]) = h_*\m ([O_V])$ and $sd([V \xrightarrow {h} X]) = h_*([\bQ_V[\op{dim}V]])$ for $V$ smooth and $h$ proper. ${T_y}_*(X) := {T_y}_*([X \xrightarrow {id_X} X])$ is called the *motivic Hirzebruch class of $X$*. For more details see [@BSY].\
Besides these characteristic classes formulated as natural transformations, there are several important homology classes which are usually not formulated or not captured as such natural transformations; for example: Chern-Mather class $c_*^M(X)$ ([@MacPherson]), Segre–Mather class $s_*^M(X)$ ([@Yokura-segre]), Fulton’s canonical Chern class $c^F_*(X)$ ([@Fulton-book]), Fulton–Johnson’s Chern class $c^{FJ}_*(X)$ ([@Fulton-Johnson; @Fulton-book]), Milnor class $\m M(X)$ (e.g., see [@Aluffi], [@BLSS], [@PP], [@Yokura-milnor], etc.), which is (up to sign) the difference of the MacPherson’s Chern class and the Fulton–Johnson’s Chern class, etc.
In [@Yokura-JS] we captured Fulton–Johnson’s Chern class and the Milnor class $\m M(X)$ as natural transformations, with the latter one as a special case of the Hirzebruch–Milnor class (also see [@CMSS]), using the motivic Hirzebruch class. In this paper we generalize the approach and results of [@Yokura-JS] to a more general abstract categorical context (cf. [@Yokura-RIMS]).\
Let $\m B$ be a category with a *coproduct* $\sqcup$. Here we assume that a fixed coproduct $X \sqcup Y$ of any two objects $X,Y\in ob(\m B)$ has been choosen, as well as an initial object $\emptyset$ in $\m B$, so that we can view $(\m B,\sqcup)$ as a symmetric monoidal category with unit $\emptyset$. Examples are the category $\mathcal{TOP}$ (resp., $\mathcal {TOP}_{lc}$) of (locally compact) topological spaces, the category $\mC^{\infty}$ of $\m C^{\infty}$-manifolds, or the category $\mathcal V_k$ of algebraic varieties (i.e., reduced separated schemes of finite type) over a base field $k$, with the coproduct $\sqcup$ given by the usual disjoint union. Moreover, for $\m B=\mathcal {TOP}_{lc}, \mC^{\infty}$ or $\mathcal V_k$, we often consider them as categories only with respect to proper morphisms. Let $\mathcal {AB}$ be the category of abelian groups with the coproduct $\sqcup$ given by the direct sum $\oplus$. Then a covariant functor $\m H : \mathcal B \to \mathcal {AB}$ is called *additive* (cf. [@Levine-Morel]), if it preserves the coproduct structure $\sqcup$, i.e., if we get the following isomorphism: $$(i_X)_*\oplus (i_Y)_*: \m H(X) \oplus \m H(Y) \xrightarrow {\cong} \m H(X \sqcup Y)$$ where $i_X: X \to X \sqcup Y$ and $i_Y:Y \to X \sqcup Y$ are the canonical injections. In particular, $\m H(\emptyset) = \{0\}$. Examples of such an additive functor $\mH$ are a generalized homology theory $H_*$ for $\m B = \m {TOP}$, the Borel–Moore homology theory $H_*^{BM}$ for $\m B = \m {TOP}_{lc}$, the functor $G_0$ of Grothendieck groups of coherent sheaves or the Chow groups $CH_*$ for $\m B = \m V_k$, as well as the functor $F$ of constructible functions or $\Omega$ of cobordism groups of self-dual constructible sheaf complexes for $\m B = \m V_{\bC}$.\
Let $\mC$ be a category of “(topological) spaces with some additional structures *stable* under $\sqcup$", such as the category of (complex) algebraic varieties, the category of (compact) topological spaces, or the category of oriented smooth (i.e., $C^{\infty}$-) manifolds, etc. More abstractly, $\mC$ is a *symmetric monoidal category* $(\mC, \sqcup)$ together with a strict monoidal forgetful functor $\frak f: (\m C, \sqcup) \to (\m B, \sqcup)$. Then an *additive invariant* on objects in $ob(\mC)$ with values in the additive functor $\m H$ is given by an element $\alp(X) \in \mH(\frak f(X))$ for all $X \in ob(\m C)$ satisfying $$\alp(X \sqcup Y) = (i_{\frak f(X)})_* \alp(X) + (i_{\frak f(Y)})_*\alp(Y),$$ where $i_{\frak f(X)}$ and $i_{\frak f(Y)}$ are the corresponding canonical inclusions. Note that the sum or difference $\alpha\pm \beta$ of two additive invariants with values in $\m H$ is again such an invariant.\
\[singular\] Let $\m B=\m V_{\bC}$ be the category of complex algebraic varieties and $\m C$ be a full subcategory of $\m B$ which is stable under disjoint union, isomorphisms and contains the initial object $\emptyset$ and the final object $pt$ (e.g., *an admissible subcategory* in the sense of Levine–Morel [@Levine-Morel]), such as the category $\m V_{\bC}^{sm}$ of smooth varieties, $\m V^{lpd}_{\bC}$ of locally pure dimensional varieties, $\m V_{\bC}^{emb}$ of varieties embeddable into smooth varieties, or $\m V_{\bC}^{lci}$ of local complete intersections. Then we have the following additive invariants, some of which are already mentioned above:
1. characteristic homology classes $c\ell_*(X) := c\ell(TX) \cap [X] \in H_*(X)\otimes R$ for $X \in ob(\m V^{sm}_{\bC})$ with $[X]$ the fundamental class and $TX$ the tangent bundle, or the corresponding virtual classes $c\ell_*(X)$ for $X \in ob(\m V^{lci}_{\bC})$, with $TX$ the virtual tangent bundle and $c\ell(TX)$ a characteristic cohomology class.
2. Fulton’s canonical Chern class or Fulton–Johnson Chern class $c^F_*(X)$, $c^{FJ}_*(X)$ for $X \in ob(\m V_{\bC}^{emb})$.
3. Mather-type characteristic homology classes $c\ell^{Ma}_*(X)$, such as Chern–Mather class and Segre–Mather class, or the local Euler obstruction $Eu_X \in F(X)$ as a constructible function for $X \in ob(\m V_{\bC}^{lpd})$.
4. the (class of the) self-dual Intersection Homology complex $[\m {IC}_X] \in \Omega (X)$ and for $X$ compact its Goresky–MacPherson’s $L$-homology class $L^{GM}(X) \in H_*(X)$ for $X \in ob(\m V^{lpd}_{\bC})$.
A *categorification of the additive invariant $\alp$* with values in the additive functor $\m H$ is meant to be an associated *natural transformation* $$\tau_{\alp}: \m H_{\alp} \to \m H\circ \frak f$$ from a *covariant functor* $\m H_{\alp}$ on the category $\m C $, such that $$\tau_{\alp}(\delta_X) = \alp(X)$$ for some *distinguished element* $\delta_X \in \m H_{\alp}(X)$ and all $X \in ob(\m C)$.\
To construct such a covariant functor $\m H_{\alp}$, we introduce *generalized relative Grothendieck groups*, using comma categories in a more abstract categorical context. The construction of such a covariant functor is hinted by the definition of the relative Grothendieck group $ K_0(\m V_{\bC}/X)$ and more clearly by the description of the oriented bordism group $\Omega^{SO}_*(X)$.
The oriented bordism group $\Omega^{SO}_m(X)$ of a topological space $X$ is defined to be the free abelian group generated by the isomorphism classes $[M \xrightarrow h X]$ of *continuous maps* $M \xrightarrow h X$ from *closed oriented smooth manifolds* $M$ of dimension $m$ to the given *topological space* $X$, modulo the following relations
1. $[M \xrightarrow h X] + [M' \xrightarrow {h'} X] =
[M \sqcup M' \xrightarrow {h + h'} X]$,
2. if $M \xrightarrow h X$ and $M' \xrightarrow {h'} X$ are *bordant*, then $[M \xrightarrow h X] = [M' \xrightarrow {h'} X].$
Note that if $M \xrightarrow h X$ and $M' \xrightarrow {h'} X$ are isomorphic to each other, i.e., there exists an orientation preserving diffeomorphism $\phi: M \cong M'$ such that the following diagram commutes *as topological spaces*: $$\xymatrix{M \ar[dr]_ {h}\ar[rr]^ {\phi} && M' \ar[dl]^{h'}\\& X,}$$ then $M \xrightarrow h X$ and $M' \xrightarrow {h'} X$ are clearly bordant to each other. Hence we can say that the bordism group $\Omega^{SO}_m(X)$ is defined to be the free abelian group generated by isomorphism classes of continuous maps $M \xrightarrow h X$ from *closed oriented smooth manifolds* $M$ of dimension $m$ to the given *topological space* $X$, modulo the relations (1) and (2) above. The bordism group $\Omega^{SO}_m(X)$ can also be described slightly differently as follows: the set of the bordism classes $[M \xrightarrow h X]$ can be turned into a commutative semi-group or monoid $\mathcal M(X)$ by the relation (1) above, with $0=[\emptyset \to X]$. Then the Grothendieck group or the group completion of the monoid $\mathcal M(X)$ is nothing but the above bordism group $\Omega^{SO}_m(X)$.\
In the definition of the bordism group two categories are involved:
- the category $\mC^{\infty}_{co}$ of closed oriented $\m C^{\infty}$-manifolds, i.e., compact oriented $\m C^{\infty}$-manifolds without boundary,
- the category $\mTOP$ of topological spaces.
It should be emphasized that *even though we consider such a finer category $\mC^{\infty}_{co}$ for a source space $M$ the map $h:M \to X$ of course has to be considered in the crude category $\mTOP$*.
The bordism group $\Omega^{SO}_*$ is a generalized homology theory, in particular $\Omega^{SO}_*$ is a covariant functor $$\Omega^{SO}_*: \mTOP \to \mAB,$$ where $\mAB$ is the category of abelian groups.
Clearly we can consider this covariant functor on a different category finer than the category $\mTOP$ of topological spaces, e.g., the category $\m V_{\bC}$ of complex algebraic varieties. Namely we consider *continuous* maps $h: M \to V$ from closed oriented manifolds $M$ to a complex algebraic variety $V$. Then we still get a covariant functor $$\Omega^{SO}_*: \m V_{\bC} \to \mAB.$$ In this set-up *three different categories* $\mC^{\infty}_{co}$, $\mTOP$ and $\m V_{\bC}$ are involved. More precisely, we have the following forgetful functors $\frak f_s: \mC^{\infty}_{co} \to \mTOP$ and $\frak f_t: \m V_{\bC} \to \mTOP$ (here “s" and “t" mean “source object" and “target object): $$\mC^{\infty}_{co} \xrightarrow {\frak f_s} \mTOP \xleftarrow {\frak f_t} \m V_{\bC}.$$ And the commutative triangle $$\xymatrix{M \ar[dr]_ {h}\ar[rr]^ {\phi} && M' \ar[dl]^{h'}\\& V}$$ really means $$\xymatrix{\frak f_s(M) \ar[dr]_ {h}\ar[rr]^ {\frak f_s(\phi)} && \frak f_s (M') \ar[dl]^{h'}\\& \frak f_t(V).}$$
Abstracting this situation, we deal in this paper with a more general situation of a cospan of categories $$\mC_s \xrightarrow {S} \mB \xleftarrow {T} \mC_t,$$ where $\m B$ is a category with (chosen) coproducts $\sqcup$, $(\m C_s, \sqcup)$ is a *symmetric monoidal category* with $S: \mC_s \to \mB$ a strict monoidal functor respecting the units $\emptyset$, and $T: \mC_t \to \mB$ just a functor.
Two triples $(V, X, h), (V', X, h')$, where $V, V' \in ob(\m C_s), X \in ob(\mC_t)$, $h \in hom_{\m B}(S(V), T(X))$ and $h' \in hom_{\m B}(S(V'), T(X))$, are called *isomorphic*, if there exists an isomorphism $\phi:V \xrightarrow {\cong} V' \in hom_{\mC_s}(V, V')$ such that the following diagram commutes, just like as above: $$\xymatrix{S(V) \ar[dr]_ {h}\ar[rr]^ {S(\phi)} && S(V') \ar[dl]^{h'}\\& T(X).}$$
Then the isomorphism classes $[(V, X, h)]$ of triples $(V, X, h)$ can be turned into a monoid $\m M(\mC_s \xrightarrow S \mB /T(X))$ by $$[(V, X,h)] \sqcup [(V', X,h')] := [(V \sqcup V, X, h + h')],$$ with unit $0=[(\emptyset, X,h)]$. The associated Grothendieck group is denoted by $K(\mC_s \xrightarrow S \mB /T(X))$ (respectively $K(\mC_s \xrightarrow S \mB /X)$ in case $\mC_t =B$ and $T=\op{id}_{\mC_t}$) and called the *canonical generalized $(S,T)$-relative Grothendieck group*. It is a covariant functor with respect to $X$ by composing $h$ with $T(f)$ for $f \in hom_{\mC_t}(X, X')$.
\[dist\] In the following two cases, a *distinguished element $\delta_X \in K(\mC_s \xrightarrow S \mB /T(X))$* is available:
1. $\delta_{S(V)} := [(V, S(V), \op{id}_{S(V)})]$ for $X = S(V)$ in the case when $T= \op{id}_{\mC_t} :\mC_t \to \mC_t =\mB$ is the identity functor.
2. $\delta_V:= [(V, V, \op{id}_{S(V)})]$ for $X =V$ in the case when $\mC_t = \mC_s$ and $T = S$.\
When it comes to a categorification of an additive invariant $\alp$ with values in an additive covariant functor $\mH: \mB \to \mathcal {AB}$, we sometimes need to consider a stronger notion of an isomorphism of triples $(V, X, h)$ (e.g. in the context with $V$ an “oriented space"): Let $(V, X, h), (V', X, h')$ be two triples as above. Then they are called *$\alp$-isomorphic* if the following holds: they are isomorphic as above by an isomorphism $\phi:V \xrightarrow {\cong} V' \in hom_{\mC_s}(V, V')$ such that $$(S(\phi))_*\alp(V) = \alp(V').$$
If this last equality holds for any isomorphism $\phi:V \xrightarrow {\cong} V' \in hom_{\mC_s}(V, V')$, then $\alp$ is also called an *isomorphism invariant*. So in this case there is no difference between $\alp$-isomorphism classes and isomorphism classes for such triples $(V, X, h)$. The $\alp$-isomorphism class of $(V, X, h)$ is denoted by $[(V, X, h)]_{\alp}$ and these can be turned as before into a monoid $\m M_{\alp}(\mC_s \xrightarrow S \mB /T(X))$, since $\alp$ is additive. The associated generalized $(S,T)$-relative Grothendieck group is denoted by $K_{\alp}(\mC_s \xrightarrow S \mB /T(X))$, respectively $K_{\alp}(\mC_s \xrightarrow S \mB /X)$ in case $\mC_t =B$ and $T=\op{id}_{\mC_t}$. If it is clear that we consider this group $K_{\alp}(\mC_s \xrightarrow S \mB /T(X))$ from the context, we sometimes omit the suffix $\alp$ from the notation, e.g. if $\alp$ is an isomorphism invariant. Similarly, if we consider $\m B=\mathcal {TOP}_{lc}, \mC^{\infty}$ or $\mathcal V_k$ only as categories with respect to proper morphisms, then this is indicated by the notation $K^{prop}_{\alp}(\mC_s \xrightarrow S \mB /T(X)),
K^{prop}_{\alp}(\mC_s \xrightarrow S \mB /X)$ etc.
(A “categorification" of an additive invariant) \[cat\] Let $\mH: \mB \to \mathcal {AB}$ be an additive functor on $\mB$ with $T' = \mH \circ T$. Then an additive invariant $\alp$ on $ob(\mC_s)$ with values in $\mH$ induces a natural transformation on $\mC_t$: $$\tau_{\alp}:K_{\alp}\bigl(\mC_s \xrightarrow S \mB/T(-) \bigr) \to T'(-); \, \tau_{\alp}([(V, X, h)]_{\alp}) := h_*(\alp(V)).$$ Assume $T$ is full, and consider one of the cases (i) or (ii) as in Remark \[dist\]. Then $\tau_{\alp}$ is the unique natural transformation satisfying $$\tau_{\alp}([(V, S(V), \op {id}_{S(V)})]_{\alp}) = \alp(V) \,\, \, \text {or} \, \, \, \tau_{\alp}([(V, V, \op {id}_{S(V)})]_{\alp}) = \alp(V).$$
Let us illustrate this result by some examples.
Let $\mTOP$ be the category of topological spaces and let $\mC^{\infty}_{co}$ be the category of closed oriented $C^{\infty}$-manifolds, whose morphisms are just differentiable maps. Consider the following cospan of these categories: $$\mC^{\infty}_{co} \xrightarrow {\frak f} \mTOP \xleftarrow {id_{\mTOP}} \mTOP,$$ where $\frak f: \mC^{\infty}_{co} \to \mTOP$ is the forgetful functor. Then the fundamental class $[M]\in H_*(M)$ for $M\in ob(\mC^{\infty}_{co})$ is an additive invariant with values in the usual homology $H_*$. More generally, let $c\ell(E)\in H^*(-;R)$ be a contravariant functorial characteristic class of (isomorphism classes of) oriented vector bundles $E$. Then also $$\alp(M):=c\ell(TM)\cap [M]\in H_*(M;R)$$ is additive, where we recover the fundamental class for $c\ell$ the unit $1\in H^*(-)$. Assume $c\ell$ is normalized in the sense that $1=c\ell^0(E)\in H^0(-;R)$, so that a diffeomorphism $\phi: M \xrightarrow {\cong} M'$ of such oriented manifolds is an $\alp$-isomorphism if and only if it is orientation preserving. Then there exists a unique natural transformation $$\tau_{c\ell}: K_{\alp}(\mC^{\infty}_{co} \xrightarrow {\frak f} \mTOP /-) \to H_*(-;R)$$ such that for a closed oriented $C^{\infty}$-manifold $M$ $$\tau_{c\ell}([(M, \frak f(M), id_{\frak f(M)})]_{\alp}) =c\ell(TM)\cap [M].$$ Let $\Omega^{SO}_*(X)$ be the oriented bordism group of a topological space $X$. Then we have $$\Omega^{SO}_*(X) \cong K_{\alp}(\mC^{\infty}_{co} \xrightarrow {\frak f} \mTOP /X) / \sim,$$ with $\sim$ the bordism relation. And $\tau_{c\ell}$ factorizes over $\Omega^{SO}_*(X)$ in the case when $c\ell$ is stable in the sense that $c\ell(E\oplus \bR)=c\ell(E)$ for $\bR$ the trivial real line bundle.
Let $\m V_{\bC}$ be the category of complex algebraic varieties, with $\m C$ the full subcategory $\m V_{\bC}^{sm},\m V^{lpd}_{\bC}, \m V_{\bC}^{emb} $ or $\m V_{\bC}^{lci}$ of smooth, locally pure dimensional, embeddable or local complete intersection varieties, with $$\alp(V):=c\ell_*(V)\in H_*(V)\otimes R$$ for $V\in ob(\m C)$ a corresponding additive characteristic homology class as in Example \[singular\]. Here the homology theory $H_*(X)$ is either Chow groups $CH_*(X)$ or the even-degree Borel–Moore homology groups $H^{BM}_{2*}(X)$. Then $\alp$ is also an isomorphism invariant. Consider the following cospan of categories, where we consider only proper morphisms: $$\mC \xrightarrow {\frak f} \m V_{\bC} \xleftarrow {id_{\m V_{\bC}}} \m V_{\bC},$$ with $\frak f: \mC \to \m V_{\bC}$ being the inclusion functor. Then there exists a unique natural transformation $$\tau_{c\ell}: K^{prop}(\mC \xrightarrow {\frak f}\m V_{\bC} /-) \to H_*(-)\otimes R,$$ such that for all $V\in ob(\mC)$ $$\tau_{c\ell}([(V, \frak f(V), id_{\frak f(V)})]) =c\ell_*(V).$$ Moreover, there is a tautological surjective natural transformation $$K^{prop}(\mC \xrightarrow {\frak f} \m V_{\bC} /-) \to K_0(\m V_{\bC}/-)$$ to the relative Grothendieck group $K_0(\m V_{\bC}/-)$ of complex algebraic varieties. And for $\mC=\m V_{\bC}^{sm}$ and $$c\ell_*(V)=T_y(TV)\cap [V]=T_{y*}(V)
\quad \text{or} \quad c\ell_*(V)=c(TV)\cap [V]=c_*(V),$$ i.e., the Hirzebruch homology class or MacPherson’s Chern class, the transformations $\tau_{T_{y*}}$ and $\tau_{c_{*}}$ factorize over $K_0(\m V_{\bC}/-)$ by [@BSY].
In §2 we explain the general categorical background and prove Theorem \[cat\], whereas in §§3-4 we apply it to many different geometric situations, e.g., for obtaining Riemann–Roch type theorems.
Generalized relative Grothendieck groups
========================================
Let $\mC$ be a symmetric monoidal category equipped with monoidal structure $\oplus$ and the unit object $\emptyset$. The Grothedieck group $K(\mC)$ is defined to be the free abelian group generated by the isomophism classes $[X]$ of objects $X \in ob(\mC)$ modulo the relations $$[X] + [Y] = [X \oplus Y], \quad \text{, with $0 = [\emptyset]$.}$$
A functor $\Phi: \mC_1 \to \mC_2$ of two symmetric monoidal categories is a functor which preserves $\oplus$ in the relaxed sense that there are natural transformations: $$\Phi(A) \oplus_{\mC_2} \Phi(B) \to \Phi(A \oplus_{\mC_1}B),$$ In some usage it requires an isomorphism $$\Phi(A) \oplus_{\mC_2} \Phi(B) \cong \Phi(A \oplus_{\mC_1}B),$$ in which case it is called a *strong* monoidal functor. Here we also assume that $\Phi$ respects the units.
Let $\m B$ be a category with a (chosen) *coproduct* $\sqcup$ and $\m H : \mathcal B \to \mathcal {AB}$ be an additive covariant functor. Then we have $$\m H_*(X \sqcup Y) \cong \m H_*(X) \oplus \m H_*(Y) \quad
\text{, with $\m H_*(\emptyset)=\{0\}$.}$$ Examples of such an additive functor $\m H$ are a generalized homology theory $H_*$ for $\m B = \m {TOP}$, the Borel–Moore homology theory $H_*^{BM}$ for $\m B = \m {TOP}_{lc}$, the functor $G_0$ of Grothendieck groups of coherent sheaves or the Chow groups $CH_*$ for $\m B = \m V_k$, as well as the functor $F$ of constructible functions or $\Omega$ of cobordism groups of self-dual constructible sheaf complexes for $\m B = \m V_{\bC}$.
Let $\Phi: \mC_1 \to \mC_2$ be a strong monoidal functor of two symmetric monoidal categories. Then the map $$\Phi_*:K(\mC_1) \to K(\mC_2), \quad \Phi_*([X]):= [\Phi(X)]$$ is a well-defined group homomorphism. Namely, the Grothendieck group $K$ is a covariant functor from the category of such categories and functors to the category of abelian groups.
Now we recall the notion of *comma category* and *fiber category* (e.g., see [@Maclane]):
Let $\mC_s$ (the suffix “s" meaning *source*), $\mC_t$ (the suffix “t" meaning *target*) and $\mB$ be categories, and let $$\mC_s \xrightarrow {S} \mB \xleftarrow {T} \mC_t$$ be functors among them (which shall be called *a cospan of categories*). The *comma category*, denoted by $(S \downarrow T)$, is formed by
- the objects of $(S \downarrow T)$ are triples $(V, X, h)$ with $V \in ob(\mC_s)$, $X \in ob(\mC_t)$ and $h \in hom_{\mB}(S(V), T(X))$,
- the morphisms from $(V, X, h)$ to $(V', X', h')$ are pairs $(g_s,g_t)$ where $g_s: V \to V'$ is a morphism in $\mC_s$ and $g_t: X \to X'$ is a morphism in $\mC_t$ such that the following diagram commutes in the category $\mB$: $$\CD
S(V) @> S(g_s)>> S(V') \\
@V h VV @VV h' V\\
T(X) @>> T(g_t) > T(X'). \endCD$$
Let $\mC_s \xrightarrow {S} \mB \xleftarrow {T} \mC_t$ be a cospan and let $(S \downarrow T)$ be the above comma category associated to the cospan. We define the canonical *projection functors* as follows:
1. $\pi_t: (S \downarrow T) \to \mC_t$ is defined by
- for an object $(V, X, h)$, $\pi_t((V, X, h)) := X,$
- for a morphism $(g_s, g_t): (V, X, h) \to (V', X', h')$, $\pi_t((g_s, g_t)) := g_t.$
2. $\pi_s: (S \downarrow T) \to \mC_s$ is defined by
- for an object $(V, X, h)$, $\pi_s((V, X, h)) := V,$
- for a morphism $(g_s, g_t): (V, X, h) \to (V', X', h')$, $\pi_s((g_s, g_t)) := g_s.$
Namely *a cospan of categories* $\mC_s \xrightarrow {S} \mB \xleftarrow {T} \mC_t$ induces *a span of categories* $\mC_s \xleftarrow {\pi_s} (S \downarrow T) \xrightarrow {\pi_t} \mC_t.$
Let $F: \mC \to \m{D}$ be a functor of two categories. Then, for an object $B \in ob(\m{D})$ *the fiber category* of $F$ over $B$, denoted by $F^{-1}(B)$, is defined to be the category consisting of
- the objects $X \in ob(\mC)$ such that $F(X) = B$,
- for such objects $X, X'$, morphisms $f:X \to X'$ such that $F(f) = id_B$.
In other words, more precisely, this category should be denoted by $F^{-1}(B, id_B)$.
As above, let us consider a cospan of categories and its associated span of categories: $$\mC_s \xrightarrow {S} \mB \xleftarrow {T} \mC_t, \quad \quad \mC_s \xleftarrow {\pi_s} (S \downarrow T) \xrightarrow {\pi_t} \mC_t.$$
1. For an object $X \in \mC_t$, the fiber category $\pi_t^{-1}(X)$ is nothing but the *S-over category* $(S \downarrow T(X))$, whose objects are *objects $S$-over $T(X)$*, i.e., the triple $(V, X, h)$, and for two triples $(V, X, h)$ and $(V', X, h')$ a morphism from $(V, X, h)$ to $(V', X, h')$ is $g_s \in hom_{\mC_s}(V, V')$ such that the following triangle commutes: $$\xymatrix{S(V) \ar[dr]_ {h}\ar[rr]^ {S(g_s)} && S(V') \ar[dl]^{h'}\\& T(X).}$$
2. Furthermore, if $\mC_s = \mB$ and $S = id_S$ is the identity functor, then the above *S-over category* $(S \downarrow X)$ is the standard *over category* $(\mB \downarrow X)$, whose objects are objects over $X$, i.e., morphisms $h:V \to X$, and for two tmorphisms $h:V \to X$ and $h:V' \to X$ a morphism from $h:V \to X$ to $h:V' \to X$ is $g \in hom_{\mB}(V, V')$ such that the following triangle commutes: $$\xymatrix{V \ar[dr]_ {h}\ar[rr]^ {g} && V' \ar[dl]^{h'}\\& X.}$$
For an object $V \in \mC_s$, the fiber category $\pi_s^{-1}(S)$ is nothing but the *T-under category* $(S(V) \downarrow T)$, whose objects are *objects $T$-under $S(V)$*, i.e., the triple $(V, X, h)$, and for two triples $(V, X, h)$ and $(V, X', h')$ a morphism from $(V, X, h)$ to $(V, X', h')$ is $g_t \in hom_{\mC_t}(X, X')$ such that the following triangle commutes: $$\xymatrix{&S(V) \ar[dl]_ {h}\ar[dr]^ {h'}&\\
T(X) \ar[rr]_{T(g_t)} && T(X').}$$
Similarly, we can think of the *T-under category* $(V \downarrow T)$ and the *under category* $(V \downarrow \mB)$, but we do not write them down here, since we do not use them below in the rest of the paper.
Let $\mC_s \xrightarrow {S} \mB \xleftarrow {T} \mC_t$ be a cospan of categories. Then a morphism $f \in hom_{\mC_t} (X_1, X_2)$ gives rise to the functor between the corresponding fiber categories: $$T(f)_*: \pi_t^{-1}(X_1) \to \pi_t^{-1}(X_2) ,$$ which is defined by
1. For an object $(V, X_1, h)$, $T(f)_*((V, X_1, h)) := (V, X_2, T(f) \circ h).$
2. For a morphism $(g_s, id_{X_1}): (V, X_1, h) \to (V', X_1, h')$ with $g_s \in hom_{\mC_s}(V, V')$, $$T(f)_*((g_s, id_{X_1})) := (g_s, id_{X_2}): (V, X_2, T(f) \circ h) \to (V', X_2, T(f) \circ h').$$ $$\xymatrix{S(V) \ar[dr]^ {h}\ar[rr]^ {g}\ar[dddr]_{T(f) \circ h} && S(V') \ar[dl]_{h'} \ar [dddl]^{T(f) \circ h'} \\
& T(X_1) \ar[dd]^(.25){T(f)}\\
&\\
&T(X_2)}$$
\[relK\] Let $\m B$ be a category with a *coproduct* $\sqcup$, with the coproduct of any two objects and an initial object chosen, and let $(\m C_s, \sqcup)$ be a symmetric monoidal category. Let $\mC_s \xrightarrow {S} \mB \xleftarrow {T} \mC_t$ be a cospan of categories with $S: (\mC_s , \sqcup)\to (\mB, \sqcup)$ a strict monoidal functor and $T:\mC_t \to \mB$ just a functor. Then for each object $X \in ob(\mC_t)$, the fiber category $\pi_t^{-1}(X)$, i.e. the $S$-over category $(S \downarrow T(X))$ becomes also a symmetric monoidal category with $$(V, X, h) \sqcup (V', X, h') := (V \sqcup V', X, h + h').$$
Let the situation be as above. A morphism $f \in hom_{\mC_t}(X_1, X_2)$ gives rise to the canonical group homomorphism $$T(f)_* : K(\pi_t^{-1}(X_1) ) \to K(\pi_t^{-1}(X_2)),$$ and $$K(\pi_t^{-1}(-)) : \mC_t \to \mAB$$ is a covariant functor from the category $\mC_t$ to the category of abelian groups.
After these simple observations, we can introduce the following notions:
$ $\
Let $\mC_s \xrightarrow {S} \mB \xleftarrow {T} \mC_t$ be a cospan of categories as in Lemma \[relK\].
1. The Grothendieck group of the fiber category of the projection functor $\pi_t: (S \downarrow T) \to \mC_t$ is denoted by $$K(\mC_s \xrightarrow {S} \mB /T(X)): = K(\pi_t^{-1}(X)))$$ and called the *generalized $(S,T)$-relative Grothendieck group of $X\in \mC_t$*.
2. If $\mC_t = \mB$ and $T = id_{\mB}$, then $K(\mC_s \xrightarrow {S} \mB /T(X))$ is simply denoted by $K(\mC_s \xrightarrow {S} \mB /X)$.
3. If $S = T= id_{\mC_s} :\mC_s \to \mC_s$ is the identity functor, then the above $id_{\mC}$-relative Grothendieck group $K(\mC_s \xrightarrow {id_{\mC_s}}\mC _s/X)$ is simply denoted by $K(\mC_s/X)$ and called the *relative Grothendieck group of $X$* or the *relative Grothendieck group of $\m C_s$ over $X$*.
All these relative Grothendieck groups are covariant functors from $\mC_t$ to $\mAB$.
If $T(X)=pt$ is a terminal object in the category $\m B$, then all the above relative Grothendieck groups are isomophic to the Grothendieck group $K(\mC_s)$ of the symmetric monoidal category $(\mC_s,\sqcup)$: $$K(\mC_s \xrightarrow {S} \mB /T(X)=pt) \cong K(\mC_s \xrightarrow {S} \mB /pt) \cong K(\mC_s/pt) \cong K(\mC_s).$$
\[comma\] Let $\mC_s \xrightarrow {S} \mB \xleftarrow {T} \mC_t$ be a cospan of categories as in Lemma \[relK\], $\Phi_b:\mB \to \mB'$ be a covariant functor preserving (chosen) coproducts and set $S':= \Phi_b \circ S$ and $T':=\Phi_b \circ T$. Then we get the canonical natural transformation from the functor $K(\mC_s \xrightarrow {S} \mB /T(-)): \mC_t \to \mAB$ to the functor $K(\mC_s \xrightarrow {S'} \mB' /T'(-)): \mC_t \to \mAB$:
$$\Phi_*: K(\mC_s \xrightarrow {S} \mB /T(-)) \to K(\mC_s \xrightarrow {S'} \mB' /T'(-)),$$ i.e., for a morphism $f \in hom_{\mC_t}(X_1, X_2)$ the following diagram commutes in the category $\mAB$: $$\CD
K(\mC_s \xrightarrow {S} \mB /T(X_1))@> \Phi_* >> K(\mC_s \xrightarrow {S'} \mB' /T'(X_1)) \\
@V T(f)_* VV @VV T'(f)_*V \\
K(\mC_s \xrightarrow {S} \mB /T(X_2)) @>> \Phi_* >K(\mC_s \xrightarrow {S'} \mB' /T'(X_2)). \endCD$$ Here $\Phi_*: K(\mC_s \xrightarrow {S} \mB /T(X)) \to K(\mC_s \xrightarrow {S'} \mB' /T'(X))$ is defined by $$\Phi_*([(V, X, h)]:= [(V, X, \Phi_b(h))].$$
It suffices to show the commutativity $f_* \Phi_* = \Phi_* f_*$ in the square diagram above. First we observe the following morphisms: $$\xymatrix{\hspace{4cm} S'(V) = \Phi_b(S(V)) \ar[d]^ {\Phi_b(h)} \hspace{2cm}\\
\hspace{4.2cm} T'(X_1) = \Phi_b(T(X_1)) \ar[d]^{T'(f)= \Phi_b(T(f))} \hspace{2cm}\\
\hspace{2.2cm} T'(X_2) =\Phi_b(T(X_2))} \hspace{2cm}$$
Then we get for $[(V, X_1, h)] \in K(\mC_s \xrightarrow {S} \mB /T(X_1))$: $$\begin{aligned}
T'(f)_*\Phi_*([(V, X_1, h)]) & = T'(f)_*([(V, X_1, \Phi_b(h))] \\
& = [(V, X_2, T'(f)\circ \Phi_b(h))] \\
& = [(V, X_2, \Phi_b(T(f) )\circ \Phi_b(h))] \\
& = [(V, X_2, \Phi_b(T(f) \circ h))] \\
& = \Phi_* ([(V, X_2, T(f) \circ h)]) \\
& = \Phi_* T(f)_* ([(V, X_1,h)]). \\\end{aligned}$$
As explained in the introduction, when it comes to a categorification of an addtitive invariant $\alp$ with values in an additive covariant functor $\m H$, we need to consider *$\alp$-isomorphism class $[(V,X,h)]_{\alp}$*. Then all the previous results hold even if we replace $[(V,X,h)]$ and $K(\mC_s \xrightarrow {S} \mB /T(-))$ with $[(V,X,h)]_{\alp}$ and $K/{\alp}(\mC_s \xrightarrow {S} \mB /T(-))$, respectively. And we get the following result:
(A “categorification" of an additive invariant) \[main\] Let $\m B$ be a category with a *coproduct* $\sqcup$, with the coproduct of any two objects and an initial object chosen, and let $(\m C_s, \sqcup)$ be a symmetric monoidal category. Consider a cospan of categories $\mC_s \xrightarrow {S} \mB \xleftarrow {T} \mC_t$, with $S: (\mC_s , \sqcup)\to (\mB, \sqcup)$ a strict monoidal functor and $T:\mC_t \to \mB$ just a functor. Let $\mH: \mB \to \mathcal {AB}$ be an additive functor on $\mB$ with $T' = \mH \circ T$. Then an additive invariant $\alp$ on $Obj(\mC_s)$ with values in $\mH$ induces a natural transformation on $\mC_t$: $$\tau_{\alp}:K_{\alp}(\mC_s \xrightarrow S \mB/T(-)) \to T'(-); \, \tau_{\alp}([(V, X, h)]_{\alp}) := h_*(\alp(V)).$$ Assume $T$ is full, and consider one of the cases (i) or (ii) as in Remark \[dist\]. Then $\tau_{\alp}$ is the unique natural transformation satisfying $$\tau_{\alp}([V, S(V), \op {id}_{S(V)}]_{\alp}) = \alp(V) \,\, \, \text {or} \, \, \, \tau_{\alp}([V, V, \op {id}_{S(V)}]_{\alp}) = \alp(V).$$
Note that $\tau_{\alp}$ is well-defined, since we consider $\alp$-isomorphism classes so that it does not depend on the chosen representative of the class $[(V,X,h)]_{\alp}$. Then $\tau_{\alp}$ becomes a group homomorphism by the definition of an additive invariant and the functoriality of $\m H$: $$\begin{aligned}
\tau_{\alp}\Bigl([(V, X, h)]_{\alp} + [(V', X, h')]_{\alp}\Bigr) & = \tau_{\alp}([(V, X, h)]_{\alp} \sqcup [(V', X, h')]_{\alp}) \\
& = \tau_{\alp}([(V \sqcup V', X, h + h')]_{\alp}) \\
& = (h+h')_*(\alp(V \sqcup V')) \\
& = (h+h')_*\Bigl( \left(i_{\frak f(V)} \right)_*\alp(V) + \left(i_{\frak f(V')} \right)_*\alp(V') \Bigr)\\
& = (h+h')_*(i_{\frak f(V)})_* (\alp(V)) + (h+h')_*(i_{\frak f(V')})_*(\alp(V'))\\
& = h_*(\alp(V)) + h'_*(\alp(V')) \\
& = \tau_{\alp}([(V, X, h)]_{\alp}) + \tau_{\alp}([(V', X, h')]_{\alp}),\end{aligned}$$ since $(h+h') \circ i_{\frak f(V)} =h$ and $(h+h') \circ i_{\frak f(V')} =h'$. Finally assume $T$ is full so that there is a morphism $f\in hom_{\mC_t}(S(V),X)$ in the case of (i), or $f\in hom_{\mC_t}(V,X)$ in the case of (ii), with $T(f)=h: S(V)\to T(X)$. Then $$f_*([(V, S(V), id_{S(V)})]_{\alp})=([(V, X, h)]_{\alp}) \quad \text{or} \quad
f_*([(V, V, id_{S(V)})]_{\alp})=([(V, X, h)]_{\alp})\:,$$ which implies the uniqueness statement.
A categorification of an additive homology class
================================================
In the rest of the paper we deal with $\mB$ a category of spaces (with some possible extra structures), with the coproduct $\sqcup$ given by the usual disjoint union. Examples are the category $\mathcal {TOP}_{(lc)}$ of (locally compact) topological spaces, the category $\mC_{(co)}^{\infty}$ of (closed oriented) $\m C^{\infty}$-manifolds, or the category $\mathcal V_k$ of algebraic varieties (i.e. reduced separated schemes of finite type) over a base field $k$. And the additive covariant functor $\m H : \mathcal B \to \mathcal {AB}$ is most of the time a suitable homology theory, like usual homology $H_*(-;R)$ with coefficients in some commutative ring $R$ or a generalized homomolgy $H_*$, in case $\mB= \mathcal {TOP}$. Moreover, for $\m B=\mathcal {TOP}_{lc}, \mC^{\infty}$ or $\mathcal V_k$, we often consider them as categories only with respect to proper morphisms, with $\m H$ the (even degree) Borel-Moore homology $H^{BM}_{(2)*}(-;R)$ with coefficients in some commutative ring $R$ for $\m B=\mathcal {TOP}_{lc}, \mC^{\infty}$, or $\m H=CH_*$ (resp. $CH_*\otimes R$) the Chow groups (with coefficients in $R$) for $\m B=\mathcal V_k$. Similarly, if we further restrict ourselfes to projective morphisms in the algebraic context, then $\m H : \mathcal V_k \to \mathcal {AB}$ could also be a suitable Borel–Moore functor in the sense of [@Levine-Morel].\
The corresponding generalized relative Grothendieck groups only with respect to proper morphisms in $\mB$ (and $\alp$ a corresponding additive invariant) are then denoted by $$K^{prop}_{(\alp)}(\mC_s \xrightarrow {S} \mB /T(-)), K^{prop}_{(\alp)}(\mC_s \xrightarrow {S} \mB /-) \quad \text{and} \quad K^{prop}_{(\alp)}(\mC_s /-).$$ Then one has a tautological group homomorphism (by just forgetting the properness condition) $$K^{prop}(\mC_s \xrightarrow {S} \mB /T(X)) \to K(\mC_s \xrightarrow {S} \mB /T(X)),$$ whose image is the subgroup of the generalized relative Grothendieck group $K(\mC_s \xrightarrow {S} \mB /T(X))$ generated by isomorphism classes $[(V, X, h)]$ with $h:S(V) \to T(X)$ being a *proper* map. If we assume that $S(V)$ is compact (resp. complete in the algebraic context) for every $V\in ob(\mC_s)$, then we have $$K^{prop}(\mC_s \xrightarrow {S} \mB /T(X)) = K(\mC_s \xrightarrow {S} \mB /T(X)).$$
Now our base category $\m B$ of spaces is not only a symmetric monoidal category with respect to the disjoint union $\sqcup$ (with unit the initial empty space $\emptyset$), but also with respect to the product of spaces $\times$ (with unit the terminal point space $\{pt\}$, given by $Spec(k)$ in the algebraic context). Moreover, these structures are compatible in the sense that $$(X\sqcup X')\times Y = (X\times Y) \sqcup (X'\times Y) \quad \text{and} \quad Y\times (X\sqcup X') = (Y\times X) \sqcup (Y\times X'),$$ with $\emptyset \times Y =\emptyset = Y\times \emptyset$. Similarly the class of proper (or projective) morphisms in $\m B$ (in the algebraic context) is stable under products $\times$. And we want to discuss the multiplicativity properties of our transformations $$\tau_{\alp}: K^{(prop)}_{\alp}(\mC_s \xrightarrow {S} \m B /-) \to \m H$$ associated to an additive invariant on objects in $ob(\mC_s)$ with values in a suitable additive functor $\m H : \mathcal B \to \mathcal {AB}$ on the category $\mB$, which may be functorial only with respect to proper (or projective) morphisms. Here we consider for simplicity only the most important case that $\mC_t =B$ and $T=\op{id}_{\mC_t}$. Then it is easy see the following:
\[mult\]
1. Assume that $\mC_s$ is also a symmetric monoidal category with respect to a product $\times$, such that $S: \mC=\mC_s \to \m B$ is strict monoidal with respect to $\sqcup$ as well as $\times$ (e.g. $S$ is the inclusion of a subcategory stable under $\sqcup$ and $\times$). Then $K^{(prop)}(\mC \xrightarrow {S} \m B /X)$ has a functorial bilinear cross product structure: $$\times : K^{(prop)}(\mC\xrightarrow {S} \mB /X) \times K^{(prop)}(\mC \xrightarrow {S} \mB /Y)\to K^{(prop)}(\mC \xrightarrow {S} \mB /X \times Y);$$ $$[(V, X, h)] \times [(W, Y, k)] := [(V \times W, X \times Y, h \times k)],$$ with $\times \circ (f_*\times g_*) = (f\times g)_*\circ \times$ for all (proper or projective) morphisms $f,g$ in $\mB$.
2. Assume that the additive functor $\m H : \mB \to \mathcal {AB}$ is endowed with a bilinear cross poduct $$\boxtimes: \mH(X)\times \mH(Y) \to \mH(X\times Y)$$ such that $\boxtimes \circ (f_*\times g_*) = (f\times g)_*\circ \boxtimes$ for all (proper or projective) morphisms $f,g$ in $\mB$. Consider an additive invariant $\alp$ on objects in $ob(\mC)$ with values in $\m H$, which is *multiplicative* in the sense that $$\alp(V \times V') = \alp(V)\boxtimes \alp(V')\quad \text{for all $V,V'\in ob(\mC)$}.$$ Then $K^{(prop)}_{\alp}(\mC \xrightarrow {S} \mB /-)$ also gets a functorial bilinear cross product structure in the same way as before and the associated natural transformation $\tau_{\alp}: K^{(prop)}_{\alp}(\mC\xrightarrow {S} \mB /-) \to \mH$ given in Theorem \[main\] commutes with the cross product, i.e., the following diagram commutes: $$\begin{CD}
K^{(prop)}_{\alp}(\mC \xrightarrow {S} \mB/X) \times K^{(prop)}_{\alp}(\mC \xrightarrow {S} \mB /Y ) @> \times >>
K^{(prop)}_{\alp}(\mC \xrightarrow {S} \mB /X \times Y ) \\
@V \tau_{\alp} \times \tau_{\alp} VV @VV \tau_{\alp} V \\
\mH(X) \times \mH(Y) @>> \boxtimes > \mH(X \times Y) \:.
\end{CD}$$
We only need to prove the commutativity of the last diagram, which follows from $$\begin{aligned}
\tau_{\alp} \left([(V, X, h)]_{\alp} \times [(W, Y, k)]_{\alp}\right) &=
\tau_{\alp} \left([(V \times W, X \times Y, h \times k)]_{\alp}\right)\\
&= ( h \times k)_*\left(\alp(V \times W)\right)\\
&= ( h \times k)_*\left(\alp(V) \boxtimes \alp(W)\right)\\
&= h_*(\alp(V))\boxtimes k_*(\alp(W))\\
&= \tau_{\alp} \left([(V, X, h)]_{\alp}\right) \boxtimes
\tau_{\alp} \left([(W, Y, k)]_{\alp}\right).\end{aligned}$$
Let us illustrate this result in some examples. First we consider the differential-topological context with $\m B = \mathcal {TOP}_{lc}$ the category of locally compact topological spaces, and $\mC$ the category $\mC_{(o)}^{\infty}$ or $\mC^{\infty}_{\bC}$ of all differentiable (oriented) or stable complex $\m C^{\infty}$-manifolds, with $S: \mC\to \mB$ the forget functor. Of course $S$ commutes with $\sqcup$ and $\times$ for $\mC=\mC^{\infty}$ the category of $\m C^{\infty}$-manifolds. Moreover, any such manifold $V$ has a fundamental class $[V]\in H_*^{BM}(V;\bZ_2)$ in Borel-Moore homology with $\bZ_2$-coefficients. And this fundamental class is additive and multiplicative: $$[V\sqcup V']= [V]+[V'] \quad \text{and} \quad [V\times V'].$$ When it comes to an oriented (or a stable complex) $\m C^{\infty}$-manifold $V$, then this has a fundamental class $$[V]\in H_*^{BM}(V;\bZ)$$ in Borel-Moore homology with $\bZ$-coefficients. And we view them as a category with the same morphisms as for the underlying $\m C^{\infty}$-manifolds, i.e. with $\m C^{\infty}$-maps between them (so that diffeomorphisms are the isomorphisms). Then the disjoint union $V\sqcup V'$ or product $V\times V'$ of two oriented (or stable complex) $\m C^{\infty}$-manifolds $V, V'$ can also be oriented (or given the structure of a stable complex) $\m C^{\infty}$-manifold. And there is a natural choice for this so that the fundamental class $[-]$ becomes additive and multiplicative as before. In this way we also get the symmetric monoidal structures $\sqcup$ and $\times$ on the category of oriented (or a stable complex) $\m C^{\infty}$-manifolds, with the forget functor $S$ to $\mathcal {TOP}_{lc}$ (or also to $\m C^{\infty}$) commuting with these structures. Also note that $$s_*[V\times V']= (-1)^{dim(V)\cdot dim(V')} [V'\times V]
\in H^{BM}_{dim(V)+dim(V')}(V'\times V;\bZ)$$ for two *connected* oriented manifolds $V,V'$ and $s: V\times V'\stackrel{\sim}{\to} V'\times V$ the symmetry isomorphism, since $H^{BM}_{*}(-;\bZ)$ is *graded-commutative* with respect to the usual cross product $\boxtimes$.
\[smooth-gen\] Let $\mC$ be the category $\mC^{\infty}_{(o)}$ or $\mC^{\infty}_{\bC}$ of all (oriented) or stable complex smooth manifolds. Consider a contravariant functorial characteristic class $$c\ell(E) \in H^*(-; \bZ_2) \quad \text{or} \quad c\ell(E) \in H^{2*}(-;R)$$ of (isomorphism classes of) real (oriented) or complex vector bundles, which is *multiplicative and normalized*, i.e.: $$c\ell(E \oplus F) = c\ell(E) \cup c\ell(F)\quad \text{and} \quad 1=c\ell^0(E)\in H^0(-;R).$$ For a smooth (oriented or stable complex) manifold $V$, let $$\alp(V) := c\ell(TV) \cap [V]\in H_*^{BM}(-;R).$$ Then the invariant $\alp$ is *additive and multiplicative*. By Theorem \[main\], there exists a unique natural transformation $$\tau_{c\ell}: K^{prop}_{\alp}(\mC \xrightarrow {S} \mTOP_{lc} /-) \to H^{BM}_*(-;R)$$ such that for a smooth (oriented or stable complex) manifold $V$ $$\tau_{c\ell}([(V, V, id_{V})]_{\alp}) = c\ell(TV) \cap [V].$$ And $\tau_{c\ell}$ is also multiplicative (by Proposition \[mult\]), i.e.: $$\tau_{c\ell}([(V, X, h)]_{\alp} \times [(W, Y, k)]_{\alp}) = \tau_{c\ell}([(V, X, h)]_{\alp}) \boxtimes \tau_{c\ell}([(W, Y, k)]_{\alp}).$$
The invariant $\alp$ is additive by the functoriality of $c\ell$ and the projection formula for the inclusions $i: V\to V\sqcup V'$ and $i': V'\to V\sqcup V'$: $$\begin{aligned}
\alp(V\sqcup V')&= c\ell(T(V\sqcup V'))\cap [V\sqcup V']\\
&= c\ell(T(V\sqcup V'))\cap (i_*[V]+ i'_*[V'])\\
&=i_*\left(i^*c\ell(T(V\sqcup V'))\cap[V]\right) +
i'_*\left(i'^*c\ell(T(V\sqcup V'))\cap[V']\right)\\
&= i_*\left(c\ell(TV)\cap[V]\right) +
i'_*\left(c\ell(T V')\cap[V']\right)\\
&= i_*\alp(V)+ i'_*\alp(V').\end{aligned}$$ The invariant $\alp$ is multiplicative by the functoriality and multiplicativity of $c\ell$: $$\begin{aligned}
\alp(V\times V')&= c\ell(T(V\times V'))\cap [V\times V']\\
&=\left( c\ell(TV)\boxtimes c\ell(TV')\right) \cap \left([V]\boxtimes [V']\right)\\
&=\left( c\ell(TV)\cap [V]\right) \boxtimes \left( c\ell(TV')\cap [V']\right)\\
&=\alp(V)\boxtimes \alp(V').\end{aligned}$$ Note that in the third equality there is no sign appearing, since we only consider even degree characteristic classes $c\ell$ in the context of oriented or stable complex manifolds.
\[or-preserving\] In the context of oriented manifolds in the above corollary, $(V, X,h)$ and $(V', X, h)$ are $\alp$-isomorphic if and only if the isomorphism $\phi: V \to V'$ (with $h=h'\circ S(\phi)$) is orientation preserving. Indeed, if $\phi: V \to V'$ is orientation preserving, then $S(\phi)_*[V] = [V']$ and $\phi^*TV' \simeq TV$ as oriented vector bundles. Hence we have $$\begin{aligned}
S(\phi)_*(\alp(V)) & = S(\phi)_*(c\ell(TV) \cap [V]) \\
& = S(\phi)_*(S(\phi)^*c\ell(TV') \cap [V])\\
& = c\ell(TV') \cap S(\phi)_*[V] \,\, \, \,\text {(by the projection formula)}\\
& = c\ell(TV') \cap [V'] = \alp(V').\end{aligned}$$ Hence $(V, X,h)$ and $(V', X, h)$ are $\alp$-isomorphic. Conversely, if $(V, X,h)$ and $(V', X, h)$ are $\alp$-isomorphic, then we have $S(\phi)_*(\alp(V)) = \alp(V')$, i.e., $S(\phi)_*(c\ell(TV) \cap [V]) = c\ell(TV') \cap [V']$. Since $c\ell$ is normalized, we have $c\ell = 1 + \cdots$, hence $c\ell(TV) \cap [V] = [V] + \text {lower dimensional classes}$. Therefore we have $$S(\phi)_*[V] = [V'].$$ Hence $\phi:V \to V'$ is orientation preserving.
Similarly $\alp$ is an isomorphism invariant in the context of unoriented manifolds. In the context of stable complex manifolds, $\alp$ is at least invariant under a diffeomorphism $\phi: V\to V'$ of stable complex manifolds who preserves the stable almost complex structure (and therefore also the orientation) in the sense that $\phi^*TV' \simeq TV$ as stable complex vector bundles.
If we consider in Corollary \[smooth-gen\] all *compact* (oriented) or stable complex smooth manifolds, then we get similar results for any *generalized homology theory* $\m H_*$, which has a corresponding fundamental class $[V]\in \m H_*(V)$ for a compact (oriented) or stable complex smooth manifold $V$, e.g. for a complex oriented (co)homology theory and $V$ a stable complex smooth manifold.\
Let us now switch to some counterparts in the algebraic geometric context, with $\mB=\m V_k$ the category of algebraic varieties (i.e. reduced separated schemes of finite type) over a base field $k$, and $S: \mC\to \m V_k$ the inclusion functor of a (full) subcategory $\mC$ stable under isomorphisms, disjoint union $\sqcup$ and product $\times$, with $\emptyset, Spec(k)\in ob(\mC)$. First we consider the subcategory $\mC= \m V_k^{sm}$ of smooth varieties. The proof of the following result is identical to that of Corollary \[smooth-gen\].
\[cor-alg\] Consider the cospan of categories $$\m V_k^{sm} \xrightarrow {S} \m V_k \xleftarrow {id_{\m V_k}} \m V_k,$$ together with a contravariant functorial characteristic class $$c\ell(E) \in CH^*(-)\otimes R\quad \text{or $\quad c\ell(E) \in H^{2*}(-;R)\:\:$ for $k=\bC$}$$ of (isomorphism classes of) algebraic vector bundles, which is *multiplicative* in the sense that $$c\ell(E) = c\ell(E') \cup c\ell(E'')$$ for any short exact sequence $0\to E'\to E\to E''\to 0$ of such vector bundles. Here $CH^*(-)$ is the operational Chow cohomology group of [@Fulton-book]. For a smooth algebraic manifold $V$, let $$\alp(V) := c\ell(TV) \cap [V]\in \mH_*(-)\otimes R,$$ with $[V]$ the fundamental class of $V$ for $\mH_*=CH_*$ the Chow group or $H_{2*}^{BM}$ the even degree Borel-Moore homology in the case of $k=\bC$. Then the isomorphism invariant $\alp$ is *additive and multiplicative*. By Theorem \[main\], there exists a unique natural transformation $$\tau_{c\ell}: K^{prop}(\m V_k^{sm} \xrightarrow {S} \m V_k /-) \to \mH_*(-)\otimes R$$ such that for a smooth algebraic manifold $V$ $$\tau_{c\ell}([(V, V, id_{V})]) = c\ell(TV) \cap [V].$$ And $\tau_{c\ell}$ is also multiplicative (by Proposition \[mult\]), i.e.: $$\tau_{c\ell}([(V, X, h)] \times [(W, Y, k)]) = \tau_{c\ell}([(V, X, h)]) \boxtimes \tau_{c\ell}([(W, Y, k)]).$$
\[rem-LM\] If we only consider projective morphisms and (pure dimensional) quasi-projective smooth varieties, then a similar result holds for $\mH_*$ an *oriented Borel-Moore weak homology theory* in the sense of [@Levine-Morel] and $c\ell$ a multiplicative characteristic class as in [@Levine-Morel §4.1.8], which is, for a line bundle $L$, given by a normalized power series in the first Chern class operator of $L$ with respect to $\mH_*$: $$c\ell(L)=f(\tilde{c}_1(L)), \quad \text{with $f(t)\in 1+t\cdot \mH_*(pt)[[t]]$.}$$ Here the fundamental class of a quasi-projective smooth variety $V$ of pure dimension $d$ is defined as $$[V]:=k^*1_{pt} \in \m H_d(V) \quad \text{for $k: V\to pt$ the constant smooth morphism.}$$
Next we consider the subcategory $ \mC= \m V^{(l)pd}_k$ of (locally) pure-dimensional algebraic varieties over $k$. Let us first recall the universal property of the *Nash blow-up* $\nu: \widehat X \to X$ of a pure $d$-dimensional algebraic variety $X\in Obj(V^{pd}_k)$, with $\widehat {TX}$ the *tautological Nash tangent bundle* over $\widehat X$: Let $\pi: \overline{X}\to X$ be a proper birational map with a surjection $\pi^*\Omega^1_X\to \overline{\Omega}$ to a locally free sheaf $\overline{\Omega}$ of rank $d$ on $\overline{X}$. Then the Nash blow-up $\nu: \widehat X \to X$ is universal in the sense that $\pi: \overline{X}\to X$ factors through $\pi': \overline{X}\to \widehat X$, with $\overline{\Omega}\simeq \pi'^*\widehat \Omega$, where the tautological Nash tangent bundle $\widehat {TX}$ over $\widehat X$ corresponds to the dual of $\widehat \Omega$.
Let $c\ell$ be a functorial characteristic class of algebraic vector bundles as in Corollary \[cor-alg\]. For a pure $d$-dimensional algebraic variety $X\in Obj(V^{pd}_k)$, the *$c\ell$-Mather homology class* $c\ell^{Ma}_*(X)\in \mH_*(X)\otimes R$ is defined to be $$c\ell^{Ma}_*(X):= \nu_*(c\ell(\widehat {TX}) \cap [\widehat X])=\pi_*(c\ell(\overline{TX}) \cap [\overline{X}]).$$ Here $\pi: \overline{X}\to X$ is any *proper birational* map with a surjection $\pi^*\Omega^1_X\to \overline{\Omega}$ to a locally free sheaf $\overline{\Omega}$ of rank $d$ on $\overline{X}$, with the vector bundle $\overline{TX}$ corresponding to the dual of $\overline{\Omega}$. This definition is extended to a locally pure-dimensional variety $X\in Obj(V^{lpd}_k)$ by additivity over the connected components of $X$.
Note that the second equality in the definition above follows from the projection formula by $\overline{TX}\simeq \pi'^*\widehat {TX}$ and $\pi'_*[\overline{X}]=[\widehat X]$, since $\pi': \overline{X}\to \widehat X$ is a proper birational map (see [@Fulton-book Example 4.2.9(b)] in the case of the Chern-Mather class $c^{Ma}$ corresponding to $c\ell=c$ the Chern class).
\[mather\] Consider the cospan of categories $$\m V_k^{(l)pd} \xrightarrow {S} \m V_k \xleftarrow {id_{\m V_k}} \m V_k,$$ together with a contravariant functorial characteristic class $$c\ell(E) \in CH^*(-)\otimes R\quad \text{or $\quad c\ell(E) \in H^{2*}(-;R)\:\:$ for $k=\bC$}$$ of (isomorphism classes of) algebraic vector bundles, which is *multiplicative*. Here $CH^*(-)$ is the operational Chow cohomology group [@Fulton-book Definition 17.3]. For a (locally) pure-dimensional algebraic variety $V$, let $$\alp(V) := c\ell^{Ma}_*(V)\in \mH_*(V)\otimes R,$$ for $\mH_*=CH_*$ the Chow group or $H_{2*}^{BM}$ the even degree Borel-Moore homology in the case of $k=\bC$.
1. The isomorphism invariant $\alp$ is *additive and multiplicative*. By Theorem \[main\], there exists a unique natural transformation $$\tau_{c\ell^{Ma}_*}: K^{prop}(\m V_k^{(l)pd} \xrightarrow {S} \m V_k /-) \to \mH_*(-)\otimes R$$ such that for a (locally) pure-dimensional algebraic variety $V$ $$\tau_{c\ell^{Ma}_*}([(V, V, id_{V})]) = c\ell^{Ma}_*(V).$$ And $\tau_{c\ell^{Ma}_*}$ is also multiplicative (by Proposition \[mult\]), i.e.: $$\tau_{c\ell^{Ma}_*}([(V, X, h)] \times [(W, Y, k)]) = \tau_{c\ell^{Ma}_*}([(V, X, h)]) \boxtimes \tau_{c\ell^{Ma}_*}([(W, Y, k)]).$$
2. When $c\ell = c$ is the Chern class, then the following diagram commutes for $k$ of characteristic zero: $$\xymatrix{K^{prop} (\m V_k^{(l)pd} \xrightarrow {S} \m V_k/X) \ar[dr]_ {\tau_{c^{Ma}_*}}\ar[rr]^ {\m{E}u} && F(X) \ar[dl]^{c_*^{Mac}}\\& \mH_*(X)\:.}$$ Here the multiplicative natural transformation $\m{E}u: K^{prop}(\m V_k^{(l)pd} \xrightarrow {C} \m V_k /X)\to F(X)$ to the group of constructible functions is defined by the isomorphism invariant $\alp(V):=Eu_V$, where $Eu_V$ is the local Euler obstruction of the (locally) pure-dimensional variety $V$. This invariant is also additive and multiplicative.
\(1) That the isomorphism invariant $\alp(V) := c\ell^{Ma}_*(V)$ is additive follows from the fact, that it commutes with restriction to open subsets (e.g. a connected component). For the multiplicativity we can then assume that $V,V'$ are pure dimensional. Let $\pi: \overline{V}\to V$ be a proper birational map with a surjection $\pi^*\Omega^1_X\to \overline{\Omega}$ to a locally free sheaf $\overline{\Omega}$, and similarly for $V'$. Then $\pi\times \pi': \overline{V}\times \overline{V'}\to V\times V'$ is a proper birational map with a surjection $$(\pi\times \pi')^*\Omega^1_{V\times V'}\simeq \pi^*\Omega^1_V\boxtimes \pi'^*\Omega^1_{V'}\to \overline{\Omega}\boxtimes \overline{\Omega'}$$ so that $$\begin{aligned}
c\ell^{Ma}_*(V\times V')&= (\pi\times \pi')_*\left( c\ell (\overline{TV}\boxtimes \overline{TV'})\cap [\overline{V}\times \overline{V'}] \right)\\
&= (\pi\times \pi')_*\left( (c\ell(\overline{TV})\cap [\overline{V}])\boxtimes ( c\ell(\overline{TV'})\cap [\overline{V'}]) \right)\\
&= \left( \pi_*(c\ell(\overline{TV})\cap [\overline{V}])\right) \boxtimes \left( \pi'_*(c\ell(\overline{TV'})\cap [\overline{V'}] )\right)\\
&= c\ell^{Ma}_*(V) \boxtimes c\ell^{Ma}_*(V').\end{aligned}$$ (2) follows from the construction [@MacPherson] of the Chern class transformation $$c_*^{Mac}: F(X) \to H^{BM}_{2*}(X;\bZ)$$ for the case $k=\bC$. MacPherson defined $c_*^{Mac}(Eu_V) = (i_V)_*c_*^{Ma}(V)$ for $i_V: V\to X$ the inclusion of a pure-dimensional subvariety, with $Eu_V$ his famous local Euler obstruction. Namely, the “constructible function" counterpart of the Chern–Mather homology class has to be the local Euler obstruction, which is one of his key observations. This is extended to locally pure-dimensional subvarieties by additivity over connected components, e.g. the local Euler obstruction is then by definition the sum of the local Euler obstructions of all connected components. The algebraic counterpart of the MacPherson Chern class transformation $$c_*^{Mac}: F(X) \to CH_*(X)$$ for a base field $k$ of characteristic zero was constructed in [@Ke] (at least if $X$ is embeddable into a smooth variety. The general case can be reduced to this using the method of Chow envelopes as in [@Fulton-book Chapter 18.3]). Moreover, Kennedy also explained in [@Ke] that $Eu_V\in F(V)$ is a constructible function, where he used the algebraic definition of the local Euler obstruction as in (due to Gonzalez-Sprinberg and Verdier): $$Eu_V(p)=\int_{\pi^{-1}(p)} c \left(\overline{TV}|_ {\pi^{-1}(p)} \right) \cap s \left(\pi^{-1}(p),\overline{V} \right),$$ with $\pi: \overline{V}\to V$ and $\overline{TV}$ are as in the definition of the $c\ell$-Mather homology classes. Here $s\bigl(\pi^{-1}(p),\overline{V} \bigr)$ is the Segre class of the fiber $\pi^{-1}(p)$ in $\overline{V}$ in the sense of [@Fulton-book Chapter 4.2]. Then the multiplicativity of $Eu_V$ follows as in (1) using [@Fulton-book Example 4.2.5]: $$s \Bigl(\pi^{-1}(p)\times \pi'^{-1}(p'),\overline{V}\times \overline{V'} \Bigr)=s \bigl(\pi^{-1}(p),\overline{V} \bigr)\boxtimes s \bigl(\pi'^{-1}(p'),\overline{V'} \bigr).$$ Here the bilinear (functorial) cross product $$\boxtimes: F(X)\times F(Y)\to F(X\times Y)$$ is just defined by $\beta\boxtimes \beta' ((p,p')):=\beta(p)\cdot \beta'(p')$.
Assume that the base field $k$ is of characteristic zero.
1. Using resolution of singularities one can show that for a given algebraic variety $X$ there are finitely many irreducible subvarieties $V$’s and integers $a_V$’s such that $$\jeden_X = \sum_{V \subset X} a_V Eu_V \quad \text{, thus} \quad c_*^{Mac}(\jeden_X) = \sum_{V \subset X} a_V c_*^{Ma}(V).$$ Whether $X$ is singular or not, $c_*^{Mac}(X):=c_*^{Mac}(\jeden_X)$ is called MacPherson’s Chern class or Chern–Schwarz–MacPherson class of $X$ (see [@BrSc; @Ke; @MacPherson; @Schw1; @Schw2]). For $X$ complete, it follows from the naturality of the transformation $c_*^{Mac}$ with respect to the proper constant map $X\to pt$, that the degree of the $0$-dimensional component of $c_0^{Mac}(X)$ is equal to the Euler–Poincaré characteristic: $$\int_X c_0^{Mac}(X) = \chi(X).$$
2. Similarly, the degree of the $0$-dimensional component of the Chern–Mather class $c_*^{Ma}(X)$ for $X$ pure-dimensional and complete is the Euler–Poincaré characteristic of $X$ weighted by the local Euler obstruction $Eu_X$: $$\int_X c_0^{Ma}(X) = \chi(X;Eu_X).$$ For $X$ a connected *complex affine* algebraic variety of pure dimension, the *global Euler obstruction $Eu(X)$* introduced and studied in [@STV] is a suitable “localization” of the $0$-dimensional component of the Chern–Mather class $c_*^{Ma}(X)$: $$H_0(X;\bZ)\simeq \bZ \ni Eu(X)\mapsto c_0^{Ma}(X) \in H^{BM}_0(X;\bZ)$$ under the natural map $H_0(X;\bZ)\to H^{BM}_0(X;\bZ) $.
3. The above “$c\ell$-Mather class" transformation $\tau_{c\ell^{Ma}_*}: K^{prop}(\m V_k^{(l)pd} \xrightarrow {S} \m V_k /-) \to \mH_*(-)\otimes R$ could be considered as a very naïve theory of characteristic classes of possibly singular algebraic varieties (in any characteristic).
So far we dealt with the covariance and multiplicativity of the functor $K_{(\alp)}(\mC_s \xrightarrow {S} \mB /T(-))$. Next we discuss the contravariance with respect to (suitable) “smooth morphisms”, where we start with the algebraic geometric context with $\mB=\m V_k$ the category of algebraic varieties over $k$ and $S: \mC\to \m V_k$ the inclusion functor of the (full) subcategories $\mC= \m V_k^{sm}$ resp. $\m V_k^{lci}$ of *smooth* resp. *local complete intersection* varieties.
Here $X$ is called a local complete intersection, if it has a regular closed embedding $i: X\to M$ into a smooth variety $M$ (i.e. the constant morphism $X\to pt$ is a local complete intersection morphism in the sense of [@Fulton-book Chapter 6.6]). Then $X$ has an intrinsic virtual tangent bundle $$TX:=i^*TM - N_XM \in K^0(X),$$ with $N_XM$ the normal bundle of the regular embedding $i: X\to M$, i.e. $TX\in K^0(X)$ doesn’t depend on the choice of this embedding (compare [@Fulton-book Appendix B.7.6]). Of course any smooth variety $M$ is local complete intersection, with $TM$ (the class of) the usual tangent bundle $TM$ (just choose $i=id_M: M\to M$). Note that the (virtual) tangent bundle commutes with restriction to open subsets (e.g. connected components). Similarly, if $f: X\to Y$ is a smooth morphism with $Y$ smooth (resp. a local complete intersection), then also $X$ is smooth (resp. a local complete intersection), and in the smooth context we have a short exact sequence of vector bundles $$\label{eq-s1}
0\to T_f \to TX \to f^*TY\to 0.$$ with $T_f$ the bundle of tangents to the fiber of the smooth morphism $f$. In particular $$\label{eq-s2}
TX = f^*TY + T_f \in K^0(X),$$ and this equality in the Grothendieck group $K^0(-)$ of algebraic vector bundles even holds for a smooth morphism $f: X\to Y$ between local complete intersections with $TX$ resp. $TY$ the corresponding virtual tangent bundle (compare e.g. with as well as [@Fulton-book Appendix B.7]). Finally, the class of smooth morphisms is stable under base-change, with $T_{f'}\simeq h'^*T_f$ for a fiber square as in the following
\[lemma\] The functor $K^{prop}(\mC \xrightarrow {S} \m V_k /-)$, with $\mC= \m V_k^{sm}$ (resp. $\m V_k^{lci}$) the subcategory of smooth (resp. local complete intersection) varieties, becomes a contravariant functor for smooth morphisms on the category $\m V_k$, where for a smooth morphism $f:X \to Y$ the pullback homomorphism $$f^*: K^{prop}(\mC \xrightarrow {S} \m V_k /Y)\to K^{prop}(\mC \xrightarrow {S} \m V_k /X)$$ is defined by $$f^*([(V, Y, h)]) := [V',X,h'],$$ using the fiber square $$\CD
V'@> {f'} >> V\\
@V h' VV @VV h V\\
X@> {f} >> Y. \endCD$$
Of course here we also use the fact that taking such fiber squares commutes with disjoint unions in $V$ resp. $V'$.
\[Verdier-RR\] Consider the cospan of categories $$\mC \xrightarrow {S} \m V_k \xleftarrow {id_{\m V_k}} \m V_k,$$ with $\mC= \m V_k^{sm}$ (resp. $\m V_k^{lci}$) the subcategory of *smooth* (resp. *local complete intersection*) varieties. Let $c\ell$ be a contravariant functorial characteristic class $$c\ell(E) \in CH^*(-)\otimes R\quad \text{or $\quad c\ell(E) \in H^{2*}(-;R)\:\:$ for $k=\bC$}$$ of (isomorphism classes of) algebraic vector bundles, which is multiplicative (and normalized in case $\mC= \m V_k^{lci}$, so that it can also be defined on virtual vector bundles in $K^0(-)$). For a smooth (resp. local complete intersection) variety $V$, let $$\alp(V) := c\ell(TV)\cap [V]\in \mH_*(V)\otimes R$$ be the corresponding (virtual) characteristic homology class, with $\mH_*=CH_*$ the Chow group or $H_{2*}^{BM}$ the even degree Borel-Moore homology in case $k=\bC$.
The isomorphism invariant $\alp$ is *additive*, so that by Theorem \[main\] there exists a unique natural transformation $$\tau_{c\ell}: K^{prop}(\mC \xrightarrow {C} \m V_k /-) \to \mH_*(-)\otimes R,$$ such that for a smooth (resp. local complete intersection) algebraic variety $V$ $$\tau_{c\ell}([(V, V, id_{V})]) = c\ell(TV)\cap [V].$$ Then this natural transformation $\tau_{c\ell}$ satisfies the following Verdier-type Riemann–Roch formula: For a smooth morphism $f:X \to Y$ the following diagram commutes: $$\CD
K^{prop}(\mC \xrightarrow {S} \m{V}_k /Y) @> {\tau_{c\ell}} >> \mH_*(Y)\otimes R\\
@V f^* VV @VV c\ell(T_f) \cap f^* V\\
K^{prop}(\mC \xrightarrow {S} \m{V}_k /X) @>> {\tau_{c\ell}} > \mH_*(X)\otimes R. \endCD$$
On one hand we have: $$\begin{aligned}
\tau_{c\ell}(f^*([(V, Y, h)])) & = \tau_{c\ell}([(V', X, h')]) \\
& = h'_*(c\ell(TV') \cap [V']). \end{aligned}$$ On the other hand we have $$\begin{aligned}
& c\ell(T_f) \cap f^*(\tau_{c\ell}([(V, Y, h)]) \\
& = c\ell(T_f) \cap f^* (h_* (c\ell (TV) \cap [V]))\end{aligned}$$
For a fiber square $$\CD
V' @> f'>> V \\
@V h' VV @VV h V\\
X@>> f > Y \endCD$$ with $h:V \to Y$ proper and $f:X \to Y$ smooth, we have the base change identity (see [@Fulton-book Proposition 1.7]): $$f^*h_* = h'_*{f'}^*: \mH_*(V)\otimes R \to \mH_*(X)\otimes R \:.$$ Hence the above equality continues as follows: $$\begin{aligned}
& = c\ell(T_f) \cap h'_*{f'}^* (c\ell (TV) \cap [V]))\\
& = h'_* ({h'}^*c\ell(T_f) \cap {f'}^* (c\ell (TV) \cap [V])) \, \text {(by the projection formula)}\\
& = h'_* (c\ell(T_{f'}) \cap {f'}^* (c\ell (TV) \cap [V])) \, \text {(by $T_{f'}\simeq h'^*T_f$)}\\
& = h'_* \left (c\ell(T_{f'}) \cap (c\ell ({f'}^*TV) \cap {f'}^*[V]) \right ) \, \text {(by functoriality of $c\ell(-)\cap$)}\\
& = h'_* \left( (c\ell(T_{f'}) \cup c\ell ({f'}^*TV)) \cap [V']\right)\\
&= h'_* \left (c\ell(TV') \cap [V'] \right ) \end{aligned}$$ by the multiplicativity of $c\ell$ and (\[eq-s1\]) resp. (\[eq-s2\]). Of course we also used the relation $[V']={f'}^*[V]$ for the fundamental classes. Therefore we get that $$\tau_{c\ell}(f^*([(V, Y, h)])) = c\ell(T_f) \cap f^*(\tau_{c\ell}([(V, Y, h)]),$$ and the above diagram of the theorem commutes.
\[rem-LM2\] If we consider only projective morphisms and (pure dimensional) quasi-projective local complete intersection (resp. smooth) varieties, then a similar Verdier-type Riemann–Roch formula holds for $\mH_*$ an *oriented Borel-Moore (weak) homology theory* in the sense of [@Levine-Morel] and $c\ell$ a multiplicative characteristic class as in [@Levine-Morel §7.4.1 resp. §4.1.8], which is, for a line bundle $L$, given by a normalized power series in the first Chern class operator of $L$ with respect to $\mH_*$: $$c\ell(L)=f(\tilde{c}_1(L)), \quad \text{with $f(t)\in 1+t\cdot \mH_*(pt)[[t]]$.}$$ Here the fundamental class of a quasi-projective local complete intersection (resp. smooth) variety $V$ of pure dimension $d$ is defined as $$[V]:=k^*1_{pt} \in \m H_d(V) \quad \text{for $k: V\to pt$ the constant local complete intersection (resp. smooth) morphism}$$ so that $[V']={f'}^*[V]$ for a smooth morphism $f': V'\to V$ by functoriality of $f'^*$.
Assume that the base field $k$ is of characteristic zero.
1. The interesting thing about the motivic Hirzebruch class or MacPherson Chern class transformation $${T_y}_*: K^{prop}(\m V^{sm}_k\xrightarrow {S} \m V_k /-) \to \m H_*(-)\otimes \bQ[y]
\quad \text{or} \quad
c^{Mac}_*: K^{prop}(\m V^{sm}_k\xrightarrow {S} \m V_k /-) \to \m H_*(-)$$ is for example, that in the above discussions, the na" ive relative Grothendieck group $K^{prop}(\m V^{sm}_k\xrightarrow {S} \m V_k /X)$ can be replaced by the much smaller and more interesting relative Grothendieck group $K_0(\m V_k/X)$, by imposing one more *additivity relation*, as recalled in the introduction: $$[V \xrightarrow h X] = [W \xrightarrow {h|_W} X] + [V \setminus W \xrightarrow {h|_{V \setminus W}} X],$$ with $W \subset V$ is a closed subvariety of $V$. So these two transformations factorize over the tautological surjective transformation $$K^{prop}(\m V^{sm}_k\xrightarrow {S} \m V_k /-) \to K_0(\m V_k/-).$$ In particular the multiplicativity result of Corollary \[cor-alg\] and the Verdier-type Riemann–Roch formula of Theorem \[Verdier-RR\] are true for the motivic Hirzebruch class and the MacPherson Chern class transformation (compare [@BSY]).
2. The corresponding Hirzebruch class $T_{y*}(X)=T_{y*}(id_X)$ and MacPherson Chern class $c_*^{Mac}(X)=c_*^{Mac}(id_X)$ of an algebraic variety $X$ is not only invariant under an isomorphism, but also under a proper *(geometric) bijection*: $$f_*\left(T_{y*}(X)\right)=T_{y*}(Y) \quad \text{and} \quad f_*\left(c_*^{Mac}(X)\right)=c_*^{Mac}(Y)$$ for a proper morphism $f: X\to Y$ such that the induced map $f: X(\bar{k}) \to Y(\bar{k})$ is a bijection of sets (with $\bar{k}$ an algebraic closure of $k$), since then $f_*([id_X])=[id_Y]$ by Noetherian induction using additivity and generic smoothness of $f$. Of course for MacPherson’s Chern class $c_*^{Mac}(X)=c_*^{Mac}(\jeden_X)$ this also follows from $f_*\jeden_X=\jeden_Y\in F(Y)$ for a proper (geometric) bijection $f: X\to Y$. Similarly, Baum–Fulton–MacPherson’s Todd class $td^{BFM}_*(X)=td^{BFM}_*([\m O_X])$ is invariant under a such a proper (geometric) bijection (for $k$ of any characteristic), since $f_*[\m O_X]=[\m O_Y]\in G_0(Y)$. Finally, Goresky–MacPherson’s or Cappell–Shaneson’s homology $L$-class $L_*(X)=L_*^{CS}([\m {IC}_X])$ of a (locally) pure-dimensional compact complex algebraic variety $X$ is not only invariant under a proper bijection, but more generally under taking the normalization of $X$.
Let us finish this section with a counterpart in the differential-topological context, with $\m B=\m C^{\infty}$ the category of $\m C^{\infty}$-manifolds and $S: \mC\to \m C^{\infty}$ the forget functor from the the category $\mC = \mC_{(o)}^{\infty}$ or $\mC^{\infty}_{\bC}$ of all differentiable (oriented) or stable complex $\m C^{\infty}$-manifolds.
\[lemma2\] The functor $K^{prop}(\mC \xrightarrow {S} \m C^{\infty}/-)$ for $\mC = \mC_{(o)}^{\infty}$ or $\mC^{\infty}_{\bC}$ becomes a contravariant functor for a ((complex) oriented) submersion $f:X \to Y$ of $\m C^{\infty}$-manifolds. Here the pullback homomorphism $$f^*: K^{prop}(\mC \xrightarrow {S} \m C^{\infty} /Y)\to K^{prop}(\mC \xrightarrow {C} \m C^{\infty}/X)$$ is defined by $$f^*([(V, Y, h)]) := [V',X,h'],$$ using the fiber square $$\CD
V'@> {f'} >> V\\
@V h' VV @VV h V\\
X@> {f} >> Y, \endCD$$ with $TV'$ oriented (or given a stable complex structure) by the short exact sequence $$0\to T_{f'} \to TV' \to f'^*TV\to 0,$$ if $TV$ and the bundle of tangents to the fibers $T_f$ (and therefore also $T_{f'}\simeq h'^*T_f$) are oriented (or stable complex).
Then the proof of the following result is identical to that in the algebraic geometric context:
\[Verdier-RR-smooth\] Consider the cospan of categories $$\mC \xrightarrow {S} \m C^{\infty} \xleftarrow {id_{\m C^{\infty}}} \m C^{\infty},$$ with $\mC = \mC_{(o)}^{\infty}$ or $\mC^{\infty}_{\bC}$ the category of all differentiable (oriented) or stable complex $\m C^{\infty}$-manifolds. Let $c\ell$ be a contravariant functorial multiplicative and normalized characteristic class $$c\ell(E) \in H^*(-;\bZ_2)\quad \text{or} \quad c\ell(E) \in H^{2*}(-;R)$$ of (isomorphism classes of) real (oriented) or complex vector bundles. For a smooth (oriented or stable complex) manifold $V$, let $$\alp(V) := c\ell(TV)\cap [V]\in H^{BM}_*(V;R).$$ The invariant $\alp$ is *additive*, so that by Theorem \[main\] there exists a unique natural transformation $$\tau_{c\ell}: K^{prop}_{\alp}(\mC \xrightarrow {S} \m C^{\infty}/-) \to H^{BM}_*(-;R),$$ such that for a smooth (oriented or stable complex) manifold $V$ $$\tau_{c\ell}([(V, V, id_{V})]) = c\ell(TV)\cap [V].$$ Then this natural transformation $\tau_{c\ell}$ satisfies the following Verdier-type Riemann–Roch formula: For a ((complex) oriented) submersion $f:X \to Y$ of $\m C^{\infty}$-manifolds, the following diagram commutes: $$\CD
K^{prop}_{\alp}(\mC \xrightarrow {S} \m C^{\infty}/Y) @> {\tau_{c\ell}} >> H^{BM}_*(Y;R)\\
@V f^* VV @VV c\ell(T_f) \cap f^* V\\
K^{prop}_{\alp}(\mC \xrightarrow {S} \m C^{\infty}/X) @>> {\tau_{c\ell}} > H^{BM}_*(X;R). \endCD$$
Examples
========
In this last section we discuss some results, questions and problems related to some very specific examples, which also fit with our natural transformations associated to some additive (homology) invariants.
The case of the fundamental class
---------------------------------
Let us first consider taking the fundamental class $[-]\in H^{BM}_*(-;R)$ for $R=\bZ_2$ (or $R=\bZ$) on the category $\mC^{\infty}_{(o)}$ of (oriented) $C^{\infty}$-manifolds. The fundamental class $\alp:=[-]$ is certainly an additive and multiplicative homology class and by Corollary \[smooth-gen\] we have a unique natural transformation $$\tau_{[-]}: K^{prop}_{\alp}(\mC \xrightarrow {S} \mTOP_{lc} /-) \to H^{BM}_*(-;R)$$ such that $\tau_{[-]}([(V, V, id_{V})]) = [V]$ for a smooth (oriented) manifold $V$. Here an $\alp$-isomorphism $\phi: V\to V'$ is just an (orientation preserving) diffeomeorphism by Remark \[or-preserving\].
Of course such an additive (and multiplicative) fundamental homology class $\alp:=[-]$ is also available on bigger categories of spaces, like the categories $\mC^{0}_{(o)}$ of (oriented) topological manifolds, or $\m C^{\text{pseudo}}_{(o)}$ the category of (oriented) triangulated or topological pseudo-manifolds, whose morphisms are by definition just continuous maps of the underlying topological spaces (so that an isomorphism is nothing but a homeomorphism). If we restrict ourselfes to compact spaces (as classically often done), then these fundamental classes live in usual homology $H_*(-;R)$ for $R=\bZ_2$ (or $R=\bZ$ in the oriented context). Moreover such an additive and multiplicative fundamental class $\alp=[-]\in H_*(-;R)$ is also available for the category $\m C^{\text{Poincar\'e}}_{(o)}$ of finite (oriented) Poincaré complexes, i.e., topological spaces $V$ (like homology-manifolds) which satisfy Poincaré duality $$\cap [V]: H^*(V;R)\stackrel{\sim}{\to} H_*(V;R)$$ for a suitable fundamental class $[V]\in H_*(V;R)$ . So we can turn these categories $\m C$ as before into symmetric monoidal categories with respect to disjoint unions $\sqcup$ (and also products $\times$) such that the forgetful functor $\frak f: \m C\to \mTOP_{lc}$ is strictly monoidal. Moreover, an $\alp$-isomorphism $\phi: V\to V'$ in this context is just an (orientation preserving) homeomorphism.\
Then the classical *Steenrod’s realization problem* can be reinterpreted as the problem of asking for the surjectivity of the group homomorphism $\tau_{[-]}: K_{\alp}(\mC \xrightarrow {\frak f} \mTOP /X) \to H_*(X;R)$ for a topological space $X$. Here the following results are well known (see [@Rudyak2; @Sullivan2]).
\[rudyak\]
1. ([@Thom] and [@Rudyak Chapter IV, Theorem 7.33]) $$\tau_{[-]}: K(\mC^{\infty}_c \xrightarrow {\frak f} \mTOP /X) \to H_*(X;\bZ_2)$$ is surjective, i.e. every $\bZ_2$-homology class can be realized by a compact smooth manifold.
2. ([@Thom] and [@Rudyak Chapter IV, Theorem 7.37]) The following composed map is surjective: $$\tau_{[-]}: K_{\alp}(\mC^{\infty}_{co} \xrightarrow {\frak f} \mTOP /X) \to H_*(X;\bZ) \stackrel{proj.}{\to}
\bigoplus _{0 \leq i \leq 6} H_i(X;\bZ),$$ i.e. every $\bZ$-homology class in degree $i\leq 6$ can be realized by a compact oriented smooth manifold.
3. ([@Levitt]) The following composed map is surjective: $$\tau_{[-]}: K_{\alp}(\m C^{\text{Poincar\'e}}_{co} \xrightarrow {\frak f} \mTOP /X) \to H_*(X;\bZ) \stackrel{proj.}{\to}
\bigoplus _{i \not =3 } H_i(X;\bZ),$$ i.e. every $\bZ$-homology class in degree $i\neq 3$ can be realized by a finite oriented Poincaré complex.
4. ([@Sullivan] and [@Rudyak Chapter VIII, Example 1.25(a)]) $$\tau_{[-]}: K_{\alp}(\m C^{\text{pseudo}}_{co} \xrightarrow {\frak f} \mTOP /X) \to H_*(X;\bZ)$$ is surjective, i.e. every $\bZ$-homology class can be realized by a compact oriented pseudo-manifold.
The case of the Stiefel–Whitney class
-------------------------------------
Let $V$ be a differentiable manifold and let $c\ell_*(V) \in H^{BM}_*(V, \bZ_2)$ be the Poincaré dual $c\ell(TV) \cap [V]$ of a multiplicative and normalized functorial characteristic class $c\ell(E)\in H^*(-, \bZ_2)$ of (isomorphism classes of) real vector bundles. $\alp(V):=c\ell_*(V)$ is clearly an additive (and multiplicative) homology class, and we have by Corollary \[smooth-gen\] a unique natural transformation $$\tau_{c\ell_*}: K^{prop}(\mC^{\infty} \xrightarrow {S} \mTOP_{lc} /-) \to H^{BM}_*(-;\bZ_2)$$ such that $\tau_{c\ell_*}([(V, V, id_{V})]) = c\ell_*(V)$ for a smooth manifold $V$. In particular the Stiefel–Whitney class $c\ell=w$ is a typical one.\
If we restrict ourselves to the category $\m V_{\bR}$ of real algebraic varieties and let $\m V^{sm}_{\bR}$ be its full subcategory of smooth real algebraic varieties, then we have a more geometric “realization" of the natural transformation $w_*$ on the category $\m V_{\bR}$ through *$\bZ_2$-valued semi-algebraic constructible functions*: $$\xymatrix{K^{prop} (\m V^{sm}_{\bR} \xrightarrow {\iota} \mathcal {V_{\bR}} /X) \ar[dr]_ {w_*}\ar[rr]^ {const} && F(X;\bZ_2) \ar[dl]^{w_*}\\&
H^{BM}_*(X, \bZ_2)\:.}$$ Here the multiplicative natural transformation $const$ is defined by the isomorphism invariant $\alp(V):=1_V$, which is additive and multiplicative. The Stiefel-Whitney class transformation $w_*: F(X;\bZ_2)\to H^{BM}_*(X, \bZ_2)$ was first constructed by Sullivan [@Sullivan] in the pl-context. For a new approach in the semi-algebraic (or subanalytic) context through the theory of “conormal or characteristic cycles” see [@FMc].
\[Thom-Whitney\] For a compact topological manifold $V$, Thom constructed a Whitney class using a relation with Steenrod squares [@Thom-7] (see [@MacPherson2]). Let us denote this Thom-Whitney class in homology by $w_*^{Th}(V) \in H_*(V;\bZ_2)$, which for a compact smooth manifold $V$ agrees with the Stiefel-Whitney class $w_*(V)$ above. Then also $\alp(V):= w_*^{Th}(V)$ is an additive (and multiplicative) invariant, so that we have a natural transformation $$\tau_{w_*}^{Th}: K(\mC^{0}_c \xrightarrow {\frak f} \mTOP /-) \to H_*(-;\bZ_2)$$ defined by $$\tau_{w_*}^{Th}([(V, X, h)]) = h_*w_*^{Th}(V).$$ If we consider the above Whitney class transformation $$w_*: K(\mC^{\infty}_c \xrightarrow {\frak f} \mTOP/-) \to H_*(-;\bZ_2)$$ for compact smooth $C^{\infty}$-manifolds, then this is surjective by Theorem \[rudyak\](i), since $w_*(V)=[V]+ \dots$ equals the fundamental class plus lower order terms. So a natural problem is to find for a given compact topological manifold $X$ a class $\beta \in K(\mC^{\infty}_c\xrightarrow {\frak f} \mTOP/X)$ such that $$w_*(\beta) = w_*^{Th}(X).$$
The case of the Pontryagin and L-class
--------------------------------------
Let $V$ be an oriented differentiable manifold and let $c\ell_*(V) \in H^{BM}_{2*}(V,\bQ)$ be the Poincaré dual $c\ell(TV) \cap [V]$ of a multiplicative and normalized functorial characteristic class $c\ell(E)\in H^{2*}(-, \bQ)$ of (isomorphism classes of) oriented real vector bundles. $\alp(V):=c\ell_*(V)$ is clearly an additive (and multiplicative) homology class, and by Corollary \[smooth-gen\] we have a unique natural transformation $$\tau_{c\ell_*}: K^{prop}_{\alp}(\mC^{\infty}_o \xrightarrow {S} \mTOP_{lc} /-) \to H^{BM}_{2*}(-;\bQ)$$ such that $\tau_{c\ell_*}([(V, V, id_{V})]) = c\ell_*(V)$ for a smooth oriented manifold $V$. In particular the Pontryagin class $c\ell=p$ and the Hirzebruch-Thom $L$-class $c\ell=L$ are typical ones.\
If we restrict ourselves to the category $\m V_{\bC,c}$ of compact complex algebraic varieties and let $\m V^{sm}_{\bC,c}$ be its full subcategory of compact smooth complex algebraic varieties, then we have a more geometric “realization" of the natural transformation $L_*$ on the category $\m V_{\bC,c}$ through *Cappell–Shaneson–Youssin’s cobordism groups $\Omega_*(X)$*(see [@CS], [@Youssin]) and the motivic relative Grothendieck group as mentioned in the introduction (compare with [@BSY] for more details): $$\begin{CD}
K(\m V^{sm}_{\bC,c} \xrightarrow {\iota} \m V_{\bC,c}/X) @> taut >> K_0(\m V_{\bC}/X) \\
@V L_* VV @VV sd V \\
H_{2*}(X, \bQ) @<< L^{CS}_* < \Omega(X) .
\end{CD}$$ Here the (composed) multiplicative natural transformation $$sd: K(\m V^{sm}_{\bC,c} \xrightarrow {\iota} \m V_{\bC,c}/X) \to \Omega(X)$$ is defined by the isomorphism invariant $\alp(V):=[\bQ_V[dim\;V]]\in \Omega(V)$, which is additive and multiplicative.
For a compact triangulated $\bQ$-homology manifold $V$, Thom constructed in [@Thom-8] an $L$-class in $H^{2*}(V;\bQ) \simeq H_{2*}(V;\bQ)$ using a relation with the signature (see [@MacPherson2; @Sullivan2]), so that for a compact oriented smooth manifold it agrees with the usual $L$-class $L(TV)$ of the tangent bundle by the famous *Hirzebruch signature theorem*. Through the development of *Intersection (co)homology*, this approach was further extended (by e.g. Goresky-MacPherson [@GM] and Cappell-Shaneson [@CS]) to more general singular spaces like compact (locally) pure-dimensional complex algebraic varieties. But then one has to view this class as a homology $L$-class in $H_{2*}(V;\bQ)$. Nevertheless, in Thom’s original approach for a compact triangulated $\bQ$-homology manifold $V$ only, one can use Poincaré duality to view this as a cohomology class in $H^{2*}(V;\bQ)$. Using the cup-product structure on cohomology, he was then able to define also a Pontryagin class $p(V)\in H^{2*}(V;\bQ)$ so that for a compact oriented smooth manifold it agrees with the usual Pontryagin class $p(TV)$ of the tangent bundle.
Let us call the corresponding homology classes the Thom–Pontryagin and Thom $L$-class, denoted by $$p_*^{Th}(X), L^{Th}_*(X) \in H_{2*}(X;\bQ).$$ If we consider the above Pontryagin- or $L$-class transformation for $c\ell=p$ or $L$ in the context of compact oriented smooth manifolds $$\tau_{c\ell_*}: K_{\alp}(\mC^{\infty}_{co} \xrightarrow {S} \mTOP /-) \to H_{2*}(-;\bQ),$$ then for a given compact triangulated $\bQ$-homology manifold $X$ it is a very interesting problem to find a class $\beta \in K(\mC^{\infty}_{co} \xrightarrow {S} \mTOP /X)$ such that $$p_*(\beta) = p_*^{Th}(X) \quad \text{or} \quad L_*(\beta)=L^{Th}_*(X).$$ Note that the rationalized group homomorphism $\tau_{c\ell_*}\otimes \bQ$ is surjective by [@Thom] (compare [@Rudyak Chapter IV, Theorem 7.36]), since $c\ell_*(V)=[V]+ \dots$ equals the fundamental class plus lower order terms.
The case of the Chern class
---------------------------
Let $V$ be a stable complex differentiable manifold and let $c\ell_*(V) \in H^{BM}_{2*}(V,R)$ be the Poincaré dual $c\ell(TV) \cap [V]$ of a multiplicative and normalized functorial characteristic class $c\ell(E)\in H^{2*}(-, R)$ of (isomorphism classes of) complex vector bundles. $\alp(V):=c\ell_*(V)$ is clearly an additive (and multiplicative) homology class, and we have by Corollary \[smooth-gen\] a unique natural transformation $$\tau_{c\ell_*}: K^{prop}(\mC^{\infty}_{\bC} \xrightarrow {S} \mTOP_{lc} /-) \to H^{BM}_{2*}(-;R)$$ such that $\tau_{c\ell_*}([(V, V, id_{V})]) = c\ell_*(V)$ for a smooth stable complex manifold $V$. In particular the Chern class $c\ell=c$ is a typical one.\
If we restrict ourselves to the category $\m V_{\bC}$ of complex algebraic varieties and let $\m V^{sm}_{\bC}$ be its full subcategory of smooth complex algebraic varieties, then we have a more geometric “realization" of the natural transformation $c_*$ on the category $\m V_{\bC}$ through constructible functions and the motivic relative Grothendieck group as mentioned in the introduction (see [@BSY] for more details): $$\begin{CD}
K^{prop}(\m V^{sm}_{\bC} \xrightarrow {\iota} \m V_{\bC}/X) @> taut >> K_0(\m V_{\bC}/X) \\
@V c_* VV @VV const V \\
H^{BM}_{2*}(X, \bZ) @<< c^{Mac}_* < F(X) .
\end{CD}$$ Here the (composed) multiplicative natural transformation $$const: K^{prop}(\m V^{sm}_{\bC} \xrightarrow {\iota} \m V_{\bC}/X) \to F(X)$$ is defined by the isomorphism invariant $\alp(V):=\jeden_V\in F(V)$, which is additive and multiplicative.
The case of Chern classes of other types
----------------------------------------
Let $\m V^{emb}_{\bC}$ be the subcategory of (complex) algebraic varieties embeddable into smooth varieties and let $c_*^{FJ}(X)$ resp., $c_*^{FJ}(X)$ be Fulton–Johnson’s Chern class resp., Fulton’s canonical class defined for such an embeddable (complex) algebraic variety: $c_*^{FJ}(X)$ ([@Fulton-book Example 4.2.6 (c)]) is defined by $$c_*^{FJ}(X):= c(TM|_X) \cap s(\m N_XM),$$ where $TM$ is the tangent bundle of $M$ and $s(\m N_XM)$ is the Segre class of the conormal sheaf $\m N_XM$ of $X$ in $M$ [@Fulton-book §4.2]. Fulton’s canonical class $c_*^F(X)$ ([@Fulton-book Example 4.2.6 (a)]) is defined by $$c_*^F(X) := c(TM|_X) \cap s(X,M),$$ where $s(X,M)$ is the relative Segre class [@Fulton-book §4.2]. For a local complete intersection variety $X$ we also have a normal bundle $N_XM$ in $M$, from which we can define the virtual tangent bundle $T_X$ of $X$ (as in §3) by $$T_X := TM|_X - N_XM \in K^0(X).$$ As shown in [@Fulton-book Example 4.2.6], for a local complete intersection variety $X$ in a non-singular variety $M$, these two Chern classes are both equal to the virtual Chern class $$c_*^{FJ}(X) = c_*^F(X) = c(T_X) \cap [X].$$
Moreover, both isomorphism invariants $\alp(V):=c_*^{FJ}(V)$ and $c_*^{F}(V)$ are additive for $V\in ob(\m V^{emb}_{\bC})$, so that there exists unique natural transformations on the category $\m V_{\bC}$: $$\tau_{c_*^F}, \tau_{c_*^{FJ}} : K^{prop}(\m V^{emb}_{\bC} \xrightarrow {\iota} \m V_{\bC}/-) \to H^{BM}_{2*}(-; \bZ),$$ such that $$\tau_{c_*^F}([V \xrightarrow{\op{id}_V} V]) = c_*^F(V) \quad \text{resp.} \quad \tau_{c_*^{FJ}} ([V \xrightarrow{\op{id}_V} V]) = c_*^{FJ}(V)$$ for $V \in ob(\m V^{emb}_{\bC})$. Finally, using Chow groups $CH_*(X)$ as a corresponding homology theory, all of this remains true over any base field $k$ (instead of working over $\bC$ with Borel-Moore homology), with the invariant $\alp(V)=c_*^F(V)$ for $V \in ob(\m V^{emb}_k)$ also multiplicative (as follows from [@Fulton-book Example 4.2.5]).\
[**Acknowlwdgements.**]{} We would like to thank Paolo Aluffi and Markus Banagl for useful discussions, and also Yuli B. Rudyak for providing the precise reference for Theorem \[rudyak\]. Finally we would like to thank Lê Dũng Tráng for useful suggestions and the referee for his/her careful reading and valuable suggestions and comments.
[99999]{}
P. Aluffi, *Singular schemes of hypersurfaces*, Duke J. Math. [**180**]{} (1995), 325 – 351.
P. Baum, W. Fulton and R. MacPherson, [*Riemann–Roch for singular varieties*]{}, Publ. Math. I.H.E.S. [**45**]{} (1975), 101–145.
J.-P. Brasselet, D. Lehman, J. Seade and T. Suwa, [*Milnor classes of local complete intersection*]{}, Tran. Amer. Math. Soc. [**354**]{} (2001), 1351–1371.
J.-P. Brasselet, J. Schürmann and S. Yokura, [*Hirzebruch classes and motivic Chern classes for singular spaces*]{}, Journal of Topology and Analysis, Vo1. [**2**]{}, No.1 (2010), 1–55.
J.-P. Brasselet and M.-H. Schwartz, [*Sur les classes de Chern d’une ensemble analytique complexe*]{}, Astérisque [**82–83**]{}(1981), 93–148.
S. E. Cappell, L. Maxim, J. L. Shaneson and J. Schürmann, [*Hirzebruch classes of hypersurfaces*]{}, Adv. in Math. [**225**]{} (2010), 2616–2647.
S. E. Cappell and J. L. Shaneson, [*Stratifiable maps and topological invariants*]{}, J. Amer. Math. Soc. [**4**]{} (1991), 521–551.
J. Fu and C. McCrory, [*Stiefel-Whitney classes and the conormal cycle of a singular variety*]{}, Trans. Amer. Math. Soc. [**349**]{} (1997), 809–835 .
W. Fulton, [*Intersection Theory*]{}, Springer Verlag, 1981.
W. Fulton and K. Johnson, [*Canonical classes on singular varieties*]{}, Manus. Math. [**32**]{} (1980), 381–389.
W. Fulton and S. Lang, [*Riemann-Roc Algebra*]{}, Springer Verlag New York, 1985.
W. Fulton and R. MacPherson, [*Categorical frameworks for the study of singular spaces*]{}, Memoirs of Amer. Math. Soc. [**243**]{}, 1981.
M. Goresky and R. MacPherson, [*Intersection homology theory*]{}, Topology, [**149**]{} (1980), 155–162.
F. Hirzebruch, [*Topological Methods in Algebraic Geometry, 3rd ed. (1st German ed. 1956)*]{}, Springer-Verlag, 1966.
F. Hirzebruch, T. Berger and R. Jung, [*Manifolds and Modular Forms*]{}, Vieweg, 1992.
G. Kennedy, [*MacPherson’s Chern classes of singular varieties*]{}, Com. Algebra. [**9**]{} (1990), 2821–2839.
M. Levine and F. Morel, [*Algebraic Cobordism*]{}, Springer–Verlag Berlin Heidelberg, 2007.
N. Levitt, [*Applications of Poincaré duality spaces to the topology of manifolds*]{}, Topology [**11**]{} (1972), 205–221.
S. MacLane, *Categories for the Working Mathematician*, Springer–Verlag New York, 1998.
R. MacPherson, [*Chern classes for singular algebraic varieties*]{}, Ann. of Math. [**100**]{} (1974), 423–432.
R. MacPherson, [*Characteristic classes for singular varieties*]{}, Proceedings of the 9th Brazilian Mathematical Colloquium (Poços de Caldas, 1973), Vol II, Instituto de Matematica Pura e Aplicada, S\~ ao Paolo (1977), 321–327.
A. Parusiński and P. Pragacz, [*Characteristic classes of hypersurfaces and characteristic cycles*]{}, J. Algebraic Geometry, [**10**]{} (2001), 63–79.
Y. B. Rudyak, [*On Thom Spectra, Orientability, and Cobordism*]{}, Springer Monographs in Mathematics, Springer, 1998.
Yu. B. Rudyak, [*Steenrod problem*]{}, Springer Online Reference Works, Encyclopaedia of Mathematics (Edited by Michiel Hazewinkel), http://eom.springer. de/S/s087540.htm
J. Schürmann, [*Characteristic classes of mixed Hodge modules*]{}, in “Topology of Stratified Spaces", MSRI Publications [**58**]{}, Cambridge Univ. Press (2010), 419–470,
J. Schürmann and S. Yokura, [*A survey of characteristic classes of singular spaces*]{}, in “Singularity Theory” (Denis Cheniot et al, ed.), Dedicated to Jean-Paul Brasselet on his 60th birthday, Proceedings of the 2005 Marseille Singularity School and Conference, World Scientific (2007), 865–952,
M.-H. Schwartz, [*Classes caract[é]{}ristiques d[é]{}finies par une stratiification d’une vari[é]{}ti[é]{} analytique complex*]{}, C. R. Acad. Sci. Paris [**t. 260**]{} (1965), 3262-3264, 3535–3537.
M.-H. Schwartz, [*Classes et caractères de Chern des espaces linéaires*]{}, C. R. Acad. Sci. Paris Sér. I. Math. [**295**]{} (1982), 399–402.
J. Seade, T. Mihai and A. Verjovsky, *Global Euler obstruction and polar invariants*, Math. Ann. [**333**]{} (2005), no.2, 393–403.
D. Sullivan, [*Geometric periodicity and the invariants of manifolds*]{}, Manifolds–Amsterdam 1970 (Proc. Nuffic Summer School), Springer Lecture Notes in Mathematics, Vol. [**197**]{} (1971), 44–75.
D. Sullivan, [*René Thom’s work on geometric homology and bordism*]{}, Bulletin Amer. Math. Soc. (New Series), [**41**]{}, No. 3 (2004), 341–350.
R. Thom, [*Quelques propriétés globales des variét’es differentiables*]{}, Comm. Math. Helv., [**28**]{} (1954), 17–86.
R. Thom, [*Espaces fibrés en sphères et carrés de Steenrod*]{}, Ann. Sci. Ecole Norm. Sup., [**69**]{} (1952), 109–182.
R. Thom, [*Les classes caracteristique de Pontrjagin des variétés triangulés*]{}, Symposium Internacional de Topologia Algebraica, UNESCO (1958), 54–67.
S. Yokura, [*Polar classes and Segre classes on singular projective varieties*]{}, Trans. Amer. Math. Soc., 298 (1986), 169–191,
S. Yokura, [*On Cappell-Shaneson’s homology L-class of singular algebraic varieties*]{}, Trans. Amer. Math. Soc. 347 (1995), 1005-1012
S. Yokura, [*On characteristic classes of complete intersections*]{}, “Algebraic Geometry - Hirzebruch 70” (ed. by Pragacz, Szurek and Wisniewski) Contemporary Mathematics Amer. Math. Soc., 241 (1999), 349–369
S. Yokura, [*Motivic characteristic classes*]{}, in “Topology of Stratified Spaces", MSRI Publications [**58**]{}, Cambridge Univ. Press (2010), 375–418,
S. Yokura, [*Motivic Milnor classes*]{}, Journal of Singularities, [**1**]{} (2010), 39–59.
S. Yokura, [*Natural transformations associated to additive homology classes*]{}, Kyoto University RIMS Kôkyûroku [**1764**]{} (2011), 30–47. B. Youssin, [*Witt Groups of Derived Categories*]{}, K-Theory, [**11**]{} (1997), 373-395.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We present spatially resolved millimeter maps of Neptune between 95 and 242 GHz taken with the Atacama Large Millimeter/submillimeter Array (ALMA) in $2016-2017$. The millimeter weighting functions peak between 1 and 10 bar on Neptune, lying in between the altitudes probed at visible/infrared and centimeter wavelengths. Thus, these observations provide important constraints on the atmospheric structure and dynamics of Neptune.
We identify seven well-resolved latitudinal bands of discrete brightness temperature variations, on the order of $0.5-3$K in all three observed ALMA spectral bands. We model Neptune’s brightness temperature using the radiative transfer code Radio-BEAR and compare how various H$_2$S, CH$_4$, and *ortho/para* H$_2$ abundance profiles can fit the observed temperature variations across the disk. We find that observed variations in brightness temperature with latitude can be explained by variations in the H$_2$S profile that range from sub- to super-saturations at altitudes above the 10-bar pressure level, while variations in CH$_4$ improve the quality of fit near the equator. At the south polar cap, our best fit model has a depleted deep atmospheric abundance of H$_2$S from 30 to only 1.5 times the protosolar value, while simultaneously depleting the CH$_4$ abundance. This pattern of enhancement and depletion of condensable species is consistent with a global circulation structure where enriched air rises at the mid-latitudes ($32^{\circ}-12^{\circ}$S) and north of the equator ($2^{\circ}-20^{\circ}$N), and dry air descends at the poles ($90^{\circ}-66^{\circ}$S) and just south of the equator ($12^{\circ}$S$-2^{\circ}$N). Our analysis finds more complex structure near the equator than accounted for in previous circulation models.
author:
- Joshua Tollefson
- Imke de Pater
- 'Statia Luszcz-Cook'
- David DeBoer
title: 'Neptune’s Latitudinal Variations as Viewed with ALMA'
---
Introduction {#S:1}
============
Millimeter continuum observations of Neptune provide a valuable bridge between visible/infrared studies of the cloud tops and above ($P<1$ bar) and deeply probing centimeter maps ($10<P<40$ bar). Visible and near-infrared (NIR) imaging from *Voyager*, the Hubble Space Telescope (HST), and Keck have shown bright methane cloud activity at Neptune’s mid-latitudes and comparatively dark regions and/or hazes at the equator and poles [@Limaye1991; @Sromovsky2001b; @Martin2012; @Fitzpatrick2013; @Tollefson2018]. Mid-infrared (MIR) images and spectra have been used to obtain zonal-mean temperature profiles which show that the equator and south pole are warmer than the southern mid-latitudes in the upper troposphere and stratosphere [@Conrath1991; @Orton2007; @Fletcher2014]. Centimeter maps obtained with the Very Large Array (VLA) probe well below Neptune’s cloud deck and indicate the presence of dry (low-opacity) air at the south pole (e.g., [@Butler2012DPS; @dePater2014]). Taken together, observations of the deep troposphere and upper atmosphere are broadly consistent with a global circulation pattern where air enriched with trace gases rises at the mid-latitudes, adiabatically cools and condenses to form clouds, dries out, and descends at the poles and equator [@Conrath1991; @dePater2014]. Millimeter observations probe altitudes in between those seen in the above maps, $1 < P < 10$ bar, making such observations vital for bridging the pressures viewed in the visible/infrared and centimeter, and for improving our overall understanding of Neptune’s atmospheric structure and dynamics. In particular, millimeter continuum observations are sensitive to variations in composition, including variations in trace condensable species such as H$_2$S and CH$_4$. Like the centimeter, millimeter observed variations in these trace gases reflect atmospheric motions; depleted, low-opacity regions (enriched, high-opacity) are consistent with downwelling (upwelling) air.
The opacity of Neptune’s continuum at millimeter observations is dominated by collision-induced absorption of H$_2$ (CIA) with hydrogen, helium, and methane. Trace gases, such as H$_2$S, PH$_3$, and NH$_3$ also contribute to the overall opacity. Since millimeter observations of these gases probe Neptune’s troposphere between 1 and 10 bar, individual lines are highly pressure-broadened, appearing as broad ‘continuum’ bands in the millimeter spectrum. As a result, it is difficult to differentiate between these opacity sources since clear line features cannot be detected.
On previous millimeter maps, Neptune’s disk was spatially resolved and regions of enhanced or diminished brightness temperatures could be distinguished. Using the Combined Array for Research in Millimeter-wave Astronomy (CARMA), @LuszczCook2013b imaged Neptune in the far wings of the CO ($2-1$) line (230.538 GHz). In their longitudinally-averaged maps, they found brightness temperatures increased by $2-3$ K from $40^{\circ}$N to the south pole. Since their observations were taken far ($\sim 5$ GHz) from the CO ($2-1$) line center, this emission is primarily from sources forming the ’quasi-continuum’. Assuming an adiabat in the troposphere, the authors showed that variations in the brightness temperature at the south pole could be explained by a 30% decrease in opacity at $P > 1$ bar. If variations in opacity would occur at altitudes only below 4 bar, the opacity needed to be decreased by a factor of 50. Hence, brightness variations at a particular wavelength are coupled to the pressure at which the opacity changes, itself dependent on the opacity source. @LuszczCook2013b could explain the latitudinal variations in brightness temperature as latitudinal gradients in any or all of: H$_2$S, CH$_4$, and/or *ortho/para* H$_2$. However, due to their limited wavelength coverage, they were unable to disentangle the true contribution to the brightness from each candidate source.
In this paper, we present millimeter maps of Neptune taken with the Atacama Large Millimeter/Submillimeter Array (ALMA). ALMA provides the wavelength coverage, sensitivity, and resolution needed to constrain Neptune’s opacity sources accurately across the disk. In Section 2, we present ALMA observations taken in three bands over the millimeter wavelength range ($1-3$ mm). The observed brightness temperature distribution is compared to model maps produced with the radiative-transfer code Radio-BEAR, described in Section 3. We generate a model for the latitudinally-varying H$_2$S and CH$_4$ abundance profiles that agree with the observations in each band in Section 4. In Section 5, we compare these results to other profiles of Neptune’s trace gases and summarize how our findings impact our understanding of the dynamics and evolution of Neptune’s upper atmosphere.
Observations {#S:2}
============
Data {#S:21}
----
We observed Neptune with ALMA, which is an interferometer located in the Atacama desert in northern Chile. A total of 66 high-precision antennas form the array: fifty-four 12m and twelve 7m. The arrangement of these antennas defines the angular resolution and maximum resolvable angular scale of the data. The tightest packed configuration, where the antennas are up to 150 m apart, allow large, faint objects to be observed. Extended arrangements, where the antennas are 16 km apart, provide a more detailed look. Our Neptune observations were taken between $2016-2017$ in Bands 3, 4, and 6, between $95-250$ GHz (1–3 mm). For each band, we selected the antenna configuration which allowed us to resolve Neptune and simultaneously see Neptune’s entire (2.3” diameter) disk. A typical resolution was 0.3” using $\sim40$ 12-m antennas. We observed in four spectral windows in continuum mode covering Neptune’s continuum spectra within each band. Table \[table1\] summarizes the observations while Table \[table2\] outlines the correlator and spectral setup.
[llllll]{}\
ALMA Band & Configuration & UT Date & Flux Calibrator & Bandpass Calibrator & T$_{\text{Source}}^{a}$\
Band 3 & C40-8 & 2017-07-27 & J0006-0623 & J2246-1206 & 726\
Band 4 & C40-7 & 2016-10-07 & J2258-2758 & J2246-1206 & 1701\
Band 6 & C40-7 & 2016-10-24 & J2258-2758 & J2246-1206 & 1032\
\[table1\]
\
[lcccccc]{}\
ALMA Band & Center Frequency & Center Wavelength & Channel Width & Total Bandwidth$^a$ & Beam Size & Position Angle$^b$\
& (GHz) & (mm) & (MHz) & (MHz) & (arcsec$^2$) & (degrees)\
Band 3 &
---------
95.012
96.970
107.000
109.000
---------
: Summary of correlator setup, spectral windows, and synthesized beams.
&
-------
3.155
3.092
2.802
2.750
-------
: Summary of correlator setup, spectral windows, and synthesized beams.
& 15.625 & 2000 &
--------------------
0.45“$\times$0.33”
0.44“$\times$0.33”
0.40“$\times$0.30”
0.39“$\times$0.29”
--------------------
: Summary of correlator setup, spectral windows, and synthesized beams.
&
-----------------
$281.5^{\circ}$
$282.1^{\circ}$
$282.9^{\circ}$
$282.9^{\circ}$
-----------------
: Summary of correlator setup, spectral windows, and synthesized beams.
\
Band 4 &
---------
135.986
137.924
147.986
149.986
---------
: Summary of correlator setup, spectral windows, and synthesized beams.
&
-------
2.205
2.174
2.026
1.999
-------
: Summary of correlator setup, spectral windows, and synthesized beams.
& 15.625 & 2000 &
--------------------
0.38“$\times$0.28”
0.38“$\times$0.28”
0.35“$\times$0.26”
0.35“$\times$0.26”
--------------------
: Summary of correlator setup, spectral windows, and synthesized beams.
&
----------------
$23.0^{\circ}$
$24.8^{\circ}$
$22.7^{\circ}$
$24.3^{\circ}$
----------------
: Summary of correlator setup, spectral windows, and synthesized beams.
\
Band 6 &
---------
223.982
225.982
239.981
241.981
---------
: Summary of correlator setup, spectral windows, and synthesized beams.
&
-------
1.338
1.327
1.249
1.239
-------
: Summary of correlator setup, spectral windows, and synthesized beams.
& 15.625 & 2000 &
--------------------
0.31“$\times$0.25”
0.32“$\times$0.24”
0.30“$\times$0.23”
0.29“$\times$0.23”
--------------------
: Summary of correlator setup, spectral windows, and synthesized beams.
&
----------------
$64.3^{\circ}$
$61.8^{\circ}$
$64.0^{\circ}$
$63.6^{\circ}$
----------------
: Summary of correlator setup, spectral windows, and synthesized beams.
\
\[table2\]
\
Calibration and Imaging {#S:22}
-----------------------
The obtained visibility data were reduced and calibrated in the Common Astronomical Software Application (CASA) version 5.1. We applied the standard ALMA pipeline to perform flagging (bad edge channels, shadowed antennas, and poor quality data), bandpass and flux calibration, and gain-time solutions. Table \[table1\] lists our calibrator sources. Finally, we applied three iterations of self calibration on our Neptune data to remove short-term phase variability caused by fast atmospheric fluctuations [@Brogan2018]. The first iteration used the entire observation range while the second and third used 350-second and 60-second intervals.
The multi-frequency synthesis mode in CASA *tclean* was used to transform the visibility data into image maps. We used natural weighting and restricted clean to a circular mask that is roughly the size of Neptune’s diameter plus twice the beam size. The natural weighting scheme gives equal weight to all baseline samples while the uniform weighting scheme gives equal weight to each spatial frequency. Since there are ‘naturally’ more short baselines than long baselines, natural weighting preserves the peak sensitivity while uniform weighting reduces the peak sensitivity as the short baselines have been weighted less. In our testing, we found that uniform and intermediate weightings produced artificial speckles on the disk and large scale structures in the sky. Natural weighting limits these artifacts since the sensitivity is highest, but sacrifices some angular resolution since the long-baselines are under-weighted relative to the uniform scheme. Despite this, we still resolve roughly one-seventh of Neptune’s disk. We also subtracted a limb-darkened model from the data to speed up the deconvolution process and reduce imaging artifacts. The number of clean iterations varied from 1000-2000, stopping once the noise within the planetary disk reached the noise in the sky.
Our resulting maps are 512$\times$512 pixels$^2$ with a cell size of 0.02". This cell size follows a common rule-of-thumb, which is to make the cell size roughly 1/10 the size of the synthesized beam. The planetary disk appears elliptically elongated due to convolution with the synthesized beam. This beam resembles a Gaussian with full-width at half maxima and position angles given in Table \[table2\].
Error estimation {#S:23}
----------------
The error in our maps is calculated by averaging over four regions of the sky with boxes equal to the diameter of Neptune and taking the root-mean-square (RMS). RMS values range from $0.1-0.6$K. Table \[table3\] lists our estimated errors in each band. This RMS does not include systematic effects, such as errors in the bandpass or flux calibration. The ALMA Calibrator Source Catalogue lists errors in the calibrator’s flux estimate being about 5$\%$ or less in each band so we use this as an estimate for the absolute error in our disk-averaged temperature data (Section 5.1).
[llllll]{}\
Center Frequency (GHz) & Observed Flux Density$^a$ (Jy) & Observed T$_{\text{b}}^{a}$ (K) & Nominal T$_{\text{b}}^b$ (K) & Noise$^c$ (K) & Factor$^d$\
95.012 & $3.4\pm0.2$ & $126.6\pm6.3$ & 119.9 & 0.1 & 0.945\
96.970 & $3.5\pm0.2$ & $126.0\pm6.3$ & 119.4 & 0.1 & 0.950\
107.000 & $4.1\pm0.2$ & $120.5\pm6.0$ & 116.7 & 0.2 & 0.975\
109.000 & $4.1\pm0.2$& $118.8\pm6.0$ & 116.1 & 0.3 & 0.975\
135.986 & $5.9\pm0.3$ & $108.5\pm5.4$ & 109.0 & 0.3 & 1.010\
137.924 & $6.1\pm0.3$ & $108.0\pm5.4$ & 108.5 & 0.2 & 1.010\
147.986 & $6.7\pm0.3$ & $104.3\pm5.2$ & 105.7 & 0.2 & 1.020\
149.986 & $6.8\pm0.3$ & $104.5\pm5.2$ & 105.1 & 0.3 & 1.025\
223.982 & $13.3\pm0.7$ & $93.4\pm4.7$ & 94.8 & 0.4 & 1.020\
225.982 & $13.5\pm0.7$ & $93.0\pm4.7$ & 94.7 & 0.4 & 1.025\
239.981 & $15.1\pm0.8$ & $93.1\pm4.7$ & 93.4 & 0.6 & 1.010\
241.981 & $15.3\pm0.8$ & $92.8\pm4.6$ & 93.2 & 0.6 & 1.010\
\[table3\]
\
\
\
Models {#S:4}
======
2D model maps of Neptune’s disk were created using our radiative transfer (RT) code Radio-BEAR described in @dePater2019[^1]. Given an atmospheric composition and thermal structure (described below), we calculate the RT-derived brightness temperatures of the planet on each location on the planet.
Radio-BEAR assumes that the atmosphere is in local thermodynamic equilibrium, where the temperature is calculated from deep in the atmosphere upwards assuming a dry or wet adiabat such that the temperature at 1 bar matches 71.5 K, that derived by *Voyager* radio occultation measurements [@Lindal1992]. At altitudes above 1 bar, we use the temperature profile from @Fletcher2010. The temperature, pressure, and altitudes are related to each other through hydrostatic equilibrium. In all of our latitude-varying models (Sections 4.2 and 4.3), we assume that the temperature-pressure profile follows a dry adiabat in the troposphere. Cases of a wet adiabat are presented in the context of the disk-averaged brightness temperature in section 4.1. Our models also allow the abundance profiles of H$_2$S, CH$_4$ and *ortho/para* H$_2$ to vary, as discussed below.
Neptune’s trace gases H$_2$S, CH$_4$, H$_2$O, and NH$_3$ in the deep atmosphere are assumed to be enhanced by $30\times$ their protosolar values[^2] in our nominal model, apart from ammonia gas as 30$\times$ protosolar values for NH$_3$ are inconsistent with previous microwave disk-averaged temperatures [@dePater1985; @dePater2014]. At higher altitudes, these gases follow the saturated vapor pressure curve with 100$\%$ relative humidity. Clouds expected to form under thermochemical equilibrium on Neptune include: an aqueous ammonia solution (H$_2$O-NH$_3$-H$_2$S) topped with water ice, ammonium-hydrosulfide (NH$_4$SH), and H$_2$S- and CH$_4$-ice [@Weidenschilling1973]. To form the NH$_4$SH cloud, H$_2$S and NH$_3$ are reduced in equal molar quantities until the product of their partial pressures reaches the equilibrium constant of the reaction forming NH$_4$SH. Once equilibrium is reached, only H$_2$S will remain to form clouds since there is practically no NH$_3$ gas remaining above the NH$_4$SH cloud. A tentative detection of H$_2$S spectral features near 1.58 $\mu$m in Neptune’s troposphere implies that the deep bulk S/N ratio is greater than one [@Irwin2019]. The cloud density might affect microwave measurements [@dePater1991; @dePater1993]. However, little is known about the cloud density on Neptune and the millimeter weighting functions peak at altitudes above the aqueous and ammonium-hydrosulfide clouds (see Fig. \[fig:contributions\]). Clouds have also been shown to not affect the microwave opacity on Jupiter [@dePater2019]. Therefore, we ignore the effect of cloud opacity and focus on the gas opacity in our models.
The gas opacity of Neptune’s millimeter spectrum is dominated by H$_2$S and the collision-induced absorption (CIA) of H$_2$ (we include: H$_2$-H$_2$, H$_2$-He, H$_2$-CH$_4$). NH$_3$ and H$_2$O affect the spectra at wavelengths longer than 10 cm; NH$_3$ would have a larger impact at millimeter wavelengths if its abundance within Neptune were larger (see Fig. 4 of @LuszczCook2013a). The effect of PH$_3$ in the millimeter is most prominent at the ($1-0$) absorption line at 266.9 GHz. The width of this line is $\sim$ 20 GHz due to pressure broadening, meaning the wings of this absorption feature will have a small effect on the highest frequency data. However, this effect is well within the estimated noise of the maps so we do not add this gas in our models.
The *ortho/para* H$_2$ fraction also influences Neptune’s millimeter brightness temperature, by modifying both the adiabatic lapse rate and the gas opacity [@Trafton1967; @Wallace1980; @dePater1985; @dePater1993]. The ratio of ortho to para hydrogen in equilibrium depends on temperature; however, fast vertical mixing could bring the ratio of ortho and para states of hydrogen away from equilibrium and towards a “normal" ratio of 3 parts ortho to 1 part para. In this paper, we assume “intermediate" H$_2$ proposed by @Trafton1967 and used by e.g., @LuszczCook2013b: the ortho and para states of hydrogen (which define the CIA opacities) are set to the equilibrium value at the local temperature, while the specific heat is set to that of ’normal’ hydrogen. For further explanation of intermediate hydrogen, see e.g., @Massie1982. @LuszczCook2013b find that assuming a normal hydrogen fraction rather than equilibrium fraction significantly increases the opacity and lowers the brightness temperature by $5-6$K in their 1.2-mm model maps.
We generate 2D model maps of Neptune as follows: at the center frequencies of each spectral window, we calculate the RT-derived brightness temperatures of the planet on 9455 points on the disk and interpolate between these points to obtain the same resolution as our CASA imaged maps (0.02”). We then convolve each RT-map with an elliptical gaussian model of the synthesized beam. Our nominal model can be summarized as follows: 1) Neptune’s trace gases are enhanced to 30$\times$ their protosolar value, except NH$_3$ which is $1\times$ protosolar; 2) ‘intermediate’ *ortho/para* H$_2$ (i.e., equilibrium *ortho/para* H$_2$; specific heat close to that of normal H$_2$); 3) the temperature-pressure profile follows a dry adiabat in the troposphere. Figure \[fig:nomprofs\] plots the temperature-pressure and abundance profiles for the nominal model. In the following section, we compare these beam-convolved model maps to the observed disk-averaged brightness temperatures and latitudinal brightness variations.
![Normalized weighting functions at: a) nadir and b) the south pole, compared to the expected clouds expected to form on Neptune (c). The weighting functions at nadir and the south pole are both computed using the nominal abundance profile depicted in Figure \[fig:nomprofs\]. Weighting functions are shown for representative frequencies in each ALMA band: 95 GHz (3.155 mm, Band 3), 136 GHz (2.205 mm, Band 4), and 224 GHz (1.338 mm, Band 6), as well as selected VLA frequencies: 4 and 43 GHz (6.2 cm and 0.7 cm, respectively). Note that the VLA frequencies probe significantly deeper into Neptune’s atmosphere than the ALMA frequencies. The rightmost plot shows the density of different clouds expect to form on Neptune under thermochemical equilibrium.[]{data-label="fig:contributions"}](cont_fns_clouds.pdf){width="0.75\linewidth"}
![Temperature and abundance profiles of trace gases whose deep atmospheric abundances are enhanced by 30$\times$ their protosolar value, apart from NH$_3$ which is $1\times$ (solid lines). In the troposphere, the temperature follows a dry adiabat. These profiles define the nominal model. The dashed line is the H$_2$S depleted profile that that best fits Neptune’s south pole from 2003 VLA data [@dePater2014].[]{data-label="fig:nomprofs"}](nominal_profiles.pdf){width="0.75\linewidth"}
Results {#S:5}
=======
In the following subsections, we investigate the distribution of Neptune’s observed brightness temperature. First, we look at the disk-averaged brightness temperature and compute a variety of RT models using dry and wet adiabats and different enhancements of the deep abundance of trace gases. From these results, we present our nominal model (the same as that described at the end of Section 3), which fits both our ALMA and 2003 VLA disk-averaged temperatures simultaneously. Second, we subtract our nominal model from the data to produce residual maps. These maps show seven distinct latitudinal bands across Neptune’s disk. In the final subsections, we find abundance profiles of H$_2$S, CH$_4$, and *ortho/para* H$_2$ which fit the observed brightness variations in each latitudinal band. As an example of how we do this, we first look at Neptune’s south polar cap, where previous VLA studies constrained the deep abundance of H$_2$S. This provides a good litmus test for our model and fitting routine. We then move on to the rest of Neptune’s disk and show that we can fit the observed brightness temperature distribution at all latitude bands by varying the abundances of the aforementioned species.
Disk-Averaged Brightness Temperatures {#S:5.1}
-------------------------------------
We calculate Neptune’s disk-averaged brightness temperature from the ALMA data in two ways. First, we sum the flux density contained within the planetary disk convolved with the model beam. Second, we fit the u-v short-spacing amplitude with a Bessel function and obtain an estimate of the zero-spacing flux density. Both results agree to well within the absolute calibration errors. We report the average of these two methods in Table \[table3\]. Since Neptune blocks the cosmic microwave background (CMB), its true brightness temperature is higher than observed in the radio. In all of our millimeter observations, we correct for the CMB by following the procedure laid out in Appendix A of @dePater2014.
Figure \[fig:diskavgtemp\] plots our observed radio disk-averaged brightness temperatures on top of RT model spectra. In this plot, we show the effect of varying the deep abundance enhancement of trace gases (H$_2$S, CH$_4$, H$_2$O, and NH$_3$) relative to their protosolar values and consider both wet and dry lapse rates. We find that the nominal model: 30$\times$ enhancement with a dry lapse rate, agrees very well with both the ALMA millimeter and the 2003 VLA centimeter data from @dePater2014. Models with temperatures following wet lapse rates are too cold relative to these data. We acknowledge, though, the following shortcomings in our models. First, these models are highly degenerate. Gas abundances are seldom at 100$\%$ humidity and so a profile with a wet adiabat where the gases are sub-saturated may result in high enough brightness temperatures to match the data. Second, our models do not account for the cooling (heating) of the atmosphere from adiabatic expansion (compression) at latitudes where the air is rising (sinking). Finally, apart from the dry and wet adiabat, we do not consider latitudinal variations in temperature. @Conrath1998 and @Orton2007 see $2-3$ K latitudinal temperature variations between $50-100$ mbar. These temperature variations are similar in strength to those we see in our residual maps (Fig. \[fig:residualmaps\]), particularly in Band 6. @LuszczCook2013b created temperature profiles matching the $2-3$ K variations described in @Conrath1998 and @Orton2007 at altitudes above 1 bar. These temperature-varying models reduced the overall $\chi^2$ in their fit to 1.2 mm continuum data. However, we expect similar profiles to have only a modest effect in Band 6, where the contribution functions peak at the highest altitudes (but still below 1 bar, see Fig. \[fig:contributions\]), and little effect in Bands 3 and 4, which probe pressures much higher than where temperature variations are seen in the infrared. Below 1 bar, the temperature profile should strictly follow an adiabat, regardless of latitude. In summary, while we do not consider temperature variations in our analysis, we do not expect them to be the primary cause of the observed brightness temperature distribution.
![Disk-averaged brightness temperatures of Neptune. The ALMA data are plotted as orange points, with 5% absolute errors estimated from the calibrators. In addition, VLA 2003 data are plotted in red [@dePater2014], along with older VLA data in gray [@dePater1991] and older single dish radio data in open circles [@dePater1989]. Overplotted are model spectra which vary the deep abundance of H$_2$S, CH$_4$, and H$_2$O (30$\times$ and 50$\times$ their protosolar values in black and light blue respectively), and compare dry and wet lapse rates (solid and dashed lines, respectively). A model for Neptune’s south polar hot spot which depletes H$_2$S above 43 bar is also shown in the dot-dash green line [@dePater2014].[]{data-label="fig:diskavgtemp"}](nep_discavg_sep2018.pdf){width="0.75\linewidth"}
Latitudinal Brightness Temperature Variations {#S:5.2}
---------------------------------------------
In Figure \[fig:residualmaps\], we show residual maps of Neptune in the first spectral window of each ALMA band. Since the absolute calibration is imperfect, each map is scaled by a factor such that the observed disk-averaged temperature matches the nominal model (see Table 3). As a result, our main assumption in this analysis is that Neptune’s disk-average matches the nominal model. Deviations from this model result in the latitudinal structure evident in the residuals. This structure is due to changes in the brightness temperature that we assume to be due to variations in the opacity. The dark areas at the southern mid-latitudes and in the northern equatorial region are interpreted as probing higher, colder altitudes due to enhancements in absorbers. Conversely, the south pole appears bright in the radio due to opacity depletions, allowing the deeper, warmer layers to be probed.
Planetary coordinates are computed for each pixel on the disk using ephemeris data from JPL Horizons, with the center pixel equaling the sub-observer latitude and longitude. Latitudes are reported in planetographic coordinates. We identify seven bands on Neptune that correspond to discrete changes in the temperature structure: 90$^{\circ}$S$-66^{\circ}$S, 66$^{\circ}$S$-55^{\circ}$S, 55$^{\circ}$S$-32^{\circ}$S, 32$^{\circ}$S$-12^{\circ}$S, 12$^{\circ}$S$-2^{\circ}$N, 2$^{\circ}$N$-20^{\circ}$N, northward of 20$^{\circ}$N. Each latitude bin is well-resolved, covering at least the total area of the ALMA beam, and is in planetographic coordinates. In Figure \[fig:residuals\_h2s30\], we plot the average temperature difference between the data and nominal model per spectral window and band. Differences that are twice the RMS noise reported in Table 3 are significant.
![ALMA residual maps where the beam-convolved nominal model has been subtracted from the data. Contour lines delineate the latitude transitions between bands. Dark bands represent cold brightness temperatures relative to the model, while bright bands are warmer than the model. The FWHM of the beam is indicated in white in the bottom left of each map.[]{data-label="fig:residualmaps"}](alma_residuals_images_all_v2.pdf){width="0.75\linewidth"}
![Residual temperatures comparing the data to the beam-convolved nominal model versus latitude (Band 3 left, Band 4 middle, Band 6 right). Points are plotted at the central latitude in the seven identified bins. The four gray dashed lines are the residual temperatures for each of the four spectral windows in a band. Red triangles are the average residual over each spectral window. The residuals in each latitude bin are calculated by averaging over each pixel within 60$^{\circ}$ of the sub-observer longitude. Error bars are the image noise divided by the square-root of the number of ALMA beams that fit into the corresponding bin.[]{data-label="fig:residuals_h2s30"}](alma_tempresiduals_allbands_h2s30.pdf){width="0.75\linewidth"}
Neptune’s South Polar Cap {#S:5.4}
-------------------------
The ALMA residual maps show clear warming on Neptune at latitudes southward of -66$^{\circ}$. This region is Neptune’s south polar cap and has been detected as a hot spot in the radio [@Butler2012DPS; @LuszczCook2013b; @dePater2014] and in the mid-infrared [@Hammel2007b; @Orton2007]. @dePater2014 found a best fit to their VLA data to the hot spot by defining a ‘plateau’ of constant low opacity from 66$^{\circ}$S to the south pole, depleting H$_2$S to 1.5$\times$ the protosolar value (or 5$\%$ their nominal 50$\times$ model) above 43 bar (i.e., above the NH$_4$SH cloud). The deep CH$_4$ abundance is also depleted to 1.5$\times$ the protosolar value in their model (equivalently, a 0.072$\%$ volume mixing ratio), but itself is not a source of opacity at cm-wavelengths and so this study did not attempt to constrain the south pole CH$_4$ content (although methane does affect the adiabat and opacity due to CIA). @LuszczCook2013b found that the high southern latitudes in their CARMA 1.3-mm map were consistent with the VLA model. In Figure \[fig:nomprofs\], we show this depleted profile and in Figure \[fig:sp-residuals\] we compare the bin-averaged residual brightness temperatures for the ALMA observations and the de Pater VLA depleted model. We find good agreement in Band 6, which covers the frequency of the CARMA map, however the model is too bright at lower frequencies. @LuszczCook2013b found that the high southern latitudes in their CARMA data could be matched if CH$_4$ were depleted to 0.55$\%$, while keeping H$_2$S at the nominal value. A comparison of our data with this CH$_4$ depleted model is also shown in Figure \[fig:sp-residuals\]. Once again, the Band 6 data agree with the 1.3-mm map of Neptune presented in @LuszczCook2013b to within their estimated uncertainties. However, the model is too cold to match the new low frequency data.
A better fit at low frequencies is obtained by setting the deep atmospheric CH$_4$ abundance to 0.55$\%$ and by depleting H$_2$S to 1.5$\times$ protosolar from 43 bar up to the saturated vapor pressure curve. The H$_2$S abundance follows the saturated vapor pressure curve below 4.5 bar; for $P < 4.5$ bar, H$_2$S is subsaturated down to $5\%$ of the saturated vapor curve. This adjusted profile is plotted in Figure \[fig:sp-profiles\] and the temperature residuals are shown in Figure \[fig:sp-residuals\]. We emphasize that some subsaturation is needed in order to fit the data well. Since the millimeter is most sensitive to the pressures where H$_2$S saturation occurs, all models in which H$_2$S profiles follow the saturated vapor pressure curve appear indistinguishable in terms of their contributions to the overall opacity and all under-predict the brightness temperatures in this region. Therefore, even models significantly depleting the deep abundance of H$_2$S will have temperatures that are too cold at the south pole and will resemble the residuals in Figure \[fig:residuals\_h2s30\].
With this new depleted model for Neptune’s south polar hot spot, we can fit the VLA, CARMA, and ALMA data simultaneously. This is because the VLA cm-data probe pressures deeper than our alterations at the H$_2$S saturation curve and we have not changed the deep H$_2$S abundance estimated from @dePater2014 nor the deep CH$_4$ abundance determined by @LuszczCook2013b. The CARMA maps have larger uncertainties ($2-4$ K) than our maps presented here, and so depleting both H$_2$S and CH$_4$ is consistent with those data.
![Residuals of the south polar data minus each of three different models, as described in Section 4.3. The dashed green line represents the residuals for the data compared with the model assuming H$_2$S subsaturation to $5\%$ of its nominal value, as suggested by the VLA cm data in @dePater2014. The orange dashed line and orange triangles are residuals to the south pole region in a model which only depletes CH$_4$ to 0.55$\%$, holding the other trace gases to their nominal values. The solid green line represents the residuals for the data compared with a model in which H$_2$S is partially subsaturated. This model depletes H$_2$S to 1.5$\times$S above 43 bar before following the saturation vapor pressure curve up to 4.5 bar and becoming subsaturated at higher altitudes (see Figure \[fig:sp-profiles\] for a plot of this profile). Simultaneously, CH$_4$ is depleted to 0.55$\%$ in the deep atmosphere. This is our preferred model and is consistent with the observations of @dePater2014 and @LuszczCook2013b. []{data-label="fig:sp-residuals"}](sp_tempresiduals_all_models.pdf){width="0.75\linewidth"}
![H$_2$S (green) and CH$_4$ (orange) abundance profiles. Models depleting one or both of these gases are used to explain radio brightness enhancements at Neptune’s South Pole. The 30$\times$S nominal profiles are shown as solid lines. The H$_2$S model fitting VLA measurements depletes H$_2$S down to $5\%$ (1.5$\times$S) of the nominal profile at $P < 43$ bar; this profile is plotted as a dot-dash green line [@dePater2014]. Our models for the H$_2$S and CH$_4$ profile, which simultaneously fits the ALMA, CARMA, and VLA measurements, are plotted as dashed lines. Note that in our H$_2S$ profile, H$_2$S partially follows the saturation pressure curve, and follows the subsaturated profile of [@dePater2014] higher than 4.5 bar.[]{data-label="fig:sp-profiles"}](southpole_profiles.pdf){width="0.75\linewidth"}
Constraining Neptune’s Variations with Latitude {#S:5.3}
-----------------------------------------------
From Figures \[fig:residualmaps\] and \[fig:residuals\_h2s30\], it is clear that the magnitude of temperature variations across Neptune’s disk varies with wavelength. In the millimeter, these variations are caused by altering the abundance of trace gases: H$_2$S and CH$_4$, or changing the fraction of *ortho/para* H$_2$. Figure \[fig:comp\_vs\_spectra\] compares the millimeter-spectrum of the nominal model to that of models depleted in trace gases. These plots show that the spectral shape varies with both wavelength and composition. Models where the H$_2$S profile follows the saturated vapor pressure curve differ little from the nominal model, no matter the abundance alterations in the deep troposphere. This is because the ALMA bands probe altitudes above 10 bar, where H$_2$S saturation begins, meaning changes in the constant deep abundance profile are undetected. On the other hand, models which subsaturate H$_2$S significantly increase the brightness temperature at short frequencies. The Band 6 (high frequencies) contribution functions peak between $1-5$ bar while Band 3 (low frequencies) probes between $5-10$ bar, altitudes close to where H$_2$S begins saturation. Thus, subsaturating H$_2$S results in a larger loss of opacity at low frequencies than the frequencies which probe higher altitudes, meaning low frequencies are able to probe deeper, warmer layers. In contrast, depleting CH$_4$ results in uniform increases across the millimeter. Substituting normal hydrogen for equilibrium hydrogen decreases the brightness temperature by 4K in Band 6 and less than 2K in Band 3. These differences with wavelength enable us to disentangle the effect of each constituent on the observed brightness temperatures in each spectral band.
In the following subsections, we present models where we vary a single constituent: H$_2$S, CH$_4$, or *ortho/para* H$_2$, while holding the others at their nominal value. We find the profiles which best match the data in each latitude bin for a single band. Then, we compare how well this matches the data in the other spectral bands. This gives a sense of how important the varied constituent is in producing model temperatures for each observation. Following this procedure, we present our best model fits to all three spectral bands where every parameter is allowed to vary. Table \[table:modeldesc\] describes how the constituent profiles change over all latitude and models.
The significance of our results within a particular latitude bin is computed with the reduced $\chi^2$: $$\chi^2 = \frac{1}{M-N}\sum_{m=0}^{M-1} \frac{\delta T_m^2}{\sigma_m^2},$$ where $\delta T_m$ is the difference between the data and model brightness temperature at each spectral window $m$, $\sigma_m$ is the image noise (see Table 3) divided by the square root of the number of beams that fit within the particular latitude bin. $M - N$ is the number of degrees of freedom in the model. We calculate the probability $p$ that this value of the reduced $\chi^2$, or a larger one, could arise by chance given $M-N$. We take the number of degrees of freedom to be eight: twelve spectral windows ($M$) minus four free parameters ($N$): the deep abundances of H$_2$S and CH$_4$, the *ortho/para* H$_2$ fraction, and the scaling factor assumption. For eight degrees of freedom, $p < 0.05$ when the reduced $\chi^2 > 15.5$. Therefore, models with reduced $\chi^2 > 15.5$ are unlikely to occur due to random chance and are inconsistent with the observations. Table \[table:chisq\] summarizes the various fits to the data across each latitude range.
![A comparison of models of Neptune’s spectrum to gauge how different opacity sources affect the brightness temperature across millimeter wavelengths. Plotted is the disk-averaged brightness temperatures for various models with the nominal model subtracted. The orange model depletes CH$_4$ to 15$\times$ the protosolar value (0.72$\%$) below saturation and the follows saturated vapor pressure curve above (as in the dashed-orange line in Fig. \[fig:sp-profiles\]). The solid green model depletes H$_2$S to 15$\times$ the protosolar value (3.52E-4 mixing ratio) in the deep atmosphere, but follows the saturation curve. The dashed green model depletes H$_2$S similarly, but subsaturates H$_2$S by the same fraction (similar to the dashed green profile in Figure \[fig:nomprofs\]). The black dot-dash model substitutes opacity by normal hydrogen for equilibrium hydrogen. Gray rectangles indicate the ALMA bands (Bands 3, 4, and 6 left-to-right).[]{data-label="fig:comp_vs_spectra"}](alma_comp_spectra_diff.pdf){width="0.75\linewidth"}
### Varying H$_2$S
We first consider variations in H$_2$S that best match the Band 3 latitude variations. For each latitude band other than the south pole, we produce a grid of models in which the H$_2$S mixing ratio above 43 bar (the NH$_4$SH cloud) is set from $5-30\times$ solar in 5$\times$S steps. The profile then follows the saturation curve up to 4.5 bar, and is either depleted above 4.5 bar or supersaturated to a higher altitude. We choose 4.5 bar as this transition pressure as this matches the data at the south pole best. For supersaturated models, the H$_2$S abundance is set to the constant deep atmospheric value until $5-7$ bar, testing a grid of models in 0.25 bar steps. At altitudes above 3 bar (the high-altitude limit of H$_2$S cloud formation), the H$_2$S profile follows the saturation curve. At pressures in between 3 bar and $5-7$ bar, the abundance is assumed to be linear in log-log space. For the south pole, we do not produce new model fits but continue to use the model from Section 4.3, which is depleted to a much higher depth to be consistent with VLA observations. The deep CH$_4$ abundance is held to its nominal value of 1.44$\%$ ($30\times$ S) and equilibrium H$_2$ is assumed. Figure \[fig:h2sprof\] plots example H$_2$S profiles used in these fits.
From the residuals plotted in Figure \[fig:residuals\_h2s\], there is general good agreement between the data and model in each band, apart from 12$^{\circ}$S$-2^{\circ}$N where the model in Bands 4 and 6 is too warm compared to the data. This suggests the need to enhance CH$_4$ or add normal H$_2$ in this area.
![Examples of H$_2$S abundance profiles used to best fit the ALMA data. The light green dashed line and solid green line are the same profiles as depicted in Fig. \[fig:sp-profiles\]. The dark green and black dashed and dot-dashed lines are supersaturated profiles (abbreviated supersat-v1 and supersat-v2).[]{data-label="fig:h2sprof"}](h2s_profiles.pdf){width="0.75\linewidth"}
![Residual temperatures comparing the data to the beam-convolved nominal model versus latitude, as in Fig. \[fig:residuals\_h2s30\], allowing H$_2$S (only) to vary with latitude to best-match the Band 3 data.[]{data-label="fig:residuals_h2s"}](alma_tempresiduals_allbands_h2s_vary.pdf){width="0.75\linewidth"}
### Varying CH$_4$
In order to determine if variations in CH$_4$ may effect the brightness temperature, we find CH$_4$ profiles which fit Band 6 to within the error bars of the average residual of the four spectral windows. The H$_2$S abundance is held at the nominal 30$\times$S profile and equilibrium H$_2$ is assumed. CH$_4$ is either enhanced or depleted from the nominal volume mixing ratio (1.44$\%$ or 30$\times$S), below the saturation curve. We create a grid of models in mixing ratio step-sizes of 0.36$\%$ (or 7.5$\times$S), testing deep abundances between $0.55-4.4\%$, the limits considered in [@LuszczCook2013b].
The residuals in each band are plotted in Figure \[fig:residuals\_ch4\]. Bands 3 and 4 do not fit a CH$_4$-varying only model well, implying H$_2$S must vary as well.
![As Fig. \[fig:residuals\_h2s\], but allowing only CH$_4$ to vary with latitude. We find CH$_4$ profiles that match the Band 6 data and compare these with the Band 3 and Band 4 data.[]{data-label="fig:residuals_ch4"}](alma_tempresiduals_allbands_ch4_vary.pdf){width="0.75\linewidth"}
### Varying *ortho/para* H$_2$
In these fits, we consider the fraction of equilibrium H$_2$ to normal H$_2$ needed to match the Band 6 brightness temperatures. Since normal H$_2$ decreases the brightness temperature, adding normal H$_2$ to latitudes with positive temperature residuals (like the south pole) is implausible. Figure \[fig:residuals\_h2n\] plots the temperature residuals in each band. We find that a hydrogen mixture of 90$\%$ equilibrium hydrogen and 10$\%$ normal hydrogen can explain the Band 6 data between 32$^{\circ}$S$-12^{\circ}$S, with equilibrium H$_2$ used everywhere else. However, the other ALMA spectral bands fail to fit the data from 32$^{\circ}$S$-12^{\circ}$S. This result is expected based on the spectral analysis in figure \[fig:comp\_vs\_spectra\], showing the minimal effect normal hydrogen has in Band 3. For the best fit model, we therefore do not consider normal hydrogen. We emphasize however that our results do not preclude a small fraction of normal H$_2$, on the order of $\leq 10\%$, existing at the pressures probed by ALMA. This is consistent with findings from *Voyager* IRIS measurements ([@Conrath1998; @Fletcher2014]. These studies find that the para H$_2$ fraction deviates from its expected equilibrium values on the order of $2-5\%$ between $0.01-1$ bar.
![As Fig. \[fig:residuals\_h2s\], but allowing only the fraction of equilibrium H$_2$ to vary with latitude. We find H$_2$ profiles that match the Band 6 data and compare these with the Band 3 and Band 4 data.[]{data-label="fig:residuals_h2n"}](alma_tempresiduals_allbands_h2n_vary.pdf){width="0.75\linewidth"}
### Best Fit Model
The first row in Table \[table:modeldesc\] describes the H$_2$S and CH$_4$ profiles that provide a best fit to the data in all ALMA bands in each latitudinal bin. Figure \[fig:best\_fit\_profs\] plots colormaps of the best fit profiles versus latitude. Figure \[fig:residuals\_bestfit\] plots the temperature residuals in each band, showing that an excellent fit is obtained at all latitudes. In general, we find that bright regions in the residual maps require depleting H$_2$S and sometimes CH$_4$ (down to $\sim0.6 - 1.1\%$), while dark regions require supersaturing H$_2$S and enhancing the deep CH$_4$ abundance ($\sim2.2 - 2.9\%$) to fit the data in every band.
### Comparison to [@Karkoschka2011]
@Karkoschka2011 determined methane mixing ratios from the HST/STIS spectrograph data between $300-1000$ nm across Neptune’s disk. Their results are consistent with a constant deep methane mixing ratio of $4\pm1 \%$ at $P > 3.3$ bar. However, between $1.2-3.3$ bar, they argued that the methane mixing ratio was depressed by a factor of $\sim3$ at Neptune’s mid-latitudes compared to the equator. Not only is their deep methane mixing ratio significantly higher than our best fit models (ranging between $0.6-2.9\%$), but their observed trend in the methane abundance across Neptune’s disk differs from ours. We find high methane mixing ratios relative to the nominal model between $32^{\circ}$S$-12^{\circ}$S and $2^{\circ}$N$-20^{\circ}$N, and low abundances from $90^{\circ}$S$-66^{\circ}$S and $12^{\circ}$S$-2^{\circ}$N. The ALMA Band 6 contribution functions peak at $P<4$ bar, so we expect some signature of the opacity from $1.2-3.3$ bar to appear in our results.
In order to test the models presented in @Karkoschka2011 to our data, we compare how the H$_2$S abundance must change to fit the data in every band. Between $32^{\circ}$S$-20^{\circ}$N, we use the @Karkoschka2011 methane profile at $6^{\circ}$S; elsewhere, we use their $45^{\circ}$S model (see their Figs. 10 and 14)[^3]. The one exception is at the south pole cap, $90^{\circ}$S$-66^{\circ}$S, where we use our best fitting profile since the deep H$_2$S abundance is fairly well constrained in VLA data.
Since the Band 6 data have the most overlap with the analysis in @Karkoschka2011 in terms of altitude, we find the H$_2$S profiles at each latitude that fit Band 6 well, assuming their adopted CH$_4$ profiles and equilibrium H$_2$. Our results are plotted in Figure \[fig:residuals\_kark\] and a full description of the model and corresponding statistics is listed in Tables \[table:modeldesc\] and \[table:chisq\]. We can find H$_2$S profiles which result in generally good agreement between the model and data only from $90^{\circ}$S$-50^{\circ}$S and northward of $20^{\circ}$N. In this case, the disk-averaged H$_2$S abundance is $\sim 10\times$S. @LuszczCook2013a computed the disk-averaged brightness temperature assuming a 10$\times$S model, showing that the brightness temperature is too high in this case compared to most of the radio data (see also Fig. \[fig:diskavgtemp\]). Therefore, our data are not consistent with the @Karkoschka2011 methane profiles. We address this in the following section.
![As Figure \[fig:residuals\_h2s\], but using the best fit H$_2$S and CH$_4$ profiles.[]{data-label="fig:residuals_bestfit"}](alma_tempresiduals_allbands_jt_model_v2.pdf){width="0.75\linewidth"}
![As Figure \[fig:residuals\_h2s30\], but using the CH$_4$ profiles from [@Karkoschka2011] and finding the H$_2$S profiles that best match Band 6.[]{data-label="fig:residuals_kark"}](alma_tempresiduals_allbands_kark.pdf){width="0.75\linewidth"}
[llllllll]{}\
Model Name & 90$^{\circ}$S$-66^{\circ}$S & 66$^{\circ}$S$-55^{\circ}$S & 55$^{\circ}$S$-32^{\circ}$S & 32$^{\circ}$S$-12^{\circ}$S & 12$^{\circ}$S$-2^{\circ}$N & 2$^{\circ}$N$-20^{\circ}$N & 20$^{\circ}$N$-50^{\circ}$N\
Best Fit &
---------------------
H$_2$S 1.5$\times$S
CH$_4$ 0.55$\%$
H$_2$ equil.
---------------------
&
--------------------
H$_2$S 10$\times$S
CH$_4$ 1.44$\%$
H$_2$ equil.
--------------------
&
--------------------
H$_2$S 30$\times$S
CH$_4$ 1.44$\%$
H$_2$ equil.
--------------------
&
--------------------
H$_2$S supersat-v1
CH$_4$ 2.88$\%$
H$_2$ equil.
--------------------
&
--------------------
H$_2$S 20$\times$S
CH$_4$ 1.08$\%$
H$_2$ equil.
--------------------
&
--------------------
H$_2$S supersat-v1
CH$_4$ 2.16$\%$
H$_2$ equil.
--------------------
&
--------------------
H$_2$S 10$\times$S
CH$_4$ 1.44$\%$
H$_2$ equil.
--------------------
\
Vary H$_2$S Only &
---------------------
H$_2$S 1.5$\times$S
CH$_4$ 0.55$\%$
H$_2$ equil.
---------------------
&
--------------------
H$_2$S 10$\times$S
CH$_4$ 1.44$\%$
H$_2$ equil.
--------------------
&
--------------------
H$_2$S 30$\times$S
CH$_4$ 1.44$\%$
H$_2$ equil.
--------------------
&
--------------------
H$_2$S supersat-v2
CH$_4$ 1.44$\%$
H$_2$ equil.
--------------------
&
-------------------
H$_2$S 3$\times$S
CH$_4$ 1.44$\%$
H$_2$ equil.
-------------------
&
--------------------
H$_2$S supersat-v2
CH$_4$ 1.44$\%$
H$_2$ equil.
--------------------
&
-------------------
H$_2$S 3$\times$S
CH$_4$ 1.44$\%$
H$_2$ equil.
-------------------
\
Vary CH$_4$ Only &
--------------------
H$_2$S 30$\times$S
CH$_4$ 0.55$\%$
H$_2$ equil.
--------------------
&
--------------------
H$_2$S 30$\times$S
CH$_4$ 0.55$\%$
H$_2$ equil.
--------------------
&
--------------------
H$_2$S 30$\times$S
CH$_4$ 1.44$\%$
H$_2$ equil.
--------------------
&
--------------------
H$_2$S 30$\times$S
CH$_4$ 2.88$\%$
H$_2$ equil.
--------------------
&
--------------------
H$_2$S 30$\times$S
CH$_4$ 1.44$\%$
H$_2$ equil.
--------------------
&
--------------------
H$_2$S 30$\times$S
CH$_4$ 1.44$\%$
H$_2$ equil.
--------------------
&
--------------------
H$_2$S 30$\times$S
CH$_4$ 0.55$\%$
H$_2$ equil.
--------------------
\
Vary H$_2$ Only &
--------------------
H$_2$S 30$\times$S
CH$_4$ 0.55$\%$
H$_2$ equil.
--------------------
&
--------------------
H$_2$S 30$\times$S
CH$_4$ 1.44$\%$
H$_2$ equil.
--------------------
&
--------------------
H$_2$S 30$\times$S
CH$_4$ 1.44$\%$
H$_2$ equil.
--------------------
&
---------------------
H$_2$S 30$\times$S
CH$_4$ 1.44$\%$
H$_2$ 90$\%$ equil.
---------------------
&
--------------------
H$_2$S 30$\times$S
CH$_4$ 1.44$\%$
H$_2$ equil.
--------------------
&
--------------------
H$_2$S 30$\times$S
CH$_4$ 1.44$\%$
H$_2$ equil.
--------------------
&
--------------------
H$_2$S 30$\times$S
CH$_4$ 1.44$\%$
H$_2$ equil.
--------------------
\
KT2011$^a$ CH$_4$ &
---------------------
H$_2$S 1.5$\times$S
CH$_4$ 0.55$\%$
H$_2$ equil.
---------------------
&
-------------------
H$_2$S 5$\times$S
CH$_4$ 45S
H$_2$ equil.
-------------------
&
--------------------
H$_2$S 15$\times$S
CH$_4$ 45S
H$_2$ equil.
--------------------
&
--------------------
H$_2$S $15\times$S
CH$_4$ 6S
H$_2$ equil.
--------------------
&
--------------------
H$_2$S 10$\times$S
CH$_4$ 6S
H$_2$ equil.
--------------------
&
--------------------
H$_2$S 10$\times$S
CH$_4$ 6S
H$_2$ equil.
--------------------
&
-------------------
H$_2$S 5$\times$S
CH$_4$ 45S
H$_2$ equil.
-------------------
\
\
\[table:modeldesc\]
[lccccccc]{}\
Model Name & 90$^{\circ}$S-66$^{\circ}$S & 66$^{\circ}$S-55$^{\circ}$S & 55$^{\circ}$S-32$^{\circ}$S & 32$^{\circ}$S-12$^{\circ}$S & 12$^{\circ}$S-2$^{\circ}$N & 2$^{\circ}$N-20$^{\circ}$N & 20$^{\circ}$N-50$^{\circ}$N\
Best Fit & 3.4 & 2.9 & 6.8 & 2.0 & 8.1 & 4.1 & 6.2\
Vary H$_2$S Only & 3.4 & 2.6 & 3.2 & 14.8 & **46.4** & 13.0 & 7.5\
Vary CH$_4$ Only & **55.8** & **20.7** & **16.5** & **56.7** & 2.3 & **25.5** & 7.3\
Vary H$_2$ Only & **95.8** & **48.0** & **16.1** & **85.6** & 2.8 & **17.2** & **29.0**\
KT2011 CH$_4$ Only & 3.9 & 4.3 & **28.9** & **100.4** & **20.9** & **50.3** & 8.8\
Discussion {#S:6}
==========
Neptune’s South Polar Cap {#neptunes-south-polar-cap}
-------------------------
Our ALMA millimeter observations provide a glimpse of Neptune’s atmosphere situated between the deep troposphere in centimeter maps ($P \geq 10$ bar) and the upper troposphere and stratosphere in the visible and infrared. The constraints on the trace gases are useful to infer the dynamics of Neptune’s atmosphere. Of particular interest is Neptune’s south pole, whose high temperatures were first published by @Hammel2007b with images taken with the Gemini north telescope at 7.7$\mu$m and 11.7$\mu$m. They suggested these bright regions are due to enhancements in ethane and methane. @Orton2007 imaged atmospheric line-free thermal emission of Neptune with the Very Large Telescope in 2006, finding temperature excesses of $10-11$K and $3-5$K near, but not at, Neptune’s south pole at 17.6$\mu$m and 18.7$\mu$m respectively. These authors suggested that seasonal warming around Neptune’s south pole could explain why the stratospheric abundance of methane is larger than expected; cold temperatures should result in methane condensing and becoming trapped below the tropopause. However, warm polar temperatures may allow methane gas to escape upward into the stratosphere and diffuse across the globe.
Warm brightness temperature measurements at high-latitudes are persistent throughout radio maps. @dePater2014 found temperature enhancements from 8$-$30K in VLA 1.3$-$6.2 cm maps, where sensitivities peak between $5-50$ bar. EVLA 1-cm maps show enhancements of similar magnitudes southward of 70$^{\circ}$S [@Butler2012DPS]. @LuszczCook2013b see southern high-latitude enhancements of $2-3$K compared to the northern mid-latitudes in 1.2 mm CARMA maps. Our ALMA maps show average enhancements of $2-3$K in Bands 3 and 4, with sensitivities peaking at $P > 1$ bar, and $1-2$K in Band 6, whose sensitivities peak at 1 bar. @Iino2018 analyzed ALMA flux calibration data of Neptune at 646 GHz (0.46 mm), peaking at 0.6 bar, and ruled out a detection of the south polar hot spot greater than 2.1K compared to the background. Combined, these data suggest that the magnitude of the south polar brightness enhancements decreases with increasing altitude. This trend is likely due to the temperature-pressure profile, which is mostly isothermal between $0.1-1$ bar. This picture also appears dynamically distinct from that described in @Orton2007, who predicted upwelling air to explain methane and ethane enrichment in the stratosphere. In the radio, brightness enhancements are consistent with low-opacity (dry) air so deeper warmer layers are probed. The air is likely subsiding, after having been dried out at other latitudes. The subsiding air causes adiabatic warming, which is sensed in the mid-IR (e.g., Fig. 16 in @dePater2014). Persistent cloud activity surrounding Neptune’s south pole may be indicative of vigorous convection and evidence of a south polar vortex [@LuszczCook2010], analogous to the polar activity seen on Saturn [@Fletcher2008; @Dyudina2008]. Such a system could explain the observed temperature, ethane and methane enhancements in the mid-IR, polar cloud features seen with *Voyager* and Keck, and high brightness temperatures in the radio.
Recent findings by @Irwin2019 find a tentative detection of an H$_2$S spectral signature between $1.57-1.58\mu$m on Neptune with Gemini-North/NIRS. They find that the signature is stronger at southern latitudes than at the equator, with H$_2$S abundances around 3 ppm and 1 ppm respectively at the top of Neptune’s H$_2$S cloud deck: $2.5-3.5$ bar. This broadly agrees with our suite of proposed H$_2$S profiles (Fig. 9), where H$_2$S is depleted to an abundance of 1 ppm between $\sim 2.5-4$ bar. The exception is our model for the south polar cap, which has less H$_2$S at the relevant altitudes than predicted by @Irwin2019. In order to match their results, we would have to lower the altitude where H$_2$S follows the saturated vapor pressure curve (to $P > 4.5$ bar) so that the abundance from $2.5-3.5$ bar could be increased to 1 ppm. Moreover, their retrieved H$_2$S abundances are most consistent with models which deplete the deep abundance of CH$_4$ at the southern high-latitudes relative to the equator. This agrees with our results and @Karkoschka2011.
Neptune’s Mid-Latitudes and Equator
-----------------------------------
Moving northward, our ALMA maps show that latitudes spanning $32^{\circ}-12^{\circ}$S and $2^{\circ}-20^{\circ}$N are colder than the background by $\sim0.5-1.5$K in all three bands. This difference can only be explained at all wavelengths with models which both increase the deep CH$_4$ abundance to $\sim2.2-2.9\%$ and supersaturate the H$_2$S profile. This is consistent with upwelling, adiabatically cooling plumes, and the observed distribution of prevalent bright cloud activity in this latitude range seen in the visible and near-IR. @Karkoschka2011 find that methane is well-mixed in the deep-atmosphere, with an abundance of $4\pm1\%$. In addition, they find that methane is depleted by a factor of 3 at the southern mid-latitudes compared to the equatorial region between 1.2 and 3.3 bar. This pressure range contributes the most to the opacity in Band 6, wavelengths where we also see the lowest temperature contrasts. If we assume their horizontal methane profiles, our models are too cold to match the ALMA data, due to their high methane abundance. A corresponding global decrease in the H$_2$S abundance to $\sim10\times$S must be made in order to decrease the total opacity and increase the brightness temperature so that the Band 6 data and model agree. However, decreasing the global H$_2$S abundance has the side effect of dramatically increasing the brightness temperatures in Bands 3 and 4, particularly near the equator. In Band 4, these profiles fit the equatorial regions poorly, yet agree at these latitudes in Band 3. For Band 4 to match here, the H$_2$S abundance would have to be increased to account for a 1 K difference. However, the Band 3 model temperatures would be decreased by at least this same amount, putting it outside a two-sigma fit to the data. Moreover, model disk-averaged brightness temperatures assuming a $10\times$ S H$_2$S abundances are too warm compared to the ALMA and 2003 VLA data (Fig. \[fig:diskavgtemp\]).
Our results are more consistent with @Baines1995, who predict disk-averaged deep-atmosphere methane molar fractions of $2.2^{+0.5}_{-0.6}\%$. @Karkoschka2011 remark that the two studies use different haze and cloud profiles and relative humidities. Correcting for these factors lowers their deep-atmosphere methane mixing ratio to within the error bars cited in @Baines1995. In addition, @Karkoschka2011 note that a significant contribution to their error bars is due to systematic variations in their center-to-limb profiles in the methane bands. Their model is too bright at disk center and too cold at the limb. Lowering the methane mixing ratio and using different values for the (poorly known) methane and hydrogen absorption coefficients would improve their fit. Alternatively, a high deep methane mixing ratio may be possible if methane is subsaturated, which we did not consider in our models. Subsaturated models will produce warmer brightness temperatures while increasing the deep methane mixing ratio will produce colder temperatures. As a result, these effects will partially cancel and combined may produce a more reasonable fit. There may also be uncertainties in the millimeter spectral line parameters that are not accounted for here. Laboratory measurements of H$_2$S absorption in the millimeter are extremely limited and not at Neptune’s tropospheric conditions . Updated measurements of the trace gas absorption lines under cold, ice giant conditions would improve, and perhaps reconcile, models of these multi-wavelength observations.
While we disagree with the deep methane mixing ratio computed by @Karkoschka2011, we find general agreement with their observed latitude trends. Both our study and theirs find enhancements in trace gases from $32^{\circ}-12^{\circ}$S and $2^{\circ}-20^{\circ}$N, depletion from $90^{\circ}-50^{\circ}$S and north of 20$^{\circ}$N, and intermediate values from $50^{\circ}-32^{\circ}$S. Our sole inconsistency is from $12^{\circ}$S$-2^{\circ}$N, where we also depleted trace gases. However, at these latitudes, @Karkoschka2011 lowered the tropospheric haze optical depth compared to the surrounding latitudes. If their haze optical depth was increased here, we would expect them to need to decrease the methane abundance from $12^{\circ}$S$-2^{\circ}$N to fit their data. @LuszczCook2016 also investigated methane profiles in cloud free regions of Neptune with the OSIRIS integral field spectrograph in the H and K broad bands. These data probe altitudes higher than ALMA and HST/STIS and so are insensitive to the deep troposphere methane abundance. These authors saw tentative evidence of meridional variations in the methane profile, qualitatively consistent with @Karkoschka2011, but remarked that their parameterizations could not fully characterize the true shape of the methane profile since they were only sensitive to $P < 2.5$ bar.
In the equatorial region, our ALMA residual maps show clear latitudinal structure undetected in prior radio studies of Neptune. North of the equator, we detect low brightness temperatures that are consistent with supersaturating H$_2$S and a slight increase in the deep CH$_4$ abundance. In the visible and near-IR, Neptune’s equatorial region is quiescent, lacking bright cloud activity compared to the dramatic stormy mid-latitudes. However, there is evidence of more cloudy activity just north of the equator than south of it, agreeing with our finding of CH$_4$ enrichment from $2^{\circ}-20^{\circ}$N and depletion from $12^{\circ}$S$-2^{\circ}$N. Figure \[fig:featurehist\] shows a histogram counting the number of bright cloud features versus latitude from *Voyager*, the Hubble Space Telescope, and H-band Keck maps which were tracked in five papers: @Limaye1991 [@Sromovsky2001b; @Martin2012; @Fitzpatrick2013; @Tollefson2018]. By eye, there appears to be a clear correlation between our ALMA defined latitude bands and latitudes where the number of features transitions from low-to-high or vice-versa, particularly in the *Voyager* data. This suggests that the banded structure seen by ALMA from $1-10$ bar exists up to the visible cloud deck. Specifically, at the central latitudes within the 12$^{\circ}$S$-$2$^{\circ}$S band, there is a persistent scarcity of activity observed over a 30 year period. In contrast, features are seen regularly between 2$^{\circ}$S$-$20$^{\circ}$N, suggesting a larger source of condensible methane. This agrees with our best fit model which depletes CH$_4$ to $\sim1.1\%$ in the cloud-free latitudes and enriches it to $\sim2.2\%$ in the cloud-prevalent latitudes. The enhancement of methane in this region is consistent with a moist convective origin of a recently discovered large bright storm centered at 2$^{\circ}$N [@Molter2019]. Moreover, this model may explain inconsistencies between the thermal wind equation and observed vertical wind shear measurements. @Fitzpatrick2013 and @Tollefson2018 detected vertical wind shear at Neptune’s equator by tracking bright cloud features in the H- and K’-bands with Keck, with the K’-band probing higher altitudes with features that have, on average, stronger retrograde velocities than features seen in the H-band. @Tollefson2018 showed that the equator must be methane-rich and warm compared to mid-latitudes if the thermal wind equation holds.
![Histogram counting the number of tracked bright cloud features versus latitude. Dashed lines delineate the latitude bands used in the modeling of the ALMA data and correspond to variations in the residual brightness temperature. Counts are in one-degree latitude bins and latitudes have been converted to planetographic coordinates. Since the *Voyager* data would otherwise dominate the count and perceived trend, these data are split apart, showing the *Voyager* counts on the left and post-*Voyager* counts on the right. These data come from five papers, labeled as follows: Li91 - [@Limaye1991]; Ma12 - [@Martin2012]; To18 - [@Tollefson2018]; Fi14 - [@Fitzpatrick2013]; Sr01 - [@Sromovsky2001b].[]{data-label="fig:featurehist"}](tracked_features_hist_split.pdf){width="0.75\linewidth"}
![Schematic of Neptune’s global circulation inferred from ALMA observations, adapted from Fig. 21 in @dePater2014. Black arrows outline the circulation pattern, which extends from the stratosphere down to 40 bar. The biggest change from the sketch in @dePater2014 is our circulation cells are narrowed near the equator and we prefer a subsiding region which is just south of the equator (aligning with the $12^{\circ}$S$-2^{\circ}$N range seen in our residual map), instead of hemispheric symmetric cells. The gray rectangle between $1-10$ bar shows the sensitivity of the ALMA spectral bands (see Fig. 1). Dark-gray (off-white) rectangles are the latitude ranges where the residual brightness temperature is colder (warmer) than the background (see Fig. 4). At these latitudes, both H$_2$S and CH$_4$ are enriched (depleted) (see Table 5). The light-gray rectangles are latitude ranges where the residual brightness temperature is similar to the background. Here, H$_2$S and CH$_4$ equal their nominal value, or only H$_2$S is depleted. The locations of Visible/NIR clouds are illustrated with blue vertical patches. Most clouds are seen at the mid-latitudes (peak cloud counts are centered near $25^{\circ}$S and $30^{\circ}$N, see Fig. 15) and are seen high in the atmosphere (e.g., those tracked in @dePater2014). However, clouds just north of the equator are also seen and are deeper in the atmosphere (@dePater2014, @Tollefson2018, and Fig. 15). Neptune’s south polar feature (SPF) and south polar (SP) cloud are also placed. Marked cloud layers (CH$_4$-ice, H$_2$S-ice, NH$_4$SH, H$_2$O-ice, Solution Cloud) on the right are assumed to be independent of latitude (see Fig. 1). Also indicated are where high and low temperatures are measured in the MIR (Fletcher et al. 2014). Poleward of $\sim50^{\circ}$N, seeing is cut off in all observations as Neptune’s north pole is tilted away from the observer (rectangular diagonal hatches). Downwelling motions are assumed at Neptune’s north pole in order to complete the circulation diagram.[]{data-label="fig:circulation"}](circulation_sketch.pdf){width="0.75\linewidth"}
The thermal profile inferred from MIR observations, the horizontal distribution of clouds, and observed regions of enriched/depleted air all relate to vertical circulation motions. Based on multi-wavelength observations, @dePater2014 suggest a vertically extended, hemispheric symmetric double-cell pattern with upwelling at the mid-latitudes and downwelling at the equator and poles from the stratosphere down to the deep troposphere ($\sim$40 bar). The ALMA, HST/STIS, and OSIRIS results suggest more detailed circulation, particularly near the equator. Our analysis is consistent with the prediction of vertically extended cells, going from the stratosphere down to $\sim40$ bar, since there is a clear alignment between the ALMA detected latitudinal bands and the observed distribution of cloud features. However, we argue that the equatorial region is more intricate than what was outlined by @dePater2014, with upwelling from 2$^{\circ}$N$-$20$^{\circ}$N and downwelling from 12$^{\circ}$S-2$^{\circ}$N. A narrower circulation cell than predicted by @dePater2014 centered just south of the equator would partially explain the differing predictions between these studies. Figure \[fig:circulation\] gives a sketch of our proposed global circulation pattern.
Conclusion {#S:7}
==========
Spatially resolved millimeter maps of Neptune are presented in three ALMA bands, spanning $95-242$ GHz. These maps have unprecedented sensitivities in the millimeter, ranging between $0.1-0.6$K, and resolutions down to $0.3"$, roughly one-eighth of Neptune’s diameter. The observed emission is mainly modulated by the opacity due to H$_2$S absorption, CIA of H$_2$ with H$_2$, He, and CH$_4$, and *ortho/para* H$_2$. We used the radiative transfer code Radio-BEAR to vary the abundance profiles of these gases in order to model the brightness temperature across Neptune’s disk. Our main conclusions are as follows:
1. The disk-averaged brightness temperature of Neptune in the millimeter and centimeter matches a model where: the temperature-pressure profile follows a dry adiabat; H$_2$S, CH$_4$, and H$_2$O are enriched by 30$\times$ their protosolar values, while NH$_3$ is held at $1\times$ its protosolar value; intermediate H$_2$ is assumed. This is referred to as the ‘nominal’ model.
2. Subtracting the nominal beam-convolved model from the data shows clear brightness temperature variations across Neptune’s disk. We identify latitudes in between significant transitions in the brightness temperature: 90$^{\circ}-66^{\circ}$S, 66$^{\circ}-55^{\circ}$S, 55$^{\circ}-32^{\circ}$S, 32$^{\circ}-12^{\circ}$S, 12$^{\circ}$S$-$2$^{\circ}$N, 2$^{\circ}-20^{\circ}$N, and northward of 20$^{\circ}$N. These bands are at least the size of the ALMA synthesized beam and are apparent in all maps. Relative to the nominal model, brightness enhancements of $1-3$K are seen at $90^{\circ}$S$-$$66^{\circ}$S and $66^{\circ}$S$-$$55^{\circ}$S. Negative temperature residuals between $0.5-1.5$K are seen from 32$^{\circ}-12^{\circ}$S and 2$^{\circ}-20^{\circ}$N. These bands align with regions transitioning from high or low counts in the number of cloud features versus latitude, suggesting that the banded structure we see in the ALMA data exists up at the visible cloud deck. As a result, these identified latitudinal bands may be indicative of a zonal wind profile that is more complex than hitherto considered.
3. At the south polar cap, our best fit model depletes the deep atmospheric abundance of both H$_2$S to 1.5$\times$ the protosolar value and CH$_4$ to 0.55$\%$ (11.5$\times$S). This is consistent with models fitting VLA and CARMA data [@dePater2014; @LuszczCook2013b]. Between 55$^{\circ}-32^{\circ}$S and northward of 20$^{\circ}$N, our best fit models deplete H$_2$S to $10\times$S while keeping CH$_4$ at the nominal value of 1.44$\%$ (30$\times$S). From 32$^{\circ}-12^{\circ}$S and 2$^{\circ}$N$-$20$^{\circ}$N, H$_2$S is supersaturated and the deep abundance of CH$_4$ is enriched to $2.2-2.9\%$ ($45-60\times$S). From 12$^{\circ}$S$-2^{\circ}$N, we deplete H$_2$S to 10$\times$ protosolar and CH$_4$ to 1.1$\%$ ($22.5\times$S). Warm brightness temperatures relative to the nominal model are consistent with dry, subsiding air. Conversely, cold brightness temperatures are consistent with moist, rising air. Our results are, therefore, consistent with an intricate global circulation system that extends from the cloud deck to deep in the atmosphere.
These ALMA maps are evidence of a more complex zone and belt structure that may be yet unresolved in the zonal wind field. An ice-giant probe and high-resolution spacecraft imaging characterizing the zonal wind structure in detail would help settle any inconsistencies between Neptune’s velocity, thermal, and compositional profiles.
Acknowledgements
================
The authors wish to thank Arielle Moullet for helping with the data reduction and imaging process. This work has been supported in part by the National Science Foundation, NSF Grant AST-1615004 to UC Berkeley and by NASA Headquarters: under the NASA Earth and Space Science Fellowship program Grant NNX16AP12H to UC Berkeley.
This paper makes use of the following ALMA data: ADS/JAO.2016.1.00859.S. ALMA is a partnership of ESO (representing its member states), NSF (USA) and NINS (Japan), together with NRC (Canada), NSC and ASIAA (Taiwan), and KASI (Republic of Korea), in cooperation with the Republic of Chile. The Joint ALMA Observatory is operated by ESO, AUI/NRAO and NAOJ. The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc.
Asplund, M., Grevesse, N., Sauval, A. J., Scott, P., 2009, Annual Review of Astronomy and Astrophysics, 47, 481 Baines, K. H., Mickelson, M. E., Larson, L. E., Ferguson, D. W., 1995, Icarus, 114, 328 Brogan, C. L., Hunter, T. R., Fomalont, E. B., 2014, in the Proceedings of the 2014 Synthesis Imaging Workshop held in Socorro, NM on May 13-20, 2014 Butler, B. J., Hofstadter, M., Gurwell, M., Orton, G., Norwood, J., 2012. in AAS/Division for Planetary Sciences Meeting Abstracts, Vol. 44, 504.06 Conrath, B., Gierasch, P., Ustinov, E. A., 1998, Icarus, 135, 501 Conrath, B. J., Flasar, F. M., Gierasch, P. J., 1991, JGRA, 96, 18931 de Pater, I., Fletcher, L. N., Luszcz-Cook, S., et al. 2014, Icarus, 237, 211 de Pater, I. & Massie, S. T., 1985, Icarus 62, 143 de Pater, I., Mitchell, D. L., 1993, JGRE, 98, 5471 de Pater, I. & Richmond, M., 1989, Icarus, 80, 1 de Pater, I., Romani, P. N., Atreya, S. K., 1991, Icarus, 91, 220 de Pater, I., Sault, R., Wong, M. H., et al. 2019, Icarus, 322, 168 Dyudina, U. A., Ingersoll, A. P., Ewald, S. P., et al. 2008. Science, 319, 1801 Fitzpatrick, P. J., de Pater, I., Luszcz-Cook, S., Wong, M. H., Hammel, H. B., 2014, Ap&SS, 350, 65 Fletcher, L. N., Achterberg, R. K., Greathouse, T. K., et al. 2010, Icarus, 208, 337 Fletcher, L. N., de Pater, I., Orton, G. S., et al. 2014, Icarus, 231, 146 Fletcher, L. N., Irwin, P. G. J., Orton, G. S., et al. 2008. Science, 319, 79 Hammel, H. B., Sitko, M. L., Lynch, D. K., et al. 2007, , 134 Iino, T. & Yamada, T., 2018, , 155, 92. Irwin, P. G., Toledo, D., Garland, R., et al. 2019. Icarus, 321, 550 Joiner, J., Steffes, P. G., Noll, K. S., 1992. ITMTT, 40, 1101 Karkoschka, E. & Tomasko, M. G., 2011, Icarus, 211, 780 Limaye, S. S. & Sromovsky, L. A., 1991, JGRA, 96, 18941 Lindal, G. F., 1992, , 103, 967 Luszcz-Cook, S., de Kleer, K., de Pater, I., Ádámkovics, M., Hammel, H., 2016, Icarus, 276, 52 Luszcz-Cook, S. & de Pater, I., 2013b, Icarus 222, 379 Luszcz-Cook, S., de Pater, I., Wright, M., 2013a, Icarus, 226, 437 Luszcz-Cook, S., de Pater, I., Ád ámkovics, M., Hammel, H., 2010, Icarus, 208, 938 Martin, S. C., de Pater, I., Marcus, P., 2012, Ap&SS 337, 65 Massie, S. T. & Hunten, D. M., 1982, Icarus 49, 213 Molter, E., de Pater, I., Luszcz-Cook, S., et al. 2019, Icarus, 321, 324 Orton, G. S., Encrenaz, T., Leyrat, C., Puetter, R., Friedson, A. J., 2007, A&A 473, L5 Sromovsky, L. A., Fry, P. M., Dowling, T. E., Baines, K. H., Limaye, S. S., 2001, Icarus, 149, 459 Tollefson, J., de Pater, I., Marcus, P. S., et al. 2018, Icarus, 311, 317 Trafton, L. M., 1967, , 147, 765 Wallace, L., 1980., Icarus, 43, 231 Weidenschilling, S. & Lewis, J., 1973, Icarus 20, 465
[^1]: This code is available at: <https://github.com/david-deboer/radiobear> .
[^2]: We use the protosolar values from @Asplund2009: C/H$_2$ = 5.90E-4; N/H$_2$ = 1.48E-4; O/H$_2$ = 1.07E-3; S/H$_2$ = 2.89E-5.
[^3]: The methane profiles in @Karkoschka2011 are constructed by increasing the methane mixing ratio at a constant rate below 1.2 bar: $\sim0.15$ bar/$\%$ at $6^{\circ}$S and $\sim0.6$ bar/$\%$ at 45$^{\circ}$S. There is a transition region using intermediate rates from $20^{\circ}$S$-45^{\circ}$S. A $4\%$ deep mixing ratio is assumed. For $P < 1.2$ bar, methane follows the saturation vapor pressure curve.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
Alice Schoenauer Sebag\
Altschuler&Wu lab, Dpt of Pharm. Chem.\
UCSF, San Francisco, CA 94158\
`[email protected]`\
Marc Schoenauer\
INRIA-CNRS-UPSud-UPSay\
TAU, U. Paris-Sud, 91405 Orsay\
`[email protected]`\
Michèle Sebag\
CNRS-UPSud-INRIA-UPSay\
TAU, U. Paris-Sud, 91405 Orsay\
`[email protected]`\
bibliography:
- 'bibliography.bib'
title: 'Stochastic Gradient Descent: Going As Fast As Possible But Not Faster'
---
Related work {#background}
============
\[sec:soa\]
{#eve}
involves two components: a learning rate adaptation scheme, which ensures that the learning system goes as fast as it can; and a catastrophic event manager, which is in charge of detecting undesirable behaviors and getting the system back on track.
Agnostic learning rate adaptation {#sec:speedup}
---------------------------------
Catastrophic event management {#sec:PH}
-----------------------------
Experiments
===========
Experimental results {#sec:experimental_results}
--------------------
Discussion {#sec:discu}
==========
### Acknowledgments {#acknowledgments .unnumbered}
We heartfully thank Steve Altschuler and Lani Wu for making this work possible, supporting it, as well as for very insightful discussions. We also thank Yann Ollivier and Sigurd Angenent for other insightful discussions, and the anonymous reviewers of a preliminary version of this paper for their accurate and constructive comments.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: '[ The use of levitated nanospheres represents a new paradigm for the optomechanical cooling of a small mechanical oscillator, with the prospect of realising quantum oscillators with unprecedentedly high quality factors. We investigate the dynamics of this system, especially in the so-called self-trapping regimes, where one or more optical fields simultaneously trap and cool the mechanical oscillator. The determining characteristic of this regime is that both the mechanical frequency $\omega_M$ and single-photon optomechanical coupling strength parameters $g$ are a function of the optical field intensities, in contrast to usual set-ups where $\omega_M$ and $g$ are constant for the given system. We also measure the characteristic transverse and axial trapping frequencies of different sized silica nanospheres in a simple optical standing wave potential, for spheres of radii $r=20-500$nm, illustrating a protocol for loading single nanospheres into a standing wave optical trap that would be formed by an optical cavity. We use this data to confirm the dependence of the effective optomechanical coupling strength on sphere radius for levitated nanospheres in an optical cavity and discuss the prospects for reaching regimes of strong light-matter coupling. Theoretical semiclassical and quantum displacement noise spectra show that for larger nanospheres with $r~\gtrsim~100$nm a range of interesting and novel dynamical regimes can be accessed. These include simultaneous hybridization of the two optical modes with the mechanical modes and parameter regimes where the system is bistable. We show that here, in contrast to typical single-optical mode optomechanical systems, bistabilities are independent of intracavity intensity and can occur for very weak laser driving amplitudes.]{}'
address:
- 'Department of Physics and Astronomy, University College London, Gower Street, London WC1E 6BT, United Kingdom'
- 'Department of Physics and Astronomy, University College London, Gower Street, London WC1E 6BT, United Kingdom'
- 'Department of Physics and Astronomy, University College London, Gower Street, London WC1E 6BT, United Kingdom'
- 'Institut for Theoretical Physics, Universität Erlangen-Nürnberg, Staudtstraße 7, 91058 Erlangen Germany'
- 'Max Planck Institute for the Science of Light, Günther-Scharowsky-Straße 1/Bau 24, 91058 Erlangen Germany'
- 'ICFO - Institut de Ciencies Fotoniques Av. Carl Friedrich Gauss, 308860 Castelldefels, (Barcelona), Spain '
- 'Department of Physics and Astronomy, University College London, Gower Street, London WC1E 6BT, United Kingdom'
author:
- 'T. S. Monteiro'
- 'J. Millen'
- 'G. A. T. Pender'
- Florian Marquardt
- 'D. Chang'
- 'P. F. Barker'
title: 'Dynamics of levitated nanospheres: towards the strong coupling regime'
---
Introduction {#introduction .unnumbered}
============
Extraordinary progress has been made in the last half-dozen years [@Review; @Kipp] towards the goal of cooling a small mechanical resonator down to its quantum ground state and hence to realise quantum behavior in a macroscopic system. Implementations include cavity cooling of micromirrors on cantilevers [@Metzger; @Arcizet; @Gigan; @Regal]; dielectric membranes in Fabry-Perot cavities [@membrane]; radial and whispering gallery modes of optical microcavities [@Schliess] and nano-electromechanical systems [@NEMS]. Indeed the realizations span 12 orders of magnitude in size [@Kipp], up to and including the LIGO gravity wave experiments. In 2011 two separate experiments [@Teufel; @Chan] achieved sideband cooling of micromechanical and nanomechanical oscillators to the quantum ground state. In Ref. [@Asymm], spectral signatures (in the form of asymmetric displacement noise spectra) of quantum ground state cooling were further investigated. Corresponding advances in the theory of optomechanical cooling have also been made [@Brag; @Paternostro; @Marquardt; @Wilson].
Over the last year or so, a promising new paradigm has been attracting much interest: several groups [@Isart; @Zoller; @Barker; @Ritsch; @Isart2; @Opto1] have now investigated schemes for optomechanical cooling of levitated dielectric particles, including nanospheres, microspheres and even viruses. The important advantage is the elimination of the mechanical support, a dominant source of environmental noise which can heat and decohere the system.
In general, these proposals involve two fields, one for trapping and one for cooling. This may involve an optical cavity mode plus a separate trap, or two optical cavity modes: the so-called “self-trapping” scenario.
Mechanical oscillators in the self-trapping regime differ from other optomechanically-cooled devices in a second fundamental respect (in addition to the absence of mechanical support): the mechanical frequency, $\omega_M$, associated with centre-of-mass oscillations is not an intrinsic feature of the resonator but is determined by the optical field. In particular, it is a function of one or both of the detuning frequencies, $\Delta_1$ and $\Delta_2$, of the optical modes. In previous work [@Opto1], we analysed cooling in the self-trapped regime and found that the optimal condition for cooling occurs where both fields competitively cool and trap the nanosphere. This happens when $\omega_M$ is resonantly red detuned from both the detuning frequencies i.e. $\omega_M(\Delta_1,\Delta_2) \sim -\Delta_{1,2}$ so the relevant resonant frequencies are mutually interdependent. Most significantly, the effective light-matter coupling strength $g$ also depends on the detunings.
The effective coupling strength, ${\tilde g} = g\sqrt{n}$ (the optomechanical coupling rescaled by the square root of photon number) determines whether one can attain strong coupling regimes in levitated systems such as recently observed in a non-levitated set-up [@strong]. It determines too whether one may access other interesting dynamics, both in the semiclassical and quantum regimes. We consider in particular the possibility of simultaneous hybridization of the two optical modes with the mechanical mode; here, we consider also the implications of a static bistability, which, unusually, occurs also in the limit of quite weak driving in the levitated self-trapped system.
In the present work, we investigate theoretically and experimentally the strength of the optomechanical coupling. In particular, we present experimental measurements of the mechanical frequency of a nanosphere trapped in an optical standing wave in order to investigate the optical coupling as a function of the size of the nanosphere.
In section 1 we review the theory of the cavity cooling and dynamics of a self-trapped system, and in section 2 we employ the experimentally measured size dependence of the coupling to determine the range of optomechanical coupling strengths accessible in a cavity. The data suggests that the most effective means to attain stronger coupling will be to employ larger nanospheres of typical radii $r=200-300$nm. Our work suggests that increasing photon number by stronger driving (and by implication increasing rescaled coupling strengths) will not prove an effective alternative, since in the present system we show ${\tilde g}
\propto {n}^{1/4}$ rather than $\sqrt{n}$, so the rescaled coupling increases very slowly with laser input power.
In section 3, we in investigate the cooling and dynamics. In sec. 3.1, we review the corresponding cooling rate expressions obtained from quantum perturbation theory (or linear response theory). In sec. 3.2 we report a study of the corresponding semiclassical Langevin equations and compare them with fully quantum noise spectra; we compare also quantum, semiclassical and perturbation theory results for levitated nanospheres. In section 4 we investigate novel regimes of triple mode hybridization, coincident with static bistabilities, which the present study shows are experimentally accessible given the large optomechanical coupling strengths associated with $r=100-300$nm nanospheres. In section 5 we describe the experimental study which provides data from which the size-dependence of the coupling may be inferred. In Section 6, we conclude.
Theory: Quantum Hamiltonian for a nanosphere in a cavity
========================================================
We approximate the equivalent cavity model by a one-dimensional system, with centre-of-mass motion confined to the axial dimension. In this simplified study, we consider only the axial dynamics: for the cavity system, we will have a much smaller tranverse frequency relative to the axial frequency, i.e. $\omega_t \ll \omega_a$ and there is little mixing between these transverse and axial degrees of freedom.
We consider the dynamics of the following Hamiltonian: $$\begin{aligned}
\frac{\hat H}{\hbar}&=& -\Delta_1{\hat a}_1^\dagger {\hat a}_1 -\Delta_2 {\hat
a_2}^\dagger {\hat a}_2
+ \frac{{\hat P}^2}{2m\hbar} -A {\hat a}_1^\dagger {\hat a}_1 \cos^2
(k_1 {\hat x}-\phi_1)
\nonumber \\
&- & A{\hat a}_2^\dagger {\hat a}_2\cos^2 (k_2{\hat x}-\phi_2)
+ E_1({\hat a_1}^\dagger+ {\hat a}_1) + E_2({\hat a}_2^\dagger + {\hat a}_2).
\label{Hamiltonian}\end{aligned}$$
Two optical field modes of a high finesse cavity ${\hat a}_{1,2}$ are coupled to a nanosphere with centre-of-mass position $x$. The parameter $A$ (dependent on the nanosphere polarizability), determines the depth of the optical standing-wave potentials. We investigate the case where both modes competitively cool and trap the nanosphere, in contrast to previous schemes [@Isart; @Zoller] where one optical field is exclusively responsible for trapping, while the other is exclusively responsible for cooling. $\hat H$ is given in the rotating frame of the laser which drives the modes with amplitudes $E_1$ and $E_2=R E_1$ respectively, where $R$ represents the ratio of driving amplitudes for the two modes. We restrict ourselves to the regime $0 \leq R \leq 1$, since we consider the most general case where both optical modes contribute to the trapping as well as the cooling. Thus we can define mode 1 simply as the mode which is more strongly driven and mode 2 as the mode which is (except where $R=1$) more weakly driven. The detunings $\Delta_{j}=\omega^j_L-\omega^{j}_c$ for $j=1,2$ are between the input lasers and the corresponding cavity mode of interest, and $\phi_{1,2}$ represents the phases of the optical potentials.
The two fields could represent two modes generated by the same laser field, or they could be generated by two independent lasers. Nonetheless, since the particle motion is confined to within one wavelegth, one can make the approximation $k_1\approx k_2 \equiv k$. Previous studies generally consider $\phi_1=0 , \phi_2=\pi/4$ to be convenient, since then the anti-node of one field coincides with a purely linear potential of the other optical field, but we may also consider general values of $\phi_1-\phi_2$. One can write corresponding equations of motion:
$$\begin{aligned}
{\ddot {\hat x}} & = & -\frac{\hbar k A}{m}\sum_j {\hat a}_j^\dagger {\hat a}_j
\sin 2(k {\hat x}-\phi_j)
-\Gamma_M {\dot {\hat x}} \nonumber\\
{\dot {\hat a}_j}& = & i \Delta_j {\hat a}_j -iE_j +i A {\hat a}_j\cos^2 (k
{\hat x}-\phi_j) -\frac{\kappa}{2} {\hat a}_j,
\label{Heisenberg}\end{aligned}$$
where $j=1,2$ for the two optical-mode realisation. Additional damping terms have also been added: $\frac{\kappa}{2} {\hat a}_i$ accounts for photon losses due to mirror imperfections and the $\Gamma_M {\dot {\hat x}}$ term for mechanical damping. The above should also include quantum noise terms arising from (say) shot noise or gas collisions: for brevity, the quantum noise terms are left out until sec.3.
We consider here the linearised dynamics; we replace operators by their expectation values and perform the shifts about equilibrium values such as ${\hat a}_j(t) \to \alpha_j
+ {\hat a}_j(t)$, and ${\hat x} \to x_0 + {\hat x}(t)$. The values for the equilibrium photon fields (e.g. for the two-mode case) are $\alpha_{1} = -iE_1 \left[\frac{\kappa}{2} - i\Delta_{1}^x \right]^{-1} \ $ and $\alpha_{2} = -iRE_1\left[\frac{\kappa}{2}- i\Delta_{2}^x \right]^{-1}$. The equilibrium position is then given by the relation $-\frac{\sin
2(kx_0-\phi_1)}{\sin 2(kx_0-\phi_2)} = {|\alpha_2|^2}/{|\alpha_1|^2}$, by numerical solution of the equation,
$$\begin{aligned}
-\frac{\sin 2(kx-\phi_1)}{\sin 2(kx-\phi_2)}= R^2\frac{|\frac{\kappa}{2} -
i\Delta^x_{1}|^2}{|\frac{\kappa}{2} - i\Delta^x_{2}|^2},
\label{equil}\end{aligned}$$
where $\Delta_j^x=\Delta_j +A \cos^2 (kx_0-\phi_j)$. As usual we consider the dynamics of the fluctuations via the linearised equations. To first order, the linearised equations of motion, in the shifted frame, are:
$$\begin{aligned}
{\ddot {\hat x}}& = & - \omega^2_M {\hat x} - \frac{\hbar kA}{m}\sum_j
(\alpha_j^* {\hat a}_j+ \alpha_j {\hat a}_j^\dagger)\sin 2(kx_0-\phi_j)
-\Gamma_M {\dot {\hat x}} \nonumber\\
{\dot {\hat a}_j} & = & i \Delta_j^x {\hat a}_j - ikA \alpha_j {\hat x} \sin
2(kx_0-\phi_j) -\frac{\kappa}{2} {\hat a}_j.
\label{a2s}\end{aligned}$$
The resulting effective mechanical harmonic oscillator frequency is:
$$\begin{aligned}
\omega_M^2= \frac{2\hbar A k^2}{m} \sum_j |\alpha_j|^2 \cos{2(kx_0-\phi_j)}.
\label{freqs}\end{aligned}$$
We can restrict ourselves to real equilibrium fields. We take $\alpha_j ={\tilde
{\alpha}}_j e^{-i\theta_j}$ then transform ${\hat a}_j \to {\hat a}_j e^{i\theta_j}$. Thus $(\alpha_j^*
{\hat a}_j+ \alpha_j {\hat a}_j^*)\equiv {\tilde {\alpha}}_j( {\hat a}_j+ {\hat
a}_j^\dagger)$. We also rescale the mechanical oscillator coordinates ${\hat x} \to
\sqrt{2}X_{zpf}{\hat x}$ and ${\hat p} \to \sqrt{\hbar m\omega_M} {\hat p}$, where $X_{zpf}=\sqrt{\frac{\hbar}{2 m\omega_M}} $ is the zero-point fluctuation length scale. Hence, $\frac{{\hat P}^2}{2m}+\frac{1}{2} m \omega_M^2 {\hat x}^2 \to \frac{\hbar
\omega_M}{2}({\hat x}^2+{\hat p}^2)$.\
Below we drop the tilde so the equilibrium field values ${\tilde {\alpha}}_j
\equiv \alpha_j$ are real. Using field operators ${\hat x}= ( {\hat b} +{\hat b}^\dagger)/\sqrt{2}$, the linearised dynamics for a two-optical mode system would correspond to an effective Hamiltonian:
$$\begin{aligned}
\frac{\hat H_{Lin}}{\hbar}&=& -\Delta_1^x{\hat a}_1^\dagger {\hat a}_1
-\Delta_2^x{\hat a}_2^\dagger {\hat a}_2
+ \omega_M(\Delta_1^x,\Delta_2^x){\hat b}^\dagger {\hat b}\nonumber\\
& +& g_1(\Delta_1^x,\Delta_2^x)\alpha_1( {\hat a}_1+ {\hat a}_1^\dagger)(
{\hat b} +{\hat b}^\dagger) +g_2(\Delta_1^x,\Delta_2^x)
\alpha_2( {\hat a}_2+ {\hat a}_2^\dagger)( {\hat b} +{\hat b}^\dagger).
\label{LinHam}\end{aligned}$$
![(Colour online) Size dependent effects in the magnitude of the optomechanical coupling parameter ${\tilde g}_{1}$. It is assumed that cavity parameters would correspond to ${\tilde
g}_{1}=10^6$Hz at $r=150$nm, for photon numbers $n_1=10^9$. For comparison, the value of $A$ is also shown, as are the experimental and simulated frequencies $\omega_a \equiv \omega_M =2\pi f_M$. The $\omega_M$ are scaled by a factor of 10 for clarity and in fact, values of $\omega_M \sim 1$MHz are quite realistic in optical cavities.[]{data-label="beadg"}](BEADG.eps){height="3.3in"}
Towards strong light-matter coupling with two optical cavity modes
==================================================================
The Hamiltonian in Eq. \[LinHam\] appears analogous in form to standard, well-studied optomechanical Hamiltonians, albeit with two optical modes rather than one. However, it differs in one important respect: in this case, both the mechanical frequency $\omega_M(\Delta_1^x, \Delta_2^x)$ and the optomechanical coupling strengths $g_{1,2} \equiv g_{1,2}(\Delta_1^x,\Delta_2^x)$ are not fixed and depend on the detunings. The fact that the frequencies of the three modes (two optical, one mechanical) are interdependent makes the dynamics different from other optomechanical set-ups, where the equilibrium mechanical frequency (i.e. excluding shifts arising from the fluctuations) is intrinsic to the mechanical oscillator.
There is considerable interest in achieving strong-coupling, which leads to regimes of light-matter hybridization. The corresponding mode splitting has been observed experimentally [@strong]. In typical set-ups, these regimes are reached if the rescaled effective optomechanical coupling exceeds the damping rates i.e. ${\tilde g} = g \sqrt{n} \gtrsim \kappa,
\Gamma_M$, where $n \sim |\alpha|^2$ is the cavity photon number. Even if the unnormalized coupling is weak, strong-coupling regimes may be achieved by increasing the driving power and thus increasing intracavity photon numbers.
In the present levitated system, a particularly interesting regime would involve triple-mode hybridization enabling, for example, the coupling of the two modes of light via the mechanical mode. However, here, mode hybridization (for which ${\tilde g}_{1,2}= g_{1,2} \alpha_{1,2} \gtrsim \kappa, \Gamma_M$) depends non-trivially on the detunings.
We argue that large coupling cannot be easily achieved by increasing the driving power, since ${\tilde g}_{1,2} \propto n_{1,2}^{1/4}$ and thus increases slowly with the driving strength. We can show, by a simple argument, that increasing the nanosphere size provides the most effective means to attain strong coupling.
For the self-trapped system, optomechanical coupling strengths are $g_j=
\sqrt{2}kA X_{zpf} \sin 2(kx_0-\phi_j)$ and depend on the detunings via $x_0$. Note also that $
X_{zpf}=\sqrt{\hbar/(2m\omega_M)}$ here too depends on the detunings via $\omega_M$.
For triple mode hybridization, $\omega_M \sim \Delta_{1} \sim \Delta_{2} $. For convenience, we also take $\phi_1=0, \phi_2=\pi/4$. Then, since $\tan 2kx_0 = {|\alpha_2|^2}/{|\alpha_1|^2}$, we can re-write Eq. \[freqs\]:
$$\begin{aligned}
\omega_M^2= \frac{2\hbar A k^2}{m \cos{2kx_0}}|\alpha_1|^2.
\label{freq2}\end{aligned}$$
We consider near symmetric driving of the two optical modes for which $R \sim 1$ and thus $kx_0 \approx \phi/2=\pi/8$ so $\omega_M \sim \left({\frac{2\hbar A
k^2}{m}}\right)^{1/2} n_{1,2}^{1/2} $. Hence,
$$\begin{aligned}
{\tilde g}_{1,2} \sim \left( {\frac{\hbar k^2}{4}}\right)^{1/4} \left( \frac{A^3
n_{1,2}}{ m}\right)^{1/4}.\end{aligned}$$
Since the optomechanical coupling increases only very slowly with cavity photon number the most effective means to reach strong coupling regimes is to increase the nanosphere size to the maximum practical value ($r\sim 200-250$nm).
For the ideal case where the nanosphere radius $r$ is small [@Zoller] ie. $\lambda \gg r$, then $A(r\ll \lambda) \equiv A_0(r)$ where the small nanosphere coupling takes the form: $$\begin{aligned}
A_0(r) = \frac{3}{2} \frac{\epsilon_r-1}{\epsilon_r-2} \frac{V_s}{V_c} \omega_L
\label{ideal}\end{aligned}$$ where $V_s= 4/3\pi r^3$ is the sphere volume (and hence $m=V_s \rho$ where the density $\rho=2000$Kgm$^{-3}$ for silica). In turn, $V_c=\pi (w/2)^2 L$ is the cavity volume, where $w\approx 40\,\mu$m is the cavity waist and $L \simeq 0.5 - 1$cm is the cavity length.
For larger nanospheres, the measured size-dependent corrections must be applied. In the experiments described below, we find that the mechanical oscillation frequency is modulated by a finite size correction $\omega_M(r) = \omega_M(r\simeq 0) f(r)$ (see Fig. \[exp\_trap\_freq\] and description of the measurement of $f(r)$ in section 5 below). Thus, since: $$\begin{aligned}
\omega_M^2 \propto \frac{A(r)}{m},
\label{freq2}\end{aligned}$$
then $A(r)\equiv A_0(r) f^2(r)$ and the coupling is in turn modulated by the finite size correction. The experimental results suggest that for $r\lesssim 200$nm, then $f(r)\sim 1$ and $A(r) \simeq A_0(r)$.
For example, for $r=150$nm, $L=1$cm and $w=40\,\mu$m, then $A_0\simeq 8
\times 10^5$Hz. For reasonable values of cavity decay constants $\kappa \simeq 2-8 \times
10^5$Hz, then for $n_1 \sim 10^{9}$, $$\begin{aligned}
{\tilde g}_{1,2} \sim 5.4 \times 10^{-6} \left({\frac{A^3
n_{1,2}}{m}}\right)^{1/4} \simeq 10^6\,\mathrm{Hz} \gg \kappa\end{aligned}$$
For $r \lesssim 200$nm, ${\tilde g}_{1,2} \propto r^{3/2}$. Thus a 200nm sphere provides an optomechanical coupling about an order of magnitude larger than a 40-50nm sphere. To achieve a comparable increase in coupling by photon number enhancement would require increasing the driving power by a factor of order $10^4$.
A more careful analysis, including the effects of the finite-size correction function $f^2(r)$ is shown in Fig. \[beadg\]. We see that ${\tilde g}_{1}$ (and for $n_1\sim
n_2$, also ${\tilde g}_{2}$) reaches a maximum value for $r\simeq 300$nm before falling to zero. Other maxima for larger $r$ do not provide a larger value of ${\tilde g_{1,2}}$. Furthermore, they have the disadvantage that they may enhance photon recoil heating effects. For comparison, the value of $A$ is also shown, overlaid on the experimental and simulated frequencies.
Dynamics
========
Optomechanical damping
----------------------
A previous study [@Opto1], using rescaled coordinates, investigated the full parameter space of two optical mode cooling. Here we investigate more carefully the effect of non-zero mechanical damping. Using linear response theory, we can extract cooling rates from the equations Eq. \[a2s\]:
$$\Gamma_{opt}=
\left[S_1(\omega_M)+S_2(\omega_M)-S_1(-\omega_M)-S_2(-\omega_M)\right],
\label{gamma}$$
where
$$S_j(\omega)=\frac{ |\alpha_j|^2 g^2_j \kappa}{[\Delta_j^x-\omega]^2 +
\frac{\kappa^2}{4}},
\label{gamma12}$$
for $j=1,2$. Net cooling occurs for $\Gamma_{opt} <0$. Although the above is quite similar in form to standard optomechanical expressions, as explained previously, rather different behaviour is observed since here $\omega_M$ and $g_j$ are both dependent on the $\Delta_j^x$.
From quantum perturbation theory we can show that $R_{n\to m}$, the rate of transition from state $n$ to $n+1$ is: $R_{n \to n+1}=(n+1) \left( S_1(\omega_M)+S_2(\omega_M) \right)$ while $R_{n \to n-1}=n \left( S_1(-\omega_M)+S_2(-\omega_M) \right)$
For $n>>1$, then $R_{n \to n+1}- R_{n \to n-1}$ gives the cooling rate of Eq.\[gamma\]. However, with the exact expressions we can show that the equilibrium mean phonon number is
$$\langle n
\rangle_{PT}=\frac{S_1(\omega_M)+S_2(\omega_M)}{
S_1(-\omega_M)+S_2(-\omega_M)-S_1(\omega_M)-S_2(\omega_M)}.
\label{neq}$$
![(Colour online) Maps of cooling rates calculated from Eq. \[gamma\] for parameters $R=1.0$ and $R=0.5$. Blue corresponds to cooling, yellow/white to heating. The white lines indicate the locus of the single field resonances (where $-\Delta_{1}^x=\omega_M$ or where $-\Delta_{2}^x=\omega_M$). The detunings are given in units of $A$ and are dimensionless. For $R=1$ it is clear that there is a deep, maximum cooling region at a double resonance where the two white lines intersect and both optical fields cool simultaneously. It is also evident that there is a strong cooling resonance for $+\Delta_{1,2\pm}^x=\omega_M$. For $R=0.5$, three cooling resonances $-\Delta_{1\pm}^x=\omega_M$, $-\Delta_{2\pm}^x=\omega_M$ and $+\Delta_{1}^x=\omega_M$ merge to give a very broad strong-cooling [*region*]{}, quite insensitive to detuning $\Delta_2$ over a range of over 1MHz. Here $A=\kappa/2= 0.3$MHz and the input power into mode 1 corresponds to 2mW.[]{data-label="Cool"}](Cool.eps){height="3.3in"}
In Fig. \[Cool\] we show colour maps comparing the cooling and minimum phonon numbers for both $R=0.5$ and $R=1$ respectively. The cooling behaviour was investigated previously in [@Opto1]. In this case, for each fixed detuning $\Delta_1$ there are up to three cooling resonances (at three different values of $\Delta_2$), where strong damping is observed (and similarly for each fixed $\Delta_2$). This is in contrast to single optical mode schemes where there is a single cooling resonance for which $\Delta_{1}=-\omega_M$ or $\Delta_{2}=-\omega_M$. For the $R=0.5$ map the three cooling resonances merge, giving a single extended cooling region of about 1MHz width.
For $R=1$ the map has a high degree of symmetry, since the role of the two optical modes is interchangeable. The figures show that the largest cooling rates are found in the double resonance region, making it the most favourable region to work in.
The equilibrium phonon number in Eq. \[neq\] concerns only the idealised situation where there is a very good vacuum, negligible photon recoil heating and thus no mechanical damping or heating effects. For small $r \lesssim 200$nm spheres, we assume recoil heating is negligible [@Zoller] and the dominant source of mechanical damping is background gas collisions, which provide an effective mechanical damping $\Gamma_M=\frac{8}{3} \pi \frac{m_g}{m_s} r^2 n_g \bar{v}_g$ [@Isart; @Zoller] where $m_g / m_s$ is ratio of the gas particle’s mass to that of the sphere, $n_g$ is the gas number density and $\bar{v}_g$ is the mean gas velocity for a room temperature thermal distribution.
It can be shown that the perturbation theory argument above can be adapted to obtain equilibrium phonon numbers for a given cooling rate $\Gamma_{opt}$:
$$\begin{aligned}
\langle{n}\rangle_{PT}=\frac{\frac{k_BT_B}{\hbar\omega_M} {\Gamma_M} +
\left[S_1(\omega_M )+
S_2(\omega_M )\right]}
{\Gamma_M + |\Gamma_{opt}|},
\label{PT}\end{aligned}$$
where $T_B\simeq 300$K. Alternatively, the final equilibrium temperatures $T_{eq} = \frac{\Gamma_M T_{B}
+ |\Gamma| T_{vac}}{\Gamma_M+|\Gamma|}$, where $T_{vac}$ is the equilibrium oscillator temperature which would have been obtained in a perfect vacuum.
Quantum and semiclassical noise spectra
---------------------------------------
Although we investigate only a two optical mode system, generalization to more optical modes is straightforward. We consider a set of equations of motion, for $j=1,...N$: $$\begin{aligned}
{ \dot { \hat b} } & = & -( i\omega_M (\Delta_1^x,...\Delta_j^x)+
\frac{\Gamma_M}{2} ) {\hat b} +
i\sum_j g_j(\Delta_1^x,...\Delta_j^x) ({\hat a}_j+{\hat a}_j^\dagger) +
\sqrt{\Gamma_M} {\hat b}_{in} \nonumber \\
{\dot {\hat a}_j} & = & (i \Delta_j^x-\frac{\kappa}{2}) {\hat a}_j + i
g_j(\Delta_1^x,...\Delta_j^x)\alpha_j({\hat b}+{\hat b}^\dagger)
+ \sqrt{\kappa}{\hat a}_{in}^{(j)}, \nonumber\\
\label{QM}\end{aligned}$$
where the optomechanical strengths $g_j(\Delta_1^x,...\Delta_j^x)= -kA X_{ZPF}
\sin 2(kx_0-\phi_j)$ depend on the detunings (as does the mechanical frequency $\omega_M$). In the two mode case we consider, we take $\phi_1=0$ and $\phi_2=\pi/4$.
The optical modes are subject to photon shot noise, while the mechanical modes are subject to Brownian noise from collisions with gas molecules in the cavity. For the photon shot noise, we assume independent lasers and uncorrelated zero temperature noise for which $ \langle {\hat a}_{in}^\dagger(t') {\hat a}_{in}(t)\rangle =0$, while $ \langle {\hat a}^{(i)}_{in}(t') {\hat a}^{(j)\dagger }_{in}(t)\rangle
=\delta(t-t')\delta_{ij}$. For the gas collisions, we take $ \langle {\hat b}_{in}(t') {\hat
b}_{in}^\dagger(t)\rangle =(n_B+1)\delta(t-t')$ and $ \langle {\hat b}_{in}^\dagger(t') {\hat b}_{in}(t)\rangle=
n_B\delta(t-t')$ where the number of surrounding bath phonons $n_B \approx \frac{k_BT}{\hbar \omega_M}$.
The above equations can be integrated in frequency space to obtain analytical expressions for the displacement noise spectra for the arbitrary mode case. We can evaluate the displacement spectrum $S_{xx}(\omega) \equiv \langle |x(\omega)|^2\rangle_{QM}= \frac{1}{2\pi} \int e^{-i\omega \tau} \langle
x(t+\tau) x(t)\rangle d\tau$. We obtain:
$$\begin{aligned}
\langle |x(\omega)|^2\rangle_{QM} |M(\omega)|^2 & = & {\Gamma_M} \left[
|\chi_M(\omega)|^2 n_B + |\chi_M(-\omega)|^2 (n_B+1)\right]\nonumber \\
& + & \frac{\kappa}{2}
|\mu(\omega)|^2 \sum_{j=1,2} g_j^2 |\chi_{jo}(-\omega)|^2,
\label{qnoise}\end{aligned}$$
where the $\chi(\omega)$ represent optical and mechanical susceptibilities: $$\begin{aligned}
\chi_{jo}(\omega)=\left[-i(\omega+\Delta_j^x)+\frac{\kappa}{2}\right]^{-1}; \
\chi_M(\omega) =\left[-i(\omega-\omega_M)+\frac{\Gamma_M}{2}\right]^{-1}.\end{aligned}$$ with $ \mu(\omega)=\chi_M(\omega)-\chi_M^*(-\omega)$ and $ \eta_j(\omega)=
\chi_{jo}(\omega)-\chi_{jo}^*(-\omega)$; then also $M(\omega)= 1+ \mu(\omega)\sum_j g_j^2 |\eta_{j}(\omega)|^2 $.
![(Colour online) Comparison between quantum and semiclassical displacement spectra for gas pressures of 1mBar and at near vacuum pressure in a strong cooling region. At high vacuum, the ground state is approached and thus, for the quantum spectrum, the blue sideband vanishes. At higher pressure there is good agreement between the quantum and classical results. Spectra near double resonance for input power $P_1=7$mW, $A=\kappa=3\times
10^5$Hz, $\Delta_1=-1.5$MHz, $\Delta_2=-0.68$MHz, $R=0.5$. Some hybridization between the mechanical mode and optical mode 1 is seen in the characteristic double-peak sideband structure.[]{data-label="Noise"}](Noise.eps){height="3in"}
, we show equilibrium phonon numbers obtained from Eqs. (\[PT\]), (\[qnoise\]) and (\[SSC\]) i.e. perturbation theory, the analytical quantum noise formula and semiclassical Langevin equations respectively. Agreement between quantum and semiclassical results is excellent, less so for perturbation theory at low pressures.[]{data-label="Phonon"}](Phonon.eps){height="3in"}
![Triple mode splitting. For $A=\kappa=3$MHz, even at quite high pressures (here 1mBar), mode splitting is seen in the noise spectra of the optical modes. In all the plots, $\Delta_1 = -1.15$MHz is held fixed while $\Delta_2$ is swept from 0 to -1.6MHz (for an input power of 2mW into mode 1, while $R=0.5$). (a) Shows noise spectra for both optical mode 1 and mode 2. Three way hybridization between the mechanical and both optical modes appears clearly (highlighted in the bold blue line). For clarity, some of the strongest peaks have been truncated in height. In (b) three avoided crossings are apparent. The dominant character of each normal mode is indicated by the colour (black is mechanical, blue is optical mode 1, red is optical mode 2). When $-\Delta_2$ is large, there is no mixing. However as $\Delta_2 \to 0$, there is strong mixing and the dominant character of each normal mode changes from light to matter (or vice versa) as an avoided crossing is encountered. Panel (c) shows the cooling and indicates strong cooling at each of the avoided crossings. []{data-label="eigensplit"}](NMODE1.eps){height="3.5in"}
We compare the quantum displacement with corresponding semiclassical solutions in the steady state. The linearised two mode system Eq. \[LinHam\], in matrix form corresponds to a standard problem [@Milburn]. Inclusion of the noise arising from gas collisions or laser shot noise yields a set of corresponding Langevin equations: $ \frac{d{\bf X(t)}}{dt}= {\bf A}{\bf X}+ {\bf B E}(t)$, where ${\bf A}$ is termed the drift matrix. Its eigenvalues give the stabilities and eigenfrequencies of the system’s normal modes, while the noise is determined by ${\bf B}$, a constant diagonal matrix. The elements of the random noise matrix are assumed to be $\delta$-correlated $\langle E_i(t) E_j(t')\rangle= \delta(t-t') \delta_{ij}$. Methods for obtaining the solution for the steady state correlation functions of this system, under conditions of stability, i.e. where all the eigenvalues of ${\bf A}$ have negative real parts, are well-known [@Milburn]. The required noise spectra, or autocorrelation functions, in frequency space are:
$$\begin{aligned}
{\bf S}(\omega)=\frac{1}{2\pi} _{-\infty}\int^{\infty} e^{-i\omega \tau} \langle
{\bf X}(t+\tau) {\bf X^T}(t)\rangle d\tau,\end{aligned}$$
where,
$$\begin{aligned}
{\bf S}(\omega)=\frac{1}{2\pi} \left({\bf A}+ i \omega{\bf I}\right)^{-1}
{\bf B}{\bf B^T}\left({\bf A^T}- i \omega{\bf I}\right)^{-1},
\label{SSC}\end{aligned}$$
where the diagonal matrix ${\bf B}{\bf B^T}$ has elements $\left((n_B+\frac{1}{2}){\Gamma_M},
(n_B+\frac{1}{2}){\Gamma_M},\frac{\kappa}{2},\frac{\kappa}{2},\frac{\kappa}{2},
\frac{\kappa}{2}\right)$. From the above, the noise spectra of all modes may be calculated. Eq.[SSC]{} yields semiclassical sideband spectra, symmetrical in $\omega$.
In Fig. \[Noise\], we compare semiclassical displacement spectra calculated from Eq. \[SSC\] with corresponding quantum results obtained from Eq. \[qnoise\]. At high pressures (and hence high phonon occupancy) there is excellent agreement between semiclassical and quantum results. At low pressures (near ground state cooling) however, the quantum spectrum shows a characteristic asymmetry, such as was observed recently in experiments on photonic cavities [@Asymm].
![Maps of the displacement noise spectra $S_{xx}(\omega)$, showing mode splitting for similar parameters to Fig. \[Noise\], near the quantum limit, except that here $\Delta_1 = -1.5$MHz is held fixed while $\Delta_2$ (vertical axes) is swept. (a) Shows the semiclassical spectrum which is symmetric in frequency $\omega$. (b) Shows the quantum spectrum which is asymmetric. In both cases, triple hybridization appears clearly and is seen near $\Delta_2 \approx -1.0$MHz. (c) shows the corresponding behaviour for $A=\kappa/2$, showing that the triple peak structure has disappeared. Note that $\log{ S_{xx}(\omega)}$ is plotted.[]{data-label="SPLIT"}](SPLIT.eps){height="3in"}
Comparison between perturbation theory, semiclassical and quantum results
--------------------------------------------------------------------------
The equilibrium variance (and hence the final phonon number) of the mechanical oscillator is,
$$\begin{aligned}
\langle x^2\rangle=\frac{1}{2\pi} _{-\infty}\int^{\infty} \langle
|x(\omega)|^2\rangle d\omega,\end{aligned}$$
thus, the final equilibrium temperature of the mechanical oscillator after optomechanical cooling is $k_BT_{eq}=1/2m \omega_M^2 \langle X^2\rangle$. Noting the rescaling $ \langle X^2\rangle= 2X_{ZPF} \langle x^2\rangle$ and setting $k_BT_{eq}=(\langle n \rangle +1/2) \hbar \omega_M$, we can write $\langle x^2\rangle= \langle n \rangle+1/2$.
Using Eqs. (\[QM\]), (\[SSC\]) and (\[PT\]), we can investigate final equilibrium phonon numbers (and the minimum achievable for levitated self-trapped spheres) comparing quantum, semiclassical and perturbation theory respectively. In Fig. \[Phonon\], we compare the corresponding equilibrium phonon numbers, $\langle n \rangle_{QM}$, $\langle n \rangle_{SC}$ and $\langle n \rangle_{PT}$ respectively for the unusual triple cooling resonance region shown in Fig. \[Cool\]. Cooling to near the ground state $ \langle n \rangle \sim 0$ is possible for a pressure of order $10^{-6}$mBar, even for modest driving powers of 2mW and values of $A \simeq 3 \times 10^5$Hz corresponding to spheres of order $r \simeq 100$nm.
![Mode mixing and bistability for $A=3\kappa=6 \times 10^{5}$Hz. We consider a relatively low input power of 0.37mW. (a) plots the optomechanical cooling rate (blue indicates cooling, brown indicates heating). The red dashed line indicates the locus of bistability as a function of the detunings $\Delta_1$ and $\Delta_2$. The discontinuity in the cooling can be discerned near the strong doubly resonant cooling region. (b) shows displacement noise spectra $S_{xx}(\omega)$ as a function of $\omega$ plotted along $\Delta_1=-1$MHz (along black horizontal line in upper panel). We sweep in increasing $\Delta_2$. At the point where $\Delta_1=-1$MHz intersects the bistability, a discontinuity in the noise spectra is apparent. On one side of the discontinuity there is strong hybridization between the mechanical mode and optical mode 2; this changes abruptly across the discontinuity to hybridization between the mechanical mode and optical mode 1 for larger $\Delta_2$. This may allow for control of entanglement between the modes. The map corresponds to near-vacuum conditions so the system is near the quantum ground state in this regime (as evidenced by the asymmetric sidebands).[]{data-label="bi"}](BISTAB2.eps)
Strong coupling regimes: triple mode splitting and bistability
==============================================================
The multimode, or at least two mode, self-trapping regime may permit new possibilities for position sensing and for controlling entanglement between two optical modes and the mechanical resonator. Here we investigate regimes where such effects are clearly apparent. The implications of the measurement for the accessible range of optomechanical coupling strengths suggests that multiple hybridization and bistability are quite accessible with reasonable cavity parameters.
In Fig. \[eigensplit\] we investigate the complex behaviour of the eigenmode frequencies of the self-trapped, levitated system. On the left panels (Fig. \[eigensplit\] (a)) we plot the noise spectra of the two optical modes, which exibit sidebands near $\omega
\simeq \omega_M$ since the corresponding optical fields are modulated by the motion of the mechanical oscillator. Here we fix one detuning ($\Delta_1=-1.15$MHz) and look at the behaviour as the other detuning is varied. The sidebands are displaced in frequency and split: one effect is simply due to the dependence of $\omega_M$ on $\Delta_j$ (unique to the levitated system); it arises from the calculation of the equilibrium fields and frequencies. The other effect is due to normal mode mixing (hybridization of light and matter modes) arising from the linearised equations. If $\omega_M \simeq \Delta_1 \simeq \Delta_2$ simultaneous hybridization is observed, provided ${\tilde g}_{1,2} \gtrsim \kappa$. Figure \[eigensplit\] (b) shows that there are several distinct avoided anti-crossings, where the dominant character of each eigenmode changes; if two crossings coincide, the spectra show a characteristic triple-peak structure (symmetric about $\omega=0$ in the semiclassical regime shown here). Panel (c) shows that the corresponding cooling rate is enhanced at each avoided level crossing.
In Fig. \[SPLIT\] we plot displacement spectra corresponding to Fig. \[Noise\], but over a range of values of $\Delta_2$. A log-scale is used for $S_{xx}(\omega)$. The triple mixing which can appear when two avoided crossings nearly coincide is clearly apparent at $\Delta_2 \simeq 1$MHz.
Static bistability in a cavity of varying length has been seen experimentally [@bistab]. The potential for generating entanglement has recently been investigated in an optomechanical system [@Ghobadi]; however, a relatively high laser power $P \sim 50$mW is required.
For the self-trapped systems, the incoherent sum of the optical standing-wave potentials $\cos^2 (kx-\phi_1)$ and $\cos^2 (kx-\phi_2)$ does not by itself produce a double-well structure; nevertheless, as we see below, in combination with optomechanical shifts, bistabilities are observed, even for weak driving. Whether a double-well structure emerges, or not, is completely independent of the driving power (where $ P \propto E_1^2$) and can emerge at very low input powers, as we demonstrate below. It is easy to see that the levitated particle moves in an effective static potential $V(x)$ where:
$$\begin{aligned}
\frac{dV(x)}{dx}= \hbar k AE_1^2 \left[\frac{\sin 2(kx-\phi_1)}{|(\kappa/2) -
i\Delta_{1}(x)|^2}+
R^2\frac{\sin 2(kx-\phi_2)}{|(\kappa/2) -
i\Delta_{2}(x)|^2}\right], \nonumber\\
\label{pot}\end{aligned}$$
and
$$\begin{aligned}
V(x)= \hbar A E_1^2\left[\tan^{-1}
\left({\frac{\Delta_{1}(x)}{\frac{\kappa}{2}}}\right)+
R^2\tan^{-1}\left({\frac{
\Delta_{2}(x)}{\frac{\kappa}{2}}}\right)\right]. \nonumber \\
\label{pot1}\end{aligned}$$
Here, note that the shifted detuning $\Delta_{j}(x)=\Delta_j+A \cos^2
(kx-\phi_j)$ is dependent on $x$, not the equilibrium displacement $x_0$. This potential admits two stable equilibrium points over parameter regimes where $A \gg \kappa$ (in practice, such bistability is observed already for $A/\kappa \simeq 3$). It is evident that the driving power factors out, so does not affect the shape of the potential, providing only a scaling factor. We show in Fig. \[bi\] that for a high $A/\kappa$ ratio simultaneous hybridisation and bistability co-exist: we show that that for $P=0.37$mW , $A=3\kappa$ , $R=0.15$ and modest photon numbers $n_1 \sim 10^8$ we can switch discontinuously from hybridisation between the mechanical mode and optical mode 1, to hybrization between the mechanical mode and mode 2. We take $\phi_1=0, \phi_2=\pi/4$. In the noise spectra, the switch is heralded by a large zero-frequency peak in the displacement spectra, which is clearly apparent in Fig. \[bi\].
Experiment
==========
Current experimental status: Loading protocols and variation of trap frequency with radius
------------------------------------------------------------------------------------------
![(Colour online) Schematic diagram of the standing wave trap. It is formed from two counter-propagating 1064nm beams focused inside a vacuum chamber. Light at 532nm enters via one fibre to image the sphere. Images and measurement of the axial and transverse position of the trapped nanosphere as a function of time are measured by a CCD camera and quadrant cell photodiode (QPD) respectively, through a long working distance microscope outside the vacuum chamber.[]{data-label="trap"}](small_diag.ps)
We have built a simple standing wave dipole trap to develop protocols for loading a single nanosphere into the trap to confirm that nanospheres with a range of radii around 100nm can be trapped. Importantly, we have measured the variation in trap frequency with sphere radius so that realistic values of optomechanical coupling strengths, for a given nanosphere radius, can be included in our models. In this section we explain how the size-dependent modulation function $f(r)$, used in the theory section above to obtain $A(r)$ and the optomechanical coupling strengths, was measured.
A schematic of the standing wave trap used in our experiments is shown in Fig. \[trap\]. The standing wave trap consists of two focused beams that counter-propagate and overlap near their foci. The two laser beams, derived from the same laser at a wavelength of 1064nm, enter the trapping region via optical fibres. The light exiting the fibres is focused using aspheric lenses (Thorlabs C140TME) with a focal length of 1.45mm and numerical aperture 0.55. The power in each trapping beam after it has passed through each lens is measured to be $150\pm10$mW, and the best focused beam waist (radius) is theoretically 1.7$\mu$m. To optimize the alignment of the trap, we maximize the light through-coupled from one fibre into the other. This is accomplished by mounting one optical fibre and its aspheric lens on an XYZ flexure stage. The alignment is done inside the vacuum chamber at atmospheric pressure.
A long working distance microscope (Navitar Zoom 6000 system, with up to 45x zoom) is used to image the trapped sphere. The image is split into two using a beamsplitting plate, with one image directed to a CCD camera for diagnostics and the other aligned onto a quadrant cell photodiode (QPD) which measures position fluctuations as a function of time in two orthogonal axes. We define the axial direction as that along which the trapping light propagates, and the transverse direction as the orthogonal axis in the focal plane of our imaging system. Light at 532nm is used to illuminate the trapped sphere, as the QPD is more sensitive at this wavelength. The green beam enters the system via one of the optical fibres, as shown in Fig. \[trap\], and a filter is used to stop 1064nm light reaching the detectors. The power of the 532nm beam is 10 mW.
![\[beads\] An image of a string of 100nm diameter beads trapped in the standing wave trap. A single bead is trapped by continually blocking and unblocking one of the trapping beams until only one sphere is trapped. ](beads)
Silica (SiO$_{2}$) nanospheres, manufactured by Microspheres-nanospheres and Bangs Laboratories, are introduced into the trapping region at atmospheric pressure via an ultrasonic nebulizer (Omron NE-U22). These spheres range in radius from 26nm to 510nm and are suspended in methanol. The nanosphere solution is sonicated using an ultrasonic bath for at least an hour before trapping to prevent clumping. Once introduced into the trapping region the methanol surrounding the spheres rapidly evaporates and the spheres are trapped over many fringes of the standing wave, as shown in Fig. \[beads\]. As our imaging system does not have single fringe resolution we cannot determine if more than one sphere is trapped in a single fringe by this method. However, this information can be inferred from the relative intensity of the light scattered from the trapped spheres and also by the reduced stability of the particles in the trap when more than one particle is trapped. To reduce the number of trapped particles the trapping light is briefly blocked and unblocked. This is repeated until a single sphere is visible in the trap. At this pressure, where there is a strong damping force from air the sphere can be held in the trap indefinitely.
To measure the trap frequency the air is pumped from the system, and at this point no more spheres enter the trap, as without air-damping their velocity is too high. The air pressure in the trap is reduced to 5mbar, so that clear trap frequencies can be obtained from the power spectrum of the position fluctuations of the trapped sphere, as recorded on the QPD. Example power spectra are shown in Fig. \[fft\]. Above 5mbar the damping of the motion in the trap due to air broadens the peak in the power spectrum so that finding an accurate trap frequency is difficult. Below pressures of 5mbar the spheres become unstable in the trap and escape. This is most likely due to radiometric forces which have been compensated for in other experiments using feedback techniques [@Ashkin; @Raizen]. At 5mbar the damping rate due to gas collisions is significantly less than our lowest measured trap frequencies, and thus the measured frequency at this pressure is a good approximation to the bare trap frequency which would be measured in vacuum without damping.
The angular axial trap frequency for a small polarizable particle in a standing wave is $\omega_a = 2\pi f_a = \sqrt{\frac{4 \alpha k^2 I_0}{m\epsilon_0 c}}$, where the polarizability of a sphere of refractive index $n$ is $\alpha = 4 \pi
\epsilon_0 r^3 \frac{n^2-1}{n^2+2}$. The maximum intensity in the radial center of each equal intensity beam is given by $I_0$, and $k$ is the magnitude of the wavevector of each beam. The sphere has mass $m= 4/3 \pi\rho r^3$, radius $r$ and density $\rho \simeq 2000$kgm$^{-3}$. The transverse trap frequency is given by $\omega_t=\sqrt{\frac{8\alpha I_0}{ m\epsilon_0 c w^2}}$, where $w$ is the focused spot size (radius) of the two counter-propagating beams. From these expressions the ratio of the trap frequencies is given by $\omega_a/\omega_t=k
w/\sqrt{2}$.
![(Colour online) Power spectra at 5mbar calculated from a measurement of the position of a trapped 200.1nm diameter nanosphere as a function of time, using a QPD. (a) The transverse frequency, and (b) an axial frequency. Outlier points are due to electronic noise. Red lines show fitted Gaussian functions, from which the trap frequencies are extracted.[]{data-label="fft"}](fts){height="2.0in"}
The trap frequencies in each axis are determined by fitting measured position fluctuation power spectra using $\frac{2 k_B
T}{m}\frac{\Gamma_0}{(\omega_a^2-\omega^2)^2+\omega^2 \Gamma_0^2}$, where $k_B$ is Boltzmann’s constant and $\Gamma_0$ is the damping rate. The fit to the data is shown in Fig. \[fft\] with $\Gamma_0/2\pi \simeq 2.4$kHz for the axial trap frequency. Several sets of data were taken for different spheres of the same nominal radius and the measured axial and transverse trap frequencies for each size sphere are shown in Fig. \[exp\_trap\_freq\]. The derived trap frequencies for each sphere radius are the average over different experiments at each radius, and the errors are the standard errors in the mean. The uncertainty in the sphere radius is taken from the information supplied by the manufacturer. Two axial (red and green data in Fig. \[exp\_trap\_freq\]) frequencies and one transverse frequency (blue data points) were measured. When a particle is tightly trapped by the optical field only one axial frequency is expected from a single sphere in a standing wave. The lower axial frequency (in green in figure \[exp\_trap\_freq\]) is always observed in the data and this is taken as the true axial frequency. The higher frequency, which is often present in the data, may be due to the trapping of two spheres in a single anti-node, with the higher frequency occurring due to optical binding, which requires further study [@Zamenek]. The higher axial frequency also changes rapidly with sphere size, indicating that it is not the true axial trap frequency, which should be almost constant for the small spheres. The presence of a single axial frequency is, we believe indicitave of having trapped a single sphere.
Although we don’t know the radial dimensions of the beam within the trap we can estimate this value from the ratio of the axial to transverse trap frequencies for small spheres. The spot size from this ratio is $w=\frac{\sqrt{2}}{k}\frac{\omega_a}{\omega_t}$ and for $\frac{\omega_a}{\omega_t}$=9.8 this gives a spot size of $w= 2.3\,\mu$m. Since the trap frequency with two overlapping beams of size 2.3$\mu$m would be equal to 207kHz with a power in each beam of 150mW, and we only measure a maximum axial trap frequency of approximately 40 kHz, we conclude that the particles are trapped in a standing wave formed where the waist of one beam is much larger. If one spot size is 2.3$\mu$m the other would have to be 15$\mu$m. A plot of the calculated axial trapping frequency, found by calculating Maxwell’s stress tensor ([@Barton; @Chang]), is also shown in Fig. \[exp\_trap\_freq\]. Like the experimental data the trapping frequency is constant for small spheres and decreases to approximately zero when the particle size is comparable to the size of the interference pattern produced by the standing wave. At larger radii the force on the particle changes sign and a stable trap is formed in a node of the standing wave, as shown for the particle of radius 510nm. Our measurements confirm that for particle radii less that 200nm the simple dipole model for the nanospheres is adequate for modelling the cooling and dynamics of the nanospheres in an optical cavity utilising 1064nm radiation.
![(Colour online) Measured trap frequencies as a function of sphere radius. Points plotted in green are the axial trap frequency, blue are the transverse trap frequency and the red data points are the higher axial trap frequencies which are believed to be due to optical binding. The solid black line is a theoretical curve derived from a numerical calculation [@Barton].[]{data-label="exp_trap_freq"}](exp_trap_freqs.eps){height="3.5in"}
Our experiments have shown that optical traps without feedback are currently limited to operation at pressures down to a few millibar for all particles that we have measured. In addition this limiting pressure did not change by reducing the intensity by 50%. This radiometric force is due to localized heating of nanosphere and the subsequent heating of the surrounding air. At low pressures, when the mean free path is comparable to the size of the nanosphere, the radiometric force competes with and eventually dominates the dipole force which traps the particle. While feedback techniques have been successful [@Raizen], decreasing the absorption of the nanospheres is another route to minimising radiometric effects. This is feasible since all the spheres we have used in this study are not made of optical quality glass but from colloidally grown nanospheres. Finally, we have also successfully trapped silica spheres in an ion trap at pressures of $10^{-6}$mbar which could be used to load an optical trap formed by a cavity at lower pressures where radiometric forces are not significant.
Conclusions {#conclusions .unnumbered}
===========
We have described a study of the dynamics and noise spectra of self-trapped levitated optomechanical systems. We have been able to show, by combining experimental measurements and theoretical calculations that strong light-matter coupling is attainable over a wide range of particle sizes, and that these can be trapped. The interdependence of the mechanical and optical mode frequencies, unique to self-trapped levitated systems provides a complex and interesting side-band structure, including multi-mode mixing and bistabilities which we aim to explore experimentally. These conclusions are supported by measurements of trap frequency made in an optical standing trap where we have demonstrated a protocol for loading a single nanosphere in a single antinode.
[*Acknowledgements*]{}: We acknowledge support for the UK’s Engineering and Physical Sciences Research Council.
References {#references .unnumbered}
==========
[99]{}
F. Marquardt and S Girvin, Physics [**2**]{} 40 (2009). T. Kippenberg and K. Vahala, Science, [**321**]{} 1172 (2008). C.H. Metzger and K. Karrai, Nature, [**432**]{} 1002 (2004). O. Arcizet et al, Nature, [**444**]{} 71 (2006). S. Gigan et al, Nature, [**444**]{} 67 (2006). C. A. Regal et al, Nature Phys,[**4**]{} 555 (2008). J. D. Thompson et al, Nature, [**452**]{} 72 (2008); A.M. Jayich et al New J. Phys, [**10**]{}, 095008 (2008). A. Schliesser et al, Nature Physics, [**5**]{} 509 (2009). A. Naik et al , [*Nature*]{} [**443**]{} 193 (2006); A. D. Armour et al Phys.Rev.Lett [**88**]{} 148301 (2002). J. D. Teufel, T. Donner, D. Li, J. H. Harlow, M. S. Allman, K. Cicak, A. J. Sirois, J. D. Whittaker, K. W. Lehnert, R. W. Simmonds, Nature, [**475**]{}, 359 (2011). J. Chan, T. P. Mayer Alegre, A. H. Safavi-Naeini, J. T. Hill, A. Krause, S. Groblacher, M. Aspelmeyer and O. Painter, Nature, [**478**]{}, 89 (2011). A. H. Safavi-Naeini, J. Chan, J. T. Hill, T. P. Mayer Alegre, A. Krause, and O. Painter, Phys. Rev. Lett., [**108**]{} 033602 (2012). V. B. Braginski and F. Y. Khalili [*Quantum Measurement*]{}, Cambridge University Press, 1992. S. Bose, K. Jacobs and P. L. Knight,Phys.Rev. A [**59**]{} 3204 (1999). M. Paternostro et al, New J. Phys, [**8**]{}, 107 (2006). F. Marquardt et al, Phys.Rev.Lett [**99**]{} 093902 (2007). I. Wilson-Rae et al, Phys.Rev.Lett [**99**]{} 093901 (2007).
O. Romero-Isart, M.L. Juan, R. Quidant and J. I. Cirac, New J. Phys, [**12**]{}, 033015 (2010). D. E. Chang, C. A. Regal, S. B. Papp, D. J. Wilson, O. Painter, H. J. Kimble and P. Zoller, Proc. Natl. Acad. Sci. USA [**107**]{}, 1005 (2010). P. F. Barker and M. N. Shneider, Phys. Rev. A [**81**]{} 023826 (2010). R. J. Schulze, C. Genes and H. Ritsch, Phys. Rev. A [**81**]{} 063820 (2011). O. Romero-Isart et al, Phys. Rev. A [**83**]{} 013803 (2011). G. A. T. Pender, P. F. Barker, J. Millen, F. Marquardt, T. S. Monteiro Phys. Rev. A [**85**]{} 021802 (2012). S. Groblacher et al, Nature, [**460**]{} 724 (2009). A. Ashkin, J. M. Dziedzic, Appl. Phys. Lett. [**28**]{}, 333 (1976) T. Li, S. Kheifets, M. G. Raizen, Nat. Phys. [**7**]{}, 527 (2011) P. Zemanek, A. Jonas, L. Sramek, M. Liska, Opt. Comms.[**151**]{} 273 (1998) J. P. Barton, D. R. Alexander, S. A. Schaub, J. Appl. Phys. [bf 66]{}, 4594 (1989) D. Chang, Matlab code provided for calculation of optical forces in a standing wave. D. F. Walls and G. J. Milburn, [*Quantum Optics*]{},Springer (Heidelberg) 2008. A. Dorsel, J. D. McCullen, P. Meystre,. E. Vignes and H. Walther, Phys. Rev. Lett [**51**]{} 1550 (1983). R. Ghobadi, A. R. Bahrampour and C.Simon, arXiv:1104.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We give a complete proof of the fact that the trace of the curvature of the connection associated to a planar d-web ($d>3$) is the sum of the Blaschke curvature of its sub 3-webs.'
author:
- 'Jean-Paul Dufour\'
title: '**La formule de la trace pour les tissus planaires**'
---
[**Keywords:**]{} planar webs
**Introduction.**
=================
Un $d$-tissu du plan est une famille de $d$ feuilletages d’un ouvert du plan qui sont deux à deux transverses. Il y a plusieurs méthodes équivalentes pour donner ces feuilletages. Il y a des méthodes “explicites” où chaque feuilletage est donné soit par les trajectoires d’un champ ou, ce qui revient au même, en se donnant les pentes $m_i(x,y)$ de chaque feuilletage dans un système de coordonnées $(x,y)$ ou bien encore en se donnant ses intégrales premières $f_i$ (les feuilletages sont donnés par les courbes de niveau des $f_i$). Alain Hénaut a développé la méthode “implicite” qui présente les feuilles comme les trajectoires d’une équation différentielle implicite du type $F(x,y,y')=0$ où $F$ est un polynôme de degré $d$ en $y'$ à coefficients dépendant de $x$ et $y.$ Pour une bibliographie relativement complète sur ce domaine nous renvoyons au livre de J.V. Pereira et L. Pirio de 2015 [@PP] ou au texte de J. V. Periera au Séminaire Bourbaki de 2007 [@JP].
Dans ce travail, on considère un $d$-tissu $W(f_1,\dots ,f_d)$ sur un ouvert $U$ du plan donné (explicitement) par ses $d$ intégrales premières $f_1,\dots ,f_d.$ Un invariant important d’un tel tissu est son “rang”, c’est la dimension de l’espace vectoriel de ses “relations abéliennes” $\sum_{i=1}^d h_i(f_i)=0$ où les $h_i$ sont des fonctions d’une variable, nulles en un point fixé. En 2004 et dans le contexte “implicite” Alain Hénaut [@AH] a montré que l’on pouvait associer au tissu un fibré vectoriel sur $U$ muni d’une connexion $\nabla$ dont l’une des propriétés est que sa courbure est nulle si et seulement si le rang du tissu a la valeur maximale possible $(d-1)(d-2)/2.$ A la même époque L. Pirio a soutenu sa thèse [@P] dans laquelle, entre autre, il revisite des travaux anciens de A. Pantazi [@AP], pour construire une connexion analogue dans le cas “explicite”. En 2005 Olivier Ripoll [@OR], sous la direction d’A.Hénaut, a soutenu une thèse sur ces sujets. Ces deux auteurs ont construit des programmes Maple qui calculent la connexion et sa courbure pour $d=3,$ $d=4$ et $d=5,$ dans le cas “explicite” pour L. Pirio et dans le cas “implicite” pour O. Ripoll. En 2007 Vincent Cavalier et Daniel Lehmann [@CL] ont généralisé les constructions précédentes aux tissus de codimension 1 en dimension $n$ arbitraire, pourvu que ceux-ci soient “ordinaires” (ce qui est toujours le cas pour $n=2$) et tels qu’il existe un entier $k_0$ tel que $d=(n-1+k_0)!/((n-1)!k_0!)$ (ce qui est toujours le cas pour $n=2$ avec $k_0=d-1$). Travaillant dans le cas “explicite” avec les pentes des feuilletages, ils ont construit un fibré vectoriel muni d’une connexion qui généralise celle de A.Pantazi. En 2014 D. Lehmann et l’auteur [@DL] ont reconstruit ce fibré vectoriel et sa connexion à partir des intégrales premières des feuilletages et rédigé un programme Maple qui, non seulement calcule la connexion et la courbure des $d$-tissus plans pour tout $d,$ mais aussi qui peut fonctionner en toute dimension. Dans la suite on note $\nabla$ cette connexion.
En 1933 W. Blaschke [@WB] avait attaché une “courbure” aux 3-tissus du plan. La courbure de Blaschke est celle de $\nabla$ dans le cas particulier $d=3.$
Dans leurs thèses et travaux suivants L. Pirio et O. Ripoll ont conjecturé le résultat suivant.
[**FORMULE DE LA TRACE.**]{} [**La trace de la courbure de $\nabla$ est la somme des courbures de Blaschke des sous-3-tissus $W(f_i,f_j,f_k)$ de $W(f_1,\dots ,f_d)$.**]{}
Nous voyons la courbure de Blaschke, et la trace de la courbure de $\nabla ,$ comme des 2-formes à valeurs scalaires sur le plan ; cela donne sens à la formule précédente.
On peut voir les prémisses de cette formule dans les travaux de N. Mihaileanu [@NM] et A. Pantazi [@AP]. L. Pirio et O. Ripoll indépendamment en ont donné des preuves pour $d=4,$ $5$ et $6$ Ils ont aussi présenté un plan de démonstration précis pour $d$ quelconque. L. Pirio a donné à cette formule le nom de “formule de Mihaleanu.”
Dans ce travail nous proposons une démonstration complète de ce résultat basée sur la méthode utilisée pour construire le programme de D. Lehmann et de l’auteur [@DL]. Elle est un peu longue mais tous les calculs sont explicités et élémentaires. La méthode “explicite” a l’avantage de permettre des raisonnements par récurrence sur le nombre $d$ car on peut comprendre plus facilement ce qui advient de la connexion quand on rajoute une ($d+1$)-ème fonction. Dans le cadre “implicite” on ne travaille qu’avec les polynômes symétriques de ces fonctions et l’effet de l’ajout d’une nouvelle fonction est un peu plus caché. Ceci dit, le détail des calculs que l’on trouvera dans certaines parties de la démonstration laisse penser que l’on pourrait avoir une démonstration plus simple dans le contexte “implicite”.
Construction de la connexion du $d$-tissu.
==========================================
Nous rappelons, sans écrire tous les détails, comment nous construisons le fibré vectoriel et la connexion dans le texte de D. Lehmann et l’auteur[@DL].
On travaille au voisinage d’un point $P$ du plan et on impose que les fonctions $f_1$,...,$f_d$ soient nulles en $P$. On choisit des coordonnées locales $x$ et $y$ qui s’annulent elles aussi en $P.$ Si $f$ est une fonction définie sur un voisinage de $P,$ $f_x$ (resp. $f_y$) désigne la dérivée de $f$ par rapport à $x$ (resp. $y$). Ainsi $f_{xx}$ désigne la dérivée seconde de $f$ par rapport à $x$....
Les relations abéliennes de notre tissu sont les relations $$\sum_{i=1}^{d}h_i(f_i)=0$$ où les $h_i$ sont des fonctions d’une variable qui s’annulent à l’origine.
Pour étudier cette relation on dérive successivement ses deux membres par rapport aux deux variables.
On note $\omega^r_i=h_i^{(r)}(f_i)$ , où $h_i^{(r)}$ désigne la dérivée $r$-ième de $h_i.$
A l’ordre 1 on a les deux équations $$\sum_{i=1}^{d}\omega^1_i{ f_x}=0$$$$\sum_{i=1}^{d}\omega^1_i{ f_y
}=0$$ que l’on récrit sous forme matricielle $$P_2(\omega^1_1,\dots ,\omega^1_d)=0$$ où $ P_2 $ est la matrice jacobienne de $(x,y)\mapsto (f_1(x,y),\dots ,f_d(x,y)).$
A l’ordre 2 on a 3 équations que l’on range en prenant pour première celle qui correspond à la dérivée $\partial\over{\partial x^2}$, la deuxième à $\partial\over{\partial x\partial y}$, la troisième à $\partial\over{\partial y^2}$. On les écrit matriciellement sous la forme $$G^2_{3}(\omega^1_1,\dots ,\omega^1_d)+P_3(\omega^2_1,\dots ,\omega^2_d)=0$$.
Plus généralement, on range les équations d’ordre $r-1$ en imposant l’ordre $${\partial\over{\partial x^{r-1}}}, {\partial\over{\partial x^{r-2}\partial y}},\dots , {\partial\over{\partial y^{r-1}}}$$ pour les dérivations et on les récrit sous la forme matricielle $$G^{r-1}_{r}(\omega^1_1,\dots ,\omega^1_d)+\cdots +G^{2}_{r}(\omega^{r-2}_1,\dots ,\omega^{r-2}_d)+P_{r}(\omega^{r-1}_1,\dots ,\omega^{r-1}_d)=0.$$
Les matrices $G_r^j$ ont des coefficients qui sont des expressions polynomiales des dérivées partielles des $f_i$ d’ordre 1 à $j.$
Lorsque cela nous paraîtra utile pour rendre notre texte plus clair, nous rajouterons l’indice $(f_1,\dots ,f_d)$ à nos matrices. Ainsi on écrira $G^i_{j ;f_1,\dots ,f_d}$ à la place de $G^i_j$ pour préciser quelles sont les fonctions en jeu.
La première étape du programme donné dans [@DL] calcule les coefficients des matrices $G_r^s$ et $P_r$ par récurrence. Nous retiendrons simplement que les matrices $P_r$ ont des colonnes de la forme $$f_i^{r-1}=\left|\matrix{( f_{ix})^{r-1} \cr ( f_{i x})^{r-2}( f_{iy})\cr .\cr .\cr ( f_{ix})^{r-j}( f_{iy})^{j-1}
\cr .\cr. \cr(f_{i y})^{r-1}\cr }\right|$$ et que, si l’on note $ G^{r}(f_i)$ les colonnes de $G_r^{2}$ et $G^{r+}(f_i)$ la $(r-1)$-colonne obtenue en supprimant la dernière composante de $G^r(f_i)$, on a les relations de récurrence $$G^{r+}(f_i)=f_{ix}.G^{r-1}(f_i)+{\partial f^{r-2}_i\over\partial x}.$$
On construit par blocs la matrice à $(d+1)(d-2)/2$ lignes et $(d-2)d$ colonnes $$MM=\left|\matrix{P_2 & 0&.&.&.&0\cr G_3^2 &P_3&0&.&.&0\cr G_4^3&G_4^2&P_4&0&.&0\cr .&.&.&.&.&.\cr .&.&.&.&.&.\cr .&.&.&.&.&.\cr G_{d-1}^{d-2}& G_{d-1}^{d-3}&.&.& G_{d-1}^{2}&P_{d-1} \cr}\right|$$ Le noyau de cette matrice donne le fibré de rang $(d-1)(d-2)/2,$ et de base le plan des $(x,y),$ sur lequel la connexion $\nabla$ sera définie.
On construit $\nabla$ comme suit.
On considère d’abord la matrice par blocs à $(d-2)d$ lignes et $(d-2)d$ colonnes $$\Delta =\left|\matrix{0 & Id_d&0&.&.&0\cr 0 &0 &Id_d&0&.&0\cr .&.&.&.&.&.\cr .&.&.&.&.&.\cr .&.&.&.&.&.\cr 0&. &.&.&0&Id_d\cr A_{d-1}& A_{d-2}&.&.& A_{3}&A_{2} \cr}\right|$$ où $Id_d$ désigne la matrice identité à $d$ lignes et $$A_j=-P_{d}^{-1}.G^j_{d}.$$ On note $D_x$ (resp. $D_y$) la matrice diagonale dont les éléments diagonaux sont $( f_{1 x},\dots ,
f_{d x})$ (resp. $( f_{1 y},\dots ,
f_{d y})).$ On considère maintenant la matrice $DD_x$ (resp. $DD_y$) diagonale par blocs, à $(d-2)d$ lignes, dont les blocs diagonaux sont tous égaux à $D_x$ (resp. $D_y$).
On considère la matrice carrée à $(d-2)d$ colonnes $\Delta_x= DD_x.\Delta$ (resp. $\Delta_y= DD_y.\Delta$).
Les coefficients des deux matrices $\Delta_x$ et $\Delta_y$ dépendent de $(x,y)$ ; elles donnent donc des morphismes du fibré trivial de rang $(d-2)d$ sur le plan. On a une connexion $\nabla^0$ sur ce fibré en prenant $\nabla^0_{\partial\over\partial x}={\partial\over\partial x}-\Delta_x $ et la même chose en remplaçant $x$ par $y.$ On peut voir que cette connexion préserve le sous-fibré donné par le noyau de $MM$.
La connexion $\nabla$ est alors la restriction de $\nabla^0$ au noyau de $MM$.
Pour construire une base du noyau de $MM$ nous procédons comme suit.
On remarque que chacun des blocs “diagonaux” $P_r$ de $MM$ est tel que ses $r$ premières colonnes forment une sous-matrice inversible. Cela nous mène à changer un peu l’écriture de nos variables : on récrit $$(\omega^1_1,...,\omega^1_d;\omega^2_1,...,\omega^2_d;...;\omega^{d-2}_1,...,\omega^{d-2}_d),$$ plutôt sous la forme $$(\omega^1_1,\omega^1_2,\beta_3^1,...,\beta_d^1;\omega^2_1,\omega_2^2,\omega^2_3,\beta_4^2,...,\beta_d^2;...;\omega^{d-2}_1,...,\omega^{d-2}_{d-1},\beta_d^{d-2}).$$ C’est à dire que l’on remplace $\omega_i^r$ par $\beta_i^r$ pour $i> r+1$.
On obtient une base $$B=\{e^1_3,\dots , e^1_{d};e^2_4,\dots ,e^2_{d}; \dots ;e^{d-3}_{d-1},e^{d-3}_d;e^{d-2}_d\}$$ du noyau de $MM$ en prenant pour $e^r_i$ le vecteur du noyau de $MM$ dont les coordonnées $\beta^s_j$ sont toutes nulles sauf $\beta^r_{i}$ qui est 1.
Notons $\Omega_x$ (resp. $\Omega_y$) les matrices de $\nabla_{\partial\over\partial x}$ (resp. $\nabla_{\partial\over\partial y}$) par rapport à la base $B$. Ses coefficients sont obtenus comme suit. On remarque d’abord que si l’on dérive n’importe quel vecteur $e^r_i$ de la base $B$ par rapport à $x$ ou $y$ on obtient un vecteur dont toutes les composantes $\beta^s_j$ sont nulles. Ainsi $\nabla_{\partial\over\partial x}(e^r_i)$ a des composantes $\beta^s_j$ qui sont celles de $-\Delta_x(e^r_i).$ On note $\Omega_{x,s,j}^{ r,i}$ sa composante $\beta^s_j$ ; c’est sa composante sur le vecteur de base $e^s_j.$ Alors $\Omega_x$ est la matrice qui a les coefficients $\Omega_{x,s,j}^{ r,i}$ ; autrement dit, on a $$-\Delta_x(e^r_i)=\sum_{s=1}^{d-2}\sum_{j=s+2}^{d} \Omega_{x,s,j}^{ r,i}e^s_j.$$ On agit de même en permutant $x$ et $y$ pour calculer $\Omega_y.$
Calcul de la trace de la courbure.
==================================
[**Définition.**]{} On considère le tissu $W(f_1,\dots ,f_s)$ donné par ses $s$ intégrales premières locales $f_1,\dots ,f_s$ ($s>2$). On lui associe les matrices $P_{s-1}=P_{s-1;f_1,\dots ,f_{s}},$ $P_{s}=P_{s;f_1,\dots ,f_{s}}$ et $G^2_s=G^{2}_{s;f_1,\dots ,f_{s}}$ définies comme dans le paragraphe précédent. La matrice $P_{s-1}$ est de rang $s-1$ et a un noyau de dimension 1 engendré par un vecteur du type $(X_1,X_2,\dots , X_{s-1},1)$. On appelle [**élément de trace**]{} de $f_1,\dots ,f_s$ et on note $\gamma(f_1,\dots ,f_s)$ la dernière composante du vecteur $$-P_{s}^{-1} G^{2}_{s}(X_1,X_2,\dots , X_{s-1},1).$$
Cet élément de trace est aussi caractérisé par le fait que la matrice par blocs $$M_{f_1,\dots ,f_{s}}=\left|\matrix{P_{s-1;f_1,\dots ,f_{s}} & 0 \cr G^{2}_{s;f_1,\dots ,f_{s}} & P_{s;f_1,\dots ,f_{s}}\cr }\right|$$ a un noyau engendré par un vecteur de la forme $$(X_1,X_2,\dots , X_{s-1},1;Y_1,Y_2,\dots , Y_{s-1},\gamma(f_1,\dots ,f_s)).$$
Etudions maintenant l’expression de la courbure de $\nabla$ dans la base donnée dans la section précédente. Elle a la matrice $$K={\partial\over\partial y}(\Omega_x)-{\partial\over\partial x}(\Omega_y)+\Omega_x.\Omega_y -\Omega_y.\Omega_x.$$ Le commutateur $\Omega_x.\Omega_y -\Omega_y.\Omega_x$ a une trace nulle, donc la trace de la courbure est la trace de la matrice $$KK={\partial\over\partial y}(\Omega_x)-{\partial\over\partial x}(\Omega_y).$$
[**Remarque.**]{} [*Dans un prochain travail avec D. Lehmann nous prouverons que la matrice $K$ a toutes ses lignes nulles sauf la dernière. Donc sa trace se réduit au seul terme diagonal sur la cette dernière ligne. On aurait pu penser que ce résultat serait un ingrédient essentiel de toute preuve de la formule de la trace. Dans la démonstration suivante nous procédons autrement, en n’utilisant que $KK.$*]{}
Pour calculer la trace de $K$, il suffit donc de calculer la trace de $\Omega_x,$ de la dériver par rapport à $y$, puis de retrancher ce que l’on obtient en échangeant les rôles de $x$ et $y$.
La trace de $\Omega_x$ est la somme des $\Omega^{r,i}_{x, r,i}.$
Dans la suite de cette section on calcule $\Omega^{r,i}_{x, r,i}$ pour $r$ et $i $ fixés. Pour cela on rappelle que $e^r_i$ est le $(d-2)d$-vecteur $$(\omega^1_1,\omega^1_2,\beta_3^1,...,\beta_d^1;\omega^2_1,\omega_2^2,\omega^2_3,\beta_4^2,...,\beta_d^2;...;\omega^{d-2}_1,...,\omega^{d-2}_{d-1},\beta_d^{d-2})$$ du noyau de $MM$ dont les coordonnées $\beta^s_j$ sont toutes nulles sauf $\beta^r_{i}$ qui est 1. La forme “triangulaire inférieure par blocs” de $MM$ implique que les $\omega^s_i$ sont tous nuls pour $s<r.$ C’est dire que $e^r_i$ a la forme $$(0,...,0;...;0,...,0;\omega^r_1,...,\omega^r_{r+1},0,...,0,1,0,...,0;\omega^{r+1}_1,...,\omega^{r+1}_{r+2},0,...,0;\omega^{r+2}_1...),$$ où le 1 est à la $i$-ème place entre les deux points virgules qui l’encadrent.
On va distinguer deux cas :
[**Le cas $r<d-2$.**]{} Etudions $-\Delta_x(e_i^r)$ ; c’est, avec les notations de la section précédente, $- DD_x.\Delta(e^r_j)$. La structure des lignes par blocs $Id_d$ de la partie supérieure de $\Delta$ fait que l’on a $$-\Delta_x(e_i^r)=(\theta^1_1,...,\theta^1_d;....;\theta^ {d-2}_1,...,\theta^ {d-2}_d)$$ où les $\theta^s_i$ sont tous nuls pour $s<r-1$, $$(\theta^{r-1}_1,...,\theta^{r-1}_d)=(-f_{1x}\omega^r_1,...,-f_{(r-1)x}\omega^r_{r+1},0,...,0,-f_{ix},0,...,0)$$ (ces deux conditions n’ayant de sens que pour $r>1$) et $$(\theta^{r}_1,...,\theta^{r}_d)=(-f_{1x}\omega^{r+1}_1,...,-f_{(r+2)x}\omega^{r+1}_{r+2},0,...,0).$$ On en déduit $$\Omega^{r,i}_{x, r,i}=0$$ si $i>r+2$ et $$\Omega^{r,r+2}_{x, r,r+2}=-f_{(r+2)x}\omega^{r+1}_{r+2}.$$ Exprimons maintenant ce qu’est $\omega^{r+1}_{r+2}$
Revenant un peu en arrière (avec $i=r+2$) nous remarquons que le fait que $e_{r+2}^r$ soit dans le noyau de $MM$ et que ses composantes $\beta^s_j$ soient nulles pour $j>r+2$ et $s=r$ ou $s=r+1$ nous donne la relation $$M_{f_1,...,f_{r+2}}(\omega^r_1,...,\omega^r_{r+1},1;\omega^{r+1}_1,...,\omega^{r+1}_{r+2})=0.$$ D’après la définition de début de cette section cela veut dire que $\omega^{r+1}_{r+2}$ [**est l’élément de trace**]{} $\gamma(f_1,\dots ,f_{r+2})$. 0n en déduit finalement
$$\Omega^{r,r+2}_{x, r,r+2}=-f_{(r+2)x}\gamma(f_1,\dots ,f_{r+2}).$$
[**Le cas $r=d-2$.**]{}
On a $e^{d-2}_d=(0;...;0;X)$ avec $X=(\omega^{d-2}_1,\cdots ,\omega^{d-2}_{d-1},1)$ et $P_{d-1}(X)=0 $ (notation de la section précédente). On a $$-\Delta_x (e^{d-2}_d)=(0;...;0;-D_x(X);-D_x.A_{2}(X).$$ Or, par définition, la dernière composante de $A_2(X)=-P^{-1}G_2(X)$ est l’élément de trace $\gamma(f_1,\dots ,f_{d})$. On en tire une formule analogue à celle des cas précédents
$$\Omega^{d-2,d}_{x, d-2,d}=-f_{dx}\gamma(f_1,\dots ,f_{d}).$$
On a le résultat analogue lorsque l’on échange $x$ et $y$ et on en tire facilement la proposition suivante.
[**Proposition.**]{} [*La trace de la courbure du tissu*]{} $W(f_1,\dots , f_d)$ [est la 2-forme]{} $$-\sum^d_{r=3} df_{r}\wedge d\gamma (f_1,\dots ,f_r).$$
Le cas des 3-tissus.
====================
On considère d’abord un 3-tissu $W(f,g,h)$ du plan. Trivialement sa courbure classique de Blaschke est aussi la trace de la courbure de la connexion associée. Suivant la procédure décrite dans la section précédente, on l’obtient en calculant d’abord l’élément de courbure $\gamma (f,g,h)$. Nous allons donner son expression précise dans le cas particulier où $h(x,y)=y$ et $f_x$ et $g_x$ ne s’annulent pas.
On adopte les notations $m_f=f_y/f_x$ et $m_g=g_y/g_x$. On prend $$X_1={1\over {(m_g-m_f)f_x}},\ \ X_2={1\over {(m_f-m_g)g_x}}$$ et on voit que $(X_1,X_2,1)$ engendre le noyau de $$P_{2;f,g,h}=\left | \matrix{f_{x} & g_{x} & 0 \cr f_{y} & g_{y} & 1 \cr}\right | .$$
Alors $\gamma (f,g,h)$ est la dernière composante du vecteur $$- P_{3;f,g,h}^{-1}G^{2}_{3;f,g,h}(X_1,X_2,1)$$
en prenant :
$$P_{3;f,g,h}=\left | \matrix{(f_{x})^2 & (g_{x})^2 & 0 \cr f_x f_{y} & g_x g_{y} & 0 \cr (f_{y})^2 & (g_{y})^2 & 1 \cr}\right | ,$$ $$G^{2}_{3;f,g,h}=\left | \matrix{f_{xx} & g_{xx} & 0 \cr f_{xy} & g_{xy} & 0 \cr f_{yy} & g_{yy} & 0 \cr }\right |.$$
Or le même argument que celui qui permet de calculer la matrice inverse d’une matrice de Vandermonde, montre que la dernière ligne de $ P_{3;f,g,h}^{-1}$ est $(m_fm_g,-(m_f+m_g),1)$ ; on en tire $$\gamma (f,g,h)=-(m_fm_g,-(m_f+m_g),1).G^{2}_{3;f,g,h}(X_1,X_2,1),$$ ce qui mène facilement à
$$\gamma (f,g,h)={1\over {m_f-m_g}}\{ m_fm_g(f_{xx}/f_x-g_{xx}/g_x)-$$$$(m_f+m_g)(f_{xy}/f_x-g_{xy}/g_x)+(f_{yy}/f_x-g_{yy}/g_x)\}.$$
On en tire une formule explicite pour la courburede Blaschke de notre tissu :$$-dy\wedge d\gamma (f,g,y).$$
La méthode de démonstration de la formule de la trace.
======================================================
Pour le tissu $W(f_1,\dots ,f_s)$ la somme des courbures des sous 3-tissus est $$SC(f_1,\dots ,f_s)=-\sum_{0<i<j<r\leq s}df_r\wedge d\gamma(f_i,f_j ,f_r).$$ On note $$Tr(f_1,\dots ,f_s)(=-\sum^s_{r=3}( df_{r}\wedge d\gamma (f_1,\dots ,f_r))$$ la trace de la courbure de la connexion $\nabla$ associée comme plus haut. La formule de la trace dit que l’on a $$SC(f_1,\dots ,f_s)=Tr(f_1,\dots ,f_s)$$ pour tout $s$ plus grand que 3.
La formule est triviale pour $s=3.$ Nous la démontrons par récurrence sur $s$ ; pour cela nous la supposons montrée à l’ordre $d-1$ et nous allons la prouver à l’ordre $d.$ Comme $SC(f_1,\dots ,f_d)$ et $Tr(f_1,\dots ,f_d)$ sont des quantités qui ne dépendent pas des coordonnées on peut choisir ces coordonnées pour avoir $f_d =y$ et la non-nullité des $f_{ix}$ pour $i$ variant de 1 à $d-1.$
La formule à l’ordre $d-1$ nous donne $$\sum_{r=3}^{d-1} (df_{r}\wedge d(\sum_{0<i<j<r}\gamma(f_i,f_j ,f_r))=\sum^{d-1}_{r=3}( df_{r}\wedge d\gamma (f_1,\dots ,f_r)).$$
Pour prouver la formule à l’ordre $d$ il suffit de prouver la relation $$\gamma(f_1,\dots ,f_d)=\sum_{0<i<j<d}\gamma(f_i,f_j ,f_d).$$ C’est ce que nous allons faire dans la suite de ce travail en utilisant l’hypothèse simplificatrice $f_d=y.$
Les quantités $\gamma(f_1,\dots ,f_d)$ et $\sum_{0<i<j<d}\gamma(f_i,f_j ,f_d)$ ont des expressions linéaires dans les dérivées secondes $f_{sxx},$ $f_{sxy}$ et $f_{syy}$ avec des coefficients qui ne dépendent que des dérivées premières. On va voir que ces coefficients sont les mêmes dans les deux expressions.
La somme des courbures des sous 3-tissus.
=========================================
La section 4 montre que l’on a la formule $$\sum_{0<i<j<d}\gamma(f_i,f_j ,y)=\sum_{0<i<j<d}{1\over {m_i-m_j}}\{ m_im_j(f_{ixx}/f_{ix}-f_{jxx}/f_{jx})-$$$$(m_i+m_j)(f_{ixy}/f_{ix}-f_{jxy}/f_{jx})+(f_{iyy}/f_{ix}-f_{jyy}/f_{jx})\}$$ avec la notation $m_k=f_{ky}/f_{kx}.$ On en déduit un développement $$\sum_{0<i<j<d}\gamma(f_i,f_j ,y)=\sum_{s<d}A_sf_{sxx}+B_sf_{sxy}+C_sf_{sxx}$$ avec $$A_s={1\over f_{sx}}\sum_{j\neq s}{m_sm_j\over{m_s-m_j}}$$ $$B_s={-1\over f_{sx}}\sum_{j\neq s}{m_s+m_j\over{m_s-m_j}}$$ $$C_s={1\over f_{sx}}\sum_{j\neq s}{1\over{m_s-m_j}}.$$ Dans les sections suivantes nous allons montrer que $\gamma(f_1,\dots ,f_d)$ a le même développement.
Calcul du noyau de $P_{d-1}.$
=============================
Il nous faut calculer l’élément de trace $\gamma(f_1,\dots ,f_d)$ (avec $f_d=y$). Si l’on revient sur sa définition, donnée dans la section 3, il nous faut d’abord calculer le vecteur $(X_1,\dots ,X_{d-1},1)$ qui engendre le noyau de $P_{d-1}$.
En tenant compte du fait que $f_d=y,$ le sytème $P_{d-1} (X_1,\dots ,X_{d-1},1)=0$ se récrit sous la forme $$\sum_{i=1}^{d-1}f_{ix}^{d-1-j}f_{iy}^{j-1}X_i=0$$ $$\sum_{i=1}^{d-1}f_{iy}^{d-2}X_i=-1$$ où $j$ varie de 1 à $d-1.$ Si l’on pose $Y_i=f_{ix}^{d-2}X_i$ et $m_i=f_{iy}/f_{ix}$ les $d-1$ premières équations donnent un système matriciel $VM((Y_1,\dots ,Y_{d-1}))=(0,\dots,0,-1)$ où $VM$ est une matrice de Vandermonde dont le coefficient sur la $s$-ième ligne et la $r$-ième colonne est $m_r^{s-1}.$ On en tire que $(Y_1, \dots ,Y_{d-1})$ est l’opposé de la transposée de la dernière colonne de $VM^{-1}.$
On a alors $$Y_i={-1\over \prod_{j\neq i}(m_i-m_j)}$$ et donc $$X_i={-1\over{f_{ix}^{d-2} \prod_{j\neq i}(m_i-m_j)}}$$ pour $i$ variant de 1 à $d-1.$
Calcul de la matrice $G^2_d$.
=============================
Dans la section 2 nous avions noté $G^d(f_1),\dots ,G^d(f_{d}),$ ses colonnes (avec $f_d=y$).
On remarque d’abord que, pour une fonction arbitraire $f$, le $i$-ème coefficient de $G^d(f)$ est de la forme $$G^d_i(f)=a_i^df_x^{d-i-2}f_y^{i-1}f_{xx}+b_i^df_x^{d-i-1}f_y^{i-2}f_{xy}+c_i^df_x^{d-i}f_y^{i-3}f_{yy},$$ avec la convention d’écriture que les puissances négatives des $f_x$ ou $f_y$ sont nulles ; les $a_i^d,$ $b_i^d$ et $c_i^d$ sont des nombres que nous allons déterminer.
Pour des raisons de symétrie par rapport aux deux dérivations $\partial /\partial x$ et $\partial /\partial y$, on a les relations suivantes : $$a_i^d=c^d_{d-i+1},\ \ \ b_i^d=b^d_{d-i+1}.$$
Nous utilisons la relation de récurrence $$G^{d+}(f)=f_x.G^{d-1}(f)+{\partial f^{d-2}_i\over\partial x}$$ que nous avions donnée en section 2. Elle nous donne $$G^d_i(f)=f_xG^{d-1}_i(f)+(d-1-i)f_x^{d-i-2}f_y^{i-1}f_{xx}+(i-1)f_x^{d-i-1}f_y^{i-2}f_{xy}$$ pour $i<d.$
On en tire les relations de récurrence : $$a_i^d=a_i^{d-1}+d-1-i$$ $$b_i^d=b_i^{d-1}+i-1$$ $$c_i^d=c_i^{k-1}$$ pour $i<d.$
On a aussi les relations évidentes : $$a_1^3=1,\ \ b_1^3=c_1^3=0$$ $$a_2^3=0,\ \ b_2^3=1,\ \ c_1^3=0$$ $$a_1^3= b_1^3=0,\ \ c_1^3=1.$$
Utilisant ces relations, les relations de symétrie et de récurrence ci-dessus on obtient :
$$a_i^d={(d-1-i)(d-i)\over 2}$$ $$b_i^d=(i-1)(d-i)$$ $$c_i^d={(i-2)(i-1))\over 2}$$ pour $ i $ compris entre 1 et $d.$
La dernière ligne de $P_d^{-1}$.
================================
Pour calculer l’élément de trace $\gamma(f_1,\dots , f_{d-1},y)$ nous aurons besoin d’un autre ingrédient : la dernière ligne $\alpha =(\alpha_1,\dots , \alpha_d)$ de $P_d^{-1}$. Cette ligne est caractérisée par le fait que le produit de cette ligne avec chacune des $r-1$ premières colonnes de $P_d$ est nul et son produit avec la dernière colonne est 1. On a donc les équations $$\sum_{j=1}^{d-1}\alpha_jf_{ix}^{d-j}f_{iy}^{j-1}=0$$ pour tout $i$ variant de 1 à $d-1$ et $$\alpha_d=1.$$ En divisant les deux membres des $d-1$ premières équations par $f_{ix}$ on peut les remplacer par $$\sum_{j=1}^{d-1}\alpha_jm_i^{j-1}=0.$$ Comme lorsque l’on calcule l’inverse d’une matrice de Vandermonde, on introduit le polynôme $P(t)=\sum_{j=1}^{d-1}\alpha_j t^{j-1}$ et les relations précédentes montrent que ce polynôme admet les racines $m_1,\dots ,m_{ d-1}$ et 1 comme coefficient du terme de plus haut degré. On en déduit $$\alpha =((-1)^{d-1}S_{d-1},(-1)^{d-2}S_{d-2}, \dots ,-S_1,1),$$ où les $S_i$ sont les polynômes symétriques en $m_ 1,m_2,\dots ,m_{d-1}.$
Le calcul de l’élément de trace $\gamma (f_1,\dots ,f_{r-1},y)$.
================================================================
Rappelons que, par définition, $\gamma (f_1,\dots ,f_{r-1},y)$ est le dernier coefficient de $-P_k^{-1}G^2_d(X_1,\dots ,X_{d-1},1)$ où les $X_i$ sont ceux de la section 7 ; donc c’est le produit scalaire usuel des deux vecteurs $\alpha$ (voir section 9) et $G^2_d(X_1,\dots ,X_{d-1},1)$. On écrit ce résultat sous la forme $$\gamma (f_1,\dots ,f_{r-1},y)=-\alpha . G^2_d(X_1,\dots ,X_{d-1},1).$$
On en déduit $$\gamma (f_1,\dots ,f_{r-1},y)=-\sum_{s=1}^{d-1}X_s(\alpha .G^d(f_s))$$ en remarquant que $G^d(y)$ est nulle.
On rappelle la formule de la section 8 : $$G^d_i(f_s)=a_i^df_{sx}^{d-i-2}f_{sy}^{i-1}f_{sxx}+b_i^df_{sx}^{d-i-1}f_{sy}^{i-2}f_{sxy}+c_i^df_{sx}^{d-i}f_{sy}^{i-3}f_{syy}$$ pour la $i$-ème composante de $G^d(f_s).$ On rappelle que, pour cette formule, les puissances négatives de dérivées de fonctions sont nulles par convention. On a donc la formule $$\alpha .G^d(f_s)=a^sf_{sxx}+b^sf_{sxy}+c^sf_{syy},$$ avec $$a^s=\sum_{i=1}^d\alpha_ia_i^df_{sx}^{d-i-2}f_{sy}^{i-1},$$ $$b^s=\sum_{i=1}^d\alpha_ib_i^df_{sx}^{d-i-1}f_{sy}^{i-2},$$ $$c^s=\sum_{i=1}^d\alpha_ic_i^df_{sx}^{d-i}f_{sy}^{i-3}.$$
Plus précisément, on a donc $$a^s=\sum_{i=1}^d (-1)^{d-i}{{(d-i-1)(d-i)}\over 2}S_{d-i}f_{sx}^{d-i-2}f_{sy}^{i-1},$$ $$b^s=\sum_{i=1}^d (-1)^{d-i}(i-1)(d-i)S_{d-i}f_{sx}^{d-i-1}f_{sy}^{i-2},$$ $$c^s=\sum_{i=1}^d (-1)^{d-i}{{(i-1)(i-2)}\over 2}S_{d-i}f_{sx}^{d-i}f_{sy}^{i-3}$$ ou encore $$a^s=f_{sx}^{d-3}\sum_{i=1}^d (-1)^{d-i}{{(d-i-1)(d-i)}\over 2}S_{d-i}m_s^{i-1},$$ $$b^s=f_{sx}^{d-3}\sum_{i=1}^d (-1)^{d-i}(i-1)(d-i)S_{d-i}m_s^{i-2},$$ $$c^s=f_{sx}^{d-3}\sum_{i=1}^d (-1)^{d-i}{{(i-1)(i-2)}\over 2}S_{d-i}m_s^{i-3}.$$
On a la formule $$S_{d-i} =m_sS^s_{d-i-1}+S^s_{d-i},$$ en notant $S^s_k$ le $k$-ième polynôme symétrique dans les variables $$m_1, \dots ,m_{s-1},m_{s+1},\dots ,m_{d-1}$$ avec, par convention, $S^s_j=0$ pour $j<0$ ou $j>d-2.$ C’est dire l’on oublie $m_s$ dans les $S^s_j.$ En portant cela dans équations précédentes on obtient $$a^s=f_{sx}^{d-3}\sum_{i=1}^d (-1)^{d-i}{{(d-i-1)(d-i)}\over 2}(S_{d-i}^sm_s^{i-1}+S^s_{d-i-1}m_s^i),$$ $$b^s=f_{sx}^{d-3}\sum_{i=1}^d (-1)^{d-i}(i-1)(d-i)(S_{d-i}^sm_s^{i-2}+S^s_{d-i-1}m_s^{i-1}),$$ $$c^s=f_{sx}^{d-3}\sum_{i=1}^d (-1)^{d-i}{{(i-1)(i-2)}\over 2}(S^s_{d-i}m_s^{i-3}+S^s_{d-i-1}m_s^{i-2}).$$ En réordonnant les termes en fonction des puissances de $m_s,$ on arrive à $$a^s=f_{sx}^{d-3}\sum_{i=1}^d (-1)^{d-i}(d-i-1)S_{d-i-1}^sm_s^i,$$ $$b^s=f_{sx}^{d-3}\sum_{i=1}^d (-1)^{d-i}(d-2i-2)S_{d-i-2}^sm_s^i,$$ $$c^s=f_{sx}^{d-3}\sum_{i=1}^d (-1)^{d-i-1}(i+1)S^s_{d-i-3}m_s^{i}.$$
Fin de la preuve de la formule de la trace.
===========================================
On a la relation $$\gamma (f_1,\dots ,f_{d-1},y)=-\sum_{s=1}^{d-1}X_s(a^sf_{sxx}+b^sf_{sxy}+c^sf_{syy}).$$ Donc nous aurons montré la formule de la trace si l’on prouve les relations $$X_sa^s=-A_s,\ \ X_sb^s=-B_s,\ \ X_sc^s=-C_s,$$ les $A_s$, $B_s$ et $C_s$ étant ceux définis dans la section 6. Il revient au même de prouver les relations $$a^s=-A_s/X_s,\ \ b^s=-B_s/X_s,\ \ c^s=-C_s/X_s.$$ Pour simplifier les notations on ne démontrera ces relations que dans le cas $s=d-1$ car exactement la même méthode fonctionne dans les autres cas. Toujours pour simplifier, nous écrirons $m$ à la place de $m_{d-1}.$ On a $$-A_{d-1}/X_{d-1}={1\over{ f_{(d-1)x}^{d-3}}}(\prod_{r=1}^{d-2}(m-m_r))\sum_{j=1}^{d-2}{{mm_j}\over{m-m_j}}.$$ On a la relation $$(\prod_{r=1}^{d-2}(m-m_r))\sum_{j=1}^{d-2}{{mm_j}\over{m-m_j}}=$$$$\sum_{j=1}^{d-2}(m-m_1)\cdots (m-m_{j-1})mm_j(m-m_{j+1})\cdots (m-m_{d-2})$$ Pour calculer cette quantité on introduit la fonction $$P(t)=\prod_{j=1}^{d-2}(t-m_j)$$ et il est facile de voir que l’on la relation $$(\prod_{r=1}^{d-2}(t-m_r))\sum_{j=1}^{d-2}{{tm_j}\over{t-m_j}}=t(tP'(t)-(d-2)P(t)).$$ Mais $P(t)$ est aussi le polynôme $$P(t)=\sum_{j=1}^{d-2}(-1)^{d-2-j}S^{d-1}_{d-2-j}t^j,$$ où $S^{d-1}_r$ est le polynôme symétrique de degré $r$ dans les variables $m_1,$ ...,$m_{d-2}.$ Cela mène à $$(\prod_{r=1}^{d-2}(t-m_r))\sum_{j=1}^{d-2}{{tm_j}\over{t-m_j}}=\sum_{i=1}^{d-2}(-1)^{d-i}(d-i-1)S_{d-i-1}^{d-1}t^i,$$ et, en posant $t=m$ et rajoutant le facteur ${1\over{ f_{(d-1)x}^{d-3}}},$ on obtient $$a^{d-1}=-A_{d-1}/X_{d-1},$$ et, par la même méthode, à $$a^{s}=-A_{s}/X_{s},$$ pour tout $s.$
Par ailleurs, on a $$-B_{d-1}/X_{d-1}={1\over{ f_{(d-1)x}^{d-3}}}(\prod_{r=1}^{d-2}(m-m_r))\sum_{j=1}^{d-2}{{m+m_j}\over{m-m_j}}.$$ Comme pour le calcul précédent on obtient facilement $$(\prod_{r=1}^{d-2}(t-m_r))\sum_{j=1}^{d-2}{{t+m_j}\over{t-m_j}}=2tP'(t)-(d-2)P(t)$$ qui, en revenant l’expression polynomiale de $P(t),$ mène à $$(\prod_{r=1}^{d-2}(t-m_r))\sum_{j=1}^{d-2}{{t+m_j}\over{t-m_j}}=\sum_{i=1}^{d-2}(-1)^{d-i-1}(d-2i+2)S_{d-i-2}^{d-1}t^i,$$ d’où il découle $$b^{d-1}=-B_{d-1}/X_{d-1}$$ et de la même manière $$b^{s}=-B_{s}/X_{s}$$ pour tout $s.$
Enfin, on a
$$-C_{d-1}/X_{d-1}={1\over{ f_{(d-1)x}^{d-3}}}(\prod_{r=1}^{d-2}(m-m_r))\sum_{j=1}^{d-2}{1\over{m-m_j}}.$$ Comme $$(\prod_{r=1}^{d-2}(t-m_r))\sum_{j=1}^{d-2}{1\over{t-m_j}}=P'(t),$$ on a $$(\prod_{r=1}^{d-2}(t-m_r))\sum_{j=1}^{d-2}{1\over{t-m_j}}=\sum_{j=1}^{d-2}(-1)^{d-2-j}jS^{d-1}_{d-2-j}t^{j-1},$$ et $$c^{d-1}=-C_{d-1}/X_{d-1}$$ puis, par la même méthode, $$c^{s}=-C_{s}/X_{s},$$ pour tout $s.$
Ceci achève notre démonstration de la formule de la trace pour les tissus planaires.
Exemple d’application de la formule de la trace.
================================================
Alain Hénaut a proposé la conjecture suivante. [*Soit $W$ un $d$-tissu du plan donné de manière implicite par le polynôme $$F=(y')^{d}+f(x,y)\ ;$$ alors la courbure de la connexion associée est nulle si et seulement si $f(x,y)$ est décomposable, c’est à dire de la forme $X(x)Y(y).$*]{}
Remarquons que dans l’ouvert où $F$ est non nul, les $d$ racines de $F$ sont toutes de la forme $m_i=R(x,y)\lambda_i$ où les $\lambda_i$ sont des constantes deux à deux différentes et la fonction $R(x,y)$ est indépendante de l’indice $i.$ De plus $f(x,y)$ est décomposable si et seulement si $R(x,y)$ l’est.
Alors la conjecture d’Hénaut est un corollaire du lemme suivant.
[**Lemme.**]{} [*Soit $W$ un $d$-tissu plan donné par ses pentes $m_i$ pour $i$ variant de 1 à $d$. On suppose que l’on a $$m_i=R(x,y)\lambda_i$$ où les $\lambda_i$ sont des constantes deux à deux différentes et la fonction $R(x,y)$ est indépendante de l’indice $i$ et non nulle. Alors la courbure de la connexion associée est nulle si et seulement si $R(x,y)$ est décomposable.*]{}
Nous allons donner une preuve de ce lemme en montrant d’abord le sens direct : si $R(x,y)$ est de la forme $X(x)Y(y)$ alors la courbure est nulle puis la réciproque.
[**1- On suppose**]{} $R(x,y)=A(x)B(y).$
Les feuilles des $d$ feuilletages sont les trajectoires des $d$ champs de vecteurs $$X_i={\partial\over\partial x}+A(x)B(y)\lambda_i{\partial\over\partial y}.$$ Ce sont aussi les trajectoires de $$Y_i=1/A(x){\partial\over\partial x}+\lambda_iB(y){\partial\over\partial y}.$$ Or des changements de la coordonnée $x,$ d’une part, et $y,$ d’autre part, permettent de rectifier les champs de vecteurs $1/A(x){\partial\over\partial x}$ et $B(y){\partial\over\partial y}.$ Ces changements nous ramènent au cas où tous les $Y_i$ sont à coefficients constants, donc au cas où les $d$ feuilletages sont tous formés de segments parallèles. Or on sait que ces tissus sont à courbure nulle.
[**2- La réciproque.**]{} On suppose que $W$ est de courbure nulle. Alors la trace de cette courbure est encore nulle, donc, par la [**formule de la trace**]{}, la somme des courbure des sous 3-tissus de $W$ est nulle. Etudions le sous 3-tissu de $W$ donné par les pentes $m_i,$ $m_j$ et $m_k.$ Un calcul élémentaire (qui peut être fait par Maple) montre que sa coubure est $${\partial^2\over {\partial x\partial y}}ln(R(x,y)).$$ Ainsi on voit que la somme des courbures des sous 3-tissus est nulle si et seulement si ${\partial^2\over {\partial x\partial y}}ln(R(x,y))=0,$ ou, ce qui est équivalent, que la fonction $R(x,y)$ est décomposable. Ceci achève la démonstration.
[dango 9999]{}
W. Blaschke, [*Uber die Tangenten einer ebenen Kurve funfter Klasse.* ]{} Abh. Math. Semin. Hamb. Univ. 9 (1933) 313-317.
V. Cavalier, D. Lehmann, [*Ordinary holomorphic webs of codimension one.* ]{} arXiv 0703596v2 \[mathsDS\], 2007, et Ann. Sc. Norm. Super. Pisa, cl. Sci (5), vol XI (2012), 197-214. . J. P. Dufour, D. Lehmann, [*Calcul explicite de la courbure des tissus calibrés ordinaires* ]{} arXiv 1408.3909v1 \[mathsDG\],18/08/2014.
A. Hénaut, [*Planar web geometry through abelian relations and connections*]{} Annals of Math. 159 (2004) 425-445.
N. Mihaileanu. [*Sur les tissus plans de première espèce.*]{} Bull. Math. Soc. Roum.Sci. 43 (1941), 23-26.
L. Pirio, [*Equations Fonctionnelies Abéliennes et Géométrie des tissus*]{} Thèse de doctorat de l’Université Paris VI, 2004.
A. Pantazi. [*Sur la détermination du rang d’un tissu plan.*]{} C.R. Acad. Sc. Roumanie 4 (1940), 108-111.
J.V. Pereira, L. Pirio,[*An Invitation to Web Geometry*]{} Series IMPA Monographs Vol.2, Springer (2015).
J.V. Pereira, [*Algebraization of codimension one webs*]{} Séminaire Bourbaki, 59ème année, 2006-2007, $n^0$974 (mars 2007).
O. Ripoll, [*Géométrie des tissus du plan et équations différentielles*]{} Thèse de doctorat de l’Université de Bordeaux 1, 2005.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'New sensitive CO(2-1) observations of the 30 Doradus region in the Large Magellanic Cloud are presented. We identify a chain of three newly discovered molecular clouds we name KN1, KN2 and KN3 lying within 2–14pc in projection from the young massive cluster R136 in 30 Doradus. Excited H$_2$2.12$\mu$m emission is spatially coincident with the molecular clouds, but ionized Br$\gamma$ emission is not. We interpret these observations as the tails of pillar-like structures whose ionized heads are pointing towards R136. Based on infrared photometry, we identify a new generation of stars forming within this structure.'
author:
- 'Venu M. Kalari'
- 'M[ó]{}nica Rubio'
- 'Bruce G. Elmegreen'
- 'Viviana V. Guzm[á]{}n'
- 'Cinthya N. Herrera'
- Hans Zinnecker
title: |
Pillars of creation amongst destruction:\
Star formation in molecular clouds near R136 in 30 Doradus
---
Introduction {#sec:intro}
============
30 Doradus is a giant H[II]{} region in the Large Magellanic Cloud (LMC). The LMC is a local group dwarf galaxy that lies at a distance of 50kpc (Pietrzy[ń]{}ski et al. 2013), and has a mean stellar metallicity ($Z$) half of the Sun (Rolleston et al. 2002). 30 Doradus hosts the young massive cluster (YMC) R136. R136 is a $\sim$1.5-3Myr YMC that encloses a total cluster mass in excess of 10$^5$$M_{\odot}$ within 10pc (Selman & Melnick 2013). The cluster contains roughly 200 massive stars ($>$8$M_{\odot}$) within a central region less than 6pc, whose radiation and mechanical feedback profoundly impact the surrounding medium (Schneider et al. 2017, subm.). R136 is the most massive YMC in our local neighbourhood that can be adequately resolved spatially (at 50kpc, the nominal distance to R136, 1$\arcsec$$\approx$0.25pc) enabling us observe individual objects at the star and molecular clump scale. This makes R136 an ideal laboratory to examine how feedback from massive stars affects further star formation (e.g. Dale et al. 2012).
The mechanical and radiation output from R136 has created a central cavity by sweeping the surrounding molecular clouds (labelled as clouds 6 and 10 in Fig.1) that extend up to 100pc along the northeast-southwest axis (Pellegrini et al. 2010). We adopt the cloud nomenclature of Johansson et al. (1998). Brightly illuminated arcs delineate the interfaces between the cold gas and the ionizing radiation, where subsequent generations of stars are thought to have been triggered (Walborn et al. 2002). Studies at optical (De Marchi et al. 2011; Kalari et al. 2014), near-infrared (nIR; Rubio et al. 1998; Brandner et al. 2001), mid-infrared (mIR; Whitney et al. 2008; Gruendl & Chu 2009; Walborn et al. 2013), far-infrared (fIR; Seale et al. 2014) and sub-millimeter (Johansson et al. 1998; Indebetouw et al. 2013) wavelengths have identified evidence for active star formation throughout the 30 Doradus nebula, consistent with the idea of multiple star formation episodes.
We focus on the stapler nebula that lies 2-14pc away from R136 (see Fig.1) in projection. The stapler nebula is the H[II]{} region including and surrounding the stapler shaped dark cloud that is seen in silhouette in the optical near R136. The nebula spans an area of 1.1$'\times$0.35$'$ centred on $\alpha$=05$^h$38$^m$40$^s$, $\delta\,=\,-$69$^{\rm \circ}$05$'$36$''$ and is elongated with a position angle of 35$^{\rm \circ}$. A candidate young stellar object (YSO) has been reported at the edge of the elongated dark cloud by Walborn et al. (2013; marked as S5 in that paper). The YSO is close to, but not coincident with a region of high density ($n>$10$^{6}$cm$^{-3}$) reported by Rubio et al. (2009) using CS line observations. Known infrared excess objects, some of which are thought to be disc/envelope bearing young stellar objects (YSOs) are dotted towards the edge of the dark cloud according to Rubio et al. (1998; their Figure 3). The literature evidence for dense molecular gas and YSOs in the stapler nebula lying near R136 indicates that star formation may be ongoing, which deserves further study. In this paper we discuss the properties of molecular clouds in the stapler nebula, and examine whether new stars are being formed in these clouds.
This paper is organised as follows. In Section 2 we describe the data used in this study. The results from the analysis of CO(2-1) line observations are presented in Section 3. We discuss the results obtained from nIR emission line images of the stapler nebula in Section 4. Based on archival infrared photometry, we identify YSOs within the stapler nebula in Section 5. The picture obtained from our results is described in Section 6. In Section 7, a brief summary of our paper is presented along with future work arising from our results.
{width="49.50000%"} {width="49.50000%"} {width="49.50000%"} {width="49.50000%"}
\[fig:COa\]
Data
====
[CO(2-1)]{} observations
------------------------
We conducted a deep CO(2-1) survey centred on the 30 Doradus Nebula using the Swedish-ESO submillimeter telescope (SEST) across March 1997- January 2001. SEST was a 15m radio telescope located at La Silla, Chile. The angular resolution at the CO(2-1) frequency of 230GHz is 23$\arcsec$, which corresponds to a projected size of 5.6pc at the distance to the LMC.
Observations were conducted in position switching mode using a reference point free of CO(2-1) (at $\alpha$=05$^{h}$37$^{m}$54$^s$, $\delta$=$-69^{\rm \circ}$04$'24''$) for sky subtraction. The backend narrow-band high-resolution spectrometer (HRS) was used with a bandwidth of 80MHz, and a frequency resolution of 41.7kHz, which translates to a velocity resolution of 0.054kms$^{-1}$ at the frequency of CO(2-1). The data were reduced using the GILDAS software [^1], with linear or third-order polynomial for baseline fitting. The resultant spectra were smoothed to a velocity resolution of 0.25kms$^{-1}$. The rms noise achieved in a single channel is 0.07K after 240s of integration. We mapped the 30 Doradus region with 10$\arcsec$ spacings and detected CO(2-1) emission across 30 Doradus, including the stapler nebula where CO(1-0) emission had not been previously detected (Pineda et al. 2009; Johansson et al. 1998). In Fig.1, the CO(2-1) contours are overlaid on a [*Hubble space telescope*]{} (HST) three colour optical $BVI$H$\alpha$ image of 30 Doradus, where the position of CO(2-1) emission with respect to the R136 cluster, and its ionized surroundings can be visualized. Our observations represent a five-fold increase in sensitivity at twice the spatial resolution of previous CO(1-0) observations across the 30 Doradus nebula (see Pineda et al. 2009). Higher angular resolution observations of the 30 Doradus nebula are presented in Indebetouw et al. (2013), Anderson et al. (2014) and Nayak et al. (2016), but those data do not cover the region studied here.
Near-infrared emission line imaging
-----------------------------------
We obtained nIR imaging of the 30 Doradus region in H$_2$ 2.12$\mu$m narrowband filter, and the $K$s broadband filter using the ISAAC (Infrared spectrometer and Array Camera) imager mounted on 8m Melipal (UT3) telescope of the Very Large Telescope (VLT) situated in Paranal, Chile (program ID 078.C-0487A). Toward the R136 region, the mosaic covered a field of 5$\arcmin$$\times$5$\arcmin$ area. The average seeing measured from the images is $\sim$0.8$''$–1.1$''$. Observations were taken in the ABBA sequence with the sky image 30$'$ from the stapler nebula at $\alpha$=05$^h$39$^m$01$^s$, $\delta\,=\,-$69$^{\rm \circ}$42$'$36$''$ in a region free of nebulosity within the ISAAC field of view, ensuring adequate sky subtraction. The total integration time on source was 1 hour for the narrowband filter, and 100s for the $K$s broadband filter. Data were reduced using the ISAAC pipelines, with flux calibration carried out using Persson (1998) standards. Astrometric calibration was refined using 2MASS. Bright stars are saturated in the emission line images and have a non-linear CCD response, meaning they cannot be completely subtracted. This leads to circular bright residuals (and in some cases vertical bleeding) in the emission line images. We excluded these regions from further analysis by masking them.
We used the Br$\gamma$ 2.165$\mu$m narrowband flux calibrated image from Yeh et al. (2015). The image was obtained using the NOAO Extremely Wide Field Infrared Imager (NEWFIRM) mounted on the 4m Victor Blanco telescope located at the Cerro Tololo Inter-American Observatory, Chile. The final Gaussian convolved image resolution for the Br$\gamma$ narrowband image is 1$\arcsec$, and is comparable to the VLT H$_2$2.12$\mu$m narrowband image.
Archival photometry
-------------------
mIR photometry of point sources in the stapler nebula were estimated by Gruendl & Chu (2009) from images at 3.6, 4.5, 5.8 and 8.0$\mu$m taken using the [*Spitzer*]{} space telescope IRAC (Infrared Array Camera) as part of the [*Spitzer*]{} legacy program SAGE (Spitzer Survey of the Large Magellanic Cloud: Surveying the Agents of a Galaxy’s Evolution; Meixner et al. 2006). The full width half maximum (FWHM) of the images is 1.6$\arcsec$, 1.7$\arcsec$, 1.7$\arcsec$ and 2$\arcsec$ respectively. Alternative photometry of point sources from the same images are also presented by Whitney et al. (2008), but Whitney et al. (2008) miss a significant fraction of point sources in the 30 Doradus region, as their study is motivated towards detecting reliable sources throughout the LMC via pipeline analysis. Gruendl & Chu (2009) detect objects in the dense and nebulous surroundings using detailed aperture photometry (see Section 6.3 of Gruendl & Chu (2009) for a comparison). Photometry in the fIR at 100 and 160$\mu$m, and 250 and 350$\mu$m of point-like and extended sources in the stapler nebula is given in Seale et al (2014), using images taken by the [*Herschel*]{} space telescope PACS (Photoconductor Array Camera and Spectrometer), and SPIRE (Spectral and Photometric Imaging Receiver) instruments respectively as part of the [*Hershel*]{} large program HERITAGE (HERschel Inventory of The Agents of Galaxy Evolution; Meixner et al. 2013). The FWHM for these images are 8$\arcsec$, 12$\arcsec$, 18$\arcsec$ and 25$\arcsec$ respectively. We utilize the photometry from Gruendl & Chu (2009) and Seale et al. (2014) in this study.
Molecular clouds
================
Figure \[fig:CO\] shows the CO(2-1) integrated line emission over the velocity interval 235–270 kms$^{-1}$ as contours, superimposed on the HST composite $BVI$H$\alpha$ image. Strong CO(2-1) emission from Cloud 10 (northeast region of the map) and Cloud 6 (southwest region of the map) previously reported by Johansson et al. (1998) is seen along the northeast-southwest axis.
We observe previously undetected CO(2-1) emission originating from the region located between Clouds 6 and 10, close to R136 (see Fig.\[fig:COa\]). The emission extends along the southeast-northwest direction in projection. The distance of the emission from R136 in projection is between 2pc (emission is located to the north of R136) to 14pc away (in the northwest direction from R136). This CO emission is approximately five times weaker than the CO emission observed in Clouds 6 and 10. This emission is coincident spatially to the stapler nebula in the optical image of Fig.\[fig:COa\]a. The emission is resolved in CO(2-1) velocity as a chain of small and weak clouds (Fig.\[fig:COa\]a,b). We define the stapler region by the extent of the CO(2-1) emission, which goes beyond the visible stapler shaped dark cloud in the optical. This boundary is marked in Fig.1 with a dashed rectangle. We named the CO clouds Knots (KN), as they form a chain separated in velocity, as demonstrated by the position velocity slice across the stapler nebula, and the CO(2-1) spectra of each individual cloud shown in Fig.\[fig:spectra\]. By analysing the radial velocities and spatial distribution we found that the KN clouds are composed of three clouds we name KN1, KN2, and KN3 in order of decreasing Right Ascension (labelled in Fig.\[fig:COa\]a).
![[*Top*]{}: The stapler nebula blown up from the HST mosaic in Fig.1, with the stapler region is outlined with the dashed white box. The outermost CO(2-1) contours integrated over the 235–240kms$^{-1}$, and 245–250kms$^{-1}$ are shown in magenta and red respectively, with the second outermost contour of the 240–245kms$^{-1}$ also shown. The solid white line marks the position of the slice shown in the middle panel. [*Middle*]{}: Position velocity slice of the CO(2-1) cube in linear scale along the direction of the slit given by the solid white line in the top panel. The slit cuts along the centre of the stapler nebula. [*Bottom*]{}: CO(2-1) spectra of each cloud extracted from the region bounded by the contours shown in top panel. The dashed lines for each cloud is it’s $V_{\rm{lsr}}$ given in Table 1. []{data-label="fig:spectra"}](30dorfigcocloseupcss.jpg "fig:"){width="49.50000%"} ![[*Top*]{}: The stapler nebula blown up from the HST mosaic in Fig.1, with the stapler region is outlined with the dashed white box. The outermost CO(2-1) contours integrated over the 235–240kms$^{-1}$, and 245–250kms$^{-1}$ are shown in magenta and red respectively, with the second outermost contour of the 240–245kms$^{-1}$ also shown. The solid white line marks the position of the slice shown in the middle panel. [*Middle*]{}: Position velocity slice of the CO(2-1) cube in linear scale along the direction of the slit given by the solid white line in the top panel. The slit cuts along the centre of the stapler nebula. [*Bottom*]{}: CO(2-1) spectra of each cloud extracted from the region bounded by the contours shown in top panel. The dashed lines for each cloud is it’s $V_{\rm{lsr}}$ given in Table 1. []{data-label="fig:spectra"}](pv2.jpg "fig:"){width="49.50000%"} ![[*Top*]{}: The stapler nebula blown up from the HST mosaic in Fig.1, with the stapler region is outlined with the dashed white box. The outermost CO(2-1) contours integrated over the 235–240kms$^{-1}$, and 245–250kms$^{-1}$ are shown in magenta and red respectively, with the second outermost contour of the 240–245kms$^{-1}$ also shown. The solid white line marks the position of the slice shown in the middle panel. [*Middle*]{}: Position velocity slice of the CO(2-1) cube in linear scale along the direction of the slit given by the solid white line in the top panel. The slit cuts along the centre of the stapler nebula. [*Bottom*]{}: CO(2-1) spectra of each cloud extracted from the region bounded by the contours shown in top panel. The dashed lines for each cloud is it’s $V_{\rm{lsr}}$ given in Table 1. []{data-label="fig:spectra"}](spectra.pdf "fig:"){width="49.50000%"}
Physical properties
-------------------
After identifying each cloud, we computed the central velocity ($V_{\rm lsr}$) in the local standard of rest frame, and velocity width ($\sigma_{v}$) by fitting a Gaussian profile to the total cloud spectrum. The major and minor axis sizes of the profiles, in conjunction with the rms size of the beam were used to compute the deconvolved radius ($r$). Given the uncertainties on the Gaussian fit of the CO spectra found for each cloud are around 30%, we estimate the uncertainties on $r$ to be 15%.
### CO luminosity and mass
The CO cloud luminosity is computed as: $$L_{\rm CO}\, [K{\rm kms}^{-1}{\rm pc}^{-2}] = \rm{D}^2 \int_{\Omega} \int_v T_{\rm mb}(\nu) \, d\nu \,d\Omega$$ where D is the distance to the source in pc (adopted as 50kpc), $T_{\rm mb}$ the main beam temperature which is the antenna temperature corrected for the efficiency of the antenna ($T_{\rm mb} = T_{\textrm{A}}/\eta$), and $\Omega$ is the solid angle of the subtended by the source.
The H$_2$ mass of the clouds can be calculated from the observed CO(1-0) luminosity assuming a linear conversion between the velocity integrated CO emission ($I_{\textrm{CO}}$) and the H$_2$ column density ($N_{{\rm H}_2}$); $$N_{\rm H{_2}} = X_{\textrm{CO}}\, [cm^{-2}(K{\rm kms^{-1}})^{-1}]\,\, I_{\textrm{CO}}\,[K{\rm kms^{-1}}],$$ where $X_{\textrm{CO}}$ is the CO-to-H$_2$ conversion factor (Bolatto et al. 2013; Roman-Duval et al. 2014). The total mass of H$_2$ ($M_{{\rm H}_2}$) is, $$M_{\textrm{H}_2} \,[M_\odot] = \alpha_{\textrm{CO}} {\rm D}^2\, [{\rm Mpc}] S_{\textrm{CO}},$$ where $$\alpha_{\textrm{CO}}\, [M_\odot {\rm Mpc}^{-2} ({\it Jy}{\rm kms}^{-1})^{-1}] = X_{\textrm{CO}} \frac{m_{\textrm{H}_2} c^2}{2 k \nu^2},$$ and the flux density $S_{\textrm{CO}}$ is, $$S_{\textrm{CO}}\,\, [{\it Jy}{\rm kms}^{-1}] = \int S_{\nu} \, dv.$$ The molecular gas mass is multiplied by 1.36 to include the Helium contribution. This method is calibrated for the $J$=1$\rightarrow$0 transition. We use a ratio between the CO$J$=2$\rightarrow$1 and $J$=1$\rightarrow$0 lines of 0.87 for 30 Doradus Cloud 10 (the North Eastern cloud; see Fig.1) found by Johannson et al. (1998) to estimate the CO(1-0) luminosity.
The conversion factor depends on both metallicity and the ambient radiation field intensity (Maloney 1988). As a consequence of strong radiation fields and poor self-shielding in low metallicity environments, the CO molecule is photo-dissociated as it does not self-shield like the H$_2$ molecule. Therefore, there is less CO compared to the H$_2$ abundance in the Magellanic Clouds than in the Galaxy. This translates into higher values of $X_{\textrm{CO}}$ in the LMC compared to the Galaxy. In general the conversion factor increases with higher radiation fields and decreases with higher metallicities. We adopt the median conversion factor in the LMC compiled from the literature by Bolatto et al. (2013) of $$X_{\textrm{CO}} = 8.8 \pm 0.3 \times 10^{20} \,[\textrm{cm}^{-2} \textrm{(K kms}^{-1})^{-1}].$$ The adopted $X_{\textrm{CO}}$ factor is 3.8 times larger than the canonical Galactic $X_{\textrm{CO}}$ of Bolatto et al. (2013). Our adopted value is similar to that found by Herrera et al. (2013) when comparing molecular and dust mass estimates in the LMC N11 region; but higher than the value of $6 \times 10^{20} \,\textrm{cm}^{-2} \textrm{(K kms}^{-1})^{-1}$ reported by Roman-Duval et al. (2014). The resulting cloud masses if we adopted the Roman-Duval et al. (2014) $X_{\textrm{CO}}$ would be reduced by $\sim20$%.
### Virial mass
The virial mass ($M_{\rm{vir}}$) was computed assuming that each cloud is spherical, is in virial equilibrium and has a density ($\rho$) profile of the form $\rho \propto
r^{-1}$. The virial mass is given by $$M_{\textrm{vir}} \,[M_\odot] = 190 \sigma_{v}^2\,[{\rm kms}^{-1}]\, r\, [{\rm pc}]
\label{eq:virial_mass2}$$ according to MacLaren et al. (1988). The results of our analysis for each cloud are given in Table 1.
From Table \[tab:clouds\_properties\] we see that the virial masses are a factor 3-6 larger than the masses derived from the integrated CO emission for the resolved molecular clouds in 30 Doradus. Therefore, the conversion factor between the H$_2$ column density and the CO intensity is, on average, 4.5 times the Galactic value. This translates into a conversion factor $X_{\textrm{CO}} = 1.0 \pm 0.4 \times 10^{21}$ cm$^{-2}$ (K km s$^{-1}$)$^{-1}$. [@israel97] estimated the H$_2$ column densities towards CO clouds in the LMC and SMC from far-infrared surface brightness, and derived, in units of $10^{21}$ cm$^{-2}$ (K km s$^{-1}$)$^{-1}$, $X_{\textrm{CO}} = 12 \pm 2$ and $X_{\textrm{CO}} = 1.3 \pm 0.2$ for the SMC and LMC, respectively. [@israel03] found, in units of $10^{20}$ cm$^{-2}$ (K km s$^{-1}$)$^{-1}$, $X_{\textrm{CO}} = 4.8 \pm 1.0$ and $X_{\textrm{CO}} =
4.3 \pm 0.6$ for the SMC and LMC, respectively. [@garay02] derived $X_{\textrm{CO}} = 6.4 \times 10^{20}$ cm$^{-2}$ (K km s$^{-1}$)$^{-1}$ for the Complex-37 in the LMC. [@johansson98] also derived conversion factors close to the canonical value for the Galaxy to a factor of a few higher, although he found these values using CO($1-0$) observations whose sensitivity is lower than our CO($2-1$) data. There is a small difference between the conversion factor for 30 Doradus found in this work and the previously mentioned values found by other studies in the Magellanic clouds. However, one would expect to find a different conversion factor in the clouds of 30 Doradus than in the rest of the clouds of the LMC due to the extreme conditions in the environment, specially the strong radiation fields that photo-dissociate the CO molecule leaving large H$_2$ envelopes untraced by CO.
Analysis
--------
### Larson’s Laws
Molecular clouds in virial equilibrium follow the empirical power law relation $\sigma_{v} \propto r^{\alpha}$ (Larson 1981). $\alpha$ is generally agreed to be between 0.4–0.5 based on numerous molecular cloud surveys of the Milky Way, and external normal and dwarf Galaxies (Bolatto et al. 2008; Heyer et al. 2009). The observed value of $\alpha$ is oft explained by turbulence (McKee & Ostriker 2007; Lombardi et al. 2010). The velocity dispersion is considered to be a measure of the internal dynamics within the clouds, because the observed line profiles, averaged over a cloud, have Gaussian shapes and the line widths are broader than the thermal line widths. It is a reasonable assumption to make that the line profiles are produced by turbulent motions of the gas inside the clouds (Solomon et al. 1987).
Figure\[fig:slw\]a displays the position of the KN clouds in the $\sigma_{v}$–$r$ diagram. Also shown are results from Heyer et al. (2009) summarising the canonical relation found for Galactic clouds of $\sigma_{v} = 0.72\,r^{0.5}$; and clouds in the 30 Doradus region (from Pineda et al. 2009; having a resolution of 43$\arcsec$; and from Nayak et al. (2016) having a resolution of 2$\arcsec$), and in the LMC (excluding 30 Doradus) from Wong et al. (2011) whose study had a spatial resolution of 45$\arcsec$. Note that the Pineda et al. (2009) and Nayak et al. (2016) clouds do not cover the central region of 30 Doradus near R136, and there are no spatial overlaps between the KN clouds and the clouds they identify. The molecular clouds associated with the stapler nebula lie above the canonical $\sigma_{v}$–$r$ relation for Galactic clouds. Interestingly, the other detected clouds in 30Doradus from Pineda et al. (2009) and Nayak et al. (2016) also lie above the canonical relation. Although the departure from the relation is not at the same scale as the KN clouds, this might still indicate that the observed $\sigma_{v}$–$r$ relation might be a function of distance from R136, and a more global property of 30Doradus.
The position of the KN molecular clouds in the $\sigma_{v}$–$r$ diagram implies either of two scenarios; the clouds are collapsing or expanding, or the observed line widths are the manifestation of external pressures that keep the clouds in equilibrium. Since collapse velocities are generally only $\sim40$% larger than equilibrium velocity dispersions for a self-gravitating cloud, the observed large linewidths for the sizes of the KN clouds are probably not the result of collapse (Ballesteros-Paredes et al. 2011). Neither are they likely to result from expansion because there is no obvious shell or hole structure that usually accompanies expansion.
----------- ------------------- -------------------------------- --------------- ------------------ --------------------- --------------- --------------------- ---------------------
Name R.A. Dec. $V_{\rm lsr}$ $\sigma_{\rm v}$ $L_{\rm CO}$ $r$ $M_{\rm H_2}$ $M_{\rm vir}$
(J2000) (J2000) (kms$^{-1}$) (kms$^{-1}$) (Kkms$^{-1}$pc$^2$) (pc) (10$^3\,M_{\odot}$) (10$^3\,M_{\odot}$)
30Dor-KN1 5$^h$38$^m$45$^s$ $-$69$^{\rm \circ}$06$'$00$''$ 237.2$\pm$0.1 4.0$\pm$0.7 154.2$\pm$21.3 3.93$\pm$0.58 2.7$\pm$0.2 10.9$\pm$3.3
30Dor-KN2 5$^h$38$^m$40$^s$ $-$69$^{\rm \circ}$05$'$30$''$ 244.4$\pm$0.2 4.4$\pm$0.1 243.6$\pm$3.7 3.35$\pm$0.5 4.2$\pm$0.3 11.25$\pm$1.8
30Dor-KN3 5$^h$38$^m$36$^s$ $-$69$^{\rm \circ}$05$'$30$''$ 250.0$\pm$0.2 5.0$\pm$0.3 320.8$\pm$44.3 2.96$\pm$0.44 5.3$\pm$0.4 12.83$\pm$2.5
----------- ------------------- -------------------------------- --------------- ------------------ --------------------- --------------- --------------------- ---------------------
\
We examine whether the observed large $\sigma_{v}$ are the manifestation of external pressures necessary to keep the clouds in equilibrium. In Fig.\[fig:slw\]b, we plot the mass surface density ($\Sigma_{\rm H_2}$) against the ${\sigma_v}^2/r$ value of the KN clouds, along with those in the Milky Way from Heyer et al. (2009); in 30 Dor from Pineda et al. (2009) and Nayak et al. (2016); and in the LMC from Wong et al. (2011). Isolated virial clouds confined by self-gravity follow a linear relation in the ${\sigma_v}^2/r$ vs. $\Sigma_{\rm H_2}$ plot (for e.g. see Heyer et al. 2009). The values of clouds from the literature fall along this expectation. However, the KN clouds alone depart from the expected relation, and are likely confined by external pressure and not in virial equilibrium. Following the simplifying assumptions of Field et al. (2011), we plot isobars of external pressure (in terms of $P/k_{\rm B}$) according to their prescription. The lines reflect the external pressure necessary to confine clouds for a given ${\sigma_v}^2/r$ assuming clouds with a centrally concentrated internal density structure approximated by hydrostatic equilibrium. From Fig.\[fig:slw\]b, we see that external pressures of $\sim\,10^6$cm$^{-3}$K are necessary to keep the KN clouds confined. These values are in general agreement with Chevance et al. (2016), who report that the stapler nebula is located in a region with gas pressure $\sim$ 0.85-1.2$\times 10^6$cm$^{-3}$K, with the peak found in KN2 (see Fig. 15 in Chevance et al. 2016).
### Variation in properties as a function of distance from R136
The $V_{\rm lsr}$, and $M_{{\rm H}_2}$ of each cloud are plotted as a function of projected distance from R136 in Fig.\[fig:distanced\]. The velocity of the clouds increases as a function of projected distance from R136. The velocity of the KN3 cloud loosely matches the radial velocities of the stars within R136[^2]. The KN1 cloud is closest in projection to R136 and is blue shifted with respect to the mean velocity of stars in the cluster. This suggests the KN1 cloud, (and likely KN2 and KN3 clouds as suggested by the lack of background stars) lie slightly in front of the cluster. The clouds appear to be moving away from the cluster as function of projected distance from it (Fig.\[fig:distanced\]a).
The $M_{{\rm H}_2}$ of each cloud also increases as the projected distance from R136 increases (Fig.\[fig:distanced\]b). The KN1 cloud mass is approximately 2 times lower than the KN3 cloud. Under the assumption that the molecular clouds detected in CO(2-1) were initially all of similar densities, and photoionization from R136 alone is evaporating the molecular cloud, then KN1 must be closer to R136 (considering it as the only source of external photoionization) because the ionizing flux decreases to the inverse square with distance.
Line of sight distances
-----------------------
Our main results concerning the detection of cold molecular gas near R136 suffers from possible projection effects. Although the cold molecular gas detected in the CO(2-1) observations lies within 2-14pc in projection of the R136 cluster, the actual distance may likely be further in the line of sight direction allowing for the clouds to possibly survive photoionisation. Chevance et al. (2016) analyse the physical distance of the CO gas in 30 Doradus to the stars, by comparing the incident radiation field on the gas modelled against fIR observations in fine structure lines of the emitted radiation field measured from the known massive star population. By comparing the luminosity of the photodissociation region (which forms the interface between the photoionizing radiation from the stars and the gas) against their predictions, they are able to constrain the line of sight distance of the photodissociation regions from R136 with uncertainties of 4pc. Based on their results (Figure 20 in Chevance et al. 2016), the stapler nebula lies less than 20pc away in the line of sight direction from R136, which itself lies at the centre of a sphere of about 6pc in radius. This distance agrees well with the line of sight distance measured from line ratios of ionized lines in optical spectra by Pellegrini et al. (2010). From their Fig.12, we find that the distance of the KN clouds is less than 20pc away in our line of sight from R136.
We also consider that if the CO gas is close to R136, and coincident with the dust, there is likely to be a gradient (reflecting the gradient in the projected $V_{\rm lsr}$) in the dust temperature, and the total fIR luminosity arising from the photodissociation region. Such a gradient is visible in both the dust temperature maps (Guzman 2010), and also in the total fIR luminosity which peaks at KN2, and decreases towards KN3 (see Fig.1 of Chevance of et al. 2016). Therefore, although our observations are unable to resolve the line of sight distances to the KN clouds from R136, based on corroboration from multiple independent sources in the literature, we find that the line of sight distance to the KN clouds from R136 is $\lesssim$20pc. The molecular clouds in the stapler nebula lie between 2-14pc in projected distance, and $\lesssim$20pc in the line of sight distance from R136.
Near-infrared emission line imaging
===================================
The H$_2$ 2.12$\mu$m emission line image is shown in Fig.\[fig:COa\]c, with the CO(2-1) contours overlaid. Strong H$_2$ emission is spatially coincident with the CO(2-1) emission of the molecular clouds, and shares similar morphology. The H$_2$ emission is clumpy, with numerous knots and a reticulated pattern. In contrast, detected ionized gas (Br$\gamma$) at the position of the CO(2-1) molecular clouds is weak and diffuse (see Fig.\[fig:COa\]d). This diffuse Br$\gamma$ emission is associated with filaments and arc-like structures of ionized gas vivid in H$\alpha$ (brown in Fig.\[fig:COa\]a). This convinces us that the strong H$_2$ nIR emission is the warmer component of the cold molecular gas traced by the CO detections, whereas the diffuse Br$\gamma$ emission lies slightly beyond the ionized surface of the molecular cloud although no clear demarcation is noted. The Br$\gamma$ morphology is not spatially coincident with the H$_2$ and CO(2-1) emission.
The H$_2$/Br$\gamma$ ratio can be used to disentangle shock/collisionally excited H$_2$ from fluorescence excitation. This is because the shocks and collisional excitation affect primarily the H$_2$ gas, leaving the ratio of H$_2$/Br$\gamma$ above unity, whereas fluorescence acts on both the molecular and ionised gas leading to a ratio below unity. Using the absolute flux ratio, we find that the H$_2$/Br$\gamma$ ratio never exceeds 0.5 at an angular resolution of 1$\arcsec$ in the stapler nebula, agreeing with the findings of Yeh et al. (2015). This indicates that the excited nIR H$_2$ is primarily excited by the ultraviolet (UV) radiation from R136 acting on the surfaces of the molecular gas, with the filamentary Br$\gamma$ arising from the same source. The KN clouds therefore must lie in front of us given the morphology of the clumped H$_2$ emission, and lack of background stars. We note that it is possible that shock excited emission is prevalent on smaller scales (a few tenths of a parsec, or $\lesssim$0.5$\arcsec$ at the distance to 30 Doradus) caused by outflows from massive protostars residing within the KN molecular clouds, but our current angular resolution limitations in the nIR narrowband images ($\sim$1$\arcsec$) prevent us from examining the same in detail. Future high angular resolution integral field unit (IFU) nIR spectroscopy would help towards constraining this further, as shocked gas will likely be offset in velocity.
A picture of ionization fronts emerges, with the photodissociation region extincted from our line of sight by the cold molecular gas. These structures could resemble the dense “pillars” or structures observed in the galaxy in regions such as M16 and NGC3603 (Sankrit & Hester 2000), but are smaller in scale at $\sim$0.1–0.3pc, and in the 30 Doradus Nebula by Pellegrini et al. (2010). From the observed emission line imaging and CO(2-1) observations it appears that the clouds are being ionised on the backside. We are viewing the KN molecular clouds face on, and they are likely the tail of pillar-like structures (we refer the reader to Pound 1998 for a description of pillar morphology; or to Fig.7) with the ionized head pointing towards R136. The observed velocity line widths of the CO(2-1) line are then likely caused due to the velocity gradient between the head and tail of pillars (e.g. Pound 1998). Following the Bertoldi (1989) analytical theory of photoevaporating clouds, during photoevaporation clouds form a cometary structure similar to pillars. Neutral gas at the head is pushed back by the ionization front stripping the outer envelope. The difference in velocity between the slowly moving head to the tail with respect to ionizing stars develops a velocity gradient, that is manifested as velocity line width seen in Fig.\[fig:slw\], also providing a natural explanation for the observed molecular cloud properties. The observed velocity line widths is in a similar range to those observed in M16 using CO observations (Pound 1998). Our results are well supported by the work presented in Chevance et al. (2016). The stapler nebula is prominent in forbidden mIR and fIR line emission from \[S[II]{}\], \[C[II]{}\], and \[O[I]{}\] (Fig. 1 in Chevance et al. 2016), which are key tracers of photodissociation regions.
Another interesting morphological observation can be made from examining the mIR imaging in Figure \[fig:ysothumb\]e. The 8$\mu$m image from [*Spitzer*]{} covers the 8.6$\mu$m Polycyclic Aromatic Hydrocarbon emission (PAH). PAHs are considered to be released from the dust exposed to UV photons in photodissociation regions, and are altered/destroyed in these energetic environments. The northern boundary of the KN2 and KN3 cloud is bright in 8$\mu$m suggesting the PAH emission occurs at this interface where the UV photons ionise the molecular cloud. In contrast, the KN1 cloud is being ionised from the south directly facing towards R136 following the strong 8$\mu$m emission.
In contrast, the 24$\mu$m imaging (Fig.\[fig:ysothumb\]f) is faint across the KN2 and KN3 cloud region, but displays a wispy structure at KN1. The 24$\mu$m emission is emitted from intermediate to larger size grains. This emission is coincident with early type stars to its immediate east, whose measured median reddening law $R_V$ is greater than the canonical average for 30 Doradus and stands $\sim$4.5. This is indicative of selective evaporation of small grains leading to a larger reddening law (e.g. De Marchi et al. 2016), and coincident with the strong 24$\mu$m emission region.
Candidate massive young stellar objects
=======================================
Identification of YSO
---------------------
Our aim is to detect any on-going star formation through the identification of YSOs using archival infrared photometry. The youngest YSOs (Class 0 objects) will only be visible at $\lambda>$10$\mu$m, while older Class I/II sources will be visible at shorter ($>$2$\mu$m) wavelengths. Using [*Spitzer*]{} 3.6-24$\mu$m, and [*Herschel*]{} 100-350$\mu$m photometry (details in Section 2.3; and images are presented in Fig.\[fig:ysothumb\]), we classify point or point-like sources based on either the spectral energy distribution (SED) slope in the 3.6-24$\mu$m wavelength range, or by examining the parameters derived from fitting modified blackbody and YSO models across the entire wavelength range. We require detection in at least three bands in each catalogue to classify the object as a source. We then check for any counterparts with the coordinates of the detection in the [*Herschel*]{} 100$\mu$m within a crossmatch radius of 6.7$''$ (which is the convolution kernel at 100$\mu$m) and the [*Spitzer*]{} coordinates. We find a known [*Spitzer*]{} source in KN2 (KN2-A, or S5 in Walborn et al. 2013), a [*Herschel*]{} source without a [*Spitzer*]{} counterpart in KN2 visible at 100-350$\mu$m (KN2-B), and one [*Herschel*]{} source with a [*Spitzer*]{} counterpart within 0.9$''$ in KN3 (KN3-A). The thumbnails of these sources in the [*Spitzer*]{} and [*Herschel*]{} bands are shown in Fig.\[fig:ysothumb\].
For SED fitting, we utilize three methods to derive the properties of the source. First, we fit the photometry of each source with the YSO models of Robitalle (2017). The model grid covers 20000 radiative transfer YSO models covering a mass range of 0.1–50$M_{\odot}$. It should be noted that the SED fits are not intended to provide accurate parameters, but as a crude guide to the nature of each sources and that these models are limited in nature compared to the available free parameters (Offner et al. 2012). Secondly, we use a modified blackbody fit to the far-IR [*Herschel*]{} sources to constrain the temperature of the emitting source, $T_{\rm b}$. The observed flux can be reproduced as a blackbody with frequency flux density $F_{\nu}$ as $$F_{\nu} =B_{\nu}(T_{\rm b})(1-e{^{-\tau_{\nu}}})\Omega.$$
Here, $B_{\nu}(T_{\rm b})$ is the blackbody emission at $T_{\rm b}$, and $\tau_{\nu}$ is the optical depth constrained by a power law at frequency $\nu$ by $\tau_{\nu} \propto \nu^{\beta}$. $\beta$ = 1.5 in the LMC (Galliano et al. 2011). Thirdly, only for the [*Spitzer*]{} sources we employ the slope of the SED fit ($\alpha_{\rm SED}$), where $$\alpha_{\rm{SED}}\,=\,\frac{d\,{\rm{log}}(\lambda F_{\lambda})}{d\,{\rm{log}}\lambda}.$$ We consider only the flux between the wavelengths 3.6–24.0$\mu$m which classify well Class I/II sources (e.g. Greene et al. 1994). The results of our analysis are presented in Table 2, and described in the following subsections.
{width="2.1\columnwidth"}
Description of YSOs in the clouds
---------------------------------
### [KN1]{} cloud
KN1 contains no identified mid or far infrared source.
### [KN2]{} cloud
KN2 is extremely promising as a future site of high mass star formation. KN2-B is a [*Herschel*]{} source detected in 100–350$\mu$m imaging (marked in Fig.\[fig:ysothumb\]). There exists no near counterpart mIR source. The nearest mIR source is KN2-A which lies to its immediate south west, and at the very edge of the CO emission. This source angularly coincides with the dense region observed in CS to the south of KN2 by Rubio et al. (2009). CS is a tracer of dense gas with densities $n$, upwards of $10^5$cm$^{-3}$.
KN2-B is classified as a high-mass YSO based on the [*Herschel*]{} classification scheme devised by Seale et al. (2014) for YSOs in the Magellanic Clouds. Following their three classification criteria (see Sec. 4.4 of that paper for further details), we find KN2-B meets all three as it is (i) the dominant source of fIR photometry and is clearly defined in at least 3 [*Herschel*]{} bands at 3$\sigma$ above the background (ii) It is not identified as a background galaxy/interloper (iii) it has 24$\mu$m emission as seen in Fig.\[fig:ysothumb\]. The source has marked PolyAromaticHydrocarbon (PAH) emission from the 8.6$\mu$m feature.
We achieve an excellent fit for KN2-B with the SED models of Robitalle et al. (2017). The best fit has a $\chi^2$/datapoint$<$3 and is shown in Fig.\[fig:sed\]. The parameters of the best fit model were a stellar temperature of 18500$\pm$1000 K, which translates to a stellar mass exceeding 20$M_{\odot}$, with the $\dot M_{\rm env}/M_{\ast}$ from the model fit suggestive of a Class 0 massive YSO (Fig.\[fig:sed\]).The temperature estimated from the blackbody fit is 28.8$\pm$4K, with the integrated logarithm of luminosity log$L/L_{\odot}$=4.59 which further indicates a Class 0 classification (Andr[é]{} et al. 2010). The luminosity of the source is much higher than expected for starless clouds heated by the ambient radiation field, considering which it is likely a true high-mass YSO (Seale et al. 2014). The estimated extinction from the SED fit (of $A_V$=6.9) suggests that the candidate protostar is shielded from external photoionization. Moreover, the external heating at wavelengths longwards of 100$\mu$m does not affect significantly the SED (Pavlyuchenkov et al. 2012). Finally, as noted in Sec.5.1, the parameters estimated from the SED fits with the Robitalle (2017) models may be ambiguous due to the limited range of parameters considered, but they do provide a crude guide to source classification. We also stress that KN2-B is classified as a high-mass YSOs based on the independent schemes of Seale et al. (2014) based on source intensity and morphology; and the SED models of Robitalle et al. (2017) based on the flux distribution of the source in the fIR. The age of KN2-B from the model fit is $\lesssim$0.1Myr, and is smaller by an order of magnitude than the crossing timescale of the cloud (the crossing timescale is the length over the velocity) of 0.66Myr.
KN2-A is likely a Class II source based on its [*Spitzer*]{} colours. The $\alpha_{\rm SED}$ slope derived from the 3.6-8$\mu$m photometry is 1.22$\pm$0.2, which falls into the Class II category (Greene et al. 1994). Our classification as a Class II YSO from the SED slope agrees with the Gruendl & Chu (2009) classification based on the position of the KN2-A in [*Spitzer*]{} colour-colour and colour-magnitude diagrams. We note the there is no 24$\mu$m counterpart to KN2-A, and it also does not have a [*Herschel*]{} counterpart according to the procedure followed Seale et al. (2014), who cross match [*Spitzer*]{} and [*Herschel*]{} catalogues in the LMC. They adopt a cross match radius of 0.5$\times$FWHM to cross match the two catalogues in crowded regions. If we still consider that some of the fIR flux from KN2-A is assigned towards KN2-B, we find that $\gtrsim$5% of the fIR flux of KN2-B is required to change the classification of KN2-A from Class II to Class I. There is no high mass Class II YSOs population on the northern side.
ID Wavelength($\mu$m) R.A. (J2000) Dec. (J2000) Class $\alpha_{\rm SED}$ Comments
------- -------------------- --------------------- ---------------------------------- --------- -------------------- ----------------------------------------------------------------
KN2-A 1.2–8 05$^h$38$^m$39$^s$7 $-$69$^{\rm \circ}$05$'$38.1$''$ II 1.22$\pm$0.2 S5 in Walborn et al. (2013)
KN2-B 100-350 05$^h$38$^m$39$^s$1 $-$69$^{\rm \circ}$05$'$33.8$''$ 0 – $T_{\rm b}$=29$\pm$4K; point like in 24$\mu$m imaging
KN3-A 3.6-500 05$^h$38$^m$36$^s$3 $-$69$^{\rm \circ}$05$'$24$''$ Dust(?) 2.65$\pm$0.7 $T_{\rm b}$=31.5$\pm$7K; extended in 24$\mu$m imaging
0.9$''$ between [*Herschel*]{} 100$\mu$m& [*Spitzer*]{} source
\
![SED of KN2-B with best-fit models from Robitalle (2017), overlaid with [*Herschel*]{} 100-350$\mu$m photometry and corresponding error bars. The best fit model (with $\chi^2$/datapoint$<$3) with a stellar temperature of $\sim$18500K is shown as a solid black line. Other models with $\chi^2$/datapoint$<$5 are also shown as solid grey lines.[]{data-label="fig:sed"}](sedA-crop.pdf){width="\columnwidth"}
### KN3
KN3-A is a bright [*Herschel*]{} source and is extended in 100–350$\mu$m imaging. The [*Herschel*]{} detection lies within 0.9$''$ of a Gruendl & Chu (2009) [*Spitzer*]{} source. We consider the mIR and fIR detection to be of the same source, as they lie well within the FWHM of the [*Herschel*]{} and [*Spitzer*]{} imaging. Gruendl & Chu (2009) classified the source as likely originating from dust, and we see that their exists no point-like or otherwise 24$\mu$m counterpart. The whole SED fits rather poorly with the models of Robitalle et al. (2017), with the best fits achieved when discarding the 4.5$\mu$m photometry of $\chi^2$/datapoint $>$100. Examination of the 24$\mu$m imaging reveals no distinct point-like source (Fig.\[fig:ysothumb\]; right dashed circle). Given the flux at long wavelengths, and the visible extended emission in the imaging we cannot classify the source as a YSO conclusively. A modified blackbody fit suggests a peak temperature of 31.5$\pm$6K, with the integrated total luminosity slightly less than 10$^4$$L_{\odot}$. This lies at the boundary of the classifying scheme between interstellar starless clouds heated by the ambient radiation field (Seale et al. 2014), and YSOs. The ambiguous classification, and lack of any clear point-like source in the nIR and fIR images suggests that this source is likely a starless dust cloud, probably heated by the ambient ionizing field.
A comparison of the multi-wavelength images in Fig.\[fig:ysothumb\] demonstrates the differences between KN2-B and KN3-A. At 24$\mu$m, diffuse emission in star-forming regions is likely to originate from heated gas from inner regions of protostars (Chambers et al. 2009). In KN2-B, the 24$\mu$m emission is clearly detected at 3$\sigma$ level with respect to the immediate surroundings. In KN3-A, 24$\mu$m emission above the 3$\sigma$ threshold is not detected. At 100$\mu$m, and 160$\mu$m, both sources appear as point sources. But, at longer wavelengths, the KN3-A source is extended or disappears (see the 250$\mu$m image), suggesting that it is more likely to be dust, rather than a protostellar candidate when compared to the nature of the source detected at the position of KN2-B in the 250$\mu$m image.
A picture of the stapler nebula in R136
=======================================
From our deep CO(2-1) observations, we identified three molecular clouds separated in velocity we name KN1-3 lying within 14pc in projection; and $\lesssim$20pc in the line of sight from the YMC R136. These clouds lie north of R136 and stretch for 14pc in projection from southeast to northwest. This axis is perpendicular to the previously identified giant molecular clouds by Johansson et al. (1998). The $V_{\rm lsr}$ of the clouds ranges from 237kms$^{-1}$ for the cloud directly above R136 in projection to 250kms$^{-1}$ for the furthest cloud, which is similar to the mean radial velocity of the stellar population in R136. These clouds display relatively large linewidths for their radii, lying above the predicted relation for clouds in virial equilibrium according to Larson’s first law. By plotting ${\sigma_{v}}^2$/$r$ of the KN clouds as a function of their $\Sigma_{\rm H_{2}}$, we show that KN clouds depart significantly from the relation for virialized, isolated clouds confined by self-gravity. The KN clouds are most likely confined by external pressures upwards of $\sim$10$^6$cm$^{-3}$K in this scenario. Nonetheless, the results from our CO(2-1) line observations show that there are three molecular clouds lying near R136, whose $V_{\rm lsr}$ increase as a function of projected distance from the cluster. The resolved clouds are not in virial equilibrium, but are highly turbulent requiring external pressure to confine them.
Br$\gamma$ and H$_2$ nIR emission line imaging reveal that the CO clouds coincide with the dense H$_2$ clumpy structures, while Br$\gamma$ is diffuse and exhibits little spatial coincidence with the CO clouds. The lack of background stars at optical wavelengths (Fig.1) suggests that these clouds lie in front of the cluster. Based on the H$_2$/Br$\gamma$ ratio, we suggest that these clouds are UV-heated by stellar radiation emanating from R136. These results suggests that both the H$_2$ nIR emission, and CO emission arise from molecular clouds near R136. The molecular clouds lie in front of R136 from our perspective. The backside of the molecular clouds is being ionised by R136 leading to diffuse Br$\gamma$ emission and weak H$\alpha$ emission. The excited H$_2$ emission coincides with the peak of CO emission further giving weight to our view of photodissociated molecular clouds near R136.
Finally, a search for results of on-going star formation yields interesting results. KN1 shows no signs of active star formation. In fIR [*Herschel*]{} 100-350$\mu$m photometry, we identify a likely Class 0/I object KN2-A corresponding to the peak of the molecular cloud KN2. It is known from previous CS observations that the densities near this source exceed $n>$10$^6$. Near this source (within 5$''$) lies a Class II object KN2-A visible in [*Spitzer*]{} imaging, which has been classified as a YSO by Gruendl & Chu (2009) and Walborn et al. (2013). Towards KN3, we detect a [*Herschel*]{} 100-160$\mu$m source (KN3-A) close to a [*Spitzer*]{} source identified by Gruendl & Chu (2009). However, inspection of both the [*Spitzer*]{} and [*Herschel*]{} imaging suggests KN3-A is likely dusty in nature, which agrees with the Gruendl & Chu (2009) classification.
The complete scenario can be visualised in the toy model shown in Fig.\[fig:model\] along the observer’s viewpoint. The natal molecular cloud of R136 has an arc like structure, protruding in front of it along our line of sight, and lying directly above it and stretching towards the north west in projection forming an arc shaped structure. The structure may have been carved out from an initially spherical cloud. After the formation of the cluster, the natal molecular cloud has been steadily ionised giving rise to excited H$_2$ emission. The excited boundary lies on the backside of the molecular cloud from the observer’s viewpoint. The radiation from R136 is eroding the molecular clouds, possibly leading to the formation of structures similar to the “pillars of creation" observed in galactic regions such as M16. The observer may be visualising the tail of such pillars face on. The observed high velocity line widths could then be explained as a manifestation of the expected velocity differences between the head and tail of a photoevaporating cloud. Alternatively, the results also support the picture that the KN clouds may likely be confined by external pressures. High angular resolution IFU spectroscopy of the clouds which will reveal their gas kinematics and ionization structure (McLeod et al. 2015) can help differentiate between the two proposed scenarios. Massive star formation is occurring in the middle molecular cloud or pillar, while the remaining two molecular clouds show no active signs of massive star formation.
Summary and future work
=======================
Our results present a picture of three CO molecular clouds separated in velocity lying $\lesssim$20pc in front of R136, and between 2–14pc away in projection from the cluster. We appear to view the tail of pillar-like structures whose ionised heads are pointing towards R136. The observed high line widths of the molecular clouds for their sizes with respect to the canonical $\sigma_{v}$-$r$ relation are likely due to the gas being pushed away from the head of the pillar-like structures by the ionization radiation produced by the massive stars in R136, or due to external pressure confining the clouds. A massive YSO (KN2-B) is detected inside the KN2 molecular cloud, indicating active star formation. These results suggest that 1.5–3Myr R136 cluster is in the process eroding the KN molecular clouds. During this process a new generation of stars is able to form from the reservoir of cold molecular gas that is in the process of being destroyed by stellar feedback in structures that may be similar to the pillars of creation seen in Galactic regions such as M16 and NGC3603 from a different perspective.
We have demonstrated that with sensitive CO(2-1) observations, molecular clouds of $\sim$10$^4$$M_{\odot}$ exist surprisingly close to the YMC R136. This observation challenges the results of most simulations of feedback from massive stars in YMC (see Dale 2015 for a review), where feedback is able to successfully expel gas close to YMCs within a few Myrs. Future high angular resolution sensitive observations of multiple molecular lines using the current generation of interferometers (for e.g. using the Atacama Large Millimeter/submillimeter Array (ALMA) which can achieve a resolution comparable to optical/nIR imaging in CO lines) may reveal in exquisite detail the structure, and densities of these clouds. Combined with observations of the photodissociation region from fine structure lines (Chevance et al. 2016), one can gather enough information about the densities, pressures, and radiation field acting on the KN clouds and the H[II]{} region to predict if the clouds are dense and massive enough to survive photoionisation and form a new massive stellar cluster, or whether star formation will be abruptly terminated due to feedback. In a broader sense, sensitive radio observations in CO lines may reveal previously undetected molecular clouds lying close to YMCs, that may nurture a new generation of star formation.
The authors thank Jes[ús]{} Ma[í]{}z Apell[á]{}niz for kindly providing us with the [*HST*]{} mosaic, and sharing with us a pictorial etymology of the stapler nebula. We thank Sherry Yeh for generously providing her Br$\gamma$ image, and Hugo Saldano for his help in visualizing the radio data. The anonymous referee is thanked for providing constructive comments, and in helping to improve the overall impact of the paper. V.M.K. acknowledges support from the FONDECYT-CHILE Fellowship grant N$^{\rm o}$3116017. M.R. acknowledges support from CONICYT (CHILE) through FONDECYT grant N$^{\rm o}$1140839, and partial support through BASALPFB-06. This work made use of the CLASS software, and the Python packages ApLpy, Numpy, Scipy and Matplotlib for analysis and presentation.
Anderson, C. N., Meier, D. S., Ott, J., et al. 2014, , 793, 37 Andr[é]{}, P., Men’shchikov, A., Bontemps, S., et al. 2010, , 518, L102 Ballesteros-Paredes, J., Hartmann, L. W., V[á]{}zquez-Semadeni, E., Heitsch, F., & Zamora-Avil[é]{}s, M. A. 2011, , 411, 65 Bertoldi, F. 1989, , 346, 735 Bolatto, A. D., Wolfire, M., & Leroy, A. K. 2013, , 51, 207 Bolatto, A. D., Leroy, A. K., Rosolowsky, E., Walter, F., & Blitz, L. 2008, , 686, 948-965 Brandner, W., Grebel, E. K., Barb[á]{}, R. H., Walborn, N. R., & Moneti, A. 2001, , 122, 858 Chambers, E. T., Jackson, J. M., Rathborne, J. M., & Simon, R. 2009, , 181, 360 Chevance, M., Madden, S. C., Lebouteiller, V., et al. 2016, , 590, A36 Dale, J. E. 2015, , 68, 1 Dale, J. E., Ercolano, B., & Bonnell, I. A. 2012, , 427, 2852 De Marchi, G., Panagia, N., Romaniello, M., et al. 2011, , 740, 11 Evans, C. J., Kennedy, M. B., Dufton, P. L., et al. 2015, , 574, A13 Field, G. B., Blackman, E. G., & Keto, E. R. 2011, , 416, 710 Galliano, F., Hony, S., Bernard, J.-P., et al. 2011, , 536, A88 Greene, T. P., Wilking, B. A., Andre, P., Young, E. T., & Lada, C. J. 1994, , 434, 614 Gruendl, R. A., & Chu, Y.-H. 2009, , 184, 172 Guzman, V. V. 2010, MSc. Thesis, Universidad de Chile Herrera, C. N., Rubio, M., Bolatto, A. D., et al. 2013, , 554, A91 Heyer, M., Krawczyk, C., Duval, J., & Jackson, J. M. 2009, , 699, 1092 Indebetouw, R., Brogan, C., Chen, C.-H. R., et al. 2013, , 774, 73 Johansson, L. E. B., Greve, A., Booth, R. S., et al. 1998, , 331, 857 Kalari, V. M., Vink, J. S., Dufton, P. L., et al. 2014, , 564, L7 Larson, R. B. 1981, , 194, 809 Lombardi, M., Alves, J., & Lada, C. J. 2010, , 519, L7 MacLaren, I., Richardson, K. M., & Wolfendale, A. W. 1988, , 333, 821 Maloney, P. 1988, , 334, 761 McKee, C. F., & Ostriker, E. C. 2007, , 45, 565 McLeod, A. F., Dale, J. E., Ginsburg, A., et al. 2015, , 450, 1057 Meixner, M., Panuzzo, P., Roman-Duval, J., et al. 2013, , 146, 62 Meixner, M., Gordon, K. D., Indebetouw, R., et al. 2006, , 132, 2268 Nayak, O., Meixner, M., Indebetouw, R., et al. 2016, , 831, 32 Offner, S. S. R., Robitaille, T. P., Hansen, C. E., McKee, C. F., & Klein, R. I. 2012, , 753, 98 Pavlyuchenkov, Y. N., Wiebe, D. S., Akimkin, V. V., Khramtsova, M. S., & Henning, T. 2012, , 421, 2430 Pellegrini, E. W., Baldwin, J. A., & Ferland, G. J. 2010, , 191, 160 Persson, S. E., Murphy, D. C., Krzeminski, W., Roth, M., & Rieke, M. J. 1998, , 116, 2475 Pietrzy[ń]{}ski, G., Graczyk, D., Gieren, W., et al. 2013, , 495, 76 Pineda, J. L., Ott, J., Klein, U., et al. 2009, , 703, 736 Pound, M. W. 1998, , 493, L113 Robitaille, T. P. 2017, , 600, A11 Rolleston, W. R. J., Trundle, C., & Dufton, P. L. 2002, , 396, 53 Roman-Duval, J., Gordon, K. D., Meixner, M., et al. 2014, , 797, 86 Rubio, M., Paron, S., & Dubner, G. 2009, , 505, 177 Rubio, M., Barb[á]{}, R. H., Walborn, N. R., et al. 1998, , 116, 1708 Salim, D. M., Federrath, C., & Kewley, L. J. 2015, , 806, L36 Sankrit, R., & Hester, J. J. 2000, , 535, 847 Seale, J. P., Meixner, M., Sewi[ł]{}o, M., et al. 2014, , 148, 124 Selman, F. J., & Melnick, J. 2013, , 552, A94 Solomon, P. M., Rivolo, A. R., Barrett, J., & Yahil, A. 1987, , 319, 730 Walborn, N. R., Barb[á]{}, R. H., & Sewi[ł]{}o, M. M. 2013, , 145, 98 Walborn, N. R., Ma[í]{}z-Apell[á]{}niz, J., & Barb[á]{}, R. H. 2002, , 124, 1601 Whitney, B. A., Sewilo, M., Indebetouw, R., et al. 2008, , 136, 18 Wong, T., Hughes, A., Ott, J., et al. 2011, , 197, 16 Yeh, S. C. C., Seaquist, E. R., Matzner, C. D., & Pellegrini, E. W. 2015, , 807, 117
[^1]: http://www.iram.fr/IRAMFR/GILDAS
[^2]: The mean radial velocity of the stars translated to local standard of rest frame is $\approx$255$\pm$5kms$^{-1}$ (Evans et al. 2015)
|
{
"pile_set_name": "ArXiv"
}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.